1 // SPDX-License-Identifier: GPL-2.0
3 * Completely Fair Scheduling (CFS) Class (SCHED_NORMAL/SCHED_BATCH)
5 * Copyright (C) 2007 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
7 * Interactivity improvements by Mike Galbraith
8 * (C) 2007 Mike Galbraith <efault@gmx.de>
10 * Various enhancements by Dmitry Adamushko.
11 * (C) 2007 Dmitry Adamushko <dmitry.adamushko@gmail.com>
13 * Group scheduling enhancements by Srivatsa Vaddagiri
14 * Copyright IBM Corporation, 2007
15 * Author: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
17 * Scaled math optimizations by Thomas Gleixner
18 * Copyright (C) 2007, Thomas Gleixner <tglx@linutronix.de>
20 * Adaptive scheduling granularity, math enhancements by Peter Zijlstra
21 * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra
25 #include <trace/events/sched.h>
28 * Targeted preemption latency for CPU-bound tasks:
30 * NOTE: this latency value is not the same as the concept of
31 * 'timeslice length' - timeslices in CFS are of variable length
32 * and have no persistent notion like in traditional, time-slice
33 * based scheduling concepts.
35 * (to see the precise effective timeslice length of your workload,
36 * run vmstat and monitor the context-switches (cs) field)
38 * (default: 6ms * (1 + ilog(ncpus)), units: nanoseconds)
40 unsigned int sysctl_sched_latency = 6000000ULL;
41 unsigned int normalized_sysctl_sched_latency = 6000000ULL;
44 * The initial- and re-scaling of tunables is configurable
48 * SCHED_TUNABLESCALING_NONE - unscaled, always *1
49 * SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus)
50 * SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus
52 * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus))
54 enum sched_tunable_scaling sysctl_sched_tunable_scaling = SCHED_TUNABLESCALING_LOG;
57 * Minimal preemption granularity for CPU-bound tasks:
59 * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds)
61 unsigned int sysctl_sched_min_granularity = 750000ULL;
62 unsigned int normalized_sysctl_sched_min_granularity = 750000ULL;
65 * This value is kept at sysctl_sched_latency/sysctl_sched_min_granularity
67 static unsigned int sched_nr_latency = 8;
70 * After fork, child runs first. If set to 0 (default) then
71 * parent will (try to) run first.
73 unsigned int sysctl_sched_child_runs_first __read_mostly;
76 * SCHED_OTHER wake-up granularity.
78 * This option delays the preemption effects of decoupled workloads
79 * and reduces their over-scheduling. Synchronous workloads will still
80 * have immediate wakeup/sleep latencies.
82 * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds)
84 unsigned int sysctl_sched_wakeup_granularity = 1000000UL;
85 unsigned int normalized_sysctl_sched_wakeup_granularity = 1000000UL;
87 const_debug unsigned int sysctl_sched_migration_cost = 500000UL;
91 * For asym packing, by default the lower numbered CPU has higher priority.
93 int __weak arch_asym_cpu_priority(int cpu)
99 #ifdef CONFIG_CFS_BANDWIDTH
101 * Amount of runtime to allocate from global (tg) to local (per-cfs_rq) pool
102 * each time a cfs_rq requests quota.
104 * Note: in the case that the slice exceeds the runtime remaining (either due
105 * to consumption or the quota being specified to be smaller than the slice)
106 * we will always only issue the remaining available time.
108 * (default: 5 msec, units: microseconds)
110 unsigned int sysctl_sched_cfs_bandwidth_slice = 5000UL;
114 * The margin used when comparing utilization with CPU capacity:
115 * util * margin < capacity * 1024
119 unsigned int capacity_margin = 1280;
121 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
127 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
133 static inline void update_load_set(struct load_weight *lw, unsigned long w)
140 * Increase the granularity value when there are more CPUs,
141 * because with more CPUs the 'effective latency' as visible
142 * to users decreases. But the relationship is not linear,
143 * so pick a second-best guess by going with the log2 of the
146 * This idea comes from the SD scheduler of Con Kolivas:
148 static unsigned int get_update_sysctl_factor(void)
150 unsigned int cpus = min_t(unsigned int, num_online_cpus(), 8);
153 switch (sysctl_sched_tunable_scaling) {
154 case SCHED_TUNABLESCALING_NONE:
157 case SCHED_TUNABLESCALING_LINEAR:
160 case SCHED_TUNABLESCALING_LOG:
162 factor = 1 + ilog2(cpus);
169 static void update_sysctl(void)
171 unsigned int factor = get_update_sysctl_factor();
173 #define SET_SYSCTL(name) \
174 (sysctl_##name = (factor) * normalized_sysctl_##name)
175 SET_SYSCTL(sched_min_granularity);
176 SET_SYSCTL(sched_latency);
177 SET_SYSCTL(sched_wakeup_granularity);
181 void sched_init_granularity(void)
186 #define WMULT_CONST (~0U)
187 #define WMULT_SHIFT 32
189 static void __update_inv_weight(struct load_weight *lw)
193 if (likely(lw->inv_weight))
196 w = scale_load_down(lw->weight);
198 if (BITS_PER_LONG > 32 && unlikely(w >= WMULT_CONST))
200 else if (unlikely(!w))
201 lw->inv_weight = WMULT_CONST;
203 lw->inv_weight = WMULT_CONST / w;
207 * delta_exec * weight / lw.weight
209 * (delta_exec * (weight * lw->inv_weight)) >> WMULT_SHIFT
211 * Either weight := NICE_0_LOAD and lw \e sched_prio_to_wmult[], in which case
212 * we're guaranteed shift stays positive because inv_weight is guaranteed to
213 * fit 32 bits, and NICE_0_LOAD gives another 10 bits; therefore shift >= 22.
215 * Or, weight =< lw.weight (because lw.weight is the runqueue weight), thus
216 * weight/lw.weight <= 1, and therefore our shift will also be positive.
218 static u64 __calc_delta(u64 delta_exec, unsigned long weight, struct load_weight *lw)
220 u64 fact = scale_load_down(weight);
221 int shift = WMULT_SHIFT;
223 __update_inv_weight(lw);
225 if (unlikely(fact >> 32)) {
232 /* hint to use a 32x32->64 mul */
233 fact = (u64)(u32)fact * lw->inv_weight;
240 return mul_u64_u32_shr(delta_exec, fact, shift);
244 const struct sched_class fair_sched_class;
246 /**************************************************************
247 * CFS operations on generic schedulable entities:
250 #ifdef CONFIG_FAIR_GROUP_SCHED
252 /* cpu runqueue to which this cfs_rq is attached */
253 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
258 static inline struct task_struct *task_of(struct sched_entity *se)
260 SCHED_WARN_ON(!entity_is_task(se));
261 return container_of(se, struct task_struct, se);
264 /* Walk up scheduling entities hierarchy */
265 #define for_each_sched_entity(se) \
266 for (; se; se = se->parent)
268 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
273 /* runqueue on which this entity is (to be) queued */
274 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
279 /* runqueue "owned" by this group */
280 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
285 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
287 if (!cfs_rq->on_list) {
288 struct rq *rq = rq_of(cfs_rq);
289 int cpu = cpu_of(rq);
291 * Ensure we either appear before our parent (if already
292 * enqueued) or force our parent to appear after us when it is
293 * enqueued. The fact that we always enqueue bottom-up
294 * reduces this to two cases and a special case for the root
295 * cfs_rq. Furthermore, it also means that we will always reset
296 * tmp_alone_branch either when the branch is connected
297 * to a tree or when we reach the beg of the tree
299 if (cfs_rq->tg->parent &&
300 cfs_rq->tg->parent->cfs_rq[cpu]->on_list) {
302 * If parent is already on the list, we add the child
303 * just before. Thanks to circular linked property of
304 * the list, this means to put the child at the tail
305 * of the list that starts by parent.
307 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
308 &(cfs_rq->tg->parent->cfs_rq[cpu]->leaf_cfs_rq_list));
310 * The branch is now connected to its tree so we can
311 * reset tmp_alone_branch to the beginning of the
314 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
315 } else if (!cfs_rq->tg->parent) {
317 * cfs rq without parent should be put
318 * at the tail of the list.
320 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
321 &rq->leaf_cfs_rq_list);
323 * We have reach the beg of a tree so we can reset
324 * tmp_alone_branch to the beginning of the list.
326 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
329 * The parent has not already been added so we want to
330 * make sure that it will be put after us.
331 * tmp_alone_branch points to the beg of the branch
332 * where we will add parent.
334 list_add_rcu(&cfs_rq->leaf_cfs_rq_list,
335 rq->tmp_alone_branch);
337 * update tmp_alone_branch to points to the new beg
340 rq->tmp_alone_branch = &cfs_rq->leaf_cfs_rq_list;
347 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
349 if (cfs_rq->on_list) {
350 list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
355 /* Iterate thr' all leaf cfs_rq's on a runqueue */
356 #define for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) \
357 list_for_each_entry_safe(cfs_rq, pos, &rq->leaf_cfs_rq_list, \
360 /* Do the two (enqueued) entities belong to the same group ? */
361 static inline struct cfs_rq *
362 is_same_group(struct sched_entity *se, struct sched_entity *pse)
364 if (se->cfs_rq == pse->cfs_rq)
370 static inline struct sched_entity *parent_entity(struct sched_entity *se)
376 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
378 int se_depth, pse_depth;
381 * preemption test can be made between sibling entities who are in the
382 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
383 * both tasks until we find their ancestors who are siblings of common
387 /* First walk up until both entities are at same depth */
388 se_depth = (*se)->depth;
389 pse_depth = (*pse)->depth;
391 while (se_depth > pse_depth) {
393 *se = parent_entity(*se);
396 while (pse_depth > se_depth) {
398 *pse = parent_entity(*pse);
401 while (!is_same_group(*se, *pse)) {
402 *se = parent_entity(*se);
403 *pse = parent_entity(*pse);
407 #else /* !CONFIG_FAIR_GROUP_SCHED */
409 static inline struct task_struct *task_of(struct sched_entity *se)
411 return container_of(se, struct task_struct, se);
414 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
416 return container_of(cfs_rq, struct rq, cfs);
420 #define for_each_sched_entity(se) \
421 for (; se; se = NULL)
423 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
425 return &task_rq(p)->cfs;
428 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
430 struct task_struct *p = task_of(se);
431 struct rq *rq = task_rq(p);
436 /* runqueue "owned" by this group */
437 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
442 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
446 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
450 #define for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) \
451 for (cfs_rq = &rq->cfs, pos = NULL; cfs_rq; cfs_rq = pos)
453 static inline struct sched_entity *parent_entity(struct sched_entity *se)
459 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
463 #endif /* CONFIG_FAIR_GROUP_SCHED */
465 static __always_inline
466 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec);
468 /**************************************************************
469 * Scheduling class tree data structure manipulation methods:
472 static inline u64 max_vruntime(u64 max_vruntime, u64 vruntime)
474 s64 delta = (s64)(vruntime - max_vruntime);
476 max_vruntime = vruntime;
481 static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime)
483 s64 delta = (s64)(vruntime - min_vruntime);
485 min_vruntime = vruntime;
490 static inline int entity_before(struct sched_entity *a,
491 struct sched_entity *b)
493 return (s64)(a->vruntime - b->vruntime) < 0;
496 static void update_min_vruntime(struct cfs_rq *cfs_rq)
498 struct sched_entity *curr = cfs_rq->curr;
499 struct rb_node *leftmost = rb_first_cached(&cfs_rq->tasks_timeline);
501 u64 vruntime = cfs_rq->min_vruntime;
505 vruntime = curr->vruntime;
510 if (leftmost) { /* non-empty tree */
511 struct sched_entity *se;
512 se = rb_entry(leftmost, struct sched_entity, run_node);
515 vruntime = se->vruntime;
517 vruntime = min_vruntime(vruntime, se->vruntime);
520 /* ensure we never gain time by being placed backwards. */
521 cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime);
524 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
529 * Enqueue an entity into the rb-tree:
531 static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
533 struct rb_node **link = &cfs_rq->tasks_timeline.rb_root.rb_node;
534 struct rb_node *parent = NULL;
535 struct sched_entity *entry;
536 bool leftmost = true;
539 * Find the right place in the rbtree:
543 entry = rb_entry(parent, struct sched_entity, run_node);
545 * We dont care about collisions. Nodes with
546 * the same key stay together.
548 if (entity_before(se, entry)) {
549 link = &parent->rb_left;
551 link = &parent->rb_right;
556 rb_link_node(&se->run_node, parent, link);
557 rb_insert_color_cached(&se->run_node,
558 &cfs_rq->tasks_timeline, leftmost);
561 static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
563 rb_erase_cached(&se->run_node, &cfs_rq->tasks_timeline);
566 struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq)
568 struct rb_node *left = rb_first_cached(&cfs_rq->tasks_timeline);
573 return rb_entry(left, struct sched_entity, run_node);
576 static struct sched_entity *__pick_next_entity(struct sched_entity *se)
578 struct rb_node *next = rb_next(&se->run_node);
583 return rb_entry(next, struct sched_entity, run_node);
586 #ifdef CONFIG_SCHED_DEBUG
587 struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
589 struct rb_node *last = rb_last(&cfs_rq->tasks_timeline.rb_root);
594 return rb_entry(last, struct sched_entity, run_node);
597 /**************************************************************
598 * Scheduling class statistics methods:
601 int sched_proc_update_handler(struct ctl_table *table, int write,
602 void __user *buffer, size_t *lenp,
605 int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
606 unsigned int factor = get_update_sysctl_factor();
611 sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency,
612 sysctl_sched_min_granularity);
614 #define WRT_SYSCTL(name) \
615 (normalized_sysctl_##name = sysctl_##name / (factor))
616 WRT_SYSCTL(sched_min_granularity);
617 WRT_SYSCTL(sched_latency);
618 WRT_SYSCTL(sched_wakeup_granularity);
628 static inline u64 calc_delta_fair(u64 delta, struct sched_entity *se)
630 if (unlikely(se->load.weight != NICE_0_LOAD))
631 delta = __calc_delta(delta, NICE_0_LOAD, &se->load);
637 * The idea is to set a period in which each task runs once.
639 * When there are too many tasks (sched_nr_latency) we have to stretch
640 * this period because otherwise the slices get too small.
642 * p = (nr <= nl) ? l : l*nr/nl
644 static u64 __sched_period(unsigned long nr_running)
646 if (unlikely(nr_running > sched_nr_latency))
647 return nr_running * sysctl_sched_min_granularity;
649 return sysctl_sched_latency;
653 * We calculate the wall-time slice from the period by taking a part
654 * proportional to the weight.
658 static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
660 u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq);
662 for_each_sched_entity(se) {
663 struct load_weight *load;
664 struct load_weight lw;
666 cfs_rq = cfs_rq_of(se);
667 load = &cfs_rq->load;
669 if (unlikely(!se->on_rq)) {
672 update_load_add(&lw, se->load.weight);
675 slice = __calc_delta(slice, se->load.weight, load);
681 * We calculate the vruntime slice of a to-be-inserted task.
685 static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se)
687 return calc_delta_fair(sched_slice(cfs_rq, se), se);
692 #include "sched-pelt.h"
694 static int select_idle_sibling(struct task_struct *p, int prev_cpu, int cpu);
695 static unsigned long task_h_load(struct task_struct *p);
697 /* Give new sched_entity start runnable values to heavy its load in infant time */
698 void init_entity_runnable_average(struct sched_entity *se)
700 struct sched_avg *sa = &se->avg;
702 memset(sa, 0, sizeof(*sa));
705 * Tasks are intialized with full load to be seen as heavy tasks until
706 * they get a chance to stabilize to their real load level.
707 * Group entities are intialized with zero load to reflect the fact that
708 * nothing has been attached to the task group yet.
710 if (entity_is_task(se))
711 sa->runnable_load_avg = sa->load_avg = scale_load_down(se->load.weight);
713 se->runnable_weight = se->load.weight;
715 /* when this task enqueue'ed, it will contribute to its cfs_rq's load_avg */
718 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq);
719 static void attach_entity_cfs_rq(struct sched_entity *se);
722 * With new tasks being created, their initial util_avgs are extrapolated
723 * based on the cfs_rq's current util_avg:
725 * util_avg = cfs_rq->util_avg / (cfs_rq->load_avg + 1) * se.load.weight
727 * However, in many cases, the above util_avg does not give a desired
728 * value. Moreover, the sum of the util_avgs may be divergent, such
729 * as when the series is a harmonic series.
731 * To solve this problem, we also cap the util_avg of successive tasks to
732 * only 1/2 of the left utilization budget:
734 * util_avg_cap = (cpu_scale - cfs_rq->avg.util_avg) / 2^n
736 * where n denotes the nth task and cpu_scale the CPU capacity.
738 * For example, for a CPU with 1024 of capacity, a simplest series from
739 * the beginning would be like:
741 * task util_avg: 512, 256, 128, 64, 32, 16, 8, ...
742 * cfs_rq util_avg: 512, 768, 896, 960, 992, 1008, 1016, ...
744 * Finally, that extrapolated util_avg is clamped to the cap (util_avg_cap)
745 * if util_avg > util_avg_cap.
747 void post_init_entity_util_avg(struct sched_entity *se)
749 struct cfs_rq *cfs_rq = cfs_rq_of(se);
750 struct sched_avg *sa = &se->avg;
751 long cpu_scale = arch_scale_cpu_capacity(NULL, cpu_of(rq_of(cfs_rq)));
752 long cap = (long)(cpu_scale - cfs_rq->avg.util_avg) / 2;
755 if (cfs_rq->avg.util_avg != 0) {
756 sa->util_avg = cfs_rq->avg.util_avg * se->load.weight;
757 sa->util_avg /= (cfs_rq->avg.load_avg + 1);
759 if (sa->util_avg > cap)
766 if (entity_is_task(se)) {
767 struct task_struct *p = task_of(se);
768 if (p->sched_class != &fair_sched_class) {
770 * For !fair tasks do:
772 update_cfs_rq_load_avg(now, cfs_rq);
773 attach_entity_load_avg(cfs_rq, se, 0);
774 switched_from_fair(rq, p);
776 * such that the next switched_to_fair() has the
779 se->avg.last_update_time = cfs_rq_clock_task(cfs_rq);
784 attach_entity_cfs_rq(se);
787 #else /* !CONFIG_SMP */
788 void init_entity_runnable_average(struct sched_entity *se)
791 void post_init_entity_util_avg(struct sched_entity *se)
794 static void update_tg_load_avg(struct cfs_rq *cfs_rq, int force)
797 #endif /* CONFIG_SMP */
800 * Update the current task's runtime statistics.
802 static void update_curr(struct cfs_rq *cfs_rq)
804 struct sched_entity *curr = cfs_rq->curr;
805 u64 now = rq_clock_task(rq_of(cfs_rq));
811 delta_exec = now - curr->exec_start;
812 if (unlikely((s64)delta_exec <= 0))
815 curr->exec_start = now;
817 schedstat_set(curr->statistics.exec_max,
818 max(delta_exec, curr->statistics.exec_max));
820 curr->sum_exec_runtime += delta_exec;
821 schedstat_add(cfs_rq->exec_clock, delta_exec);
823 curr->vruntime += calc_delta_fair(delta_exec, curr);
824 update_min_vruntime(cfs_rq);
826 if (entity_is_task(curr)) {
827 struct task_struct *curtask = task_of(curr);
829 trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime);
830 cgroup_account_cputime(curtask, delta_exec);
831 account_group_exec_runtime(curtask, delta_exec);
834 account_cfs_rq_runtime(cfs_rq, delta_exec);
837 static void update_curr_fair(struct rq *rq)
839 update_curr(cfs_rq_of(&rq->curr->se));
843 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
845 u64 wait_start, prev_wait_start;
847 if (!schedstat_enabled())
850 wait_start = rq_clock(rq_of(cfs_rq));
851 prev_wait_start = schedstat_val(se->statistics.wait_start);
853 if (entity_is_task(se) && task_on_rq_migrating(task_of(se)) &&
854 likely(wait_start > prev_wait_start))
855 wait_start -= prev_wait_start;
857 __schedstat_set(se->statistics.wait_start, wait_start);
861 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
863 struct task_struct *p;
866 if (!schedstat_enabled())
869 delta = rq_clock(rq_of(cfs_rq)) - schedstat_val(se->statistics.wait_start);
871 if (entity_is_task(se)) {
873 if (task_on_rq_migrating(p)) {
875 * Preserve migrating task's wait time so wait_start
876 * time stamp can be adjusted to accumulate wait time
877 * prior to migration.
879 __schedstat_set(se->statistics.wait_start, delta);
882 trace_sched_stat_wait(p, delta);
885 __schedstat_set(se->statistics.wait_max,
886 max(schedstat_val(se->statistics.wait_max), delta));
887 __schedstat_inc(se->statistics.wait_count);
888 __schedstat_add(se->statistics.wait_sum, delta);
889 __schedstat_set(se->statistics.wait_start, 0);
893 update_stats_enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
895 struct task_struct *tsk = NULL;
896 u64 sleep_start, block_start;
898 if (!schedstat_enabled())
901 sleep_start = schedstat_val(se->statistics.sleep_start);
902 block_start = schedstat_val(se->statistics.block_start);
904 if (entity_is_task(se))
908 u64 delta = rq_clock(rq_of(cfs_rq)) - sleep_start;
913 if (unlikely(delta > schedstat_val(se->statistics.sleep_max)))
914 __schedstat_set(se->statistics.sleep_max, delta);
916 __schedstat_set(se->statistics.sleep_start, 0);
917 __schedstat_add(se->statistics.sum_sleep_runtime, delta);
920 account_scheduler_latency(tsk, delta >> 10, 1);
921 trace_sched_stat_sleep(tsk, delta);
925 u64 delta = rq_clock(rq_of(cfs_rq)) - block_start;
930 if (unlikely(delta > schedstat_val(se->statistics.block_max)))
931 __schedstat_set(se->statistics.block_max, delta);
933 __schedstat_set(se->statistics.block_start, 0);
934 __schedstat_add(se->statistics.sum_sleep_runtime, delta);
937 if (tsk->in_iowait) {
938 __schedstat_add(se->statistics.iowait_sum, delta);
939 __schedstat_inc(se->statistics.iowait_count);
940 trace_sched_stat_iowait(tsk, delta);
943 trace_sched_stat_blocked(tsk, delta);
946 * Blocking time is in units of nanosecs, so shift by
947 * 20 to get a milliseconds-range estimation of the
948 * amount of time that the task spent sleeping:
950 if (unlikely(prof_on == SLEEP_PROFILING)) {
951 profile_hits(SLEEP_PROFILING,
952 (void *)get_wchan(tsk),
955 account_scheduler_latency(tsk, delta >> 10, 0);
961 * Task is being enqueued - update stats:
964 update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
966 if (!schedstat_enabled())
970 * Are we enqueueing a waiting task? (for current tasks
971 * a dequeue/enqueue event is a NOP)
973 if (se != cfs_rq->curr)
974 update_stats_wait_start(cfs_rq, se);
976 if (flags & ENQUEUE_WAKEUP)
977 update_stats_enqueue_sleeper(cfs_rq, se);
981 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
984 if (!schedstat_enabled())
988 * Mark the end of the wait period if dequeueing a
991 if (se != cfs_rq->curr)
992 update_stats_wait_end(cfs_rq, se);
994 if ((flags & DEQUEUE_SLEEP) && entity_is_task(se)) {
995 struct task_struct *tsk = task_of(se);
997 if (tsk->state & TASK_INTERRUPTIBLE)
998 __schedstat_set(se->statistics.sleep_start,
999 rq_clock(rq_of(cfs_rq)));
1000 if (tsk->state & TASK_UNINTERRUPTIBLE)
1001 __schedstat_set(se->statistics.block_start,
1002 rq_clock(rq_of(cfs_rq)));
1007 * We are picking a new current task - update its stats:
1010 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
1013 * We are starting a new run period:
1015 se->exec_start = rq_clock_task(rq_of(cfs_rq));
1018 /**************************************************
1019 * Scheduling class queueing methods:
1022 #ifdef CONFIG_NUMA_BALANCING
1024 * Approximate time to scan a full NUMA task in ms. The task scan period is
1025 * calculated based on the tasks virtual memory size and
1026 * numa_balancing_scan_size.
1028 unsigned int sysctl_numa_balancing_scan_period_min = 1000;
1029 unsigned int sysctl_numa_balancing_scan_period_max = 60000;
1031 /* Portion of address space to scan in MB */
1032 unsigned int sysctl_numa_balancing_scan_size = 256;
1034 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
1035 unsigned int sysctl_numa_balancing_scan_delay = 1000;
1040 spinlock_t lock; /* nr_tasks, tasks */
1045 struct rcu_head rcu;
1046 unsigned long total_faults;
1047 unsigned long max_faults_cpu;
1049 * Faults_cpu is used to decide whether memory should move
1050 * towards the CPU. As a consequence, these stats are weighted
1051 * more by CPU use than by memory faults.
1053 unsigned long *faults_cpu;
1054 unsigned long faults[0];
1057 static inline unsigned long group_faults_priv(struct numa_group *ng);
1058 static inline unsigned long group_faults_shared(struct numa_group *ng);
1060 static unsigned int task_nr_scan_windows(struct task_struct *p)
1062 unsigned long rss = 0;
1063 unsigned long nr_scan_pages;
1066 * Calculations based on RSS as non-present and empty pages are skipped
1067 * by the PTE scanner and NUMA hinting faults should be trapped based
1070 nr_scan_pages = sysctl_numa_balancing_scan_size << (20 - PAGE_SHIFT);
1071 rss = get_mm_rss(p->mm);
1073 rss = nr_scan_pages;
1075 rss = round_up(rss, nr_scan_pages);
1076 return rss / nr_scan_pages;
1079 /* For sanitys sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */
1080 #define MAX_SCAN_WINDOW 2560
1082 static unsigned int task_scan_min(struct task_struct *p)
1084 unsigned int scan_size = READ_ONCE(sysctl_numa_balancing_scan_size);
1085 unsigned int scan, floor;
1086 unsigned int windows = 1;
1088 if (scan_size < MAX_SCAN_WINDOW)
1089 windows = MAX_SCAN_WINDOW / scan_size;
1090 floor = 1000 / windows;
1092 scan = sysctl_numa_balancing_scan_period_min / task_nr_scan_windows(p);
1093 return max_t(unsigned int, floor, scan);
1096 static unsigned int task_scan_start(struct task_struct *p)
1098 unsigned long smin = task_scan_min(p);
1099 unsigned long period = smin;
1101 /* Scale the maximum scan period with the amount of shared memory. */
1102 if (p->numa_group) {
1103 struct numa_group *ng = p->numa_group;
1104 unsigned long shared = group_faults_shared(ng);
1105 unsigned long private = group_faults_priv(ng);
1107 period *= atomic_read(&ng->refcount);
1108 period *= shared + 1;
1109 period /= private + shared + 1;
1112 return max(smin, period);
1115 static unsigned int task_scan_max(struct task_struct *p)
1117 unsigned long smin = task_scan_min(p);
1120 /* Watch for min being lower than max due to floor calculations */
1121 smax = sysctl_numa_balancing_scan_period_max / task_nr_scan_windows(p);
1123 /* Scale the maximum scan period with the amount of shared memory. */
1124 if (p->numa_group) {
1125 struct numa_group *ng = p->numa_group;
1126 unsigned long shared = group_faults_shared(ng);
1127 unsigned long private = group_faults_priv(ng);
1128 unsigned long period = smax;
1130 period *= atomic_read(&ng->refcount);
1131 period *= shared + 1;
1132 period /= private + shared + 1;
1134 smax = max(smax, period);
1137 return max(smin, smax);
1140 void init_numa_balancing(unsigned long clone_flags, struct task_struct *p)
1143 struct mm_struct *mm = p->mm;
1146 mm_users = atomic_read(&mm->mm_users);
1147 if (mm_users == 1) {
1148 mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
1149 mm->numa_scan_seq = 0;
1153 p->numa_scan_seq = mm ? mm->numa_scan_seq : 0;
1154 p->numa_scan_period = sysctl_numa_balancing_scan_delay;
1155 p->numa_work.next = &p->numa_work;
1156 p->numa_faults = NULL;
1157 p->numa_group = NULL;
1158 p->last_task_numa_placement = 0;
1159 p->last_sum_exec_runtime = 0;
1161 /* New address space, reset the preferred nid */
1162 if (!(clone_flags & CLONE_VM)) {
1163 p->numa_preferred_nid = -1;
1168 * New thread, keep existing numa_preferred_nid which should be copied
1169 * already by arch_dup_task_struct but stagger when scans start.
1174 delay = min_t(unsigned int, task_scan_max(current),
1175 current->numa_scan_period * mm_users * NSEC_PER_MSEC);
1176 delay += 2 * TICK_NSEC;
1177 p->node_stamp = delay;
1181 static void account_numa_enqueue(struct rq *rq, struct task_struct *p)
1183 rq->nr_numa_running += (p->numa_preferred_nid != -1);
1184 rq->nr_preferred_running += (p->numa_preferred_nid == task_node(p));
1187 static void account_numa_dequeue(struct rq *rq, struct task_struct *p)
1189 rq->nr_numa_running -= (p->numa_preferred_nid != -1);
1190 rq->nr_preferred_running -= (p->numa_preferred_nid == task_node(p));
1193 /* Shared or private faults. */
1194 #define NR_NUMA_HINT_FAULT_TYPES 2
1196 /* Memory and CPU locality */
1197 #define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2)
1199 /* Averaged statistics, and temporary buffers. */
1200 #define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2)
1202 pid_t task_numa_group_id(struct task_struct *p)
1204 return p->numa_group ? p->numa_group->gid : 0;
1208 * The averaged statistics, shared & private, memory & CPU,
1209 * occupy the first half of the array. The second half of the
1210 * array is for current counters, which are averaged into the
1211 * first set by task_numa_placement.
1213 static inline int task_faults_idx(enum numa_faults_stats s, int nid, int priv)
1215 return NR_NUMA_HINT_FAULT_TYPES * (s * nr_node_ids + nid) + priv;
1218 static inline unsigned long task_faults(struct task_struct *p, int nid)
1220 if (!p->numa_faults)
1223 return p->numa_faults[task_faults_idx(NUMA_MEM, nid, 0)] +
1224 p->numa_faults[task_faults_idx(NUMA_MEM, nid, 1)];
1227 static inline unsigned long group_faults(struct task_struct *p, int nid)
1232 return p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 0)] +
1233 p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 1)];
1236 static inline unsigned long group_faults_cpu(struct numa_group *group, int nid)
1238 return group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 0)] +
1239 group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 1)];
1242 static inline unsigned long group_faults_priv(struct numa_group *ng)
1244 unsigned long faults = 0;
1247 for_each_online_node(node) {
1248 faults += ng->faults[task_faults_idx(NUMA_MEM, node, 1)];
1254 static inline unsigned long group_faults_shared(struct numa_group *ng)
1256 unsigned long faults = 0;
1259 for_each_online_node(node) {
1260 faults += ng->faults[task_faults_idx(NUMA_MEM, node, 0)];
1267 * A node triggering more than 1/3 as many NUMA faults as the maximum is
1268 * considered part of a numa group's pseudo-interleaving set. Migrations
1269 * between these nodes are slowed down, to allow things to settle down.
1271 #define ACTIVE_NODE_FRACTION 3
1273 static bool numa_is_active_node(int nid, struct numa_group *ng)
1275 return group_faults_cpu(ng, nid) * ACTIVE_NODE_FRACTION > ng->max_faults_cpu;
1278 /* Handle placement on systems where not all nodes are directly connected. */
1279 static unsigned long score_nearby_nodes(struct task_struct *p, int nid,
1280 int maxdist, bool task)
1282 unsigned long score = 0;
1286 * All nodes are directly connected, and the same distance
1287 * from each other. No need for fancy placement algorithms.
1289 if (sched_numa_topology_type == NUMA_DIRECT)
1293 * This code is called for each node, introducing N^2 complexity,
1294 * which should be ok given the number of nodes rarely exceeds 8.
1296 for_each_online_node(node) {
1297 unsigned long faults;
1298 int dist = node_distance(nid, node);
1301 * The furthest away nodes in the system are not interesting
1302 * for placement; nid was already counted.
1304 if (dist == sched_max_numa_distance || node == nid)
1308 * On systems with a backplane NUMA topology, compare groups
1309 * of nodes, and move tasks towards the group with the most
1310 * memory accesses. When comparing two nodes at distance
1311 * "hoplimit", only nodes closer by than "hoplimit" are part
1312 * of each group. Skip other nodes.
1314 if (sched_numa_topology_type == NUMA_BACKPLANE &&
1318 /* Add up the faults from nearby nodes. */
1320 faults = task_faults(p, node);
1322 faults = group_faults(p, node);
1325 * On systems with a glueless mesh NUMA topology, there are
1326 * no fixed "groups of nodes". Instead, nodes that are not
1327 * directly connected bounce traffic through intermediate
1328 * nodes; a numa_group can occupy any set of nodes.
1329 * The further away a node is, the less the faults count.
1330 * This seems to result in good task placement.
1332 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
1333 faults *= (sched_max_numa_distance - dist);
1334 faults /= (sched_max_numa_distance - LOCAL_DISTANCE);
1344 * These return the fraction of accesses done by a particular task, or
1345 * task group, on a particular numa node. The group weight is given a
1346 * larger multiplier, in order to group tasks together that are almost
1347 * evenly spread out between numa nodes.
1349 static inline unsigned long task_weight(struct task_struct *p, int nid,
1352 unsigned long faults, total_faults;
1354 if (!p->numa_faults)
1357 total_faults = p->total_numa_faults;
1362 faults = task_faults(p, nid);
1363 faults += score_nearby_nodes(p, nid, dist, true);
1365 return 1000 * faults / total_faults;
1368 static inline unsigned long group_weight(struct task_struct *p, int nid,
1371 unsigned long faults, total_faults;
1376 total_faults = p->numa_group->total_faults;
1381 faults = group_faults(p, nid);
1382 faults += score_nearby_nodes(p, nid, dist, false);
1384 return 1000 * faults / total_faults;
1387 bool should_numa_migrate_memory(struct task_struct *p, struct page * page,
1388 int src_nid, int dst_cpu)
1390 struct numa_group *ng = p->numa_group;
1391 int dst_nid = cpu_to_node(dst_cpu);
1392 int last_cpupid, this_cpupid;
1394 this_cpupid = cpu_pid_to_cpupid(dst_cpu, current->pid);
1397 * Multi-stage node selection is used in conjunction with a periodic
1398 * migration fault to build a temporal task<->page relation. By using
1399 * a two-stage filter we remove short/unlikely relations.
1401 * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate
1402 * a task's usage of a particular page (n_p) per total usage of this
1403 * page (n_t) (in a given time-span) to a probability.
1405 * Our periodic faults will sample this probability and getting the
1406 * same result twice in a row, given these samples are fully
1407 * independent, is then given by P(n)^2, provided our sample period
1408 * is sufficiently short compared to the usage pattern.
1410 * This quadric squishes small probabilities, making it less likely we
1411 * act on an unlikely task<->page relation.
1413 last_cpupid = page_cpupid_xchg_last(page, this_cpupid);
1414 if (!cpupid_pid_unset(last_cpupid) &&
1415 cpupid_to_nid(last_cpupid) != dst_nid)
1418 /* Always allow migrate on private faults */
1419 if (cpupid_match_pid(p, last_cpupid))
1422 /* A shared fault, but p->numa_group has not been set up yet. */
1427 * Destination node is much more heavily used than the source
1428 * node? Allow migration.
1430 if (group_faults_cpu(ng, dst_nid) > group_faults_cpu(ng, src_nid) *
1431 ACTIVE_NODE_FRACTION)
1435 * Distribute memory according to CPU & memory use on each node,
1436 * with 3/4 hysteresis to avoid unnecessary memory migrations:
1438 * faults_cpu(dst) 3 faults_cpu(src)
1439 * --------------- * - > ---------------
1440 * faults_mem(dst) 4 faults_mem(src)
1442 return group_faults_cpu(ng, dst_nid) * group_faults(p, src_nid) * 3 >
1443 group_faults_cpu(ng, src_nid) * group_faults(p, dst_nid) * 4;
1446 static unsigned long weighted_cpuload(struct rq *rq);
1447 static unsigned long source_load(int cpu, int type);
1448 static unsigned long target_load(int cpu, int type);
1449 static unsigned long capacity_of(int cpu);
1451 /* Cached statistics for all CPUs within a node */
1455 /* Total compute capacity of CPUs on a node */
1456 unsigned long compute_capacity;
1458 unsigned int nr_running;
1462 * XXX borrowed from update_sg_lb_stats
1464 static void update_numa_stats(struct numa_stats *ns, int nid)
1466 int smt, cpu, cpus = 0;
1467 unsigned long capacity;
1469 memset(ns, 0, sizeof(*ns));
1470 for_each_cpu(cpu, cpumask_of_node(nid)) {
1471 struct rq *rq = cpu_rq(cpu);
1473 ns->nr_running += rq->nr_running;
1474 ns->load += weighted_cpuload(rq);
1475 ns->compute_capacity += capacity_of(cpu);
1481 * If we raced with hotplug and there are no CPUs left in our mask
1482 * the @ns structure is NULL'ed and task_numa_compare() will
1483 * not find this node attractive.
1485 * We'll detect a huge imbalance and bail there.
1490 /* smt := ceil(cpus / capacity), assumes: 1 < smt_power < 2 */
1491 smt = DIV_ROUND_UP(SCHED_CAPACITY_SCALE * cpus, ns->compute_capacity);
1492 capacity = cpus / smt; /* cores */
1494 capacity = min_t(unsigned, capacity,
1495 DIV_ROUND_CLOSEST(ns->compute_capacity, SCHED_CAPACITY_SCALE));
1498 struct task_numa_env {
1499 struct task_struct *p;
1501 int src_cpu, src_nid;
1502 int dst_cpu, dst_nid;
1504 struct numa_stats src_stats, dst_stats;
1509 struct task_struct *best_task;
1514 static void task_numa_assign(struct task_numa_env *env,
1515 struct task_struct *p, long imp)
1517 struct rq *rq = cpu_rq(env->dst_cpu);
1519 /* Bail out if run-queue part of active NUMA balance. */
1520 if (xchg(&rq->numa_migrate_on, 1))
1524 * Clear previous best_cpu/rq numa-migrate flag, since task now
1525 * found a better CPU to move/swap.
1527 if (env->best_cpu != -1) {
1528 rq = cpu_rq(env->best_cpu);
1529 WRITE_ONCE(rq->numa_migrate_on, 0);
1533 put_task_struct(env->best_task);
1538 env->best_imp = imp;
1539 env->best_cpu = env->dst_cpu;
1542 static bool load_too_imbalanced(long src_load, long dst_load,
1543 struct task_numa_env *env)
1546 long orig_src_load, orig_dst_load;
1547 long src_capacity, dst_capacity;
1550 * The load is corrected for the CPU capacity available on each node.
1553 * ------------ vs ---------
1554 * src_capacity dst_capacity
1556 src_capacity = env->src_stats.compute_capacity;
1557 dst_capacity = env->dst_stats.compute_capacity;
1559 imb = abs(dst_load * src_capacity - src_load * dst_capacity);
1561 orig_src_load = env->src_stats.load;
1562 orig_dst_load = env->dst_stats.load;
1564 old_imb = abs(orig_dst_load * src_capacity - orig_src_load * dst_capacity);
1566 /* Would this change make things worse? */
1567 return (imb > old_imb);
1571 * This checks if the overall compute and NUMA accesses of the system would
1572 * be improved if the source tasks was migrated to the target dst_cpu taking
1573 * into account that it might be best if task running on the dst_cpu should
1574 * be exchanged with the source task
1576 static void task_numa_compare(struct task_numa_env *env,
1577 long taskimp, long groupimp, bool maymove)
1579 struct rq *dst_rq = cpu_rq(env->dst_cpu);
1580 struct task_struct *cur;
1581 long src_load, dst_load;
1583 long imp = env->p->numa_group ? groupimp : taskimp;
1585 int dist = env->dist;
1587 if (READ_ONCE(dst_rq->numa_migrate_on))
1591 cur = task_rcu_dereference(&dst_rq->curr);
1592 if (cur && ((cur->flags & PF_EXITING) || is_idle_task(cur)))
1596 * Because we have preemption enabled we can get migrated around and
1597 * end try selecting ourselves (current == env->p) as a swap candidate.
1603 if (maymove || imp > env->best_imp)
1610 * "imp" is the fault differential for the source task between the
1611 * source and destination node. Calculate the total differential for
1612 * the source task and potential destination task. The more negative
1613 * the value is, the more remote accesses that would be expected to
1614 * be incurred if the tasks were swapped.
1616 /* Skip this swap candidate if cannot move to the source cpu */
1617 if (!cpumask_test_cpu(env->src_cpu, &cur->cpus_allowed))
1621 * If dst and source tasks are in the same NUMA group, or not
1622 * in any group then look only at task weights.
1624 if (cur->numa_group == env->p->numa_group) {
1625 imp = taskimp + task_weight(cur, env->src_nid, dist) -
1626 task_weight(cur, env->dst_nid, dist);
1628 * Add some hysteresis to prevent swapping the
1629 * tasks within a group over tiny differences.
1631 if (cur->numa_group)
1635 * Compare the group weights. If a task is all by itself
1636 * (not part of a group), use the task weight instead.
1638 if (cur->numa_group && env->p->numa_group)
1639 imp += group_weight(cur, env->src_nid, dist) -
1640 group_weight(cur, env->dst_nid, dist);
1642 imp += task_weight(cur, env->src_nid, dist) -
1643 task_weight(cur, env->dst_nid, dist);
1646 if (imp <= env->best_imp)
1649 if (maymove && moveimp > imp && moveimp > env->best_imp) {
1656 * In the overloaded case, try and keep the load balanced.
1658 load = task_h_load(env->p) - task_h_load(cur);
1662 dst_load = env->dst_stats.load + load;
1663 src_load = env->src_stats.load - load;
1665 if (load_too_imbalanced(src_load, dst_load, env))
1670 * One idle CPU per node is evaluated for a task numa move.
1671 * Call select_idle_sibling to maybe find a better one.
1675 * select_idle_siblings() uses an per-CPU cpumask that
1676 * can be used from IRQ context.
1678 local_irq_disable();
1679 env->dst_cpu = select_idle_sibling(env->p, env->src_cpu,
1684 task_numa_assign(env, cur, imp);
1689 static void task_numa_find_cpu(struct task_numa_env *env,
1690 long taskimp, long groupimp)
1692 long src_load, dst_load, load;
1693 bool maymove = false;
1696 load = task_h_load(env->p);
1697 dst_load = env->dst_stats.load + load;
1698 src_load = env->src_stats.load - load;
1701 * If the improvement from just moving env->p direction is better
1702 * than swapping tasks around, check if a move is possible.
1704 maymove = !load_too_imbalanced(src_load, dst_load, env);
1706 for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) {
1707 /* Skip this CPU if the source task cannot migrate */
1708 if (!cpumask_test_cpu(cpu, &env->p->cpus_allowed))
1712 task_numa_compare(env, taskimp, groupimp, maymove);
1716 static int task_numa_migrate(struct task_struct *p)
1718 struct task_numa_env env = {
1721 .src_cpu = task_cpu(p),
1722 .src_nid = task_node(p),
1724 .imbalance_pct = 112,
1730 struct sched_domain *sd;
1732 unsigned long taskweight, groupweight;
1734 long taskimp, groupimp;
1737 * Pick the lowest SD_NUMA domain, as that would have the smallest
1738 * imbalance and would be the first to start moving tasks about.
1740 * And we want to avoid any moving of tasks about, as that would create
1741 * random movement of tasks -- counter the numa conditions we're trying
1745 sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu));
1747 env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2;
1751 * Cpusets can break the scheduler domain tree into smaller
1752 * balance domains, some of which do not cross NUMA boundaries.
1753 * Tasks that are "trapped" in such domains cannot be migrated
1754 * elsewhere, so there is no point in (re)trying.
1756 if (unlikely(!sd)) {
1757 sched_setnuma(p, task_node(p));
1761 env.dst_nid = p->numa_preferred_nid;
1762 dist = env.dist = node_distance(env.src_nid, env.dst_nid);
1763 taskweight = task_weight(p, env.src_nid, dist);
1764 groupweight = group_weight(p, env.src_nid, dist);
1765 update_numa_stats(&env.src_stats, env.src_nid);
1766 taskimp = task_weight(p, env.dst_nid, dist) - taskweight;
1767 groupimp = group_weight(p, env.dst_nid, dist) - groupweight;
1768 update_numa_stats(&env.dst_stats, env.dst_nid);
1770 /* Try to find a spot on the preferred nid. */
1771 task_numa_find_cpu(&env, taskimp, groupimp);
1774 * Look at other nodes in these cases:
1775 * - there is no space available on the preferred_nid
1776 * - the task is part of a numa_group that is interleaved across
1777 * multiple NUMA nodes; in order to better consolidate the group,
1778 * we need to check other locations.
1780 if (env.best_cpu == -1 || (p->numa_group && p->numa_group->active_nodes > 1)) {
1781 for_each_online_node(nid) {
1782 if (nid == env.src_nid || nid == p->numa_preferred_nid)
1785 dist = node_distance(env.src_nid, env.dst_nid);
1786 if (sched_numa_topology_type == NUMA_BACKPLANE &&
1788 taskweight = task_weight(p, env.src_nid, dist);
1789 groupweight = group_weight(p, env.src_nid, dist);
1792 /* Only consider nodes where both task and groups benefit */
1793 taskimp = task_weight(p, nid, dist) - taskweight;
1794 groupimp = group_weight(p, nid, dist) - groupweight;
1795 if (taskimp < 0 && groupimp < 0)
1800 update_numa_stats(&env.dst_stats, env.dst_nid);
1801 task_numa_find_cpu(&env, taskimp, groupimp);
1806 * If the task is part of a workload that spans multiple NUMA nodes,
1807 * and is migrating into one of the workload's active nodes, remember
1808 * this node as the task's preferred numa node, so the workload can
1810 * A task that migrated to a second choice node will be better off
1811 * trying for a better one later. Do not set the preferred node here.
1813 if (p->numa_group) {
1814 if (env.best_cpu == -1)
1817 nid = cpu_to_node(env.best_cpu);
1819 if (nid != p->numa_preferred_nid)
1820 sched_setnuma(p, nid);
1823 /* No better CPU than the current one was found. */
1824 if (env.best_cpu == -1)
1828 * Reset the scan period if the task is being rescheduled on an
1829 * alternative node to recheck if the tasks is now properly placed.
1831 p->numa_scan_period = task_scan_start(p);
1833 best_rq = cpu_rq(env.best_cpu);
1834 if (env.best_task == NULL) {
1835 ret = migrate_task_to(p, env.best_cpu);
1836 WRITE_ONCE(best_rq->numa_migrate_on, 0);
1838 trace_sched_stick_numa(p, env.src_cpu, env.best_cpu);
1842 ret = migrate_swap(p, env.best_task, env.best_cpu, env.src_cpu);
1843 WRITE_ONCE(best_rq->numa_migrate_on, 0);
1846 trace_sched_stick_numa(p, env.src_cpu, task_cpu(env.best_task));
1847 put_task_struct(env.best_task);
1851 /* Attempt to migrate a task to a CPU on the preferred node. */
1852 static void numa_migrate_preferred(struct task_struct *p)
1854 unsigned long interval = HZ;
1856 /* This task has no NUMA fault statistics yet */
1857 if (unlikely(p->numa_preferred_nid == -1 || !p->numa_faults))
1860 /* Periodically retry migrating the task to the preferred node */
1861 interval = min(interval, msecs_to_jiffies(p->numa_scan_period) / 16);
1862 p->numa_migrate_retry = jiffies + interval;
1864 /* Success if task is already running on preferred CPU */
1865 if (task_node(p) == p->numa_preferred_nid)
1868 /* Otherwise, try migrate to a CPU on the preferred node */
1869 task_numa_migrate(p);
1873 * Find out how many nodes on the workload is actively running on. Do this by
1874 * tracking the nodes from which NUMA hinting faults are triggered. This can
1875 * be different from the set of nodes where the workload's memory is currently
1878 static void numa_group_count_active_nodes(struct numa_group *numa_group)
1880 unsigned long faults, max_faults = 0;
1881 int nid, active_nodes = 0;
1883 for_each_online_node(nid) {
1884 faults = group_faults_cpu(numa_group, nid);
1885 if (faults > max_faults)
1886 max_faults = faults;
1889 for_each_online_node(nid) {
1890 faults = group_faults_cpu(numa_group, nid);
1891 if (faults * ACTIVE_NODE_FRACTION > max_faults)
1895 numa_group->max_faults_cpu = max_faults;
1896 numa_group->active_nodes = active_nodes;
1900 * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
1901 * increments. The more local the fault statistics are, the higher the scan
1902 * period will be for the next scan window. If local/(local+remote) ratio is
1903 * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS)
1904 * the scan period will decrease. Aim for 70% local accesses.
1906 #define NUMA_PERIOD_SLOTS 10
1907 #define NUMA_PERIOD_THRESHOLD 7
1910 * Increase the scan period (slow down scanning) if the majority of
1911 * our memory is already on our local node, or if the majority of
1912 * the page accesses are shared with other processes.
1913 * Otherwise, decrease the scan period.
1915 static void update_task_scan_period(struct task_struct *p,
1916 unsigned long shared, unsigned long private)
1918 unsigned int period_slot;
1919 int lr_ratio, ps_ratio;
1922 unsigned long remote = p->numa_faults_locality[0];
1923 unsigned long local = p->numa_faults_locality[1];
1926 * If there were no record hinting faults then either the task is
1927 * completely idle or all activity is areas that are not of interest
1928 * to automatic numa balancing. Related to that, if there were failed
1929 * migration then it implies we are migrating too quickly or the local
1930 * node is overloaded. In either case, scan slower
1932 if (local + shared == 0 || p->numa_faults_locality[2]) {
1933 p->numa_scan_period = min(p->numa_scan_period_max,
1934 p->numa_scan_period << 1);
1936 p->mm->numa_next_scan = jiffies +
1937 msecs_to_jiffies(p->numa_scan_period);
1943 * Prepare to scale scan period relative to the current period.
1944 * == NUMA_PERIOD_THRESHOLD scan period stays the same
1945 * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
1946 * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
1948 period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS);
1949 lr_ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote);
1950 ps_ratio = (private * NUMA_PERIOD_SLOTS) / (private + shared);
1952 if (ps_ratio >= NUMA_PERIOD_THRESHOLD) {
1954 * Most memory accesses are local. There is no need to
1955 * do fast NUMA scanning, since memory is already local.
1957 int slot = ps_ratio - NUMA_PERIOD_THRESHOLD;
1960 diff = slot * period_slot;
1961 } else if (lr_ratio >= NUMA_PERIOD_THRESHOLD) {
1963 * Most memory accesses are shared with other tasks.
1964 * There is no point in continuing fast NUMA scanning,
1965 * since other tasks may just move the memory elsewhere.
1967 int slot = lr_ratio - NUMA_PERIOD_THRESHOLD;
1970 diff = slot * period_slot;
1973 * Private memory faults exceed (SLOTS-THRESHOLD)/SLOTS,
1974 * yet they are not on the local NUMA node. Speed up
1975 * NUMA scanning to get the memory moved over.
1977 int ratio = max(lr_ratio, ps_ratio);
1978 diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot;
1981 p->numa_scan_period = clamp(p->numa_scan_period + diff,
1982 task_scan_min(p), task_scan_max(p));
1983 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
1987 * Get the fraction of time the task has been running since the last
1988 * NUMA placement cycle. The scheduler keeps similar statistics, but
1989 * decays those on a 32ms period, which is orders of magnitude off
1990 * from the dozens-of-seconds NUMA balancing period. Use the scheduler
1991 * stats only if the task is so new there are no NUMA statistics yet.
1993 static u64 numa_get_avg_runtime(struct task_struct *p, u64 *period)
1995 u64 runtime, delta, now;
1996 /* Use the start of this time slice to avoid calculations. */
1997 now = p->se.exec_start;
1998 runtime = p->se.sum_exec_runtime;
2000 if (p->last_task_numa_placement) {
2001 delta = runtime - p->last_sum_exec_runtime;
2002 *period = now - p->last_task_numa_placement;
2004 delta = p->se.avg.load_sum;
2005 *period = LOAD_AVG_MAX;
2008 p->last_sum_exec_runtime = runtime;
2009 p->last_task_numa_placement = now;
2015 * Determine the preferred nid for a task in a numa_group. This needs to
2016 * be done in a way that produces consistent results with group_weight,
2017 * otherwise workloads might not converge.
2019 static int preferred_group_nid(struct task_struct *p, int nid)
2024 /* Direct connections between all NUMA nodes. */
2025 if (sched_numa_topology_type == NUMA_DIRECT)
2029 * On a system with glueless mesh NUMA topology, group_weight
2030 * scores nodes according to the number of NUMA hinting faults on
2031 * both the node itself, and on nearby nodes.
2033 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
2034 unsigned long score, max_score = 0;
2035 int node, max_node = nid;
2037 dist = sched_max_numa_distance;
2039 for_each_online_node(node) {
2040 score = group_weight(p, node, dist);
2041 if (score > max_score) {
2050 * Finding the preferred nid in a system with NUMA backplane
2051 * interconnect topology is more involved. The goal is to locate
2052 * tasks from numa_groups near each other in the system, and
2053 * untangle workloads from different sides of the system. This requires
2054 * searching down the hierarchy of node groups, recursively searching
2055 * inside the highest scoring group of nodes. The nodemask tricks
2056 * keep the complexity of the search down.
2058 nodes = node_online_map;
2059 for (dist = sched_max_numa_distance; dist > LOCAL_DISTANCE; dist--) {
2060 unsigned long max_faults = 0;
2061 nodemask_t max_group = NODE_MASK_NONE;
2064 /* Are there nodes at this distance from each other? */
2065 if (!find_numa_distance(dist))
2068 for_each_node_mask(a, nodes) {
2069 unsigned long faults = 0;
2070 nodemask_t this_group;
2071 nodes_clear(this_group);
2073 /* Sum group's NUMA faults; includes a==b case. */
2074 for_each_node_mask(b, nodes) {
2075 if (node_distance(a, b) < dist) {
2076 faults += group_faults(p, b);
2077 node_set(b, this_group);
2078 node_clear(b, nodes);
2082 /* Remember the top group. */
2083 if (faults > max_faults) {
2084 max_faults = faults;
2085 max_group = this_group;
2087 * subtle: at the smallest distance there is
2088 * just one node left in each "group", the
2089 * winner is the preferred nid.
2094 /* Next round, evaluate the nodes within max_group. */
2102 static void task_numa_placement(struct task_struct *p)
2104 int seq, nid, max_nid = -1;
2105 unsigned long max_faults = 0;
2106 unsigned long fault_types[2] = { 0, 0 };
2107 unsigned long total_faults;
2108 u64 runtime, period;
2109 spinlock_t *group_lock = NULL;
2112 * The p->mm->numa_scan_seq field gets updated without
2113 * exclusive access. Use READ_ONCE() here to ensure
2114 * that the field is read in a single access:
2116 seq = READ_ONCE(p->mm->numa_scan_seq);
2117 if (p->numa_scan_seq == seq)
2119 p->numa_scan_seq = seq;
2120 p->numa_scan_period_max = task_scan_max(p);
2122 total_faults = p->numa_faults_locality[0] +
2123 p->numa_faults_locality[1];
2124 runtime = numa_get_avg_runtime(p, &period);
2126 /* If the task is part of a group prevent parallel updates to group stats */
2127 if (p->numa_group) {
2128 group_lock = &p->numa_group->lock;
2129 spin_lock_irq(group_lock);
2132 /* Find the node with the highest number of faults */
2133 for_each_online_node(nid) {
2134 /* Keep track of the offsets in numa_faults array */
2135 int mem_idx, membuf_idx, cpu_idx, cpubuf_idx;
2136 unsigned long faults = 0, group_faults = 0;
2139 for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) {
2140 long diff, f_diff, f_weight;
2142 mem_idx = task_faults_idx(NUMA_MEM, nid, priv);
2143 membuf_idx = task_faults_idx(NUMA_MEMBUF, nid, priv);
2144 cpu_idx = task_faults_idx(NUMA_CPU, nid, priv);
2145 cpubuf_idx = task_faults_idx(NUMA_CPUBUF, nid, priv);
2147 /* Decay existing window, copy faults since last scan */
2148 diff = p->numa_faults[membuf_idx] - p->numa_faults[mem_idx] / 2;
2149 fault_types[priv] += p->numa_faults[membuf_idx];
2150 p->numa_faults[membuf_idx] = 0;
2153 * Normalize the faults_from, so all tasks in a group
2154 * count according to CPU use, instead of by the raw
2155 * number of faults. Tasks with little runtime have
2156 * little over-all impact on throughput, and thus their
2157 * faults are less important.
2159 f_weight = div64_u64(runtime << 16, period + 1);
2160 f_weight = (f_weight * p->numa_faults[cpubuf_idx]) /
2162 f_diff = f_weight - p->numa_faults[cpu_idx] / 2;
2163 p->numa_faults[cpubuf_idx] = 0;
2165 p->numa_faults[mem_idx] += diff;
2166 p->numa_faults[cpu_idx] += f_diff;
2167 faults += p->numa_faults[mem_idx];
2168 p->total_numa_faults += diff;
2169 if (p->numa_group) {
2171 * safe because we can only change our own group
2173 * mem_idx represents the offset for a given
2174 * nid and priv in a specific region because it
2175 * is at the beginning of the numa_faults array.
2177 p->numa_group->faults[mem_idx] += diff;
2178 p->numa_group->faults_cpu[mem_idx] += f_diff;
2179 p->numa_group->total_faults += diff;
2180 group_faults += p->numa_group->faults[mem_idx];
2184 if (!p->numa_group) {
2185 if (faults > max_faults) {
2186 max_faults = faults;
2189 } else if (group_faults > max_faults) {
2190 max_faults = group_faults;
2195 if (p->numa_group) {
2196 numa_group_count_active_nodes(p->numa_group);
2197 spin_unlock_irq(group_lock);
2198 max_nid = preferred_group_nid(p, max_nid);
2202 /* Set the new preferred node */
2203 if (max_nid != p->numa_preferred_nid)
2204 sched_setnuma(p, max_nid);
2207 update_task_scan_period(p, fault_types[0], fault_types[1]);
2210 static inline int get_numa_group(struct numa_group *grp)
2212 return atomic_inc_not_zero(&grp->refcount);
2215 static inline void put_numa_group(struct numa_group *grp)
2217 if (atomic_dec_and_test(&grp->refcount))
2218 kfree_rcu(grp, rcu);
2221 static void task_numa_group(struct task_struct *p, int cpupid, int flags,
2224 struct numa_group *grp, *my_grp;
2225 struct task_struct *tsk;
2227 int cpu = cpupid_to_cpu(cpupid);
2230 if (unlikely(!p->numa_group)) {
2231 unsigned int size = sizeof(struct numa_group) +
2232 4*nr_node_ids*sizeof(unsigned long);
2234 grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
2238 atomic_set(&grp->refcount, 1);
2239 grp->active_nodes = 1;
2240 grp->max_faults_cpu = 0;
2241 spin_lock_init(&grp->lock);
2243 /* Second half of the array tracks nids where faults happen */
2244 grp->faults_cpu = grp->faults + NR_NUMA_HINT_FAULT_TYPES *
2247 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2248 grp->faults[i] = p->numa_faults[i];
2250 grp->total_faults = p->total_numa_faults;
2253 rcu_assign_pointer(p->numa_group, grp);
2257 tsk = READ_ONCE(cpu_rq(cpu)->curr);
2259 if (!cpupid_match_pid(tsk, cpupid))
2262 grp = rcu_dereference(tsk->numa_group);
2266 my_grp = p->numa_group;
2271 * Only join the other group if its bigger; if we're the bigger group,
2272 * the other task will join us.
2274 if (my_grp->nr_tasks > grp->nr_tasks)
2278 * Tie-break on the grp address.
2280 if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp)
2283 /* Always join threads in the same process. */
2284 if (tsk->mm == current->mm)
2287 /* Simple filter to avoid false positives due to PID collisions */
2288 if (flags & TNF_SHARED)
2291 /* Update priv based on whether false sharing was detected */
2294 if (join && !get_numa_group(grp))
2302 BUG_ON(irqs_disabled());
2303 double_lock_irq(&my_grp->lock, &grp->lock);
2305 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) {
2306 my_grp->faults[i] -= p->numa_faults[i];
2307 grp->faults[i] += p->numa_faults[i];
2309 my_grp->total_faults -= p->total_numa_faults;
2310 grp->total_faults += p->total_numa_faults;
2315 spin_unlock(&my_grp->lock);
2316 spin_unlock_irq(&grp->lock);
2318 rcu_assign_pointer(p->numa_group, grp);
2320 put_numa_group(my_grp);
2328 void task_numa_free(struct task_struct *p)
2330 struct numa_group *grp = p->numa_group;
2331 void *numa_faults = p->numa_faults;
2332 unsigned long flags;
2336 spin_lock_irqsave(&grp->lock, flags);
2337 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2338 grp->faults[i] -= p->numa_faults[i];
2339 grp->total_faults -= p->total_numa_faults;
2342 spin_unlock_irqrestore(&grp->lock, flags);
2343 RCU_INIT_POINTER(p->numa_group, NULL);
2344 put_numa_group(grp);
2347 p->numa_faults = NULL;
2352 * Got a PROT_NONE fault for a page on @node.
2354 void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags)
2356 struct task_struct *p = current;
2357 bool migrated = flags & TNF_MIGRATED;
2358 int cpu_node = task_node(current);
2359 int local = !!(flags & TNF_FAULT_LOCAL);
2360 struct numa_group *ng;
2363 if (!static_branch_likely(&sched_numa_balancing))
2366 /* for example, ksmd faulting in a user's mm */
2370 /* Allocate buffer to track faults on a per-node basis */
2371 if (unlikely(!p->numa_faults)) {
2372 int size = sizeof(*p->numa_faults) *
2373 NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids;
2375 p->numa_faults = kzalloc(size, GFP_KERNEL|__GFP_NOWARN);
2376 if (!p->numa_faults)
2379 p->total_numa_faults = 0;
2380 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
2384 * First accesses are treated as private, otherwise consider accesses
2385 * to be private if the accessing pid has not changed
2387 if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) {
2390 priv = cpupid_match_pid(p, last_cpupid);
2391 if (!priv && !(flags & TNF_NO_GROUP))
2392 task_numa_group(p, last_cpupid, flags, &priv);
2396 * If a workload spans multiple NUMA nodes, a shared fault that
2397 * occurs wholly within the set of nodes that the workload is
2398 * actively using should be counted as local. This allows the
2399 * scan rate to slow down when a workload has settled down.
2402 if (!priv && !local && ng && ng->active_nodes > 1 &&
2403 numa_is_active_node(cpu_node, ng) &&
2404 numa_is_active_node(mem_node, ng))
2408 * Retry task to preferred node migration periodically, in case it
2409 * case it previously failed, or the scheduler moved us.
2411 if (time_after(jiffies, p->numa_migrate_retry)) {
2412 task_numa_placement(p);
2413 numa_migrate_preferred(p);
2417 p->numa_pages_migrated += pages;
2418 if (flags & TNF_MIGRATE_FAIL)
2419 p->numa_faults_locality[2] += pages;
2421 p->numa_faults[task_faults_idx(NUMA_MEMBUF, mem_node, priv)] += pages;
2422 p->numa_faults[task_faults_idx(NUMA_CPUBUF, cpu_node, priv)] += pages;
2423 p->numa_faults_locality[local] += pages;
2426 static void reset_ptenuma_scan(struct task_struct *p)
2429 * We only did a read acquisition of the mmap sem, so
2430 * p->mm->numa_scan_seq is written to without exclusive access
2431 * and the update is not guaranteed to be atomic. That's not
2432 * much of an issue though, since this is just used for
2433 * statistical sampling. Use READ_ONCE/WRITE_ONCE, which are not
2434 * expensive, to avoid any form of compiler optimizations:
2436 WRITE_ONCE(p->mm->numa_scan_seq, READ_ONCE(p->mm->numa_scan_seq) + 1);
2437 p->mm->numa_scan_offset = 0;
2441 * The expensive part of numa migration is done from task_work context.
2442 * Triggered from task_tick_numa().
2444 void task_numa_work(struct callback_head *work)
2446 unsigned long migrate, next_scan, now = jiffies;
2447 struct task_struct *p = current;
2448 struct mm_struct *mm = p->mm;
2449 u64 runtime = p->se.sum_exec_runtime;
2450 struct vm_area_struct *vma;
2451 unsigned long start, end;
2452 unsigned long nr_pte_updates = 0;
2453 long pages, virtpages;
2455 SCHED_WARN_ON(p != container_of(work, struct task_struct, numa_work));
2457 work->next = work; /* protect against double add */
2459 * Who cares about NUMA placement when they're dying.
2461 * NOTE: make sure not to dereference p->mm before this check,
2462 * exit_task_work() happens _after_ exit_mm() so we could be called
2463 * without p->mm even though we still had it when we enqueued this
2466 if (p->flags & PF_EXITING)
2469 if (!mm->numa_next_scan) {
2470 mm->numa_next_scan = now +
2471 msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2475 * Enforce maximal scan/migration frequency..
2477 migrate = mm->numa_next_scan;
2478 if (time_before(now, migrate))
2481 if (p->numa_scan_period == 0) {
2482 p->numa_scan_period_max = task_scan_max(p);
2483 p->numa_scan_period = task_scan_start(p);
2486 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
2487 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
2491 * Delay this task enough that another task of this mm will likely win
2492 * the next time around.
2494 p->node_stamp += 2 * TICK_NSEC;
2496 start = mm->numa_scan_offset;
2497 pages = sysctl_numa_balancing_scan_size;
2498 pages <<= 20 - PAGE_SHIFT; /* MB in pages */
2499 virtpages = pages * 8; /* Scan up to this much virtual space */
2504 if (!down_read_trylock(&mm->mmap_sem))
2506 vma = find_vma(mm, start);
2508 reset_ptenuma_scan(p);
2512 for (; vma; vma = vma->vm_next) {
2513 if (!vma_migratable(vma) || !vma_policy_mof(vma) ||
2514 is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_MIXEDMAP)) {
2519 * Shared library pages mapped by multiple processes are not
2520 * migrated as it is expected they are cache replicated. Avoid
2521 * hinting faults in read-only file-backed mappings or the vdso
2522 * as migrating the pages will be of marginal benefit.
2525 (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ)))
2529 * Skip inaccessible VMAs to avoid any confusion between
2530 * PROT_NONE and NUMA hinting ptes
2532 if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)))
2536 start = max(start, vma->vm_start);
2537 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
2538 end = min(end, vma->vm_end);
2539 nr_pte_updates = change_prot_numa(vma, start, end);
2542 * Try to scan sysctl_numa_balancing_size worth of
2543 * hpages that have at least one present PTE that
2544 * is not already pte-numa. If the VMA contains
2545 * areas that are unused or already full of prot_numa
2546 * PTEs, scan up to virtpages, to skip through those
2550 pages -= (end - start) >> PAGE_SHIFT;
2551 virtpages -= (end - start) >> PAGE_SHIFT;
2554 if (pages <= 0 || virtpages <= 0)
2558 } while (end != vma->vm_end);
2563 * It is possible to reach the end of the VMA list but the last few
2564 * VMAs are not guaranteed to the vma_migratable. If they are not, we
2565 * would find the !migratable VMA on the next scan but not reset the
2566 * scanner to the start so check it now.
2569 mm->numa_scan_offset = start;
2571 reset_ptenuma_scan(p);
2572 up_read(&mm->mmap_sem);
2575 * Make sure tasks use at least 32x as much time to run other code
2576 * than they used here, to limit NUMA PTE scanning overhead to 3% max.
2577 * Usually update_task_scan_period slows down scanning enough; on an
2578 * overloaded system we need to limit overhead on a per task basis.
2580 if (unlikely(p->se.sum_exec_runtime != runtime)) {
2581 u64 diff = p->se.sum_exec_runtime - runtime;
2582 p->node_stamp += 32 * diff;
2587 * Drive the periodic memory faults..
2589 void task_tick_numa(struct rq *rq, struct task_struct *curr)
2591 struct callback_head *work = &curr->numa_work;
2595 * We don't care about NUMA placement if we don't have memory.
2597 if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
2601 * Using runtime rather than walltime has the dual advantage that
2602 * we (mostly) drive the selection from busy threads and that the
2603 * task needs to have done some actual work before we bother with
2606 now = curr->se.sum_exec_runtime;
2607 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
2609 if (now > curr->node_stamp + period) {
2610 if (!curr->node_stamp)
2611 curr->numa_scan_period = task_scan_start(curr);
2612 curr->node_stamp += period;
2614 if (!time_before(jiffies, curr->mm->numa_next_scan)) {
2615 init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */
2616 task_work_add(curr, work, true);
2622 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
2626 static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p)
2630 static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p)
2634 #endif /* CONFIG_NUMA_BALANCING */
2637 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2639 update_load_add(&cfs_rq->load, se->load.weight);
2640 if (!parent_entity(se))
2641 update_load_add(&rq_of(cfs_rq)->load, se->load.weight);
2643 if (entity_is_task(se)) {
2644 struct rq *rq = rq_of(cfs_rq);
2646 account_numa_enqueue(rq, task_of(se));
2647 list_add(&se->group_node, &rq->cfs_tasks);
2650 cfs_rq->nr_running++;
2654 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2656 update_load_sub(&cfs_rq->load, se->load.weight);
2657 if (!parent_entity(se))
2658 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
2660 if (entity_is_task(se)) {
2661 account_numa_dequeue(rq_of(cfs_rq), task_of(se));
2662 list_del_init(&se->group_node);
2665 cfs_rq->nr_running--;
2669 * Signed add and clamp on underflow.
2671 * Explicitly do a load-store to ensure the intermediate value never hits
2672 * memory. This allows lockless observations without ever seeing the negative
2675 #define add_positive(_ptr, _val) do { \
2676 typeof(_ptr) ptr = (_ptr); \
2677 typeof(_val) val = (_val); \
2678 typeof(*ptr) res, var = READ_ONCE(*ptr); \
2682 if (val < 0 && res > var) \
2685 WRITE_ONCE(*ptr, res); \
2689 * Unsigned subtract and clamp on underflow.
2691 * Explicitly do a load-store to ensure the intermediate value never hits
2692 * memory. This allows lockless observations without ever seeing the negative
2695 #define sub_positive(_ptr, _val) do { \
2696 typeof(_ptr) ptr = (_ptr); \
2697 typeof(*ptr) val = (_val); \
2698 typeof(*ptr) res, var = READ_ONCE(*ptr); \
2702 WRITE_ONCE(*ptr, res); \
2707 enqueue_runnable_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2709 cfs_rq->runnable_weight += se->runnable_weight;
2711 cfs_rq->avg.runnable_load_avg += se->avg.runnable_load_avg;
2712 cfs_rq->avg.runnable_load_sum += se_runnable(se) * se->avg.runnable_load_sum;
2716 dequeue_runnable_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2718 cfs_rq->runnable_weight -= se->runnable_weight;
2720 sub_positive(&cfs_rq->avg.runnable_load_avg, se->avg.runnable_load_avg);
2721 sub_positive(&cfs_rq->avg.runnable_load_sum,
2722 se_runnable(se) * se->avg.runnable_load_sum);
2726 enqueue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2728 cfs_rq->avg.load_avg += se->avg.load_avg;
2729 cfs_rq->avg.load_sum += se_weight(se) * se->avg.load_sum;
2733 dequeue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2735 sub_positive(&cfs_rq->avg.load_avg, se->avg.load_avg);
2736 sub_positive(&cfs_rq->avg.load_sum, se_weight(se) * se->avg.load_sum);
2740 enqueue_runnable_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
2742 dequeue_runnable_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
2744 enqueue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
2746 dequeue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
2749 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
2750 unsigned long weight, unsigned long runnable)
2753 /* commit outstanding execution time */
2754 if (cfs_rq->curr == se)
2755 update_curr(cfs_rq);
2756 account_entity_dequeue(cfs_rq, se);
2757 dequeue_runnable_load_avg(cfs_rq, se);
2759 dequeue_load_avg(cfs_rq, se);
2761 se->runnable_weight = runnable;
2762 update_load_set(&se->load, weight);
2766 u32 divider = LOAD_AVG_MAX - 1024 + se->avg.period_contrib;
2768 se->avg.load_avg = div_u64(se_weight(se) * se->avg.load_sum, divider);
2769 se->avg.runnable_load_avg =
2770 div_u64(se_runnable(se) * se->avg.runnable_load_sum, divider);
2774 enqueue_load_avg(cfs_rq, se);
2776 account_entity_enqueue(cfs_rq, se);
2777 enqueue_runnable_load_avg(cfs_rq, se);
2781 void reweight_task(struct task_struct *p, int prio)
2783 struct sched_entity *se = &p->se;
2784 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2785 struct load_weight *load = &se->load;
2786 unsigned long weight = scale_load(sched_prio_to_weight[prio]);
2788 reweight_entity(cfs_rq, se, weight, weight);
2789 load->inv_weight = sched_prio_to_wmult[prio];
2792 #ifdef CONFIG_FAIR_GROUP_SCHED
2795 * All this does is approximate the hierarchical proportion which includes that
2796 * global sum we all love to hate.
2798 * That is, the weight of a group entity, is the proportional share of the
2799 * group weight based on the group runqueue weights. That is:
2801 * tg->weight * grq->load.weight
2802 * ge->load.weight = ----------------------------- (1)
2803 * \Sum grq->load.weight
2805 * Now, because computing that sum is prohibitively expensive to compute (been
2806 * there, done that) we approximate it with this average stuff. The average
2807 * moves slower and therefore the approximation is cheaper and more stable.
2809 * So instead of the above, we substitute:
2811 * grq->load.weight -> grq->avg.load_avg (2)
2813 * which yields the following:
2815 * tg->weight * grq->avg.load_avg
2816 * ge->load.weight = ------------------------------ (3)
2819 * Where: tg->load_avg ~= \Sum grq->avg.load_avg
2821 * That is shares_avg, and it is right (given the approximation (2)).
2823 * The problem with it is that because the average is slow -- it was designed
2824 * to be exactly that of course -- this leads to transients in boundary
2825 * conditions. In specific, the case where the group was idle and we start the
2826 * one task. It takes time for our CPU's grq->avg.load_avg to build up,
2827 * yielding bad latency etc..
2829 * Now, in that special case (1) reduces to:
2831 * tg->weight * grq->load.weight
2832 * ge->load.weight = ----------------------------- = tg->weight (4)
2835 * That is, the sum collapses because all other CPUs are idle; the UP scenario.
2837 * So what we do is modify our approximation (3) to approach (4) in the (near)
2842 * tg->weight * grq->load.weight
2843 * --------------------------------------------------- (5)
2844 * tg->load_avg - grq->avg.load_avg + grq->load.weight
2846 * But because grq->load.weight can drop to 0, resulting in a divide by zero,
2847 * we need to use grq->avg.load_avg as its lower bound, which then gives:
2850 * tg->weight * grq->load.weight
2851 * ge->load.weight = ----------------------------- (6)
2856 * tg_load_avg' = tg->load_avg - grq->avg.load_avg +
2857 * max(grq->load.weight, grq->avg.load_avg)
2859 * And that is shares_weight and is icky. In the (near) UP case it approaches
2860 * (4) while in the normal case it approaches (3). It consistently
2861 * overestimates the ge->load.weight and therefore:
2863 * \Sum ge->load.weight >= tg->weight
2867 static long calc_group_shares(struct cfs_rq *cfs_rq)
2869 long tg_weight, tg_shares, load, shares;
2870 struct task_group *tg = cfs_rq->tg;
2872 tg_shares = READ_ONCE(tg->shares);
2874 load = max(scale_load_down(cfs_rq->load.weight), cfs_rq->avg.load_avg);
2876 tg_weight = atomic_long_read(&tg->load_avg);
2878 /* Ensure tg_weight >= load */
2879 tg_weight -= cfs_rq->tg_load_avg_contrib;
2882 shares = (tg_shares * load);
2884 shares /= tg_weight;
2887 * MIN_SHARES has to be unscaled here to support per-CPU partitioning
2888 * of a group with small tg->shares value. It is a floor value which is
2889 * assigned as a minimum load.weight to the sched_entity representing
2890 * the group on a CPU.
2892 * E.g. on 64-bit for a group with tg->shares of scale_load(15)=15*1024
2893 * on an 8-core system with 8 tasks each runnable on one CPU shares has
2894 * to be 15*1024*1/8=1920 instead of scale_load(MIN_SHARES)=2*1024. In
2895 * case no task is runnable on a CPU MIN_SHARES=2 should be returned
2898 return clamp_t(long, shares, MIN_SHARES, tg_shares);
2902 * This calculates the effective runnable weight for a group entity based on
2903 * the group entity weight calculated above.
2905 * Because of the above approximation (2), our group entity weight is
2906 * an load_avg based ratio (3). This means that it includes blocked load and
2907 * does not represent the runnable weight.
2909 * Approximate the group entity's runnable weight per ratio from the group
2912 * grq->avg.runnable_load_avg
2913 * ge->runnable_weight = ge->load.weight * -------------------------- (7)
2916 * However, analogous to above, since the avg numbers are slow, this leads to
2917 * transients in the from-idle case. Instead we use:
2919 * ge->runnable_weight = ge->load.weight *
2921 * max(grq->avg.runnable_load_avg, grq->runnable_weight)
2922 * ----------------------------------------------------- (8)
2923 * max(grq->avg.load_avg, grq->load.weight)
2925 * Where these max() serve both to use the 'instant' values to fix the slow
2926 * from-idle and avoid the /0 on to-idle, similar to (6).
2928 static long calc_group_runnable(struct cfs_rq *cfs_rq, long shares)
2930 long runnable, load_avg;
2932 load_avg = max(cfs_rq->avg.load_avg,
2933 scale_load_down(cfs_rq->load.weight));
2935 runnable = max(cfs_rq->avg.runnable_load_avg,
2936 scale_load_down(cfs_rq->runnable_weight));
2940 runnable /= load_avg;
2942 return clamp_t(long, runnable, MIN_SHARES, shares);
2944 #endif /* CONFIG_SMP */
2946 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
2949 * Recomputes the group entity based on the current state of its group
2952 static void update_cfs_group(struct sched_entity *se)
2954 struct cfs_rq *gcfs_rq = group_cfs_rq(se);
2955 long shares, runnable;
2960 if (throttled_hierarchy(gcfs_rq))
2964 runnable = shares = READ_ONCE(gcfs_rq->tg->shares);
2966 if (likely(se->load.weight == shares))
2969 shares = calc_group_shares(gcfs_rq);
2970 runnable = calc_group_runnable(gcfs_rq, shares);
2973 reweight_entity(cfs_rq_of(se), se, shares, runnable);
2976 #else /* CONFIG_FAIR_GROUP_SCHED */
2977 static inline void update_cfs_group(struct sched_entity *se)
2980 #endif /* CONFIG_FAIR_GROUP_SCHED */
2982 static inline void cfs_rq_util_change(struct cfs_rq *cfs_rq, int flags)
2984 struct rq *rq = rq_of(cfs_rq);
2986 if (&rq->cfs == cfs_rq || (flags & SCHED_CPUFREQ_MIGRATION)) {
2988 * There are a few boundary cases this might miss but it should
2989 * get called often enough that that should (hopefully) not be
2992 * It will not get called when we go idle, because the idle
2993 * thread is a different class (!fair), nor will the utilization
2994 * number include things like RT tasks.
2996 * As is, the util number is not freq-invariant (we'd have to
2997 * implement arch_scale_freq_capacity() for that).
3001 cpufreq_update_util(rq, flags);
3006 #ifdef CONFIG_FAIR_GROUP_SCHED
3008 * update_tg_load_avg - update the tg's load avg
3009 * @cfs_rq: the cfs_rq whose avg changed
3010 * @force: update regardless of how small the difference
3012 * This function 'ensures': tg->load_avg := \Sum tg->cfs_rq[]->avg.load.
3013 * However, because tg->load_avg is a global value there are performance
3016 * In order to avoid having to look at the other cfs_rq's, we use a
3017 * differential update where we store the last value we propagated. This in
3018 * turn allows skipping updates if the differential is 'small'.
3020 * Updating tg's load_avg is necessary before update_cfs_share().
3022 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force)
3024 long delta = cfs_rq->avg.load_avg - cfs_rq->tg_load_avg_contrib;
3027 * No need to update load_avg for root_task_group as it is not used.
3029 if (cfs_rq->tg == &root_task_group)
3032 if (force || abs(delta) > cfs_rq->tg_load_avg_contrib / 64) {
3033 atomic_long_add(delta, &cfs_rq->tg->load_avg);
3034 cfs_rq->tg_load_avg_contrib = cfs_rq->avg.load_avg;
3039 * Called within set_task_rq() right before setting a task's CPU. The
3040 * caller only guarantees p->pi_lock is held; no other assumptions,
3041 * including the state of rq->lock, should be made.
3043 void set_task_rq_fair(struct sched_entity *se,
3044 struct cfs_rq *prev, struct cfs_rq *next)
3046 u64 p_last_update_time;
3047 u64 n_last_update_time;
3049 if (!sched_feat(ATTACH_AGE_LOAD))
3053 * We are supposed to update the task to "current" time, then its up to
3054 * date and ready to go to new CPU/cfs_rq. But we have difficulty in
3055 * getting what current time is, so simply throw away the out-of-date
3056 * time. This will result in the wakee task is less decayed, but giving
3057 * the wakee more load sounds not bad.
3059 if (!(se->avg.last_update_time && prev))
3062 #ifndef CONFIG_64BIT
3064 u64 p_last_update_time_copy;
3065 u64 n_last_update_time_copy;
3068 p_last_update_time_copy = prev->load_last_update_time_copy;
3069 n_last_update_time_copy = next->load_last_update_time_copy;
3073 p_last_update_time = prev->avg.last_update_time;
3074 n_last_update_time = next->avg.last_update_time;
3076 } while (p_last_update_time != p_last_update_time_copy ||
3077 n_last_update_time != n_last_update_time_copy);
3080 p_last_update_time = prev->avg.last_update_time;
3081 n_last_update_time = next->avg.last_update_time;
3083 __update_load_avg_blocked_se(p_last_update_time, cpu_of(rq_of(prev)), se);
3084 se->avg.last_update_time = n_last_update_time;
3089 * When on migration a sched_entity joins/leaves the PELT hierarchy, we need to
3090 * propagate its contribution. The key to this propagation is the invariant
3091 * that for each group:
3093 * ge->avg == grq->avg (1)
3095 * _IFF_ we look at the pure running and runnable sums. Because they
3096 * represent the very same entity, just at different points in the hierarchy.
3098 * Per the above update_tg_cfs_util() is trivial and simply copies the running
3099 * sum over (but still wrong, because the group entity and group rq do not have
3100 * their PELT windows aligned).
3102 * However, update_tg_cfs_runnable() is more complex. So we have:
3104 * ge->avg.load_avg = ge->load.weight * ge->avg.runnable_avg (2)
3106 * And since, like util, the runnable part should be directly transferable,
3107 * the following would _appear_ to be the straight forward approach:
3109 * grq->avg.load_avg = grq->load.weight * grq->avg.runnable_avg (3)
3111 * And per (1) we have:
3113 * ge->avg.runnable_avg == grq->avg.runnable_avg
3117 * ge->load.weight * grq->avg.load_avg
3118 * ge->avg.load_avg = ----------------------------------- (4)
3121 * Except that is wrong!
3123 * Because while for entities historical weight is not important and we
3124 * really only care about our future and therefore can consider a pure
3125 * runnable sum, runqueues can NOT do this.
3127 * We specifically want runqueues to have a load_avg that includes
3128 * historical weights. Those represent the blocked load, the load we expect
3129 * to (shortly) return to us. This only works by keeping the weights as
3130 * integral part of the sum. We therefore cannot decompose as per (3).
3132 * Another reason this doesn't work is that runnable isn't a 0-sum entity.
3133 * Imagine a rq with 2 tasks that each are runnable 2/3 of the time. Then the
3134 * rq itself is runnable anywhere between 2/3 and 1 depending on how the
3135 * runnable section of these tasks overlap (or not). If they were to perfectly
3136 * align the rq as a whole would be runnable 2/3 of the time. If however we
3137 * always have at least 1 runnable task, the rq as a whole is always runnable.
3139 * So we'll have to approximate.. :/
3141 * Given the constraint:
3143 * ge->avg.running_sum <= ge->avg.runnable_sum <= LOAD_AVG_MAX
3145 * We can construct a rule that adds runnable to a rq by assuming minimal
3148 * On removal, we'll assume each task is equally runnable; which yields:
3150 * grq->avg.runnable_sum = grq->avg.load_sum / grq->load.weight
3152 * XXX: only do this for the part of runnable > running ?
3157 update_tg_cfs_util(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq)
3159 long delta = gcfs_rq->avg.util_avg - se->avg.util_avg;
3161 /* Nothing to update */
3166 * The relation between sum and avg is:
3168 * LOAD_AVG_MAX - 1024 + sa->period_contrib
3170 * however, the PELT windows are not aligned between grq and gse.
3173 /* Set new sched_entity's utilization */
3174 se->avg.util_avg = gcfs_rq->avg.util_avg;
3175 se->avg.util_sum = se->avg.util_avg * LOAD_AVG_MAX;
3177 /* Update parent cfs_rq utilization */
3178 add_positive(&cfs_rq->avg.util_avg, delta);
3179 cfs_rq->avg.util_sum = cfs_rq->avg.util_avg * LOAD_AVG_MAX;
3183 update_tg_cfs_runnable(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq)
3185 long delta_avg, running_sum, runnable_sum = gcfs_rq->prop_runnable_sum;
3186 unsigned long runnable_load_avg, load_avg;
3187 u64 runnable_load_sum, load_sum = 0;
3193 gcfs_rq->prop_runnable_sum = 0;
3195 if (runnable_sum >= 0) {
3197 * Add runnable; clip at LOAD_AVG_MAX. Reflects that until
3198 * the CPU is saturated running == runnable.
3200 runnable_sum += se->avg.load_sum;
3201 runnable_sum = min(runnable_sum, (long)LOAD_AVG_MAX);
3204 * Estimate the new unweighted runnable_sum of the gcfs_rq by
3205 * assuming all tasks are equally runnable.
3207 if (scale_load_down(gcfs_rq->load.weight)) {
3208 load_sum = div_s64(gcfs_rq->avg.load_sum,
3209 scale_load_down(gcfs_rq->load.weight));
3212 /* But make sure to not inflate se's runnable */
3213 runnable_sum = min(se->avg.load_sum, load_sum);
3217 * runnable_sum can't be lower than running_sum
3218 * As running sum is scale with CPU capacity wehreas the runnable sum
3219 * is not we rescale running_sum 1st
3221 running_sum = se->avg.util_sum /
3222 arch_scale_cpu_capacity(NULL, cpu_of(rq_of(cfs_rq)));
3223 runnable_sum = max(runnable_sum, running_sum);
3225 load_sum = (s64)se_weight(se) * runnable_sum;
3226 load_avg = div_s64(load_sum, LOAD_AVG_MAX);
3228 delta_sum = load_sum - (s64)se_weight(se) * se->avg.load_sum;
3229 delta_avg = load_avg - se->avg.load_avg;
3231 se->avg.load_sum = runnable_sum;
3232 se->avg.load_avg = load_avg;
3233 add_positive(&cfs_rq->avg.load_avg, delta_avg);
3234 add_positive(&cfs_rq->avg.load_sum, delta_sum);
3236 runnable_load_sum = (s64)se_runnable(se) * runnable_sum;
3237 runnable_load_avg = div_s64(runnable_load_sum, LOAD_AVG_MAX);
3238 delta_sum = runnable_load_sum - se_weight(se) * se->avg.runnable_load_sum;
3239 delta_avg = runnable_load_avg - se->avg.runnable_load_avg;
3241 se->avg.runnable_load_sum = runnable_sum;
3242 se->avg.runnable_load_avg = runnable_load_avg;
3245 add_positive(&cfs_rq->avg.runnable_load_avg, delta_avg);
3246 add_positive(&cfs_rq->avg.runnable_load_sum, delta_sum);
3250 static inline void add_tg_cfs_propagate(struct cfs_rq *cfs_rq, long runnable_sum)
3252 cfs_rq->propagate = 1;
3253 cfs_rq->prop_runnable_sum += runnable_sum;
3256 /* Update task and its cfs_rq load average */
3257 static inline int propagate_entity_load_avg(struct sched_entity *se)
3259 struct cfs_rq *cfs_rq, *gcfs_rq;
3261 if (entity_is_task(se))
3264 gcfs_rq = group_cfs_rq(se);
3265 if (!gcfs_rq->propagate)
3268 gcfs_rq->propagate = 0;
3270 cfs_rq = cfs_rq_of(se);
3272 add_tg_cfs_propagate(cfs_rq, gcfs_rq->prop_runnable_sum);
3274 update_tg_cfs_util(cfs_rq, se, gcfs_rq);
3275 update_tg_cfs_runnable(cfs_rq, se, gcfs_rq);
3281 * Check if we need to update the load and the utilization of a blocked
3284 static inline bool skip_blocked_update(struct sched_entity *se)
3286 struct cfs_rq *gcfs_rq = group_cfs_rq(se);
3289 * If sched_entity still have not zero load or utilization, we have to
3292 if (se->avg.load_avg || se->avg.util_avg)
3296 * If there is a pending propagation, we have to update the load and
3297 * the utilization of the sched_entity:
3299 if (gcfs_rq->propagate)
3303 * Otherwise, the load and the utilization of the sched_entity is
3304 * already zero and there is no pending propagation, so it will be a
3305 * waste of time to try to decay it:
3310 #else /* CONFIG_FAIR_GROUP_SCHED */
3312 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force) {}
3314 static inline int propagate_entity_load_avg(struct sched_entity *se)
3319 static inline void add_tg_cfs_propagate(struct cfs_rq *cfs_rq, long runnable_sum) {}
3321 #endif /* CONFIG_FAIR_GROUP_SCHED */
3324 * update_cfs_rq_load_avg - update the cfs_rq's load/util averages
3325 * @now: current time, as per cfs_rq_clock_task()
3326 * @cfs_rq: cfs_rq to update
3328 * The cfs_rq avg is the direct sum of all its entities (blocked and runnable)
3329 * avg. The immediate corollary is that all (fair) tasks must be attached, see
3330 * post_init_entity_util_avg().
3332 * cfs_rq->avg is used for task_h_load() and update_cfs_share() for example.
3334 * Returns true if the load decayed or we removed load.
3336 * Since both these conditions indicate a changed cfs_rq->avg.load we should
3337 * call update_tg_load_avg() when this function returns true.
3340 update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq)
3342 unsigned long removed_load = 0, removed_util = 0, removed_runnable_sum = 0;
3343 struct sched_avg *sa = &cfs_rq->avg;
3346 if (cfs_rq->removed.nr) {
3348 u32 divider = LOAD_AVG_MAX - 1024 + sa->period_contrib;
3350 raw_spin_lock(&cfs_rq->removed.lock);
3351 swap(cfs_rq->removed.util_avg, removed_util);
3352 swap(cfs_rq->removed.load_avg, removed_load);
3353 swap(cfs_rq->removed.runnable_sum, removed_runnable_sum);
3354 cfs_rq->removed.nr = 0;
3355 raw_spin_unlock(&cfs_rq->removed.lock);
3358 sub_positive(&sa->load_avg, r);
3359 sub_positive(&sa->load_sum, r * divider);
3362 sub_positive(&sa->util_avg, r);
3363 sub_positive(&sa->util_sum, r * divider);
3365 add_tg_cfs_propagate(cfs_rq, -(long)removed_runnable_sum);
3370 decayed |= __update_load_avg_cfs_rq(now, cpu_of(rq_of(cfs_rq)), cfs_rq);
3372 #ifndef CONFIG_64BIT
3374 cfs_rq->load_last_update_time_copy = sa->last_update_time;
3378 cfs_rq_util_change(cfs_rq, 0);
3384 * attach_entity_load_avg - attach this entity to its cfs_rq load avg
3385 * @cfs_rq: cfs_rq to attach to
3386 * @se: sched_entity to attach
3387 * @flags: migration hints
3389 * Must call update_cfs_rq_load_avg() before this, since we rely on
3390 * cfs_rq->avg.last_update_time being current.
3392 static void attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3394 u32 divider = LOAD_AVG_MAX - 1024 + cfs_rq->avg.period_contrib;
3397 * When we attach the @se to the @cfs_rq, we must align the decay
3398 * window because without that, really weird and wonderful things can
3403 se->avg.last_update_time = cfs_rq->avg.last_update_time;
3404 se->avg.period_contrib = cfs_rq->avg.period_contrib;
3407 * Hell(o) Nasty stuff.. we need to recompute _sum based on the new
3408 * period_contrib. This isn't strictly correct, but since we're
3409 * entirely outside of the PELT hierarchy, nobody cares if we truncate
3412 se->avg.util_sum = se->avg.util_avg * divider;
3414 se->avg.load_sum = divider;
3415 if (se_weight(se)) {
3417 div_u64(se->avg.load_avg * se->avg.load_sum, se_weight(se));
3420 se->avg.runnable_load_sum = se->avg.load_sum;
3422 enqueue_load_avg(cfs_rq, se);
3423 cfs_rq->avg.util_avg += se->avg.util_avg;
3424 cfs_rq->avg.util_sum += se->avg.util_sum;
3426 add_tg_cfs_propagate(cfs_rq, se->avg.load_sum);
3428 cfs_rq_util_change(cfs_rq, flags);
3432 * detach_entity_load_avg - detach this entity from its cfs_rq load avg
3433 * @cfs_rq: cfs_rq to detach from
3434 * @se: sched_entity to detach
3436 * Must call update_cfs_rq_load_avg() before this, since we rely on
3437 * cfs_rq->avg.last_update_time being current.
3439 static void detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3441 dequeue_load_avg(cfs_rq, se);
3442 sub_positive(&cfs_rq->avg.util_avg, se->avg.util_avg);
3443 sub_positive(&cfs_rq->avg.util_sum, se->avg.util_sum);
3445 add_tg_cfs_propagate(cfs_rq, -se->avg.load_sum);
3447 cfs_rq_util_change(cfs_rq, 0);
3451 * Optional action to be done while updating the load average
3453 #define UPDATE_TG 0x1
3454 #define SKIP_AGE_LOAD 0x2
3455 #define DO_ATTACH 0x4
3457 /* Update task and its cfs_rq load average */
3458 static inline void update_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3460 u64 now = cfs_rq_clock_task(cfs_rq);
3461 struct rq *rq = rq_of(cfs_rq);
3462 int cpu = cpu_of(rq);
3466 * Track task load average for carrying it to new CPU after migrated, and
3467 * track group sched_entity load average for task_h_load calc in migration
3469 if (se->avg.last_update_time && !(flags & SKIP_AGE_LOAD))
3470 __update_load_avg_se(now, cpu, cfs_rq, se);
3472 decayed = update_cfs_rq_load_avg(now, cfs_rq);
3473 decayed |= propagate_entity_load_avg(se);
3475 if (!se->avg.last_update_time && (flags & DO_ATTACH)) {
3478 * DO_ATTACH means we're here from enqueue_entity().
3479 * !last_update_time means we've passed through
3480 * migrate_task_rq_fair() indicating we migrated.
3482 * IOW we're enqueueing a task on a new CPU.
3484 attach_entity_load_avg(cfs_rq, se, SCHED_CPUFREQ_MIGRATION);
3485 update_tg_load_avg(cfs_rq, 0);
3487 } else if (decayed && (flags & UPDATE_TG))
3488 update_tg_load_avg(cfs_rq, 0);
3491 #ifndef CONFIG_64BIT
3492 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
3494 u64 last_update_time_copy;
3495 u64 last_update_time;
3498 last_update_time_copy = cfs_rq->load_last_update_time_copy;
3500 last_update_time = cfs_rq->avg.last_update_time;
3501 } while (last_update_time != last_update_time_copy);
3503 return last_update_time;
3506 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
3508 return cfs_rq->avg.last_update_time;
3513 * Synchronize entity load avg of dequeued entity without locking
3516 void sync_entity_load_avg(struct sched_entity *se)
3518 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3519 u64 last_update_time;
3521 last_update_time = cfs_rq_last_update_time(cfs_rq);
3522 __update_load_avg_blocked_se(last_update_time, cpu_of(rq_of(cfs_rq)), se);
3526 * Task first catches up with cfs_rq, and then subtract
3527 * itself from the cfs_rq (task must be off the queue now).
3529 void remove_entity_load_avg(struct sched_entity *se)
3531 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3532 unsigned long flags;
3535 * tasks cannot exit without having gone through wake_up_new_task() ->
3536 * post_init_entity_util_avg() which will have added things to the
3537 * cfs_rq, so we can remove unconditionally.
3539 * Similarly for groups, they will have passed through
3540 * post_init_entity_util_avg() before unregister_sched_fair_group()
3544 sync_entity_load_avg(se);
3546 raw_spin_lock_irqsave(&cfs_rq->removed.lock, flags);
3547 ++cfs_rq->removed.nr;
3548 cfs_rq->removed.util_avg += se->avg.util_avg;
3549 cfs_rq->removed.load_avg += se->avg.load_avg;
3550 cfs_rq->removed.runnable_sum += se->avg.load_sum; /* == runnable_sum */
3551 raw_spin_unlock_irqrestore(&cfs_rq->removed.lock, flags);
3554 static inline unsigned long cfs_rq_runnable_load_avg(struct cfs_rq *cfs_rq)
3556 return cfs_rq->avg.runnable_load_avg;
3559 static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq)
3561 return cfs_rq->avg.load_avg;
3564 static int idle_balance(struct rq *this_rq, struct rq_flags *rf);
3566 static inline unsigned long task_util(struct task_struct *p)
3568 return READ_ONCE(p->se.avg.util_avg);
3571 static inline unsigned long _task_util_est(struct task_struct *p)
3573 struct util_est ue = READ_ONCE(p->se.avg.util_est);
3575 return max(ue.ewma, ue.enqueued);
3578 static inline unsigned long task_util_est(struct task_struct *p)
3580 return max(task_util(p), _task_util_est(p));
3583 static inline void util_est_enqueue(struct cfs_rq *cfs_rq,
3584 struct task_struct *p)
3586 unsigned int enqueued;
3588 if (!sched_feat(UTIL_EST))
3591 /* Update root cfs_rq's estimated utilization */
3592 enqueued = cfs_rq->avg.util_est.enqueued;
3593 enqueued += (_task_util_est(p) | UTIL_AVG_UNCHANGED);
3594 WRITE_ONCE(cfs_rq->avg.util_est.enqueued, enqueued);
3598 * Check if a (signed) value is within a specified (unsigned) margin,
3599 * based on the observation that:
3601 * abs(x) < y := (unsigned)(x + y - 1) < (2 * y - 1)
3603 * NOTE: this only works when value + maring < INT_MAX.
3605 static inline bool within_margin(int value, int margin)
3607 return ((unsigned int)(value + margin - 1) < (2 * margin - 1));
3611 util_est_dequeue(struct cfs_rq *cfs_rq, struct task_struct *p, bool task_sleep)
3613 long last_ewma_diff;
3616 if (!sched_feat(UTIL_EST))
3619 /* Update root cfs_rq's estimated utilization */
3620 ue.enqueued = cfs_rq->avg.util_est.enqueued;
3621 ue.enqueued -= min_t(unsigned int, ue.enqueued,
3622 (_task_util_est(p) | UTIL_AVG_UNCHANGED));
3623 WRITE_ONCE(cfs_rq->avg.util_est.enqueued, ue.enqueued);
3626 * Skip update of task's estimated utilization when the task has not
3627 * yet completed an activation, e.g. being migrated.
3633 * If the PELT values haven't changed since enqueue time,
3634 * skip the util_est update.
3636 ue = p->se.avg.util_est;
3637 if (ue.enqueued & UTIL_AVG_UNCHANGED)
3641 * Skip update of task's estimated utilization when its EWMA is
3642 * already ~1% close to its last activation value.
3644 ue.enqueued = (task_util(p) | UTIL_AVG_UNCHANGED);
3645 last_ewma_diff = ue.enqueued - ue.ewma;
3646 if (within_margin(last_ewma_diff, (SCHED_CAPACITY_SCALE / 100)))
3650 * Update Task's estimated utilization
3652 * When *p completes an activation we can consolidate another sample
3653 * of the task size. This is done by storing the current PELT value
3654 * as ue.enqueued and by using this value to update the Exponential
3655 * Weighted Moving Average (EWMA):
3657 * ewma(t) = w * task_util(p) + (1-w) * ewma(t-1)
3658 * = w * task_util(p) + ewma(t-1) - w * ewma(t-1)
3659 * = w * (task_util(p) - ewma(t-1)) + ewma(t-1)
3660 * = w * ( last_ewma_diff ) + ewma(t-1)
3661 * = w * (last_ewma_diff + ewma(t-1) / w)
3663 * Where 'w' is the weight of new samples, which is configured to be
3664 * 0.25, thus making w=1/4 ( >>= UTIL_EST_WEIGHT_SHIFT)
3666 ue.ewma <<= UTIL_EST_WEIGHT_SHIFT;
3667 ue.ewma += last_ewma_diff;
3668 ue.ewma >>= UTIL_EST_WEIGHT_SHIFT;
3669 WRITE_ONCE(p->se.avg.util_est, ue);
3672 #else /* CONFIG_SMP */
3674 #define UPDATE_TG 0x0
3675 #define SKIP_AGE_LOAD 0x0
3676 #define DO_ATTACH 0x0
3678 static inline void update_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int not_used1)
3680 cfs_rq_util_change(cfs_rq, 0);
3683 static inline void remove_entity_load_avg(struct sched_entity *se) {}
3686 attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) {}
3688 detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3690 static inline int idle_balance(struct rq *rq, struct rq_flags *rf)
3696 util_est_enqueue(struct cfs_rq *cfs_rq, struct task_struct *p) {}
3699 util_est_dequeue(struct cfs_rq *cfs_rq, struct task_struct *p,
3702 #endif /* CONFIG_SMP */
3704 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
3706 #ifdef CONFIG_SCHED_DEBUG
3707 s64 d = se->vruntime - cfs_rq->min_vruntime;
3712 if (d > 3*sysctl_sched_latency)
3713 schedstat_inc(cfs_rq->nr_spread_over);
3718 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
3720 u64 vruntime = cfs_rq->min_vruntime;
3723 * The 'current' period is already promised to the current tasks,
3724 * however the extra weight of the new task will slow them down a
3725 * little, place the new task so that it fits in the slot that
3726 * stays open at the end.
3728 if (initial && sched_feat(START_DEBIT))
3729 vruntime += sched_vslice(cfs_rq, se);
3731 /* sleeps up to a single latency don't count. */
3733 unsigned long thresh = sysctl_sched_latency;
3736 * Halve their sleep time's effect, to allow
3737 * for a gentler effect of sleepers:
3739 if (sched_feat(GENTLE_FAIR_SLEEPERS))
3745 /* ensure we never gain time by being placed backwards. */
3746 se->vruntime = max_vruntime(se->vruntime, vruntime);
3749 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
3751 static inline void check_schedstat_required(void)
3753 #ifdef CONFIG_SCHEDSTATS
3754 if (schedstat_enabled())
3757 /* Force schedstat enabled if a dependent tracepoint is active */
3758 if (trace_sched_stat_wait_enabled() ||
3759 trace_sched_stat_sleep_enabled() ||
3760 trace_sched_stat_iowait_enabled() ||
3761 trace_sched_stat_blocked_enabled() ||
3762 trace_sched_stat_runtime_enabled()) {
3763 printk_deferred_once("Scheduler tracepoints stat_sleep, stat_iowait, "
3764 "stat_blocked and stat_runtime require the "
3765 "kernel parameter schedstats=enable or "
3766 "kernel.sched_schedstats=1\n");
3777 * update_min_vruntime()
3778 * vruntime -= min_vruntime
3782 * update_min_vruntime()
3783 * vruntime += min_vruntime
3785 * this way the vruntime transition between RQs is done when both
3786 * min_vruntime are up-to-date.
3790 * ->migrate_task_rq_fair() (p->state == TASK_WAKING)
3791 * vruntime -= min_vruntime
3795 * update_min_vruntime()
3796 * vruntime += min_vruntime
3798 * this way we don't have the most up-to-date min_vruntime on the originating
3799 * CPU and an up-to-date min_vruntime on the destination CPU.
3803 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3805 bool renorm = !(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_MIGRATED);
3806 bool curr = cfs_rq->curr == se;
3809 * If we're the current task, we must renormalise before calling
3813 se->vruntime += cfs_rq->min_vruntime;
3815 update_curr(cfs_rq);
3818 * Otherwise, renormalise after, such that we're placed at the current
3819 * moment in time, instead of some random moment in the past. Being
3820 * placed in the past could significantly boost this task to the
3821 * fairness detriment of existing tasks.
3823 if (renorm && !curr)
3824 se->vruntime += cfs_rq->min_vruntime;
3827 * When enqueuing a sched_entity, we must:
3828 * - Update loads to have both entity and cfs_rq synced with now.
3829 * - Add its load to cfs_rq->runnable_avg
3830 * - For group_entity, update its weight to reflect the new share of
3832 * - Add its new weight to cfs_rq->load.weight
3834 update_load_avg(cfs_rq, se, UPDATE_TG | DO_ATTACH);
3835 update_cfs_group(se);
3836 enqueue_runnable_load_avg(cfs_rq, se);
3837 account_entity_enqueue(cfs_rq, se);
3839 if (flags & ENQUEUE_WAKEUP)
3840 place_entity(cfs_rq, se, 0);
3842 check_schedstat_required();
3843 update_stats_enqueue(cfs_rq, se, flags);
3844 check_spread(cfs_rq, se);
3846 __enqueue_entity(cfs_rq, se);
3849 if (cfs_rq->nr_running == 1) {
3850 list_add_leaf_cfs_rq(cfs_rq);
3851 check_enqueue_throttle(cfs_rq);
3855 static void __clear_buddies_last(struct sched_entity *se)
3857 for_each_sched_entity(se) {
3858 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3859 if (cfs_rq->last != se)
3862 cfs_rq->last = NULL;
3866 static void __clear_buddies_next(struct sched_entity *se)
3868 for_each_sched_entity(se) {
3869 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3870 if (cfs_rq->next != se)
3873 cfs_rq->next = NULL;
3877 static void __clear_buddies_skip(struct sched_entity *se)
3879 for_each_sched_entity(se) {
3880 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3881 if (cfs_rq->skip != se)
3884 cfs_rq->skip = NULL;
3888 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
3890 if (cfs_rq->last == se)
3891 __clear_buddies_last(se);
3893 if (cfs_rq->next == se)
3894 __clear_buddies_next(se);
3896 if (cfs_rq->skip == se)
3897 __clear_buddies_skip(se);
3900 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
3903 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3906 * Update run-time statistics of the 'current'.
3908 update_curr(cfs_rq);
3911 * When dequeuing a sched_entity, we must:
3912 * - Update loads to have both entity and cfs_rq synced with now.
3913 * - Substract its load from the cfs_rq->runnable_avg.
3914 * - Substract its previous weight from cfs_rq->load.weight.
3915 * - For group entity, update its weight to reflect the new share
3916 * of its group cfs_rq.
3918 update_load_avg(cfs_rq, se, UPDATE_TG);
3919 dequeue_runnable_load_avg(cfs_rq, se);
3921 update_stats_dequeue(cfs_rq, se, flags);
3923 clear_buddies(cfs_rq, se);
3925 if (se != cfs_rq->curr)
3926 __dequeue_entity(cfs_rq, se);
3928 account_entity_dequeue(cfs_rq, se);
3931 * Normalize after update_curr(); which will also have moved
3932 * min_vruntime if @se is the one holding it back. But before doing
3933 * update_min_vruntime() again, which will discount @se's position and
3934 * can move min_vruntime forward still more.
3936 if (!(flags & DEQUEUE_SLEEP))
3937 se->vruntime -= cfs_rq->min_vruntime;
3939 /* return excess runtime on last dequeue */
3940 return_cfs_rq_runtime(cfs_rq);
3942 update_cfs_group(se);
3945 * Now advance min_vruntime if @se was the entity holding it back,
3946 * except when: DEQUEUE_SAVE && !DEQUEUE_MOVE, in this case we'll be
3947 * put back on, and if we advance min_vruntime, we'll be placed back
3948 * further than we started -- ie. we'll be penalized.
3950 if ((flags & (DEQUEUE_SAVE | DEQUEUE_MOVE)) == DEQUEUE_SAVE)
3951 update_min_vruntime(cfs_rq);
3955 * Preempt the current task with a newly woken task if needed:
3958 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
3960 unsigned long ideal_runtime, delta_exec;
3961 struct sched_entity *se;
3964 ideal_runtime = sched_slice(cfs_rq, curr);
3965 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
3966 if (delta_exec > ideal_runtime) {
3967 resched_curr(rq_of(cfs_rq));
3969 * The current task ran long enough, ensure it doesn't get
3970 * re-elected due to buddy favours.
3972 clear_buddies(cfs_rq, curr);
3977 * Ensure that a task that missed wakeup preemption by a
3978 * narrow margin doesn't have to wait for a full slice.
3979 * This also mitigates buddy induced latencies under load.
3981 if (delta_exec < sysctl_sched_min_granularity)
3984 se = __pick_first_entity(cfs_rq);
3985 delta = curr->vruntime - se->vruntime;
3990 if (delta > ideal_runtime)
3991 resched_curr(rq_of(cfs_rq));
3995 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
3997 /* 'current' is not kept within the tree. */
4000 * Any task has to be enqueued before it get to execute on
4001 * a CPU. So account for the time it spent waiting on the
4004 update_stats_wait_end(cfs_rq, se);
4005 __dequeue_entity(cfs_rq, se);
4006 update_load_avg(cfs_rq, se, UPDATE_TG);
4009 update_stats_curr_start(cfs_rq, se);
4013 * Track our maximum slice length, if the CPU's load is at
4014 * least twice that of our own weight (i.e. dont track it
4015 * when there are only lesser-weight tasks around):
4017 if (schedstat_enabled() && rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
4018 schedstat_set(se->statistics.slice_max,
4019 max((u64)schedstat_val(se->statistics.slice_max),
4020 se->sum_exec_runtime - se->prev_sum_exec_runtime));
4023 se->prev_sum_exec_runtime = se->sum_exec_runtime;
4027 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
4030 * Pick the next process, keeping these things in mind, in this order:
4031 * 1) keep things fair between processes/task groups
4032 * 2) pick the "next" process, since someone really wants that to run
4033 * 3) pick the "last" process, for cache locality
4034 * 4) do not run the "skip" process, if something else is available
4036 static struct sched_entity *
4037 pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr)
4039 struct sched_entity *left = __pick_first_entity(cfs_rq);
4040 struct sched_entity *se;
4043 * If curr is set we have to see if its left of the leftmost entity
4044 * still in the tree, provided there was anything in the tree at all.
4046 if (!left || (curr && entity_before(curr, left)))
4049 se = left; /* ideally we run the leftmost entity */
4052 * Avoid running the skip buddy, if running something else can
4053 * be done without getting too unfair.
4055 if (cfs_rq->skip == se) {
4056 struct sched_entity *second;
4059 second = __pick_first_entity(cfs_rq);
4061 second = __pick_next_entity(se);
4062 if (!second || (curr && entity_before(curr, second)))
4066 if (second && wakeup_preempt_entity(second, left) < 1)
4071 * Prefer last buddy, try to return the CPU to a preempted task.
4073 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
4077 * Someone really wants this to run. If it's not unfair, run it.
4079 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
4082 clear_buddies(cfs_rq, se);
4087 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
4089 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
4092 * If still on the runqueue then deactivate_task()
4093 * was not called and update_curr() has to be done:
4096 update_curr(cfs_rq);
4098 /* throttle cfs_rqs exceeding runtime */
4099 check_cfs_rq_runtime(cfs_rq);
4101 check_spread(cfs_rq, prev);
4104 update_stats_wait_start(cfs_rq, prev);
4105 /* Put 'current' back into the tree. */
4106 __enqueue_entity(cfs_rq, prev);
4107 /* in !on_rq case, update occurred at dequeue */
4108 update_load_avg(cfs_rq, prev, 0);
4110 cfs_rq->curr = NULL;
4114 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
4117 * Update run-time statistics of the 'current'.
4119 update_curr(cfs_rq);
4122 * Ensure that runnable average is periodically updated.
4124 update_load_avg(cfs_rq, curr, UPDATE_TG);
4125 update_cfs_group(curr);
4127 #ifdef CONFIG_SCHED_HRTICK
4129 * queued ticks are scheduled to match the slice, so don't bother
4130 * validating it and just reschedule.
4133 resched_curr(rq_of(cfs_rq));
4137 * don't let the period tick interfere with the hrtick preemption
4139 if (!sched_feat(DOUBLE_TICK) &&
4140 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
4144 if (cfs_rq->nr_running > 1)
4145 check_preempt_tick(cfs_rq, curr);
4149 /**************************************************
4150 * CFS bandwidth control machinery
4153 #ifdef CONFIG_CFS_BANDWIDTH
4155 #ifdef HAVE_JUMP_LABEL
4156 static struct static_key __cfs_bandwidth_used;
4158 static inline bool cfs_bandwidth_used(void)
4160 return static_key_false(&__cfs_bandwidth_used);
4163 void cfs_bandwidth_usage_inc(void)
4165 static_key_slow_inc_cpuslocked(&__cfs_bandwidth_used);
4168 void cfs_bandwidth_usage_dec(void)
4170 static_key_slow_dec_cpuslocked(&__cfs_bandwidth_used);
4172 #else /* HAVE_JUMP_LABEL */
4173 static bool cfs_bandwidth_used(void)
4178 void cfs_bandwidth_usage_inc(void) {}
4179 void cfs_bandwidth_usage_dec(void) {}
4180 #endif /* HAVE_JUMP_LABEL */
4183 * default period for cfs group bandwidth.
4184 * default: 0.1s, units: nanoseconds
4186 static inline u64 default_cfs_period(void)
4188 return 100000000ULL;
4191 static inline u64 sched_cfs_bandwidth_slice(void)
4193 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
4197 * Replenish runtime according to assigned quota and update expiration time.
4198 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
4199 * additional synchronization around rq->lock.
4201 * requires cfs_b->lock
4203 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
4207 if (cfs_b->quota == RUNTIME_INF)
4210 now = sched_clock_cpu(smp_processor_id());
4211 cfs_b->runtime = cfs_b->quota;
4212 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
4213 cfs_b->expires_seq++;
4216 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
4218 return &tg->cfs_bandwidth;
4221 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
4222 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
4224 if (unlikely(cfs_rq->throttle_count))
4225 return cfs_rq->throttled_clock_task - cfs_rq->throttled_clock_task_time;
4227 return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time;
4230 /* returns 0 on failure to allocate runtime */
4231 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4233 struct task_group *tg = cfs_rq->tg;
4234 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
4235 u64 amount = 0, min_amount, expires;
4238 /* note: this is a positive sum as runtime_remaining <= 0 */
4239 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
4241 raw_spin_lock(&cfs_b->lock);
4242 if (cfs_b->quota == RUNTIME_INF)
4243 amount = min_amount;
4245 start_cfs_bandwidth(cfs_b);
4247 if (cfs_b->runtime > 0) {
4248 amount = min(cfs_b->runtime, min_amount);
4249 cfs_b->runtime -= amount;
4253 expires_seq = cfs_b->expires_seq;
4254 expires = cfs_b->runtime_expires;
4255 raw_spin_unlock(&cfs_b->lock);
4257 cfs_rq->runtime_remaining += amount;
4259 * we may have advanced our local expiration to account for allowed
4260 * spread between our sched_clock and the one on which runtime was
4263 if (cfs_rq->expires_seq != expires_seq) {
4264 cfs_rq->expires_seq = expires_seq;
4265 cfs_rq->runtime_expires = expires;
4268 return cfs_rq->runtime_remaining > 0;
4272 * Note: This depends on the synchronization provided by sched_clock and the
4273 * fact that rq->clock snapshots this value.
4275 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4277 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4279 /* if the deadline is ahead of our clock, nothing to do */
4280 if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0))
4283 if (cfs_rq->runtime_remaining < 0)
4287 * If the local deadline has passed we have to consider the
4288 * possibility that our sched_clock is 'fast' and the global deadline
4289 * has not truly expired.
4291 * Fortunately we can check determine whether this the case by checking
4292 * whether the global deadline(cfs_b->expires_seq) has advanced.
4294 if (cfs_rq->expires_seq == cfs_b->expires_seq) {
4295 /* extend local deadline, drift is bounded above by 2 ticks */
4296 cfs_rq->runtime_expires += TICK_NSEC;
4298 /* global deadline is ahead, expiration has passed */
4299 cfs_rq->runtime_remaining = 0;
4303 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
4305 /* dock delta_exec before expiring quota (as it could span periods) */
4306 cfs_rq->runtime_remaining -= delta_exec;
4307 expire_cfs_rq_runtime(cfs_rq);
4309 if (likely(cfs_rq->runtime_remaining > 0))
4313 * if we're unable to extend our runtime we resched so that the active
4314 * hierarchy can be throttled
4316 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
4317 resched_curr(rq_of(cfs_rq));
4320 static __always_inline
4321 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
4323 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
4326 __account_cfs_rq_runtime(cfs_rq, delta_exec);
4329 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
4331 return cfs_bandwidth_used() && cfs_rq->throttled;
4334 /* check whether cfs_rq, or any parent, is throttled */
4335 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
4337 return cfs_bandwidth_used() && cfs_rq->throttle_count;
4341 * Ensure that neither of the group entities corresponding to src_cpu or
4342 * dest_cpu are members of a throttled hierarchy when performing group
4343 * load-balance operations.
4345 static inline int throttled_lb_pair(struct task_group *tg,
4346 int src_cpu, int dest_cpu)
4348 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
4350 src_cfs_rq = tg->cfs_rq[src_cpu];
4351 dest_cfs_rq = tg->cfs_rq[dest_cpu];
4353 return throttled_hierarchy(src_cfs_rq) ||
4354 throttled_hierarchy(dest_cfs_rq);
4357 static int tg_unthrottle_up(struct task_group *tg, void *data)
4359 struct rq *rq = data;
4360 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
4362 cfs_rq->throttle_count--;
4363 if (!cfs_rq->throttle_count) {
4364 /* adjust cfs_rq_clock_task() */
4365 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
4366 cfs_rq->throttled_clock_task;
4372 static int tg_throttle_down(struct task_group *tg, void *data)
4374 struct rq *rq = data;
4375 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
4377 /* group is entering throttled state, stop time */
4378 if (!cfs_rq->throttle_count)
4379 cfs_rq->throttled_clock_task = rq_clock_task(rq);
4380 cfs_rq->throttle_count++;
4385 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
4387 struct rq *rq = rq_of(cfs_rq);
4388 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4389 struct sched_entity *se;
4390 long task_delta, dequeue = 1;
4393 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
4395 /* freeze hierarchy runnable averages while throttled */
4397 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
4400 task_delta = cfs_rq->h_nr_running;
4401 for_each_sched_entity(se) {
4402 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
4403 /* throttled entity or throttle-on-deactivate */
4408 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
4409 qcfs_rq->h_nr_running -= task_delta;
4411 if (qcfs_rq->load.weight)
4416 sub_nr_running(rq, task_delta);
4418 cfs_rq->throttled = 1;
4419 cfs_rq->throttled_clock = rq_clock(rq);
4420 raw_spin_lock(&cfs_b->lock);
4421 empty = list_empty(&cfs_b->throttled_cfs_rq);
4424 * Add to the _head_ of the list, so that an already-started
4425 * distribute_cfs_runtime will not see us
4427 list_add_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
4430 * If we're the first throttled task, make sure the bandwidth
4434 start_cfs_bandwidth(cfs_b);
4436 raw_spin_unlock(&cfs_b->lock);
4439 void unthrottle_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;
4447 se = cfs_rq->tg->se[cpu_of(rq)];
4449 cfs_rq->throttled = 0;
4451 update_rq_clock(rq);
4453 raw_spin_lock(&cfs_b->lock);
4454 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
4455 list_del_rcu(&cfs_rq->throttled_list);
4456 raw_spin_unlock(&cfs_b->lock);
4458 /* update hierarchical throttle state */
4459 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
4461 if (!cfs_rq->load.weight)
4464 task_delta = cfs_rq->h_nr_running;
4465 for_each_sched_entity(se) {
4469 cfs_rq = cfs_rq_of(se);
4471 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
4472 cfs_rq->h_nr_running += task_delta;
4474 if (cfs_rq_throttled(cfs_rq))
4479 add_nr_running(rq, task_delta);
4481 /* Determine whether we need to wake up potentially idle CPU: */
4482 if (rq->curr == rq->idle && rq->cfs.nr_running)
4486 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
4487 u64 remaining, u64 expires)
4489 struct cfs_rq *cfs_rq;
4491 u64 starting_runtime = remaining;
4494 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
4496 struct rq *rq = rq_of(cfs_rq);
4500 if (!cfs_rq_throttled(cfs_rq))
4503 runtime = -cfs_rq->runtime_remaining + 1;
4504 if (runtime > remaining)
4505 runtime = remaining;
4506 remaining -= runtime;
4508 cfs_rq->runtime_remaining += runtime;
4509 cfs_rq->runtime_expires = expires;
4511 /* we check whether we're throttled above */
4512 if (cfs_rq->runtime_remaining > 0)
4513 unthrottle_cfs_rq(cfs_rq);
4523 return starting_runtime - remaining;
4527 * Responsible for refilling a task_group's bandwidth and unthrottling its
4528 * cfs_rqs as appropriate. If there has been no activity within the last
4529 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
4530 * used to track this state.
4532 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
4534 u64 runtime, runtime_expires;
4537 /* no need to continue the timer with no bandwidth constraint */
4538 if (cfs_b->quota == RUNTIME_INF)
4539 goto out_deactivate;
4541 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
4542 cfs_b->nr_periods += overrun;
4545 * idle depends on !throttled (for the case of a large deficit), and if
4546 * we're going inactive then everything else can be deferred
4548 if (cfs_b->idle && !throttled)
4549 goto out_deactivate;
4551 __refill_cfs_bandwidth_runtime(cfs_b);
4554 /* mark as potentially idle for the upcoming period */
4559 /* account preceding periods in which throttling occurred */
4560 cfs_b->nr_throttled += overrun;
4562 runtime_expires = cfs_b->runtime_expires;
4565 * This check is repeated as we are holding onto the new bandwidth while
4566 * we unthrottle. This can potentially race with an unthrottled group
4567 * trying to acquire new bandwidth from the global pool. This can result
4568 * in us over-using our runtime if it is all used during this loop, but
4569 * only by limited amounts in that extreme case.
4571 while (throttled && cfs_b->runtime > 0) {
4572 runtime = cfs_b->runtime;
4573 raw_spin_unlock(&cfs_b->lock);
4574 /* we can't nest cfs_b->lock while distributing bandwidth */
4575 runtime = distribute_cfs_runtime(cfs_b, runtime,
4577 raw_spin_lock(&cfs_b->lock);
4579 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
4581 cfs_b->runtime -= min(runtime, cfs_b->runtime);
4585 * While we are ensured activity in the period following an
4586 * unthrottle, this also covers the case in which the new bandwidth is
4587 * insufficient to cover the existing bandwidth deficit. (Forcing the
4588 * timer to remain active while there are any throttled entities.)
4598 /* a cfs_rq won't donate quota below this amount */
4599 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
4600 /* minimum remaining period time to redistribute slack quota */
4601 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
4602 /* how long we wait to gather additional slack before distributing */
4603 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
4606 * Are we near the end of the current quota period?
4608 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
4609 * hrtimer base being cleared by hrtimer_start. In the case of
4610 * migrate_hrtimers, base is never cleared, so we are fine.
4612 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
4614 struct hrtimer *refresh_timer = &cfs_b->period_timer;
4617 /* if the call-back is running a quota refresh is already occurring */
4618 if (hrtimer_callback_running(refresh_timer))
4621 /* is a quota refresh about to occur? */
4622 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
4623 if (remaining < min_expire)
4629 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
4631 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
4633 /* if there's a quota refresh soon don't bother with slack */
4634 if (runtime_refresh_within(cfs_b, min_left))
4637 hrtimer_start(&cfs_b->slack_timer,
4638 ns_to_ktime(cfs_bandwidth_slack_period),
4642 /* we know any runtime found here is valid as update_curr() precedes return */
4643 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4645 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4646 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
4648 if (slack_runtime <= 0)
4651 raw_spin_lock(&cfs_b->lock);
4652 if (cfs_b->quota != RUNTIME_INF &&
4653 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
4654 cfs_b->runtime += slack_runtime;
4656 /* we are under rq->lock, defer unthrottling using a timer */
4657 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
4658 !list_empty(&cfs_b->throttled_cfs_rq))
4659 start_cfs_slack_bandwidth(cfs_b);
4661 raw_spin_unlock(&cfs_b->lock);
4663 /* even if it's not valid for return we don't want to try again */
4664 cfs_rq->runtime_remaining -= slack_runtime;
4667 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4669 if (!cfs_bandwidth_used())
4672 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
4675 __return_cfs_rq_runtime(cfs_rq);
4679 * This is done with a timer (instead of inline with bandwidth return) since
4680 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
4682 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
4684 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
4687 /* confirm we're still not at a refresh boundary */
4688 raw_spin_lock(&cfs_b->lock);
4689 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
4690 raw_spin_unlock(&cfs_b->lock);
4694 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice)
4695 runtime = cfs_b->runtime;
4697 expires = cfs_b->runtime_expires;
4698 raw_spin_unlock(&cfs_b->lock);
4703 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
4705 raw_spin_lock(&cfs_b->lock);
4706 if (expires == cfs_b->runtime_expires)
4707 cfs_b->runtime -= min(runtime, cfs_b->runtime);
4708 raw_spin_unlock(&cfs_b->lock);
4712 * When a group wakes up we want to make sure that its quota is not already
4713 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
4714 * runtime as update_curr() throttling can not not trigger until it's on-rq.
4716 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
4718 if (!cfs_bandwidth_used())
4721 /* an active group must be handled by the update_curr()->put() path */
4722 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
4725 /* ensure the group is not already throttled */
4726 if (cfs_rq_throttled(cfs_rq))
4729 /* update runtime allocation */
4730 account_cfs_rq_runtime(cfs_rq, 0);
4731 if (cfs_rq->runtime_remaining <= 0)
4732 throttle_cfs_rq(cfs_rq);
4735 static void sync_throttle(struct task_group *tg, int cpu)
4737 struct cfs_rq *pcfs_rq, *cfs_rq;
4739 if (!cfs_bandwidth_used())
4745 cfs_rq = tg->cfs_rq[cpu];
4746 pcfs_rq = tg->parent->cfs_rq[cpu];
4748 cfs_rq->throttle_count = pcfs_rq->throttle_count;
4749 cfs_rq->throttled_clock_task = rq_clock_task(cpu_rq(cpu));
4752 /* conditionally throttle active cfs_rq's from put_prev_entity() */
4753 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4755 if (!cfs_bandwidth_used())
4758 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
4762 * it's possible for a throttled entity to be forced into a running
4763 * state (e.g. set_curr_task), in this case we're finished.
4765 if (cfs_rq_throttled(cfs_rq))
4768 throttle_cfs_rq(cfs_rq);
4772 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
4774 struct cfs_bandwidth *cfs_b =
4775 container_of(timer, struct cfs_bandwidth, slack_timer);
4777 do_sched_cfs_slack_timer(cfs_b);
4779 return HRTIMER_NORESTART;
4782 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
4784 struct cfs_bandwidth *cfs_b =
4785 container_of(timer, struct cfs_bandwidth, period_timer);
4789 raw_spin_lock(&cfs_b->lock);
4791 overrun = hrtimer_forward_now(timer, cfs_b->period);
4795 idle = do_sched_cfs_period_timer(cfs_b, overrun);
4798 cfs_b->period_active = 0;
4799 raw_spin_unlock(&cfs_b->lock);
4801 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
4804 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4806 raw_spin_lock_init(&cfs_b->lock);
4808 cfs_b->quota = RUNTIME_INF;
4809 cfs_b->period = ns_to_ktime(default_cfs_period());
4811 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
4812 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED);
4813 cfs_b->period_timer.function = sched_cfs_period_timer;
4814 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
4815 cfs_b->slack_timer.function = sched_cfs_slack_timer;
4818 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4820 cfs_rq->runtime_enabled = 0;
4821 INIT_LIST_HEAD(&cfs_rq->throttled_list);
4824 void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4828 lockdep_assert_held(&cfs_b->lock);
4830 if (cfs_b->period_active)
4833 cfs_b->period_active = 1;
4834 overrun = hrtimer_forward_now(&cfs_b->period_timer, cfs_b->period);
4835 cfs_b->runtime_expires += (overrun + 1) * ktime_to_ns(cfs_b->period);
4836 cfs_b->expires_seq++;
4837 hrtimer_start_expires(&cfs_b->period_timer, HRTIMER_MODE_ABS_PINNED);
4840 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4842 /* init_cfs_bandwidth() was not called */
4843 if (!cfs_b->throttled_cfs_rq.next)
4846 hrtimer_cancel(&cfs_b->period_timer);
4847 hrtimer_cancel(&cfs_b->slack_timer);
4851 * Both these CPU hotplug callbacks race against unregister_fair_sched_group()
4853 * The race is harmless, since modifying bandwidth settings of unhooked group
4854 * bits doesn't do much.
4857 /* cpu online calback */
4858 static void __maybe_unused update_runtime_enabled(struct rq *rq)
4860 struct task_group *tg;
4862 lockdep_assert_held(&rq->lock);
4865 list_for_each_entry_rcu(tg, &task_groups, list) {
4866 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
4867 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
4869 raw_spin_lock(&cfs_b->lock);
4870 cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF;
4871 raw_spin_unlock(&cfs_b->lock);
4876 /* cpu offline callback */
4877 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
4879 struct task_group *tg;
4881 lockdep_assert_held(&rq->lock);
4884 list_for_each_entry_rcu(tg, &task_groups, list) {
4885 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
4887 if (!cfs_rq->runtime_enabled)
4891 * clock_task is not advancing so we just need to make sure
4892 * there's some valid quota amount
4894 cfs_rq->runtime_remaining = 1;
4896 * Offline rq is schedulable till CPU is completely disabled
4897 * in take_cpu_down(), so we prevent new cfs throttling here.
4899 cfs_rq->runtime_enabled = 0;
4901 if (cfs_rq_throttled(cfs_rq))
4902 unthrottle_cfs_rq(cfs_rq);
4907 #else /* CONFIG_CFS_BANDWIDTH */
4908 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
4910 return rq_clock_task(rq_of(cfs_rq));
4913 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {}
4914 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; }
4915 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
4916 static inline void sync_throttle(struct task_group *tg, int cpu) {}
4917 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
4919 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
4924 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
4929 static inline int throttled_lb_pair(struct task_group *tg,
4930 int src_cpu, int dest_cpu)
4935 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
4937 #ifdef CONFIG_FAIR_GROUP_SCHED
4938 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
4941 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
4945 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
4946 static inline void update_runtime_enabled(struct rq *rq) {}
4947 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
4949 #endif /* CONFIG_CFS_BANDWIDTH */
4951 /**************************************************
4952 * CFS operations on tasks:
4955 #ifdef CONFIG_SCHED_HRTICK
4956 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
4958 struct sched_entity *se = &p->se;
4959 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4961 SCHED_WARN_ON(task_rq(p) != rq);
4963 if (rq->cfs.h_nr_running > 1) {
4964 u64 slice = sched_slice(cfs_rq, se);
4965 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
4966 s64 delta = slice - ran;
4973 hrtick_start(rq, delta);
4978 * called from enqueue/dequeue and updates the hrtick when the
4979 * current task is from our class and nr_running is low enough
4982 static void hrtick_update(struct rq *rq)
4984 struct task_struct *curr = rq->curr;
4986 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
4989 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
4990 hrtick_start_fair(rq, curr);
4992 #else /* !CONFIG_SCHED_HRTICK */
4994 hrtick_start_fair(struct rq *rq, struct task_struct *p)
4998 static inline void hrtick_update(struct rq *rq)
5004 * The enqueue_task method is called before nr_running is
5005 * increased. Here we update the fair scheduling stats and
5006 * then put the task into the rbtree:
5009 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
5011 struct cfs_rq *cfs_rq;
5012 struct sched_entity *se = &p->se;
5015 * The code below (indirectly) updates schedutil which looks at
5016 * the cfs_rq utilization to select a frequency.
5017 * Let's add the task's estimated utilization to the cfs_rq's
5018 * estimated utilization, before we update schedutil.
5020 util_est_enqueue(&rq->cfs, p);
5023 * If in_iowait is set, the code below may not trigger any cpufreq
5024 * utilization updates, so do it here explicitly with the IOWAIT flag
5028 cpufreq_update_util(rq, SCHED_CPUFREQ_IOWAIT);
5030 for_each_sched_entity(se) {
5033 cfs_rq = cfs_rq_of(se);
5034 enqueue_entity(cfs_rq, se, flags);
5037 * end evaluation on encountering a throttled cfs_rq
5039 * note: in the case of encountering a throttled cfs_rq we will
5040 * post the final h_nr_running increment below.
5042 if (cfs_rq_throttled(cfs_rq))
5044 cfs_rq->h_nr_running++;
5046 flags = ENQUEUE_WAKEUP;
5049 for_each_sched_entity(se) {
5050 cfs_rq = cfs_rq_of(se);
5051 cfs_rq->h_nr_running++;
5053 if (cfs_rq_throttled(cfs_rq))
5056 update_load_avg(cfs_rq, se, UPDATE_TG);
5057 update_cfs_group(se);
5061 add_nr_running(rq, 1);
5066 static void set_next_buddy(struct sched_entity *se);
5069 * The dequeue_task method is called before nr_running is
5070 * decreased. We remove the task from the rbtree and
5071 * update the fair scheduling stats:
5073 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
5075 struct cfs_rq *cfs_rq;
5076 struct sched_entity *se = &p->se;
5077 int task_sleep = flags & DEQUEUE_SLEEP;
5079 for_each_sched_entity(se) {
5080 cfs_rq = cfs_rq_of(se);
5081 dequeue_entity(cfs_rq, se, flags);
5084 * end evaluation on encountering a throttled cfs_rq
5086 * note: in the case of encountering a throttled cfs_rq we will
5087 * post the final h_nr_running decrement below.
5089 if (cfs_rq_throttled(cfs_rq))
5091 cfs_rq->h_nr_running--;
5093 /* Don't dequeue parent if it has other entities besides us */
5094 if (cfs_rq->load.weight) {
5095 /* Avoid re-evaluating load for this entity: */
5096 se = parent_entity(se);
5098 * Bias pick_next to pick a task from this cfs_rq, as
5099 * p is sleeping when it is within its sched_slice.
5101 if (task_sleep && se && !throttled_hierarchy(cfs_rq))
5105 flags |= DEQUEUE_SLEEP;
5108 for_each_sched_entity(se) {
5109 cfs_rq = cfs_rq_of(se);
5110 cfs_rq->h_nr_running--;
5112 if (cfs_rq_throttled(cfs_rq))
5115 update_load_avg(cfs_rq, se, UPDATE_TG);
5116 update_cfs_group(se);
5120 sub_nr_running(rq, 1);
5122 util_est_dequeue(&rq->cfs, p, task_sleep);
5128 /* Working cpumask for: load_balance, load_balance_newidle. */
5129 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
5130 DEFINE_PER_CPU(cpumask_var_t, select_idle_mask);
5132 #ifdef CONFIG_NO_HZ_COMMON
5134 * per rq 'load' arrray crap; XXX kill this.
5138 * The exact cpuload calculated at every tick would be:
5140 * load' = (1 - 1/2^i) * load + (1/2^i) * cur_load
5142 * If a CPU misses updates for n ticks (as it was idle) and update gets
5143 * called on the n+1-th tick when CPU may be busy, then we have:
5145 * load_n = (1 - 1/2^i)^n * load_0
5146 * load_n+1 = (1 - 1/2^i) * load_n + (1/2^i) * cur_load
5148 * decay_load_missed() below does efficient calculation of
5150 * load' = (1 - 1/2^i)^n * load
5152 * Because x^(n+m) := x^n * x^m we can decompose any x^n in power-of-2 factors.
5153 * This allows us to precompute the above in said factors, thereby allowing the
5154 * reduction of an arbitrary n in O(log_2 n) steps. (See also
5155 * fixed_power_int())
5157 * The calculation is approximated on a 128 point scale.
5159 #define DEGRADE_SHIFT 7
5161 static const u8 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
5162 static const u8 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
5163 { 0, 0, 0, 0, 0, 0, 0, 0 },
5164 { 64, 32, 8, 0, 0, 0, 0, 0 },
5165 { 96, 72, 40, 12, 1, 0, 0, 0 },
5166 { 112, 98, 75, 43, 15, 1, 0, 0 },
5167 { 120, 112, 98, 76, 45, 16, 2, 0 }
5171 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
5172 * would be when CPU is idle and so we just decay the old load without
5173 * adding any new load.
5175 static unsigned long
5176 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
5180 if (!missed_updates)
5183 if (missed_updates >= degrade_zero_ticks[idx])
5187 return load >> missed_updates;
5189 while (missed_updates) {
5190 if (missed_updates % 2)
5191 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
5193 missed_updates >>= 1;
5200 cpumask_var_t idle_cpus_mask;
5202 int has_blocked; /* Idle CPUS has blocked load */
5203 unsigned long next_balance; /* in jiffy units */
5204 unsigned long next_blocked; /* Next update of blocked load in jiffies */
5205 } nohz ____cacheline_aligned;
5207 #endif /* CONFIG_NO_HZ_COMMON */
5210 * __cpu_load_update - update the rq->cpu_load[] statistics
5211 * @this_rq: The rq to update statistics for
5212 * @this_load: The current load
5213 * @pending_updates: The number of missed updates
5215 * Update rq->cpu_load[] statistics. This function is usually called every
5216 * scheduler tick (TICK_NSEC).
5218 * This function computes a decaying average:
5220 * load[i]' = (1 - 1/2^i) * load[i] + (1/2^i) * load
5222 * Because of NOHZ it might not get called on every tick which gives need for
5223 * the @pending_updates argument.
5225 * load[i]_n = (1 - 1/2^i) * load[i]_n-1 + (1/2^i) * load_n-1
5226 * = A * load[i]_n-1 + B ; A := (1 - 1/2^i), B := (1/2^i) * load
5227 * = A * (A * load[i]_n-2 + B) + B
5228 * = A * (A * (A * load[i]_n-3 + B) + B) + B
5229 * = A^3 * load[i]_n-3 + (A^2 + A + 1) * B
5230 * = A^n * load[i]_0 + (A^(n-1) + A^(n-2) + ... + 1) * B
5231 * = A^n * load[i]_0 + ((1 - A^n) / (1 - A)) * B
5232 * = (1 - 1/2^i)^n * (load[i]_0 - load) + load
5234 * In the above we've assumed load_n := load, which is true for NOHZ_FULL as
5235 * any change in load would have resulted in the tick being turned back on.
5237 * For regular NOHZ, this reduces to:
5239 * load[i]_n = (1 - 1/2^i)^n * load[i]_0
5241 * see decay_load_misses(). For NOHZ_FULL we get to subtract and add the extra
5244 static void cpu_load_update(struct rq *this_rq, unsigned long this_load,
5245 unsigned long pending_updates)
5247 unsigned long __maybe_unused tickless_load = this_rq->cpu_load[0];
5250 this_rq->nr_load_updates++;
5252 /* Update our load: */
5253 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
5254 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
5255 unsigned long old_load, new_load;
5257 /* scale is effectively 1 << i now, and >> i divides by scale */
5259 old_load = this_rq->cpu_load[i];
5260 #ifdef CONFIG_NO_HZ_COMMON
5261 old_load = decay_load_missed(old_load, pending_updates - 1, i);
5262 if (tickless_load) {
5263 old_load -= decay_load_missed(tickless_load, pending_updates - 1, i);
5265 * old_load can never be a negative value because a
5266 * decayed tickless_load cannot be greater than the
5267 * original tickless_load.
5269 old_load += tickless_load;
5272 new_load = this_load;
5274 * Round up the averaging division if load is increasing. This
5275 * prevents us from getting stuck on 9 if the load is 10, for
5278 if (new_load > old_load)
5279 new_load += scale - 1;
5281 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
5285 /* Used instead of source_load when we know the type == 0 */
5286 static unsigned long weighted_cpuload(struct rq *rq)
5288 return cfs_rq_runnable_load_avg(&rq->cfs);
5291 #ifdef CONFIG_NO_HZ_COMMON
5293 * There is no sane way to deal with nohz on smp when using jiffies because the
5294 * CPU doing the jiffies update might drift wrt the CPU doing the jiffy reading
5295 * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}.
5297 * Therefore we need to avoid the delta approach from the regular tick when
5298 * possible since that would seriously skew the load calculation. This is why we
5299 * use cpu_load_update_periodic() for CPUs out of nohz. However we'll rely on
5300 * jiffies deltas for updates happening while in nohz mode (idle ticks, idle
5301 * loop exit, nohz_idle_balance, nohz full exit...)
5303 * This means we might still be one tick off for nohz periods.
5306 static void cpu_load_update_nohz(struct rq *this_rq,
5307 unsigned long curr_jiffies,
5310 unsigned long pending_updates;
5312 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
5313 if (pending_updates) {
5314 this_rq->last_load_update_tick = curr_jiffies;
5316 * In the regular NOHZ case, we were idle, this means load 0.
5317 * In the NOHZ_FULL case, we were non-idle, we should consider
5318 * its weighted load.
5320 cpu_load_update(this_rq, load, pending_updates);
5325 * Called from nohz_idle_balance() to update the load ratings before doing the
5328 static void cpu_load_update_idle(struct rq *this_rq)
5331 * bail if there's load or we're actually up-to-date.
5333 if (weighted_cpuload(this_rq))
5336 cpu_load_update_nohz(this_rq, READ_ONCE(jiffies), 0);
5340 * Record CPU load on nohz entry so we know the tickless load to account
5341 * on nohz exit. cpu_load[0] happens then to be updated more frequently
5342 * than other cpu_load[idx] but it should be fine as cpu_load readers
5343 * shouldn't rely into synchronized cpu_load[*] updates.
5345 void cpu_load_update_nohz_start(void)
5347 struct rq *this_rq = this_rq();
5350 * This is all lockless but should be fine. If weighted_cpuload changes
5351 * concurrently we'll exit nohz. And cpu_load write can race with
5352 * cpu_load_update_idle() but both updater would be writing the same.
5354 this_rq->cpu_load[0] = weighted_cpuload(this_rq);
5358 * Account the tickless load in the end of a nohz frame.
5360 void cpu_load_update_nohz_stop(void)
5362 unsigned long curr_jiffies = READ_ONCE(jiffies);
5363 struct rq *this_rq = this_rq();
5367 if (curr_jiffies == this_rq->last_load_update_tick)
5370 load = weighted_cpuload(this_rq);
5371 rq_lock(this_rq, &rf);
5372 update_rq_clock(this_rq);
5373 cpu_load_update_nohz(this_rq, curr_jiffies, load);
5374 rq_unlock(this_rq, &rf);
5376 #else /* !CONFIG_NO_HZ_COMMON */
5377 static inline void cpu_load_update_nohz(struct rq *this_rq,
5378 unsigned long curr_jiffies,
5379 unsigned long load) { }
5380 #endif /* CONFIG_NO_HZ_COMMON */
5382 static void cpu_load_update_periodic(struct rq *this_rq, unsigned long load)
5384 #ifdef CONFIG_NO_HZ_COMMON
5385 /* See the mess around cpu_load_update_nohz(). */
5386 this_rq->last_load_update_tick = READ_ONCE(jiffies);
5388 cpu_load_update(this_rq, load, 1);
5392 * Called from scheduler_tick()
5394 void cpu_load_update_active(struct rq *this_rq)
5396 unsigned long load = weighted_cpuload(this_rq);
5398 if (tick_nohz_tick_stopped())
5399 cpu_load_update_nohz(this_rq, READ_ONCE(jiffies), load);
5401 cpu_load_update_periodic(this_rq, load);
5405 * Return a low guess at the load of a migration-source CPU weighted
5406 * according to the scheduling class and "nice" value.
5408 * We want to under-estimate the load of migration sources, to
5409 * balance conservatively.
5411 static unsigned long source_load(int cpu, int type)
5413 struct rq *rq = cpu_rq(cpu);
5414 unsigned long total = weighted_cpuload(rq);
5416 if (type == 0 || !sched_feat(LB_BIAS))
5419 return min(rq->cpu_load[type-1], total);
5423 * Return a high guess at the load of a migration-target CPU weighted
5424 * according to the scheduling class and "nice" value.
5426 static unsigned long target_load(int cpu, int type)
5428 struct rq *rq = cpu_rq(cpu);
5429 unsigned long total = weighted_cpuload(rq);
5431 if (type == 0 || !sched_feat(LB_BIAS))
5434 return max(rq->cpu_load[type-1], total);
5437 static unsigned long capacity_of(int cpu)
5439 return cpu_rq(cpu)->cpu_capacity;
5442 static unsigned long capacity_orig_of(int cpu)
5444 return cpu_rq(cpu)->cpu_capacity_orig;
5447 static unsigned long cpu_avg_load_per_task(int cpu)
5449 struct rq *rq = cpu_rq(cpu);
5450 unsigned long nr_running = READ_ONCE(rq->cfs.h_nr_running);
5451 unsigned long load_avg = weighted_cpuload(rq);
5454 return load_avg / nr_running;
5459 static void record_wakee(struct task_struct *p)
5462 * Only decay a single time; tasks that have less then 1 wakeup per
5463 * jiffy will not have built up many flips.
5465 if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) {
5466 current->wakee_flips >>= 1;
5467 current->wakee_flip_decay_ts = jiffies;
5470 if (current->last_wakee != p) {
5471 current->last_wakee = p;
5472 current->wakee_flips++;
5477 * Detect M:N waker/wakee relationships via a switching-frequency heuristic.
5479 * A waker of many should wake a different task than the one last awakened
5480 * at a frequency roughly N times higher than one of its wakees.
5482 * In order to determine whether we should let the load spread vs consolidating
5483 * to shared cache, we look for a minimum 'flip' frequency of llc_size in one
5484 * partner, and a factor of lls_size higher frequency in the other.
5486 * With both conditions met, we can be relatively sure that the relationship is
5487 * non-monogamous, with partner count exceeding socket size.
5489 * Waker/wakee being client/server, worker/dispatcher, interrupt source or
5490 * whatever is irrelevant, spread criteria is apparent partner count exceeds
5493 static int wake_wide(struct task_struct *p)
5495 unsigned int master = current->wakee_flips;
5496 unsigned int slave = p->wakee_flips;
5497 int factor = this_cpu_read(sd_llc_size);
5500 swap(master, slave);
5501 if (slave < factor || master < slave * factor)
5507 * The purpose of wake_affine() is to quickly determine on which CPU we can run
5508 * soonest. For the purpose of speed we only consider the waking and previous
5511 * wake_affine_idle() - only considers 'now', it check if the waking CPU is
5512 * cache-affine and is (or will be) idle.
5514 * wake_affine_weight() - considers the weight to reflect the average
5515 * scheduling latency of the CPUs. This seems to work
5516 * for the overloaded case.
5519 wake_affine_idle(int this_cpu, int prev_cpu, int sync)
5522 * If this_cpu is idle, it implies the wakeup is from interrupt
5523 * context. Only allow the move if cache is shared. Otherwise an
5524 * interrupt intensive workload could force all tasks onto one
5525 * node depending on the IO topology or IRQ affinity settings.
5527 * If the prev_cpu is idle and cache affine then avoid a migration.
5528 * There is no guarantee that the cache hot data from an interrupt
5529 * is more important than cache hot data on the prev_cpu and from
5530 * a cpufreq perspective, it's better to have higher utilisation
5533 if (available_idle_cpu(this_cpu) && cpus_share_cache(this_cpu, prev_cpu))
5534 return available_idle_cpu(prev_cpu) ? prev_cpu : this_cpu;
5536 if (sync && cpu_rq(this_cpu)->nr_running == 1)
5539 return nr_cpumask_bits;
5543 wake_affine_weight(struct sched_domain *sd, struct task_struct *p,
5544 int this_cpu, int prev_cpu, int sync)
5546 s64 this_eff_load, prev_eff_load;
5547 unsigned long task_load;
5549 this_eff_load = target_load(this_cpu, sd->wake_idx);
5552 unsigned long current_load = task_h_load(current);
5554 if (current_load > this_eff_load)
5557 this_eff_load -= current_load;
5560 task_load = task_h_load(p);
5562 this_eff_load += task_load;
5563 if (sched_feat(WA_BIAS))
5564 this_eff_load *= 100;
5565 this_eff_load *= capacity_of(prev_cpu);
5567 prev_eff_load = source_load(prev_cpu, sd->wake_idx);
5568 prev_eff_load -= task_load;
5569 if (sched_feat(WA_BIAS))
5570 prev_eff_load *= 100 + (sd->imbalance_pct - 100) / 2;
5571 prev_eff_load *= capacity_of(this_cpu);
5574 * If sync, adjust the weight of prev_eff_load such that if
5575 * prev_eff == this_eff that select_idle_sibling() will consider
5576 * stacking the wakee on top of the waker if no other CPU is
5582 return this_eff_load < prev_eff_load ? this_cpu : nr_cpumask_bits;
5585 static int wake_affine(struct sched_domain *sd, struct task_struct *p,
5586 int this_cpu, int prev_cpu, int sync)
5588 int target = nr_cpumask_bits;
5590 if (sched_feat(WA_IDLE))
5591 target = wake_affine_idle(this_cpu, prev_cpu, sync);
5593 if (sched_feat(WA_WEIGHT) && target == nr_cpumask_bits)
5594 target = wake_affine_weight(sd, p, this_cpu, prev_cpu, sync);
5596 schedstat_inc(p->se.statistics.nr_wakeups_affine_attempts);
5597 if (target == nr_cpumask_bits)
5600 schedstat_inc(sd->ttwu_move_affine);
5601 schedstat_inc(p->se.statistics.nr_wakeups_affine);
5605 static unsigned long cpu_util_wake(int cpu, struct task_struct *p);
5607 static unsigned long capacity_spare_wake(int cpu, struct task_struct *p)
5609 return max_t(long, capacity_of(cpu) - cpu_util_wake(cpu, p), 0);
5613 * find_idlest_group finds and returns the least busy CPU group within the
5616 * Assumes p is allowed on at least one CPU in sd.
5618 static struct sched_group *
5619 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
5620 int this_cpu, int sd_flag)
5622 struct sched_group *idlest = NULL, *group = sd->groups;
5623 struct sched_group *most_spare_sg = NULL;
5624 unsigned long min_runnable_load = ULONG_MAX;
5625 unsigned long this_runnable_load = ULONG_MAX;
5626 unsigned long min_avg_load = ULONG_MAX, this_avg_load = ULONG_MAX;
5627 unsigned long most_spare = 0, this_spare = 0;
5628 int load_idx = sd->forkexec_idx;
5629 int imbalance_scale = 100 + (sd->imbalance_pct-100)/2;
5630 unsigned long imbalance = scale_load_down(NICE_0_LOAD) *
5631 (sd->imbalance_pct-100) / 100;
5633 if (sd_flag & SD_BALANCE_WAKE)
5634 load_idx = sd->wake_idx;
5637 unsigned long load, avg_load, runnable_load;
5638 unsigned long spare_cap, max_spare_cap;
5642 /* Skip over this group if it has no CPUs allowed */
5643 if (!cpumask_intersects(sched_group_span(group),
5647 local_group = cpumask_test_cpu(this_cpu,
5648 sched_group_span(group));
5651 * Tally up the load of all CPUs in the group and find
5652 * the group containing the CPU with most spare capacity.
5658 for_each_cpu(i, sched_group_span(group)) {
5659 /* Bias balancing toward CPUs of our domain */
5661 load = source_load(i, load_idx);
5663 load = target_load(i, load_idx);
5665 runnable_load += load;
5667 avg_load += cfs_rq_load_avg(&cpu_rq(i)->cfs);
5669 spare_cap = capacity_spare_wake(i, p);
5671 if (spare_cap > max_spare_cap)
5672 max_spare_cap = spare_cap;
5675 /* Adjust by relative CPU capacity of the group */
5676 avg_load = (avg_load * SCHED_CAPACITY_SCALE) /
5677 group->sgc->capacity;
5678 runnable_load = (runnable_load * SCHED_CAPACITY_SCALE) /
5679 group->sgc->capacity;
5682 this_runnable_load = runnable_load;
5683 this_avg_load = avg_load;
5684 this_spare = max_spare_cap;
5686 if (min_runnable_load > (runnable_load + imbalance)) {
5688 * The runnable load is significantly smaller
5689 * so we can pick this new CPU:
5691 min_runnable_load = runnable_load;
5692 min_avg_load = avg_load;
5694 } else if ((runnable_load < (min_runnable_load + imbalance)) &&
5695 (100*min_avg_load > imbalance_scale*avg_load)) {
5697 * The runnable loads are close so take the
5698 * blocked load into account through avg_load:
5700 min_avg_load = avg_load;
5704 if (most_spare < max_spare_cap) {
5705 most_spare = max_spare_cap;
5706 most_spare_sg = group;
5709 } while (group = group->next, group != sd->groups);
5712 * The cross-over point between using spare capacity or least load
5713 * is too conservative for high utilization tasks on partially
5714 * utilized systems if we require spare_capacity > task_util(p),
5715 * so we allow for some task stuffing by using
5716 * spare_capacity > task_util(p)/2.
5718 * Spare capacity can't be used for fork because the utilization has
5719 * not been set yet, we must first select a rq to compute the initial
5722 if (sd_flag & SD_BALANCE_FORK)
5725 if (this_spare > task_util(p) / 2 &&
5726 imbalance_scale*this_spare > 100*most_spare)
5729 if (most_spare > task_util(p) / 2)
5730 return most_spare_sg;
5737 * When comparing groups across NUMA domains, it's possible for the
5738 * local domain to be very lightly loaded relative to the remote
5739 * domains but "imbalance" skews the comparison making remote CPUs
5740 * look much more favourable. When considering cross-domain, add
5741 * imbalance to the runnable load on the remote node and consider
5744 if ((sd->flags & SD_NUMA) &&
5745 min_runnable_load + imbalance >= this_runnable_load)
5748 if (min_runnable_load > (this_runnable_load + imbalance))
5751 if ((this_runnable_load < (min_runnable_load + imbalance)) &&
5752 (100*this_avg_load < imbalance_scale*min_avg_load))
5759 * find_idlest_group_cpu - find the idlest CPU among the CPUs in the group.
5762 find_idlest_group_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
5764 unsigned long load, min_load = ULONG_MAX;
5765 unsigned int min_exit_latency = UINT_MAX;
5766 u64 latest_idle_timestamp = 0;
5767 int least_loaded_cpu = this_cpu;
5768 int shallowest_idle_cpu = -1;
5771 /* Check if we have any choice: */
5772 if (group->group_weight == 1)
5773 return cpumask_first(sched_group_span(group));
5775 /* Traverse only the allowed CPUs */
5776 for_each_cpu_and(i, sched_group_span(group), &p->cpus_allowed) {
5777 if (available_idle_cpu(i)) {
5778 struct rq *rq = cpu_rq(i);
5779 struct cpuidle_state *idle = idle_get_state(rq);
5780 if (idle && idle->exit_latency < min_exit_latency) {
5782 * We give priority to a CPU whose idle state
5783 * has the smallest exit latency irrespective
5784 * of any idle timestamp.
5786 min_exit_latency = idle->exit_latency;
5787 latest_idle_timestamp = rq->idle_stamp;
5788 shallowest_idle_cpu = i;
5789 } else if ((!idle || idle->exit_latency == min_exit_latency) &&
5790 rq->idle_stamp > latest_idle_timestamp) {
5792 * If equal or no active idle state, then
5793 * the most recently idled CPU might have
5796 latest_idle_timestamp = rq->idle_stamp;
5797 shallowest_idle_cpu = i;
5799 } else if (shallowest_idle_cpu == -1) {
5800 load = weighted_cpuload(cpu_rq(i));
5801 if (load < min_load) {
5803 least_loaded_cpu = i;
5808 return shallowest_idle_cpu != -1 ? shallowest_idle_cpu : least_loaded_cpu;
5811 static inline int find_idlest_cpu(struct sched_domain *sd, struct task_struct *p,
5812 int cpu, int prev_cpu, int sd_flag)
5816 if (!cpumask_intersects(sched_domain_span(sd), &p->cpus_allowed))
5820 * We need task's util for capacity_spare_wake, sync it up to prev_cpu's
5823 if (!(sd_flag & SD_BALANCE_FORK))
5824 sync_entity_load_avg(&p->se);
5827 struct sched_group *group;
5828 struct sched_domain *tmp;
5831 if (!(sd->flags & sd_flag)) {
5836 group = find_idlest_group(sd, p, cpu, sd_flag);
5842 new_cpu = find_idlest_group_cpu(group, p, cpu);
5843 if (new_cpu == cpu) {
5844 /* Now try balancing at a lower domain level of 'cpu': */
5849 /* Now try balancing at a lower domain level of 'new_cpu': */
5851 weight = sd->span_weight;
5853 for_each_domain(cpu, tmp) {
5854 if (weight <= tmp->span_weight)
5856 if (tmp->flags & sd_flag)
5864 #ifdef CONFIG_SCHED_SMT
5865 DEFINE_STATIC_KEY_FALSE(sched_smt_present);
5867 static inline void set_idle_cores(int cpu, int val)
5869 struct sched_domain_shared *sds;
5871 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
5873 WRITE_ONCE(sds->has_idle_cores, val);
5876 static inline bool test_idle_cores(int cpu, bool def)
5878 struct sched_domain_shared *sds;
5880 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
5882 return READ_ONCE(sds->has_idle_cores);
5888 * Scans the local SMT mask to see if the entire core is idle, and records this
5889 * information in sd_llc_shared->has_idle_cores.
5891 * Since SMT siblings share all cache levels, inspecting this limited remote
5892 * state should be fairly cheap.
5894 void __update_idle_core(struct rq *rq)
5896 int core = cpu_of(rq);
5900 if (test_idle_cores(core, true))
5903 for_each_cpu(cpu, cpu_smt_mask(core)) {
5907 if (!available_idle_cpu(cpu))
5911 set_idle_cores(core, 1);
5917 * Scan the entire LLC domain for idle cores; this dynamically switches off if
5918 * there are no idle cores left in the system; tracked through
5919 * sd_llc->shared->has_idle_cores and enabled through update_idle_core() above.
5921 static int select_idle_core(struct task_struct *p, struct sched_domain *sd, int target)
5923 struct cpumask *cpus = this_cpu_cpumask_var_ptr(select_idle_mask);
5926 if (!static_branch_likely(&sched_smt_present))
5929 if (!test_idle_cores(target, false))
5932 cpumask_and(cpus, sched_domain_span(sd), &p->cpus_allowed);
5934 for_each_cpu_wrap(core, cpus, target) {
5937 for_each_cpu(cpu, cpu_smt_mask(core)) {
5938 cpumask_clear_cpu(cpu, cpus);
5939 if (!available_idle_cpu(cpu))
5948 * Failed to find an idle core; stop looking for one.
5950 set_idle_cores(target, 0);
5956 * Scan the local SMT mask for idle CPUs.
5958 static int select_idle_smt(struct task_struct *p, struct sched_domain *sd, int target)
5962 if (!static_branch_likely(&sched_smt_present))
5965 for_each_cpu(cpu, cpu_smt_mask(target)) {
5966 if (!cpumask_test_cpu(cpu, &p->cpus_allowed))
5968 if (available_idle_cpu(cpu))
5975 #else /* CONFIG_SCHED_SMT */
5977 static inline int select_idle_core(struct task_struct *p, struct sched_domain *sd, int target)
5982 static inline int select_idle_smt(struct task_struct *p, struct sched_domain *sd, int target)
5987 #endif /* CONFIG_SCHED_SMT */
5990 * Scan the LLC domain for idle CPUs; this is dynamically regulated by
5991 * comparing the average scan cost (tracked in sd->avg_scan_cost) against the
5992 * average idle time for this rq (as found in rq->avg_idle).
5994 static int select_idle_cpu(struct task_struct *p, struct sched_domain *sd, int target)
5996 struct sched_domain *this_sd;
5997 u64 avg_cost, avg_idle;
6000 int cpu, nr = INT_MAX;
6002 this_sd = rcu_dereference(*this_cpu_ptr(&sd_llc));
6007 * Due to large variance we need a large fuzz factor; hackbench in
6008 * particularly is sensitive here.
6010 avg_idle = this_rq()->avg_idle / 512;
6011 avg_cost = this_sd->avg_scan_cost + 1;
6013 if (sched_feat(SIS_AVG_CPU) && avg_idle < avg_cost)
6016 if (sched_feat(SIS_PROP)) {
6017 u64 span_avg = sd->span_weight * avg_idle;
6018 if (span_avg > 4*avg_cost)
6019 nr = div_u64(span_avg, avg_cost);
6024 time = local_clock();
6026 for_each_cpu_wrap(cpu, sched_domain_span(sd), target) {
6029 if (!cpumask_test_cpu(cpu, &p->cpus_allowed))
6031 if (available_idle_cpu(cpu))
6035 time = local_clock() - time;
6036 cost = this_sd->avg_scan_cost;
6037 delta = (s64)(time - cost) / 8;
6038 this_sd->avg_scan_cost += delta;
6044 * Try and locate an idle core/thread in the LLC cache domain.
6046 static int select_idle_sibling(struct task_struct *p, int prev, int target)
6048 struct sched_domain *sd;
6049 int i, recent_used_cpu;
6051 if (available_idle_cpu(target))
6055 * If the previous CPU is cache affine and idle, don't be stupid:
6057 if (prev != target && cpus_share_cache(prev, target) && available_idle_cpu(prev))
6060 /* Check a recently used CPU as a potential idle candidate: */
6061 recent_used_cpu = p->recent_used_cpu;
6062 if (recent_used_cpu != prev &&
6063 recent_used_cpu != target &&
6064 cpus_share_cache(recent_used_cpu, target) &&
6065 available_idle_cpu(recent_used_cpu) &&
6066 cpumask_test_cpu(p->recent_used_cpu, &p->cpus_allowed)) {
6068 * Replace recent_used_cpu with prev as it is a potential
6069 * candidate for the next wake:
6071 p->recent_used_cpu = prev;
6072 return recent_used_cpu;
6075 sd = rcu_dereference(per_cpu(sd_llc, target));
6079 i = select_idle_core(p, sd, target);
6080 if ((unsigned)i < nr_cpumask_bits)
6083 i = select_idle_cpu(p, sd, target);
6084 if ((unsigned)i < nr_cpumask_bits)
6087 i = select_idle_smt(p, sd, target);
6088 if ((unsigned)i < nr_cpumask_bits)
6095 * Amount of capacity of a CPU that is (estimated to be) used by CFS tasks
6096 * @cpu: the CPU to get the utilization of
6098 * The unit of the return value must be the one of capacity so we can compare
6099 * the utilization with the capacity of the CPU that is available for CFS task
6100 * (ie cpu_capacity).
6102 * cfs_rq.avg.util_avg is the sum of running time of runnable tasks plus the
6103 * recent utilization of currently non-runnable tasks on a CPU. It represents
6104 * the amount of utilization of a CPU in the range [0..capacity_orig] where
6105 * capacity_orig is the cpu_capacity available at the highest frequency
6106 * (arch_scale_freq_capacity()).
6107 * The utilization of a CPU converges towards a sum equal to or less than the
6108 * current capacity (capacity_curr <= capacity_orig) of the CPU because it is
6109 * the running time on this CPU scaled by capacity_curr.
6111 * The estimated utilization of a CPU is defined to be the maximum between its
6112 * cfs_rq.avg.util_avg and the sum of the estimated utilization of the tasks
6113 * currently RUNNABLE on that CPU.
6114 * This allows to properly represent the expected utilization of a CPU which
6115 * has just got a big task running since a long sleep period. At the same time
6116 * however it preserves the benefits of the "blocked utilization" in
6117 * describing the potential for other tasks waking up on the same CPU.
6119 * Nevertheless, cfs_rq.avg.util_avg can be higher than capacity_curr or even
6120 * higher than capacity_orig because of unfortunate rounding in
6121 * cfs.avg.util_avg or just after migrating tasks and new task wakeups until
6122 * the average stabilizes with the new running time. We need to check that the
6123 * utilization stays within the range of [0..capacity_orig] and cap it if
6124 * necessary. Without utilization capping, a group could be seen as overloaded
6125 * (CPU0 utilization at 121% + CPU1 utilization at 80%) whereas CPU1 has 20% of
6126 * available capacity. We allow utilization to overshoot capacity_curr (but not
6127 * capacity_orig) as it useful for predicting the capacity required after task
6128 * migrations (scheduler-driven DVFS).
6130 * Return: the (estimated) utilization for the specified CPU
6132 static inline unsigned long cpu_util(int cpu)
6134 struct cfs_rq *cfs_rq;
6137 cfs_rq = &cpu_rq(cpu)->cfs;
6138 util = READ_ONCE(cfs_rq->avg.util_avg);
6140 if (sched_feat(UTIL_EST))
6141 util = max(util, READ_ONCE(cfs_rq->avg.util_est.enqueued));
6143 return min_t(unsigned long, util, capacity_orig_of(cpu));
6147 * cpu_util_wake: Compute CPU utilization with any contributions from
6148 * the waking task p removed.
6150 static unsigned long cpu_util_wake(int cpu, struct task_struct *p)
6152 struct cfs_rq *cfs_rq;
6155 /* Task has no contribution or is new */
6156 if (cpu != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time))
6157 return cpu_util(cpu);
6159 cfs_rq = &cpu_rq(cpu)->cfs;
6160 util = READ_ONCE(cfs_rq->avg.util_avg);
6162 /* Discount task's blocked util from CPU's util */
6163 util -= min_t(unsigned int, util, task_util(p));
6168 * a) if *p is the only task sleeping on this CPU, then:
6169 * cpu_util (== task_util) > util_est (== 0)
6170 * and thus we return:
6171 * cpu_util_wake = (cpu_util - task_util) = 0
6173 * b) if other tasks are SLEEPING on this CPU, which is now exiting
6175 * cpu_util >= task_util
6176 * cpu_util > util_est (== 0)
6177 * and thus we discount *p's blocked utilization to return:
6178 * cpu_util_wake = (cpu_util - task_util) >= 0
6180 * c) if other tasks are RUNNABLE on that CPU and
6181 * util_est > cpu_util
6182 * then we use util_est since it returns a more restrictive
6183 * estimation of the spare capacity on that CPU, by just
6184 * considering the expected utilization of tasks already
6185 * runnable on that CPU.
6187 * Cases a) and b) are covered by the above code, while case c) is
6188 * covered by the following code when estimated utilization is
6191 if (sched_feat(UTIL_EST))
6192 util = max(util, READ_ONCE(cfs_rq->avg.util_est.enqueued));
6195 * Utilization (estimated) can exceed the CPU capacity, thus let's
6196 * clamp to the maximum CPU capacity to ensure consistency with
6197 * the cpu_util call.
6199 return min_t(unsigned long, util, capacity_orig_of(cpu));
6203 * Disable WAKE_AFFINE in the case where task @p doesn't fit in the
6204 * capacity of either the waking CPU @cpu or the previous CPU @prev_cpu.
6206 * In that case WAKE_AFFINE doesn't make sense and we'll let
6207 * BALANCE_WAKE sort things out.
6209 static int wake_cap(struct task_struct *p, int cpu, int prev_cpu)
6211 long min_cap, max_cap;
6213 min_cap = min(capacity_orig_of(prev_cpu), capacity_orig_of(cpu));
6214 max_cap = cpu_rq(cpu)->rd->max_cpu_capacity;
6216 /* Minimum capacity is close to max, no need to abort wake_affine */
6217 if (max_cap - min_cap < max_cap >> 3)
6220 /* Bring task utilization in sync with prev_cpu */
6221 sync_entity_load_avg(&p->se);
6223 return min_cap * 1024 < task_util(p) * capacity_margin;
6227 * select_task_rq_fair: Select target runqueue for the waking task in domains
6228 * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE,
6229 * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
6231 * Balances load by selecting the idlest CPU in the idlest group, or under
6232 * certain conditions an idle sibling CPU if the domain has SD_WAKE_AFFINE set.
6234 * Returns the target CPU number.
6236 * preempt must be disabled.
6239 select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags)
6241 struct sched_domain *tmp, *sd = NULL;
6242 int cpu = smp_processor_id();
6243 int new_cpu = prev_cpu;
6244 int want_affine = 0;
6245 int sync = (wake_flags & WF_SYNC) && !(current->flags & PF_EXITING);
6247 if (sd_flag & SD_BALANCE_WAKE) {
6249 want_affine = !wake_wide(p) && !wake_cap(p, cpu, prev_cpu)
6250 && cpumask_test_cpu(cpu, &p->cpus_allowed);
6254 for_each_domain(cpu, tmp) {
6255 if (!(tmp->flags & SD_LOAD_BALANCE))
6259 * If both 'cpu' and 'prev_cpu' are part of this domain,
6260 * cpu is a valid SD_WAKE_AFFINE target.
6262 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
6263 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
6264 if (cpu != prev_cpu)
6265 new_cpu = wake_affine(tmp, p, cpu, prev_cpu, sync);
6267 sd = NULL; /* Prefer wake_affine over balance flags */
6271 if (tmp->flags & sd_flag)
6273 else if (!want_affine)
6279 new_cpu = find_idlest_cpu(sd, p, cpu, prev_cpu, sd_flag);
6280 } else if (sd_flag & SD_BALANCE_WAKE) { /* XXX always ? */
6283 new_cpu = select_idle_sibling(p, prev_cpu, new_cpu);
6286 current->recent_used_cpu = cpu;
6293 static void detach_entity_cfs_rq(struct sched_entity *se);
6296 * Called immediately before a task is migrated to a new CPU; task_cpu(p) and
6297 * cfs_rq_of(p) references at time of call are still valid and identify the
6298 * previous CPU. The caller guarantees p->pi_lock or task_rq(p)->lock is held.
6300 static void migrate_task_rq_fair(struct task_struct *p)
6303 * As blocked tasks retain absolute vruntime the migration needs to
6304 * deal with this by subtracting the old and adding the new
6305 * min_vruntime -- the latter is done by enqueue_entity() when placing
6306 * the task on the new runqueue.
6308 if (p->state == TASK_WAKING) {
6309 struct sched_entity *se = &p->se;
6310 struct cfs_rq *cfs_rq = cfs_rq_of(se);
6313 #ifndef CONFIG_64BIT
6314 u64 min_vruntime_copy;
6317 min_vruntime_copy = cfs_rq->min_vruntime_copy;
6319 min_vruntime = cfs_rq->min_vruntime;
6320 } while (min_vruntime != min_vruntime_copy);
6322 min_vruntime = cfs_rq->min_vruntime;
6325 se->vruntime -= min_vruntime;
6328 if (p->on_rq == TASK_ON_RQ_MIGRATING) {
6330 * In case of TASK_ON_RQ_MIGRATING we in fact hold the 'old'
6331 * rq->lock and can modify state directly.
6333 lockdep_assert_held(&task_rq(p)->lock);
6334 detach_entity_cfs_rq(&p->se);
6338 * We are supposed to update the task to "current" time, then
6339 * its up to date and ready to go to new CPU/cfs_rq. But we
6340 * have difficulty in getting what current time is, so simply
6341 * throw away the out-of-date time. This will result in the
6342 * wakee task is less decayed, but giving the wakee more load
6345 remove_entity_load_avg(&p->se);
6348 /* Tell new CPU we are migrated */
6349 p->se.avg.last_update_time = 0;
6351 /* We have migrated, no longer consider this task hot */
6352 p->se.exec_start = 0;
6355 static void task_dead_fair(struct task_struct *p)
6357 remove_entity_load_avg(&p->se);
6359 #endif /* CONFIG_SMP */
6361 static unsigned long wakeup_gran(struct sched_entity *se)
6363 unsigned long gran = sysctl_sched_wakeup_granularity;
6366 * Since its curr running now, convert the gran from real-time
6367 * to virtual-time in his units.
6369 * By using 'se' instead of 'curr' we penalize light tasks, so
6370 * they get preempted easier. That is, if 'se' < 'curr' then
6371 * the resulting gran will be larger, therefore penalizing the
6372 * lighter, if otoh 'se' > 'curr' then the resulting gran will
6373 * be smaller, again penalizing the lighter task.
6375 * This is especially important for buddies when the leftmost
6376 * task is higher priority than the buddy.
6378 return calc_delta_fair(gran, se);
6382 * Should 'se' preempt 'curr'.
6396 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
6398 s64 gran, vdiff = curr->vruntime - se->vruntime;
6403 gran = wakeup_gran(se);
6410 static void set_last_buddy(struct sched_entity *se)
6412 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
6415 for_each_sched_entity(se) {
6416 if (SCHED_WARN_ON(!se->on_rq))
6418 cfs_rq_of(se)->last = se;
6422 static void set_next_buddy(struct sched_entity *se)
6424 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
6427 for_each_sched_entity(se) {
6428 if (SCHED_WARN_ON(!se->on_rq))
6430 cfs_rq_of(se)->next = se;
6434 static void set_skip_buddy(struct sched_entity *se)
6436 for_each_sched_entity(se)
6437 cfs_rq_of(se)->skip = se;
6441 * Preempt the current task with a newly woken task if needed:
6443 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
6445 struct task_struct *curr = rq->curr;
6446 struct sched_entity *se = &curr->se, *pse = &p->se;
6447 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
6448 int scale = cfs_rq->nr_running >= sched_nr_latency;
6449 int next_buddy_marked = 0;
6451 if (unlikely(se == pse))
6455 * This is possible from callers such as attach_tasks(), in which we
6456 * unconditionally check_prempt_curr() after an enqueue (which may have
6457 * lead to a throttle). This both saves work and prevents false
6458 * next-buddy nomination below.
6460 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
6463 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
6464 set_next_buddy(pse);
6465 next_buddy_marked = 1;
6469 * We can come here with TIF_NEED_RESCHED already set from new task
6472 * Note: this also catches the edge-case of curr being in a throttled
6473 * group (e.g. via set_curr_task), since update_curr() (in the
6474 * enqueue of curr) will have resulted in resched being set. This
6475 * prevents us from potentially nominating it as a false LAST_BUDDY
6478 if (test_tsk_need_resched(curr))
6481 /* Idle tasks are by definition preempted by non-idle tasks. */
6482 if (unlikely(curr->policy == SCHED_IDLE) &&
6483 likely(p->policy != SCHED_IDLE))
6487 * Batch and idle tasks do not preempt non-idle tasks (their preemption
6488 * is driven by the tick):
6490 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
6493 find_matching_se(&se, &pse);
6494 update_curr(cfs_rq_of(se));
6496 if (wakeup_preempt_entity(se, pse) == 1) {
6498 * Bias pick_next to pick the sched entity that is
6499 * triggering this preemption.
6501 if (!next_buddy_marked)
6502 set_next_buddy(pse);
6511 * Only set the backward buddy when the current task is still
6512 * on the rq. This can happen when a wakeup gets interleaved
6513 * with schedule on the ->pre_schedule() or idle_balance()
6514 * point, either of which can * drop the rq lock.
6516 * Also, during early boot the idle thread is in the fair class,
6517 * for obvious reasons its a bad idea to schedule back to it.
6519 if (unlikely(!se->on_rq || curr == rq->idle))
6522 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
6526 static struct task_struct *
6527 pick_next_task_fair(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
6529 struct cfs_rq *cfs_rq = &rq->cfs;
6530 struct sched_entity *se;
6531 struct task_struct *p;
6535 if (!cfs_rq->nr_running)
6538 #ifdef CONFIG_FAIR_GROUP_SCHED
6539 if (prev->sched_class != &fair_sched_class)
6543 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
6544 * likely that a next task is from the same cgroup as the current.
6546 * Therefore attempt to avoid putting and setting the entire cgroup
6547 * hierarchy, only change the part that actually changes.
6551 struct sched_entity *curr = cfs_rq->curr;
6554 * Since we got here without doing put_prev_entity() we also
6555 * have to consider cfs_rq->curr. If it is still a runnable
6556 * entity, update_curr() will update its vruntime, otherwise
6557 * forget we've ever seen it.
6561 update_curr(cfs_rq);
6566 * This call to check_cfs_rq_runtime() will do the
6567 * throttle and dequeue its entity in the parent(s).
6568 * Therefore the nr_running test will indeed
6571 if (unlikely(check_cfs_rq_runtime(cfs_rq))) {
6574 if (!cfs_rq->nr_running)
6581 se = pick_next_entity(cfs_rq, curr);
6582 cfs_rq = group_cfs_rq(se);
6588 * Since we haven't yet done put_prev_entity and if the selected task
6589 * is a different task than we started out with, try and touch the
6590 * least amount of cfs_rqs.
6593 struct sched_entity *pse = &prev->se;
6595 while (!(cfs_rq = is_same_group(se, pse))) {
6596 int se_depth = se->depth;
6597 int pse_depth = pse->depth;
6599 if (se_depth <= pse_depth) {
6600 put_prev_entity(cfs_rq_of(pse), pse);
6601 pse = parent_entity(pse);
6603 if (se_depth >= pse_depth) {
6604 set_next_entity(cfs_rq_of(se), se);
6605 se = parent_entity(se);
6609 put_prev_entity(cfs_rq, pse);
6610 set_next_entity(cfs_rq, se);
6617 put_prev_task(rq, prev);
6620 se = pick_next_entity(cfs_rq, NULL);
6621 set_next_entity(cfs_rq, se);
6622 cfs_rq = group_cfs_rq(se);
6627 done: __maybe_unused;
6630 * Move the next running task to the front of
6631 * the list, so our cfs_tasks list becomes MRU
6634 list_move(&p->se.group_node, &rq->cfs_tasks);
6637 if (hrtick_enabled(rq))
6638 hrtick_start_fair(rq, p);
6643 new_tasks = idle_balance(rq, rf);
6646 * Because idle_balance() releases (and re-acquires) rq->lock, it is
6647 * possible for any higher priority task to appear. In that case we
6648 * must re-start the pick_next_entity() loop.
6660 * Account for a descheduled task:
6662 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
6664 struct sched_entity *se = &prev->se;
6665 struct cfs_rq *cfs_rq;
6667 for_each_sched_entity(se) {
6668 cfs_rq = cfs_rq_of(se);
6669 put_prev_entity(cfs_rq, se);
6674 * sched_yield() is very simple
6676 * The magic of dealing with the ->skip buddy is in pick_next_entity.
6678 static void yield_task_fair(struct rq *rq)
6680 struct task_struct *curr = rq->curr;
6681 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
6682 struct sched_entity *se = &curr->se;
6685 * Are we the only task in the tree?
6687 if (unlikely(rq->nr_running == 1))
6690 clear_buddies(cfs_rq, se);
6692 if (curr->policy != SCHED_BATCH) {
6693 update_rq_clock(rq);
6695 * Update run-time statistics of the 'current'.
6697 update_curr(cfs_rq);
6699 * Tell update_rq_clock() that we've just updated,
6700 * so we don't do microscopic update in schedule()
6701 * and double the fastpath cost.
6703 rq_clock_skip_update(rq);
6709 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
6711 struct sched_entity *se = &p->se;
6713 /* throttled hierarchies are not runnable */
6714 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
6717 /* Tell the scheduler that we'd really like pse to run next. */
6720 yield_task_fair(rq);
6726 /**************************************************
6727 * Fair scheduling class load-balancing methods.
6731 * The purpose of load-balancing is to achieve the same basic fairness the
6732 * per-CPU scheduler provides, namely provide a proportional amount of compute
6733 * time to each task. This is expressed in the following equation:
6735 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
6737 * Where W_i,n is the n-th weight average for CPU i. The instantaneous weight
6738 * W_i,0 is defined as:
6740 * W_i,0 = \Sum_j w_i,j (2)
6742 * Where w_i,j is the weight of the j-th runnable task on CPU i. This weight
6743 * is derived from the nice value as per sched_prio_to_weight[].
6745 * The weight average is an exponential decay average of the instantaneous
6748 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
6750 * C_i is the compute capacity of CPU i, typically it is the
6751 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
6752 * can also include other factors [XXX].
6754 * To achieve this balance we define a measure of imbalance which follows
6755 * directly from (1):
6757 * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4)
6759 * We them move tasks around to minimize the imbalance. In the continuous
6760 * function space it is obvious this converges, in the discrete case we get
6761 * a few fun cases generally called infeasible weight scenarios.
6764 * - infeasible weights;
6765 * - local vs global optima in the discrete case. ]
6770 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
6771 * for all i,j solution, we create a tree of CPUs that follows the hardware
6772 * topology where each level pairs two lower groups (or better). This results
6773 * in O(log n) layers. Furthermore we reduce the number of CPUs going up the
6774 * tree to only the first of the previous level and we decrease the frequency
6775 * of load-balance at each level inv. proportional to the number of CPUs in
6781 * \Sum { --- * --- * 2^i } = O(n) (5)
6783 * `- size of each group
6784 * | | `- number of CPUs doing load-balance
6786 * `- sum over all levels
6788 * Coupled with a limit on how many tasks we can migrate every balance pass,
6789 * this makes (5) the runtime complexity of the balancer.
6791 * An important property here is that each CPU is still (indirectly) connected
6792 * to every other CPU in at most O(log n) steps:
6794 * The adjacency matrix of the resulting graph is given by:
6797 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
6800 * And you'll find that:
6802 * A^(log_2 n)_i,j != 0 for all i,j (7)
6804 * Showing there's indeed a path between every CPU in at most O(log n) steps.
6805 * The task movement gives a factor of O(m), giving a convergence complexity
6808 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
6813 * In order to avoid CPUs going idle while there's still work to do, new idle
6814 * balancing is more aggressive and has the newly idle CPU iterate up the domain
6815 * tree itself instead of relying on other CPUs to bring it work.
6817 * This adds some complexity to both (5) and (8) but it reduces the total idle
6825 * Cgroups make a horror show out of (2), instead of a simple sum we get:
6828 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
6833 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
6835 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on CPU i.
6837 * The big problem is S_k, its a global sum needed to compute a local (W_i)
6840 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
6841 * rewrite all of this once again.]
6844 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
6846 enum fbq_type { regular, remote, all };
6848 #define LBF_ALL_PINNED 0x01
6849 #define LBF_NEED_BREAK 0x02
6850 #define LBF_DST_PINNED 0x04
6851 #define LBF_SOME_PINNED 0x08
6852 #define LBF_NOHZ_STATS 0x10
6853 #define LBF_NOHZ_AGAIN 0x20
6856 struct sched_domain *sd;
6864 struct cpumask *dst_grpmask;
6866 enum cpu_idle_type idle;
6868 /* The set of CPUs under consideration for load-balancing */
6869 struct cpumask *cpus;
6874 unsigned int loop_break;
6875 unsigned int loop_max;
6877 enum fbq_type fbq_type;
6878 struct list_head tasks;
6882 * Is this task likely cache-hot:
6884 static int task_hot(struct task_struct *p, struct lb_env *env)
6888 lockdep_assert_held(&env->src_rq->lock);
6890 if (p->sched_class != &fair_sched_class)
6893 if (unlikely(p->policy == SCHED_IDLE))
6897 * Buddy candidates are cache hot:
6899 if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running &&
6900 (&p->se == cfs_rq_of(&p->se)->next ||
6901 &p->se == cfs_rq_of(&p->se)->last))
6904 if (sysctl_sched_migration_cost == -1)
6906 if (sysctl_sched_migration_cost == 0)
6909 delta = rq_clock_task(env->src_rq) - p->se.exec_start;
6911 return delta < (s64)sysctl_sched_migration_cost;
6914 #ifdef CONFIG_NUMA_BALANCING
6916 * Returns 1, if task migration degrades locality
6917 * Returns 0, if task migration improves locality i.e migration preferred.
6918 * Returns -1, if task migration is not affected by locality.
6920 static int migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
6922 struct numa_group *numa_group = rcu_dereference(p->numa_group);
6923 unsigned long src_weight, dst_weight;
6924 int src_nid, dst_nid, dist;
6926 if (!static_branch_likely(&sched_numa_balancing))
6929 if (!p->numa_faults || !(env->sd->flags & SD_NUMA))
6932 src_nid = cpu_to_node(env->src_cpu);
6933 dst_nid = cpu_to_node(env->dst_cpu);
6935 if (src_nid == dst_nid)
6938 /* Migrating away from the preferred node is always bad. */
6939 if (src_nid == p->numa_preferred_nid) {
6940 if (env->src_rq->nr_running > env->src_rq->nr_preferred_running)
6946 /* Encourage migration to the preferred node. */
6947 if (dst_nid == p->numa_preferred_nid)
6950 /* Leaving a core idle is often worse than degrading locality. */
6951 if (env->idle == CPU_IDLE)
6954 dist = node_distance(src_nid, dst_nid);
6956 src_weight = group_weight(p, src_nid, dist);
6957 dst_weight = group_weight(p, dst_nid, dist);
6959 src_weight = task_weight(p, src_nid, dist);
6960 dst_weight = task_weight(p, dst_nid, dist);
6963 return dst_weight < src_weight;
6967 static inline int migrate_degrades_locality(struct task_struct *p,
6975 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
6978 int can_migrate_task(struct task_struct *p, struct lb_env *env)
6982 lockdep_assert_held(&env->src_rq->lock);
6985 * We do not migrate tasks that are:
6986 * 1) throttled_lb_pair, or
6987 * 2) cannot be migrated to this CPU due to cpus_allowed, or
6988 * 3) running (obviously), or
6989 * 4) are cache-hot on their current CPU.
6991 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
6994 if (!cpumask_test_cpu(env->dst_cpu, &p->cpus_allowed)) {
6997 schedstat_inc(p->se.statistics.nr_failed_migrations_affine);
6999 env->flags |= LBF_SOME_PINNED;
7002 * Remember if this task can be migrated to any other CPU in
7003 * our sched_group. We may want to revisit it if we couldn't
7004 * meet load balance goals by pulling other tasks on src_cpu.
7006 * Avoid computing new_dst_cpu for NEWLY_IDLE or if we have
7007 * already computed one in current iteration.
7009 if (env->idle == CPU_NEWLY_IDLE || (env->flags & LBF_DST_PINNED))
7012 /* Prevent to re-select dst_cpu via env's CPUs: */
7013 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
7014 if (cpumask_test_cpu(cpu, &p->cpus_allowed)) {
7015 env->flags |= LBF_DST_PINNED;
7016 env->new_dst_cpu = cpu;
7024 /* Record that we found atleast one task that could run on dst_cpu */
7025 env->flags &= ~LBF_ALL_PINNED;
7027 if (task_running(env->src_rq, p)) {
7028 schedstat_inc(p->se.statistics.nr_failed_migrations_running);
7033 * Aggressive migration if:
7034 * 1) destination numa is preferred
7035 * 2) task is cache cold, or
7036 * 3) too many balance attempts have failed.
7038 tsk_cache_hot = migrate_degrades_locality(p, env);
7039 if (tsk_cache_hot == -1)
7040 tsk_cache_hot = task_hot(p, env);
7042 if (tsk_cache_hot <= 0 ||
7043 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
7044 if (tsk_cache_hot == 1) {
7045 schedstat_inc(env->sd->lb_hot_gained[env->idle]);
7046 schedstat_inc(p->se.statistics.nr_forced_migrations);
7051 schedstat_inc(p->se.statistics.nr_failed_migrations_hot);
7056 * detach_task() -- detach the task for the migration specified in env
7058 static void detach_task(struct task_struct *p, struct lb_env *env)
7060 lockdep_assert_held(&env->src_rq->lock);
7062 p->on_rq = TASK_ON_RQ_MIGRATING;
7063 deactivate_task(env->src_rq, p, DEQUEUE_NOCLOCK);
7064 set_task_cpu(p, env->dst_cpu);
7068 * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
7069 * part of active balancing operations within "domain".
7071 * Returns a task if successful and NULL otherwise.
7073 static struct task_struct *detach_one_task(struct lb_env *env)
7075 struct task_struct *p;
7077 lockdep_assert_held(&env->src_rq->lock);
7079 list_for_each_entry_reverse(p,
7080 &env->src_rq->cfs_tasks, se.group_node) {
7081 if (!can_migrate_task(p, env))
7084 detach_task(p, env);
7087 * Right now, this is only the second place where
7088 * lb_gained[env->idle] is updated (other is detach_tasks)
7089 * so we can safely collect stats here rather than
7090 * inside detach_tasks().
7092 schedstat_inc(env->sd->lb_gained[env->idle]);
7098 static const unsigned int sched_nr_migrate_break = 32;
7101 * detach_tasks() -- tries to detach up to imbalance weighted load from
7102 * busiest_rq, as part of a balancing operation within domain "sd".
7104 * Returns number of detached tasks if successful and 0 otherwise.
7106 static int detach_tasks(struct lb_env *env)
7108 struct list_head *tasks = &env->src_rq->cfs_tasks;
7109 struct task_struct *p;
7113 lockdep_assert_held(&env->src_rq->lock);
7115 if (env->imbalance <= 0)
7118 while (!list_empty(tasks)) {
7120 * We don't want to steal all, otherwise we may be treated likewise,
7121 * which could at worst lead to a livelock crash.
7123 if (env->idle != CPU_NOT_IDLE && env->src_rq->nr_running <= 1)
7126 p = list_last_entry(tasks, struct task_struct, se.group_node);
7129 /* We've more or less seen every task there is, call it quits */
7130 if (env->loop > env->loop_max)
7133 /* take a breather every nr_migrate tasks */
7134 if (env->loop > env->loop_break) {
7135 env->loop_break += sched_nr_migrate_break;
7136 env->flags |= LBF_NEED_BREAK;
7140 if (!can_migrate_task(p, env))
7143 load = task_h_load(p);
7145 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
7148 if ((load / 2) > env->imbalance)
7151 detach_task(p, env);
7152 list_add(&p->se.group_node, &env->tasks);
7155 env->imbalance -= load;
7157 #ifdef CONFIG_PREEMPT
7159 * NEWIDLE balancing is a source of latency, so preemptible
7160 * kernels will stop after the first task is detached to minimize
7161 * the critical section.
7163 if (env->idle == CPU_NEWLY_IDLE)
7168 * We only want to steal up to the prescribed amount of
7171 if (env->imbalance <= 0)
7176 list_move(&p->se.group_node, tasks);
7180 * Right now, this is one of only two places we collect this stat
7181 * so we can safely collect detach_one_task() stats here rather
7182 * than inside detach_one_task().
7184 schedstat_add(env->sd->lb_gained[env->idle], detached);
7190 * attach_task() -- attach the task detached by detach_task() to its new rq.
7192 static void attach_task(struct rq *rq, struct task_struct *p)
7194 lockdep_assert_held(&rq->lock);
7196 BUG_ON(task_rq(p) != rq);
7197 activate_task(rq, p, ENQUEUE_NOCLOCK);
7198 p->on_rq = TASK_ON_RQ_QUEUED;
7199 check_preempt_curr(rq, p, 0);
7203 * attach_one_task() -- attaches the task returned from detach_one_task() to
7206 static void attach_one_task(struct rq *rq, struct task_struct *p)
7211 update_rq_clock(rq);
7217 * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
7220 static void attach_tasks(struct lb_env *env)
7222 struct list_head *tasks = &env->tasks;
7223 struct task_struct *p;
7226 rq_lock(env->dst_rq, &rf);
7227 update_rq_clock(env->dst_rq);
7229 while (!list_empty(tasks)) {
7230 p = list_first_entry(tasks, struct task_struct, se.group_node);
7231 list_del_init(&p->se.group_node);
7233 attach_task(env->dst_rq, p);
7236 rq_unlock(env->dst_rq, &rf);
7239 static inline bool cfs_rq_has_blocked(struct cfs_rq *cfs_rq)
7241 if (cfs_rq->avg.load_avg)
7244 if (cfs_rq->avg.util_avg)
7250 static inline bool others_have_blocked(struct rq *rq)
7252 if (READ_ONCE(rq->avg_rt.util_avg))
7255 if (READ_ONCE(rq->avg_dl.util_avg))
7258 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
7259 if (READ_ONCE(rq->avg_irq.util_avg))
7266 #ifdef CONFIG_FAIR_GROUP_SCHED
7268 static inline bool cfs_rq_is_decayed(struct cfs_rq *cfs_rq)
7270 if (cfs_rq->load.weight)
7273 if (cfs_rq->avg.load_sum)
7276 if (cfs_rq->avg.util_sum)
7279 if (cfs_rq->avg.runnable_load_sum)
7285 static void update_blocked_averages(int cpu)
7287 struct rq *rq = cpu_rq(cpu);
7288 struct cfs_rq *cfs_rq, *pos;
7289 const struct sched_class *curr_class;
7293 rq_lock_irqsave(rq, &rf);
7294 update_rq_clock(rq);
7297 * Iterates the task_group tree in a bottom up fashion, see
7298 * list_add_leaf_cfs_rq() for details.
7300 for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) {
7301 struct sched_entity *se;
7303 /* throttled entities do not contribute to load */
7304 if (throttled_hierarchy(cfs_rq))
7307 if (update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq))
7308 update_tg_load_avg(cfs_rq, 0);
7310 /* Propagate pending load changes to the parent, if any: */
7311 se = cfs_rq->tg->se[cpu];
7312 if (se && !skip_blocked_update(se))
7313 update_load_avg(cfs_rq_of(se), se, 0);
7316 * There can be a lot of idle CPU cgroups. Don't let fully
7317 * decayed cfs_rqs linger on the list.
7319 if (cfs_rq_is_decayed(cfs_rq))
7320 list_del_leaf_cfs_rq(cfs_rq);
7322 /* Don't need periodic decay once load/util_avg are null */
7323 if (cfs_rq_has_blocked(cfs_rq))
7327 curr_class = rq->curr->sched_class;
7328 update_rt_rq_load_avg(rq_clock_task(rq), rq, curr_class == &rt_sched_class);
7329 update_dl_rq_load_avg(rq_clock_task(rq), rq, curr_class == &dl_sched_class);
7330 update_irq_load_avg(rq, 0);
7331 /* Don't need periodic decay once load/util_avg are null */
7332 if (others_have_blocked(rq))
7335 #ifdef CONFIG_NO_HZ_COMMON
7336 rq->last_blocked_load_update_tick = jiffies;
7338 rq->has_blocked_load = 0;
7340 rq_unlock_irqrestore(rq, &rf);
7344 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
7345 * This needs to be done in a top-down fashion because the load of a child
7346 * group is a fraction of its parents load.
7348 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
7350 struct rq *rq = rq_of(cfs_rq);
7351 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
7352 unsigned long now = jiffies;
7355 if (cfs_rq->last_h_load_update == now)
7358 cfs_rq->h_load_next = NULL;
7359 for_each_sched_entity(se) {
7360 cfs_rq = cfs_rq_of(se);
7361 cfs_rq->h_load_next = se;
7362 if (cfs_rq->last_h_load_update == now)
7367 cfs_rq->h_load = cfs_rq_load_avg(cfs_rq);
7368 cfs_rq->last_h_load_update = now;
7371 while ((se = cfs_rq->h_load_next) != NULL) {
7372 load = cfs_rq->h_load;
7373 load = div64_ul(load * se->avg.load_avg,
7374 cfs_rq_load_avg(cfs_rq) + 1);
7375 cfs_rq = group_cfs_rq(se);
7376 cfs_rq->h_load = load;
7377 cfs_rq->last_h_load_update = now;
7381 static unsigned long task_h_load(struct task_struct *p)
7383 struct cfs_rq *cfs_rq = task_cfs_rq(p);
7385 update_cfs_rq_h_load(cfs_rq);
7386 return div64_ul(p->se.avg.load_avg * cfs_rq->h_load,
7387 cfs_rq_load_avg(cfs_rq) + 1);
7390 static inline void update_blocked_averages(int cpu)
7392 struct rq *rq = cpu_rq(cpu);
7393 struct cfs_rq *cfs_rq = &rq->cfs;
7394 const struct sched_class *curr_class;
7397 rq_lock_irqsave(rq, &rf);
7398 update_rq_clock(rq);
7399 update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq);
7401 curr_class = rq->curr->sched_class;
7402 update_rt_rq_load_avg(rq_clock_task(rq), rq, curr_class == &rt_sched_class);
7403 update_dl_rq_load_avg(rq_clock_task(rq), rq, curr_class == &dl_sched_class);
7404 update_irq_load_avg(rq, 0);
7405 #ifdef CONFIG_NO_HZ_COMMON
7406 rq->last_blocked_load_update_tick = jiffies;
7407 if (!cfs_rq_has_blocked(cfs_rq) && !others_have_blocked(rq))
7408 rq->has_blocked_load = 0;
7410 rq_unlock_irqrestore(rq, &rf);
7413 static unsigned long task_h_load(struct task_struct *p)
7415 return p->se.avg.load_avg;
7419 /********** Helpers for find_busiest_group ************************/
7428 * sg_lb_stats - stats of a sched_group required for load_balancing
7430 struct sg_lb_stats {
7431 unsigned long avg_load; /*Avg load across the CPUs of the group */
7432 unsigned long group_load; /* Total load over the CPUs of the group */
7433 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
7434 unsigned long load_per_task;
7435 unsigned long group_capacity;
7436 unsigned long group_util; /* Total utilization of the group */
7437 unsigned int sum_nr_running; /* Nr tasks running in the group */
7438 unsigned int idle_cpus;
7439 unsigned int group_weight;
7440 enum group_type group_type;
7441 int group_no_capacity;
7442 #ifdef CONFIG_NUMA_BALANCING
7443 unsigned int nr_numa_running;
7444 unsigned int nr_preferred_running;
7449 * sd_lb_stats - Structure to store the statistics of a sched_domain
7450 * during load balancing.
7452 struct sd_lb_stats {
7453 struct sched_group *busiest; /* Busiest group in this sd */
7454 struct sched_group *local; /* Local group in this sd */
7455 unsigned long total_running;
7456 unsigned long total_load; /* Total load of all groups in sd */
7457 unsigned long total_capacity; /* Total capacity of all groups in sd */
7458 unsigned long avg_load; /* Average load across all groups in sd */
7460 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
7461 struct sg_lb_stats local_stat; /* Statistics of the local group */
7464 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
7467 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
7468 * local_stat because update_sg_lb_stats() does a full clear/assignment.
7469 * We must however clear busiest_stat::avg_load because
7470 * update_sd_pick_busiest() reads this before assignment.
7472 *sds = (struct sd_lb_stats){
7475 .total_running = 0UL,
7477 .total_capacity = 0UL,
7480 .sum_nr_running = 0,
7481 .group_type = group_other,
7487 * get_sd_load_idx - Obtain the load index for a given sched domain.
7488 * @sd: The sched_domain whose load_idx is to be obtained.
7489 * @idle: The idle status of the CPU for whose sd load_idx is obtained.
7491 * Return: The load index.
7493 static inline int get_sd_load_idx(struct sched_domain *sd,
7494 enum cpu_idle_type idle)
7500 load_idx = sd->busy_idx;
7503 case CPU_NEWLY_IDLE:
7504 load_idx = sd->newidle_idx;
7507 load_idx = sd->idle_idx;
7514 static unsigned long scale_rt_capacity(struct sched_domain *sd, int cpu)
7516 struct rq *rq = cpu_rq(cpu);
7517 unsigned long max = arch_scale_cpu_capacity(sd, cpu);
7518 unsigned long used, free;
7521 irq = cpu_util_irq(rq);
7523 if (unlikely(irq >= max))
7526 used = READ_ONCE(rq->avg_rt.util_avg);
7527 used += READ_ONCE(rq->avg_dl.util_avg);
7529 if (unlikely(used >= max))
7534 return scale_irq_capacity(free, irq, max);
7537 static void update_cpu_capacity(struct sched_domain *sd, int cpu)
7539 unsigned long capacity = scale_rt_capacity(sd, cpu);
7540 struct sched_group *sdg = sd->groups;
7542 cpu_rq(cpu)->cpu_capacity_orig = arch_scale_cpu_capacity(sd, cpu);
7547 cpu_rq(cpu)->cpu_capacity = capacity;
7548 sdg->sgc->capacity = capacity;
7549 sdg->sgc->min_capacity = capacity;
7552 void update_group_capacity(struct sched_domain *sd, int cpu)
7554 struct sched_domain *child = sd->child;
7555 struct sched_group *group, *sdg = sd->groups;
7556 unsigned long capacity, min_capacity;
7557 unsigned long interval;
7559 interval = msecs_to_jiffies(sd->balance_interval);
7560 interval = clamp(interval, 1UL, max_load_balance_interval);
7561 sdg->sgc->next_update = jiffies + interval;
7564 update_cpu_capacity(sd, cpu);
7569 min_capacity = ULONG_MAX;
7571 if (child->flags & SD_OVERLAP) {
7573 * SD_OVERLAP domains cannot assume that child groups
7574 * span the current group.
7577 for_each_cpu(cpu, sched_group_span(sdg)) {
7578 struct sched_group_capacity *sgc;
7579 struct rq *rq = cpu_rq(cpu);
7582 * build_sched_domains() -> init_sched_groups_capacity()
7583 * gets here before we've attached the domains to the
7586 * Use capacity_of(), which is set irrespective of domains
7587 * in update_cpu_capacity().
7589 * This avoids capacity from being 0 and
7590 * causing divide-by-zero issues on boot.
7592 if (unlikely(!rq->sd)) {
7593 capacity += capacity_of(cpu);
7595 sgc = rq->sd->groups->sgc;
7596 capacity += sgc->capacity;
7599 min_capacity = min(capacity, min_capacity);
7603 * !SD_OVERLAP domains can assume that child groups
7604 * span the current group.
7607 group = child->groups;
7609 struct sched_group_capacity *sgc = group->sgc;
7611 capacity += sgc->capacity;
7612 min_capacity = min(sgc->min_capacity, min_capacity);
7613 group = group->next;
7614 } while (group != child->groups);
7617 sdg->sgc->capacity = capacity;
7618 sdg->sgc->min_capacity = min_capacity;
7622 * Check whether the capacity of the rq has been noticeably reduced by side
7623 * activity. The imbalance_pct is used for the threshold.
7624 * Return true is the capacity is reduced
7627 check_cpu_capacity(struct rq *rq, struct sched_domain *sd)
7629 return ((rq->cpu_capacity * sd->imbalance_pct) <
7630 (rq->cpu_capacity_orig * 100));
7634 * Group imbalance indicates (and tries to solve) the problem where balancing
7635 * groups is inadequate due to ->cpus_allowed constraints.
7637 * Imagine a situation of two groups of 4 CPUs each and 4 tasks each with a
7638 * cpumask covering 1 CPU of the first group and 3 CPUs of the second group.
7641 * { 0 1 2 3 } { 4 5 6 7 }
7644 * If we were to balance group-wise we'd place two tasks in the first group and
7645 * two tasks in the second group. Clearly this is undesired as it will overload
7646 * cpu 3 and leave one of the CPUs in the second group unused.
7648 * The current solution to this issue is detecting the skew in the first group
7649 * by noticing the lower domain failed to reach balance and had difficulty
7650 * moving tasks due to affinity constraints.
7652 * When this is so detected; this group becomes a candidate for busiest; see
7653 * update_sd_pick_busiest(). And calculate_imbalance() and
7654 * find_busiest_group() avoid some of the usual balance conditions to allow it
7655 * to create an effective group imbalance.
7657 * This is a somewhat tricky proposition since the next run might not find the
7658 * group imbalance and decide the groups need to be balanced again. A most
7659 * subtle and fragile situation.
7662 static inline int sg_imbalanced(struct sched_group *group)
7664 return group->sgc->imbalance;
7668 * group_has_capacity returns true if the group has spare capacity that could
7669 * be used by some tasks.
7670 * We consider that a group has spare capacity if the * number of task is
7671 * smaller than the number of CPUs or if the utilization is lower than the
7672 * available capacity for CFS tasks.
7673 * For the latter, we use a threshold to stabilize the state, to take into
7674 * account the variance of the tasks' load and to return true if the available
7675 * capacity in meaningful for the load balancer.
7676 * As an example, an available capacity of 1% can appear but it doesn't make
7677 * any benefit for the load balance.
7680 group_has_capacity(struct lb_env *env, struct sg_lb_stats *sgs)
7682 if (sgs->sum_nr_running < sgs->group_weight)
7685 if ((sgs->group_capacity * 100) >
7686 (sgs->group_util * env->sd->imbalance_pct))
7693 * group_is_overloaded returns true if the group has more tasks than it can
7695 * group_is_overloaded is not equals to !group_has_capacity because a group
7696 * with the exact right number of tasks, has no more spare capacity but is not
7697 * overloaded so both group_has_capacity and group_is_overloaded return
7701 group_is_overloaded(struct lb_env *env, struct sg_lb_stats *sgs)
7703 if (sgs->sum_nr_running <= sgs->group_weight)
7706 if ((sgs->group_capacity * 100) <
7707 (sgs->group_util * env->sd->imbalance_pct))
7714 * group_smaller_cpu_capacity: Returns true if sched_group sg has smaller
7715 * per-CPU capacity than sched_group ref.
7718 group_smaller_cpu_capacity(struct sched_group *sg, struct sched_group *ref)
7720 return sg->sgc->min_capacity * capacity_margin <
7721 ref->sgc->min_capacity * 1024;
7725 group_type group_classify(struct sched_group *group,
7726 struct sg_lb_stats *sgs)
7728 if (sgs->group_no_capacity)
7729 return group_overloaded;
7731 if (sg_imbalanced(group))
7732 return group_imbalanced;
7737 static bool update_nohz_stats(struct rq *rq, bool force)
7739 #ifdef CONFIG_NO_HZ_COMMON
7740 unsigned int cpu = rq->cpu;
7742 if (!rq->has_blocked_load)
7745 if (!cpumask_test_cpu(cpu, nohz.idle_cpus_mask))
7748 if (!force && !time_after(jiffies, rq->last_blocked_load_update_tick))
7751 update_blocked_averages(cpu);
7753 return rq->has_blocked_load;
7760 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
7761 * @env: The load balancing environment.
7762 * @group: sched_group whose statistics are to be updated.
7763 * @load_idx: Load index of sched_domain of this_cpu for load calc.
7764 * @local_group: Does group contain this_cpu.
7765 * @sgs: variable to hold the statistics for this group.
7766 * @overload: Indicate more than one runnable task for any CPU.
7768 static inline void update_sg_lb_stats(struct lb_env *env,
7769 struct sched_group *group, int load_idx,
7770 int local_group, struct sg_lb_stats *sgs,
7776 memset(sgs, 0, sizeof(*sgs));
7778 for_each_cpu_and(i, sched_group_span(group), env->cpus) {
7779 struct rq *rq = cpu_rq(i);
7781 if ((env->flags & LBF_NOHZ_STATS) && update_nohz_stats(rq, false))
7782 env->flags |= LBF_NOHZ_AGAIN;
7784 /* Bias balancing toward CPUs of our domain: */
7786 load = target_load(i, load_idx);
7788 load = source_load(i, load_idx);
7790 sgs->group_load += load;
7791 sgs->group_util += cpu_util(i);
7792 sgs->sum_nr_running += rq->cfs.h_nr_running;
7794 nr_running = rq->nr_running;
7798 #ifdef CONFIG_NUMA_BALANCING
7799 sgs->nr_numa_running += rq->nr_numa_running;
7800 sgs->nr_preferred_running += rq->nr_preferred_running;
7802 sgs->sum_weighted_load += weighted_cpuload(rq);
7804 * No need to call idle_cpu() if nr_running is not 0
7806 if (!nr_running && idle_cpu(i))
7810 /* Adjust by relative CPU capacity of the group */
7811 sgs->group_capacity = group->sgc->capacity;
7812 sgs->avg_load = (sgs->group_load*SCHED_CAPACITY_SCALE) / sgs->group_capacity;
7814 if (sgs->sum_nr_running)
7815 sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
7817 sgs->group_weight = group->group_weight;
7819 sgs->group_no_capacity = group_is_overloaded(env, sgs);
7820 sgs->group_type = group_classify(group, sgs);
7824 * update_sd_pick_busiest - return 1 on busiest group
7825 * @env: The load balancing environment.
7826 * @sds: sched_domain statistics
7827 * @sg: sched_group candidate to be checked for being the busiest
7828 * @sgs: sched_group statistics
7830 * Determine if @sg is a busier group than the previously selected
7833 * Return: %true if @sg is a busier group than the previously selected
7834 * busiest group. %false otherwise.
7836 static bool update_sd_pick_busiest(struct lb_env *env,
7837 struct sd_lb_stats *sds,
7838 struct sched_group *sg,
7839 struct sg_lb_stats *sgs)
7841 struct sg_lb_stats *busiest = &sds->busiest_stat;
7843 if (sgs->group_type > busiest->group_type)
7846 if (sgs->group_type < busiest->group_type)
7849 if (sgs->avg_load <= busiest->avg_load)
7852 if (!(env->sd->flags & SD_ASYM_CPUCAPACITY))
7856 * Candidate sg has no more than one task per CPU and
7857 * has higher per-CPU capacity. Migrating tasks to less
7858 * capable CPUs may harm throughput. Maximize throughput,
7859 * power/energy consequences are not considered.
7861 if (sgs->sum_nr_running <= sgs->group_weight &&
7862 group_smaller_cpu_capacity(sds->local, sg))
7866 /* This is the busiest node in its class. */
7867 if (!(env->sd->flags & SD_ASYM_PACKING))
7870 /* No ASYM_PACKING if target CPU is already busy */
7871 if (env->idle == CPU_NOT_IDLE)
7874 * ASYM_PACKING needs to move all the work to the highest
7875 * prority CPUs in the group, therefore mark all groups
7876 * of lower priority than ourself as busy.
7878 if (sgs->sum_nr_running &&
7879 sched_asym_prefer(env->dst_cpu, sg->asym_prefer_cpu)) {
7883 /* Prefer to move from lowest priority CPU's work */
7884 if (sched_asym_prefer(sds->busiest->asym_prefer_cpu,
7885 sg->asym_prefer_cpu))
7892 #ifdef CONFIG_NUMA_BALANCING
7893 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
7895 if (sgs->sum_nr_running > sgs->nr_numa_running)
7897 if (sgs->sum_nr_running > sgs->nr_preferred_running)
7902 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
7904 if (rq->nr_running > rq->nr_numa_running)
7906 if (rq->nr_running > rq->nr_preferred_running)
7911 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
7916 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
7920 #endif /* CONFIG_NUMA_BALANCING */
7923 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
7924 * @env: The load balancing environment.
7925 * @sds: variable to hold the statistics for this sched_domain.
7927 static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
7929 struct sched_domain *child = env->sd->child;
7930 struct sched_group *sg = env->sd->groups;
7931 struct sg_lb_stats *local = &sds->local_stat;
7932 struct sg_lb_stats tmp_sgs;
7933 int load_idx, prefer_sibling = 0;
7934 bool overload = false;
7936 if (child && child->flags & SD_PREFER_SIBLING)
7939 #ifdef CONFIG_NO_HZ_COMMON
7940 if (env->idle == CPU_NEWLY_IDLE && READ_ONCE(nohz.has_blocked))
7941 env->flags |= LBF_NOHZ_STATS;
7944 load_idx = get_sd_load_idx(env->sd, env->idle);
7947 struct sg_lb_stats *sgs = &tmp_sgs;
7950 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_span(sg));
7955 if (env->idle != CPU_NEWLY_IDLE ||
7956 time_after_eq(jiffies, sg->sgc->next_update))
7957 update_group_capacity(env->sd, env->dst_cpu);
7960 update_sg_lb_stats(env, sg, load_idx, local_group, sgs,
7967 * In case the child domain prefers tasks go to siblings
7968 * first, lower the sg capacity so that we'll try
7969 * and move all the excess tasks away. We lower the capacity
7970 * of a group only if the local group has the capacity to fit
7971 * these excess tasks. The extra check prevents the case where
7972 * you always pull from the heaviest group when it is already
7973 * under-utilized (possible with a large weight task outweighs
7974 * the tasks on the system).
7976 if (prefer_sibling && sds->local &&
7977 group_has_capacity(env, local) &&
7978 (sgs->sum_nr_running > local->sum_nr_running + 1)) {
7979 sgs->group_no_capacity = 1;
7980 sgs->group_type = group_classify(sg, sgs);
7983 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
7985 sds->busiest_stat = *sgs;
7989 /* Now, start updating sd_lb_stats */
7990 sds->total_running += sgs->sum_nr_running;
7991 sds->total_load += sgs->group_load;
7992 sds->total_capacity += sgs->group_capacity;
7995 } while (sg != env->sd->groups);
7997 #ifdef CONFIG_NO_HZ_COMMON
7998 if ((env->flags & LBF_NOHZ_AGAIN) &&
7999 cpumask_subset(nohz.idle_cpus_mask, sched_domain_span(env->sd))) {
8001 WRITE_ONCE(nohz.next_blocked,
8002 jiffies + msecs_to_jiffies(LOAD_AVG_PERIOD));
8006 if (env->sd->flags & SD_NUMA)
8007 env->fbq_type = fbq_classify_group(&sds->busiest_stat);
8009 if (!env->sd->parent) {
8010 /* update overload indicator if we are at root domain */
8011 if (env->dst_rq->rd->overload != overload)
8012 env->dst_rq->rd->overload = overload;
8017 * check_asym_packing - Check to see if the group is packed into the
8020 * This is primarily intended to used at the sibling level. Some
8021 * cores like POWER7 prefer to use lower numbered SMT threads. In the
8022 * case of POWER7, it can move to lower SMT modes only when higher
8023 * threads are idle. When in lower SMT modes, the threads will
8024 * perform better since they share less core resources. Hence when we
8025 * have idle threads, we want them to be the higher ones.
8027 * This packing function is run on idle threads. It checks to see if
8028 * the busiest CPU in this domain (core in the P7 case) has a higher
8029 * CPU number than the packing function is being run on. Here we are
8030 * assuming lower CPU number will be equivalent to lower a SMT thread
8033 * Return: 1 when packing is required and a task should be moved to
8034 * this CPU. The amount of the imbalance is returned in env->imbalance.
8036 * @env: The load balancing environment.
8037 * @sds: Statistics of the sched_domain which is to be packed
8039 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
8043 if (!(env->sd->flags & SD_ASYM_PACKING))
8046 if (env->idle == CPU_NOT_IDLE)
8052 busiest_cpu = sds->busiest->asym_prefer_cpu;
8053 if (sched_asym_prefer(busiest_cpu, env->dst_cpu))
8056 env->imbalance = DIV_ROUND_CLOSEST(
8057 sds->busiest_stat.avg_load * sds->busiest_stat.group_capacity,
8058 SCHED_CAPACITY_SCALE);
8064 * fix_small_imbalance - Calculate the minor imbalance that exists
8065 * amongst the groups of a sched_domain, during
8067 * @env: The load balancing environment.
8068 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
8071 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
8073 unsigned long tmp, capa_now = 0, capa_move = 0;
8074 unsigned int imbn = 2;
8075 unsigned long scaled_busy_load_per_task;
8076 struct sg_lb_stats *local, *busiest;
8078 local = &sds->local_stat;
8079 busiest = &sds->busiest_stat;
8081 if (!local->sum_nr_running)
8082 local->load_per_task = cpu_avg_load_per_task(env->dst_cpu);
8083 else if (busiest->load_per_task > local->load_per_task)
8086 scaled_busy_load_per_task =
8087 (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
8088 busiest->group_capacity;
8090 if (busiest->avg_load + scaled_busy_load_per_task >=
8091 local->avg_load + (scaled_busy_load_per_task * imbn)) {
8092 env->imbalance = busiest->load_per_task;
8097 * OK, we don't have enough imbalance to justify moving tasks,
8098 * however we may be able to increase total CPU capacity used by
8102 capa_now += busiest->group_capacity *
8103 min(busiest->load_per_task, busiest->avg_load);
8104 capa_now += local->group_capacity *
8105 min(local->load_per_task, local->avg_load);
8106 capa_now /= SCHED_CAPACITY_SCALE;
8108 /* Amount of load we'd subtract */
8109 if (busiest->avg_load > scaled_busy_load_per_task) {
8110 capa_move += busiest->group_capacity *
8111 min(busiest->load_per_task,
8112 busiest->avg_load - scaled_busy_load_per_task);
8115 /* Amount of load we'd add */
8116 if (busiest->avg_load * busiest->group_capacity <
8117 busiest->load_per_task * SCHED_CAPACITY_SCALE) {
8118 tmp = (busiest->avg_load * busiest->group_capacity) /
8119 local->group_capacity;
8121 tmp = (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
8122 local->group_capacity;
8124 capa_move += local->group_capacity *
8125 min(local->load_per_task, local->avg_load + tmp);
8126 capa_move /= SCHED_CAPACITY_SCALE;
8128 /* Move if we gain throughput */
8129 if (capa_move > capa_now)
8130 env->imbalance = busiest->load_per_task;
8134 * calculate_imbalance - Calculate the amount of imbalance present within the
8135 * groups of a given sched_domain during load balance.
8136 * @env: load balance environment
8137 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
8139 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
8141 unsigned long max_pull, load_above_capacity = ~0UL;
8142 struct sg_lb_stats *local, *busiest;
8144 local = &sds->local_stat;
8145 busiest = &sds->busiest_stat;
8147 if (busiest->group_type == group_imbalanced) {
8149 * In the group_imb case we cannot rely on group-wide averages
8150 * to ensure CPU-load equilibrium, look at wider averages. XXX
8152 busiest->load_per_task =
8153 min(busiest->load_per_task, sds->avg_load);
8157 * Avg load of busiest sg can be less and avg load of local sg can
8158 * be greater than avg load across all sgs of sd because avg load
8159 * factors in sg capacity and sgs with smaller group_type are
8160 * skipped when updating the busiest sg:
8162 if (busiest->avg_load <= sds->avg_load ||
8163 local->avg_load >= sds->avg_load) {
8165 return fix_small_imbalance(env, sds);
8169 * If there aren't any idle CPUs, avoid creating some.
8171 if (busiest->group_type == group_overloaded &&
8172 local->group_type == group_overloaded) {
8173 load_above_capacity = busiest->sum_nr_running * SCHED_CAPACITY_SCALE;
8174 if (load_above_capacity > busiest->group_capacity) {
8175 load_above_capacity -= busiest->group_capacity;
8176 load_above_capacity *= scale_load_down(NICE_0_LOAD);
8177 load_above_capacity /= busiest->group_capacity;
8179 load_above_capacity = ~0UL;
8183 * We're trying to get all the CPUs to the average_load, so we don't
8184 * want to push ourselves above the average load, nor do we wish to
8185 * reduce the max loaded CPU below the average load. At the same time,
8186 * we also don't want to reduce the group load below the group
8187 * capacity. Thus we look for the minimum possible imbalance.
8189 max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity);
8191 /* How much load to actually move to equalise the imbalance */
8192 env->imbalance = min(
8193 max_pull * busiest->group_capacity,
8194 (sds->avg_load - local->avg_load) * local->group_capacity
8195 ) / SCHED_CAPACITY_SCALE;
8198 * if *imbalance is less than the average load per runnable task
8199 * there is no guarantee that any tasks will be moved so we'll have
8200 * a think about bumping its value to force at least one task to be
8203 if (env->imbalance < busiest->load_per_task)
8204 return fix_small_imbalance(env, sds);
8207 /******* find_busiest_group() helpers end here *********************/
8210 * find_busiest_group - Returns the busiest group within the sched_domain
8211 * if there is an imbalance.
8213 * Also calculates the amount of weighted load which should be moved
8214 * to restore balance.
8216 * @env: The load balancing environment.
8218 * Return: - The busiest group if imbalance exists.
8220 static struct sched_group *find_busiest_group(struct lb_env *env)
8222 struct sg_lb_stats *local, *busiest;
8223 struct sd_lb_stats sds;
8225 init_sd_lb_stats(&sds);
8228 * Compute the various statistics relavent for load balancing at
8231 update_sd_lb_stats(env, &sds);
8232 local = &sds.local_stat;
8233 busiest = &sds.busiest_stat;
8235 /* ASYM feature bypasses nice load balance check */
8236 if (check_asym_packing(env, &sds))
8239 /* There is no busy sibling group to pull tasks from */
8240 if (!sds.busiest || busiest->sum_nr_running == 0)
8243 /* XXX broken for overlapping NUMA groups */
8244 sds.avg_load = (SCHED_CAPACITY_SCALE * sds.total_load)
8245 / sds.total_capacity;
8248 * If the busiest group is imbalanced the below checks don't
8249 * work because they assume all things are equal, which typically
8250 * isn't true due to cpus_allowed constraints and the like.
8252 if (busiest->group_type == group_imbalanced)
8256 * When dst_cpu is idle, prevent SMP nice and/or asymmetric group
8257 * capacities from resulting in underutilization due to avg_load.
8259 if (env->idle != CPU_NOT_IDLE && group_has_capacity(env, local) &&
8260 busiest->group_no_capacity)
8264 * If the local group is busier than the selected busiest group
8265 * don't try and pull any tasks.
8267 if (local->avg_load >= busiest->avg_load)
8271 * Don't pull any tasks if this group is already above the domain
8274 if (local->avg_load >= sds.avg_load)
8277 if (env->idle == CPU_IDLE) {
8279 * This CPU is idle. If the busiest group is not overloaded
8280 * and there is no imbalance between this and busiest group
8281 * wrt idle CPUs, it is balanced. The imbalance becomes
8282 * significant if the diff is greater than 1 otherwise we
8283 * might end up to just move the imbalance on another group
8285 if ((busiest->group_type != group_overloaded) &&
8286 (local->idle_cpus <= (busiest->idle_cpus + 1)))
8290 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
8291 * imbalance_pct to be conservative.
8293 if (100 * busiest->avg_load <=
8294 env->sd->imbalance_pct * local->avg_load)
8299 /* Looks like there is an imbalance. Compute it */
8300 calculate_imbalance(env, &sds);
8301 return env->imbalance ? sds.busiest : NULL;
8309 * find_busiest_queue - find the busiest runqueue among the CPUs in the group.
8311 static struct rq *find_busiest_queue(struct lb_env *env,
8312 struct sched_group *group)
8314 struct rq *busiest = NULL, *rq;
8315 unsigned long busiest_load = 0, busiest_capacity = 1;
8318 for_each_cpu_and(i, sched_group_span(group), env->cpus) {
8319 unsigned long capacity, wl;
8323 rt = fbq_classify_rq(rq);
8326 * We classify groups/runqueues into three groups:
8327 * - regular: there are !numa tasks
8328 * - remote: there are numa tasks that run on the 'wrong' node
8329 * - all: there is no distinction
8331 * In order to avoid migrating ideally placed numa tasks,
8332 * ignore those when there's better options.
8334 * If we ignore the actual busiest queue to migrate another
8335 * task, the next balance pass can still reduce the busiest
8336 * queue by moving tasks around inside the node.
8338 * If we cannot move enough load due to this classification
8339 * the next pass will adjust the group classification and
8340 * allow migration of more tasks.
8342 * Both cases only affect the total convergence complexity.
8344 if (rt > env->fbq_type)
8347 capacity = capacity_of(i);
8349 wl = weighted_cpuload(rq);
8352 * When comparing with imbalance, use weighted_cpuload()
8353 * which is not scaled with the CPU capacity.
8356 if (rq->nr_running == 1 && wl > env->imbalance &&
8357 !check_cpu_capacity(rq, env->sd))
8361 * For the load comparisons with the other CPU's, consider
8362 * the weighted_cpuload() scaled with the CPU capacity, so
8363 * that the load can be moved away from the CPU that is
8364 * potentially running at a lower capacity.
8366 * Thus we're looking for max(wl_i / capacity_i), crosswise
8367 * multiplication to rid ourselves of the division works out
8368 * to: wl_i * capacity_j > wl_j * capacity_i; where j is
8369 * our previous maximum.
8371 if (wl * busiest_capacity > busiest_load * capacity) {
8373 busiest_capacity = capacity;
8382 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
8383 * so long as it is large enough.
8385 #define MAX_PINNED_INTERVAL 512
8387 static int need_active_balance(struct lb_env *env)
8389 struct sched_domain *sd = env->sd;
8391 if (env->idle == CPU_NEWLY_IDLE) {
8394 * ASYM_PACKING needs to force migrate tasks from busy but
8395 * lower priority CPUs in order to pack all tasks in the
8396 * highest priority CPUs.
8398 if ((sd->flags & SD_ASYM_PACKING) &&
8399 sched_asym_prefer(env->dst_cpu, env->src_cpu))
8404 * The dst_cpu is idle and the src_cpu CPU has only 1 CFS task.
8405 * It's worth migrating the task if the src_cpu's capacity is reduced
8406 * because of other sched_class or IRQs if more capacity stays
8407 * available on dst_cpu.
8409 if ((env->idle != CPU_NOT_IDLE) &&
8410 (env->src_rq->cfs.h_nr_running == 1)) {
8411 if ((check_cpu_capacity(env->src_rq, sd)) &&
8412 (capacity_of(env->src_cpu)*sd->imbalance_pct < capacity_of(env->dst_cpu)*100))
8416 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
8419 static int active_load_balance_cpu_stop(void *data);
8421 static int should_we_balance(struct lb_env *env)
8423 struct sched_group *sg = env->sd->groups;
8424 int cpu, balance_cpu = -1;
8427 * Ensure the balancing environment is consistent; can happen
8428 * when the softirq triggers 'during' hotplug.
8430 if (!cpumask_test_cpu(env->dst_cpu, env->cpus))
8434 * In the newly idle case, we will allow all the CPUs
8435 * to do the newly idle load balance.
8437 if (env->idle == CPU_NEWLY_IDLE)
8440 /* Try to find first idle CPU */
8441 for_each_cpu_and(cpu, group_balance_mask(sg), env->cpus) {
8449 if (balance_cpu == -1)
8450 balance_cpu = group_balance_cpu(sg);
8453 * First idle CPU or the first CPU(busiest) in this sched group
8454 * is eligible for doing load balancing at this and above domains.
8456 return balance_cpu == env->dst_cpu;
8460 * Check this_cpu to ensure it is balanced within domain. Attempt to move
8461 * tasks if there is an imbalance.
8463 static int load_balance(int this_cpu, struct rq *this_rq,
8464 struct sched_domain *sd, enum cpu_idle_type idle,
8465 int *continue_balancing)
8467 int ld_moved, cur_ld_moved, active_balance = 0;
8468 struct sched_domain *sd_parent = sd->parent;
8469 struct sched_group *group;
8472 struct cpumask *cpus = this_cpu_cpumask_var_ptr(load_balance_mask);
8474 struct lb_env env = {
8476 .dst_cpu = this_cpu,
8478 .dst_grpmask = sched_group_span(sd->groups),
8480 .loop_break = sched_nr_migrate_break,
8483 .tasks = LIST_HEAD_INIT(env.tasks),
8486 cpumask_and(cpus, sched_domain_span(sd), cpu_active_mask);
8488 schedstat_inc(sd->lb_count[idle]);
8491 if (!should_we_balance(&env)) {
8492 *continue_balancing = 0;
8496 group = find_busiest_group(&env);
8498 schedstat_inc(sd->lb_nobusyg[idle]);
8502 busiest = find_busiest_queue(&env, group);
8504 schedstat_inc(sd->lb_nobusyq[idle]);
8508 BUG_ON(busiest == env.dst_rq);
8510 schedstat_add(sd->lb_imbalance[idle], env.imbalance);
8512 env.src_cpu = busiest->cpu;
8513 env.src_rq = busiest;
8516 if (busiest->nr_running > 1) {
8518 * Attempt to move tasks. If find_busiest_group has found
8519 * an imbalance but busiest->nr_running <= 1, the group is
8520 * still unbalanced. ld_moved simply stays zero, so it is
8521 * correctly treated as an imbalance.
8523 env.flags |= LBF_ALL_PINNED;
8524 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
8527 rq_lock_irqsave(busiest, &rf);
8528 update_rq_clock(busiest);
8531 * cur_ld_moved - load moved in current iteration
8532 * ld_moved - cumulative load moved across iterations
8534 cur_ld_moved = detach_tasks(&env);
8537 * We've detached some tasks from busiest_rq. Every
8538 * task is masked "TASK_ON_RQ_MIGRATING", so we can safely
8539 * unlock busiest->lock, and we are able to be sure
8540 * that nobody can manipulate the tasks in parallel.
8541 * See task_rq_lock() family for the details.
8544 rq_unlock(busiest, &rf);
8548 ld_moved += cur_ld_moved;
8551 local_irq_restore(rf.flags);
8553 if (env.flags & LBF_NEED_BREAK) {
8554 env.flags &= ~LBF_NEED_BREAK;
8559 * Revisit (affine) tasks on src_cpu that couldn't be moved to
8560 * us and move them to an alternate dst_cpu in our sched_group
8561 * where they can run. The upper limit on how many times we
8562 * iterate on same src_cpu is dependent on number of CPUs in our
8565 * This changes load balance semantics a bit on who can move
8566 * load to a given_cpu. In addition to the given_cpu itself
8567 * (or a ilb_cpu acting on its behalf where given_cpu is
8568 * nohz-idle), we now have balance_cpu in a position to move
8569 * load to given_cpu. In rare situations, this may cause
8570 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
8571 * _independently_ and at _same_ time to move some load to
8572 * given_cpu) causing exceess load to be moved to given_cpu.
8573 * This however should not happen so much in practice and
8574 * moreover subsequent load balance cycles should correct the
8575 * excess load moved.
8577 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
8579 /* Prevent to re-select dst_cpu via env's CPUs */
8580 cpumask_clear_cpu(env.dst_cpu, env.cpus);
8582 env.dst_rq = cpu_rq(env.new_dst_cpu);
8583 env.dst_cpu = env.new_dst_cpu;
8584 env.flags &= ~LBF_DST_PINNED;
8586 env.loop_break = sched_nr_migrate_break;
8589 * Go back to "more_balance" rather than "redo" since we
8590 * need to continue with same src_cpu.
8596 * We failed to reach balance because of affinity.
8599 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
8601 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0)
8602 *group_imbalance = 1;
8605 /* All tasks on this runqueue were pinned by CPU affinity */
8606 if (unlikely(env.flags & LBF_ALL_PINNED)) {
8607 cpumask_clear_cpu(cpu_of(busiest), cpus);
8609 * Attempting to continue load balancing at the current
8610 * sched_domain level only makes sense if there are
8611 * active CPUs remaining as possible busiest CPUs to
8612 * pull load from which are not contained within the
8613 * destination group that is receiving any migrated
8616 if (!cpumask_subset(cpus, env.dst_grpmask)) {
8618 env.loop_break = sched_nr_migrate_break;
8621 goto out_all_pinned;
8626 schedstat_inc(sd->lb_failed[idle]);
8628 * Increment the failure counter only on periodic balance.
8629 * We do not want newidle balance, which can be very
8630 * frequent, pollute the failure counter causing
8631 * excessive cache_hot migrations and active balances.
8633 if (idle != CPU_NEWLY_IDLE)
8634 sd->nr_balance_failed++;
8636 if (need_active_balance(&env)) {
8637 unsigned long flags;
8639 raw_spin_lock_irqsave(&busiest->lock, flags);
8642 * Don't kick the active_load_balance_cpu_stop,
8643 * if the curr task on busiest CPU can't be
8644 * moved to this_cpu:
8646 if (!cpumask_test_cpu(this_cpu, &busiest->curr->cpus_allowed)) {
8647 raw_spin_unlock_irqrestore(&busiest->lock,
8649 env.flags |= LBF_ALL_PINNED;
8650 goto out_one_pinned;
8654 * ->active_balance synchronizes accesses to
8655 * ->active_balance_work. Once set, it's cleared
8656 * only after active load balance is finished.
8658 if (!busiest->active_balance) {
8659 busiest->active_balance = 1;
8660 busiest->push_cpu = this_cpu;
8663 raw_spin_unlock_irqrestore(&busiest->lock, flags);
8665 if (active_balance) {
8666 stop_one_cpu_nowait(cpu_of(busiest),
8667 active_load_balance_cpu_stop, busiest,
8668 &busiest->active_balance_work);
8671 /* We've kicked active balancing, force task migration. */
8672 sd->nr_balance_failed = sd->cache_nice_tries+1;
8675 sd->nr_balance_failed = 0;
8677 if (likely(!active_balance)) {
8678 /* We were unbalanced, so reset the balancing interval */
8679 sd->balance_interval = sd->min_interval;
8682 * If we've begun active balancing, start to back off. This
8683 * case may not be covered by the all_pinned logic if there
8684 * is only 1 task on the busy runqueue (because we don't call
8687 if (sd->balance_interval < sd->max_interval)
8688 sd->balance_interval *= 2;
8695 * We reach balance although we may have faced some affinity
8696 * constraints. Clear the imbalance flag if it was set.
8699 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
8701 if (*group_imbalance)
8702 *group_imbalance = 0;
8707 * We reach balance because all tasks are pinned at this level so
8708 * we can't migrate them. Let the imbalance flag set so parent level
8709 * can try to migrate them.
8711 schedstat_inc(sd->lb_balanced[idle]);
8713 sd->nr_balance_failed = 0;
8716 /* tune up the balancing interval */
8717 if (((env.flags & LBF_ALL_PINNED) &&
8718 sd->balance_interval < MAX_PINNED_INTERVAL) ||
8719 (sd->balance_interval < sd->max_interval))
8720 sd->balance_interval *= 2;
8727 static inline unsigned long
8728 get_sd_balance_interval(struct sched_domain *sd, int cpu_busy)
8730 unsigned long interval = sd->balance_interval;
8733 interval *= sd->busy_factor;
8735 /* scale ms to jiffies */
8736 interval = msecs_to_jiffies(interval);
8737 interval = clamp(interval, 1UL, max_load_balance_interval);
8743 update_next_balance(struct sched_domain *sd, unsigned long *next_balance)
8745 unsigned long interval, next;
8747 /* used by idle balance, so cpu_busy = 0 */
8748 interval = get_sd_balance_interval(sd, 0);
8749 next = sd->last_balance + interval;
8751 if (time_after(*next_balance, next))
8752 *next_balance = next;
8756 * active_load_balance_cpu_stop is run by the CPU stopper. It pushes
8757 * running tasks off the busiest CPU onto idle CPUs. It requires at
8758 * least 1 task to be running on each physical CPU where possible, and
8759 * avoids physical / logical imbalances.
8761 static int active_load_balance_cpu_stop(void *data)
8763 struct rq *busiest_rq = data;
8764 int busiest_cpu = cpu_of(busiest_rq);
8765 int target_cpu = busiest_rq->push_cpu;
8766 struct rq *target_rq = cpu_rq(target_cpu);
8767 struct sched_domain *sd;
8768 struct task_struct *p = NULL;
8771 rq_lock_irq(busiest_rq, &rf);
8773 * Between queueing the stop-work and running it is a hole in which
8774 * CPUs can become inactive. We should not move tasks from or to
8777 if (!cpu_active(busiest_cpu) || !cpu_active(target_cpu))
8780 /* Make sure the requested CPU hasn't gone down in the meantime: */
8781 if (unlikely(busiest_cpu != smp_processor_id() ||
8782 !busiest_rq->active_balance))
8785 /* Is there any task to move? */
8786 if (busiest_rq->nr_running <= 1)
8790 * This condition is "impossible", if it occurs
8791 * we need to fix it. Originally reported by
8792 * Bjorn Helgaas on a 128-CPU setup.
8794 BUG_ON(busiest_rq == target_rq);
8796 /* Search for an sd spanning us and the target CPU. */
8798 for_each_domain(target_cpu, sd) {
8799 if ((sd->flags & SD_LOAD_BALANCE) &&
8800 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
8805 struct lb_env env = {
8807 .dst_cpu = target_cpu,
8808 .dst_rq = target_rq,
8809 .src_cpu = busiest_rq->cpu,
8810 .src_rq = busiest_rq,
8813 * can_migrate_task() doesn't need to compute new_dst_cpu
8814 * for active balancing. Since we have CPU_IDLE, but no
8815 * @dst_grpmask we need to make that test go away with lying
8818 .flags = LBF_DST_PINNED,
8821 schedstat_inc(sd->alb_count);
8822 update_rq_clock(busiest_rq);
8824 p = detach_one_task(&env);
8826 schedstat_inc(sd->alb_pushed);
8827 /* Active balancing done, reset the failure counter. */
8828 sd->nr_balance_failed = 0;
8830 schedstat_inc(sd->alb_failed);
8835 busiest_rq->active_balance = 0;
8836 rq_unlock(busiest_rq, &rf);
8839 attach_one_task(target_rq, p);
8846 static DEFINE_SPINLOCK(balancing);
8849 * Scale the max load_balance interval with the number of CPUs in the system.
8850 * This trades load-balance latency on larger machines for less cross talk.
8852 void update_max_interval(void)
8854 max_load_balance_interval = HZ*num_online_cpus()/10;
8858 * It checks each scheduling domain to see if it is due to be balanced,
8859 * and initiates a balancing operation if so.
8861 * Balancing parameters are set up in init_sched_domains.
8863 static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle)
8865 int continue_balancing = 1;
8867 unsigned long interval;
8868 struct sched_domain *sd;
8869 /* Earliest time when we have to do rebalance again */
8870 unsigned long next_balance = jiffies + 60*HZ;
8871 int update_next_balance = 0;
8872 int need_serialize, need_decay = 0;
8876 for_each_domain(cpu, sd) {
8878 * Decay the newidle max times here because this is a regular
8879 * visit to all the domains. Decay ~1% per second.
8881 if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
8882 sd->max_newidle_lb_cost =
8883 (sd->max_newidle_lb_cost * 253) / 256;
8884 sd->next_decay_max_lb_cost = jiffies + HZ;
8887 max_cost += sd->max_newidle_lb_cost;
8889 if (!(sd->flags & SD_LOAD_BALANCE))
8893 * Stop the load balance at this level. There is another
8894 * CPU in our sched group which is doing load balancing more
8897 if (!continue_balancing) {
8903 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
8905 need_serialize = sd->flags & SD_SERIALIZE;
8906 if (need_serialize) {
8907 if (!spin_trylock(&balancing))
8911 if (time_after_eq(jiffies, sd->last_balance + interval)) {
8912 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
8914 * The LBF_DST_PINNED logic could have changed
8915 * env->dst_cpu, so we can't know our idle
8916 * state even if we migrated tasks. Update it.
8918 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
8920 sd->last_balance = jiffies;
8921 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
8924 spin_unlock(&balancing);
8926 if (time_after(next_balance, sd->last_balance + interval)) {
8927 next_balance = sd->last_balance + interval;
8928 update_next_balance = 1;
8933 * Ensure the rq-wide value also decays but keep it at a
8934 * reasonable floor to avoid funnies with rq->avg_idle.
8936 rq->max_idle_balance_cost =
8937 max((u64)sysctl_sched_migration_cost, max_cost);
8942 * next_balance will be updated only when there is a need.
8943 * When the cpu is attached to null domain for ex, it will not be
8946 if (likely(update_next_balance)) {
8947 rq->next_balance = next_balance;
8949 #ifdef CONFIG_NO_HZ_COMMON
8951 * If this CPU has been elected to perform the nohz idle
8952 * balance. Other idle CPUs have already rebalanced with
8953 * nohz_idle_balance() and nohz.next_balance has been
8954 * updated accordingly. This CPU is now running the idle load
8955 * balance for itself and we need to update the
8956 * nohz.next_balance accordingly.
8958 if ((idle == CPU_IDLE) && time_after(nohz.next_balance, rq->next_balance))
8959 nohz.next_balance = rq->next_balance;
8964 static inline int on_null_domain(struct rq *rq)
8966 return unlikely(!rcu_dereference_sched(rq->sd));
8969 #ifdef CONFIG_NO_HZ_COMMON
8971 * idle load balancing details
8972 * - When one of the busy CPUs notice that there may be an idle rebalancing
8973 * needed, they will kick the idle load balancer, which then does idle
8974 * load balancing for all the idle CPUs.
8977 static inline int find_new_ilb(void)
8979 int ilb = cpumask_first(nohz.idle_cpus_mask);
8981 if (ilb < nr_cpu_ids && idle_cpu(ilb))
8988 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
8989 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
8990 * CPU (if there is one).
8992 static void kick_ilb(unsigned int flags)
8996 nohz.next_balance++;
8998 ilb_cpu = find_new_ilb();
9000 if (ilb_cpu >= nr_cpu_ids)
9003 flags = atomic_fetch_or(flags, nohz_flags(ilb_cpu));
9004 if (flags & NOHZ_KICK_MASK)
9008 * Use smp_send_reschedule() instead of resched_cpu().
9009 * This way we generate a sched IPI on the target CPU which
9010 * is idle. And the softirq performing nohz idle load balance
9011 * will be run before returning from the IPI.
9013 smp_send_reschedule(ilb_cpu);
9017 * Current heuristic for kicking the idle load balancer in the presence
9018 * of an idle cpu in the system.
9019 * - This rq has more than one task.
9020 * - This rq has at least one CFS task and the capacity of the CPU is
9021 * significantly reduced because of RT tasks or IRQs.
9022 * - At parent of LLC scheduler domain level, this cpu's scheduler group has
9023 * multiple busy cpu.
9024 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
9025 * domain span are idle.
9027 static void nohz_balancer_kick(struct rq *rq)
9029 unsigned long now = jiffies;
9030 struct sched_domain_shared *sds;
9031 struct sched_domain *sd;
9032 int nr_busy, i, cpu = rq->cpu;
9033 unsigned int flags = 0;
9035 if (unlikely(rq->idle_balance))
9039 * We may be recently in ticked or tickless idle mode. At the first
9040 * busy tick after returning from idle, we will update the busy stats.
9042 nohz_balance_exit_idle(rq);
9045 * None are in tickless mode and hence no need for NOHZ idle load
9048 if (likely(!atomic_read(&nohz.nr_cpus)))
9051 if (READ_ONCE(nohz.has_blocked) &&
9052 time_after(now, READ_ONCE(nohz.next_blocked)))
9053 flags = NOHZ_STATS_KICK;
9055 if (time_before(now, nohz.next_balance))
9058 if (rq->nr_running >= 2) {
9059 flags = NOHZ_KICK_MASK;
9064 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
9067 * XXX: write a coherent comment on why we do this.
9068 * See also: http://lkml.kernel.org/r/20111202010832.602203411@sbsiddha-desk.sc.intel.com
9070 nr_busy = atomic_read(&sds->nr_busy_cpus);
9072 flags = NOHZ_KICK_MASK;
9078 sd = rcu_dereference(rq->sd);
9080 if ((rq->cfs.h_nr_running >= 1) &&
9081 check_cpu_capacity(rq, sd)) {
9082 flags = NOHZ_KICK_MASK;
9087 sd = rcu_dereference(per_cpu(sd_asym, cpu));
9089 for_each_cpu(i, sched_domain_span(sd)) {
9091 !cpumask_test_cpu(i, nohz.idle_cpus_mask))
9094 if (sched_asym_prefer(i, cpu)) {
9095 flags = NOHZ_KICK_MASK;
9107 static void set_cpu_sd_state_busy(int cpu)
9109 struct sched_domain *sd;
9112 sd = rcu_dereference(per_cpu(sd_llc, cpu));
9114 if (!sd || !sd->nohz_idle)
9118 atomic_inc(&sd->shared->nr_busy_cpus);
9123 void nohz_balance_exit_idle(struct rq *rq)
9125 SCHED_WARN_ON(rq != this_rq());
9127 if (likely(!rq->nohz_tick_stopped))
9130 rq->nohz_tick_stopped = 0;
9131 cpumask_clear_cpu(rq->cpu, nohz.idle_cpus_mask);
9132 atomic_dec(&nohz.nr_cpus);
9134 set_cpu_sd_state_busy(rq->cpu);
9137 static void set_cpu_sd_state_idle(int cpu)
9139 struct sched_domain *sd;
9142 sd = rcu_dereference(per_cpu(sd_llc, cpu));
9144 if (!sd || sd->nohz_idle)
9148 atomic_dec(&sd->shared->nr_busy_cpus);
9154 * This routine will record that the CPU is going idle with tick stopped.
9155 * This info will be used in performing idle load balancing in the future.
9157 void nohz_balance_enter_idle(int cpu)
9159 struct rq *rq = cpu_rq(cpu);
9161 SCHED_WARN_ON(cpu != smp_processor_id());
9163 /* If this CPU is going down, then nothing needs to be done: */
9164 if (!cpu_active(cpu))
9167 /* Spare idle load balancing on CPUs that don't want to be disturbed: */
9168 if (!housekeeping_cpu(cpu, HK_FLAG_SCHED))
9172 * Can be set safely without rq->lock held
9173 * If a clear happens, it will have evaluated last additions because
9174 * rq->lock is held during the check and the clear
9176 rq->has_blocked_load = 1;
9179 * The tick is still stopped but load could have been added in the
9180 * meantime. We set the nohz.has_blocked flag to trig a check of the
9181 * *_avg. The CPU is already part of nohz.idle_cpus_mask so the clear
9182 * of nohz.has_blocked can only happen after checking the new load
9184 if (rq->nohz_tick_stopped)
9187 /* If we're a completely isolated CPU, we don't play: */
9188 if (on_null_domain(rq))
9191 rq->nohz_tick_stopped = 1;
9193 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
9194 atomic_inc(&nohz.nr_cpus);
9197 * Ensures that if nohz_idle_balance() fails to observe our
9198 * @idle_cpus_mask store, it must observe the @has_blocked
9201 smp_mb__after_atomic();
9203 set_cpu_sd_state_idle(cpu);
9207 * Each time a cpu enter idle, we assume that it has blocked load and
9208 * enable the periodic update of the load of idle cpus
9210 WRITE_ONCE(nohz.has_blocked, 1);
9214 * Internal function that runs load balance for all idle cpus. The load balance
9215 * can be a simple update of blocked load or a complete load balance with
9216 * tasks movement depending of flags.
9217 * The function returns false if the loop has stopped before running
9218 * through all idle CPUs.
9220 static bool _nohz_idle_balance(struct rq *this_rq, unsigned int flags,
9221 enum cpu_idle_type idle)
9223 /* Earliest time when we have to do rebalance again */
9224 unsigned long now = jiffies;
9225 unsigned long next_balance = now + 60*HZ;
9226 bool has_blocked_load = false;
9227 int update_next_balance = 0;
9228 int this_cpu = this_rq->cpu;
9233 SCHED_WARN_ON((flags & NOHZ_KICK_MASK) == NOHZ_BALANCE_KICK);
9236 * We assume there will be no idle load after this update and clear
9237 * the has_blocked flag. If a cpu enters idle in the mean time, it will
9238 * set the has_blocked flag and trig another update of idle load.
9239 * Because a cpu that becomes idle, is added to idle_cpus_mask before
9240 * setting the flag, we are sure to not clear the state and not
9241 * check the load of an idle cpu.
9243 WRITE_ONCE(nohz.has_blocked, 0);
9246 * Ensures that if we miss the CPU, we must see the has_blocked
9247 * store from nohz_balance_enter_idle().
9251 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
9252 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
9256 * If this CPU gets work to do, stop the load balancing
9257 * work being done for other CPUs. Next load
9258 * balancing owner will pick it up.
9260 if (need_resched()) {
9261 has_blocked_load = true;
9265 rq = cpu_rq(balance_cpu);
9267 has_blocked_load |= update_nohz_stats(rq, true);
9270 * If time for next balance is due,
9273 if (time_after_eq(jiffies, rq->next_balance)) {
9276 rq_lock_irqsave(rq, &rf);
9277 update_rq_clock(rq);
9278 cpu_load_update_idle(rq);
9279 rq_unlock_irqrestore(rq, &rf);
9281 if (flags & NOHZ_BALANCE_KICK)
9282 rebalance_domains(rq, CPU_IDLE);
9285 if (time_after(next_balance, rq->next_balance)) {
9286 next_balance = rq->next_balance;
9287 update_next_balance = 1;
9291 /* Newly idle CPU doesn't need an update */
9292 if (idle != CPU_NEWLY_IDLE) {
9293 update_blocked_averages(this_cpu);
9294 has_blocked_load |= this_rq->has_blocked_load;
9297 if (flags & NOHZ_BALANCE_KICK)
9298 rebalance_domains(this_rq, CPU_IDLE);
9300 WRITE_ONCE(nohz.next_blocked,
9301 now + msecs_to_jiffies(LOAD_AVG_PERIOD));
9303 /* The full idle balance loop has been done */
9307 /* There is still blocked load, enable periodic update */
9308 if (has_blocked_load)
9309 WRITE_ONCE(nohz.has_blocked, 1);
9312 * next_balance will be updated only when there is a need.
9313 * When the CPU is attached to null domain for ex, it will not be
9316 if (likely(update_next_balance))
9317 nohz.next_balance = next_balance;
9323 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
9324 * rebalancing for all the cpus for whom scheduler ticks are stopped.
9326 static bool nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
9328 int this_cpu = this_rq->cpu;
9331 if (!(atomic_read(nohz_flags(this_cpu)) & NOHZ_KICK_MASK))
9334 if (idle != CPU_IDLE) {
9335 atomic_andnot(NOHZ_KICK_MASK, nohz_flags(this_cpu));
9340 * barrier, pairs with nohz_balance_enter_idle(), ensures ...
9342 flags = atomic_fetch_andnot(NOHZ_KICK_MASK, nohz_flags(this_cpu));
9343 if (!(flags & NOHZ_KICK_MASK))
9346 _nohz_idle_balance(this_rq, flags, idle);
9351 static void nohz_newidle_balance(struct rq *this_rq)
9353 int this_cpu = this_rq->cpu;
9356 * This CPU doesn't want to be disturbed by scheduler
9359 if (!housekeeping_cpu(this_cpu, HK_FLAG_SCHED))
9362 /* Will wake up very soon. No time for doing anything else*/
9363 if (this_rq->avg_idle < sysctl_sched_migration_cost)
9366 /* Don't need to update blocked load of idle CPUs*/
9367 if (!READ_ONCE(nohz.has_blocked) ||
9368 time_before(jiffies, READ_ONCE(nohz.next_blocked)))
9371 raw_spin_unlock(&this_rq->lock);
9373 * This CPU is going to be idle and blocked load of idle CPUs
9374 * need to be updated. Run the ilb locally as it is a good
9375 * candidate for ilb instead of waking up another idle CPU.
9376 * Kick an normal ilb if we failed to do the update.
9378 if (!_nohz_idle_balance(this_rq, NOHZ_STATS_KICK, CPU_NEWLY_IDLE))
9379 kick_ilb(NOHZ_STATS_KICK);
9380 raw_spin_lock(&this_rq->lock);
9383 #else /* !CONFIG_NO_HZ_COMMON */
9384 static inline void nohz_balancer_kick(struct rq *rq) { }
9386 static inline bool nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
9391 static inline void nohz_newidle_balance(struct rq *this_rq) { }
9392 #endif /* CONFIG_NO_HZ_COMMON */
9395 * idle_balance is called by schedule() if this_cpu is about to become
9396 * idle. Attempts to pull tasks from other CPUs.
9398 static int idle_balance(struct rq *this_rq, struct rq_flags *rf)
9400 unsigned long next_balance = jiffies + HZ;
9401 int this_cpu = this_rq->cpu;
9402 struct sched_domain *sd;
9403 int pulled_task = 0;
9407 * We must set idle_stamp _before_ calling idle_balance(), such that we
9408 * measure the duration of idle_balance() as idle time.
9410 this_rq->idle_stamp = rq_clock(this_rq);
9413 * Do not pull tasks towards !active CPUs...
9415 if (!cpu_active(this_cpu))
9419 * This is OK, because current is on_cpu, which avoids it being picked
9420 * for load-balance and preemption/IRQs are still disabled avoiding
9421 * further scheduler activity on it and we're being very careful to
9422 * re-start the picking loop.
9424 rq_unpin_lock(this_rq, rf);
9426 if (this_rq->avg_idle < sysctl_sched_migration_cost ||
9427 !this_rq->rd->overload) {
9430 sd = rcu_dereference_check_sched_domain(this_rq->sd);
9432 update_next_balance(sd, &next_balance);
9435 nohz_newidle_balance(this_rq);
9440 raw_spin_unlock(&this_rq->lock);
9442 update_blocked_averages(this_cpu);
9444 for_each_domain(this_cpu, sd) {
9445 int continue_balancing = 1;
9446 u64 t0, domain_cost;
9448 if (!(sd->flags & SD_LOAD_BALANCE))
9451 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost) {
9452 update_next_balance(sd, &next_balance);
9456 if (sd->flags & SD_BALANCE_NEWIDLE) {
9457 t0 = sched_clock_cpu(this_cpu);
9459 pulled_task = load_balance(this_cpu, this_rq,
9461 &continue_balancing);
9463 domain_cost = sched_clock_cpu(this_cpu) - t0;
9464 if (domain_cost > sd->max_newidle_lb_cost)
9465 sd->max_newidle_lb_cost = domain_cost;
9467 curr_cost += domain_cost;
9470 update_next_balance(sd, &next_balance);
9473 * Stop searching for tasks to pull if there are
9474 * now runnable tasks on this rq.
9476 if (pulled_task || this_rq->nr_running > 0)
9481 raw_spin_lock(&this_rq->lock);
9483 if (curr_cost > this_rq->max_idle_balance_cost)
9484 this_rq->max_idle_balance_cost = curr_cost;
9488 * While browsing the domains, we released the rq lock, a task could
9489 * have been enqueued in the meantime. Since we're not going idle,
9490 * pretend we pulled a task.
9492 if (this_rq->cfs.h_nr_running && !pulled_task)
9495 /* Move the next balance forward */
9496 if (time_after(this_rq->next_balance, next_balance))
9497 this_rq->next_balance = next_balance;
9499 /* Is there a task of a high priority class? */
9500 if (this_rq->nr_running != this_rq->cfs.h_nr_running)
9504 this_rq->idle_stamp = 0;
9506 rq_repin_lock(this_rq, rf);
9512 * run_rebalance_domains is triggered when needed from the scheduler tick.
9513 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
9515 static __latent_entropy void run_rebalance_domains(struct softirq_action *h)
9517 struct rq *this_rq = this_rq();
9518 enum cpu_idle_type idle = this_rq->idle_balance ?
9519 CPU_IDLE : CPU_NOT_IDLE;
9522 * If this CPU has a pending nohz_balance_kick, then do the
9523 * balancing on behalf of the other idle CPUs whose ticks are
9524 * stopped. Do nohz_idle_balance *before* rebalance_domains to
9525 * give the idle CPUs a chance to load balance. Else we may
9526 * load balance only within the local sched_domain hierarchy
9527 * and abort nohz_idle_balance altogether if we pull some load.
9529 if (nohz_idle_balance(this_rq, idle))
9532 /* normal load balance */
9533 update_blocked_averages(this_rq->cpu);
9534 rebalance_domains(this_rq, idle);
9538 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
9540 void trigger_load_balance(struct rq *rq)
9542 /* Don't need to rebalance while attached to NULL domain */
9543 if (unlikely(on_null_domain(rq)))
9546 if (time_after_eq(jiffies, rq->next_balance))
9547 raise_softirq(SCHED_SOFTIRQ);
9549 nohz_balancer_kick(rq);
9552 static void rq_online_fair(struct rq *rq)
9556 update_runtime_enabled(rq);
9559 static void rq_offline_fair(struct rq *rq)
9563 /* Ensure any throttled groups are reachable by pick_next_task */
9564 unthrottle_offline_cfs_rqs(rq);
9567 #endif /* CONFIG_SMP */
9570 * scheduler tick hitting a task of our scheduling class.
9572 * NOTE: This function can be called remotely by the tick offload that
9573 * goes along full dynticks. Therefore no local assumption can be made
9574 * and everything must be accessed through the @rq and @curr passed in
9577 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
9579 struct cfs_rq *cfs_rq;
9580 struct sched_entity *se = &curr->se;
9582 for_each_sched_entity(se) {
9583 cfs_rq = cfs_rq_of(se);
9584 entity_tick(cfs_rq, se, queued);
9587 if (static_branch_unlikely(&sched_numa_balancing))
9588 task_tick_numa(rq, curr);
9592 * called on fork with the child task as argument from the parent's context
9593 * - child not yet on the tasklist
9594 * - preemption disabled
9596 static void task_fork_fair(struct task_struct *p)
9598 struct cfs_rq *cfs_rq;
9599 struct sched_entity *se = &p->se, *curr;
9600 struct rq *rq = this_rq();
9604 update_rq_clock(rq);
9606 cfs_rq = task_cfs_rq(current);
9607 curr = cfs_rq->curr;
9609 update_curr(cfs_rq);
9610 se->vruntime = curr->vruntime;
9612 place_entity(cfs_rq, se, 1);
9614 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
9616 * Upon rescheduling, sched_class::put_prev_task() will place
9617 * 'current' within the tree based on its new key value.
9619 swap(curr->vruntime, se->vruntime);
9623 se->vruntime -= cfs_rq->min_vruntime;
9628 * Priority of the task has changed. Check to see if we preempt
9632 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
9634 if (!task_on_rq_queued(p))
9638 * Reschedule if we are currently running on this runqueue and
9639 * our priority decreased, or if we are not currently running on
9640 * this runqueue and our priority is higher than the current's
9642 if (rq->curr == p) {
9643 if (p->prio > oldprio)
9646 check_preempt_curr(rq, p, 0);
9649 static inline bool vruntime_normalized(struct task_struct *p)
9651 struct sched_entity *se = &p->se;
9654 * In both the TASK_ON_RQ_QUEUED and TASK_ON_RQ_MIGRATING cases,
9655 * the dequeue_entity(.flags=0) will already have normalized the
9662 * When !on_rq, vruntime of the task has usually NOT been normalized.
9663 * But there are some cases where it has already been normalized:
9665 * - A forked child which is waiting for being woken up by
9666 * wake_up_new_task().
9667 * - A task which has been woken up by try_to_wake_up() and
9668 * waiting for actually being woken up by sched_ttwu_pending().
9670 if (!se->sum_exec_runtime ||
9671 (p->state == TASK_WAKING && p->sched_remote_wakeup))
9677 #ifdef CONFIG_FAIR_GROUP_SCHED
9679 * Propagate the changes of the sched_entity across the tg tree to make it
9680 * visible to the root
9682 static void propagate_entity_cfs_rq(struct sched_entity *se)
9684 struct cfs_rq *cfs_rq;
9686 /* Start to propagate at parent */
9689 for_each_sched_entity(se) {
9690 cfs_rq = cfs_rq_of(se);
9692 if (cfs_rq_throttled(cfs_rq))
9695 update_load_avg(cfs_rq, se, UPDATE_TG);
9699 static void propagate_entity_cfs_rq(struct sched_entity *se) { }
9702 static void detach_entity_cfs_rq(struct sched_entity *se)
9704 struct cfs_rq *cfs_rq = cfs_rq_of(se);
9706 /* Catch up with the cfs_rq and remove our load when we leave */
9707 update_load_avg(cfs_rq, se, 0);
9708 detach_entity_load_avg(cfs_rq, se);
9709 update_tg_load_avg(cfs_rq, false);
9710 propagate_entity_cfs_rq(se);
9713 static void attach_entity_cfs_rq(struct sched_entity *se)
9715 struct cfs_rq *cfs_rq = cfs_rq_of(se);
9717 #ifdef CONFIG_FAIR_GROUP_SCHED
9719 * Since the real-depth could have been changed (only FAIR
9720 * class maintain depth value), reset depth properly.
9722 se->depth = se->parent ? se->parent->depth + 1 : 0;
9725 /* Synchronize entity with its cfs_rq */
9726 update_load_avg(cfs_rq, se, sched_feat(ATTACH_AGE_LOAD) ? 0 : SKIP_AGE_LOAD);
9727 attach_entity_load_avg(cfs_rq, se, 0);
9728 update_tg_load_avg(cfs_rq, false);
9729 propagate_entity_cfs_rq(se);
9732 static void detach_task_cfs_rq(struct task_struct *p)
9734 struct sched_entity *se = &p->se;
9735 struct cfs_rq *cfs_rq = cfs_rq_of(se);
9737 if (!vruntime_normalized(p)) {
9739 * Fix up our vruntime so that the current sleep doesn't
9740 * cause 'unlimited' sleep bonus.
9742 place_entity(cfs_rq, se, 0);
9743 se->vruntime -= cfs_rq->min_vruntime;
9746 detach_entity_cfs_rq(se);
9749 static void attach_task_cfs_rq(struct task_struct *p)
9751 struct sched_entity *se = &p->se;
9752 struct cfs_rq *cfs_rq = cfs_rq_of(se);
9754 attach_entity_cfs_rq(se);
9756 if (!vruntime_normalized(p))
9757 se->vruntime += cfs_rq->min_vruntime;
9760 static void switched_from_fair(struct rq *rq, struct task_struct *p)
9762 detach_task_cfs_rq(p);
9765 static void switched_to_fair(struct rq *rq, struct task_struct *p)
9767 attach_task_cfs_rq(p);
9769 if (task_on_rq_queued(p)) {
9771 * We were most likely switched from sched_rt, so
9772 * kick off the schedule if running, otherwise just see
9773 * if we can still preempt the current task.
9778 check_preempt_curr(rq, p, 0);
9782 /* Account for a task changing its policy or group.
9784 * This routine is mostly called to set cfs_rq->curr field when a task
9785 * migrates between groups/classes.
9787 static void set_curr_task_fair(struct rq *rq)
9789 struct sched_entity *se = &rq->curr->se;
9791 for_each_sched_entity(se) {
9792 struct cfs_rq *cfs_rq = cfs_rq_of(se);
9794 set_next_entity(cfs_rq, se);
9795 /* ensure bandwidth has been allocated on our new cfs_rq */
9796 account_cfs_rq_runtime(cfs_rq, 0);
9800 void init_cfs_rq(struct cfs_rq *cfs_rq)
9802 cfs_rq->tasks_timeline = RB_ROOT_CACHED;
9803 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
9804 #ifndef CONFIG_64BIT
9805 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
9808 raw_spin_lock_init(&cfs_rq->removed.lock);
9812 #ifdef CONFIG_FAIR_GROUP_SCHED
9813 static void task_set_group_fair(struct task_struct *p)
9815 struct sched_entity *se = &p->se;
9817 set_task_rq(p, task_cpu(p));
9818 se->depth = se->parent ? se->parent->depth + 1 : 0;
9821 static void task_move_group_fair(struct task_struct *p)
9823 detach_task_cfs_rq(p);
9824 set_task_rq(p, task_cpu(p));
9827 /* Tell se's cfs_rq has been changed -- migrated */
9828 p->se.avg.last_update_time = 0;
9830 attach_task_cfs_rq(p);
9833 static void task_change_group_fair(struct task_struct *p, int type)
9836 case TASK_SET_GROUP:
9837 task_set_group_fair(p);
9840 case TASK_MOVE_GROUP:
9841 task_move_group_fair(p);
9846 void free_fair_sched_group(struct task_group *tg)
9850 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
9852 for_each_possible_cpu(i) {
9854 kfree(tg->cfs_rq[i]);
9863 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9865 struct sched_entity *se;
9866 struct cfs_rq *cfs_rq;
9869 tg->cfs_rq = kcalloc(nr_cpu_ids, sizeof(cfs_rq), GFP_KERNEL);
9872 tg->se = kcalloc(nr_cpu_ids, sizeof(se), GFP_KERNEL);
9876 tg->shares = NICE_0_LOAD;
9878 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
9880 for_each_possible_cpu(i) {
9881 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
9882 GFP_KERNEL, cpu_to_node(i));
9886 se = kzalloc_node(sizeof(struct sched_entity),
9887 GFP_KERNEL, cpu_to_node(i));
9891 init_cfs_rq(cfs_rq);
9892 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
9893 init_entity_runnable_average(se);
9904 void online_fair_sched_group(struct task_group *tg)
9906 struct sched_entity *se;
9910 for_each_possible_cpu(i) {
9914 raw_spin_lock_irq(&rq->lock);
9915 update_rq_clock(rq);
9916 attach_entity_cfs_rq(se);
9917 sync_throttle(tg, i);
9918 raw_spin_unlock_irq(&rq->lock);
9922 void unregister_fair_sched_group(struct task_group *tg)
9924 unsigned long flags;
9928 for_each_possible_cpu(cpu) {
9930 remove_entity_load_avg(tg->se[cpu]);
9933 * Only empty task groups can be destroyed; so we can speculatively
9934 * check on_list without danger of it being re-added.
9936 if (!tg->cfs_rq[cpu]->on_list)
9941 raw_spin_lock_irqsave(&rq->lock, flags);
9942 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
9943 raw_spin_unlock_irqrestore(&rq->lock, flags);
9947 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
9948 struct sched_entity *se, int cpu,
9949 struct sched_entity *parent)
9951 struct rq *rq = cpu_rq(cpu);
9955 init_cfs_rq_runtime(cfs_rq);
9957 tg->cfs_rq[cpu] = cfs_rq;
9960 /* se could be NULL for root_task_group */
9965 se->cfs_rq = &rq->cfs;
9968 se->cfs_rq = parent->my_q;
9969 se->depth = parent->depth + 1;
9973 /* guarantee group entities always have weight */
9974 update_load_set(&se->load, NICE_0_LOAD);
9975 se->parent = parent;
9978 static DEFINE_MUTEX(shares_mutex);
9980 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
9985 * We can't change the weight of the root cgroup.
9990 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
9992 mutex_lock(&shares_mutex);
9993 if (tg->shares == shares)
9996 tg->shares = shares;
9997 for_each_possible_cpu(i) {
9998 struct rq *rq = cpu_rq(i);
9999 struct sched_entity *se = tg->se[i];
10000 struct rq_flags rf;
10002 /* Propagate contribution to hierarchy */
10003 rq_lock_irqsave(rq, &rf);
10004 update_rq_clock(rq);
10005 for_each_sched_entity(se) {
10006 update_load_avg(cfs_rq_of(se), se, UPDATE_TG);
10007 update_cfs_group(se);
10009 rq_unlock_irqrestore(rq, &rf);
10013 mutex_unlock(&shares_mutex);
10016 #else /* CONFIG_FAIR_GROUP_SCHED */
10018 void free_fair_sched_group(struct task_group *tg) { }
10020 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
10025 void online_fair_sched_group(struct task_group *tg) { }
10027 void unregister_fair_sched_group(struct task_group *tg) { }
10029 #endif /* CONFIG_FAIR_GROUP_SCHED */
10032 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
10034 struct sched_entity *se = &task->se;
10035 unsigned int rr_interval = 0;
10038 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
10041 if (rq->cfs.load.weight)
10042 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
10044 return rr_interval;
10048 * All the scheduling class methods:
10050 const struct sched_class fair_sched_class = {
10051 .next = &idle_sched_class,
10052 .enqueue_task = enqueue_task_fair,
10053 .dequeue_task = dequeue_task_fair,
10054 .yield_task = yield_task_fair,
10055 .yield_to_task = yield_to_task_fair,
10057 .check_preempt_curr = check_preempt_wakeup,
10059 .pick_next_task = pick_next_task_fair,
10060 .put_prev_task = put_prev_task_fair,
10063 .select_task_rq = select_task_rq_fair,
10064 .migrate_task_rq = migrate_task_rq_fair,
10066 .rq_online = rq_online_fair,
10067 .rq_offline = rq_offline_fair,
10069 .task_dead = task_dead_fair,
10070 .set_cpus_allowed = set_cpus_allowed_common,
10073 .set_curr_task = set_curr_task_fair,
10074 .task_tick = task_tick_fair,
10075 .task_fork = task_fork_fair,
10077 .prio_changed = prio_changed_fair,
10078 .switched_from = switched_from_fair,
10079 .switched_to = switched_to_fair,
10081 .get_rr_interval = get_rr_interval_fair,
10083 .update_curr = update_curr_fair,
10085 #ifdef CONFIG_FAIR_GROUP_SCHED
10086 .task_change_group = task_change_group_fair,
10090 #ifdef CONFIG_SCHED_DEBUG
10091 void print_cfs_stats(struct seq_file *m, int cpu)
10093 struct cfs_rq *cfs_rq, *pos;
10096 for_each_leaf_cfs_rq_safe(cpu_rq(cpu), cfs_rq, pos)
10097 print_cfs_rq(m, cpu, cfs_rq);
10101 #ifdef CONFIG_NUMA_BALANCING
10102 void show_numa_stats(struct task_struct *p, struct seq_file *m)
10105 unsigned long tsf = 0, tpf = 0, gsf = 0, gpf = 0;
10107 for_each_online_node(node) {
10108 if (p->numa_faults) {
10109 tsf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 0)];
10110 tpf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 1)];
10112 if (p->numa_group) {
10113 gsf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 0)],
10114 gpf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 1)];
10116 print_numa_stats(m, node, tsf, tpf, gsf, gpf);
10119 #endif /* CONFIG_NUMA_BALANCING */
10120 #endif /* CONFIG_SCHED_DEBUG */
10122 __init void init_sched_fair_class(void)
10125 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
10127 #ifdef CONFIG_NO_HZ_COMMON
10128 nohz.next_balance = jiffies;
10129 nohz.next_blocked = jiffies;
10130 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);