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 /* An entity is a task if it doesn't "own" a runqueue */
259 #define entity_is_task(se) (!se->my_q)
261 static inline struct task_struct *task_of(struct sched_entity *se)
263 SCHED_WARN_ON(!entity_is_task(se));
264 return container_of(se, struct task_struct, se);
267 /* Walk up scheduling entities hierarchy */
268 #define for_each_sched_entity(se) \
269 for (; se; se = se->parent)
271 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
276 /* runqueue on which this entity is (to be) queued */
277 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
282 /* runqueue "owned" by this group */
283 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
288 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
290 if (!cfs_rq->on_list) {
291 struct rq *rq = rq_of(cfs_rq);
292 int cpu = cpu_of(rq);
294 * Ensure we either appear before our parent (if already
295 * enqueued) or force our parent to appear after us when it is
296 * enqueued. The fact that we always enqueue bottom-up
297 * reduces this to two cases and a special case for the root
298 * cfs_rq. Furthermore, it also means that we will always reset
299 * tmp_alone_branch either when the branch is connected
300 * to a tree or when we reach the beg of the tree
302 if (cfs_rq->tg->parent &&
303 cfs_rq->tg->parent->cfs_rq[cpu]->on_list) {
305 * If parent is already on the list, we add the child
306 * just before. Thanks to circular linked property of
307 * the list, this means to put the child at the tail
308 * of the list that starts by parent.
310 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
311 &(cfs_rq->tg->parent->cfs_rq[cpu]->leaf_cfs_rq_list));
313 * The branch is now connected to its tree so we can
314 * reset tmp_alone_branch to the beginning of the
317 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
318 } else if (!cfs_rq->tg->parent) {
320 * cfs rq without parent should be put
321 * at the tail of the list.
323 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
324 &rq->leaf_cfs_rq_list);
326 * We have reach the beg of a tree so we can reset
327 * tmp_alone_branch to the beginning of the list.
329 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
332 * The parent has not already been added so we want to
333 * make sure that it will be put after us.
334 * tmp_alone_branch points to the beg of the branch
335 * where we will add parent.
337 list_add_rcu(&cfs_rq->leaf_cfs_rq_list,
338 rq->tmp_alone_branch);
340 * update tmp_alone_branch to points to the new beg
343 rq->tmp_alone_branch = &cfs_rq->leaf_cfs_rq_list;
350 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
352 if (cfs_rq->on_list) {
353 list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
358 /* Iterate thr' all leaf cfs_rq's on a runqueue */
359 #define for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) \
360 list_for_each_entry_safe(cfs_rq, pos, &rq->leaf_cfs_rq_list, \
363 /* Do the two (enqueued) entities belong to the same group ? */
364 static inline struct cfs_rq *
365 is_same_group(struct sched_entity *se, struct sched_entity *pse)
367 if (se->cfs_rq == pse->cfs_rq)
373 static inline struct sched_entity *parent_entity(struct sched_entity *se)
379 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
381 int se_depth, pse_depth;
384 * preemption test can be made between sibling entities who are in the
385 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
386 * both tasks until we find their ancestors who are siblings of common
390 /* First walk up until both entities are at same depth */
391 se_depth = (*se)->depth;
392 pse_depth = (*pse)->depth;
394 while (se_depth > pse_depth) {
396 *se = parent_entity(*se);
399 while (pse_depth > se_depth) {
401 *pse = parent_entity(*pse);
404 while (!is_same_group(*se, *pse)) {
405 *se = parent_entity(*se);
406 *pse = parent_entity(*pse);
410 #else /* !CONFIG_FAIR_GROUP_SCHED */
412 static inline struct task_struct *task_of(struct sched_entity *se)
414 return container_of(se, struct task_struct, se);
417 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
419 return container_of(cfs_rq, struct rq, cfs);
422 #define entity_is_task(se) 1
424 #define for_each_sched_entity(se) \
425 for (; se; se = NULL)
427 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
429 return &task_rq(p)->cfs;
432 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
434 struct task_struct *p = task_of(se);
435 struct rq *rq = task_rq(p);
440 /* runqueue "owned" by this group */
441 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
446 static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
450 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
454 #define for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) \
455 for (cfs_rq = &rq->cfs, pos = NULL; cfs_rq; cfs_rq = pos)
457 static inline struct sched_entity *parent_entity(struct sched_entity *se)
463 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
467 #endif /* CONFIG_FAIR_GROUP_SCHED */
469 static __always_inline
470 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec);
472 /**************************************************************
473 * Scheduling class tree data structure manipulation methods:
476 static inline u64 max_vruntime(u64 max_vruntime, u64 vruntime)
478 s64 delta = (s64)(vruntime - max_vruntime);
480 max_vruntime = vruntime;
485 static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime)
487 s64 delta = (s64)(vruntime - min_vruntime);
489 min_vruntime = vruntime;
494 static inline int entity_before(struct sched_entity *a,
495 struct sched_entity *b)
497 return (s64)(a->vruntime - b->vruntime) < 0;
500 static void update_min_vruntime(struct cfs_rq *cfs_rq)
502 struct sched_entity *curr = cfs_rq->curr;
503 struct rb_node *leftmost = rb_first_cached(&cfs_rq->tasks_timeline);
505 u64 vruntime = cfs_rq->min_vruntime;
509 vruntime = curr->vruntime;
514 if (leftmost) { /* non-empty tree */
515 struct sched_entity *se;
516 se = rb_entry(leftmost, struct sched_entity, run_node);
519 vruntime = se->vruntime;
521 vruntime = min_vruntime(vruntime, se->vruntime);
524 /* ensure we never gain time by being placed backwards. */
525 cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime);
528 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
533 * Enqueue an entity into the rb-tree:
535 static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
537 struct rb_node **link = &cfs_rq->tasks_timeline.rb_root.rb_node;
538 struct rb_node *parent = NULL;
539 struct sched_entity *entry;
540 bool leftmost = true;
543 * Find the right place in the rbtree:
547 entry = rb_entry(parent, struct sched_entity, run_node);
549 * We dont care about collisions. Nodes with
550 * the same key stay together.
552 if (entity_before(se, entry)) {
553 link = &parent->rb_left;
555 link = &parent->rb_right;
560 rb_link_node(&se->run_node, parent, link);
561 rb_insert_color_cached(&se->run_node,
562 &cfs_rq->tasks_timeline, leftmost);
565 static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
567 rb_erase_cached(&se->run_node, &cfs_rq->tasks_timeline);
570 struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq)
572 struct rb_node *left = rb_first_cached(&cfs_rq->tasks_timeline);
577 return rb_entry(left, struct sched_entity, run_node);
580 static struct sched_entity *__pick_next_entity(struct sched_entity *se)
582 struct rb_node *next = rb_next(&se->run_node);
587 return rb_entry(next, struct sched_entity, run_node);
590 #ifdef CONFIG_SCHED_DEBUG
591 struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
593 struct rb_node *last = rb_last(&cfs_rq->tasks_timeline.rb_root);
598 return rb_entry(last, struct sched_entity, run_node);
601 /**************************************************************
602 * Scheduling class statistics methods:
605 int sched_proc_update_handler(struct ctl_table *table, int write,
606 void __user *buffer, size_t *lenp,
609 int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
610 unsigned int factor = get_update_sysctl_factor();
615 sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency,
616 sysctl_sched_min_granularity);
618 #define WRT_SYSCTL(name) \
619 (normalized_sysctl_##name = sysctl_##name / (factor))
620 WRT_SYSCTL(sched_min_granularity);
621 WRT_SYSCTL(sched_latency);
622 WRT_SYSCTL(sched_wakeup_granularity);
632 static inline u64 calc_delta_fair(u64 delta, struct sched_entity *se)
634 if (unlikely(se->load.weight != NICE_0_LOAD))
635 delta = __calc_delta(delta, NICE_0_LOAD, &se->load);
641 * The idea is to set a period in which each task runs once.
643 * When there are too many tasks (sched_nr_latency) we have to stretch
644 * this period because otherwise the slices get too small.
646 * p = (nr <= nl) ? l : l*nr/nl
648 static u64 __sched_period(unsigned long nr_running)
650 if (unlikely(nr_running > sched_nr_latency))
651 return nr_running * sysctl_sched_min_granularity;
653 return sysctl_sched_latency;
657 * We calculate the wall-time slice from the period by taking a part
658 * proportional to the weight.
662 static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
664 u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq);
666 for_each_sched_entity(se) {
667 struct load_weight *load;
668 struct load_weight lw;
670 cfs_rq = cfs_rq_of(se);
671 load = &cfs_rq->load;
673 if (unlikely(!se->on_rq)) {
676 update_load_add(&lw, se->load.weight);
679 slice = __calc_delta(slice, se->load.weight, load);
685 * We calculate the vruntime slice of a to-be-inserted task.
689 static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se)
691 return calc_delta_fair(sched_slice(cfs_rq, se), se);
696 #include "sched-pelt.h"
698 static int select_idle_sibling(struct task_struct *p, int prev_cpu, int cpu);
699 static unsigned long task_h_load(struct task_struct *p);
701 /* Give new sched_entity start runnable values to heavy its load in infant time */
702 void init_entity_runnable_average(struct sched_entity *se)
704 struct sched_avg *sa = &se->avg;
706 memset(sa, 0, sizeof(*sa));
709 * Tasks are intialized with full load to be seen as heavy tasks until
710 * they get a chance to stabilize to their real load level.
711 * Group entities are intialized with zero load to reflect the fact that
712 * nothing has been attached to the task group yet.
714 if (entity_is_task(se))
715 sa->runnable_load_avg = sa->load_avg = scale_load_down(se->load.weight);
717 se->runnable_weight = se->load.weight;
719 /* when this task enqueue'ed, it will contribute to its cfs_rq's load_avg */
722 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq);
723 static void attach_entity_cfs_rq(struct sched_entity *se);
726 * With new tasks being created, their initial util_avgs are extrapolated
727 * based on the cfs_rq's current util_avg:
729 * util_avg = cfs_rq->util_avg / (cfs_rq->load_avg + 1) * se.load.weight
731 * However, in many cases, the above util_avg does not give a desired
732 * value. Moreover, the sum of the util_avgs may be divergent, such
733 * as when the series is a harmonic series.
735 * To solve this problem, we also cap the util_avg of successive tasks to
736 * only 1/2 of the left utilization budget:
738 * util_avg_cap = (1024 - cfs_rq->avg.util_avg) / 2^n
740 * where n denotes the nth task.
742 * For example, a simplest series from the beginning would be like:
744 * task util_avg: 512, 256, 128, 64, 32, 16, 8, ...
745 * cfs_rq util_avg: 512, 768, 896, 960, 992, 1008, 1016, ...
747 * Finally, that extrapolated util_avg is clamped to the cap (util_avg_cap)
748 * if util_avg > util_avg_cap.
750 void post_init_entity_util_avg(struct sched_entity *se)
752 struct cfs_rq *cfs_rq = cfs_rq_of(se);
753 struct sched_avg *sa = &se->avg;
754 long cap = (long)(SCHED_CAPACITY_SCALE - cfs_rq->avg.util_avg) / 2;
757 if (cfs_rq->avg.util_avg != 0) {
758 sa->util_avg = cfs_rq->avg.util_avg * se->load.weight;
759 sa->util_avg /= (cfs_rq->avg.load_avg + 1);
761 if (sa->util_avg > cap)
768 if (entity_is_task(se)) {
769 struct task_struct *p = task_of(se);
770 if (p->sched_class != &fair_sched_class) {
772 * For !fair tasks do:
774 update_cfs_rq_load_avg(now, cfs_rq);
775 attach_entity_load_avg(cfs_rq, se, 0);
776 switched_from_fair(rq, p);
778 * such that the next switched_to_fair() has the
781 se->avg.last_update_time = cfs_rq_clock_task(cfs_rq);
786 attach_entity_cfs_rq(se);
789 #else /* !CONFIG_SMP */
790 void init_entity_runnable_average(struct sched_entity *se)
793 void post_init_entity_util_avg(struct sched_entity *se)
796 static void update_tg_load_avg(struct cfs_rq *cfs_rq, int force)
799 #endif /* CONFIG_SMP */
802 * Update the current task's runtime statistics.
804 static void update_curr(struct cfs_rq *cfs_rq)
806 struct sched_entity *curr = cfs_rq->curr;
807 u64 now = rq_clock_task(rq_of(cfs_rq));
813 delta_exec = now - curr->exec_start;
814 if (unlikely((s64)delta_exec <= 0))
817 curr->exec_start = now;
819 schedstat_set(curr->statistics.exec_max,
820 max(delta_exec, curr->statistics.exec_max));
822 curr->sum_exec_runtime += delta_exec;
823 schedstat_add(cfs_rq->exec_clock, delta_exec);
825 curr->vruntime += calc_delta_fair(delta_exec, curr);
826 update_min_vruntime(cfs_rq);
828 if (entity_is_task(curr)) {
829 struct task_struct *curtask = task_of(curr);
831 trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime);
832 cgroup_account_cputime(curtask, delta_exec);
833 account_group_exec_runtime(curtask, delta_exec);
836 account_cfs_rq_runtime(cfs_rq, delta_exec);
839 static void update_curr_fair(struct rq *rq)
841 update_curr(cfs_rq_of(&rq->curr->se));
845 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
847 u64 wait_start, prev_wait_start;
849 if (!schedstat_enabled())
852 wait_start = rq_clock(rq_of(cfs_rq));
853 prev_wait_start = schedstat_val(se->statistics.wait_start);
855 if (entity_is_task(se) && task_on_rq_migrating(task_of(se)) &&
856 likely(wait_start > prev_wait_start))
857 wait_start -= prev_wait_start;
859 __schedstat_set(se->statistics.wait_start, wait_start);
863 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
865 struct task_struct *p;
868 if (!schedstat_enabled())
871 delta = rq_clock(rq_of(cfs_rq)) - schedstat_val(se->statistics.wait_start);
873 if (entity_is_task(se)) {
875 if (task_on_rq_migrating(p)) {
877 * Preserve migrating task's wait time so wait_start
878 * time stamp can be adjusted to accumulate wait time
879 * prior to migration.
881 __schedstat_set(se->statistics.wait_start, delta);
884 trace_sched_stat_wait(p, delta);
887 __schedstat_set(se->statistics.wait_max,
888 max(schedstat_val(se->statistics.wait_max), delta));
889 __schedstat_inc(se->statistics.wait_count);
890 __schedstat_add(se->statistics.wait_sum, delta);
891 __schedstat_set(se->statistics.wait_start, 0);
895 update_stats_enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
897 struct task_struct *tsk = NULL;
898 u64 sleep_start, block_start;
900 if (!schedstat_enabled())
903 sleep_start = schedstat_val(se->statistics.sleep_start);
904 block_start = schedstat_val(se->statistics.block_start);
906 if (entity_is_task(se))
910 u64 delta = rq_clock(rq_of(cfs_rq)) - sleep_start;
915 if (unlikely(delta > schedstat_val(se->statistics.sleep_max)))
916 __schedstat_set(se->statistics.sleep_max, delta);
918 __schedstat_set(se->statistics.sleep_start, 0);
919 __schedstat_add(se->statistics.sum_sleep_runtime, delta);
922 account_scheduler_latency(tsk, delta >> 10, 1);
923 trace_sched_stat_sleep(tsk, delta);
927 u64 delta = rq_clock(rq_of(cfs_rq)) - block_start;
932 if (unlikely(delta > schedstat_val(se->statistics.block_max)))
933 __schedstat_set(se->statistics.block_max, delta);
935 __schedstat_set(se->statistics.block_start, 0);
936 __schedstat_add(se->statistics.sum_sleep_runtime, delta);
939 if (tsk->in_iowait) {
940 __schedstat_add(se->statistics.iowait_sum, delta);
941 __schedstat_inc(se->statistics.iowait_count);
942 trace_sched_stat_iowait(tsk, delta);
945 trace_sched_stat_blocked(tsk, delta);
948 * Blocking time is in units of nanosecs, so shift by
949 * 20 to get a milliseconds-range estimation of the
950 * amount of time that the task spent sleeping:
952 if (unlikely(prof_on == SLEEP_PROFILING)) {
953 profile_hits(SLEEP_PROFILING,
954 (void *)get_wchan(tsk),
957 account_scheduler_latency(tsk, delta >> 10, 0);
963 * Task is being enqueued - update stats:
966 update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
968 if (!schedstat_enabled())
972 * Are we enqueueing a waiting task? (for current tasks
973 * a dequeue/enqueue event is a NOP)
975 if (se != cfs_rq->curr)
976 update_stats_wait_start(cfs_rq, se);
978 if (flags & ENQUEUE_WAKEUP)
979 update_stats_enqueue_sleeper(cfs_rq, se);
983 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
986 if (!schedstat_enabled())
990 * Mark the end of the wait period if dequeueing a
993 if (se != cfs_rq->curr)
994 update_stats_wait_end(cfs_rq, se);
996 if ((flags & DEQUEUE_SLEEP) && entity_is_task(se)) {
997 struct task_struct *tsk = task_of(se);
999 if (tsk->state & TASK_INTERRUPTIBLE)
1000 __schedstat_set(se->statistics.sleep_start,
1001 rq_clock(rq_of(cfs_rq)));
1002 if (tsk->state & TASK_UNINTERRUPTIBLE)
1003 __schedstat_set(se->statistics.block_start,
1004 rq_clock(rq_of(cfs_rq)));
1009 * We are picking a new current task - update its stats:
1012 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
1015 * We are starting a new run period:
1017 se->exec_start = rq_clock_task(rq_of(cfs_rq));
1020 /**************************************************
1021 * Scheduling class queueing methods:
1024 #ifdef CONFIG_NUMA_BALANCING
1026 * Approximate time to scan a full NUMA task in ms. The task scan period is
1027 * calculated based on the tasks virtual memory size and
1028 * numa_balancing_scan_size.
1030 unsigned int sysctl_numa_balancing_scan_period_min = 1000;
1031 unsigned int sysctl_numa_balancing_scan_period_max = 60000;
1033 /* Portion of address space to scan in MB */
1034 unsigned int sysctl_numa_balancing_scan_size = 256;
1036 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
1037 unsigned int sysctl_numa_balancing_scan_delay = 1000;
1042 spinlock_t lock; /* nr_tasks, tasks */
1047 struct rcu_head rcu;
1048 unsigned long total_faults;
1049 unsigned long max_faults_cpu;
1051 * Faults_cpu is used to decide whether memory should move
1052 * towards the CPU. As a consequence, these stats are weighted
1053 * more by CPU use than by memory faults.
1055 unsigned long *faults_cpu;
1056 unsigned long faults[0];
1059 static inline unsigned long group_faults_priv(struct numa_group *ng);
1060 static inline unsigned long group_faults_shared(struct numa_group *ng);
1062 static unsigned int task_nr_scan_windows(struct task_struct *p)
1064 unsigned long rss = 0;
1065 unsigned long nr_scan_pages;
1068 * Calculations based on RSS as non-present and empty pages are skipped
1069 * by the PTE scanner and NUMA hinting faults should be trapped based
1072 nr_scan_pages = sysctl_numa_balancing_scan_size << (20 - PAGE_SHIFT);
1073 rss = get_mm_rss(p->mm);
1075 rss = nr_scan_pages;
1077 rss = round_up(rss, nr_scan_pages);
1078 return rss / nr_scan_pages;
1081 /* For sanitys sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */
1082 #define MAX_SCAN_WINDOW 2560
1084 static unsigned int task_scan_min(struct task_struct *p)
1086 unsigned int scan_size = READ_ONCE(sysctl_numa_balancing_scan_size);
1087 unsigned int scan, floor;
1088 unsigned int windows = 1;
1090 if (scan_size < MAX_SCAN_WINDOW)
1091 windows = MAX_SCAN_WINDOW / scan_size;
1092 floor = 1000 / windows;
1094 scan = sysctl_numa_balancing_scan_period_min / task_nr_scan_windows(p);
1095 return max_t(unsigned int, floor, scan);
1098 static unsigned int task_scan_start(struct task_struct *p)
1100 unsigned long smin = task_scan_min(p);
1101 unsigned long period = smin;
1103 /* Scale the maximum scan period with the amount of shared memory. */
1104 if (p->numa_group) {
1105 struct numa_group *ng = p->numa_group;
1106 unsigned long shared = group_faults_shared(ng);
1107 unsigned long private = group_faults_priv(ng);
1109 period *= atomic_read(&ng->refcount);
1110 period *= shared + 1;
1111 period /= private + shared + 1;
1114 return max(smin, period);
1117 static unsigned int task_scan_max(struct task_struct *p)
1119 unsigned long smin = task_scan_min(p);
1122 /* Watch for min being lower than max due to floor calculations */
1123 smax = sysctl_numa_balancing_scan_period_max / task_nr_scan_windows(p);
1125 /* Scale the maximum scan period with the amount of shared memory. */
1126 if (p->numa_group) {
1127 struct numa_group *ng = p->numa_group;
1128 unsigned long shared = group_faults_shared(ng);
1129 unsigned long private = group_faults_priv(ng);
1130 unsigned long period = smax;
1132 period *= atomic_read(&ng->refcount);
1133 period *= shared + 1;
1134 period /= private + shared + 1;
1136 smax = max(smax, period);
1139 return max(smin, smax);
1142 void init_numa_balancing(unsigned long clone_flags, struct task_struct *p)
1145 struct mm_struct *mm = p->mm;
1148 mm_users = atomic_read(&mm->mm_users);
1149 if (mm_users == 1) {
1150 mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
1151 mm->numa_scan_seq = 0;
1155 p->numa_scan_seq = mm ? mm->numa_scan_seq : 0;
1156 p->numa_scan_period = sysctl_numa_balancing_scan_delay;
1157 p->numa_work.next = &p->numa_work;
1158 p->numa_faults = NULL;
1159 p->numa_group = NULL;
1160 p->last_task_numa_placement = 0;
1161 p->last_sum_exec_runtime = 0;
1163 /* New address space, reset the preferred nid */
1164 if (!(clone_flags & CLONE_VM)) {
1165 p->numa_preferred_nid = -1;
1170 * New thread, keep existing numa_preferred_nid which should be copied
1171 * already by arch_dup_task_struct but stagger when scans start.
1176 delay = min_t(unsigned int, task_scan_max(current),
1177 current->numa_scan_period * mm_users * NSEC_PER_MSEC);
1178 delay += 2 * TICK_NSEC;
1179 p->node_stamp = delay;
1183 static void account_numa_enqueue(struct rq *rq, struct task_struct *p)
1185 rq->nr_numa_running += (p->numa_preferred_nid != -1);
1186 rq->nr_preferred_running += (p->numa_preferred_nid == task_node(p));
1189 static void account_numa_dequeue(struct rq *rq, struct task_struct *p)
1191 rq->nr_numa_running -= (p->numa_preferred_nid != -1);
1192 rq->nr_preferred_running -= (p->numa_preferred_nid == task_node(p));
1195 /* Shared or private faults. */
1196 #define NR_NUMA_HINT_FAULT_TYPES 2
1198 /* Memory and CPU locality */
1199 #define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2)
1201 /* Averaged statistics, and temporary buffers. */
1202 #define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2)
1204 pid_t task_numa_group_id(struct task_struct *p)
1206 return p->numa_group ? p->numa_group->gid : 0;
1210 * The averaged statistics, shared & private, memory & CPU,
1211 * occupy the first half of the array. The second half of the
1212 * array is for current counters, which are averaged into the
1213 * first set by task_numa_placement.
1215 static inline int task_faults_idx(enum numa_faults_stats s, int nid, int priv)
1217 return NR_NUMA_HINT_FAULT_TYPES * (s * nr_node_ids + nid) + priv;
1220 static inline unsigned long task_faults(struct task_struct *p, int nid)
1222 if (!p->numa_faults)
1225 return p->numa_faults[task_faults_idx(NUMA_MEM, nid, 0)] +
1226 p->numa_faults[task_faults_idx(NUMA_MEM, nid, 1)];
1229 static inline unsigned long group_faults(struct task_struct *p, int nid)
1234 return p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 0)] +
1235 p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 1)];
1238 static inline unsigned long group_faults_cpu(struct numa_group *group, int nid)
1240 return group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 0)] +
1241 group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 1)];
1244 static inline unsigned long group_faults_priv(struct numa_group *ng)
1246 unsigned long faults = 0;
1249 for_each_online_node(node) {
1250 faults += ng->faults[task_faults_idx(NUMA_MEM, node, 1)];
1256 static inline unsigned long group_faults_shared(struct numa_group *ng)
1258 unsigned long faults = 0;
1261 for_each_online_node(node) {
1262 faults += ng->faults[task_faults_idx(NUMA_MEM, node, 0)];
1269 * A node triggering more than 1/3 as many NUMA faults as the maximum is
1270 * considered part of a numa group's pseudo-interleaving set. Migrations
1271 * between these nodes are slowed down, to allow things to settle down.
1273 #define ACTIVE_NODE_FRACTION 3
1275 static bool numa_is_active_node(int nid, struct numa_group *ng)
1277 return group_faults_cpu(ng, nid) * ACTIVE_NODE_FRACTION > ng->max_faults_cpu;
1280 /* Handle placement on systems where not all nodes are directly connected. */
1281 static unsigned long score_nearby_nodes(struct task_struct *p, int nid,
1282 int maxdist, bool task)
1284 unsigned long score = 0;
1288 * All nodes are directly connected, and the same distance
1289 * from each other. No need for fancy placement algorithms.
1291 if (sched_numa_topology_type == NUMA_DIRECT)
1295 * This code is called for each node, introducing N^2 complexity,
1296 * which should be ok given the number of nodes rarely exceeds 8.
1298 for_each_online_node(node) {
1299 unsigned long faults;
1300 int dist = node_distance(nid, node);
1303 * The furthest away nodes in the system are not interesting
1304 * for placement; nid was already counted.
1306 if (dist == sched_max_numa_distance || node == nid)
1310 * On systems with a backplane NUMA topology, compare groups
1311 * of nodes, and move tasks towards the group with the most
1312 * memory accesses. When comparing two nodes at distance
1313 * "hoplimit", only nodes closer by than "hoplimit" are part
1314 * of each group. Skip other nodes.
1316 if (sched_numa_topology_type == NUMA_BACKPLANE &&
1320 /* Add up the faults from nearby nodes. */
1322 faults = task_faults(p, node);
1324 faults = group_faults(p, node);
1327 * On systems with a glueless mesh NUMA topology, there are
1328 * no fixed "groups of nodes". Instead, nodes that are not
1329 * directly connected bounce traffic through intermediate
1330 * nodes; a numa_group can occupy any set of nodes.
1331 * The further away a node is, the less the faults count.
1332 * This seems to result in good task placement.
1334 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
1335 faults *= (sched_max_numa_distance - dist);
1336 faults /= (sched_max_numa_distance - LOCAL_DISTANCE);
1346 * These return the fraction of accesses done by a particular task, or
1347 * task group, on a particular numa node. The group weight is given a
1348 * larger multiplier, in order to group tasks together that are almost
1349 * evenly spread out between numa nodes.
1351 static inline unsigned long task_weight(struct task_struct *p, int nid,
1354 unsigned long faults, total_faults;
1356 if (!p->numa_faults)
1359 total_faults = p->total_numa_faults;
1364 faults = task_faults(p, nid);
1365 faults += score_nearby_nodes(p, nid, dist, true);
1367 return 1000 * faults / total_faults;
1370 static inline unsigned long group_weight(struct task_struct *p, int nid,
1373 unsigned long faults, total_faults;
1378 total_faults = p->numa_group->total_faults;
1383 faults = group_faults(p, nid);
1384 faults += score_nearby_nodes(p, nid, dist, false);
1386 return 1000 * faults / total_faults;
1389 bool should_numa_migrate_memory(struct task_struct *p, struct page * page,
1390 int src_nid, int dst_cpu)
1392 struct numa_group *ng = p->numa_group;
1393 int dst_nid = cpu_to_node(dst_cpu);
1394 int last_cpupid, this_cpupid;
1396 this_cpupid = cpu_pid_to_cpupid(dst_cpu, current->pid);
1399 * Multi-stage node selection is used in conjunction with a periodic
1400 * migration fault to build a temporal task<->page relation. By using
1401 * a two-stage filter we remove short/unlikely relations.
1403 * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate
1404 * a task's usage of a particular page (n_p) per total usage of this
1405 * page (n_t) (in a given time-span) to a probability.
1407 * Our periodic faults will sample this probability and getting the
1408 * same result twice in a row, given these samples are fully
1409 * independent, is then given by P(n)^2, provided our sample period
1410 * is sufficiently short compared to the usage pattern.
1412 * This quadric squishes small probabilities, making it less likely we
1413 * act on an unlikely task<->page relation.
1415 last_cpupid = page_cpupid_xchg_last(page, this_cpupid);
1416 if (!cpupid_pid_unset(last_cpupid) &&
1417 cpupid_to_nid(last_cpupid) != dst_nid)
1420 /* Always allow migrate on private faults */
1421 if (cpupid_match_pid(p, last_cpupid))
1424 /* A shared fault, but p->numa_group has not been set up yet. */
1429 * Destination node is much more heavily used than the source
1430 * node? Allow migration.
1432 if (group_faults_cpu(ng, dst_nid) > group_faults_cpu(ng, src_nid) *
1433 ACTIVE_NODE_FRACTION)
1437 * Distribute memory according to CPU & memory use on each node,
1438 * with 3/4 hysteresis to avoid unnecessary memory migrations:
1440 * faults_cpu(dst) 3 faults_cpu(src)
1441 * --------------- * - > ---------------
1442 * faults_mem(dst) 4 faults_mem(src)
1444 return group_faults_cpu(ng, dst_nid) * group_faults(p, src_nid) * 3 >
1445 group_faults_cpu(ng, src_nid) * group_faults(p, dst_nid) * 4;
1448 static unsigned long weighted_cpuload(struct rq *rq);
1449 static unsigned long source_load(int cpu, int type);
1450 static unsigned long target_load(int cpu, int type);
1451 static unsigned long capacity_of(int cpu);
1453 /* Cached statistics for all CPUs within a node */
1455 unsigned long nr_running;
1458 /* Total compute capacity of CPUs on a node */
1459 unsigned long compute_capacity;
1461 /* Approximate capacity in terms of runnable tasks on a node */
1462 unsigned long task_capacity;
1463 int has_free_capacity;
1467 * XXX borrowed from update_sg_lb_stats
1469 static void update_numa_stats(struct numa_stats *ns, int nid)
1471 int smt, cpu, cpus = 0;
1472 unsigned long capacity;
1474 memset(ns, 0, sizeof(*ns));
1475 for_each_cpu(cpu, cpumask_of_node(nid)) {
1476 struct rq *rq = cpu_rq(cpu);
1478 ns->nr_running += rq->nr_running;
1479 ns->load += weighted_cpuload(rq);
1480 ns->compute_capacity += capacity_of(cpu);
1486 * If we raced with hotplug and there are no CPUs left in our mask
1487 * the @ns structure is NULL'ed and task_numa_compare() will
1488 * not find this node attractive.
1490 * We'll either bail at !has_free_capacity, or we'll detect a huge
1491 * imbalance and bail there.
1496 /* smt := ceil(cpus / capacity), assumes: 1 < smt_power < 2 */
1497 smt = DIV_ROUND_UP(SCHED_CAPACITY_SCALE * cpus, ns->compute_capacity);
1498 capacity = cpus / smt; /* cores */
1500 ns->task_capacity = min_t(unsigned, capacity,
1501 DIV_ROUND_CLOSEST(ns->compute_capacity, SCHED_CAPACITY_SCALE));
1502 ns->has_free_capacity = (ns->nr_running < ns->task_capacity);
1505 struct task_numa_env {
1506 struct task_struct *p;
1508 int src_cpu, src_nid;
1509 int dst_cpu, dst_nid;
1511 struct numa_stats src_stats, dst_stats;
1516 struct task_struct *best_task;
1521 static void task_numa_assign(struct task_numa_env *env,
1522 struct task_struct *p, long imp)
1525 put_task_struct(env->best_task);
1530 env->best_imp = imp;
1531 env->best_cpu = env->dst_cpu;
1534 static bool load_too_imbalanced(long src_load, long dst_load,
1535 struct task_numa_env *env)
1538 long orig_src_load, orig_dst_load;
1539 long src_capacity, dst_capacity;
1542 * The load is corrected for the CPU capacity available on each node.
1545 * ------------ vs ---------
1546 * src_capacity dst_capacity
1548 src_capacity = env->src_stats.compute_capacity;
1549 dst_capacity = env->dst_stats.compute_capacity;
1551 /* We care about the slope of the imbalance, not the direction. */
1552 if (dst_load < src_load)
1553 swap(dst_load, src_load);
1555 /* Is the difference below the threshold? */
1556 imb = dst_load * src_capacity * 100 -
1557 src_load * dst_capacity * env->imbalance_pct;
1562 * The imbalance is above the allowed threshold.
1563 * Compare it with the old imbalance.
1565 orig_src_load = env->src_stats.load;
1566 orig_dst_load = env->dst_stats.load;
1568 if (orig_dst_load < orig_src_load)
1569 swap(orig_dst_load, orig_src_load);
1571 old_imb = orig_dst_load * src_capacity * 100 -
1572 orig_src_load * dst_capacity * env->imbalance_pct;
1574 /* Would this change make things worse? */
1575 return (imb > old_imb);
1579 * This checks if the overall compute and NUMA accesses of the system would
1580 * be improved if the source tasks was migrated to the target dst_cpu taking
1581 * into account that it might be best if task running on the dst_cpu should
1582 * be exchanged with the source task
1584 static void task_numa_compare(struct task_numa_env *env,
1585 long taskimp, long groupimp)
1587 struct rq *src_rq = cpu_rq(env->src_cpu);
1588 struct rq *dst_rq = cpu_rq(env->dst_cpu);
1589 struct task_struct *cur;
1590 long src_load, dst_load;
1592 long imp = env->p->numa_group ? groupimp : taskimp;
1594 int dist = env->dist;
1597 cur = task_rcu_dereference(&dst_rq->curr);
1598 if (cur && ((cur->flags & PF_EXITING) || is_idle_task(cur)))
1602 * Because we have preemption enabled we can get migrated around and
1603 * end try selecting ourselves (current == env->p) as a swap candidate.
1609 * "imp" is the fault differential for the source task between the
1610 * source and destination node. Calculate the total differential for
1611 * the source task and potential destination task. The more negative
1612 * the value is, the more rmeote accesses that would be expected to
1613 * 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
1636 * itself (not part of a group), use the task weight
1639 if (cur->numa_group)
1640 imp += group_weight(cur, env->src_nid, dist) -
1641 group_weight(cur, env->dst_nid, dist);
1643 imp += task_weight(cur, env->src_nid, dist) -
1644 task_weight(cur, env->dst_nid, dist);
1648 if (imp <= env->best_imp && moveimp <= env->best_imp)
1652 /* Is there capacity at our destination? */
1653 if (env->src_stats.nr_running <= env->src_stats.task_capacity &&
1654 !env->dst_stats.has_free_capacity)
1660 /* Balance doesn't matter much if we're running a task per CPU: */
1661 if (imp > env->best_imp && src_rq->nr_running == 1 &&
1662 dst_rq->nr_running == 1)
1666 * In the overloaded case, try and keep the load balanced.
1669 load = task_h_load(env->p);
1670 dst_load = env->dst_stats.load + load;
1671 src_load = env->src_stats.load - load;
1673 if (moveimp > imp && moveimp > env->best_imp) {
1675 * If the improvement from just moving env->p direction is
1676 * better than swapping tasks around, check if a move is
1677 * possible. Store a slightly smaller score than moveimp,
1678 * so an actually idle CPU will win.
1680 if (!load_too_imbalanced(src_load, dst_load, env)) {
1687 if (imp <= env->best_imp)
1691 load = task_h_load(cur);
1696 if (load_too_imbalanced(src_load, dst_load, env))
1700 * One idle CPU per node is evaluated for a task numa move.
1701 * Call select_idle_sibling to maybe find a better one.
1705 * select_idle_siblings() uses an per-CPU cpumask that
1706 * can be used from IRQ context.
1708 local_irq_disable();
1709 env->dst_cpu = select_idle_sibling(env->p, env->src_cpu,
1715 task_numa_assign(env, cur, imp);
1720 static void task_numa_find_cpu(struct task_numa_env *env,
1721 long taskimp, long groupimp)
1725 for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) {
1726 /* Skip this CPU if the source task cannot migrate */
1727 if (!cpumask_test_cpu(cpu, &env->p->cpus_allowed))
1731 task_numa_compare(env, taskimp, groupimp);
1735 /* Only move tasks to a NUMA node less busy than the current node. */
1736 static bool numa_has_capacity(struct task_numa_env *env)
1738 struct numa_stats *src = &env->src_stats;
1739 struct numa_stats *dst = &env->dst_stats;
1741 if (src->has_free_capacity && !dst->has_free_capacity)
1745 * Only consider a task move if the source has a higher load
1746 * than the destination, corrected for CPU capacity on each node.
1748 * src->load dst->load
1749 * --------------------- vs ---------------------
1750 * src->compute_capacity dst->compute_capacity
1752 if (src->load * dst->compute_capacity * env->imbalance_pct >
1754 dst->load * src->compute_capacity * 100)
1760 static int task_numa_migrate(struct task_struct *p)
1762 struct task_numa_env env = {
1765 .src_cpu = task_cpu(p),
1766 .src_nid = task_node(p),
1768 .imbalance_pct = 112,
1774 struct sched_domain *sd;
1775 unsigned long taskweight, groupweight;
1777 long taskimp, groupimp;
1780 * Pick the lowest SD_NUMA domain, as that would have the smallest
1781 * imbalance and would be the first to start moving tasks about.
1783 * And we want to avoid any moving of tasks about, as that would create
1784 * random movement of tasks -- counter the numa conditions we're trying
1788 sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu));
1790 env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2;
1794 * Cpusets can break the scheduler domain tree into smaller
1795 * balance domains, some of which do not cross NUMA boundaries.
1796 * Tasks that are "trapped" in such domains cannot be migrated
1797 * elsewhere, so there is no point in (re)trying.
1799 if (unlikely(!sd)) {
1800 p->numa_preferred_nid = task_node(p);
1804 env.dst_nid = p->numa_preferred_nid;
1805 dist = env.dist = node_distance(env.src_nid, env.dst_nid);
1806 taskweight = task_weight(p, env.src_nid, dist);
1807 groupweight = group_weight(p, env.src_nid, dist);
1808 update_numa_stats(&env.src_stats, env.src_nid);
1809 taskimp = task_weight(p, env.dst_nid, dist) - taskweight;
1810 groupimp = group_weight(p, env.dst_nid, dist) - groupweight;
1811 update_numa_stats(&env.dst_stats, env.dst_nid);
1813 /* Try to find a spot on the preferred nid. */
1814 if (numa_has_capacity(&env))
1815 task_numa_find_cpu(&env, taskimp, groupimp);
1818 * Look at other nodes in these cases:
1819 * - there is no space available on the preferred_nid
1820 * - the task is part of a numa_group that is interleaved across
1821 * multiple NUMA nodes; in order to better consolidate the group,
1822 * we need to check other locations.
1824 if (env.best_cpu == -1 || (p->numa_group && p->numa_group->active_nodes > 1)) {
1825 for_each_online_node(nid) {
1826 if (nid == env.src_nid || nid == p->numa_preferred_nid)
1829 dist = node_distance(env.src_nid, env.dst_nid);
1830 if (sched_numa_topology_type == NUMA_BACKPLANE &&
1832 taskweight = task_weight(p, env.src_nid, dist);
1833 groupweight = group_weight(p, env.src_nid, dist);
1836 /* Only consider nodes where both task and groups benefit */
1837 taskimp = task_weight(p, nid, dist) - taskweight;
1838 groupimp = group_weight(p, nid, dist) - groupweight;
1839 if (taskimp < 0 && groupimp < 0)
1844 update_numa_stats(&env.dst_stats, env.dst_nid);
1845 if (numa_has_capacity(&env))
1846 task_numa_find_cpu(&env, taskimp, groupimp);
1851 * If the task is part of a workload that spans multiple NUMA nodes,
1852 * and is migrating into one of the workload's active nodes, remember
1853 * this node as the task's preferred numa node, so the workload can
1855 * A task that migrated to a second choice node will be better off
1856 * trying for a better one later. Do not set the preferred node here.
1858 if (p->numa_group) {
1859 struct numa_group *ng = p->numa_group;
1861 if (env.best_cpu == -1)
1866 if (ng->active_nodes > 1 && numa_is_active_node(env.dst_nid, ng))
1867 sched_setnuma(p, env.dst_nid);
1870 /* No better CPU than the current one was found. */
1871 if (env.best_cpu == -1)
1875 * Reset the scan period if the task is being rescheduled on an
1876 * alternative node to recheck if the tasks is now properly placed.
1878 p->numa_scan_period = task_scan_start(p);
1880 if (env.best_task == NULL) {
1881 ret = migrate_task_to(p, env.best_cpu);
1883 trace_sched_stick_numa(p, env.src_cpu, env.best_cpu);
1887 ret = migrate_swap(p, env.best_task);
1889 trace_sched_stick_numa(p, env.src_cpu, task_cpu(env.best_task));
1890 put_task_struct(env.best_task);
1894 /* Attempt to migrate a task to a CPU on the preferred node. */
1895 static void numa_migrate_preferred(struct task_struct *p)
1897 unsigned long interval = HZ;
1899 /* This task has no NUMA fault statistics yet */
1900 if (unlikely(p->numa_preferred_nid == -1 || !p->numa_faults))
1903 /* Periodically retry migrating the task to the preferred node */
1904 interval = min(interval, msecs_to_jiffies(p->numa_scan_period) / 16);
1905 p->numa_migrate_retry = jiffies + interval;
1907 /* Success if task is already running on preferred CPU */
1908 if (task_node(p) == p->numa_preferred_nid)
1911 /* Otherwise, try migrate to a CPU on the preferred node */
1912 task_numa_migrate(p);
1916 * Find out how many nodes on the workload is actively running on. Do this by
1917 * tracking the nodes from which NUMA hinting faults are triggered. This can
1918 * be different from the set of nodes where the workload's memory is currently
1921 static void numa_group_count_active_nodes(struct numa_group *numa_group)
1923 unsigned long faults, max_faults = 0;
1924 int nid, active_nodes = 0;
1926 for_each_online_node(nid) {
1927 faults = group_faults_cpu(numa_group, nid);
1928 if (faults > max_faults)
1929 max_faults = faults;
1932 for_each_online_node(nid) {
1933 faults = group_faults_cpu(numa_group, nid);
1934 if (faults * ACTIVE_NODE_FRACTION > max_faults)
1938 numa_group->max_faults_cpu = max_faults;
1939 numa_group->active_nodes = active_nodes;
1943 * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
1944 * increments. The more local the fault statistics are, the higher the scan
1945 * period will be for the next scan window. If local/(local+remote) ratio is
1946 * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS)
1947 * the scan period will decrease. Aim for 70% local accesses.
1949 #define NUMA_PERIOD_SLOTS 10
1950 #define NUMA_PERIOD_THRESHOLD 7
1953 * Increase the scan period (slow down scanning) if the majority of
1954 * our memory is already on our local node, or if the majority of
1955 * the page accesses are shared with other processes.
1956 * Otherwise, decrease the scan period.
1958 static void update_task_scan_period(struct task_struct *p,
1959 unsigned long shared, unsigned long private)
1961 unsigned int period_slot;
1962 int lr_ratio, ps_ratio;
1965 unsigned long remote = p->numa_faults_locality[0];
1966 unsigned long local = p->numa_faults_locality[1];
1969 * If there were no record hinting faults then either the task is
1970 * completely idle or all activity is areas that are not of interest
1971 * to automatic numa balancing. Related to that, if there were failed
1972 * migration then it implies we are migrating too quickly or the local
1973 * node is overloaded. In either case, scan slower
1975 if (local + shared == 0 || p->numa_faults_locality[2]) {
1976 p->numa_scan_period = min(p->numa_scan_period_max,
1977 p->numa_scan_period << 1);
1979 p->mm->numa_next_scan = jiffies +
1980 msecs_to_jiffies(p->numa_scan_period);
1986 * Prepare to scale scan period relative to the current period.
1987 * == NUMA_PERIOD_THRESHOLD scan period stays the same
1988 * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
1989 * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
1991 period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS);
1992 lr_ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote);
1993 ps_ratio = (private * NUMA_PERIOD_SLOTS) / (private + shared);
1995 if (ps_ratio >= NUMA_PERIOD_THRESHOLD) {
1997 * Most memory accesses are local. There is no need to
1998 * do fast NUMA scanning, since memory is already local.
2000 int slot = ps_ratio - NUMA_PERIOD_THRESHOLD;
2003 diff = slot * period_slot;
2004 } else if (lr_ratio >= NUMA_PERIOD_THRESHOLD) {
2006 * Most memory accesses are shared with other tasks.
2007 * There is no point in continuing fast NUMA scanning,
2008 * since other tasks may just move the memory elsewhere.
2010 int slot = lr_ratio - NUMA_PERIOD_THRESHOLD;
2013 diff = slot * period_slot;
2016 * Private memory faults exceed (SLOTS-THRESHOLD)/SLOTS,
2017 * yet they are not on the local NUMA node. Speed up
2018 * NUMA scanning to get the memory moved over.
2020 int ratio = max(lr_ratio, ps_ratio);
2021 diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot;
2024 p->numa_scan_period = clamp(p->numa_scan_period + diff,
2025 task_scan_min(p), task_scan_max(p));
2026 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
2030 * Get the fraction of time the task has been running since the last
2031 * NUMA placement cycle. The scheduler keeps similar statistics, but
2032 * decays those on a 32ms period, which is orders of magnitude off
2033 * from the dozens-of-seconds NUMA balancing period. Use the scheduler
2034 * stats only if the task is so new there are no NUMA statistics yet.
2036 static u64 numa_get_avg_runtime(struct task_struct *p, u64 *period)
2038 u64 runtime, delta, now;
2039 /* Use the start of this time slice to avoid calculations. */
2040 now = p->se.exec_start;
2041 runtime = p->se.sum_exec_runtime;
2043 if (p->last_task_numa_placement) {
2044 delta = runtime - p->last_sum_exec_runtime;
2045 *period = now - p->last_task_numa_placement;
2047 delta = p->se.avg.load_sum;
2048 *period = LOAD_AVG_MAX;
2051 p->last_sum_exec_runtime = runtime;
2052 p->last_task_numa_placement = now;
2058 * Determine the preferred nid for a task in a numa_group. This needs to
2059 * be done in a way that produces consistent results with group_weight,
2060 * otherwise workloads might not converge.
2062 static int preferred_group_nid(struct task_struct *p, int nid)
2067 /* Direct connections between all NUMA nodes. */
2068 if (sched_numa_topology_type == NUMA_DIRECT)
2072 * On a system with glueless mesh NUMA topology, group_weight
2073 * scores nodes according to the number of NUMA hinting faults on
2074 * both the node itself, and on nearby nodes.
2076 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
2077 unsigned long score, max_score = 0;
2078 int node, max_node = nid;
2080 dist = sched_max_numa_distance;
2082 for_each_online_node(node) {
2083 score = group_weight(p, node, dist);
2084 if (score > max_score) {
2093 * Finding the preferred nid in a system with NUMA backplane
2094 * interconnect topology is more involved. The goal is to locate
2095 * tasks from numa_groups near each other in the system, and
2096 * untangle workloads from different sides of the system. This requires
2097 * searching down the hierarchy of node groups, recursively searching
2098 * inside the highest scoring group of nodes. The nodemask tricks
2099 * keep the complexity of the search down.
2101 nodes = node_online_map;
2102 for (dist = sched_max_numa_distance; dist > LOCAL_DISTANCE; dist--) {
2103 unsigned long max_faults = 0;
2104 nodemask_t max_group = NODE_MASK_NONE;
2107 /* Are there nodes at this distance from each other? */
2108 if (!find_numa_distance(dist))
2111 for_each_node_mask(a, nodes) {
2112 unsigned long faults = 0;
2113 nodemask_t this_group;
2114 nodes_clear(this_group);
2116 /* Sum group's NUMA faults; includes a==b case. */
2117 for_each_node_mask(b, nodes) {
2118 if (node_distance(a, b) < dist) {
2119 faults += group_faults(p, b);
2120 node_set(b, this_group);
2121 node_clear(b, nodes);
2125 /* Remember the top group. */
2126 if (faults > max_faults) {
2127 max_faults = faults;
2128 max_group = this_group;
2130 * subtle: at the smallest distance there is
2131 * just one node left in each "group", the
2132 * winner is the preferred nid.
2137 /* Next round, evaluate the nodes within max_group. */
2145 static void task_numa_placement(struct task_struct *p)
2147 int seq, nid, max_nid = -1, max_group_nid = -1;
2148 unsigned long max_faults = 0, max_group_faults = 0;
2149 unsigned long fault_types[2] = { 0, 0 };
2150 unsigned long total_faults;
2151 u64 runtime, period;
2152 spinlock_t *group_lock = NULL;
2155 * The p->mm->numa_scan_seq field gets updated without
2156 * exclusive access. Use READ_ONCE() here to ensure
2157 * that the field is read in a single access:
2159 seq = READ_ONCE(p->mm->numa_scan_seq);
2160 if (p->numa_scan_seq == seq)
2162 p->numa_scan_seq = seq;
2163 p->numa_scan_period_max = task_scan_max(p);
2165 total_faults = p->numa_faults_locality[0] +
2166 p->numa_faults_locality[1];
2167 runtime = numa_get_avg_runtime(p, &period);
2169 /* If the task is part of a group prevent parallel updates to group stats */
2170 if (p->numa_group) {
2171 group_lock = &p->numa_group->lock;
2172 spin_lock_irq(group_lock);
2175 /* Find the node with the highest number of faults */
2176 for_each_online_node(nid) {
2177 /* Keep track of the offsets in numa_faults array */
2178 int mem_idx, membuf_idx, cpu_idx, cpubuf_idx;
2179 unsigned long faults = 0, group_faults = 0;
2182 for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) {
2183 long diff, f_diff, f_weight;
2185 mem_idx = task_faults_idx(NUMA_MEM, nid, priv);
2186 membuf_idx = task_faults_idx(NUMA_MEMBUF, nid, priv);
2187 cpu_idx = task_faults_idx(NUMA_CPU, nid, priv);
2188 cpubuf_idx = task_faults_idx(NUMA_CPUBUF, nid, priv);
2190 /* Decay existing window, copy faults since last scan */
2191 diff = p->numa_faults[membuf_idx] - p->numa_faults[mem_idx] / 2;
2192 fault_types[priv] += p->numa_faults[membuf_idx];
2193 p->numa_faults[membuf_idx] = 0;
2196 * Normalize the faults_from, so all tasks in a group
2197 * count according to CPU use, instead of by the raw
2198 * number of faults. Tasks with little runtime have
2199 * little over-all impact on throughput, and thus their
2200 * faults are less important.
2202 f_weight = div64_u64(runtime << 16, period + 1);
2203 f_weight = (f_weight * p->numa_faults[cpubuf_idx]) /
2205 f_diff = f_weight - p->numa_faults[cpu_idx] / 2;
2206 p->numa_faults[cpubuf_idx] = 0;
2208 p->numa_faults[mem_idx] += diff;
2209 p->numa_faults[cpu_idx] += f_diff;
2210 faults += p->numa_faults[mem_idx];
2211 p->total_numa_faults += diff;
2212 if (p->numa_group) {
2214 * safe because we can only change our own group
2216 * mem_idx represents the offset for a given
2217 * nid and priv in a specific region because it
2218 * is at the beginning of the numa_faults array.
2220 p->numa_group->faults[mem_idx] += diff;
2221 p->numa_group->faults_cpu[mem_idx] += f_diff;
2222 p->numa_group->total_faults += diff;
2223 group_faults += p->numa_group->faults[mem_idx];
2227 if (faults > max_faults) {
2228 max_faults = faults;
2232 if (group_faults > max_group_faults) {
2233 max_group_faults = group_faults;
2234 max_group_nid = nid;
2238 update_task_scan_period(p, fault_types[0], fault_types[1]);
2240 if (p->numa_group) {
2241 numa_group_count_active_nodes(p->numa_group);
2242 spin_unlock_irq(group_lock);
2243 max_nid = preferred_group_nid(p, max_group_nid);
2247 /* Set the new preferred node */
2248 if (max_nid != p->numa_preferred_nid)
2249 sched_setnuma(p, max_nid);
2251 if (task_node(p) != p->numa_preferred_nid)
2252 numa_migrate_preferred(p);
2256 static inline int get_numa_group(struct numa_group *grp)
2258 return atomic_inc_not_zero(&grp->refcount);
2261 static inline void put_numa_group(struct numa_group *grp)
2263 if (atomic_dec_and_test(&grp->refcount))
2264 kfree_rcu(grp, rcu);
2267 static void task_numa_group(struct task_struct *p, int cpupid, int flags,
2270 struct numa_group *grp, *my_grp;
2271 struct task_struct *tsk;
2273 int cpu = cpupid_to_cpu(cpupid);
2276 if (unlikely(!p->numa_group)) {
2277 unsigned int size = sizeof(struct numa_group) +
2278 4*nr_node_ids*sizeof(unsigned long);
2280 grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
2284 atomic_set(&grp->refcount, 1);
2285 grp->active_nodes = 1;
2286 grp->max_faults_cpu = 0;
2287 spin_lock_init(&grp->lock);
2289 /* Second half of the array tracks nids where faults happen */
2290 grp->faults_cpu = grp->faults + NR_NUMA_HINT_FAULT_TYPES *
2293 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2294 grp->faults[i] = p->numa_faults[i];
2296 grp->total_faults = p->total_numa_faults;
2299 rcu_assign_pointer(p->numa_group, grp);
2303 tsk = READ_ONCE(cpu_rq(cpu)->curr);
2305 if (!cpupid_match_pid(tsk, cpupid))
2308 grp = rcu_dereference(tsk->numa_group);
2312 my_grp = p->numa_group;
2317 * Only join the other group if its bigger; if we're the bigger group,
2318 * the other task will join us.
2320 if (my_grp->nr_tasks > grp->nr_tasks)
2324 * Tie-break on the grp address.
2326 if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp)
2329 /* Always join threads in the same process. */
2330 if (tsk->mm == current->mm)
2333 /* Simple filter to avoid false positives due to PID collisions */
2334 if (flags & TNF_SHARED)
2337 /* Update priv based on whether false sharing was detected */
2340 if (join && !get_numa_group(grp))
2348 BUG_ON(irqs_disabled());
2349 double_lock_irq(&my_grp->lock, &grp->lock);
2351 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) {
2352 my_grp->faults[i] -= p->numa_faults[i];
2353 grp->faults[i] += p->numa_faults[i];
2355 my_grp->total_faults -= p->total_numa_faults;
2356 grp->total_faults += p->total_numa_faults;
2361 spin_unlock(&my_grp->lock);
2362 spin_unlock_irq(&grp->lock);
2364 rcu_assign_pointer(p->numa_group, grp);
2366 put_numa_group(my_grp);
2374 void task_numa_free(struct task_struct *p)
2376 struct numa_group *grp = p->numa_group;
2377 void *numa_faults = p->numa_faults;
2378 unsigned long flags;
2382 spin_lock_irqsave(&grp->lock, flags);
2383 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2384 grp->faults[i] -= p->numa_faults[i];
2385 grp->total_faults -= p->total_numa_faults;
2388 spin_unlock_irqrestore(&grp->lock, flags);
2389 RCU_INIT_POINTER(p->numa_group, NULL);
2390 put_numa_group(grp);
2393 p->numa_faults = NULL;
2398 * Got a PROT_NONE fault for a page on @node.
2400 void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags)
2402 struct task_struct *p = current;
2403 bool migrated = flags & TNF_MIGRATED;
2404 int cpu_node = task_node(current);
2405 int local = !!(flags & TNF_FAULT_LOCAL);
2406 struct numa_group *ng;
2409 if (!static_branch_likely(&sched_numa_balancing))
2412 /* for example, ksmd faulting in a user's mm */
2416 /* Allocate buffer to track faults on a per-node basis */
2417 if (unlikely(!p->numa_faults)) {
2418 int size = sizeof(*p->numa_faults) *
2419 NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids;
2421 p->numa_faults = kzalloc(size, GFP_KERNEL|__GFP_NOWARN);
2422 if (!p->numa_faults)
2425 p->total_numa_faults = 0;
2426 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
2430 * First accesses are treated as private, otherwise consider accesses
2431 * to be private if the accessing pid has not changed
2433 if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) {
2436 priv = cpupid_match_pid(p, last_cpupid);
2437 if (!priv && !(flags & TNF_NO_GROUP))
2438 task_numa_group(p, last_cpupid, flags, &priv);
2442 * If a workload spans multiple NUMA nodes, a shared fault that
2443 * occurs wholly within the set of nodes that the workload is
2444 * actively using should be counted as local. This allows the
2445 * scan rate to slow down when a workload has settled down.
2448 if (!priv && !local && ng && ng->active_nodes > 1 &&
2449 numa_is_active_node(cpu_node, ng) &&
2450 numa_is_active_node(mem_node, ng))
2453 task_numa_placement(p);
2456 * Retry task to preferred node migration periodically, in case it
2457 * case it previously failed, or the scheduler moved us.
2459 if (time_after(jiffies, p->numa_migrate_retry))
2460 numa_migrate_preferred(p);
2463 p->numa_pages_migrated += pages;
2464 if (flags & TNF_MIGRATE_FAIL)
2465 p->numa_faults_locality[2] += pages;
2467 p->numa_faults[task_faults_idx(NUMA_MEMBUF, mem_node, priv)] += pages;
2468 p->numa_faults[task_faults_idx(NUMA_CPUBUF, cpu_node, priv)] += pages;
2469 p->numa_faults_locality[local] += pages;
2472 static void reset_ptenuma_scan(struct task_struct *p)
2475 * We only did a read acquisition of the mmap sem, so
2476 * p->mm->numa_scan_seq is written to without exclusive access
2477 * and the update is not guaranteed to be atomic. That's not
2478 * much of an issue though, since this is just used for
2479 * statistical sampling. Use READ_ONCE/WRITE_ONCE, which are not
2480 * expensive, to avoid any form of compiler optimizations:
2482 WRITE_ONCE(p->mm->numa_scan_seq, READ_ONCE(p->mm->numa_scan_seq) + 1);
2483 p->mm->numa_scan_offset = 0;
2487 * The expensive part of numa migration is done from task_work context.
2488 * Triggered from task_tick_numa().
2490 void task_numa_work(struct callback_head *work)
2492 unsigned long migrate, next_scan, now = jiffies;
2493 struct task_struct *p = current;
2494 struct mm_struct *mm = p->mm;
2495 u64 runtime = p->se.sum_exec_runtime;
2496 struct vm_area_struct *vma;
2497 unsigned long start, end;
2498 unsigned long nr_pte_updates = 0;
2499 long pages, virtpages;
2501 SCHED_WARN_ON(p != container_of(work, struct task_struct, numa_work));
2503 work->next = work; /* protect against double add */
2505 * Who cares about NUMA placement when they're dying.
2507 * NOTE: make sure not to dereference p->mm before this check,
2508 * exit_task_work() happens _after_ exit_mm() so we could be called
2509 * without p->mm even though we still had it when we enqueued this
2512 if (p->flags & PF_EXITING)
2515 if (!mm->numa_next_scan) {
2516 mm->numa_next_scan = now +
2517 msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2521 * Enforce maximal scan/migration frequency..
2523 migrate = mm->numa_next_scan;
2524 if (time_before(now, migrate))
2527 if (p->numa_scan_period == 0) {
2528 p->numa_scan_period_max = task_scan_max(p);
2529 p->numa_scan_period = task_scan_start(p);
2532 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
2533 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
2537 * Delay this task enough that another task of this mm will likely win
2538 * the next time around.
2540 p->node_stamp += 2 * TICK_NSEC;
2542 start = mm->numa_scan_offset;
2543 pages = sysctl_numa_balancing_scan_size;
2544 pages <<= 20 - PAGE_SHIFT; /* MB in pages */
2545 virtpages = pages * 8; /* Scan up to this much virtual space */
2550 if (!down_read_trylock(&mm->mmap_sem))
2552 vma = find_vma(mm, start);
2554 reset_ptenuma_scan(p);
2558 for (; vma; vma = vma->vm_next) {
2559 if (!vma_migratable(vma) || !vma_policy_mof(vma) ||
2560 is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_MIXEDMAP)) {
2565 * Shared library pages mapped by multiple processes are not
2566 * migrated as it is expected they are cache replicated. Avoid
2567 * hinting faults in read-only file-backed mappings or the vdso
2568 * as migrating the pages will be of marginal benefit.
2571 (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ)))
2575 * Skip inaccessible VMAs to avoid any confusion between
2576 * PROT_NONE and NUMA hinting ptes
2578 if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)))
2582 start = max(start, vma->vm_start);
2583 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
2584 end = min(end, vma->vm_end);
2585 nr_pte_updates = change_prot_numa(vma, start, end);
2588 * Try to scan sysctl_numa_balancing_size worth of
2589 * hpages that have at least one present PTE that
2590 * is not already pte-numa. If the VMA contains
2591 * areas that are unused or already full of prot_numa
2592 * PTEs, scan up to virtpages, to skip through those
2596 pages -= (end - start) >> PAGE_SHIFT;
2597 virtpages -= (end - start) >> PAGE_SHIFT;
2600 if (pages <= 0 || virtpages <= 0)
2604 } while (end != vma->vm_end);
2609 * It is possible to reach the end of the VMA list but the last few
2610 * VMAs are not guaranteed to the vma_migratable. If they are not, we
2611 * would find the !migratable VMA on the next scan but not reset the
2612 * scanner to the start so check it now.
2615 mm->numa_scan_offset = start;
2617 reset_ptenuma_scan(p);
2618 up_read(&mm->mmap_sem);
2621 * Make sure tasks use at least 32x as much time to run other code
2622 * than they used here, to limit NUMA PTE scanning overhead to 3% max.
2623 * Usually update_task_scan_period slows down scanning enough; on an
2624 * overloaded system we need to limit overhead on a per task basis.
2626 if (unlikely(p->se.sum_exec_runtime != runtime)) {
2627 u64 diff = p->se.sum_exec_runtime - runtime;
2628 p->node_stamp += 32 * diff;
2633 * Drive the periodic memory faults..
2635 void task_tick_numa(struct rq *rq, struct task_struct *curr)
2637 struct callback_head *work = &curr->numa_work;
2641 * We don't care about NUMA placement if we don't have memory.
2643 if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
2647 * Using runtime rather than walltime has the dual advantage that
2648 * we (mostly) drive the selection from busy threads and that the
2649 * task needs to have done some actual work before we bother with
2652 now = curr->se.sum_exec_runtime;
2653 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
2655 if (now > curr->node_stamp + period) {
2656 if (!curr->node_stamp)
2657 curr->numa_scan_period = task_scan_start(curr);
2658 curr->node_stamp += period;
2660 if (!time_before(jiffies, curr->mm->numa_next_scan)) {
2661 init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */
2662 task_work_add(curr, work, true);
2668 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
2672 static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p)
2676 static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p)
2680 #endif /* CONFIG_NUMA_BALANCING */
2683 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2685 update_load_add(&cfs_rq->load, se->load.weight);
2686 if (!parent_entity(se))
2687 update_load_add(&rq_of(cfs_rq)->load, se->load.weight);
2689 if (entity_is_task(se)) {
2690 struct rq *rq = rq_of(cfs_rq);
2692 account_numa_enqueue(rq, task_of(se));
2693 list_add(&se->group_node, &rq->cfs_tasks);
2696 cfs_rq->nr_running++;
2700 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2702 update_load_sub(&cfs_rq->load, se->load.weight);
2703 if (!parent_entity(se))
2704 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
2706 if (entity_is_task(se)) {
2707 account_numa_dequeue(rq_of(cfs_rq), task_of(se));
2708 list_del_init(&se->group_node);
2711 cfs_rq->nr_running--;
2715 * Signed add and clamp on underflow.
2717 * Explicitly do a load-store to ensure the intermediate value never hits
2718 * memory. This allows lockless observations without ever seeing the negative
2721 #define add_positive(_ptr, _val) do { \
2722 typeof(_ptr) ptr = (_ptr); \
2723 typeof(_val) val = (_val); \
2724 typeof(*ptr) res, var = READ_ONCE(*ptr); \
2728 if (val < 0 && res > var) \
2731 WRITE_ONCE(*ptr, res); \
2735 * Unsigned subtract and clamp on underflow.
2737 * Explicitly do a load-store to ensure the intermediate value never hits
2738 * memory. This allows lockless observations without ever seeing the negative
2741 #define sub_positive(_ptr, _val) do { \
2742 typeof(_ptr) ptr = (_ptr); \
2743 typeof(*ptr) val = (_val); \
2744 typeof(*ptr) res, var = READ_ONCE(*ptr); \
2748 WRITE_ONCE(*ptr, res); \
2753 * XXX we want to get rid of these helpers and use the full load resolution.
2755 static inline long se_weight(struct sched_entity *se)
2757 return scale_load_down(se->load.weight);
2760 static inline long se_runnable(struct sched_entity *se)
2762 return scale_load_down(se->runnable_weight);
2766 enqueue_runnable_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2768 cfs_rq->runnable_weight += se->runnable_weight;
2770 cfs_rq->avg.runnable_load_avg += se->avg.runnable_load_avg;
2771 cfs_rq->avg.runnable_load_sum += se_runnable(se) * se->avg.runnable_load_sum;
2775 dequeue_runnable_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2777 cfs_rq->runnable_weight -= se->runnable_weight;
2779 sub_positive(&cfs_rq->avg.runnable_load_avg, se->avg.runnable_load_avg);
2780 sub_positive(&cfs_rq->avg.runnable_load_sum,
2781 se_runnable(se) * se->avg.runnable_load_sum);
2785 enqueue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2787 cfs_rq->avg.load_avg += se->avg.load_avg;
2788 cfs_rq->avg.load_sum += se_weight(se) * se->avg.load_sum;
2792 dequeue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2794 sub_positive(&cfs_rq->avg.load_avg, se->avg.load_avg);
2795 sub_positive(&cfs_rq->avg.load_sum, se_weight(se) * se->avg.load_sum);
2799 enqueue_runnable_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
2801 dequeue_runnable_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
2803 enqueue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
2805 dequeue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
2808 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
2809 unsigned long weight, unsigned long runnable)
2812 /* commit outstanding execution time */
2813 if (cfs_rq->curr == se)
2814 update_curr(cfs_rq);
2815 account_entity_dequeue(cfs_rq, se);
2816 dequeue_runnable_load_avg(cfs_rq, se);
2818 dequeue_load_avg(cfs_rq, se);
2820 se->runnable_weight = runnable;
2821 update_load_set(&se->load, weight);
2825 u32 divider = LOAD_AVG_MAX - 1024 + se->avg.period_contrib;
2827 se->avg.load_avg = div_u64(se_weight(se) * se->avg.load_sum, divider);
2828 se->avg.runnable_load_avg =
2829 div_u64(se_runnable(se) * se->avg.runnable_load_sum, divider);
2833 enqueue_load_avg(cfs_rq, se);
2835 account_entity_enqueue(cfs_rq, se);
2836 enqueue_runnable_load_avg(cfs_rq, se);
2840 void reweight_task(struct task_struct *p, int prio)
2842 struct sched_entity *se = &p->se;
2843 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2844 struct load_weight *load = &se->load;
2845 unsigned long weight = scale_load(sched_prio_to_weight[prio]);
2847 reweight_entity(cfs_rq, se, weight, weight);
2848 load->inv_weight = sched_prio_to_wmult[prio];
2851 #ifdef CONFIG_FAIR_GROUP_SCHED
2854 * All this does is approximate the hierarchical proportion which includes that
2855 * global sum we all love to hate.
2857 * That is, the weight of a group entity, is the proportional share of the
2858 * group weight based on the group runqueue weights. That is:
2860 * tg->weight * grq->load.weight
2861 * ge->load.weight = ----------------------------- (1)
2862 * \Sum grq->load.weight
2864 * Now, because computing that sum is prohibitively expensive to compute (been
2865 * there, done that) we approximate it with this average stuff. The average
2866 * moves slower and therefore the approximation is cheaper and more stable.
2868 * So instead of the above, we substitute:
2870 * grq->load.weight -> grq->avg.load_avg (2)
2872 * which yields the following:
2874 * tg->weight * grq->avg.load_avg
2875 * ge->load.weight = ------------------------------ (3)
2878 * Where: tg->load_avg ~= \Sum grq->avg.load_avg
2880 * That is shares_avg, and it is right (given the approximation (2)).
2882 * The problem with it is that because the average is slow -- it was designed
2883 * to be exactly that of course -- this leads to transients in boundary
2884 * conditions. In specific, the case where the group was idle and we start the
2885 * one task. It takes time for our CPU's grq->avg.load_avg to build up,
2886 * yielding bad latency etc..
2888 * Now, in that special case (1) reduces to:
2890 * tg->weight * grq->load.weight
2891 * ge->load.weight = ----------------------------- = tg->weight (4)
2894 * That is, the sum collapses because all other CPUs are idle; the UP scenario.
2896 * So what we do is modify our approximation (3) to approach (4) in the (near)
2901 * tg->weight * grq->load.weight
2902 * --------------------------------------------------- (5)
2903 * tg->load_avg - grq->avg.load_avg + grq->load.weight
2905 * But because grq->load.weight can drop to 0, resulting in a divide by zero,
2906 * we need to use grq->avg.load_avg as its lower bound, which then gives:
2909 * tg->weight * grq->load.weight
2910 * ge->load.weight = ----------------------------- (6)
2915 * tg_load_avg' = tg->load_avg - grq->avg.load_avg +
2916 * max(grq->load.weight, grq->avg.load_avg)
2918 * And that is shares_weight and is icky. In the (near) UP case it approaches
2919 * (4) while in the normal case it approaches (3). It consistently
2920 * overestimates the ge->load.weight and therefore:
2922 * \Sum ge->load.weight >= tg->weight
2926 static long calc_group_shares(struct cfs_rq *cfs_rq)
2928 long tg_weight, tg_shares, load, shares;
2929 struct task_group *tg = cfs_rq->tg;
2931 tg_shares = READ_ONCE(tg->shares);
2933 load = max(scale_load_down(cfs_rq->load.weight), cfs_rq->avg.load_avg);
2935 tg_weight = atomic_long_read(&tg->load_avg);
2937 /* Ensure tg_weight >= load */
2938 tg_weight -= cfs_rq->tg_load_avg_contrib;
2941 shares = (tg_shares * load);
2943 shares /= tg_weight;
2946 * MIN_SHARES has to be unscaled here to support per-CPU partitioning
2947 * of a group with small tg->shares value. It is a floor value which is
2948 * assigned as a minimum load.weight to the sched_entity representing
2949 * the group on a CPU.
2951 * E.g. on 64-bit for a group with tg->shares of scale_load(15)=15*1024
2952 * on an 8-core system with 8 tasks each runnable on one CPU shares has
2953 * to be 15*1024*1/8=1920 instead of scale_load(MIN_SHARES)=2*1024. In
2954 * case no task is runnable on a CPU MIN_SHARES=2 should be returned
2957 return clamp_t(long, shares, MIN_SHARES, tg_shares);
2961 * This calculates the effective runnable weight for a group entity based on
2962 * the group entity weight calculated above.
2964 * Because of the above approximation (2), our group entity weight is
2965 * an load_avg based ratio (3). This means that it includes blocked load and
2966 * does not represent the runnable weight.
2968 * Approximate the group entity's runnable weight per ratio from the group
2971 * grq->avg.runnable_load_avg
2972 * ge->runnable_weight = ge->load.weight * -------------------------- (7)
2975 * However, analogous to above, since the avg numbers are slow, this leads to
2976 * transients in the from-idle case. Instead we use:
2978 * ge->runnable_weight = ge->load.weight *
2980 * max(grq->avg.runnable_load_avg, grq->runnable_weight)
2981 * ----------------------------------------------------- (8)
2982 * max(grq->avg.load_avg, grq->load.weight)
2984 * Where these max() serve both to use the 'instant' values to fix the slow
2985 * from-idle and avoid the /0 on to-idle, similar to (6).
2987 static long calc_group_runnable(struct cfs_rq *cfs_rq, long shares)
2989 long runnable, load_avg;
2991 load_avg = max(cfs_rq->avg.load_avg,
2992 scale_load_down(cfs_rq->load.weight));
2994 runnable = max(cfs_rq->avg.runnable_load_avg,
2995 scale_load_down(cfs_rq->runnable_weight));
2999 runnable /= load_avg;
3001 return clamp_t(long, runnable, MIN_SHARES, shares);
3003 #endif /* CONFIG_SMP */
3005 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
3008 * Recomputes the group entity based on the current state of its group
3011 static void update_cfs_group(struct sched_entity *se)
3013 struct cfs_rq *gcfs_rq = group_cfs_rq(se);
3014 long shares, runnable;
3019 if (throttled_hierarchy(gcfs_rq))
3023 runnable = shares = READ_ONCE(gcfs_rq->tg->shares);
3025 if (likely(se->load.weight == shares))
3028 shares = calc_group_shares(gcfs_rq);
3029 runnable = calc_group_runnable(gcfs_rq, shares);
3032 reweight_entity(cfs_rq_of(se), se, shares, runnable);
3035 #else /* CONFIG_FAIR_GROUP_SCHED */
3036 static inline void update_cfs_group(struct sched_entity *se)
3039 #endif /* CONFIG_FAIR_GROUP_SCHED */
3041 static inline void cfs_rq_util_change(struct cfs_rq *cfs_rq, int flags)
3043 struct rq *rq = rq_of(cfs_rq);
3045 if (&rq->cfs == cfs_rq || (flags & SCHED_CPUFREQ_MIGRATION)) {
3047 * There are a few boundary cases this might miss but it should
3048 * get called often enough that that should (hopefully) not be
3051 * It will not get called when we go idle, because the idle
3052 * thread is a different class (!fair), nor will the utilization
3053 * number include things like RT tasks.
3055 * As is, the util number is not freq-invariant (we'd have to
3056 * implement arch_scale_freq_capacity() for that).
3060 cpufreq_update_util(rq, flags);
3067 * val * y^n, where y^32 ~= 0.5 (~1 scheduling period)
3069 static u64 decay_load(u64 val, u64 n)
3071 unsigned int local_n;
3073 if (unlikely(n > LOAD_AVG_PERIOD * 63))
3076 /* after bounds checking we can collapse to 32-bit */
3080 * As y^PERIOD = 1/2, we can combine
3081 * y^n = 1/2^(n/PERIOD) * y^(n%PERIOD)
3082 * With a look-up table which covers y^n (n<PERIOD)
3084 * To achieve constant time decay_load.
3086 if (unlikely(local_n >= LOAD_AVG_PERIOD)) {
3087 val >>= local_n / LOAD_AVG_PERIOD;
3088 local_n %= LOAD_AVG_PERIOD;
3091 val = mul_u64_u32_shr(val, runnable_avg_yN_inv[local_n], 32);
3095 static u32 __accumulate_pelt_segments(u64 periods, u32 d1, u32 d3)
3097 u32 c1, c2, c3 = d3; /* y^0 == 1 */
3102 c1 = decay_load((u64)d1, periods);
3106 * c2 = 1024 \Sum y^n
3110 * = 1024 ( \Sum y^n - \Sum y^n - y^0 )
3113 c2 = LOAD_AVG_MAX - decay_load(LOAD_AVG_MAX, periods) - 1024;
3115 return c1 + c2 + c3;
3119 * Accumulate the three separate parts of the sum; d1 the remainder
3120 * of the last (incomplete) period, d2 the span of full periods and d3
3121 * the remainder of the (incomplete) current period.
3126 * |<->|<----------------->|<--->|
3127 * ... |---x---|------| ... |------|-----x (now)
3130 * u' = (u + d1) y^p + 1024 \Sum y^n + d3 y^0
3133 * = u y^p + (Step 1)
3136 * d1 y^p + 1024 \Sum y^n + d3 y^0 (Step 2)
3139 static __always_inline u32
3140 accumulate_sum(u64 delta, int cpu, struct sched_avg *sa,
3141 unsigned long load, unsigned long runnable, int running)
3143 unsigned long scale_freq, scale_cpu;
3144 u32 contrib = (u32)delta; /* p == 0 -> delta < 1024 */
3147 scale_freq = arch_scale_freq_capacity(cpu);
3148 scale_cpu = arch_scale_cpu_capacity(NULL, cpu);
3150 delta += sa->period_contrib;
3151 periods = delta / 1024; /* A period is 1024us (~1ms) */
3154 * Step 1: decay old *_sum if we crossed period boundaries.
3157 sa->load_sum = decay_load(sa->load_sum, periods);
3158 sa->runnable_load_sum =
3159 decay_load(sa->runnable_load_sum, periods);
3160 sa->util_sum = decay_load((u64)(sa->util_sum), periods);
3166 contrib = __accumulate_pelt_segments(periods,
3167 1024 - sa->period_contrib, delta);
3169 sa->period_contrib = delta;
3171 contrib = cap_scale(contrib, scale_freq);
3173 sa->load_sum += load * contrib;
3175 sa->runnable_load_sum += runnable * contrib;
3177 sa->util_sum += contrib * scale_cpu;
3183 * We can represent the historical contribution to runnable average as the
3184 * coefficients of a geometric series. To do this we sub-divide our runnable
3185 * history into segments of approximately 1ms (1024us); label the segment that
3186 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
3188 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
3190 * (now) (~1ms ago) (~2ms ago)
3192 * Let u_i denote the fraction of p_i that the entity was runnable.
3194 * We then designate the fractions u_i as our co-efficients, yielding the
3195 * following representation of historical load:
3196 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
3198 * We choose y based on the with of a reasonably scheduling period, fixing:
3201 * This means that the contribution to load ~32ms ago (u_32) will be weighted
3202 * approximately half as much as the contribution to load within the last ms
3205 * When a period "rolls over" and we have new u_0`, multiplying the previous
3206 * sum again by y is sufficient to update:
3207 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
3208 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
3210 static __always_inline int
3211 ___update_load_sum(u64 now, int cpu, struct sched_avg *sa,
3212 unsigned long load, unsigned long runnable, int running)
3216 delta = now - sa->last_update_time;
3218 * This should only happen when time goes backwards, which it
3219 * unfortunately does during sched clock init when we swap over to TSC.
3221 if ((s64)delta < 0) {
3222 sa->last_update_time = now;
3227 * Use 1024ns as the unit of measurement since it's a reasonable
3228 * approximation of 1us and fast to compute.
3234 sa->last_update_time += delta << 10;
3237 * running is a subset of runnable (weight) so running can't be set if
3238 * runnable is clear. But there are some corner cases where the current
3239 * se has been already dequeued but cfs_rq->curr still points to it.
3240 * This means that weight will be 0 but not running for a sched_entity
3241 * but also for a cfs_rq if the latter becomes idle. As an example,
3242 * this happens during idle_balance() which calls
3243 * update_blocked_averages()
3246 runnable = running = 0;
3249 * Now we know we crossed measurement unit boundaries. The *_avg
3250 * accrues by two steps:
3252 * Step 1: accumulate *_sum since last_update_time. If we haven't
3253 * crossed period boundaries, finish.
3255 if (!accumulate_sum(delta, cpu, sa, load, runnable, running))
3261 static __always_inline void
3262 ___update_load_avg(struct sched_avg *sa, unsigned long load, unsigned long runnable)
3264 u32 divider = LOAD_AVG_MAX - 1024 + sa->period_contrib;
3267 * Step 2: update *_avg.
3269 sa->load_avg = div_u64(load * sa->load_sum, divider);
3270 sa->runnable_load_avg = div_u64(runnable * sa->runnable_load_sum, divider);
3271 sa->util_avg = sa->util_sum / divider;
3275 * When a task is dequeued, its estimated utilization should not be update if
3276 * its util_avg has not been updated at least once.
3277 * This flag is used to synchronize util_avg updates with util_est updates.
3278 * We map this information into the LSB bit of the utilization saved at
3279 * dequeue time (i.e. util_est.dequeued).
3281 #define UTIL_AVG_UNCHANGED 0x1
3283 static inline void cfs_se_util_change(struct sched_avg *avg)
3285 unsigned int enqueued;
3287 if (!sched_feat(UTIL_EST))
3290 /* Avoid store if the flag has been already set */
3291 enqueued = avg->util_est.enqueued;
3292 if (!(enqueued & UTIL_AVG_UNCHANGED))
3295 /* Reset flag to report util_avg has been updated */
3296 enqueued &= ~UTIL_AVG_UNCHANGED;
3297 WRITE_ONCE(avg->util_est.enqueued, enqueued);
3304 * se_runnable() == se_weight()
3306 * group: [ see update_cfs_group() ]
3307 * se_weight() = tg->weight * grq->load_avg / tg->load_avg
3308 * se_runnable() = se_weight(se) * grq->runnable_load_avg / grq->load_avg
3310 * load_sum := runnable_sum
3311 * load_avg = se_weight(se) * runnable_avg
3313 * runnable_load_sum := runnable_sum
3314 * runnable_load_avg = se_runnable(se) * runnable_avg
3316 * XXX collapse load_sum and runnable_load_sum
3320 * load_sum = \Sum se_weight(se) * se->avg.load_sum
3321 * load_avg = \Sum se->avg.load_avg
3323 * runnable_load_sum = \Sum se_runnable(se) * se->avg.runnable_load_sum
3324 * runnable_load_avg = \Sum se->avg.runable_load_avg
3328 __update_load_avg_blocked_se(u64 now, int cpu, struct sched_entity *se)
3330 if (entity_is_task(se))
3331 se->runnable_weight = se->load.weight;
3333 if (___update_load_sum(now, cpu, &se->avg, 0, 0, 0)) {
3334 ___update_load_avg(&se->avg, se_weight(se), se_runnable(se));
3342 __update_load_avg_se(u64 now, int cpu, struct cfs_rq *cfs_rq, struct sched_entity *se)
3344 if (entity_is_task(se))
3345 se->runnable_weight = se->load.weight;
3347 if (___update_load_sum(now, cpu, &se->avg, !!se->on_rq, !!se->on_rq,
3348 cfs_rq->curr == se)) {
3350 ___update_load_avg(&se->avg, se_weight(se), se_runnable(se));
3351 cfs_se_util_change(&se->avg);
3359 __update_load_avg_cfs_rq(u64 now, int cpu, struct cfs_rq *cfs_rq)
3361 if (___update_load_sum(now, cpu, &cfs_rq->avg,
3362 scale_load_down(cfs_rq->load.weight),
3363 scale_load_down(cfs_rq->runnable_weight),
3364 cfs_rq->curr != NULL)) {
3366 ___update_load_avg(&cfs_rq->avg, 1, 1);
3373 #ifdef CONFIG_FAIR_GROUP_SCHED
3375 * update_tg_load_avg - update the tg's load avg
3376 * @cfs_rq: the cfs_rq whose avg changed
3377 * @force: update regardless of how small the difference
3379 * This function 'ensures': tg->load_avg := \Sum tg->cfs_rq[]->avg.load.
3380 * However, because tg->load_avg is a global value there are performance
3383 * In order to avoid having to look at the other cfs_rq's, we use a
3384 * differential update where we store the last value we propagated. This in
3385 * turn allows skipping updates if the differential is 'small'.
3387 * Updating tg's load_avg is necessary before update_cfs_share().
3389 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force)
3391 long delta = cfs_rq->avg.load_avg - cfs_rq->tg_load_avg_contrib;
3394 * No need to update load_avg for root_task_group as it is not used.
3396 if (cfs_rq->tg == &root_task_group)
3399 if (force || abs(delta) > cfs_rq->tg_load_avg_contrib / 64) {
3400 atomic_long_add(delta, &cfs_rq->tg->load_avg);
3401 cfs_rq->tg_load_avg_contrib = cfs_rq->avg.load_avg;
3406 * Called within set_task_rq() right before setting a task's CPU. The
3407 * caller only guarantees p->pi_lock is held; no other assumptions,
3408 * including the state of rq->lock, should be made.
3410 void set_task_rq_fair(struct sched_entity *se,
3411 struct cfs_rq *prev, struct cfs_rq *next)
3413 u64 p_last_update_time;
3414 u64 n_last_update_time;
3416 if (!sched_feat(ATTACH_AGE_LOAD))
3420 * We are supposed to update the task to "current" time, then its up to
3421 * date and ready to go to new CPU/cfs_rq. But we have difficulty in
3422 * getting what current time is, so simply throw away the out-of-date
3423 * time. This will result in the wakee task is less decayed, but giving
3424 * the wakee more load sounds not bad.
3426 if (!(se->avg.last_update_time && prev))
3429 #ifndef CONFIG_64BIT
3431 u64 p_last_update_time_copy;
3432 u64 n_last_update_time_copy;
3435 p_last_update_time_copy = prev->load_last_update_time_copy;
3436 n_last_update_time_copy = next->load_last_update_time_copy;
3440 p_last_update_time = prev->avg.last_update_time;
3441 n_last_update_time = next->avg.last_update_time;
3443 } while (p_last_update_time != p_last_update_time_copy ||
3444 n_last_update_time != n_last_update_time_copy);
3447 p_last_update_time = prev->avg.last_update_time;
3448 n_last_update_time = next->avg.last_update_time;
3450 __update_load_avg_blocked_se(p_last_update_time, cpu_of(rq_of(prev)), se);
3451 se->avg.last_update_time = n_last_update_time;
3456 * When on migration a sched_entity joins/leaves the PELT hierarchy, we need to
3457 * propagate its contribution. The key to this propagation is the invariant
3458 * that for each group:
3460 * ge->avg == grq->avg (1)
3462 * _IFF_ we look at the pure running and runnable sums. Because they
3463 * represent the very same entity, just at different points in the hierarchy.
3465 * Per the above update_tg_cfs_util() is trivial and simply copies the running
3466 * sum over (but still wrong, because the group entity and group rq do not have
3467 * their PELT windows aligned).
3469 * However, update_tg_cfs_runnable() is more complex. So we have:
3471 * ge->avg.load_avg = ge->load.weight * ge->avg.runnable_avg (2)
3473 * And since, like util, the runnable part should be directly transferable,
3474 * the following would _appear_ to be the straight forward approach:
3476 * grq->avg.load_avg = grq->load.weight * grq->avg.runnable_avg (3)
3478 * And per (1) we have:
3480 * ge->avg.runnable_avg == grq->avg.runnable_avg
3484 * ge->load.weight * grq->avg.load_avg
3485 * ge->avg.load_avg = ----------------------------------- (4)
3488 * Except that is wrong!
3490 * Because while for entities historical weight is not important and we
3491 * really only care about our future and therefore can consider a pure
3492 * runnable sum, runqueues can NOT do this.
3494 * We specifically want runqueues to have a load_avg that includes
3495 * historical weights. Those represent the blocked load, the load we expect
3496 * to (shortly) return to us. This only works by keeping the weights as
3497 * integral part of the sum. We therefore cannot decompose as per (3).
3499 * Another reason this doesn't work is that runnable isn't a 0-sum entity.
3500 * Imagine a rq with 2 tasks that each are runnable 2/3 of the time. Then the
3501 * rq itself is runnable anywhere between 2/3 and 1 depending on how the
3502 * runnable section of these tasks overlap (or not). If they were to perfectly
3503 * align the rq as a whole would be runnable 2/3 of the time. If however we
3504 * always have at least 1 runnable task, the rq as a whole is always runnable.
3506 * So we'll have to approximate.. :/
3508 * Given the constraint:
3510 * ge->avg.running_sum <= ge->avg.runnable_sum <= LOAD_AVG_MAX
3512 * We can construct a rule that adds runnable to a rq by assuming minimal
3515 * On removal, we'll assume each task is equally runnable; which yields:
3517 * grq->avg.runnable_sum = grq->avg.load_sum / grq->load.weight
3519 * XXX: only do this for the part of runnable > running ?
3524 update_tg_cfs_util(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq)
3526 long delta = gcfs_rq->avg.util_avg - se->avg.util_avg;
3528 /* Nothing to update */
3533 * The relation between sum and avg is:
3535 * LOAD_AVG_MAX - 1024 + sa->period_contrib
3537 * however, the PELT windows are not aligned between grq and gse.
3540 /* Set new sched_entity's utilization */
3541 se->avg.util_avg = gcfs_rq->avg.util_avg;
3542 se->avg.util_sum = se->avg.util_avg * LOAD_AVG_MAX;
3544 /* Update parent cfs_rq utilization */
3545 add_positive(&cfs_rq->avg.util_avg, delta);
3546 cfs_rq->avg.util_sum = cfs_rq->avg.util_avg * LOAD_AVG_MAX;
3550 update_tg_cfs_runnable(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq)
3552 long delta_avg, running_sum, runnable_sum = gcfs_rq->prop_runnable_sum;
3553 unsigned long runnable_load_avg, load_avg;
3554 u64 runnable_load_sum, load_sum = 0;
3560 gcfs_rq->prop_runnable_sum = 0;
3562 if (runnable_sum >= 0) {
3564 * Add runnable; clip at LOAD_AVG_MAX. Reflects that until
3565 * the CPU is saturated running == runnable.
3567 runnable_sum += se->avg.load_sum;
3568 runnable_sum = min(runnable_sum, (long)LOAD_AVG_MAX);
3571 * Estimate the new unweighted runnable_sum of the gcfs_rq by
3572 * assuming all tasks are equally runnable.
3574 if (scale_load_down(gcfs_rq->load.weight)) {
3575 load_sum = div_s64(gcfs_rq->avg.load_sum,
3576 scale_load_down(gcfs_rq->load.weight));
3579 /* But make sure to not inflate se's runnable */
3580 runnable_sum = min(se->avg.load_sum, load_sum);
3584 * runnable_sum can't be lower than running_sum
3585 * As running sum is scale with CPU capacity wehreas the runnable sum
3586 * is not we rescale running_sum 1st
3588 running_sum = se->avg.util_sum /
3589 arch_scale_cpu_capacity(NULL, cpu_of(rq_of(cfs_rq)));
3590 runnable_sum = max(runnable_sum, running_sum);
3592 load_sum = (s64)se_weight(se) * runnable_sum;
3593 load_avg = div_s64(load_sum, LOAD_AVG_MAX);
3595 delta_sum = load_sum - (s64)se_weight(se) * se->avg.load_sum;
3596 delta_avg = load_avg - se->avg.load_avg;
3598 se->avg.load_sum = runnable_sum;
3599 se->avg.load_avg = load_avg;
3600 add_positive(&cfs_rq->avg.load_avg, delta_avg);
3601 add_positive(&cfs_rq->avg.load_sum, delta_sum);
3603 runnable_load_sum = (s64)se_runnable(se) * runnable_sum;
3604 runnable_load_avg = div_s64(runnable_load_sum, LOAD_AVG_MAX);
3605 delta_sum = runnable_load_sum - se_weight(se) * se->avg.runnable_load_sum;
3606 delta_avg = runnable_load_avg - se->avg.runnable_load_avg;
3608 se->avg.runnable_load_sum = runnable_sum;
3609 se->avg.runnable_load_avg = runnable_load_avg;
3612 add_positive(&cfs_rq->avg.runnable_load_avg, delta_avg);
3613 add_positive(&cfs_rq->avg.runnable_load_sum, delta_sum);
3617 static inline void add_tg_cfs_propagate(struct cfs_rq *cfs_rq, long runnable_sum)
3619 cfs_rq->propagate = 1;
3620 cfs_rq->prop_runnable_sum += runnable_sum;
3623 /* Update task and its cfs_rq load average */
3624 static inline int propagate_entity_load_avg(struct sched_entity *se)
3626 struct cfs_rq *cfs_rq, *gcfs_rq;
3628 if (entity_is_task(se))
3631 gcfs_rq = group_cfs_rq(se);
3632 if (!gcfs_rq->propagate)
3635 gcfs_rq->propagate = 0;
3637 cfs_rq = cfs_rq_of(se);
3639 add_tg_cfs_propagate(cfs_rq, gcfs_rq->prop_runnable_sum);
3641 update_tg_cfs_util(cfs_rq, se, gcfs_rq);
3642 update_tg_cfs_runnable(cfs_rq, se, gcfs_rq);
3648 * Check if we need to update the load and the utilization of a blocked
3651 static inline bool skip_blocked_update(struct sched_entity *se)
3653 struct cfs_rq *gcfs_rq = group_cfs_rq(se);
3656 * If sched_entity still have not zero load or utilization, we have to
3659 if (se->avg.load_avg || se->avg.util_avg)
3663 * If there is a pending propagation, we have to update the load and
3664 * the utilization of the sched_entity:
3666 if (gcfs_rq->propagate)
3670 * Otherwise, the load and the utilization of the sched_entity is
3671 * already zero and there is no pending propagation, so it will be a
3672 * waste of time to try to decay it:
3677 #else /* CONFIG_FAIR_GROUP_SCHED */
3679 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force) {}
3681 static inline int propagate_entity_load_avg(struct sched_entity *se)
3686 static inline void add_tg_cfs_propagate(struct cfs_rq *cfs_rq, long runnable_sum) {}
3688 #endif /* CONFIG_FAIR_GROUP_SCHED */
3691 * update_cfs_rq_load_avg - update the cfs_rq's load/util averages
3692 * @now: current time, as per cfs_rq_clock_task()
3693 * @cfs_rq: cfs_rq to update
3695 * The cfs_rq avg is the direct sum of all its entities (blocked and runnable)
3696 * avg. The immediate corollary is that all (fair) tasks must be attached, see
3697 * post_init_entity_util_avg().
3699 * cfs_rq->avg is used for task_h_load() and update_cfs_share() for example.
3701 * Returns true if the load decayed or we removed load.
3703 * Since both these conditions indicate a changed cfs_rq->avg.load we should
3704 * call update_tg_load_avg() when this function returns true.
3707 update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq)
3709 unsigned long removed_load = 0, removed_util = 0, removed_runnable_sum = 0;
3710 struct sched_avg *sa = &cfs_rq->avg;
3713 if (cfs_rq->removed.nr) {
3715 u32 divider = LOAD_AVG_MAX - 1024 + sa->period_contrib;
3717 raw_spin_lock(&cfs_rq->removed.lock);
3718 swap(cfs_rq->removed.util_avg, removed_util);
3719 swap(cfs_rq->removed.load_avg, removed_load);
3720 swap(cfs_rq->removed.runnable_sum, removed_runnable_sum);
3721 cfs_rq->removed.nr = 0;
3722 raw_spin_unlock(&cfs_rq->removed.lock);
3725 sub_positive(&sa->load_avg, r);
3726 sub_positive(&sa->load_sum, r * divider);
3729 sub_positive(&sa->util_avg, r);
3730 sub_positive(&sa->util_sum, r * divider);
3732 add_tg_cfs_propagate(cfs_rq, -(long)removed_runnable_sum);
3737 decayed |= __update_load_avg_cfs_rq(now, cpu_of(rq_of(cfs_rq)), cfs_rq);
3739 #ifndef CONFIG_64BIT
3741 cfs_rq->load_last_update_time_copy = sa->last_update_time;
3745 cfs_rq_util_change(cfs_rq, 0);
3751 * attach_entity_load_avg - attach this entity to its cfs_rq load avg
3752 * @cfs_rq: cfs_rq to attach to
3753 * @se: sched_entity to attach
3755 * Must call update_cfs_rq_load_avg() before this, since we rely on
3756 * cfs_rq->avg.last_update_time being current.
3758 static void attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3760 u32 divider = LOAD_AVG_MAX - 1024 + cfs_rq->avg.period_contrib;
3763 * When we attach the @se to the @cfs_rq, we must align the decay
3764 * window because without that, really weird and wonderful things can
3769 se->avg.last_update_time = cfs_rq->avg.last_update_time;
3770 se->avg.period_contrib = cfs_rq->avg.period_contrib;
3773 * Hell(o) Nasty stuff.. we need to recompute _sum based on the new
3774 * period_contrib. This isn't strictly correct, but since we're
3775 * entirely outside of the PELT hierarchy, nobody cares if we truncate
3778 se->avg.util_sum = se->avg.util_avg * divider;
3780 se->avg.load_sum = divider;
3781 if (se_weight(se)) {
3783 div_u64(se->avg.load_avg * se->avg.load_sum, se_weight(se));
3786 se->avg.runnable_load_sum = se->avg.load_sum;
3788 enqueue_load_avg(cfs_rq, se);
3789 cfs_rq->avg.util_avg += se->avg.util_avg;
3790 cfs_rq->avg.util_sum += se->avg.util_sum;
3792 add_tg_cfs_propagate(cfs_rq, se->avg.load_sum);
3794 cfs_rq_util_change(cfs_rq, flags);
3798 * detach_entity_load_avg - detach this entity from its cfs_rq load avg
3799 * @cfs_rq: cfs_rq to detach from
3800 * @se: sched_entity to detach
3802 * Must call update_cfs_rq_load_avg() before this, since we rely on
3803 * cfs_rq->avg.last_update_time being current.
3805 static void detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3807 dequeue_load_avg(cfs_rq, se);
3808 sub_positive(&cfs_rq->avg.util_avg, se->avg.util_avg);
3809 sub_positive(&cfs_rq->avg.util_sum, se->avg.util_sum);
3811 add_tg_cfs_propagate(cfs_rq, -se->avg.load_sum);
3813 cfs_rq_util_change(cfs_rq, 0);
3817 * Optional action to be done while updating the load average
3819 #define UPDATE_TG 0x1
3820 #define SKIP_AGE_LOAD 0x2
3821 #define DO_ATTACH 0x4
3823 /* Update task and its cfs_rq load average */
3824 static inline void update_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3826 u64 now = cfs_rq_clock_task(cfs_rq);
3827 struct rq *rq = rq_of(cfs_rq);
3828 int cpu = cpu_of(rq);
3832 * Track task load average for carrying it to new CPU after migrated, and
3833 * track group sched_entity load average for task_h_load calc in migration
3835 if (se->avg.last_update_time && !(flags & SKIP_AGE_LOAD))
3836 __update_load_avg_se(now, cpu, cfs_rq, se);
3838 decayed = update_cfs_rq_load_avg(now, cfs_rq);
3839 decayed |= propagate_entity_load_avg(se);
3841 if (!se->avg.last_update_time && (flags & DO_ATTACH)) {
3844 * DO_ATTACH means we're here from enqueue_entity().
3845 * !last_update_time means we've passed through
3846 * migrate_task_rq_fair() indicating we migrated.
3848 * IOW we're enqueueing a task on a new CPU.
3850 attach_entity_load_avg(cfs_rq, se, SCHED_CPUFREQ_MIGRATION);
3851 update_tg_load_avg(cfs_rq, 0);
3853 } else if (decayed && (flags & UPDATE_TG))
3854 update_tg_load_avg(cfs_rq, 0);
3857 #ifndef CONFIG_64BIT
3858 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
3860 u64 last_update_time_copy;
3861 u64 last_update_time;
3864 last_update_time_copy = cfs_rq->load_last_update_time_copy;
3866 last_update_time = cfs_rq->avg.last_update_time;
3867 } while (last_update_time != last_update_time_copy);
3869 return last_update_time;
3872 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
3874 return cfs_rq->avg.last_update_time;
3879 * Synchronize entity load avg of dequeued entity without locking
3882 void sync_entity_load_avg(struct sched_entity *se)
3884 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3885 u64 last_update_time;
3887 last_update_time = cfs_rq_last_update_time(cfs_rq);
3888 __update_load_avg_blocked_se(last_update_time, cpu_of(rq_of(cfs_rq)), se);
3892 * Task first catches up with cfs_rq, and then subtract
3893 * itself from the cfs_rq (task must be off the queue now).
3895 void remove_entity_load_avg(struct sched_entity *se)
3897 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3898 unsigned long flags;
3901 * tasks cannot exit without having gone through wake_up_new_task() ->
3902 * post_init_entity_util_avg() which will have added things to the
3903 * cfs_rq, so we can remove unconditionally.
3905 * Similarly for groups, they will have passed through
3906 * post_init_entity_util_avg() before unregister_sched_fair_group()
3910 sync_entity_load_avg(se);
3912 raw_spin_lock_irqsave(&cfs_rq->removed.lock, flags);
3913 ++cfs_rq->removed.nr;
3914 cfs_rq->removed.util_avg += se->avg.util_avg;
3915 cfs_rq->removed.load_avg += se->avg.load_avg;
3916 cfs_rq->removed.runnable_sum += se->avg.load_sum; /* == runnable_sum */
3917 raw_spin_unlock_irqrestore(&cfs_rq->removed.lock, flags);
3920 static inline unsigned long cfs_rq_runnable_load_avg(struct cfs_rq *cfs_rq)
3922 return cfs_rq->avg.runnable_load_avg;
3925 static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq)
3927 return cfs_rq->avg.load_avg;
3930 static int idle_balance(struct rq *this_rq, struct rq_flags *rf);
3932 static inline unsigned long task_util(struct task_struct *p)
3934 return READ_ONCE(p->se.avg.util_avg);
3937 static inline unsigned long _task_util_est(struct task_struct *p)
3939 struct util_est ue = READ_ONCE(p->se.avg.util_est);
3941 return max(ue.ewma, ue.enqueued);
3944 static inline unsigned long task_util_est(struct task_struct *p)
3946 return max(task_util(p), _task_util_est(p));
3949 static inline void util_est_enqueue(struct cfs_rq *cfs_rq,
3950 struct task_struct *p)
3952 unsigned int enqueued;
3954 if (!sched_feat(UTIL_EST))
3957 /* Update root cfs_rq's estimated utilization */
3958 enqueued = cfs_rq->avg.util_est.enqueued;
3959 enqueued += (_task_util_est(p) | UTIL_AVG_UNCHANGED);
3960 WRITE_ONCE(cfs_rq->avg.util_est.enqueued, enqueued);
3964 * Check if a (signed) value is within a specified (unsigned) margin,
3965 * based on the observation that:
3967 * abs(x) < y := (unsigned)(x + y - 1) < (2 * y - 1)
3969 * NOTE: this only works when value + maring < INT_MAX.
3971 static inline bool within_margin(int value, int margin)
3973 return ((unsigned int)(value + margin - 1) < (2 * margin - 1));
3977 util_est_dequeue(struct cfs_rq *cfs_rq, struct task_struct *p, bool task_sleep)
3979 long last_ewma_diff;
3982 if (!sched_feat(UTIL_EST))
3985 /* Update root cfs_rq's estimated utilization */
3986 ue.enqueued = cfs_rq->avg.util_est.enqueued;
3987 ue.enqueued -= min_t(unsigned int, ue.enqueued,
3988 (_task_util_est(p) | UTIL_AVG_UNCHANGED));
3989 WRITE_ONCE(cfs_rq->avg.util_est.enqueued, ue.enqueued);
3992 * Skip update of task's estimated utilization when the task has not
3993 * yet completed an activation, e.g. being migrated.
3999 * If the PELT values haven't changed since enqueue time,
4000 * skip the util_est update.
4002 ue = p->se.avg.util_est;
4003 if (ue.enqueued & UTIL_AVG_UNCHANGED)
4007 * Skip update of task's estimated utilization when its EWMA is
4008 * already ~1% close to its last activation value.
4010 ue.enqueued = (task_util(p) | UTIL_AVG_UNCHANGED);
4011 last_ewma_diff = ue.enqueued - ue.ewma;
4012 if (within_margin(last_ewma_diff, (SCHED_CAPACITY_SCALE / 100)))
4016 * Update Task's estimated utilization
4018 * When *p completes an activation we can consolidate another sample
4019 * of the task size. This is done by storing the current PELT value
4020 * as ue.enqueued and by using this value to update the Exponential
4021 * Weighted Moving Average (EWMA):
4023 * ewma(t) = w * task_util(p) + (1-w) * ewma(t-1)
4024 * = w * task_util(p) + ewma(t-1) - w * ewma(t-1)
4025 * = w * (task_util(p) - ewma(t-1)) + ewma(t-1)
4026 * = w * ( last_ewma_diff ) + ewma(t-1)
4027 * = w * (last_ewma_diff + ewma(t-1) / w)
4029 * Where 'w' is the weight of new samples, which is configured to be
4030 * 0.25, thus making w=1/4 ( >>= UTIL_EST_WEIGHT_SHIFT)
4032 ue.ewma <<= UTIL_EST_WEIGHT_SHIFT;
4033 ue.ewma += last_ewma_diff;
4034 ue.ewma >>= UTIL_EST_WEIGHT_SHIFT;
4035 WRITE_ONCE(p->se.avg.util_est, ue);
4038 #else /* CONFIG_SMP */
4041 update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq)
4046 #define UPDATE_TG 0x0
4047 #define SKIP_AGE_LOAD 0x0
4048 #define DO_ATTACH 0x0
4050 static inline void update_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int not_used1)
4052 cfs_rq_util_change(cfs_rq, 0);
4055 static inline void remove_entity_load_avg(struct sched_entity *se) {}
4058 attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) {}
4060 detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
4062 static inline int idle_balance(struct rq *rq, struct rq_flags *rf)
4068 util_est_enqueue(struct cfs_rq *cfs_rq, struct task_struct *p) {}
4071 util_est_dequeue(struct cfs_rq *cfs_rq, struct task_struct *p,
4074 #endif /* CONFIG_SMP */
4076 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
4078 #ifdef CONFIG_SCHED_DEBUG
4079 s64 d = se->vruntime - cfs_rq->min_vruntime;
4084 if (d > 3*sysctl_sched_latency)
4085 schedstat_inc(cfs_rq->nr_spread_over);
4090 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
4092 u64 vruntime = cfs_rq->min_vruntime;
4095 * The 'current' period is already promised to the current tasks,
4096 * however the extra weight of the new task will slow them down a
4097 * little, place the new task so that it fits in the slot that
4098 * stays open at the end.
4100 if (initial && sched_feat(START_DEBIT))
4101 vruntime += sched_vslice(cfs_rq, se);
4103 /* sleeps up to a single latency don't count. */
4105 unsigned long thresh = sysctl_sched_latency;
4108 * Halve their sleep time's effect, to allow
4109 * for a gentler effect of sleepers:
4111 if (sched_feat(GENTLE_FAIR_SLEEPERS))
4117 /* ensure we never gain time by being placed backwards. */
4118 se->vruntime = max_vruntime(se->vruntime, vruntime);
4121 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
4123 static inline void check_schedstat_required(void)
4125 #ifdef CONFIG_SCHEDSTATS
4126 if (schedstat_enabled())
4129 /* Force schedstat enabled if a dependent tracepoint is active */
4130 if (trace_sched_stat_wait_enabled() ||
4131 trace_sched_stat_sleep_enabled() ||
4132 trace_sched_stat_iowait_enabled() ||
4133 trace_sched_stat_blocked_enabled() ||
4134 trace_sched_stat_runtime_enabled()) {
4135 printk_deferred_once("Scheduler tracepoints stat_sleep, stat_iowait, "
4136 "stat_blocked and stat_runtime require the "
4137 "kernel parameter schedstats=enable or "
4138 "kernel.sched_schedstats=1\n");
4149 * update_min_vruntime()
4150 * vruntime -= min_vruntime
4154 * update_min_vruntime()
4155 * vruntime += min_vruntime
4157 * this way the vruntime transition between RQs is done when both
4158 * min_vruntime are up-to-date.
4162 * ->migrate_task_rq_fair() (p->state == TASK_WAKING)
4163 * vruntime -= min_vruntime
4167 * update_min_vruntime()
4168 * vruntime += min_vruntime
4170 * this way we don't have the most up-to-date min_vruntime on the originating
4171 * CPU and an up-to-date min_vruntime on the destination CPU.
4175 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
4177 bool renorm = !(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_MIGRATED);
4178 bool curr = cfs_rq->curr == se;
4181 * If we're the current task, we must renormalise before calling
4185 se->vruntime += cfs_rq->min_vruntime;
4187 update_curr(cfs_rq);
4190 * Otherwise, renormalise after, such that we're placed at the current
4191 * moment in time, instead of some random moment in the past. Being
4192 * placed in the past could significantly boost this task to the
4193 * fairness detriment of existing tasks.
4195 if (renorm && !curr)
4196 se->vruntime += cfs_rq->min_vruntime;
4199 * When enqueuing a sched_entity, we must:
4200 * - Update loads to have both entity and cfs_rq synced with now.
4201 * - Add its load to cfs_rq->runnable_avg
4202 * - For group_entity, update its weight to reflect the new share of
4204 * - Add its new weight to cfs_rq->load.weight
4206 update_load_avg(cfs_rq, se, UPDATE_TG | DO_ATTACH);
4207 update_cfs_group(se);
4208 enqueue_runnable_load_avg(cfs_rq, se);
4209 account_entity_enqueue(cfs_rq, se);
4211 if (flags & ENQUEUE_WAKEUP)
4212 place_entity(cfs_rq, se, 0);
4214 check_schedstat_required();
4215 update_stats_enqueue(cfs_rq, se, flags);
4216 check_spread(cfs_rq, se);
4218 __enqueue_entity(cfs_rq, se);
4221 if (cfs_rq->nr_running == 1) {
4222 list_add_leaf_cfs_rq(cfs_rq);
4223 check_enqueue_throttle(cfs_rq);
4227 static void __clear_buddies_last(struct sched_entity *se)
4229 for_each_sched_entity(se) {
4230 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4231 if (cfs_rq->last != se)
4234 cfs_rq->last = NULL;
4238 static void __clear_buddies_next(struct sched_entity *se)
4240 for_each_sched_entity(se) {
4241 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4242 if (cfs_rq->next != se)
4245 cfs_rq->next = NULL;
4249 static void __clear_buddies_skip(struct sched_entity *se)
4251 for_each_sched_entity(se) {
4252 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4253 if (cfs_rq->skip != se)
4256 cfs_rq->skip = NULL;
4260 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
4262 if (cfs_rq->last == se)
4263 __clear_buddies_last(se);
4265 if (cfs_rq->next == se)
4266 __clear_buddies_next(se);
4268 if (cfs_rq->skip == se)
4269 __clear_buddies_skip(se);
4272 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
4275 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
4278 * Update run-time statistics of the 'current'.
4280 update_curr(cfs_rq);
4283 * When dequeuing a sched_entity, we must:
4284 * - Update loads to have both entity and cfs_rq synced with now.
4285 * - Substract its load from the cfs_rq->runnable_avg.
4286 * - Substract its previous weight from cfs_rq->load.weight.
4287 * - For group entity, update its weight to reflect the new share
4288 * of its group cfs_rq.
4290 update_load_avg(cfs_rq, se, UPDATE_TG);
4291 dequeue_runnable_load_avg(cfs_rq, se);
4293 update_stats_dequeue(cfs_rq, se, flags);
4295 clear_buddies(cfs_rq, se);
4297 if (se != cfs_rq->curr)
4298 __dequeue_entity(cfs_rq, se);
4300 account_entity_dequeue(cfs_rq, se);
4303 * Normalize after update_curr(); which will also have moved
4304 * min_vruntime if @se is the one holding it back. But before doing
4305 * update_min_vruntime() again, which will discount @se's position and
4306 * can move min_vruntime forward still more.
4308 if (!(flags & DEQUEUE_SLEEP))
4309 se->vruntime -= cfs_rq->min_vruntime;
4311 /* return excess runtime on last dequeue */
4312 return_cfs_rq_runtime(cfs_rq);
4314 update_cfs_group(se);
4317 * Now advance min_vruntime if @se was the entity holding it back,
4318 * except when: DEQUEUE_SAVE && !DEQUEUE_MOVE, in this case we'll be
4319 * put back on, and if we advance min_vruntime, we'll be placed back
4320 * further than we started -- ie. we'll be penalized.
4322 if ((flags & (DEQUEUE_SAVE | DEQUEUE_MOVE)) == DEQUEUE_SAVE)
4323 update_min_vruntime(cfs_rq);
4327 * Preempt the current task with a newly woken task if needed:
4330 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
4332 unsigned long ideal_runtime, delta_exec;
4333 struct sched_entity *se;
4336 ideal_runtime = sched_slice(cfs_rq, curr);
4337 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
4338 if (delta_exec > ideal_runtime) {
4339 resched_curr(rq_of(cfs_rq));
4341 * The current task ran long enough, ensure it doesn't get
4342 * re-elected due to buddy favours.
4344 clear_buddies(cfs_rq, curr);
4349 * Ensure that a task that missed wakeup preemption by a
4350 * narrow margin doesn't have to wait for a full slice.
4351 * This also mitigates buddy induced latencies under load.
4353 if (delta_exec < sysctl_sched_min_granularity)
4356 se = __pick_first_entity(cfs_rq);
4357 delta = curr->vruntime - se->vruntime;
4362 if (delta > ideal_runtime)
4363 resched_curr(rq_of(cfs_rq));
4367 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
4369 /* 'current' is not kept within the tree. */
4372 * Any task has to be enqueued before it get to execute on
4373 * a CPU. So account for the time it spent waiting on the
4376 update_stats_wait_end(cfs_rq, se);
4377 __dequeue_entity(cfs_rq, se);
4378 update_load_avg(cfs_rq, se, UPDATE_TG);
4381 update_stats_curr_start(cfs_rq, se);
4385 * Track our maximum slice length, if the CPU's load is at
4386 * least twice that of our own weight (i.e. dont track it
4387 * when there are only lesser-weight tasks around):
4389 if (schedstat_enabled() && rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
4390 schedstat_set(se->statistics.slice_max,
4391 max((u64)schedstat_val(se->statistics.slice_max),
4392 se->sum_exec_runtime - se->prev_sum_exec_runtime));
4395 se->prev_sum_exec_runtime = se->sum_exec_runtime;
4399 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
4402 * Pick the next process, keeping these things in mind, in this order:
4403 * 1) keep things fair between processes/task groups
4404 * 2) pick the "next" process, since someone really wants that to run
4405 * 3) pick the "last" process, for cache locality
4406 * 4) do not run the "skip" process, if something else is available
4408 static struct sched_entity *
4409 pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr)
4411 struct sched_entity *left = __pick_first_entity(cfs_rq);
4412 struct sched_entity *se;
4415 * If curr is set we have to see if its left of the leftmost entity
4416 * still in the tree, provided there was anything in the tree at all.
4418 if (!left || (curr && entity_before(curr, left)))
4421 se = left; /* ideally we run the leftmost entity */
4424 * Avoid running the skip buddy, if running something else can
4425 * be done without getting too unfair.
4427 if (cfs_rq->skip == se) {
4428 struct sched_entity *second;
4431 second = __pick_first_entity(cfs_rq);
4433 second = __pick_next_entity(se);
4434 if (!second || (curr && entity_before(curr, second)))
4438 if (second && wakeup_preempt_entity(second, left) < 1)
4443 * Prefer last buddy, try to return the CPU to a preempted task.
4445 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
4449 * Someone really wants this to run. If it's not unfair, run it.
4451 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
4454 clear_buddies(cfs_rq, se);
4459 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
4461 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
4464 * If still on the runqueue then deactivate_task()
4465 * was not called and update_curr() has to be done:
4468 update_curr(cfs_rq);
4470 /* throttle cfs_rqs exceeding runtime */
4471 check_cfs_rq_runtime(cfs_rq);
4473 check_spread(cfs_rq, prev);
4476 update_stats_wait_start(cfs_rq, prev);
4477 /* Put 'current' back into the tree. */
4478 __enqueue_entity(cfs_rq, prev);
4479 /* in !on_rq case, update occurred at dequeue */
4480 update_load_avg(cfs_rq, prev, 0);
4482 cfs_rq->curr = NULL;
4486 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
4489 * Update run-time statistics of the 'current'.
4491 update_curr(cfs_rq);
4494 * Ensure that runnable average is periodically updated.
4496 update_load_avg(cfs_rq, curr, UPDATE_TG);
4497 update_cfs_group(curr);
4499 #ifdef CONFIG_SCHED_HRTICK
4501 * queued ticks are scheduled to match the slice, so don't bother
4502 * validating it and just reschedule.
4505 resched_curr(rq_of(cfs_rq));
4509 * don't let the period tick interfere with the hrtick preemption
4511 if (!sched_feat(DOUBLE_TICK) &&
4512 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
4516 if (cfs_rq->nr_running > 1)
4517 check_preempt_tick(cfs_rq, curr);
4521 /**************************************************
4522 * CFS bandwidth control machinery
4525 #ifdef CONFIG_CFS_BANDWIDTH
4527 #ifdef HAVE_JUMP_LABEL
4528 static struct static_key __cfs_bandwidth_used;
4530 static inline bool cfs_bandwidth_used(void)
4532 return static_key_false(&__cfs_bandwidth_used);
4535 void cfs_bandwidth_usage_inc(void)
4537 static_key_slow_inc_cpuslocked(&__cfs_bandwidth_used);
4540 void cfs_bandwidth_usage_dec(void)
4542 static_key_slow_dec_cpuslocked(&__cfs_bandwidth_used);
4544 #else /* HAVE_JUMP_LABEL */
4545 static bool cfs_bandwidth_used(void)
4550 void cfs_bandwidth_usage_inc(void) {}
4551 void cfs_bandwidth_usage_dec(void) {}
4552 #endif /* HAVE_JUMP_LABEL */
4555 * default period for cfs group bandwidth.
4556 * default: 0.1s, units: nanoseconds
4558 static inline u64 default_cfs_period(void)
4560 return 100000000ULL;
4563 static inline u64 sched_cfs_bandwidth_slice(void)
4565 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
4569 * Replenish runtime according to assigned quota and update expiration time.
4570 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
4571 * additional synchronization around rq->lock.
4573 * requires cfs_b->lock
4575 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
4579 if (cfs_b->quota == RUNTIME_INF)
4582 now = sched_clock_cpu(smp_processor_id());
4583 cfs_b->runtime = cfs_b->quota;
4584 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
4585 cfs_b->expires_seq++;
4588 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
4590 return &tg->cfs_bandwidth;
4593 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
4594 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
4596 if (unlikely(cfs_rq->throttle_count))
4597 return cfs_rq->throttled_clock_task - cfs_rq->throttled_clock_task_time;
4599 return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time;
4602 /* returns 0 on failure to allocate runtime */
4603 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4605 struct task_group *tg = cfs_rq->tg;
4606 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
4607 u64 amount = 0, min_amount, expires;
4610 /* note: this is a positive sum as runtime_remaining <= 0 */
4611 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
4613 raw_spin_lock(&cfs_b->lock);
4614 if (cfs_b->quota == RUNTIME_INF)
4615 amount = min_amount;
4617 start_cfs_bandwidth(cfs_b);
4619 if (cfs_b->runtime > 0) {
4620 amount = min(cfs_b->runtime, min_amount);
4621 cfs_b->runtime -= amount;
4625 expires_seq = cfs_b->expires_seq;
4626 expires = cfs_b->runtime_expires;
4627 raw_spin_unlock(&cfs_b->lock);
4629 cfs_rq->runtime_remaining += amount;
4631 * we may have advanced our local expiration to account for allowed
4632 * spread between our sched_clock and the one on which runtime was
4635 if (cfs_rq->expires_seq != expires_seq) {
4636 cfs_rq->expires_seq = expires_seq;
4637 cfs_rq->runtime_expires = expires;
4640 return cfs_rq->runtime_remaining > 0;
4644 * Note: This depends on the synchronization provided by sched_clock and the
4645 * fact that rq->clock snapshots this value.
4647 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4649 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4651 /* if the deadline is ahead of our clock, nothing to do */
4652 if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0))
4655 if (cfs_rq->runtime_remaining < 0)
4659 * If the local deadline has passed we have to consider the
4660 * possibility that our sched_clock is 'fast' and the global deadline
4661 * has not truly expired.
4663 * Fortunately we can check determine whether this the case by checking
4664 * whether the global deadline(cfs_b->expires_seq) has advanced.
4666 if (cfs_rq->expires_seq == cfs_b->expires_seq) {
4667 /* extend local deadline, drift is bounded above by 2 ticks */
4668 cfs_rq->runtime_expires += TICK_NSEC;
4670 /* global deadline is ahead, expiration has passed */
4671 cfs_rq->runtime_remaining = 0;
4675 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
4677 /* dock delta_exec before expiring quota (as it could span periods) */
4678 cfs_rq->runtime_remaining -= delta_exec;
4679 expire_cfs_rq_runtime(cfs_rq);
4681 if (likely(cfs_rq->runtime_remaining > 0))
4685 * if we're unable to extend our runtime we resched so that the active
4686 * hierarchy can be throttled
4688 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
4689 resched_curr(rq_of(cfs_rq));
4692 static __always_inline
4693 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
4695 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
4698 __account_cfs_rq_runtime(cfs_rq, delta_exec);
4701 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
4703 return cfs_bandwidth_used() && cfs_rq->throttled;
4706 /* check whether cfs_rq, or any parent, is throttled */
4707 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
4709 return cfs_bandwidth_used() && cfs_rq->throttle_count;
4713 * Ensure that neither of the group entities corresponding to src_cpu or
4714 * dest_cpu are members of a throttled hierarchy when performing group
4715 * load-balance operations.
4717 static inline int throttled_lb_pair(struct task_group *tg,
4718 int src_cpu, int dest_cpu)
4720 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
4722 src_cfs_rq = tg->cfs_rq[src_cpu];
4723 dest_cfs_rq = tg->cfs_rq[dest_cpu];
4725 return throttled_hierarchy(src_cfs_rq) ||
4726 throttled_hierarchy(dest_cfs_rq);
4729 /* updated child weight may affect parent so we have to do this bottom up */
4730 static int tg_unthrottle_up(struct task_group *tg, void *data)
4732 struct rq *rq = data;
4733 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
4735 cfs_rq->throttle_count--;
4736 if (!cfs_rq->throttle_count) {
4737 /* adjust cfs_rq_clock_task() */
4738 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
4739 cfs_rq->throttled_clock_task;
4745 static int tg_throttle_down(struct task_group *tg, void *data)
4747 struct rq *rq = data;
4748 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
4750 /* group is entering throttled state, stop time */
4751 if (!cfs_rq->throttle_count)
4752 cfs_rq->throttled_clock_task = rq_clock_task(rq);
4753 cfs_rq->throttle_count++;
4758 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
4760 struct rq *rq = rq_of(cfs_rq);
4761 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4762 struct sched_entity *se;
4763 long task_delta, dequeue = 1;
4766 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
4768 /* freeze hierarchy runnable averages while throttled */
4770 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
4773 task_delta = cfs_rq->h_nr_running;
4774 for_each_sched_entity(se) {
4775 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
4776 /* throttled entity or throttle-on-deactivate */
4781 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
4782 qcfs_rq->h_nr_running -= task_delta;
4784 if (qcfs_rq->load.weight)
4789 sub_nr_running(rq, task_delta);
4791 cfs_rq->throttled = 1;
4792 cfs_rq->throttled_clock = rq_clock(rq);
4793 raw_spin_lock(&cfs_b->lock);
4794 empty = list_empty(&cfs_b->throttled_cfs_rq);
4797 * Add to the _head_ of the list, so that an already-started
4798 * distribute_cfs_runtime will not see us
4800 list_add_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
4803 * If we're the first throttled task, make sure the bandwidth
4807 start_cfs_bandwidth(cfs_b);
4809 raw_spin_unlock(&cfs_b->lock);
4812 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
4814 struct rq *rq = rq_of(cfs_rq);
4815 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4816 struct sched_entity *se;
4820 se = cfs_rq->tg->se[cpu_of(rq)];
4822 cfs_rq->throttled = 0;
4824 update_rq_clock(rq);
4826 raw_spin_lock(&cfs_b->lock);
4827 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
4828 list_del_rcu(&cfs_rq->throttled_list);
4829 raw_spin_unlock(&cfs_b->lock);
4831 /* update hierarchical throttle state */
4832 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
4834 if (!cfs_rq->load.weight)
4837 task_delta = cfs_rq->h_nr_running;
4838 for_each_sched_entity(se) {
4842 cfs_rq = cfs_rq_of(se);
4844 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
4845 cfs_rq->h_nr_running += task_delta;
4847 if (cfs_rq_throttled(cfs_rq))
4852 add_nr_running(rq, task_delta);
4854 /* Determine whether we need to wake up potentially idle CPU: */
4855 if (rq->curr == rq->idle && rq->cfs.nr_running)
4859 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
4860 u64 remaining, u64 expires)
4862 struct cfs_rq *cfs_rq;
4864 u64 starting_runtime = remaining;
4867 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
4869 struct rq *rq = rq_of(cfs_rq);
4873 if (!cfs_rq_throttled(cfs_rq))
4876 runtime = -cfs_rq->runtime_remaining + 1;
4877 if (runtime > remaining)
4878 runtime = remaining;
4879 remaining -= runtime;
4881 cfs_rq->runtime_remaining += runtime;
4882 cfs_rq->runtime_expires = expires;
4884 /* we check whether we're throttled above */
4885 if (cfs_rq->runtime_remaining > 0)
4886 unthrottle_cfs_rq(cfs_rq);
4896 return starting_runtime - remaining;
4900 * Responsible for refilling a task_group's bandwidth and unthrottling its
4901 * cfs_rqs as appropriate. If there has been no activity within the last
4902 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
4903 * used to track this state.
4905 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
4907 u64 runtime, runtime_expires;
4910 /* no need to continue the timer with no bandwidth constraint */
4911 if (cfs_b->quota == RUNTIME_INF)
4912 goto out_deactivate;
4914 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
4915 cfs_b->nr_periods += overrun;
4918 * idle depends on !throttled (for the case of a large deficit), and if
4919 * we're going inactive then everything else can be deferred
4921 if (cfs_b->idle && !throttled)
4922 goto out_deactivate;
4924 __refill_cfs_bandwidth_runtime(cfs_b);
4927 /* mark as potentially idle for the upcoming period */
4932 /* account preceding periods in which throttling occurred */
4933 cfs_b->nr_throttled += overrun;
4935 runtime_expires = cfs_b->runtime_expires;
4938 * This check is repeated as we are holding onto the new bandwidth while
4939 * we unthrottle. This can potentially race with an unthrottled group
4940 * trying to acquire new bandwidth from the global pool. This can result
4941 * in us over-using our runtime if it is all used during this loop, but
4942 * only by limited amounts in that extreme case.
4944 while (throttled && cfs_b->runtime > 0) {
4945 runtime = cfs_b->runtime;
4946 raw_spin_unlock(&cfs_b->lock);
4947 /* we can't nest cfs_b->lock while distributing bandwidth */
4948 runtime = distribute_cfs_runtime(cfs_b, runtime,
4950 raw_spin_lock(&cfs_b->lock);
4952 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
4954 cfs_b->runtime -= min(runtime, cfs_b->runtime);
4958 * While we are ensured activity in the period following an
4959 * unthrottle, this also covers the case in which the new bandwidth is
4960 * insufficient to cover the existing bandwidth deficit. (Forcing the
4961 * timer to remain active while there are any throttled entities.)
4971 /* a cfs_rq won't donate quota below this amount */
4972 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
4973 /* minimum remaining period time to redistribute slack quota */
4974 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
4975 /* how long we wait to gather additional slack before distributing */
4976 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
4979 * Are we near the end of the current quota period?
4981 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
4982 * hrtimer base being cleared by hrtimer_start. In the case of
4983 * migrate_hrtimers, base is never cleared, so we are fine.
4985 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
4987 struct hrtimer *refresh_timer = &cfs_b->period_timer;
4990 /* if the call-back is running a quota refresh is already occurring */
4991 if (hrtimer_callback_running(refresh_timer))
4994 /* is a quota refresh about to occur? */
4995 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
4996 if (remaining < min_expire)
5002 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
5004 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
5006 /* if there's a quota refresh soon don't bother with slack */
5007 if (runtime_refresh_within(cfs_b, min_left))
5010 hrtimer_start(&cfs_b->slack_timer,
5011 ns_to_ktime(cfs_bandwidth_slack_period),
5015 /* we know any runtime found here is valid as update_curr() precedes return */
5016 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
5018 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
5019 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
5021 if (slack_runtime <= 0)
5024 raw_spin_lock(&cfs_b->lock);
5025 if (cfs_b->quota != RUNTIME_INF &&
5026 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
5027 cfs_b->runtime += slack_runtime;
5029 /* we are under rq->lock, defer unthrottling using a timer */
5030 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
5031 !list_empty(&cfs_b->throttled_cfs_rq))
5032 start_cfs_slack_bandwidth(cfs_b);
5034 raw_spin_unlock(&cfs_b->lock);
5036 /* even if it's not valid for return we don't want to try again */
5037 cfs_rq->runtime_remaining -= slack_runtime;
5040 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
5042 if (!cfs_bandwidth_used())
5045 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
5048 __return_cfs_rq_runtime(cfs_rq);
5052 * This is done with a timer (instead of inline with bandwidth return) since
5053 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
5055 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
5057 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
5060 /* confirm we're still not at a refresh boundary */
5061 raw_spin_lock(&cfs_b->lock);
5062 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
5063 raw_spin_unlock(&cfs_b->lock);
5067 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice)
5068 runtime = cfs_b->runtime;
5070 expires = cfs_b->runtime_expires;
5071 raw_spin_unlock(&cfs_b->lock);
5076 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
5078 raw_spin_lock(&cfs_b->lock);
5079 if (expires == cfs_b->runtime_expires)
5080 cfs_b->runtime -= min(runtime, cfs_b->runtime);
5081 raw_spin_unlock(&cfs_b->lock);
5085 * When a group wakes up we want to make sure that its quota is not already
5086 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
5087 * runtime as update_curr() throttling can not not trigger until it's on-rq.
5089 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
5091 if (!cfs_bandwidth_used())
5094 /* an active group must be handled by the update_curr()->put() path */
5095 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
5098 /* ensure the group is not already throttled */
5099 if (cfs_rq_throttled(cfs_rq))
5102 /* update runtime allocation */
5103 account_cfs_rq_runtime(cfs_rq, 0);
5104 if (cfs_rq->runtime_remaining <= 0)
5105 throttle_cfs_rq(cfs_rq);
5108 static void sync_throttle(struct task_group *tg, int cpu)
5110 struct cfs_rq *pcfs_rq, *cfs_rq;
5112 if (!cfs_bandwidth_used())
5118 cfs_rq = tg->cfs_rq[cpu];
5119 pcfs_rq = tg->parent->cfs_rq[cpu];
5121 cfs_rq->throttle_count = pcfs_rq->throttle_count;
5122 cfs_rq->throttled_clock_task = rq_clock_task(cpu_rq(cpu));
5125 /* conditionally throttle active cfs_rq's from put_prev_entity() */
5126 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
5128 if (!cfs_bandwidth_used())
5131 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
5135 * it's possible for a throttled entity to be forced into a running
5136 * state (e.g. set_curr_task), in this case we're finished.
5138 if (cfs_rq_throttled(cfs_rq))
5141 throttle_cfs_rq(cfs_rq);
5145 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
5147 struct cfs_bandwidth *cfs_b =
5148 container_of(timer, struct cfs_bandwidth, slack_timer);
5150 do_sched_cfs_slack_timer(cfs_b);
5152 return HRTIMER_NORESTART;
5155 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
5157 struct cfs_bandwidth *cfs_b =
5158 container_of(timer, struct cfs_bandwidth, period_timer);
5162 raw_spin_lock(&cfs_b->lock);
5164 overrun = hrtimer_forward_now(timer, cfs_b->period);
5168 idle = do_sched_cfs_period_timer(cfs_b, overrun);
5171 cfs_b->period_active = 0;
5172 raw_spin_unlock(&cfs_b->lock);
5174 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
5177 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
5179 raw_spin_lock_init(&cfs_b->lock);
5181 cfs_b->quota = RUNTIME_INF;
5182 cfs_b->period = ns_to_ktime(default_cfs_period());
5184 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
5185 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED);
5186 cfs_b->period_timer.function = sched_cfs_period_timer;
5187 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
5188 cfs_b->slack_timer.function = sched_cfs_slack_timer;
5191 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
5193 cfs_rq->runtime_enabled = 0;
5194 INIT_LIST_HEAD(&cfs_rq->throttled_list);
5197 void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
5201 lockdep_assert_held(&cfs_b->lock);
5203 if (cfs_b->period_active)
5206 cfs_b->period_active = 1;
5207 overrun = hrtimer_forward_now(&cfs_b->period_timer, cfs_b->period);
5208 cfs_b->runtime_expires += (overrun + 1) * ktime_to_ns(cfs_b->period);
5209 cfs_b->expires_seq++;
5210 hrtimer_start_expires(&cfs_b->period_timer, HRTIMER_MODE_ABS_PINNED);
5213 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
5215 /* init_cfs_bandwidth() was not called */
5216 if (!cfs_b->throttled_cfs_rq.next)
5219 hrtimer_cancel(&cfs_b->period_timer);
5220 hrtimer_cancel(&cfs_b->slack_timer);
5224 * Both these CPU hotplug callbacks race against unregister_fair_sched_group()
5226 * The race is harmless, since modifying bandwidth settings of unhooked group
5227 * bits doesn't do much.
5230 /* cpu online calback */
5231 static void __maybe_unused update_runtime_enabled(struct rq *rq)
5233 struct task_group *tg;
5235 lockdep_assert_held(&rq->lock);
5238 list_for_each_entry_rcu(tg, &task_groups, list) {
5239 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
5240 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
5242 raw_spin_lock(&cfs_b->lock);
5243 cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF;
5244 raw_spin_unlock(&cfs_b->lock);
5249 /* cpu offline callback */
5250 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
5252 struct task_group *tg;
5254 lockdep_assert_held(&rq->lock);
5257 list_for_each_entry_rcu(tg, &task_groups, list) {
5258 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
5260 if (!cfs_rq->runtime_enabled)
5264 * clock_task is not advancing so we just need to make sure
5265 * there's some valid quota amount
5267 cfs_rq->runtime_remaining = 1;
5269 * Offline rq is schedulable till CPU is completely disabled
5270 * in take_cpu_down(), so we prevent new cfs throttling here.
5272 cfs_rq->runtime_enabled = 0;
5274 if (cfs_rq_throttled(cfs_rq))
5275 unthrottle_cfs_rq(cfs_rq);
5280 #else /* CONFIG_CFS_BANDWIDTH */
5281 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
5283 return rq_clock_task(rq_of(cfs_rq));
5286 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {}
5287 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; }
5288 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
5289 static inline void sync_throttle(struct task_group *tg, int cpu) {}
5290 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
5292 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
5297 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
5302 static inline int throttled_lb_pair(struct task_group *tg,
5303 int src_cpu, int dest_cpu)
5308 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
5310 #ifdef CONFIG_FAIR_GROUP_SCHED
5311 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
5314 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
5318 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
5319 static inline void update_runtime_enabled(struct rq *rq) {}
5320 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
5322 #endif /* CONFIG_CFS_BANDWIDTH */
5324 /**************************************************
5325 * CFS operations on tasks:
5328 #ifdef CONFIG_SCHED_HRTICK
5329 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
5331 struct sched_entity *se = &p->se;
5332 struct cfs_rq *cfs_rq = cfs_rq_of(se);
5334 SCHED_WARN_ON(task_rq(p) != rq);
5336 if (rq->cfs.h_nr_running > 1) {
5337 u64 slice = sched_slice(cfs_rq, se);
5338 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
5339 s64 delta = slice - ran;
5346 hrtick_start(rq, delta);
5351 * called from enqueue/dequeue and updates the hrtick when the
5352 * current task is from our class and nr_running is low enough
5355 static void hrtick_update(struct rq *rq)
5357 struct task_struct *curr = rq->curr;
5359 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
5362 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
5363 hrtick_start_fair(rq, curr);
5365 #else /* !CONFIG_SCHED_HRTICK */
5367 hrtick_start_fair(struct rq *rq, struct task_struct *p)
5371 static inline void hrtick_update(struct rq *rq)
5377 * The enqueue_task method is called before nr_running is
5378 * increased. Here we update the fair scheduling stats and
5379 * then put the task into the rbtree:
5382 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
5384 struct cfs_rq *cfs_rq;
5385 struct sched_entity *se = &p->se;
5388 * The code below (indirectly) updates schedutil which looks at
5389 * the cfs_rq utilization to select a frequency.
5390 * Let's add the task's estimated utilization to the cfs_rq's
5391 * estimated utilization, before we update schedutil.
5393 util_est_enqueue(&rq->cfs, p);
5396 * If in_iowait is set, the code below may not trigger any cpufreq
5397 * utilization updates, so do it here explicitly with the IOWAIT flag
5401 cpufreq_update_util(rq, SCHED_CPUFREQ_IOWAIT);
5403 for_each_sched_entity(se) {
5406 cfs_rq = cfs_rq_of(se);
5407 enqueue_entity(cfs_rq, se, flags);
5410 * end evaluation on encountering a throttled cfs_rq
5412 * note: in the case of encountering a throttled cfs_rq we will
5413 * post the final h_nr_running increment below.
5415 if (cfs_rq_throttled(cfs_rq))
5417 cfs_rq->h_nr_running++;
5419 flags = ENQUEUE_WAKEUP;
5422 for_each_sched_entity(se) {
5423 cfs_rq = cfs_rq_of(se);
5424 cfs_rq->h_nr_running++;
5426 if (cfs_rq_throttled(cfs_rq))
5429 update_load_avg(cfs_rq, se, UPDATE_TG);
5430 update_cfs_group(se);
5434 add_nr_running(rq, 1);
5439 static void set_next_buddy(struct sched_entity *se);
5442 * The dequeue_task method is called before nr_running is
5443 * decreased. We remove the task from the rbtree and
5444 * update the fair scheduling stats:
5446 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
5448 struct cfs_rq *cfs_rq;
5449 struct sched_entity *se = &p->se;
5450 int task_sleep = flags & DEQUEUE_SLEEP;
5452 for_each_sched_entity(se) {
5453 cfs_rq = cfs_rq_of(se);
5454 dequeue_entity(cfs_rq, se, flags);
5457 * end evaluation on encountering a throttled cfs_rq
5459 * note: in the case of encountering a throttled cfs_rq we will
5460 * post the final h_nr_running decrement below.
5462 if (cfs_rq_throttled(cfs_rq))
5464 cfs_rq->h_nr_running--;
5466 /* Don't dequeue parent if it has other entities besides us */
5467 if (cfs_rq->load.weight) {
5468 /* Avoid re-evaluating load for this entity: */
5469 se = parent_entity(se);
5471 * Bias pick_next to pick a task from this cfs_rq, as
5472 * p is sleeping when it is within its sched_slice.
5474 if (task_sleep && se && !throttled_hierarchy(cfs_rq))
5478 flags |= DEQUEUE_SLEEP;
5481 for_each_sched_entity(se) {
5482 cfs_rq = cfs_rq_of(se);
5483 cfs_rq->h_nr_running--;
5485 if (cfs_rq_throttled(cfs_rq))
5488 update_load_avg(cfs_rq, se, UPDATE_TG);
5489 update_cfs_group(se);
5493 sub_nr_running(rq, 1);
5495 util_est_dequeue(&rq->cfs, p, task_sleep);
5501 /* Working cpumask for: load_balance, load_balance_newidle. */
5502 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
5503 DEFINE_PER_CPU(cpumask_var_t, select_idle_mask);
5505 #ifdef CONFIG_NO_HZ_COMMON
5507 * per rq 'load' arrray crap; XXX kill this.
5511 * The exact cpuload calculated at every tick would be:
5513 * load' = (1 - 1/2^i) * load + (1/2^i) * cur_load
5515 * If a CPU misses updates for n ticks (as it was idle) and update gets
5516 * called on the n+1-th tick when CPU may be busy, then we have:
5518 * load_n = (1 - 1/2^i)^n * load_0
5519 * load_n+1 = (1 - 1/2^i) * load_n + (1/2^i) * cur_load
5521 * decay_load_missed() below does efficient calculation of
5523 * load' = (1 - 1/2^i)^n * load
5525 * Because x^(n+m) := x^n * x^m we can decompose any x^n in power-of-2 factors.
5526 * This allows us to precompute the above in said factors, thereby allowing the
5527 * reduction of an arbitrary n in O(log_2 n) steps. (See also
5528 * fixed_power_int())
5530 * The calculation is approximated on a 128 point scale.
5532 #define DEGRADE_SHIFT 7
5534 static const u8 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
5535 static const u8 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
5536 { 0, 0, 0, 0, 0, 0, 0, 0 },
5537 { 64, 32, 8, 0, 0, 0, 0, 0 },
5538 { 96, 72, 40, 12, 1, 0, 0, 0 },
5539 { 112, 98, 75, 43, 15, 1, 0, 0 },
5540 { 120, 112, 98, 76, 45, 16, 2, 0 }
5544 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
5545 * would be when CPU is idle and so we just decay the old load without
5546 * adding any new load.
5548 static unsigned long
5549 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
5553 if (!missed_updates)
5556 if (missed_updates >= degrade_zero_ticks[idx])
5560 return load >> missed_updates;
5562 while (missed_updates) {
5563 if (missed_updates % 2)
5564 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
5566 missed_updates >>= 1;
5573 cpumask_var_t idle_cpus_mask;
5575 int has_blocked; /* Idle CPUS has blocked load */
5576 unsigned long next_balance; /* in jiffy units */
5577 unsigned long next_blocked; /* Next update of blocked load in jiffies */
5578 } nohz ____cacheline_aligned;
5580 #endif /* CONFIG_NO_HZ_COMMON */
5583 * __cpu_load_update - update the rq->cpu_load[] statistics
5584 * @this_rq: The rq to update statistics for
5585 * @this_load: The current load
5586 * @pending_updates: The number of missed updates
5588 * Update rq->cpu_load[] statistics. This function is usually called every
5589 * scheduler tick (TICK_NSEC).
5591 * This function computes a decaying average:
5593 * load[i]' = (1 - 1/2^i) * load[i] + (1/2^i) * load
5595 * Because of NOHZ it might not get called on every tick which gives need for
5596 * the @pending_updates argument.
5598 * load[i]_n = (1 - 1/2^i) * load[i]_n-1 + (1/2^i) * load_n-1
5599 * = A * load[i]_n-1 + B ; A := (1 - 1/2^i), B := (1/2^i) * load
5600 * = A * (A * load[i]_n-2 + B) + B
5601 * = A * (A * (A * load[i]_n-3 + B) + B) + B
5602 * = A^3 * load[i]_n-3 + (A^2 + A + 1) * B
5603 * = A^n * load[i]_0 + (A^(n-1) + A^(n-2) + ... + 1) * B
5604 * = A^n * load[i]_0 + ((1 - A^n) / (1 - A)) * B
5605 * = (1 - 1/2^i)^n * (load[i]_0 - load) + load
5607 * In the above we've assumed load_n := load, which is true for NOHZ_FULL as
5608 * any change in load would have resulted in the tick being turned back on.
5610 * For regular NOHZ, this reduces to:
5612 * load[i]_n = (1 - 1/2^i)^n * load[i]_0
5614 * see decay_load_misses(). For NOHZ_FULL we get to subtract and add the extra
5617 static void cpu_load_update(struct rq *this_rq, unsigned long this_load,
5618 unsigned long pending_updates)
5620 unsigned long __maybe_unused tickless_load = this_rq->cpu_load[0];
5623 this_rq->nr_load_updates++;
5625 /* Update our load: */
5626 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
5627 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
5628 unsigned long old_load, new_load;
5630 /* scale is effectively 1 << i now, and >> i divides by scale */
5632 old_load = this_rq->cpu_load[i];
5633 #ifdef CONFIG_NO_HZ_COMMON
5634 old_load = decay_load_missed(old_load, pending_updates - 1, i);
5635 if (tickless_load) {
5636 old_load -= decay_load_missed(tickless_load, pending_updates - 1, i);
5638 * old_load can never be a negative value because a
5639 * decayed tickless_load cannot be greater than the
5640 * original tickless_load.
5642 old_load += tickless_load;
5645 new_load = this_load;
5647 * Round up the averaging division if load is increasing. This
5648 * prevents us from getting stuck on 9 if the load is 10, for
5651 if (new_load > old_load)
5652 new_load += scale - 1;
5654 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
5657 sched_avg_update(this_rq);
5660 /* Used instead of source_load when we know the type == 0 */
5661 static unsigned long weighted_cpuload(struct rq *rq)
5663 return cfs_rq_runnable_load_avg(&rq->cfs);
5666 #ifdef CONFIG_NO_HZ_COMMON
5668 * There is no sane way to deal with nohz on smp when using jiffies because the
5669 * CPU doing the jiffies update might drift wrt the CPU doing the jiffy reading
5670 * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}.
5672 * Therefore we need to avoid the delta approach from the regular tick when
5673 * possible since that would seriously skew the load calculation. This is why we
5674 * use cpu_load_update_periodic() for CPUs out of nohz. However we'll rely on
5675 * jiffies deltas for updates happening while in nohz mode (idle ticks, idle
5676 * loop exit, nohz_idle_balance, nohz full exit...)
5678 * This means we might still be one tick off for nohz periods.
5681 static void cpu_load_update_nohz(struct rq *this_rq,
5682 unsigned long curr_jiffies,
5685 unsigned long pending_updates;
5687 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
5688 if (pending_updates) {
5689 this_rq->last_load_update_tick = curr_jiffies;
5691 * In the regular NOHZ case, we were idle, this means load 0.
5692 * In the NOHZ_FULL case, we were non-idle, we should consider
5693 * its weighted load.
5695 cpu_load_update(this_rq, load, pending_updates);
5700 * Called from nohz_idle_balance() to update the load ratings before doing the
5703 static void cpu_load_update_idle(struct rq *this_rq)
5706 * bail if there's load or we're actually up-to-date.
5708 if (weighted_cpuload(this_rq))
5711 cpu_load_update_nohz(this_rq, READ_ONCE(jiffies), 0);
5715 * Record CPU load on nohz entry so we know the tickless load to account
5716 * on nohz exit. cpu_load[0] happens then to be updated more frequently
5717 * than other cpu_load[idx] but it should be fine as cpu_load readers
5718 * shouldn't rely into synchronized cpu_load[*] updates.
5720 void cpu_load_update_nohz_start(void)
5722 struct rq *this_rq = this_rq();
5725 * This is all lockless but should be fine. If weighted_cpuload changes
5726 * concurrently we'll exit nohz. And cpu_load write can race with
5727 * cpu_load_update_idle() but both updater would be writing the same.
5729 this_rq->cpu_load[0] = weighted_cpuload(this_rq);
5733 * Account the tickless load in the end of a nohz frame.
5735 void cpu_load_update_nohz_stop(void)
5737 unsigned long curr_jiffies = READ_ONCE(jiffies);
5738 struct rq *this_rq = this_rq();
5742 if (curr_jiffies == this_rq->last_load_update_tick)
5745 load = weighted_cpuload(this_rq);
5746 rq_lock(this_rq, &rf);
5747 update_rq_clock(this_rq);
5748 cpu_load_update_nohz(this_rq, curr_jiffies, load);
5749 rq_unlock(this_rq, &rf);
5751 #else /* !CONFIG_NO_HZ_COMMON */
5752 static inline void cpu_load_update_nohz(struct rq *this_rq,
5753 unsigned long curr_jiffies,
5754 unsigned long load) { }
5755 #endif /* CONFIG_NO_HZ_COMMON */
5757 static void cpu_load_update_periodic(struct rq *this_rq, unsigned long load)
5759 #ifdef CONFIG_NO_HZ_COMMON
5760 /* See the mess around cpu_load_update_nohz(). */
5761 this_rq->last_load_update_tick = READ_ONCE(jiffies);
5763 cpu_load_update(this_rq, load, 1);
5767 * Called from scheduler_tick()
5769 void cpu_load_update_active(struct rq *this_rq)
5771 unsigned long load = weighted_cpuload(this_rq);
5773 if (tick_nohz_tick_stopped())
5774 cpu_load_update_nohz(this_rq, READ_ONCE(jiffies), load);
5776 cpu_load_update_periodic(this_rq, load);
5780 * Return a low guess at the load of a migration-source CPU weighted
5781 * according to the scheduling class and "nice" value.
5783 * We want to under-estimate the load of migration sources, to
5784 * balance conservatively.
5786 static unsigned long source_load(int cpu, int type)
5788 struct rq *rq = cpu_rq(cpu);
5789 unsigned long total = weighted_cpuload(rq);
5791 if (type == 0 || !sched_feat(LB_BIAS))
5794 return min(rq->cpu_load[type-1], total);
5798 * Return a high guess at the load of a migration-target CPU weighted
5799 * according to the scheduling class and "nice" value.
5801 static unsigned long target_load(int cpu, int type)
5803 struct rq *rq = cpu_rq(cpu);
5804 unsigned long total = weighted_cpuload(rq);
5806 if (type == 0 || !sched_feat(LB_BIAS))
5809 return max(rq->cpu_load[type-1], total);
5812 static unsigned long capacity_of(int cpu)
5814 return cpu_rq(cpu)->cpu_capacity;
5817 static unsigned long capacity_orig_of(int cpu)
5819 return cpu_rq(cpu)->cpu_capacity_orig;
5822 static unsigned long cpu_avg_load_per_task(int cpu)
5824 struct rq *rq = cpu_rq(cpu);
5825 unsigned long nr_running = READ_ONCE(rq->cfs.h_nr_running);
5826 unsigned long load_avg = weighted_cpuload(rq);
5829 return load_avg / nr_running;
5834 static void record_wakee(struct task_struct *p)
5837 * Only decay a single time; tasks that have less then 1 wakeup per
5838 * jiffy will not have built up many flips.
5840 if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) {
5841 current->wakee_flips >>= 1;
5842 current->wakee_flip_decay_ts = jiffies;
5845 if (current->last_wakee != p) {
5846 current->last_wakee = p;
5847 current->wakee_flips++;
5852 * Detect M:N waker/wakee relationships via a switching-frequency heuristic.
5854 * A waker of many should wake a different task than the one last awakened
5855 * at a frequency roughly N times higher than one of its wakees.
5857 * In order to determine whether we should let the load spread vs consolidating
5858 * to shared cache, we look for a minimum 'flip' frequency of llc_size in one
5859 * partner, and a factor of lls_size higher frequency in the other.
5861 * With both conditions met, we can be relatively sure that the relationship is
5862 * non-monogamous, with partner count exceeding socket size.
5864 * Waker/wakee being client/server, worker/dispatcher, interrupt source or
5865 * whatever is irrelevant, spread criteria is apparent partner count exceeds
5868 static int wake_wide(struct task_struct *p)
5870 unsigned int master = current->wakee_flips;
5871 unsigned int slave = p->wakee_flips;
5872 int factor = this_cpu_read(sd_llc_size);
5875 swap(master, slave);
5876 if (slave < factor || master < slave * factor)
5882 * The purpose of wake_affine() is to quickly determine on which CPU we can run
5883 * soonest. For the purpose of speed we only consider the waking and previous
5886 * wake_affine_idle() - only considers 'now', it check if the waking CPU is
5887 * cache-affine and is (or will be) idle.
5889 * wake_affine_weight() - considers the weight to reflect the average
5890 * scheduling latency of the CPUs. This seems to work
5891 * for the overloaded case.
5894 wake_affine_idle(int this_cpu, int prev_cpu, int sync)
5897 * If this_cpu is idle, it implies the wakeup is from interrupt
5898 * context. Only allow the move if cache is shared. Otherwise an
5899 * interrupt intensive workload could force all tasks onto one
5900 * node depending on the IO topology or IRQ affinity settings.
5902 * If the prev_cpu is idle and cache affine then avoid a migration.
5903 * There is no guarantee that the cache hot data from an interrupt
5904 * is more important than cache hot data on the prev_cpu and from
5905 * a cpufreq perspective, it's better to have higher utilisation
5908 if (available_idle_cpu(this_cpu) && cpus_share_cache(this_cpu, prev_cpu))
5909 return available_idle_cpu(prev_cpu) ? prev_cpu : this_cpu;
5911 if (sync && cpu_rq(this_cpu)->nr_running == 1)
5914 return nr_cpumask_bits;
5918 wake_affine_weight(struct sched_domain *sd, struct task_struct *p,
5919 int this_cpu, int prev_cpu, int sync)
5921 s64 this_eff_load, prev_eff_load;
5922 unsigned long task_load;
5924 this_eff_load = target_load(this_cpu, sd->wake_idx);
5927 unsigned long current_load = task_h_load(current);
5929 if (current_load > this_eff_load)
5932 this_eff_load -= current_load;
5935 task_load = task_h_load(p);
5937 this_eff_load += task_load;
5938 if (sched_feat(WA_BIAS))
5939 this_eff_load *= 100;
5940 this_eff_load *= capacity_of(prev_cpu);
5942 prev_eff_load = source_load(prev_cpu, sd->wake_idx);
5943 prev_eff_load -= task_load;
5944 if (sched_feat(WA_BIAS))
5945 prev_eff_load *= 100 + (sd->imbalance_pct - 100) / 2;
5946 prev_eff_load *= capacity_of(this_cpu);
5949 * If sync, adjust the weight of prev_eff_load such that if
5950 * prev_eff == this_eff that select_idle_sibling() will consider
5951 * stacking the wakee on top of the waker if no other CPU is
5957 return this_eff_load < prev_eff_load ? this_cpu : nr_cpumask_bits;
5960 static int wake_affine(struct sched_domain *sd, struct task_struct *p,
5961 int this_cpu, int prev_cpu, int sync)
5963 int target = nr_cpumask_bits;
5965 if (sched_feat(WA_IDLE))
5966 target = wake_affine_idle(this_cpu, prev_cpu, sync);
5968 if (sched_feat(WA_WEIGHT) && target == nr_cpumask_bits)
5969 target = wake_affine_weight(sd, p, this_cpu, prev_cpu, sync);
5971 schedstat_inc(p->se.statistics.nr_wakeups_affine_attempts);
5972 if (target == nr_cpumask_bits)
5975 schedstat_inc(sd->ttwu_move_affine);
5976 schedstat_inc(p->se.statistics.nr_wakeups_affine);
5980 static unsigned long cpu_util_wake(int cpu, struct task_struct *p);
5982 static unsigned long capacity_spare_wake(int cpu, struct task_struct *p)
5984 return max_t(long, capacity_of(cpu) - cpu_util_wake(cpu, p), 0);
5988 * find_idlest_group finds and returns the least busy CPU group within the
5991 * Assumes p is allowed on at least one CPU in sd.
5993 static struct sched_group *
5994 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
5995 int this_cpu, int sd_flag)
5997 struct sched_group *idlest = NULL, *group = sd->groups;
5998 struct sched_group *most_spare_sg = NULL;
5999 unsigned long min_runnable_load = ULONG_MAX;
6000 unsigned long this_runnable_load = ULONG_MAX;
6001 unsigned long min_avg_load = ULONG_MAX, this_avg_load = ULONG_MAX;
6002 unsigned long most_spare = 0, this_spare = 0;
6003 int load_idx = sd->forkexec_idx;
6004 int imbalance_scale = 100 + (sd->imbalance_pct-100)/2;
6005 unsigned long imbalance = scale_load_down(NICE_0_LOAD) *
6006 (sd->imbalance_pct-100) / 100;
6008 if (sd_flag & SD_BALANCE_WAKE)
6009 load_idx = sd->wake_idx;
6012 unsigned long load, avg_load, runnable_load;
6013 unsigned long spare_cap, max_spare_cap;
6017 /* Skip over this group if it has no CPUs allowed */
6018 if (!cpumask_intersects(sched_group_span(group),
6022 local_group = cpumask_test_cpu(this_cpu,
6023 sched_group_span(group));
6026 * Tally up the load of all CPUs in the group and find
6027 * the group containing the CPU with most spare capacity.
6033 for_each_cpu(i, sched_group_span(group)) {
6034 /* Bias balancing toward CPUs of our domain */
6036 load = source_load(i, load_idx);
6038 load = target_load(i, load_idx);
6040 runnable_load += load;
6042 avg_load += cfs_rq_load_avg(&cpu_rq(i)->cfs);
6044 spare_cap = capacity_spare_wake(i, p);
6046 if (spare_cap > max_spare_cap)
6047 max_spare_cap = spare_cap;
6050 /* Adjust by relative CPU capacity of the group */
6051 avg_load = (avg_load * SCHED_CAPACITY_SCALE) /
6052 group->sgc->capacity;
6053 runnable_load = (runnable_load * SCHED_CAPACITY_SCALE) /
6054 group->sgc->capacity;
6057 this_runnable_load = runnable_load;
6058 this_avg_load = avg_load;
6059 this_spare = max_spare_cap;
6061 if (min_runnable_load > (runnable_load + imbalance)) {
6063 * The runnable load is significantly smaller
6064 * so we can pick this new CPU:
6066 min_runnable_load = runnable_load;
6067 min_avg_load = avg_load;
6069 } else if ((runnable_load < (min_runnable_load + imbalance)) &&
6070 (100*min_avg_load > imbalance_scale*avg_load)) {
6072 * The runnable loads are close so take the
6073 * blocked load into account through avg_load:
6075 min_avg_load = avg_load;
6079 if (most_spare < max_spare_cap) {
6080 most_spare = max_spare_cap;
6081 most_spare_sg = group;
6084 } while (group = group->next, group != sd->groups);
6087 * The cross-over point between using spare capacity or least load
6088 * is too conservative for high utilization tasks on partially
6089 * utilized systems if we require spare_capacity > task_util(p),
6090 * so we allow for some task stuffing by using
6091 * spare_capacity > task_util(p)/2.
6093 * Spare capacity can't be used for fork because the utilization has
6094 * not been set yet, we must first select a rq to compute the initial
6097 if (sd_flag & SD_BALANCE_FORK)
6100 if (this_spare > task_util(p) / 2 &&
6101 imbalance_scale*this_spare > 100*most_spare)
6104 if (most_spare > task_util(p) / 2)
6105 return most_spare_sg;
6112 * When comparing groups across NUMA domains, it's possible for the
6113 * local domain to be very lightly loaded relative to the remote
6114 * domains but "imbalance" skews the comparison making remote CPUs
6115 * look much more favourable. When considering cross-domain, add
6116 * imbalance to the runnable load on the remote node and consider
6119 if ((sd->flags & SD_NUMA) &&
6120 min_runnable_load + imbalance >= this_runnable_load)
6123 if (min_runnable_load > (this_runnable_load + imbalance))
6126 if ((this_runnable_load < (min_runnable_load + imbalance)) &&
6127 (100*this_avg_load < imbalance_scale*min_avg_load))
6134 * find_idlest_group_cpu - find the idlest CPU among the CPUs in the group.
6137 find_idlest_group_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
6139 unsigned long load, min_load = ULONG_MAX;
6140 unsigned int min_exit_latency = UINT_MAX;
6141 u64 latest_idle_timestamp = 0;
6142 int least_loaded_cpu = this_cpu;
6143 int shallowest_idle_cpu = -1;
6146 /* Check if we have any choice: */
6147 if (group->group_weight == 1)
6148 return cpumask_first(sched_group_span(group));
6150 /* Traverse only the allowed CPUs */
6151 for_each_cpu_and(i, sched_group_span(group), &p->cpus_allowed) {
6152 if (available_idle_cpu(i)) {
6153 struct rq *rq = cpu_rq(i);
6154 struct cpuidle_state *idle = idle_get_state(rq);
6155 if (idle && idle->exit_latency < min_exit_latency) {
6157 * We give priority to a CPU whose idle state
6158 * has the smallest exit latency irrespective
6159 * of any idle timestamp.
6161 min_exit_latency = idle->exit_latency;
6162 latest_idle_timestamp = rq->idle_stamp;
6163 shallowest_idle_cpu = i;
6164 } else if ((!idle || idle->exit_latency == min_exit_latency) &&
6165 rq->idle_stamp > latest_idle_timestamp) {
6167 * If equal or no active idle state, then
6168 * the most recently idled CPU might have
6171 latest_idle_timestamp = rq->idle_stamp;
6172 shallowest_idle_cpu = i;
6174 } else if (shallowest_idle_cpu == -1) {
6175 load = weighted_cpuload(cpu_rq(i));
6176 if (load < min_load) {
6178 least_loaded_cpu = i;
6183 return shallowest_idle_cpu != -1 ? shallowest_idle_cpu : least_loaded_cpu;
6186 static inline int find_idlest_cpu(struct sched_domain *sd, struct task_struct *p,
6187 int cpu, int prev_cpu, int sd_flag)
6191 if (!cpumask_intersects(sched_domain_span(sd), &p->cpus_allowed))
6195 * We need task's util for capacity_spare_wake, sync it up to prev_cpu's
6198 if (!(sd_flag & SD_BALANCE_FORK))
6199 sync_entity_load_avg(&p->se);
6202 struct sched_group *group;
6203 struct sched_domain *tmp;
6206 if (!(sd->flags & sd_flag)) {
6211 group = find_idlest_group(sd, p, cpu, sd_flag);
6217 new_cpu = find_idlest_group_cpu(group, p, cpu);
6218 if (new_cpu == cpu) {
6219 /* Now try balancing at a lower domain level of 'cpu': */
6224 /* Now try balancing at a lower domain level of 'new_cpu': */
6226 weight = sd->span_weight;
6228 for_each_domain(cpu, tmp) {
6229 if (weight <= tmp->span_weight)
6231 if (tmp->flags & sd_flag)
6239 #ifdef CONFIG_SCHED_SMT
6241 static inline void set_idle_cores(int cpu, int val)
6243 struct sched_domain_shared *sds;
6245 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
6247 WRITE_ONCE(sds->has_idle_cores, val);
6250 static inline bool test_idle_cores(int cpu, bool def)
6252 struct sched_domain_shared *sds;
6254 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
6256 return READ_ONCE(sds->has_idle_cores);
6262 * Scans the local SMT mask to see if the entire core is idle, and records this
6263 * information in sd_llc_shared->has_idle_cores.
6265 * Since SMT siblings share all cache levels, inspecting this limited remote
6266 * state should be fairly cheap.
6268 void __update_idle_core(struct rq *rq)
6270 int core = cpu_of(rq);
6274 if (test_idle_cores(core, true))
6277 for_each_cpu(cpu, cpu_smt_mask(core)) {
6281 if (!available_idle_cpu(cpu))
6285 set_idle_cores(core, 1);
6291 * Scan the entire LLC domain for idle cores; this dynamically switches off if
6292 * there are no idle cores left in the system; tracked through
6293 * sd_llc->shared->has_idle_cores and enabled through update_idle_core() above.
6295 static int select_idle_core(struct task_struct *p, struct sched_domain *sd, int target)
6297 struct cpumask *cpus = this_cpu_cpumask_var_ptr(select_idle_mask);
6300 if (!static_branch_likely(&sched_smt_present))
6303 if (!test_idle_cores(target, false))
6306 cpumask_and(cpus, sched_domain_span(sd), &p->cpus_allowed);
6308 for_each_cpu_wrap(core, cpus, target) {
6311 for_each_cpu(cpu, cpu_smt_mask(core)) {
6312 cpumask_clear_cpu(cpu, cpus);
6313 if (!available_idle_cpu(cpu))
6322 * Failed to find an idle core; stop looking for one.
6324 set_idle_cores(target, 0);
6330 * Scan the local SMT mask for idle CPUs.
6332 static int select_idle_smt(struct task_struct *p, struct sched_domain *sd, int target)
6336 if (!static_branch_likely(&sched_smt_present))
6339 for_each_cpu(cpu, cpu_smt_mask(target)) {
6340 if (!cpumask_test_cpu(cpu, &p->cpus_allowed))
6342 if (available_idle_cpu(cpu))
6349 #else /* CONFIG_SCHED_SMT */
6351 static inline int select_idle_core(struct task_struct *p, struct sched_domain *sd, int target)
6356 static inline int select_idle_smt(struct task_struct *p, struct sched_domain *sd, int target)
6361 #endif /* CONFIG_SCHED_SMT */
6364 * Scan the LLC domain for idle CPUs; this is dynamically regulated by
6365 * comparing the average scan cost (tracked in sd->avg_scan_cost) against the
6366 * average idle time for this rq (as found in rq->avg_idle).
6368 static int select_idle_cpu(struct task_struct *p, struct sched_domain *sd, int target)
6370 struct sched_domain *this_sd;
6371 u64 avg_cost, avg_idle;
6374 int cpu, nr = INT_MAX;
6376 this_sd = rcu_dereference(*this_cpu_ptr(&sd_llc));
6381 * Due to large variance we need a large fuzz factor; hackbench in
6382 * particularly is sensitive here.
6384 avg_idle = this_rq()->avg_idle / 512;
6385 avg_cost = this_sd->avg_scan_cost + 1;
6387 if (sched_feat(SIS_AVG_CPU) && avg_idle < avg_cost)
6390 if (sched_feat(SIS_PROP)) {
6391 u64 span_avg = sd->span_weight * avg_idle;
6392 if (span_avg > 4*avg_cost)
6393 nr = div_u64(span_avg, avg_cost);
6398 time = local_clock();
6400 for_each_cpu_wrap(cpu, sched_domain_span(sd), target) {
6403 if (!cpumask_test_cpu(cpu, &p->cpus_allowed))
6405 if (available_idle_cpu(cpu))
6409 time = local_clock() - time;
6410 cost = this_sd->avg_scan_cost;
6411 delta = (s64)(time - cost) / 8;
6412 this_sd->avg_scan_cost += delta;
6418 * Try and locate an idle core/thread in the LLC cache domain.
6420 static int select_idle_sibling(struct task_struct *p, int prev, int target)
6422 struct sched_domain *sd;
6423 int i, recent_used_cpu;
6425 if (available_idle_cpu(target))
6429 * If the previous CPU is cache affine and idle, don't be stupid:
6431 if (prev != target && cpus_share_cache(prev, target) && available_idle_cpu(prev))
6434 /* Check a recently used CPU as a potential idle candidate: */
6435 recent_used_cpu = p->recent_used_cpu;
6436 if (recent_used_cpu != prev &&
6437 recent_used_cpu != target &&
6438 cpus_share_cache(recent_used_cpu, target) &&
6439 available_idle_cpu(recent_used_cpu) &&
6440 cpumask_test_cpu(p->recent_used_cpu, &p->cpus_allowed)) {
6442 * Replace recent_used_cpu with prev as it is a potential
6443 * candidate for the next wake:
6445 p->recent_used_cpu = prev;
6446 return recent_used_cpu;
6449 sd = rcu_dereference(per_cpu(sd_llc, target));
6453 i = select_idle_core(p, sd, target);
6454 if ((unsigned)i < nr_cpumask_bits)
6457 i = select_idle_cpu(p, sd, target);
6458 if ((unsigned)i < nr_cpumask_bits)
6461 i = select_idle_smt(p, sd, target);
6462 if ((unsigned)i < nr_cpumask_bits)
6469 * Amount of capacity of a CPU that is (estimated to be) used by CFS tasks
6470 * @cpu: the CPU to get the utilization of
6472 * The unit of the return value must be the one of capacity so we can compare
6473 * the utilization with the capacity of the CPU that is available for CFS task
6474 * (ie cpu_capacity).
6476 * cfs_rq.avg.util_avg is the sum of running time of runnable tasks plus the
6477 * recent utilization of currently non-runnable tasks on a CPU. It represents
6478 * the amount of utilization of a CPU in the range [0..capacity_orig] where
6479 * capacity_orig is the cpu_capacity available at the highest frequency
6480 * (arch_scale_freq_capacity()).
6481 * The utilization of a CPU converges towards a sum equal to or less than the
6482 * current capacity (capacity_curr <= capacity_orig) of the CPU because it is
6483 * the running time on this CPU scaled by capacity_curr.
6485 * The estimated utilization of a CPU is defined to be the maximum between its
6486 * cfs_rq.avg.util_avg and the sum of the estimated utilization of the tasks
6487 * currently RUNNABLE on that CPU.
6488 * This allows to properly represent the expected utilization of a CPU which
6489 * has just got a big task running since a long sleep period. At the same time
6490 * however it preserves the benefits of the "blocked utilization" in
6491 * describing the potential for other tasks waking up on the same CPU.
6493 * Nevertheless, cfs_rq.avg.util_avg can be higher than capacity_curr or even
6494 * higher than capacity_orig because of unfortunate rounding in
6495 * cfs.avg.util_avg or just after migrating tasks and new task wakeups until
6496 * the average stabilizes with the new running time. We need to check that the
6497 * utilization stays within the range of [0..capacity_orig] and cap it if
6498 * necessary. Without utilization capping, a group could be seen as overloaded
6499 * (CPU0 utilization at 121% + CPU1 utilization at 80%) whereas CPU1 has 20% of
6500 * available capacity. We allow utilization to overshoot capacity_curr (but not
6501 * capacity_orig) as it useful for predicting the capacity required after task
6502 * migrations (scheduler-driven DVFS).
6504 * Return: the (estimated) utilization for the specified CPU
6506 static inline unsigned long cpu_util(int cpu)
6508 struct cfs_rq *cfs_rq;
6511 cfs_rq = &cpu_rq(cpu)->cfs;
6512 util = READ_ONCE(cfs_rq->avg.util_avg);
6514 if (sched_feat(UTIL_EST))
6515 util = max(util, READ_ONCE(cfs_rq->avg.util_est.enqueued));
6517 return min_t(unsigned long, util, capacity_orig_of(cpu));
6521 * cpu_util_wake: Compute CPU utilization with any contributions from
6522 * the waking task p removed.
6524 static unsigned long cpu_util_wake(int cpu, struct task_struct *p)
6526 struct cfs_rq *cfs_rq;
6529 /* Task has no contribution or is new */
6530 if (cpu != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time))
6531 return cpu_util(cpu);
6533 cfs_rq = &cpu_rq(cpu)->cfs;
6534 util = READ_ONCE(cfs_rq->avg.util_avg);
6536 /* Discount task's blocked util from CPU's util */
6537 util -= min_t(unsigned int, util, task_util(p));
6542 * a) if *p is the only task sleeping on this CPU, then:
6543 * cpu_util (== task_util) > util_est (== 0)
6544 * and thus we return:
6545 * cpu_util_wake = (cpu_util - task_util) = 0
6547 * b) if other tasks are SLEEPING on this CPU, which is now exiting
6549 * cpu_util >= task_util
6550 * cpu_util > util_est (== 0)
6551 * and thus we discount *p's blocked utilization to return:
6552 * cpu_util_wake = (cpu_util - task_util) >= 0
6554 * c) if other tasks are RUNNABLE on that CPU and
6555 * util_est > cpu_util
6556 * then we use util_est since it returns a more restrictive
6557 * estimation of the spare capacity on that CPU, by just
6558 * considering the expected utilization of tasks already
6559 * runnable on that CPU.
6561 * Cases a) and b) are covered by the above code, while case c) is
6562 * covered by the following code when estimated utilization is
6565 if (sched_feat(UTIL_EST))
6566 util = max(util, READ_ONCE(cfs_rq->avg.util_est.enqueued));
6569 * Utilization (estimated) can exceed the CPU capacity, thus let's
6570 * clamp to the maximum CPU capacity to ensure consistency with
6571 * the cpu_util call.
6573 return min_t(unsigned long, util, capacity_orig_of(cpu));
6577 * Disable WAKE_AFFINE in the case where task @p doesn't fit in the
6578 * capacity of either the waking CPU @cpu or the previous CPU @prev_cpu.
6580 * In that case WAKE_AFFINE doesn't make sense and we'll let
6581 * BALANCE_WAKE sort things out.
6583 static int wake_cap(struct task_struct *p, int cpu, int prev_cpu)
6585 long min_cap, max_cap;
6587 min_cap = min(capacity_orig_of(prev_cpu), capacity_orig_of(cpu));
6588 max_cap = cpu_rq(cpu)->rd->max_cpu_capacity;
6590 /* Minimum capacity is close to max, no need to abort wake_affine */
6591 if (max_cap - min_cap < max_cap >> 3)
6594 /* Bring task utilization in sync with prev_cpu */
6595 sync_entity_load_avg(&p->se);
6597 return min_cap * 1024 < task_util(p) * capacity_margin;
6601 * select_task_rq_fair: Select target runqueue for the waking task in domains
6602 * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE,
6603 * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
6605 * Balances load by selecting the idlest CPU in the idlest group, or under
6606 * certain conditions an idle sibling CPU if the domain has SD_WAKE_AFFINE set.
6608 * Returns the target CPU number.
6610 * preempt must be disabled.
6613 select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags)
6615 struct sched_domain *tmp, *sd = NULL;
6616 int cpu = smp_processor_id();
6617 int new_cpu = prev_cpu;
6618 int want_affine = 0;
6619 int sync = (wake_flags & WF_SYNC) && !(current->flags & PF_EXITING);
6621 if (sd_flag & SD_BALANCE_WAKE) {
6623 want_affine = !wake_wide(p) && !wake_cap(p, cpu, prev_cpu)
6624 && cpumask_test_cpu(cpu, &p->cpus_allowed);
6628 for_each_domain(cpu, tmp) {
6629 if (!(tmp->flags & SD_LOAD_BALANCE))
6633 * If both 'cpu' and 'prev_cpu' are part of this domain,
6634 * cpu is a valid SD_WAKE_AFFINE target.
6636 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
6637 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
6638 if (cpu != prev_cpu)
6639 new_cpu = wake_affine(tmp, p, cpu, prev_cpu, sync);
6641 sd = NULL; /* Prefer wake_affine over balance flags */
6645 if (tmp->flags & sd_flag)
6647 else if (!want_affine)
6653 new_cpu = find_idlest_cpu(sd, p, cpu, prev_cpu, sd_flag);
6654 } else if (sd_flag & SD_BALANCE_WAKE) { /* XXX always ? */
6657 new_cpu = select_idle_sibling(p, prev_cpu, new_cpu);
6660 current->recent_used_cpu = cpu;
6667 static void detach_entity_cfs_rq(struct sched_entity *se);
6670 * Called immediately before a task is migrated to a new CPU; task_cpu(p) and
6671 * cfs_rq_of(p) references at time of call are still valid and identify the
6672 * previous CPU. The caller guarantees p->pi_lock or task_rq(p)->lock is held.
6674 static void migrate_task_rq_fair(struct task_struct *p)
6677 * As blocked tasks retain absolute vruntime the migration needs to
6678 * deal with this by subtracting the old and adding the new
6679 * min_vruntime -- the latter is done by enqueue_entity() when placing
6680 * the task on the new runqueue.
6682 if (p->state == TASK_WAKING) {
6683 struct sched_entity *se = &p->se;
6684 struct cfs_rq *cfs_rq = cfs_rq_of(se);
6687 #ifndef CONFIG_64BIT
6688 u64 min_vruntime_copy;
6691 min_vruntime_copy = cfs_rq->min_vruntime_copy;
6693 min_vruntime = cfs_rq->min_vruntime;
6694 } while (min_vruntime != min_vruntime_copy);
6696 min_vruntime = cfs_rq->min_vruntime;
6699 se->vruntime -= min_vruntime;
6702 if (p->on_rq == TASK_ON_RQ_MIGRATING) {
6704 * In case of TASK_ON_RQ_MIGRATING we in fact hold the 'old'
6705 * rq->lock and can modify state directly.
6707 lockdep_assert_held(&task_rq(p)->lock);
6708 detach_entity_cfs_rq(&p->se);
6712 * We are supposed to update the task to "current" time, then
6713 * its up to date and ready to go to new CPU/cfs_rq. But we
6714 * have difficulty in getting what current time is, so simply
6715 * throw away the out-of-date time. This will result in the
6716 * wakee task is less decayed, but giving the wakee more load
6719 remove_entity_load_avg(&p->se);
6722 /* Tell new CPU we are migrated */
6723 p->se.avg.last_update_time = 0;
6725 /* We have migrated, no longer consider this task hot */
6726 p->se.exec_start = 0;
6729 static void task_dead_fair(struct task_struct *p)
6731 remove_entity_load_avg(&p->se);
6733 #endif /* CONFIG_SMP */
6735 static unsigned long wakeup_gran(struct sched_entity *se)
6737 unsigned long gran = sysctl_sched_wakeup_granularity;
6740 * Since its curr running now, convert the gran from real-time
6741 * to virtual-time in his units.
6743 * By using 'se' instead of 'curr' we penalize light tasks, so
6744 * they get preempted easier. That is, if 'se' < 'curr' then
6745 * the resulting gran will be larger, therefore penalizing the
6746 * lighter, if otoh 'se' > 'curr' then the resulting gran will
6747 * be smaller, again penalizing the lighter task.
6749 * This is especially important for buddies when the leftmost
6750 * task is higher priority than the buddy.
6752 return calc_delta_fair(gran, se);
6756 * Should 'se' preempt 'curr'.
6770 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
6772 s64 gran, vdiff = curr->vruntime - se->vruntime;
6777 gran = wakeup_gran(se);
6784 static void set_last_buddy(struct sched_entity *se)
6786 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
6789 for_each_sched_entity(se) {
6790 if (SCHED_WARN_ON(!se->on_rq))
6792 cfs_rq_of(se)->last = se;
6796 static void set_next_buddy(struct sched_entity *se)
6798 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
6801 for_each_sched_entity(se) {
6802 if (SCHED_WARN_ON(!se->on_rq))
6804 cfs_rq_of(se)->next = se;
6808 static void set_skip_buddy(struct sched_entity *se)
6810 for_each_sched_entity(se)
6811 cfs_rq_of(se)->skip = se;
6815 * Preempt the current task with a newly woken task if needed:
6817 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
6819 struct task_struct *curr = rq->curr;
6820 struct sched_entity *se = &curr->se, *pse = &p->se;
6821 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
6822 int scale = cfs_rq->nr_running >= sched_nr_latency;
6823 int next_buddy_marked = 0;
6825 if (unlikely(se == pse))
6829 * This is possible from callers such as attach_tasks(), in which we
6830 * unconditionally check_prempt_curr() after an enqueue (which may have
6831 * lead to a throttle). This both saves work and prevents false
6832 * next-buddy nomination below.
6834 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
6837 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
6838 set_next_buddy(pse);
6839 next_buddy_marked = 1;
6843 * We can come here with TIF_NEED_RESCHED already set from new task
6846 * Note: this also catches the edge-case of curr being in a throttled
6847 * group (e.g. via set_curr_task), since update_curr() (in the
6848 * enqueue of curr) will have resulted in resched being set. This
6849 * prevents us from potentially nominating it as a false LAST_BUDDY
6852 if (test_tsk_need_resched(curr))
6855 /* Idle tasks are by definition preempted by non-idle tasks. */
6856 if (unlikely(curr->policy == SCHED_IDLE) &&
6857 likely(p->policy != SCHED_IDLE))
6861 * Batch and idle tasks do not preempt non-idle tasks (their preemption
6862 * is driven by the tick):
6864 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
6867 find_matching_se(&se, &pse);
6868 update_curr(cfs_rq_of(se));
6870 if (wakeup_preempt_entity(se, pse) == 1) {
6872 * Bias pick_next to pick the sched entity that is
6873 * triggering this preemption.
6875 if (!next_buddy_marked)
6876 set_next_buddy(pse);
6885 * Only set the backward buddy when the current task is still
6886 * on the rq. This can happen when a wakeup gets interleaved
6887 * with schedule on the ->pre_schedule() or idle_balance()
6888 * point, either of which can * drop the rq lock.
6890 * Also, during early boot the idle thread is in the fair class,
6891 * for obvious reasons its a bad idea to schedule back to it.
6893 if (unlikely(!se->on_rq || curr == rq->idle))
6896 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
6900 static struct task_struct *
6901 pick_next_task_fair(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
6903 struct cfs_rq *cfs_rq = &rq->cfs;
6904 struct sched_entity *se;
6905 struct task_struct *p;
6909 if (!cfs_rq->nr_running)
6912 #ifdef CONFIG_FAIR_GROUP_SCHED
6913 if (prev->sched_class != &fair_sched_class)
6917 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
6918 * likely that a next task is from the same cgroup as the current.
6920 * Therefore attempt to avoid putting and setting the entire cgroup
6921 * hierarchy, only change the part that actually changes.
6925 struct sched_entity *curr = cfs_rq->curr;
6928 * Since we got here without doing put_prev_entity() we also
6929 * have to consider cfs_rq->curr. If it is still a runnable
6930 * entity, update_curr() will update its vruntime, otherwise
6931 * forget we've ever seen it.
6935 update_curr(cfs_rq);
6940 * This call to check_cfs_rq_runtime() will do the
6941 * throttle and dequeue its entity in the parent(s).
6942 * Therefore the nr_running test will indeed
6945 if (unlikely(check_cfs_rq_runtime(cfs_rq))) {
6948 if (!cfs_rq->nr_running)
6955 se = pick_next_entity(cfs_rq, curr);
6956 cfs_rq = group_cfs_rq(se);
6962 * Since we haven't yet done put_prev_entity and if the selected task
6963 * is a different task than we started out with, try and touch the
6964 * least amount of cfs_rqs.
6967 struct sched_entity *pse = &prev->se;
6969 while (!(cfs_rq = is_same_group(se, pse))) {
6970 int se_depth = se->depth;
6971 int pse_depth = pse->depth;
6973 if (se_depth <= pse_depth) {
6974 put_prev_entity(cfs_rq_of(pse), pse);
6975 pse = parent_entity(pse);
6977 if (se_depth >= pse_depth) {
6978 set_next_entity(cfs_rq_of(se), se);
6979 se = parent_entity(se);
6983 put_prev_entity(cfs_rq, pse);
6984 set_next_entity(cfs_rq, se);
6991 put_prev_task(rq, prev);
6994 se = pick_next_entity(cfs_rq, NULL);
6995 set_next_entity(cfs_rq, se);
6996 cfs_rq = group_cfs_rq(se);
7001 done: __maybe_unused;
7004 * Move the next running task to the front of
7005 * the list, so our cfs_tasks list becomes MRU
7008 list_move(&p->se.group_node, &rq->cfs_tasks);
7011 if (hrtick_enabled(rq))
7012 hrtick_start_fair(rq, p);
7017 new_tasks = idle_balance(rq, rf);
7020 * Because idle_balance() releases (and re-acquires) rq->lock, it is
7021 * possible for any higher priority task to appear. In that case we
7022 * must re-start the pick_next_entity() loop.
7034 * Account for a descheduled task:
7036 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
7038 struct sched_entity *se = &prev->se;
7039 struct cfs_rq *cfs_rq;
7041 for_each_sched_entity(se) {
7042 cfs_rq = cfs_rq_of(se);
7043 put_prev_entity(cfs_rq, se);
7048 * sched_yield() is very simple
7050 * The magic of dealing with the ->skip buddy is in pick_next_entity.
7052 static void yield_task_fair(struct rq *rq)
7054 struct task_struct *curr = rq->curr;
7055 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
7056 struct sched_entity *se = &curr->se;
7059 * Are we the only task in the tree?
7061 if (unlikely(rq->nr_running == 1))
7064 clear_buddies(cfs_rq, se);
7066 if (curr->policy != SCHED_BATCH) {
7067 update_rq_clock(rq);
7069 * Update run-time statistics of the 'current'.
7071 update_curr(cfs_rq);
7073 * Tell update_rq_clock() that we've just updated,
7074 * so we don't do microscopic update in schedule()
7075 * and double the fastpath cost.
7077 rq_clock_skip_update(rq);
7083 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
7085 struct sched_entity *se = &p->se;
7087 /* throttled hierarchies are not runnable */
7088 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
7091 /* Tell the scheduler that we'd really like pse to run next. */
7094 yield_task_fair(rq);
7100 /**************************************************
7101 * Fair scheduling class load-balancing methods.
7105 * The purpose of load-balancing is to achieve the same basic fairness the
7106 * per-CPU scheduler provides, namely provide a proportional amount of compute
7107 * time to each task. This is expressed in the following equation:
7109 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
7111 * Where W_i,n is the n-th weight average for CPU i. The instantaneous weight
7112 * W_i,0 is defined as:
7114 * W_i,0 = \Sum_j w_i,j (2)
7116 * Where w_i,j is the weight of the j-th runnable task on CPU i. This weight
7117 * is derived from the nice value as per sched_prio_to_weight[].
7119 * The weight average is an exponential decay average of the instantaneous
7122 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
7124 * C_i is the compute capacity of CPU i, typically it is the
7125 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
7126 * can also include other factors [XXX].
7128 * To achieve this balance we define a measure of imbalance which follows
7129 * directly from (1):
7131 * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4)
7133 * We them move tasks around to minimize the imbalance. In the continuous
7134 * function space it is obvious this converges, in the discrete case we get
7135 * a few fun cases generally called infeasible weight scenarios.
7138 * - infeasible weights;
7139 * - local vs global optima in the discrete case. ]
7144 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
7145 * for all i,j solution, we create a tree of CPUs that follows the hardware
7146 * topology where each level pairs two lower groups (or better). This results
7147 * in O(log n) layers. Furthermore we reduce the number of CPUs going up the
7148 * tree to only the first of the previous level and we decrease the frequency
7149 * of load-balance at each level inv. proportional to the number of CPUs in
7155 * \Sum { --- * --- * 2^i } = O(n) (5)
7157 * `- size of each group
7158 * | | `- number of CPUs doing load-balance
7160 * `- sum over all levels
7162 * Coupled with a limit on how many tasks we can migrate every balance pass,
7163 * this makes (5) the runtime complexity of the balancer.
7165 * An important property here is that each CPU is still (indirectly) connected
7166 * to every other CPU in at most O(log n) steps:
7168 * The adjacency matrix of the resulting graph is given by:
7171 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
7174 * And you'll find that:
7176 * A^(log_2 n)_i,j != 0 for all i,j (7)
7178 * Showing there's indeed a path between every CPU in at most O(log n) steps.
7179 * The task movement gives a factor of O(m), giving a convergence complexity
7182 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
7187 * In order to avoid CPUs going idle while there's still work to do, new idle
7188 * balancing is more aggressive and has the newly idle CPU iterate up the domain
7189 * tree itself instead of relying on other CPUs to bring it work.
7191 * This adds some complexity to both (5) and (8) but it reduces the total idle
7199 * Cgroups make a horror show out of (2), instead of a simple sum we get:
7202 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
7207 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
7209 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on CPU i.
7211 * The big problem is S_k, its a global sum needed to compute a local (W_i)
7214 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
7215 * rewrite all of this once again.]
7218 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
7220 enum fbq_type { regular, remote, all };
7222 #define LBF_ALL_PINNED 0x01
7223 #define LBF_NEED_BREAK 0x02
7224 #define LBF_DST_PINNED 0x04
7225 #define LBF_SOME_PINNED 0x08
7226 #define LBF_NOHZ_STATS 0x10
7227 #define LBF_NOHZ_AGAIN 0x20
7230 struct sched_domain *sd;
7238 struct cpumask *dst_grpmask;
7240 enum cpu_idle_type idle;
7242 /* The set of CPUs under consideration for load-balancing */
7243 struct cpumask *cpus;
7248 unsigned int loop_break;
7249 unsigned int loop_max;
7251 enum fbq_type fbq_type;
7252 struct list_head tasks;
7256 * Is this task likely cache-hot:
7258 static int task_hot(struct task_struct *p, struct lb_env *env)
7262 lockdep_assert_held(&env->src_rq->lock);
7264 if (p->sched_class != &fair_sched_class)
7267 if (unlikely(p->policy == SCHED_IDLE))
7271 * Buddy candidates are cache hot:
7273 if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running &&
7274 (&p->se == cfs_rq_of(&p->se)->next ||
7275 &p->se == cfs_rq_of(&p->se)->last))
7278 if (sysctl_sched_migration_cost == -1)
7280 if (sysctl_sched_migration_cost == 0)
7283 delta = rq_clock_task(env->src_rq) - p->se.exec_start;
7285 return delta < (s64)sysctl_sched_migration_cost;
7288 #ifdef CONFIG_NUMA_BALANCING
7290 * Returns 1, if task migration degrades locality
7291 * Returns 0, if task migration improves locality i.e migration preferred.
7292 * Returns -1, if task migration is not affected by locality.
7294 static int migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
7296 struct numa_group *numa_group = rcu_dereference(p->numa_group);
7297 unsigned long src_faults, dst_faults;
7298 int src_nid, dst_nid;
7300 if (!static_branch_likely(&sched_numa_balancing))
7303 if (!p->numa_faults || !(env->sd->flags & SD_NUMA))
7306 src_nid = cpu_to_node(env->src_cpu);
7307 dst_nid = cpu_to_node(env->dst_cpu);
7309 if (src_nid == dst_nid)
7312 /* Migrating away from the preferred node is always bad. */
7313 if (src_nid == p->numa_preferred_nid) {
7314 if (env->src_rq->nr_running > env->src_rq->nr_preferred_running)
7320 /* Encourage migration to the preferred node. */
7321 if (dst_nid == p->numa_preferred_nid)
7324 /* Leaving a core idle is often worse than degrading locality. */
7325 if (env->idle != CPU_NOT_IDLE)
7329 src_faults = group_faults(p, src_nid);
7330 dst_faults = group_faults(p, dst_nid);
7332 src_faults = task_faults(p, src_nid);
7333 dst_faults = task_faults(p, dst_nid);
7336 return dst_faults < src_faults;
7340 static inline int migrate_degrades_locality(struct task_struct *p,
7348 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
7351 int can_migrate_task(struct task_struct *p, struct lb_env *env)
7355 lockdep_assert_held(&env->src_rq->lock);
7358 * We do not migrate tasks that are:
7359 * 1) throttled_lb_pair, or
7360 * 2) cannot be migrated to this CPU due to cpus_allowed, or
7361 * 3) running (obviously), or
7362 * 4) are cache-hot on their current CPU.
7364 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
7367 if (!cpumask_test_cpu(env->dst_cpu, &p->cpus_allowed)) {
7370 schedstat_inc(p->se.statistics.nr_failed_migrations_affine);
7372 env->flags |= LBF_SOME_PINNED;
7375 * Remember if this task can be migrated to any other CPU in
7376 * our sched_group. We may want to revisit it if we couldn't
7377 * meet load balance goals by pulling other tasks on src_cpu.
7379 * Avoid computing new_dst_cpu for NEWLY_IDLE or if we have
7380 * already computed one in current iteration.
7382 if (env->idle == CPU_NEWLY_IDLE || (env->flags & LBF_DST_PINNED))
7385 /* Prevent to re-select dst_cpu via env's CPUs: */
7386 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
7387 if (cpumask_test_cpu(cpu, &p->cpus_allowed)) {
7388 env->flags |= LBF_DST_PINNED;
7389 env->new_dst_cpu = cpu;
7397 /* Record that we found atleast one task that could run on dst_cpu */
7398 env->flags &= ~LBF_ALL_PINNED;
7400 if (task_running(env->src_rq, p)) {
7401 schedstat_inc(p->se.statistics.nr_failed_migrations_running);
7406 * Aggressive migration if:
7407 * 1) destination numa is preferred
7408 * 2) task is cache cold, or
7409 * 3) too many balance attempts have failed.
7411 tsk_cache_hot = migrate_degrades_locality(p, env);
7412 if (tsk_cache_hot == -1)
7413 tsk_cache_hot = task_hot(p, env);
7415 if (tsk_cache_hot <= 0 ||
7416 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
7417 if (tsk_cache_hot == 1) {
7418 schedstat_inc(env->sd->lb_hot_gained[env->idle]);
7419 schedstat_inc(p->se.statistics.nr_forced_migrations);
7424 schedstat_inc(p->se.statistics.nr_failed_migrations_hot);
7429 * detach_task() -- detach the task for the migration specified in env
7431 static void detach_task(struct task_struct *p, struct lb_env *env)
7433 lockdep_assert_held(&env->src_rq->lock);
7435 p->on_rq = TASK_ON_RQ_MIGRATING;
7436 deactivate_task(env->src_rq, p, DEQUEUE_NOCLOCK);
7437 set_task_cpu(p, env->dst_cpu);
7441 * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
7442 * part of active balancing operations within "domain".
7444 * Returns a task if successful and NULL otherwise.
7446 static struct task_struct *detach_one_task(struct lb_env *env)
7448 struct task_struct *p;
7450 lockdep_assert_held(&env->src_rq->lock);
7452 list_for_each_entry_reverse(p,
7453 &env->src_rq->cfs_tasks, se.group_node) {
7454 if (!can_migrate_task(p, env))
7457 detach_task(p, env);
7460 * Right now, this is only the second place where
7461 * lb_gained[env->idle] is updated (other is detach_tasks)
7462 * so we can safely collect stats here rather than
7463 * inside detach_tasks().
7465 schedstat_inc(env->sd->lb_gained[env->idle]);
7471 static const unsigned int sched_nr_migrate_break = 32;
7474 * detach_tasks() -- tries to detach up to imbalance weighted load from
7475 * busiest_rq, as part of a balancing operation within domain "sd".
7477 * Returns number of detached tasks if successful and 0 otherwise.
7479 static int detach_tasks(struct lb_env *env)
7481 struct list_head *tasks = &env->src_rq->cfs_tasks;
7482 struct task_struct *p;
7486 lockdep_assert_held(&env->src_rq->lock);
7488 if (env->imbalance <= 0)
7491 while (!list_empty(tasks)) {
7493 * We don't want to steal all, otherwise we may be treated likewise,
7494 * which could at worst lead to a livelock crash.
7496 if (env->idle != CPU_NOT_IDLE && env->src_rq->nr_running <= 1)
7499 p = list_last_entry(tasks, struct task_struct, se.group_node);
7502 /* We've more or less seen every task there is, call it quits */
7503 if (env->loop > env->loop_max)
7506 /* take a breather every nr_migrate tasks */
7507 if (env->loop > env->loop_break) {
7508 env->loop_break += sched_nr_migrate_break;
7509 env->flags |= LBF_NEED_BREAK;
7513 if (!can_migrate_task(p, env))
7516 load = task_h_load(p);
7518 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
7521 if ((load / 2) > env->imbalance)
7524 detach_task(p, env);
7525 list_add(&p->se.group_node, &env->tasks);
7528 env->imbalance -= load;
7530 #ifdef CONFIG_PREEMPT
7532 * NEWIDLE balancing is a source of latency, so preemptible
7533 * kernels will stop after the first task is detached to minimize
7534 * the critical section.
7536 if (env->idle == CPU_NEWLY_IDLE)
7541 * We only want to steal up to the prescribed amount of
7544 if (env->imbalance <= 0)
7549 list_move(&p->se.group_node, tasks);
7553 * Right now, this is one of only two places we collect this stat
7554 * so we can safely collect detach_one_task() stats here rather
7555 * than inside detach_one_task().
7557 schedstat_add(env->sd->lb_gained[env->idle], detached);
7563 * attach_task() -- attach the task detached by detach_task() to its new rq.
7565 static void attach_task(struct rq *rq, struct task_struct *p)
7567 lockdep_assert_held(&rq->lock);
7569 BUG_ON(task_rq(p) != rq);
7570 activate_task(rq, p, ENQUEUE_NOCLOCK);
7571 p->on_rq = TASK_ON_RQ_QUEUED;
7572 check_preempt_curr(rq, p, 0);
7576 * attach_one_task() -- attaches the task returned from detach_one_task() to
7579 static void attach_one_task(struct rq *rq, struct task_struct *p)
7584 update_rq_clock(rq);
7590 * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
7593 static void attach_tasks(struct lb_env *env)
7595 struct list_head *tasks = &env->tasks;
7596 struct task_struct *p;
7599 rq_lock(env->dst_rq, &rf);
7600 update_rq_clock(env->dst_rq);
7602 while (!list_empty(tasks)) {
7603 p = list_first_entry(tasks, struct task_struct, se.group_node);
7604 list_del_init(&p->se.group_node);
7606 attach_task(env->dst_rq, p);
7609 rq_unlock(env->dst_rq, &rf);
7612 static inline bool cfs_rq_has_blocked(struct cfs_rq *cfs_rq)
7614 if (cfs_rq->avg.load_avg)
7617 if (cfs_rq->avg.util_avg)
7623 #ifdef CONFIG_FAIR_GROUP_SCHED
7625 static inline bool cfs_rq_is_decayed(struct cfs_rq *cfs_rq)
7627 if (cfs_rq->load.weight)
7630 if (cfs_rq->avg.load_sum)
7633 if (cfs_rq->avg.util_sum)
7636 if (cfs_rq->avg.runnable_load_sum)
7642 static void update_blocked_averages(int cpu)
7644 struct rq *rq = cpu_rq(cpu);
7645 struct cfs_rq *cfs_rq, *pos;
7649 rq_lock_irqsave(rq, &rf);
7650 update_rq_clock(rq);
7653 * Iterates the task_group tree in a bottom up fashion, see
7654 * list_add_leaf_cfs_rq() for details.
7656 for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) {
7657 struct sched_entity *se;
7659 /* throttled entities do not contribute to load */
7660 if (throttled_hierarchy(cfs_rq))
7663 if (update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq))
7664 update_tg_load_avg(cfs_rq, 0);
7666 /* Propagate pending load changes to the parent, if any: */
7667 se = cfs_rq->tg->se[cpu];
7668 if (se && !skip_blocked_update(se))
7669 update_load_avg(cfs_rq_of(se), se, 0);
7672 * There can be a lot of idle CPU cgroups. Don't let fully
7673 * decayed cfs_rqs linger on the list.
7675 if (cfs_rq_is_decayed(cfs_rq))
7676 list_del_leaf_cfs_rq(cfs_rq);
7678 /* Don't need periodic decay once load/util_avg are null */
7679 if (cfs_rq_has_blocked(cfs_rq))
7683 #ifdef CONFIG_NO_HZ_COMMON
7684 rq->last_blocked_load_update_tick = jiffies;
7686 rq->has_blocked_load = 0;
7688 rq_unlock_irqrestore(rq, &rf);
7692 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
7693 * This needs to be done in a top-down fashion because the load of a child
7694 * group is a fraction of its parents load.
7696 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
7698 struct rq *rq = rq_of(cfs_rq);
7699 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
7700 unsigned long now = jiffies;
7703 if (cfs_rq->last_h_load_update == now)
7706 cfs_rq->h_load_next = NULL;
7707 for_each_sched_entity(se) {
7708 cfs_rq = cfs_rq_of(se);
7709 cfs_rq->h_load_next = se;
7710 if (cfs_rq->last_h_load_update == now)
7715 cfs_rq->h_load = cfs_rq_load_avg(cfs_rq);
7716 cfs_rq->last_h_load_update = now;
7719 while ((se = cfs_rq->h_load_next) != NULL) {
7720 load = cfs_rq->h_load;
7721 load = div64_ul(load * se->avg.load_avg,
7722 cfs_rq_load_avg(cfs_rq) + 1);
7723 cfs_rq = group_cfs_rq(se);
7724 cfs_rq->h_load = load;
7725 cfs_rq->last_h_load_update = now;
7729 static unsigned long task_h_load(struct task_struct *p)
7731 struct cfs_rq *cfs_rq = task_cfs_rq(p);
7733 update_cfs_rq_h_load(cfs_rq);
7734 return div64_ul(p->se.avg.load_avg * cfs_rq->h_load,
7735 cfs_rq_load_avg(cfs_rq) + 1);
7738 static inline void update_blocked_averages(int cpu)
7740 struct rq *rq = cpu_rq(cpu);
7741 struct cfs_rq *cfs_rq = &rq->cfs;
7744 rq_lock_irqsave(rq, &rf);
7745 update_rq_clock(rq);
7746 update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq);
7747 #ifdef CONFIG_NO_HZ_COMMON
7748 rq->last_blocked_load_update_tick = jiffies;
7749 if (!cfs_rq_has_blocked(cfs_rq))
7750 rq->has_blocked_load = 0;
7752 rq_unlock_irqrestore(rq, &rf);
7755 static unsigned long task_h_load(struct task_struct *p)
7757 return p->se.avg.load_avg;
7761 /********** Helpers for find_busiest_group ************************/
7770 * sg_lb_stats - stats of a sched_group required for load_balancing
7772 struct sg_lb_stats {
7773 unsigned long avg_load; /*Avg load across the CPUs of the group */
7774 unsigned long group_load; /* Total load over the CPUs of the group */
7775 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
7776 unsigned long load_per_task;
7777 unsigned long group_capacity;
7778 unsigned long group_util; /* Total utilization of the group */
7779 unsigned int sum_nr_running; /* Nr tasks running in the group */
7780 unsigned int idle_cpus;
7781 unsigned int group_weight;
7782 enum group_type group_type;
7783 int group_no_capacity;
7784 #ifdef CONFIG_NUMA_BALANCING
7785 unsigned int nr_numa_running;
7786 unsigned int nr_preferred_running;
7791 * sd_lb_stats - Structure to store the statistics of a sched_domain
7792 * during load balancing.
7794 struct sd_lb_stats {
7795 struct sched_group *busiest; /* Busiest group in this sd */
7796 struct sched_group *local; /* Local group in this sd */
7797 unsigned long total_running;
7798 unsigned long total_load; /* Total load of all groups in sd */
7799 unsigned long total_capacity; /* Total capacity of all groups in sd */
7800 unsigned long avg_load; /* Average load across all groups in sd */
7802 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
7803 struct sg_lb_stats local_stat; /* Statistics of the local group */
7806 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
7809 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
7810 * local_stat because update_sg_lb_stats() does a full clear/assignment.
7811 * We must however clear busiest_stat::avg_load because
7812 * update_sd_pick_busiest() reads this before assignment.
7814 *sds = (struct sd_lb_stats){
7817 .total_running = 0UL,
7819 .total_capacity = 0UL,
7822 .sum_nr_running = 0,
7823 .group_type = group_other,
7829 * get_sd_load_idx - Obtain the load index for a given sched domain.
7830 * @sd: The sched_domain whose load_idx is to be obtained.
7831 * @idle: The idle status of the CPU for whose sd load_idx is obtained.
7833 * Return: The load index.
7835 static inline int get_sd_load_idx(struct sched_domain *sd,
7836 enum cpu_idle_type idle)
7842 load_idx = sd->busy_idx;
7845 case CPU_NEWLY_IDLE:
7846 load_idx = sd->newidle_idx;
7849 load_idx = sd->idle_idx;
7856 static unsigned long scale_rt_capacity(int cpu)
7858 struct rq *rq = cpu_rq(cpu);
7859 u64 total, used, age_stamp, avg;
7863 * Since we're reading these variables without serialization make sure
7864 * we read them once before doing sanity checks on them.
7866 age_stamp = READ_ONCE(rq->age_stamp);
7867 avg = READ_ONCE(rq->rt_avg);
7868 delta = __rq_clock_broken(rq) - age_stamp;
7870 if (unlikely(delta < 0))
7873 total = sched_avg_period() + delta;
7875 used = div_u64(avg, total);
7877 if (likely(used < SCHED_CAPACITY_SCALE))
7878 return SCHED_CAPACITY_SCALE - used;
7883 static void update_cpu_capacity(struct sched_domain *sd, int cpu)
7885 unsigned long capacity = arch_scale_cpu_capacity(sd, cpu);
7886 struct sched_group *sdg = sd->groups;
7888 cpu_rq(cpu)->cpu_capacity_orig = capacity;
7890 capacity *= scale_rt_capacity(cpu);
7891 capacity >>= SCHED_CAPACITY_SHIFT;
7896 cpu_rq(cpu)->cpu_capacity = capacity;
7897 sdg->sgc->capacity = capacity;
7898 sdg->sgc->min_capacity = capacity;
7901 void update_group_capacity(struct sched_domain *sd, int cpu)
7903 struct sched_domain *child = sd->child;
7904 struct sched_group *group, *sdg = sd->groups;
7905 unsigned long capacity, min_capacity;
7906 unsigned long interval;
7908 interval = msecs_to_jiffies(sd->balance_interval);
7909 interval = clamp(interval, 1UL, max_load_balance_interval);
7910 sdg->sgc->next_update = jiffies + interval;
7913 update_cpu_capacity(sd, cpu);
7918 min_capacity = ULONG_MAX;
7920 if (child->flags & SD_OVERLAP) {
7922 * SD_OVERLAP domains cannot assume that child groups
7923 * span the current group.
7926 for_each_cpu(cpu, sched_group_span(sdg)) {
7927 struct sched_group_capacity *sgc;
7928 struct rq *rq = cpu_rq(cpu);
7931 * build_sched_domains() -> init_sched_groups_capacity()
7932 * gets here before we've attached the domains to the
7935 * Use capacity_of(), which is set irrespective of domains
7936 * in update_cpu_capacity().
7938 * This avoids capacity from being 0 and
7939 * causing divide-by-zero issues on boot.
7941 if (unlikely(!rq->sd)) {
7942 capacity += capacity_of(cpu);
7944 sgc = rq->sd->groups->sgc;
7945 capacity += sgc->capacity;
7948 min_capacity = min(capacity, min_capacity);
7952 * !SD_OVERLAP domains can assume that child groups
7953 * span the current group.
7956 group = child->groups;
7958 struct sched_group_capacity *sgc = group->sgc;
7960 capacity += sgc->capacity;
7961 min_capacity = min(sgc->min_capacity, min_capacity);
7962 group = group->next;
7963 } while (group != child->groups);
7966 sdg->sgc->capacity = capacity;
7967 sdg->sgc->min_capacity = min_capacity;
7971 * Check whether the capacity of the rq has been noticeably reduced by side
7972 * activity. The imbalance_pct is used for the threshold.
7973 * Return true is the capacity is reduced
7976 check_cpu_capacity(struct rq *rq, struct sched_domain *sd)
7978 return ((rq->cpu_capacity * sd->imbalance_pct) <
7979 (rq->cpu_capacity_orig * 100));
7983 * Group imbalance indicates (and tries to solve) the problem where balancing
7984 * groups is inadequate due to ->cpus_allowed constraints.
7986 * Imagine a situation of two groups of 4 CPUs each and 4 tasks each with a
7987 * cpumask covering 1 CPU of the first group and 3 CPUs of the second group.
7990 * { 0 1 2 3 } { 4 5 6 7 }
7993 * If we were to balance group-wise we'd place two tasks in the first group and
7994 * two tasks in the second group. Clearly this is undesired as it will overload
7995 * cpu 3 and leave one of the CPUs in the second group unused.
7997 * The current solution to this issue is detecting the skew in the first group
7998 * by noticing the lower domain failed to reach balance and had difficulty
7999 * moving tasks due to affinity constraints.
8001 * When this is so detected; this group becomes a candidate for busiest; see
8002 * update_sd_pick_busiest(). And calculate_imbalance() and
8003 * find_busiest_group() avoid some of the usual balance conditions to allow it
8004 * to create an effective group imbalance.
8006 * This is a somewhat tricky proposition since the next run might not find the
8007 * group imbalance and decide the groups need to be balanced again. A most
8008 * subtle and fragile situation.
8011 static inline int sg_imbalanced(struct sched_group *group)
8013 return group->sgc->imbalance;
8017 * group_has_capacity returns true if the group has spare capacity that could
8018 * be used by some tasks.
8019 * We consider that a group has spare capacity if the * number of task is
8020 * smaller than the number of CPUs or if the utilization is lower than the
8021 * available capacity for CFS tasks.
8022 * For the latter, we use a threshold to stabilize the state, to take into
8023 * account the variance of the tasks' load and to return true if the available
8024 * capacity in meaningful for the load balancer.
8025 * As an example, an available capacity of 1% can appear but it doesn't make
8026 * any benefit for the load balance.
8029 group_has_capacity(struct lb_env *env, struct sg_lb_stats *sgs)
8031 if (sgs->sum_nr_running < sgs->group_weight)
8034 if ((sgs->group_capacity * 100) >
8035 (sgs->group_util * env->sd->imbalance_pct))
8042 * group_is_overloaded returns true if the group has more tasks than it can
8044 * group_is_overloaded is not equals to !group_has_capacity because a group
8045 * with the exact right number of tasks, has no more spare capacity but is not
8046 * overloaded so both group_has_capacity and group_is_overloaded return
8050 group_is_overloaded(struct lb_env *env, struct sg_lb_stats *sgs)
8052 if (sgs->sum_nr_running <= sgs->group_weight)
8055 if ((sgs->group_capacity * 100) <
8056 (sgs->group_util * env->sd->imbalance_pct))
8063 * group_smaller_cpu_capacity: Returns true if sched_group sg has smaller
8064 * per-CPU capacity than sched_group ref.
8067 group_smaller_cpu_capacity(struct sched_group *sg, struct sched_group *ref)
8069 return sg->sgc->min_capacity * capacity_margin <
8070 ref->sgc->min_capacity * 1024;
8074 group_type group_classify(struct sched_group *group,
8075 struct sg_lb_stats *sgs)
8077 if (sgs->group_no_capacity)
8078 return group_overloaded;
8080 if (sg_imbalanced(group))
8081 return group_imbalanced;
8086 static bool update_nohz_stats(struct rq *rq, bool force)
8088 #ifdef CONFIG_NO_HZ_COMMON
8089 unsigned int cpu = rq->cpu;
8091 if (!rq->has_blocked_load)
8094 if (!cpumask_test_cpu(cpu, nohz.idle_cpus_mask))
8097 if (!force && !time_after(jiffies, rq->last_blocked_load_update_tick))
8100 update_blocked_averages(cpu);
8102 return rq->has_blocked_load;
8109 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
8110 * @env: The load balancing environment.
8111 * @group: sched_group whose statistics are to be updated.
8112 * @load_idx: Load index of sched_domain of this_cpu for load calc.
8113 * @local_group: Does group contain this_cpu.
8114 * @sgs: variable to hold the statistics for this group.
8115 * @overload: Indicate more than one runnable task for any CPU.
8117 static inline void update_sg_lb_stats(struct lb_env *env,
8118 struct sched_group *group, int load_idx,
8119 int local_group, struct sg_lb_stats *sgs,
8125 memset(sgs, 0, sizeof(*sgs));
8127 for_each_cpu_and(i, sched_group_span(group), env->cpus) {
8128 struct rq *rq = cpu_rq(i);
8130 if ((env->flags & LBF_NOHZ_STATS) && update_nohz_stats(rq, false))
8131 env->flags |= LBF_NOHZ_AGAIN;
8133 /* Bias balancing toward CPUs of our domain: */
8135 load = target_load(i, load_idx);
8137 load = source_load(i, load_idx);
8139 sgs->group_load += load;
8140 sgs->group_util += cpu_util(i);
8141 sgs->sum_nr_running += rq->cfs.h_nr_running;
8143 nr_running = rq->nr_running;
8147 #ifdef CONFIG_NUMA_BALANCING
8148 sgs->nr_numa_running += rq->nr_numa_running;
8149 sgs->nr_preferred_running += rq->nr_preferred_running;
8151 sgs->sum_weighted_load += weighted_cpuload(rq);
8153 * No need to call idle_cpu() if nr_running is not 0
8155 if (!nr_running && idle_cpu(i))
8159 /* Adjust by relative CPU capacity of the group */
8160 sgs->group_capacity = group->sgc->capacity;
8161 sgs->avg_load = (sgs->group_load*SCHED_CAPACITY_SCALE) / sgs->group_capacity;
8163 if (sgs->sum_nr_running)
8164 sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
8166 sgs->group_weight = group->group_weight;
8168 sgs->group_no_capacity = group_is_overloaded(env, sgs);
8169 sgs->group_type = group_classify(group, sgs);
8173 * update_sd_pick_busiest - return 1 on busiest group
8174 * @env: The load balancing environment.
8175 * @sds: sched_domain statistics
8176 * @sg: sched_group candidate to be checked for being the busiest
8177 * @sgs: sched_group statistics
8179 * Determine if @sg is a busier group than the previously selected
8182 * Return: %true if @sg is a busier group than the previously selected
8183 * busiest group. %false otherwise.
8185 static bool update_sd_pick_busiest(struct lb_env *env,
8186 struct sd_lb_stats *sds,
8187 struct sched_group *sg,
8188 struct sg_lb_stats *sgs)
8190 struct sg_lb_stats *busiest = &sds->busiest_stat;
8192 if (sgs->group_type > busiest->group_type)
8195 if (sgs->group_type < busiest->group_type)
8198 if (sgs->avg_load <= busiest->avg_load)
8201 if (!(env->sd->flags & SD_ASYM_CPUCAPACITY))
8205 * Candidate sg has no more than one task per CPU and
8206 * has higher per-CPU capacity. Migrating tasks to less
8207 * capable CPUs may harm throughput. Maximize throughput,
8208 * power/energy consequences are not considered.
8210 if (sgs->sum_nr_running <= sgs->group_weight &&
8211 group_smaller_cpu_capacity(sds->local, sg))
8215 /* This is the busiest node in its class. */
8216 if (!(env->sd->flags & SD_ASYM_PACKING))
8219 /* No ASYM_PACKING if target CPU is already busy */
8220 if (env->idle == CPU_NOT_IDLE)
8223 * ASYM_PACKING needs to move all the work to the highest
8224 * prority CPUs in the group, therefore mark all groups
8225 * of lower priority than ourself as busy.
8227 if (sgs->sum_nr_running &&
8228 sched_asym_prefer(env->dst_cpu, sg->asym_prefer_cpu)) {
8232 /* Prefer to move from lowest priority CPU's work */
8233 if (sched_asym_prefer(sds->busiest->asym_prefer_cpu,
8234 sg->asym_prefer_cpu))
8241 #ifdef CONFIG_NUMA_BALANCING
8242 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
8244 if (sgs->sum_nr_running > sgs->nr_numa_running)
8246 if (sgs->sum_nr_running > sgs->nr_preferred_running)
8251 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
8253 if (rq->nr_running > rq->nr_numa_running)
8255 if (rq->nr_running > rq->nr_preferred_running)
8260 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
8265 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
8269 #endif /* CONFIG_NUMA_BALANCING */
8272 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
8273 * @env: The load balancing environment.
8274 * @sds: variable to hold the statistics for this sched_domain.
8276 static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
8278 struct sched_domain *child = env->sd->child;
8279 struct sched_group *sg = env->sd->groups;
8280 struct sg_lb_stats *local = &sds->local_stat;
8281 struct sg_lb_stats tmp_sgs;
8282 int load_idx, prefer_sibling = 0;
8283 bool overload = false;
8285 if (child && child->flags & SD_PREFER_SIBLING)
8288 #ifdef CONFIG_NO_HZ_COMMON
8289 if (env->idle == CPU_NEWLY_IDLE && READ_ONCE(nohz.has_blocked))
8290 env->flags |= LBF_NOHZ_STATS;
8293 load_idx = get_sd_load_idx(env->sd, env->idle);
8296 struct sg_lb_stats *sgs = &tmp_sgs;
8299 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_span(sg));
8304 if (env->idle != CPU_NEWLY_IDLE ||
8305 time_after_eq(jiffies, sg->sgc->next_update))
8306 update_group_capacity(env->sd, env->dst_cpu);
8309 update_sg_lb_stats(env, sg, load_idx, local_group, sgs,
8316 * In case the child domain prefers tasks go to siblings
8317 * first, lower the sg capacity so that we'll try
8318 * and move all the excess tasks away. We lower the capacity
8319 * of a group only if the local group has the capacity to fit
8320 * these excess tasks. The extra check prevents the case where
8321 * you always pull from the heaviest group when it is already
8322 * under-utilized (possible with a large weight task outweighs
8323 * the tasks on the system).
8325 if (prefer_sibling && sds->local &&
8326 group_has_capacity(env, local) &&
8327 (sgs->sum_nr_running > local->sum_nr_running + 1)) {
8328 sgs->group_no_capacity = 1;
8329 sgs->group_type = group_classify(sg, sgs);
8332 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
8334 sds->busiest_stat = *sgs;
8338 /* Now, start updating sd_lb_stats */
8339 sds->total_running += sgs->sum_nr_running;
8340 sds->total_load += sgs->group_load;
8341 sds->total_capacity += sgs->group_capacity;
8344 } while (sg != env->sd->groups);
8346 #ifdef CONFIG_NO_HZ_COMMON
8347 if ((env->flags & LBF_NOHZ_AGAIN) &&
8348 cpumask_subset(nohz.idle_cpus_mask, sched_domain_span(env->sd))) {
8350 WRITE_ONCE(nohz.next_blocked,
8351 jiffies + msecs_to_jiffies(LOAD_AVG_PERIOD));
8355 if (env->sd->flags & SD_NUMA)
8356 env->fbq_type = fbq_classify_group(&sds->busiest_stat);
8358 if (!env->sd->parent) {
8359 /* update overload indicator if we are at root domain */
8360 if (env->dst_rq->rd->overload != overload)
8361 env->dst_rq->rd->overload = overload;
8366 * check_asym_packing - Check to see if the group is packed into the
8369 * This is primarily intended to used at the sibling level. Some
8370 * cores like POWER7 prefer to use lower numbered SMT threads. In the
8371 * case of POWER7, it can move to lower SMT modes only when higher
8372 * threads are idle. When in lower SMT modes, the threads will
8373 * perform better since they share less core resources. Hence when we
8374 * have idle threads, we want them to be the higher ones.
8376 * This packing function is run on idle threads. It checks to see if
8377 * the busiest CPU in this domain (core in the P7 case) has a higher
8378 * CPU number than the packing function is being run on. Here we are
8379 * assuming lower CPU number will be equivalent to lower a SMT thread
8382 * Return: 1 when packing is required and a task should be moved to
8383 * this CPU. The amount of the imbalance is returned in env->imbalance.
8385 * @env: The load balancing environment.
8386 * @sds: Statistics of the sched_domain which is to be packed
8388 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
8392 if (!(env->sd->flags & SD_ASYM_PACKING))
8395 if (env->idle == CPU_NOT_IDLE)
8401 busiest_cpu = sds->busiest->asym_prefer_cpu;
8402 if (sched_asym_prefer(busiest_cpu, env->dst_cpu))
8405 env->imbalance = DIV_ROUND_CLOSEST(
8406 sds->busiest_stat.avg_load * sds->busiest_stat.group_capacity,
8407 SCHED_CAPACITY_SCALE);
8413 * fix_small_imbalance - Calculate the minor imbalance that exists
8414 * amongst the groups of a sched_domain, during
8416 * @env: The load balancing environment.
8417 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
8420 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
8422 unsigned long tmp, capa_now = 0, capa_move = 0;
8423 unsigned int imbn = 2;
8424 unsigned long scaled_busy_load_per_task;
8425 struct sg_lb_stats *local, *busiest;
8427 local = &sds->local_stat;
8428 busiest = &sds->busiest_stat;
8430 if (!local->sum_nr_running)
8431 local->load_per_task = cpu_avg_load_per_task(env->dst_cpu);
8432 else if (busiest->load_per_task > local->load_per_task)
8435 scaled_busy_load_per_task =
8436 (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
8437 busiest->group_capacity;
8439 if (busiest->avg_load + scaled_busy_load_per_task >=
8440 local->avg_load + (scaled_busy_load_per_task * imbn)) {
8441 env->imbalance = busiest->load_per_task;
8446 * OK, we don't have enough imbalance to justify moving tasks,
8447 * however we may be able to increase total CPU capacity used by
8451 capa_now += busiest->group_capacity *
8452 min(busiest->load_per_task, busiest->avg_load);
8453 capa_now += local->group_capacity *
8454 min(local->load_per_task, local->avg_load);
8455 capa_now /= SCHED_CAPACITY_SCALE;
8457 /* Amount of load we'd subtract */
8458 if (busiest->avg_load > scaled_busy_load_per_task) {
8459 capa_move += busiest->group_capacity *
8460 min(busiest->load_per_task,
8461 busiest->avg_load - scaled_busy_load_per_task);
8464 /* Amount of load we'd add */
8465 if (busiest->avg_load * busiest->group_capacity <
8466 busiest->load_per_task * SCHED_CAPACITY_SCALE) {
8467 tmp = (busiest->avg_load * busiest->group_capacity) /
8468 local->group_capacity;
8470 tmp = (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
8471 local->group_capacity;
8473 capa_move += local->group_capacity *
8474 min(local->load_per_task, local->avg_load + tmp);
8475 capa_move /= SCHED_CAPACITY_SCALE;
8477 /* Move if we gain throughput */
8478 if (capa_move > capa_now)
8479 env->imbalance = busiest->load_per_task;
8483 * calculate_imbalance - Calculate the amount of imbalance present within the
8484 * groups of a given sched_domain during load balance.
8485 * @env: load balance environment
8486 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
8488 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
8490 unsigned long max_pull, load_above_capacity = ~0UL;
8491 struct sg_lb_stats *local, *busiest;
8493 local = &sds->local_stat;
8494 busiest = &sds->busiest_stat;
8496 if (busiest->group_type == group_imbalanced) {
8498 * In the group_imb case we cannot rely on group-wide averages
8499 * to ensure CPU-load equilibrium, look at wider averages. XXX
8501 busiest->load_per_task =
8502 min(busiest->load_per_task, sds->avg_load);
8506 * Avg load of busiest sg can be less and avg load of local sg can
8507 * be greater than avg load across all sgs of sd because avg load
8508 * factors in sg capacity and sgs with smaller group_type are
8509 * skipped when updating the busiest sg:
8511 if (busiest->avg_load <= sds->avg_load ||
8512 local->avg_load >= sds->avg_load) {
8514 return fix_small_imbalance(env, sds);
8518 * If there aren't any idle CPUs, avoid creating some.
8520 if (busiest->group_type == group_overloaded &&
8521 local->group_type == group_overloaded) {
8522 load_above_capacity = busiest->sum_nr_running * SCHED_CAPACITY_SCALE;
8523 if (load_above_capacity > busiest->group_capacity) {
8524 load_above_capacity -= busiest->group_capacity;
8525 load_above_capacity *= scale_load_down(NICE_0_LOAD);
8526 load_above_capacity /= busiest->group_capacity;
8528 load_above_capacity = ~0UL;
8532 * We're trying to get all the CPUs to the average_load, so we don't
8533 * want to push ourselves above the average load, nor do we wish to
8534 * reduce the max loaded CPU below the average load. At the same time,
8535 * we also don't want to reduce the group load below the group
8536 * capacity. Thus we look for the minimum possible imbalance.
8538 max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity);
8540 /* How much load to actually move to equalise the imbalance */
8541 env->imbalance = min(
8542 max_pull * busiest->group_capacity,
8543 (sds->avg_load - local->avg_load) * local->group_capacity
8544 ) / SCHED_CAPACITY_SCALE;
8547 * if *imbalance is less than the average load per runnable task
8548 * there is no guarantee that any tasks will be moved so we'll have
8549 * a think about bumping its value to force at least one task to be
8552 if (env->imbalance < busiest->load_per_task)
8553 return fix_small_imbalance(env, sds);
8556 /******* find_busiest_group() helpers end here *********************/
8559 * find_busiest_group - Returns the busiest group within the sched_domain
8560 * if there is an imbalance.
8562 * Also calculates the amount of weighted load which should be moved
8563 * to restore balance.
8565 * @env: The load balancing environment.
8567 * Return: - The busiest group if imbalance exists.
8569 static struct sched_group *find_busiest_group(struct lb_env *env)
8571 struct sg_lb_stats *local, *busiest;
8572 struct sd_lb_stats sds;
8574 init_sd_lb_stats(&sds);
8577 * Compute the various statistics relavent for load balancing at
8580 update_sd_lb_stats(env, &sds);
8581 local = &sds.local_stat;
8582 busiest = &sds.busiest_stat;
8584 /* ASYM feature bypasses nice load balance check */
8585 if (check_asym_packing(env, &sds))
8588 /* There is no busy sibling group to pull tasks from */
8589 if (!sds.busiest || busiest->sum_nr_running == 0)
8592 /* XXX broken for overlapping NUMA groups */
8593 sds.avg_load = (SCHED_CAPACITY_SCALE * sds.total_load)
8594 / sds.total_capacity;
8597 * If the busiest group is imbalanced the below checks don't
8598 * work because they assume all things are equal, which typically
8599 * isn't true due to cpus_allowed constraints and the like.
8601 if (busiest->group_type == group_imbalanced)
8605 * When dst_cpu is idle, prevent SMP nice and/or asymmetric group
8606 * capacities from resulting in underutilization due to avg_load.
8608 if (env->idle != CPU_NOT_IDLE && group_has_capacity(env, local) &&
8609 busiest->group_no_capacity)
8613 * If the local group is busier than the selected busiest group
8614 * don't try and pull any tasks.
8616 if (local->avg_load >= busiest->avg_load)
8620 * Don't pull any tasks if this group is already above the domain
8623 if (local->avg_load >= sds.avg_load)
8626 if (env->idle == CPU_IDLE) {
8628 * This CPU is idle. If the busiest group is not overloaded
8629 * and there is no imbalance between this and busiest group
8630 * wrt idle CPUs, it is balanced. The imbalance becomes
8631 * significant if the diff is greater than 1 otherwise we
8632 * might end up to just move the imbalance on another group
8634 if ((busiest->group_type != group_overloaded) &&
8635 (local->idle_cpus <= (busiest->idle_cpus + 1)))
8639 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
8640 * imbalance_pct to be conservative.
8642 if (100 * busiest->avg_load <=
8643 env->sd->imbalance_pct * local->avg_load)
8648 /* Looks like there is an imbalance. Compute it */
8649 calculate_imbalance(env, &sds);
8658 * find_busiest_queue - find the busiest runqueue among the CPUs in the group.
8660 static struct rq *find_busiest_queue(struct lb_env *env,
8661 struct sched_group *group)
8663 struct rq *busiest = NULL, *rq;
8664 unsigned long busiest_load = 0, busiest_capacity = 1;
8667 for_each_cpu_and(i, sched_group_span(group), env->cpus) {
8668 unsigned long capacity, wl;
8672 rt = fbq_classify_rq(rq);
8675 * We classify groups/runqueues into three groups:
8676 * - regular: there are !numa tasks
8677 * - remote: there are numa tasks that run on the 'wrong' node
8678 * - all: there is no distinction
8680 * In order to avoid migrating ideally placed numa tasks,
8681 * ignore those when there's better options.
8683 * If we ignore the actual busiest queue to migrate another
8684 * task, the next balance pass can still reduce the busiest
8685 * queue by moving tasks around inside the node.
8687 * If we cannot move enough load due to this classification
8688 * the next pass will adjust the group classification and
8689 * allow migration of more tasks.
8691 * Both cases only affect the total convergence complexity.
8693 if (rt > env->fbq_type)
8696 capacity = capacity_of(i);
8698 wl = weighted_cpuload(rq);
8701 * When comparing with imbalance, use weighted_cpuload()
8702 * which is not scaled with the CPU capacity.
8705 if (rq->nr_running == 1 && wl > env->imbalance &&
8706 !check_cpu_capacity(rq, env->sd))
8710 * For the load comparisons with the other CPU's, consider
8711 * the weighted_cpuload() scaled with the CPU capacity, so
8712 * that the load can be moved away from the CPU that is
8713 * potentially running at a lower capacity.
8715 * Thus we're looking for max(wl_i / capacity_i), crosswise
8716 * multiplication to rid ourselves of the division works out
8717 * to: wl_i * capacity_j > wl_j * capacity_i; where j is
8718 * our previous maximum.
8720 if (wl * busiest_capacity > busiest_load * capacity) {
8722 busiest_capacity = capacity;
8731 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
8732 * so long as it is large enough.
8734 #define MAX_PINNED_INTERVAL 512
8736 static int need_active_balance(struct lb_env *env)
8738 struct sched_domain *sd = env->sd;
8740 if (env->idle == CPU_NEWLY_IDLE) {
8743 * ASYM_PACKING needs to force migrate tasks from busy but
8744 * lower priority CPUs in order to pack all tasks in the
8745 * highest priority CPUs.
8747 if ((sd->flags & SD_ASYM_PACKING) &&
8748 sched_asym_prefer(env->dst_cpu, env->src_cpu))
8753 * The dst_cpu is idle and the src_cpu CPU has only 1 CFS task.
8754 * It's worth migrating the task if the src_cpu's capacity is reduced
8755 * because of other sched_class or IRQs if more capacity stays
8756 * available on dst_cpu.
8758 if ((env->idle != CPU_NOT_IDLE) &&
8759 (env->src_rq->cfs.h_nr_running == 1)) {
8760 if ((check_cpu_capacity(env->src_rq, sd)) &&
8761 (capacity_of(env->src_cpu)*sd->imbalance_pct < capacity_of(env->dst_cpu)*100))
8765 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
8768 static int active_load_balance_cpu_stop(void *data);
8770 static int should_we_balance(struct lb_env *env)
8772 struct sched_group *sg = env->sd->groups;
8773 int cpu, balance_cpu = -1;
8776 * Ensure the balancing environment is consistent; can happen
8777 * when the softirq triggers 'during' hotplug.
8779 if (!cpumask_test_cpu(env->dst_cpu, env->cpus))
8783 * In the newly idle case, we will allow all the CPUs
8784 * to do the newly idle load balance.
8786 if (env->idle == CPU_NEWLY_IDLE)
8789 /* Try to find first idle CPU */
8790 for_each_cpu_and(cpu, group_balance_mask(sg), env->cpus) {
8798 if (balance_cpu == -1)
8799 balance_cpu = group_balance_cpu(sg);
8802 * First idle CPU or the first CPU(busiest) in this sched group
8803 * is eligible for doing load balancing at this and above domains.
8805 return balance_cpu == env->dst_cpu;
8809 * Check this_cpu to ensure it is balanced within domain. Attempt to move
8810 * tasks if there is an imbalance.
8812 static int load_balance(int this_cpu, struct rq *this_rq,
8813 struct sched_domain *sd, enum cpu_idle_type idle,
8814 int *continue_balancing)
8816 int ld_moved, cur_ld_moved, active_balance = 0;
8817 struct sched_domain *sd_parent = sd->parent;
8818 struct sched_group *group;
8821 struct cpumask *cpus = this_cpu_cpumask_var_ptr(load_balance_mask);
8823 struct lb_env env = {
8825 .dst_cpu = this_cpu,
8827 .dst_grpmask = sched_group_span(sd->groups),
8829 .loop_break = sched_nr_migrate_break,
8832 .tasks = LIST_HEAD_INIT(env.tasks),
8835 cpumask_and(cpus, sched_domain_span(sd), cpu_active_mask);
8837 schedstat_inc(sd->lb_count[idle]);
8840 if (!should_we_balance(&env)) {
8841 *continue_balancing = 0;
8845 group = find_busiest_group(&env);
8847 schedstat_inc(sd->lb_nobusyg[idle]);
8851 busiest = find_busiest_queue(&env, group);
8853 schedstat_inc(sd->lb_nobusyq[idle]);
8857 BUG_ON(busiest == env.dst_rq);
8859 schedstat_add(sd->lb_imbalance[idle], env.imbalance);
8861 env.src_cpu = busiest->cpu;
8862 env.src_rq = busiest;
8865 if (busiest->nr_running > 1) {
8867 * Attempt to move tasks. If find_busiest_group has found
8868 * an imbalance but busiest->nr_running <= 1, the group is
8869 * still unbalanced. ld_moved simply stays zero, so it is
8870 * correctly treated as an imbalance.
8872 env.flags |= LBF_ALL_PINNED;
8873 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
8876 rq_lock_irqsave(busiest, &rf);
8877 update_rq_clock(busiest);
8880 * cur_ld_moved - load moved in current iteration
8881 * ld_moved - cumulative load moved across iterations
8883 cur_ld_moved = detach_tasks(&env);
8886 * We've detached some tasks from busiest_rq. Every
8887 * task is masked "TASK_ON_RQ_MIGRATING", so we can safely
8888 * unlock busiest->lock, and we are able to be sure
8889 * that nobody can manipulate the tasks in parallel.
8890 * See task_rq_lock() family for the details.
8893 rq_unlock(busiest, &rf);
8897 ld_moved += cur_ld_moved;
8900 local_irq_restore(rf.flags);
8902 if (env.flags & LBF_NEED_BREAK) {
8903 env.flags &= ~LBF_NEED_BREAK;
8908 * Revisit (affine) tasks on src_cpu that couldn't be moved to
8909 * us and move them to an alternate dst_cpu in our sched_group
8910 * where they can run. The upper limit on how many times we
8911 * iterate on same src_cpu is dependent on number of CPUs in our
8914 * This changes load balance semantics a bit on who can move
8915 * load to a given_cpu. In addition to the given_cpu itself
8916 * (or a ilb_cpu acting on its behalf where given_cpu is
8917 * nohz-idle), we now have balance_cpu in a position to move
8918 * load to given_cpu. In rare situations, this may cause
8919 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
8920 * _independently_ and at _same_ time to move some load to
8921 * given_cpu) causing exceess load to be moved to given_cpu.
8922 * This however should not happen so much in practice and
8923 * moreover subsequent load balance cycles should correct the
8924 * excess load moved.
8926 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
8928 /* Prevent to re-select dst_cpu via env's CPUs */
8929 cpumask_clear_cpu(env.dst_cpu, env.cpus);
8931 env.dst_rq = cpu_rq(env.new_dst_cpu);
8932 env.dst_cpu = env.new_dst_cpu;
8933 env.flags &= ~LBF_DST_PINNED;
8935 env.loop_break = sched_nr_migrate_break;
8938 * Go back to "more_balance" rather than "redo" since we
8939 * need to continue with same src_cpu.
8945 * We failed to reach balance because of affinity.
8948 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
8950 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0)
8951 *group_imbalance = 1;
8954 /* All tasks on this runqueue were pinned by CPU affinity */
8955 if (unlikely(env.flags & LBF_ALL_PINNED)) {
8956 cpumask_clear_cpu(cpu_of(busiest), cpus);
8958 * Attempting to continue load balancing at the current
8959 * sched_domain level only makes sense if there are
8960 * active CPUs remaining as possible busiest CPUs to
8961 * pull load from which are not contained within the
8962 * destination group that is receiving any migrated
8965 if (!cpumask_subset(cpus, env.dst_grpmask)) {
8967 env.loop_break = sched_nr_migrate_break;
8970 goto out_all_pinned;
8975 schedstat_inc(sd->lb_failed[idle]);
8977 * Increment the failure counter only on periodic balance.
8978 * We do not want newidle balance, which can be very
8979 * frequent, pollute the failure counter causing
8980 * excessive cache_hot migrations and active balances.
8982 if (idle != CPU_NEWLY_IDLE)
8983 sd->nr_balance_failed++;
8985 if (need_active_balance(&env)) {
8986 unsigned long flags;
8988 raw_spin_lock_irqsave(&busiest->lock, flags);
8991 * Don't kick the active_load_balance_cpu_stop,
8992 * if the curr task on busiest CPU can't be
8993 * moved to this_cpu:
8995 if (!cpumask_test_cpu(this_cpu, &busiest->curr->cpus_allowed)) {
8996 raw_spin_unlock_irqrestore(&busiest->lock,
8998 env.flags |= LBF_ALL_PINNED;
8999 goto out_one_pinned;
9003 * ->active_balance synchronizes accesses to
9004 * ->active_balance_work. Once set, it's cleared
9005 * only after active load balance is finished.
9007 if (!busiest->active_balance) {
9008 busiest->active_balance = 1;
9009 busiest->push_cpu = this_cpu;
9012 raw_spin_unlock_irqrestore(&busiest->lock, flags);
9014 if (active_balance) {
9015 stop_one_cpu_nowait(cpu_of(busiest),
9016 active_load_balance_cpu_stop, busiest,
9017 &busiest->active_balance_work);
9020 /* We've kicked active balancing, force task migration. */
9021 sd->nr_balance_failed = sd->cache_nice_tries+1;
9024 sd->nr_balance_failed = 0;
9026 if (likely(!active_balance)) {
9027 /* We were unbalanced, so reset the balancing interval */
9028 sd->balance_interval = sd->min_interval;
9031 * If we've begun active balancing, start to back off. This
9032 * case may not be covered by the all_pinned logic if there
9033 * is only 1 task on the busy runqueue (because we don't call
9036 if (sd->balance_interval < sd->max_interval)
9037 sd->balance_interval *= 2;
9044 * We reach balance although we may have faced some affinity
9045 * constraints. Clear the imbalance flag if it was set.
9048 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
9050 if (*group_imbalance)
9051 *group_imbalance = 0;
9056 * We reach balance because all tasks are pinned at this level so
9057 * we can't migrate them. Let the imbalance flag set so parent level
9058 * can try to migrate them.
9060 schedstat_inc(sd->lb_balanced[idle]);
9062 sd->nr_balance_failed = 0;
9065 /* tune up the balancing interval */
9066 if (((env.flags & LBF_ALL_PINNED) &&
9067 sd->balance_interval < MAX_PINNED_INTERVAL) ||
9068 (sd->balance_interval < sd->max_interval))
9069 sd->balance_interval *= 2;
9076 static inline unsigned long
9077 get_sd_balance_interval(struct sched_domain *sd, int cpu_busy)
9079 unsigned long interval = sd->balance_interval;
9082 interval *= sd->busy_factor;
9084 /* scale ms to jiffies */
9085 interval = msecs_to_jiffies(interval);
9086 interval = clamp(interval, 1UL, max_load_balance_interval);
9092 update_next_balance(struct sched_domain *sd, unsigned long *next_balance)
9094 unsigned long interval, next;
9096 /* used by idle balance, so cpu_busy = 0 */
9097 interval = get_sd_balance_interval(sd, 0);
9098 next = sd->last_balance + interval;
9100 if (time_after(*next_balance, next))
9101 *next_balance = next;
9105 * active_load_balance_cpu_stop is run by the CPU stopper. It pushes
9106 * running tasks off the busiest CPU onto idle CPUs. It requires at
9107 * least 1 task to be running on each physical CPU where possible, and
9108 * avoids physical / logical imbalances.
9110 static int active_load_balance_cpu_stop(void *data)
9112 struct rq *busiest_rq = data;
9113 int busiest_cpu = cpu_of(busiest_rq);
9114 int target_cpu = busiest_rq->push_cpu;
9115 struct rq *target_rq = cpu_rq(target_cpu);
9116 struct sched_domain *sd;
9117 struct task_struct *p = NULL;
9120 rq_lock_irq(busiest_rq, &rf);
9122 * Between queueing the stop-work and running it is a hole in which
9123 * CPUs can become inactive. We should not move tasks from or to
9126 if (!cpu_active(busiest_cpu) || !cpu_active(target_cpu))
9129 /* Make sure the requested CPU hasn't gone down in the meantime: */
9130 if (unlikely(busiest_cpu != smp_processor_id() ||
9131 !busiest_rq->active_balance))
9134 /* Is there any task to move? */
9135 if (busiest_rq->nr_running <= 1)
9139 * This condition is "impossible", if it occurs
9140 * we need to fix it. Originally reported by
9141 * Bjorn Helgaas on a 128-CPU setup.
9143 BUG_ON(busiest_rq == target_rq);
9145 /* Search for an sd spanning us and the target CPU. */
9147 for_each_domain(target_cpu, sd) {
9148 if ((sd->flags & SD_LOAD_BALANCE) &&
9149 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
9154 struct lb_env env = {
9156 .dst_cpu = target_cpu,
9157 .dst_rq = target_rq,
9158 .src_cpu = busiest_rq->cpu,
9159 .src_rq = busiest_rq,
9162 * can_migrate_task() doesn't need to compute new_dst_cpu
9163 * for active balancing. Since we have CPU_IDLE, but no
9164 * @dst_grpmask we need to make that test go away with lying
9167 .flags = LBF_DST_PINNED,
9170 schedstat_inc(sd->alb_count);
9171 update_rq_clock(busiest_rq);
9173 p = detach_one_task(&env);
9175 schedstat_inc(sd->alb_pushed);
9176 /* Active balancing done, reset the failure counter. */
9177 sd->nr_balance_failed = 0;
9179 schedstat_inc(sd->alb_failed);
9184 busiest_rq->active_balance = 0;
9185 rq_unlock(busiest_rq, &rf);
9188 attach_one_task(target_rq, p);
9195 static DEFINE_SPINLOCK(balancing);
9198 * Scale the max load_balance interval with the number of CPUs in the system.
9199 * This trades load-balance latency on larger machines for less cross talk.
9201 void update_max_interval(void)
9203 max_load_balance_interval = HZ*num_online_cpus()/10;
9207 * It checks each scheduling domain to see if it is due to be balanced,
9208 * and initiates a balancing operation if so.
9210 * Balancing parameters are set up in init_sched_domains.
9212 static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle)
9214 int continue_balancing = 1;
9216 unsigned long interval;
9217 struct sched_domain *sd;
9218 /* Earliest time when we have to do rebalance again */
9219 unsigned long next_balance = jiffies + 60*HZ;
9220 int update_next_balance = 0;
9221 int need_serialize, need_decay = 0;
9225 for_each_domain(cpu, sd) {
9227 * Decay the newidle max times here because this is a regular
9228 * visit to all the domains. Decay ~1% per second.
9230 if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
9231 sd->max_newidle_lb_cost =
9232 (sd->max_newidle_lb_cost * 253) / 256;
9233 sd->next_decay_max_lb_cost = jiffies + HZ;
9236 max_cost += sd->max_newidle_lb_cost;
9238 if (!(sd->flags & SD_LOAD_BALANCE))
9242 * Stop the load balance at this level. There is another
9243 * CPU in our sched group which is doing load balancing more
9246 if (!continue_balancing) {
9252 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
9254 need_serialize = sd->flags & SD_SERIALIZE;
9255 if (need_serialize) {
9256 if (!spin_trylock(&balancing))
9260 if (time_after_eq(jiffies, sd->last_balance + interval)) {
9261 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
9263 * The LBF_DST_PINNED logic could have changed
9264 * env->dst_cpu, so we can't know our idle
9265 * state even if we migrated tasks. Update it.
9267 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
9269 sd->last_balance = jiffies;
9270 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
9273 spin_unlock(&balancing);
9275 if (time_after(next_balance, sd->last_balance + interval)) {
9276 next_balance = sd->last_balance + interval;
9277 update_next_balance = 1;
9282 * Ensure the rq-wide value also decays but keep it at a
9283 * reasonable floor to avoid funnies with rq->avg_idle.
9285 rq->max_idle_balance_cost =
9286 max((u64)sysctl_sched_migration_cost, max_cost);
9291 * next_balance will be updated only when there is a need.
9292 * When the cpu is attached to null domain for ex, it will not be
9295 if (likely(update_next_balance)) {
9296 rq->next_balance = next_balance;
9298 #ifdef CONFIG_NO_HZ_COMMON
9300 * If this CPU has been elected to perform the nohz idle
9301 * balance. Other idle CPUs have already rebalanced with
9302 * nohz_idle_balance() and nohz.next_balance has been
9303 * updated accordingly. This CPU is now running the idle load
9304 * balance for itself and we need to update the
9305 * nohz.next_balance accordingly.
9307 if ((idle == CPU_IDLE) && time_after(nohz.next_balance, rq->next_balance))
9308 nohz.next_balance = rq->next_balance;
9313 static inline int on_null_domain(struct rq *rq)
9315 return unlikely(!rcu_dereference_sched(rq->sd));
9318 #ifdef CONFIG_NO_HZ_COMMON
9320 * idle load balancing details
9321 * - When one of the busy CPUs notice that there may be an idle rebalancing
9322 * needed, they will kick the idle load balancer, which then does idle
9323 * load balancing for all the idle CPUs.
9326 static inline int find_new_ilb(void)
9328 int ilb = cpumask_first(nohz.idle_cpus_mask);
9330 if (ilb < nr_cpu_ids && idle_cpu(ilb))
9337 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
9338 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
9339 * CPU (if there is one).
9341 static void kick_ilb(unsigned int flags)
9345 nohz.next_balance++;
9347 ilb_cpu = find_new_ilb();
9349 if (ilb_cpu >= nr_cpu_ids)
9352 flags = atomic_fetch_or(flags, nohz_flags(ilb_cpu));
9353 if (flags & NOHZ_KICK_MASK)
9357 * Use smp_send_reschedule() instead of resched_cpu().
9358 * This way we generate a sched IPI on the target CPU which
9359 * is idle. And the softirq performing nohz idle load balance
9360 * will be run before returning from the IPI.
9362 smp_send_reschedule(ilb_cpu);
9366 * Current heuristic for kicking the idle load balancer in the presence
9367 * of an idle cpu in the system.
9368 * - This rq has more than one task.
9369 * - This rq has at least one CFS task and the capacity of the CPU is
9370 * significantly reduced because of RT tasks or IRQs.
9371 * - At parent of LLC scheduler domain level, this cpu's scheduler group has
9372 * multiple busy cpu.
9373 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
9374 * domain span are idle.
9376 static void nohz_balancer_kick(struct rq *rq)
9378 unsigned long now = jiffies;
9379 struct sched_domain_shared *sds;
9380 struct sched_domain *sd;
9381 int nr_busy, i, cpu = rq->cpu;
9382 unsigned int flags = 0;
9384 if (unlikely(rq->idle_balance))
9388 * We may be recently in ticked or tickless idle mode. At the first
9389 * busy tick after returning from idle, we will update the busy stats.
9391 nohz_balance_exit_idle(rq);
9394 * None are in tickless mode and hence no need for NOHZ idle load
9397 if (likely(!atomic_read(&nohz.nr_cpus)))
9400 if (READ_ONCE(nohz.has_blocked) &&
9401 time_after(now, READ_ONCE(nohz.next_blocked)))
9402 flags = NOHZ_STATS_KICK;
9404 if (time_before(now, nohz.next_balance))
9407 if (rq->nr_running >= 2) {
9408 flags = NOHZ_KICK_MASK;
9413 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
9416 * XXX: write a coherent comment on why we do this.
9417 * See also: http://lkml.kernel.org/r/20111202010832.602203411@sbsiddha-desk.sc.intel.com
9419 nr_busy = atomic_read(&sds->nr_busy_cpus);
9421 flags = NOHZ_KICK_MASK;
9427 sd = rcu_dereference(rq->sd);
9429 if ((rq->cfs.h_nr_running >= 1) &&
9430 check_cpu_capacity(rq, sd)) {
9431 flags = NOHZ_KICK_MASK;
9436 sd = rcu_dereference(per_cpu(sd_asym, cpu));
9438 for_each_cpu(i, sched_domain_span(sd)) {
9440 !cpumask_test_cpu(i, nohz.idle_cpus_mask))
9443 if (sched_asym_prefer(i, cpu)) {
9444 flags = NOHZ_KICK_MASK;
9456 static void set_cpu_sd_state_busy(int cpu)
9458 struct sched_domain *sd;
9461 sd = rcu_dereference(per_cpu(sd_llc, cpu));
9463 if (!sd || !sd->nohz_idle)
9467 atomic_inc(&sd->shared->nr_busy_cpus);
9472 void nohz_balance_exit_idle(struct rq *rq)
9474 SCHED_WARN_ON(rq != this_rq());
9476 if (likely(!rq->nohz_tick_stopped))
9479 rq->nohz_tick_stopped = 0;
9480 cpumask_clear_cpu(rq->cpu, nohz.idle_cpus_mask);
9481 atomic_dec(&nohz.nr_cpus);
9483 set_cpu_sd_state_busy(rq->cpu);
9486 static void set_cpu_sd_state_idle(int cpu)
9488 struct sched_domain *sd;
9491 sd = rcu_dereference(per_cpu(sd_llc, cpu));
9493 if (!sd || sd->nohz_idle)
9497 atomic_dec(&sd->shared->nr_busy_cpus);
9503 * This routine will record that the CPU is going idle with tick stopped.
9504 * This info will be used in performing idle load balancing in the future.
9506 void nohz_balance_enter_idle(int cpu)
9508 struct rq *rq = cpu_rq(cpu);
9510 SCHED_WARN_ON(cpu != smp_processor_id());
9512 /* If this CPU is going down, then nothing needs to be done: */
9513 if (!cpu_active(cpu))
9516 /* Spare idle load balancing on CPUs that don't want to be disturbed: */
9517 if (!housekeeping_cpu(cpu, HK_FLAG_SCHED))
9521 * Can be set safely without rq->lock held
9522 * If a clear happens, it will have evaluated last additions because
9523 * rq->lock is held during the check and the clear
9525 rq->has_blocked_load = 1;
9528 * The tick is still stopped but load could have been added in the
9529 * meantime. We set the nohz.has_blocked flag to trig a check of the
9530 * *_avg. The CPU is already part of nohz.idle_cpus_mask so the clear
9531 * of nohz.has_blocked can only happen after checking the new load
9533 if (rq->nohz_tick_stopped)
9536 /* If we're a completely isolated CPU, we don't play: */
9537 if (on_null_domain(rq))
9540 rq->nohz_tick_stopped = 1;
9542 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
9543 atomic_inc(&nohz.nr_cpus);
9546 * Ensures that if nohz_idle_balance() fails to observe our
9547 * @idle_cpus_mask store, it must observe the @has_blocked
9550 smp_mb__after_atomic();
9552 set_cpu_sd_state_idle(cpu);
9556 * Each time a cpu enter idle, we assume that it has blocked load and
9557 * enable the periodic update of the load of idle cpus
9559 WRITE_ONCE(nohz.has_blocked, 1);
9563 * Internal function that runs load balance for all idle cpus. The load balance
9564 * can be a simple update of blocked load or a complete load balance with
9565 * tasks movement depending of flags.
9566 * The function returns false if the loop has stopped before running
9567 * through all idle CPUs.
9569 static bool _nohz_idle_balance(struct rq *this_rq, unsigned int flags,
9570 enum cpu_idle_type idle)
9572 /* Earliest time when we have to do rebalance again */
9573 unsigned long now = jiffies;
9574 unsigned long next_balance = now + 60*HZ;
9575 bool has_blocked_load = false;
9576 int update_next_balance = 0;
9577 int this_cpu = this_rq->cpu;
9582 SCHED_WARN_ON((flags & NOHZ_KICK_MASK) == NOHZ_BALANCE_KICK);
9585 * We assume there will be no idle load after this update and clear
9586 * the has_blocked flag. If a cpu enters idle in the mean time, it will
9587 * set the has_blocked flag and trig another update of idle load.
9588 * Because a cpu that becomes idle, is added to idle_cpus_mask before
9589 * setting the flag, we are sure to not clear the state and not
9590 * check the load of an idle cpu.
9592 WRITE_ONCE(nohz.has_blocked, 0);
9595 * Ensures that if we miss the CPU, we must see the has_blocked
9596 * store from nohz_balance_enter_idle().
9600 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
9601 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
9605 * If this CPU gets work to do, stop the load balancing
9606 * work being done for other CPUs. Next load
9607 * balancing owner will pick it up.
9609 if (need_resched()) {
9610 has_blocked_load = true;
9614 rq = cpu_rq(balance_cpu);
9616 has_blocked_load |= update_nohz_stats(rq, true);
9619 * If time for next balance is due,
9622 if (time_after_eq(jiffies, rq->next_balance)) {
9625 rq_lock_irqsave(rq, &rf);
9626 update_rq_clock(rq);
9627 cpu_load_update_idle(rq);
9628 rq_unlock_irqrestore(rq, &rf);
9630 if (flags & NOHZ_BALANCE_KICK)
9631 rebalance_domains(rq, CPU_IDLE);
9634 if (time_after(next_balance, rq->next_balance)) {
9635 next_balance = rq->next_balance;
9636 update_next_balance = 1;
9640 /* Newly idle CPU doesn't need an update */
9641 if (idle != CPU_NEWLY_IDLE) {
9642 update_blocked_averages(this_cpu);
9643 has_blocked_load |= this_rq->has_blocked_load;
9646 if (flags & NOHZ_BALANCE_KICK)
9647 rebalance_domains(this_rq, CPU_IDLE);
9649 WRITE_ONCE(nohz.next_blocked,
9650 now + msecs_to_jiffies(LOAD_AVG_PERIOD));
9652 /* The full idle balance loop has been done */
9656 /* There is still blocked load, enable periodic update */
9657 if (has_blocked_load)
9658 WRITE_ONCE(nohz.has_blocked, 1);
9661 * next_balance will be updated only when there is a need.
9662 * When the CPU is attached to null domain for ex, it will not be
9665 if (likely(update_next_balance))
9666 nohz.next_balance = next_balance;
9672 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
9673 * rebalancing for all the cpus for whom scheduler ticks are stopped.
9675 static bool nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
9677 int this_cpu = this_rq->cpu;
9680 if (!(atomic_read(nohz_flags(this_cpu)) & NOHZ_KICK_MASK))
9683 if (idle != CPU_IDLE) {
9684 atomic_andnot(NOHZ_KICK_MASK, nohz_flags(this_cpu));
9689 * barrier, pairs with nohz_balance_enter_idle(), ensures ...
9691 flags = atomic_fetch_andnot(NOHZ_KICK_MASK, nohz_flags(this_cpu));
9692 if (!(flags & NOHZ_KICK_MASK))
9695 _nohz_idle_balance(this_rq, flags, idle);
9700 static void nohz_newidle_balance(struct rq *this_rq)
9702 int this_cpu = this_rq->cpu;
9705 * This CPU doesn't want to be disturbed by scheduler
9708 if (!housekeeping_cpu(this_cpu, HK_FLAG_SCHED))
9711 /* Will wake up very soon. No time for doing anything else*/
9712 if (this_rq->avg_idle < sysctl_sched_migration_cost)
9715 /* Don't need to update blocked load of idle CPUs*/
9716 if (!READ_ONCE(nohz.has_blocked) ||
9717 time_before(jiffies, READ_ONCE(nohz.next_blocked)))
9720 raw_spin_unlock(&this_rq->lock);
9722 * This CPU is going to be idle and blocked load of idle CPUs
9723 * need to be updated. Run the ilb locally as it is a good
9724 * candidate for ilb instead of waking up another idle CPU.
9725 * Kick an normal ilb if we failed to do the update.
9727 if (!_nohz_idle_balance(this_rq, NOHZ_STATS_KICK, CPU_NEWLY_IDLE))
9728 kick_ilb(NOHZ_STATS_KICK);
9729 raw_spin_lock(&this_rq->lock);
9732 #else /* !CONFIG_NO_HZ_COMMON */
9733 static inline void nohz_balancer_kick(struct rq *rq) { }
9735 static inline bool nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
9740 static inline void nohz_newidle_balance(struct rq *this_rq) { }
9741 #endif /* CONFIG_NO_HZ_COMMON */
9744 * idle_balance is called by schedule() if this_cpu is about to become
9745 * idle. Attempts to pull tasks from other CPUs.
9747 static int idle_balance(struct rq *this_rq, struct rq_flags *rf)
9749 unsigned long next_balance = jiffies + HZ;
9750 int this_cpu = this_rq->cpu;
9751 struct sched_domain *sd;
9752 int pulled_task = 0;
9756 * We must set idle_stamp _before_ calling idle_balance(), such that we
9757 * measure the duration of idle_balance() as idle time.
9759 this_rq->idle_stamp = rq_clock(this_rq);
9762 * Do not pull tasks towards !active CPUs...
9764 if (!cpu_active(this_cpu))
9768 * This is OK, because current is on_cpu, which avoids it being picked
9769 * for load-balance and preemption/IRQs are still disabled avoiding
9770 * further scheduler activity on it and we're being very careful to
9771 * re-start the picking loop.
9773 rq_unpin_lock(this_rq, rf);
9775 if (this_rq->avg_idle < sysctl_sched_migration_cost ||
9776 !this_rq->rd->overload) {
9779 sd = rcu_dereference_check_sched_domain(this_rq->sd);
9781 update_next_balance(sd, &next_balance);
9784 nohz_newidle_balance(this_rq);
9789 raw_spin_unlock(&this_rq->lock);
9791 update_blocked_averages(this_cpu);
9793 for_each_domain(this_cpu, sd) {
9794 int continue_balancing = 1;
9795 u64 t0, domain_cost;
9797 if (!(sd->flags & SD_LOAD_BALANCE))
9800 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost) {
9801 update_next_balance(sd, &next_balance);
9805 if (sd->flags & SD_BALANCE_NEWIDLE) {
9806 t0 = sched_clock_cpu(this_cpu);
9808 pulled_task = load_balance(this_cpu, this_rq,
9810 &continue_balancing);
9812 domain_cost = sched_clock_cpu(this_cpu) - t0;
9813 if (domain_cost > sd->max_newidle_lb_cost)
9814 sd->max_newidle_lb_cost = domain_cost;
9816 curr_cost += domain_cost;
9819 update_next_balance(sd, &next_balance);
9822 * Stop searching for tasks to pull if there are
9823 * now runnable tasks on this rq.
9825 if (pulled_task || this_rq->nr_running > 0)
9830 raw_spin_lock(&this_rq->lock);
9832 if (curr_cost > this_rq->max_idle_balance_cost)
9833 this_rq->max_idle_balance_cost = curr_cost;
9837 * While browsing the domains, we released the rq lock, a task could
9838 * have been enqueued in the meantime. Since we're not going idle,
9839 * pretend we pulled a task.
9841 if (this_rq->cfs.h_nr_running && !pulled_task)
9844 /* Move the next balance forward */
9845 if (time_after(this_rq->next_balance, next_balance))
9846 this_rq->next_balance = next_balance;
9848 /* Is there a task of a high priority class? */
9849 if (this_rq->nr_running != this_rq->cfs.h_nr_running)
9853 this_rq->idle_stamp = 0;
9855 rq_repin_lock(this_rq, rf);
9861 * run_rebalance_domains is triggered when needed from the scheduler tick.
9862 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
9864 static __latent_entropy void run_rebalance_domains(struct softirq_action *h)
9866 struct rq *this_rq = this_rq();
9867 enum cpu_idle_type idle = this_rq->idle_balance ?
9868 CPU_IDLE : CPU_NOT_IDLE;
9871 * If this CPU has a pending nohz_balance_kick, then do the
9872 * balancing on behalf of the other idle CPUs whose ticks are
9873 * stopped. Do nohz_idle_balance *before* rebalance_domains to
9874 * give the idle CPUs a chance to load balance. Else we may
9875 * load balance only within the local sched_domain hierarchy
9876 * and abort nohz_idle_balance altogether if we pull some load.
9878 if (nohz_idle_balance(this_rq, idle))
9881 /* normal load balance */
9882 update_blocked_averages(this_rq->cpu);
9883 rebalance_domains(this_rq, idle);
9887 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
9889 void trigger_load_balance(struct rq *rq)
9891 /* Don't need to rebalance while attached to NULL domain */
9892 if (unlikely(on_null_domain(rq)))
9895 if (time_after_eq(jiffies, rq->next_balance))
9896 raise_softirq(SCHED_SOFTIRQ);
9898 nohz_balancer_kick(rq);
9901 static void rq_online_fair(struct rq *rq)
9905 update_runtime_enabled(rq);
9908 static void rq_offline_fair(struct rq *rq)
9912 /* Ensure any throttled groups are reachable by pick_next_task */
9913 unthrottle_offline_cfs_rqs(rq);
9916 #endif /* CONFIG_SMP */
9919 * scheduler tick hitting a task of our scheduling class.
9921 * NOTE: This function can be called remotely by the tick offload that
9922 * goes along full dynticks. Therefore no local assumption can be made
9923 * and everything must be accessed through the @rq and @curr passed in
9926 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
9928 struct cfs_rq *cfs_rq;
9929 struct sched_entity *se = &curr->se;
9931 for_each_sched_entity(se) {
9932 cfs_rq = cfs_rq_of(se);
9933 entity_tick(cfs_rq, se, queued);
9936 if (static_branch_unlikely(&sched_numa_balancing))
9937 task_tick_numa(rq, curr);
9941 * called on fork with the child task as argument from the parent's context
9942 * - child not yet on the tasklist
9943 * - preemption disabled
9945 static void task_fork_fair(struct task_struct *p)
9947 struct cfs_rq *cfs_rq;
9948 struct sched_entity *se = &p->se, *curr;
9949 struct rq *rq = this_rq();
9953 update_rq_clock(rq);
9955 cfs_rq = task_cfs_rq(current);
9956 curr = cfs_rq->curr;
9958 update_curr(cfs_rq);
9959 se->vruntime = curr->vruntime;
9961 place_entity(cfs_rq, se, 1);
9963 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
9965 * Upon rescheduling, sched_class::put_prev_task() will place
9966 * 'current' within the tree based on its new key value.
9968 swap(curr->vruntime, se->vruntime);
9972 se->vruntime -= cfs_rq->min_vruntime;
9977 * Priority of the task has changed. Check to see if we preempt
9981 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
9983 if (!task_on_rq_queued(p))
9987 * Reschedule if we are currently running on this runqueue and
9988 * our priority decreased, or if we are not currently running on
9989 * this runqueue and our priority is higher than the current's
9991 if (rq->curr == p) {
9992 if (p->prio > oldprio)
9995 check_preempt_curr(rq, p, 0);
9998 static inline bool vruntime_normalized(struct task_struct *p)
10000 struct sched_entity *se = &p->se;
10003 * In both the TASK_ON_RQ_QUEUED and TASK_ON_RQ_MIGRATING cases,
10004 * the dequeue_entity(.flags=0) will already have normalized the
10011 * When !on_rq, vruntime of the task has usually NOT been normalized.
10012 * But there are some cases where it has already been normalized:
10014 * - A forked child which is waiting for being woken up by
10015 * wake_up_new_task().
10016 * - A task which has been woken up by try_to_wake_up() and
10017 * waiting for actually being woken up by sched_ttwu_pending().
10019 if (!se->sum_exec_runtime || p->state == TASK_WAKING)
10025 #ifdef CONFIG_FAIR_GROUP_SCHED
10027 * Propagate the changes of the sched_entity across the tg tree to make it
10028 * visible to the root
10030 static void propagate_entity_cfs_rq(struct sched_entity *se)
10032 struct cfs_rq *cfs_rq;
10034 /* Start to propagate at parent */
10037 for_each_sched_entity(se) {
10038 cfs_rq = cfs_rq_of(se);
10040 if (cfs_rq_throttled(cfs_rq))
10043 update_load_avg(cfs_rq, se, UPDATE_TG);
10047 static void propagate_entity_cfs_rq(struct sched_entity *se) { }
10050 static void detach_entity_cfs_rq(struct sched_entity *se)
10052 struct cfs_rq *cfs_rq = cfs_rq_of(se);
10054 /* Catch up with the cfs_rq and remove our load when we leave */
10055 update_load_avg(cfs_rq, se, 0);
10056 detach_entity_load_avg(cfs_rq, se);
10057 update_tg_load_avg(cfs_rq, false);
10058 propagate_entity_cfs_rq(se);
10061 static void attach_entity_cfs_rq(struct sched_entity *se)
10063 struct cfs_rq *cfs_rq = cfs_rq_of(se);
10065 #ifdef CONFIG_FAIR_GROUP_SCHED
10067 * Since the real-depth could have been changed (only FAIR
10068 * class maintain depth value), reset depth properly.
10070 se->depth = se->parent ? se->parent->depth + 1 : 0;
10073 /* Synchronize entity with its cfs_rq */
10074 update_load_avg(cfs_rq, se, sched_feat(ATTACH_AGE_LOAD) ? 0 : SKIP_AGE_LOAD);
10075 attach_entity_load_avg(cfs_rq, se, 0);
10076 update_tg_load_avg(cfs_rq, false);
10077 propagate_entity_cfs_rq(se);
10080 static void detach_task_cfs_rq(struct task_struct *p)
10082 struct sched_entity *se = &p->se;
10083 struct cfs_rq *cfs_rq = cfs_rq_of(se);
10085 if (!vruntime_normalized(p)) {
10087 * Fix up our vruntime so that the current sleep doesn't
10088 * cause 'unlimited' sleep bonus.
10090 place_entity(cfs_rq, se, 0);
10091 se->vruntime -= cfs_rq->min_vruntime;
10094 detach_entity_cfs_rq(se);
10097 static void attach_task_cfs_rq(struct task_struct *p)
10099 struct sched_entity *se = &p->se;
10100 struct cfs_rq *cfs_rq = cfs_rq_of(se);
10102 attach_entity_cfs_rq(se);
10104 if (!vruntime_normalized(p))
10105 se->vruntime += cfs_rq->min_vruntime;
10108 static void switched_from_fair(struct rq *rq, struct task_struct *p)
10110 detach_task_cfs_rq(p);
10113 static void switched_to_fair(struct rq *rq, struct task_struct *p)
10115 attach_task_cfs_rq(p);
10117 if (task_on_rq_queued(p)) {
10119 * We were most likely switched from sched_rt, so
10120 * kick off the schedule if running, otherwise just see
10121 * if we can still preempt the current task.
10126 check_preempt_curr(rq, p, 0);
10130 /* Account for a task changing its policy or group.
10132 * This routine is mostly called to set cfs_rq->curr field when a task
10133 * migrates between groups/classes.
10135 static void set_curr_task_fair(struct rq *rq)
10137 struct sched_entity *se = &rq->curr->se;
10139 for_each_sched_entity(se) {
10140 struct cfs_rq *cfs_rq = cfs_rq_of(se);
10142 set_next_entity(cfs_rq, se);
10143 /* ensure bandwidth has been allocated on our new cfs_rq */
10144 account_cfs_rq_runtime(cfs_rq, 0);
10148 void init_cfs_rq(struct cfs_rq *cfs_rq)
10150 cfs_rq->tasks_timeline = RB_ROOT_CACHED;
10151 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
10152 #ifndef CONFIG_64BIT
10153 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
10156 raw_spin_lock_init(&cfs_rq->removed.lock);
10160 #ifdef CONFIG_FAIR_GROUP_SCHED
10161 static void task_set_group_fair(struct task_struct *p)
10163 struct sched_entity *se = &p->se;
10165 set_task_rq(p, task_cpu(p));
10166 se->depth = se->parent ? se->parent->depth + 1 : 0;
10169 static void task_move_group_fair(struct task_struct *p)
10171 detach_task_cfs_rq(p);
10172 set_task_rq(p, task_cpu(p));
10175 /* Tell se's cfs_rq has been changed -- migrated */
10176 p->se.avg.last_update_time = 0;
10178 attach_task_cfs_rq(p);
10181 static void task_change_group_fair(struct task_struct *p, int type)
10184 case TASK_SET_GROUP:
10185 task_set_group_fair(p);
10188 case TASK_MOVE_GROUP:
10189 task_move_group_fair(p);
10194 void free_fair_sched_group(struct task_group *tg)
10198 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
10200 for_each_possible_cpu(i) {
10202 kfree(tg->cfs_rq[i]);
10211 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
10213 struct sched_entity *se;
10214 struct cfs_rq *cfs_rq;
10217 tg->cfs_rq = kcalloc(nr_cpu_ids, sizeof(cfs_rq), GFP_KERNEL);
10220 tg->se = kcalloc(nr_cpu_ids, sizeof(se), GFP_KERNEL);
10224 tg->shares = NICE_0_LOAD;
10226 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
10228 for_each_possible_cpu(i) {
10229 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
10230 GFP_KERNEL, cpu_to_node(i));
10234 se = kzalloc_node(sizeof(struct sched_entity),
10235 GFP_KERNEL, cpu_to_node(i));
10239 init_cfs_rq(cfs_rq);
10240 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
10241 init_entity_runnable_average(se);
10252 void online_fair_sched_group(struct task_group *tg)
10254 struct sched_entity *se;
10258 for_each_possible_cpu(i) {
10262 raw_spin_lock_irq(&rq->lock);
10263 update_rq_clock(rq);
10264 attach_entity_cfs_rq(se);
10265 sync_throttle(tg, i);
10266 raw_spin_unlock_irq(&rq->lock);
10270 void unregister_fair_sched_group(struct task_group *tg)
10272 unsigned long flags;
10276 for_each_possible_cpu(cpu) {
10278 remove_entity_load_avg(tg->se[cpu]);
10281 * Only empty task groups can be destroyed; so we can speculatively
10282 * check on_list without danger of it being re-added.
10284 if (!tg->cfs_rq[cpu]->on_list)
10289 raw_spin_lock_irqsave(&rq->lock, flags);
10290 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
10291 raw_spin_unlock_irqrestore(&rq->lock, flags);
10295 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
10296 struct sched_entity *se, int cpu,
10297 struct sched_entity *parent)
10299 struct rq *rq = cpu_rq(cpu);
10303 init_cfs_rq_runtime(cfs_rq);
10305 tg->cfs_rq[cpu] = cfs_rq;
10308 /* se could be NULL for root_task_group */
10313 se->cfs_rq = &rq->cfs;
10316 se->cfs_rq = parent->my_q;
10317 se->depth = parent->depth + 1;
10321 /* guarantee group entities always have weight */
10322 update_load_set(&se->load, NICE_0_LOAD);
10323 se->parent = parent;
10326 static DEFINE_MUTEX(shares_mutex);
10328 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
10333 * We can't change the weight of the root cgroup.
10338 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
10340 mutex_lock(&shares_mutex);
10341 if (tg->shares == shares)
10344 tg->shares = shares;
10345 for_each_possible_cpu(i) {
10346 struct rq *rq = cpu_rq(i);
10347 struct sched_entity *se = tg->se[i];
10348 struct rq_flags rf;
10350 /* Propagate contribution to hierarchy */
10351 rq_lock_irqsave(rq, &rf);
10352 update_rq_clock(rq);
10353 for_each_sched_entity(se) {
10354 update_load_avg(cfs_rq_of(se), se, UPDATE_TG);
10355 update_cfs_group(se);
10357 rq_unlock_irqrestore(rq, &rf);
10361 mutex_unlock(&shares_mutex);
10364 #else /* CONFIG_FAIR_GROUP_SCHED */
10366 void free_fair_sched_group(struct task_group *tg) { }
10368 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
10373 void online_fair_sched_group(struct task_group *tg) { }
10375 void unregister_fair_sched_group(struct task_group *tg) { }
10377 #endif /* CONFIG_FAIR_GROUP_SCHED */
10380 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
10382 struct sched_entity *se = &task->se;
10383 unsigned int rr_interval = 0;
10386 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
10389 if (rq->cfs.load.weight)
10390 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
10392 return rr_interval;
10396 * All the scheduling class methods:
10398 const struct sched_class fair_sched_class = {
10399 .next = &idle_sched_class,
10400 .enqueue_task = enqueue_task_fair,
10401 .dequeue_task = dequeue_task_fair,
10402 .yield_task = yield_task_fair,
10403 .yield_to_task = yield_to_task_fair,
10405 .check_preempt_curr = check_preempt_wakeup,
10407 .pick_next_task = pick_next_task_fair,
10408 .put_prev_task = put_prev_task_fair,
10411 .select_task_rq = select_task_rq_fair,
10412 .migrate_task_rq = migrate_task_rq_fair,
10414 .rq_online = rq_online_fair,
10415 .rq_offline = rq_offline_fair,
10417 .task_dead = task_dead_fair,
10418 .set_cpus_allowed = set_cpus_allowed_common,
10421 .set_curr_task = set_curr_task_fair,
10422 .task_tick = task_tick_fair,
10423 .task_fork = task_fork_fair,
10425 .prio_changed = prio_changed_fair,
10426 .switched_from = switched_from_fair,
10427 .switched_to = switched_to_fair,
10429 .get_rr_interval = get_rr_interval_fair,
10431 .update_curr = update_curr_fair,
10433 #ifdef CONFIG_FAIR_GROUP_SCHED
10434 .task_change_group = task_change_group_fair,
10438 #ifdef CONFIG_SCHED_DEBUG
10439 void print_cfs_stats(struct seq_file *m, int cpu)
10441 struct cfs_rq *cfs_rq, *pos;
10444 for_each_leaf_cfs_rq_safe(cpu_rq(cpu), cfs_rq, pos)
10445 print_cfs_rq(m, cpu, cfs_rq);
10449 #ifdef CONFIG_NUMA_BALANCING
10450 void show_numa_stats(struct task_struct *p, struct seq_file *m)
10453 unsigned long tsf = 0, tpf = 0, gsf = 0, gpf = 0;
10455 for_each_online_node(node) {
10456 if (p->numa_faults) {
10457 tsf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 0)];
10458 tpf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 1)];
10460 if (p->numa_group) {
10461 gsf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 0)],
10462 gpf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 1)];
10464 print_numa_stats(m, node, tsf, tpf, gsf, gpf);
10467 #endif /* CONFIG_NUMA_BALANCING */
10468 #endif /* CONFIG_SCHED_DEBUG */
10470 __init void init_sched_fair_class(void)
10473 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
10475 #ifdef CONFIG_NO_HZ_COMMON
10476 nohz.next_balance = jiffies;
10477 nohz.next_blocked = jiffies;
10478 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);