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 static void account_numa_enqueue(struct rq *rq, struct task_struct *p)
1144 rq->nr_numa_running += (p->numa_preferred_nid != -1);
1145 rq->nr_preferred_running += (p->numa_preferred_nid == task_node(p));
1148 static void account_numa_dequeue(struct rq *rq, struct task_struct *p)
1150 rq->nr_numa_running -= (p->numa_preferred_nid != -1);
1151 rq->nr_preferred_running -= (p->numa_preferred_nid == task_node(p));
1154 /* Shared or private faults. */
1155 #define NR_NUMA_HINT_FAULT_TYPES 2
1157 /* Memory and CPU locality */
1158 #define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2)
1160 /* Averaged statistics, and temporary buffers. */
1161 #define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2)
1163 pid_t task_numa_group_id(struct task_struct *p)
1165 return p->numa_group ? p->numa_group->gid : 0;
1169 * The averaged statistics, shared & private, memory & CPU,
1170 * occupy the first half of the array. The second half of the
1171 * array is for current counters, which are averaged into the
1172 * first set by task_numa_placement.
1174 static inline int task_faults_idx(enum numa_faults_stats s, int nid, int priv)
1176 return NR_NUMA_HINT_FAULT_TYPES * (s * nr_node_ids + nid) + priv;
1179 static inline unsigned long task_faults(struct task_struct *p, int nid)
1181 if (!p->numa_faults)
1184 return p->numa_faults[task_faults_idx(NUMA_MEM, nid, 0)] +
1185 p->numa_faults[task_faults_idx(NUMA_MEM, nid, 1)];
1188 static inline unsigned long group_faults(struct task_struct *p, int nid)
1193 return p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 0)] +
1194 p->numa_group->faults[task_faults_idx(NUMA_MEM, nid, 1)];
1197 static inline unsigned long group_faults_cpu(struct numa_group *group, int nid)
1199 return group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 0)] +
1200 group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 1)];
1203 static inline unsigned long group_faults_priv(struct numa_group *ng)
1205 unsigned long faults = 0;
1208 for_each_online_node(node) {
1209 faults += ng->faults[task_faults_idx(NUMA_MEM, node, 1)];
1215 static inline unsigned long group_faults_shared(struct numa_group *ng)
1217 unsigned long faults = 0;
1220 for_each_online_node(node) {
1221 faults += ng->faults[task_faults_idx(NUMA_MEM, node, 0)];
1228 * A node triggering more than 1/3 as many NUMA faults as the maximum is
1229 * considered part of a numa group's pseudo-interleaving set. Migrations
1230 * between these nodes are slowed down, to allow things to settle down.
1232 #define ACTIVE_NODE_FRACTION 3
1234 static bool numa_is_active_node(int nid, struct numa_group *ng)
1236 return group_faults_cpu(ng, nid) * ACTIVE_NODE_FRACTION > ng->max_faults_cpu;
1239 /* Handle placement on systems where not all nodes are directly connected. */
1240 static unsigned long score_nearby_nodes(struct task_struct *p, int nid,
1241 int maxdist, bool task)
1243 unsigned long score = 0;
1247 * All nodes are directly connected, and the same distance
1248 * from each other. No need for fancy placement algorithms.
1250 if (sched_numa_topology_type == NUMA_DIRECT)
1254 * This code is called for each node, introducing N^2 complexity,
1255 * which should be ok given the number of nodes rarely exceeds 8.
1257 for_each_online_node(node) {
1258 unsigned long faults;
1259 int dist = node_distance(nid, node);
1262 * The furthest away nodes in the system are not interesting
1263 * for placement; nid was already counted.
1265 if (dist == sched_max_numa_distance || node == nid)
1269 * On systems with a backplane NUMA topology, compare groups
1270 * of nodes, and move tasks towards the group with the most
1271 * memory accesses. When comparing two nodes at distance
1272 * "hoplimit", only nodes closer by than "hoplimit" are part
1273 * of each group. Skip other nodes.
1275 if (sched_numa_topology_type == NUMA_BACKPLANE &&
1279 /* Add up the faults from nearby nodes. */
1281 faults = task_faults(p, node);
1283 faults = group_faults(p, node);
1286 * On systems with a glueless mesh NUMA topology, there are
1287 * no fixed "groups of nodes". Instead, nodes that are not
1288 * directly connected bounce traffic through intermediate
1289 * nodes; a numa_group can occupy any set of nodes.
1290 * The further away a node is, the less the faults count.
1291 * This seems to result in good task placement.
1293 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
1294 faults *= (sched_max_numa_distance - dist);
1295 faults /= (sched_max_numa_distance - LOCAL_DISTANCE);
1305 * These return the fraction of accesses done by a particular task, or
1306 * task group, on a particular numa node. The group weight is given a
1307 * larger multiplier, in order to group tasks together that are almost
1308 * evenly spread out between numa nodes.
1310 static inline unsigned long task_weight(struct task_struct *p, int nid,
1313 unsigned long faults, total_faults;
1315 if (!p->numa_faults)
1318 total_faults = p->total_numa_faults;
1323 faults = task_faults(p, nid);
1324 faults += score_nearby_nodes(p, nid, dist, true);
1326 return 1000 * faults / total_faults;
1329 static inline unsigned long group_weight(struct task_struct *p, int nid,
1332 unsigned long faults, total_faults;
1337 total_faults = p->numa_group->total_faults;
1342 faults = group_faults(p, nid);
1343 faults += score_nearby_nodes(p, nid, dist, false);
1345 return 1000 * faults / total_faults;
1348 bool should_numa_migrate_memory(struct task_struct *p, struct page * page,
1349 int src_nid, int dst_cpu)
1351 struct numa_group *ng = p->numa_group;
1352 int dst_nid = cpu_to_node(dst_cpu);
1353 int last_cpupid, this_cpupid;
1355 this_cpupid = cpu_pid_to_cpupid(dst_cpu, current->pid);
1358 * Multi-stage node selection is used in conjunction with a periodic
1359 * migration fault to build a temporal task<->page relation. By using
1360 * a two-stage filter we remove short/unlikely relations.
1362 * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate
1363 * a task's usage of a particular page (n_p) per total usage of this
1364 * page (n_t) (in a given time-span) to a probability.
1366 * Our periodic faults will sample this probability and getting the
1367 * same result twice in a row, given these samples are fully
1368 * independent, is then given by P(n)^2, provided our sample period
1369 * is sufficiently short compared to the usage pattern.
1371 * This quadric squishes small probabilities, making it less likely we
1372 * act on an unlikely task<->page relation.
1374 last_cpupid = page_cpupid_xchg_last(page, this_cpupid);
1375 if (!cpupid_pid_unset(last_cpupid) &&
1376 cpupid_to_nid(last_cpupid) != dst_nid)
1379 /* Always allow migrate on private faults */
1380 if (cpupid_match_pid(p, last_cpupid))
1383 /* A shared fault, but p->numa_group has not been set up yet. */
1388 * Destination node is much more heavily used than the source
1389 * node? Allow migration.
1391 if (group_faults_cpu(ng, dst_nid) > group_faults_cpu(ng, src_nid) *
1392 ACTIVE_NODE_FRACTION)
1396 * Distribute memory according to CPU & memory use on each node,
1397 * with 3/4 hysteresis to avoid unnecessary memory migrations:
1399 * faults_cpu(dst) 3 faults_cpu(src)
1400 * --------------- * - > ---------------
1401 * faults_mem(dst) 4 faults_mem(src)
1403 return group_faults_cpu(ng, dst_nid) * group_faults(p, src_nid) * 3 >
1404 group_faults_cpu(ng, src_nid) * group_faults(p, dst_nid) * 4;
1407 static unsigned long weighted_cpuload(struct rq *rq);
1408 static unsigned long source_load(int cpu, int type);
1409 static unsigned long target_load(int cpu, int type);
1410 static unsigned long capacity_of(int cpu);
1412 /* Cached statistics for all CPUs within a node */
1414 unsigned long nr_running;
1417 /* Total compute capacity of CPUs on a node */
1418 unsigned long compute_capacity;
1420 /* Approximate capacity in terms of runnable tasks on a node */
1421 unsigned long task_capacity;
1422 int has_free_capacity;
1426 * XXX borrowed from update_sg_lb_stats
1428 static void update_numa_stats(struct numa_stats *ns, int nid)
1430 int smt, cpu, cpus = 0;
1431 unsigned long capacity;
1433 memset(ns, 0, sizeof(*ns));
1434 for_each_cpu(cpu, cpumask_of_node(nid)) {
1435 struct rq *rq = cpu_rq(cpu);
1437 ns->nr_running += rq->nr_running;
1438 ns->load += weighted_cpuload(rq);
1439 ns->compute_capacity += capacity_of(cpu);
1445 * If we raced with hotplug and there are no CPUs left in our mask
1446 * the @ns structure is NULL'ed and task_numa_compare() will
1447 * not find this node attractive.
1449 * We'll either bail at !has_free_capacity, or we'll detect a huge
1450 * imbalance and bail there.
1455 /* smt := ceil(cpus / capacity), assumes: 1 < smt_power < 2 */
1456 smt = DIV_ROUND_UP(SCHED_CAPACITY_SCALE * cpus, ns->compute_capacity);
1457 capacity = cpus / smt; /* cores */
1459 ns->task_capacity = min_t(unsigned, capacity,
1460 DIV_ROUND_CLOSEST(ns->compute_capacity, SCHED_CAPACITY_SCALE));
1461 ns->has_free_capacity = (ns->nr_running < ns->task_capacity);
1464 struct task_numa_env {
1465 struct task_struct *p;
1467 int src_cpu, src_nid;
1468 int dst_cpu, dst_nid;
1470 struct numa_stats src_stats, dst_stats;
1475 struct task_struct *best_task;
1480 static void task_numa_assign(struct task_numa_env *env,
1481 struct task_struct *p, long imp)
1484 put_task_struct(env->best_task);
1489 env->best_imp = imp;
1490 env->best_cpu = env->dst_cpu;
1493 static bool load_too_imbalanced(long src_load, long dst_load,
1494 struct task_numa_env *env)
1497 long orig_src_load, orig_dst_load;
1498 long src_capacity, dst_capacity;
1501 * The load is corrected for the CPU capacity available on each node.
1504 * ------------ vs ---------
1505 * src_capacity dst_capacity
1507 src_capacity = env->src_stats.compute_capacity;
1508 dst_capacity = env->dst_stats.compute_capacity;
1510 /* We care about the slope of the imbalance, not the direction. */
1511 if (dst_load < src_load)
1512 swap(dst_load, src_load);
1514 /* Is the difference below the threshold? */
1515 imb = dst_load * src_capacity * 100 -
1516 src_load * dst_capacity * env->imbalance_pct;
1521 * The imbalance is above the allowed threshold.
1522 * Compare it with the old imbalance.
1524 orig_src_load = env->src_stats.load;
1525 orig_dst_load = env->dst_stats.load;
1527 if (orig_dst_load < orig_src_load)
1528 swap(orig_dst_load, orig_src_load);
1530 old_imb = orig_dst_load * src_capacity * 100 -
1531 orig_src_load * dst_capacity * env->imbalance_pct;
1533 /* Would this change make things worse? */
1534 return (imb > old_imb);
1538 * This checks if the overall compute and NUMA accesses of the system would
1539 * be improved if the source tasks was migrated to the target dst_cpu taking
1540 * into account that it might be best if task running on the dst_cpu should
1541 * be exchanged with the source task
1543 static void task_numa_compare(struct task_numa_env *env,
1544 long taskimp, long groupimp)
1546 struct rq *src_rq = cpu_rq(env->src_cpu);
1547 struct rq *dst_rq = cpu_rq(env->dst_cpu);
1548 struct task_struct *cur;
1549 long src_load, dst_load;
1551 long imp = env->p->numa_group ? groupimp : taskimp;
1553 int dist = env->dist;
1556 cur = task_rcu_dereference(&dst_rq->curr);
1557 if (cur && ((cur->flags & PF_EXITING) || is_idle_task(cur)))
1561 * Because we have preemption enabled we can get migrated around and
1562 * end try selecting ourselves (current == env->p) as a swap candidate.
1568 * "imp" is the fault differential for the source task between the
1569 * source and destination node. Calculate the total differential for
1570 * the source task and potential destination task. The more negative
1571 * the value is, the more rmeote accesses that would be expected to
1572 * be incurred if the tasks were swapped.
1575 /* Skip this swap candidate if cannot move to the source CPU: */
1576 if (!cpumask_test_cpu(env->src_cpu, &cur->cpus_allowed))
1580 * If dst and source tasks are in the same NUMA group, or not
1581 * in any group then look only at task weights.
1583 if (cur->numa_group == env->p->numa_group) {
1584 imp = taskimp + task_weight(cur, env->src_nid, dist) -
1585 task_weight(cur, env->dst_nid, dist);
1587 * Add some hysteresis to prevent swapping the
1588 * tasks within a group over tiny differences.
1590 if (cur->numa_group)
1594 * Compare the group weights. If a task is all by
1595 * itself (not part of a group), use the task weight
1598 if (cur->numa_group)
1599 imp += group_weight(cur, env->src_nid, dist) -
1600 group_weight(cur, env->dst_nid, dist);
1602 imp += task_weight(cur, env->src_nid, dist) -
1603 task_weight(cur, env->dst_nid, dist);
1607 if (imp <= env->best_imp && moveimp <= env->best_imp)
1611 /* Is there capacity at our destination? */
1612 if (env->src_stats.nr_running <= env->src_stats.task_capacity &&
1613 !env->dst_stats.has_free_capacity)
1619 /* Balance doesn't matter much if we're running a task per CPU: */
1620 if (imp > env->best_imp && src_rq->nr_running == 1 &&
1621 dst_rq->nr_running == 1)
1625 * In the overloaded case, try and keep the load balanced.
1628 load = task_h_load(env->p);
1629 dst_load = env->dst_stats.load + load;
1630 src_load = env->src_stats.load - load;
1632 if (moveimp > imp && moveimp > env->best_imp) {
1634 * If the improvement from just moving env->p direction is
1635 * better than swapping tasks around, check if a move is
1636 * possible. Store a slightly smaller score than moveimp,
1637 * so an actually idle CPU will win.
1639 if (!load_too_imbalanced(src_load, dst_load, env)) {
1646 if (imp <= env->best_imp)
1650 load = task_h_load(cur);
1655 if (load_too_imbalanced(src_load, dst_load, env))
1659 * One idle CPU per node is evaluated for a task numa move.
1660 * Call select_idle_sibling to maybe find a better one.
1664 * select_idle_siblings() uses an per-CPU cpumask that
1665 * can be used from IRQ context.
1667 local_irq_disable();
1668 env->dst_cpu = select_idle_sibling(env->p, env->src_cpu,
1674 task_numa_assign(env, cur, imp);
1679 static void task_numa_find_cpu(struct task_numa_env *env,
1680 long taskimp, long groupimp)
1684 for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) {
1685 /* Skip this CPU if the source task cannot migrate */
1686 if (!cpumask_test_cpu(cpu, &env->p->cpus_allowed))
1690 task_numa_compare(env, taskimp, groupimp);
1694 /* Only move tasks to a NUMA node less busy than the current node. */
1695 static bool numa_has_capacity(struct task_numa_env *env)
1697 struct numa_stats *src = &env->src_stats;
1698 struct numa_stats *dst = &env->dst_stats;
1700 if (src->has_free_capacity && !dst->has_free_capacity)
1704 * Only consider a task move if the source has a higher load
1705 * than the destination, corrected for CPU capacity on each node.
1707 * src->load dst->load
1708 * --------------------- vs ---------------------
1709 * src->compute_capacity dst->compute_capacity
1711 if (src->load * dst->compute_capacity * env->imbalance_pct >
1713 dst->load * src->compute_capacity * 100)
1719 static int task_numa_migrate(struct task_struct *p)
1721 struct task_numa_env env = {
1724 .src_cpu = task_cpu(p),
1725 .src_nid = task_node(p),
1727 .imbalance_pct = 112,
1733 struct sched_domain *sd;
1734 unsigned long taskweight, groupweight;
1736 long taskimp, groupimp;
1739 * Pick the lowest SD_NUMA domain, as that would have the smallest
1740 * imbalance and would be the first to start moving tasks about.
1742 * And we want to avoid any moving of tasks about, as that would create
1743 * random movement of tasks -- counter the numa conditions we're trying
1747 sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu));
1749 env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2;
1753 * Cpusets can break the scheduler domain tree into smaller
1754 * balance domains, some of which do not cross NUMA boundaries.
1755 * Tasks that are "trapped" in such domains cannot be migrated
1756 * elsewhere, so there is no point in (re)trying.
1758 if (unlikely(!sd)) {
1759 p->numa_preferred_nid = task_node(p);
1763 env.dst_nid = p->numa_preferred_nid;
1764 dist = env.dist = node_distance(env.src_nid, env.dst_nid);
1765 taskweight = task_weight(p, env.src_nid, dist);
1766 groupweight = group_weight(p, env.src_nid, dist);
1767 update_numa_stats(&env.src_stats, env.src_nid);
1768 taskimp = task_weight(p, env.dst_nid, dist) - taskweight;
1769 groupimp = group_weight(p, env.dst_nid, dist) - groupweight;
1770 update_numa_stats(&env.dst_stats, env.dst_nid);
1772 /* Try to find a spot on the preferred nid. */
1773 if (numa_has_capacity(&env))
1774 task_numa_find_cpu(&env, taskimp, groupimp);
1777 * Look at other nodes in these cases:
1778 * - there is no space available on the preferred_nid
1779 * - the task is part of a numa_group that is interleaved across
1780 * multiple NUMA nodes; in order to better consolidate the group,
1781 * we need to check other locations.
1783 if (env.best_cpu == -1 || (p->numa_group && p->numa_group->active_nodes > 1)) {
1784 for_each_online_node(nid) {
1785 if (nid == env.src_nid || nid == p->numa_preferred_nid)
1788 dist = node_distance(env.src_nid, env.dst_nid);
1789 if (sched_numa_topology_type == NUMA_BACKPLANE &&
1791 taskweight = task_weight(p, env.src_nid, dist);
1792 groupweight = group_weight(p, env.src_nid, dist);
1795 /* Only consider nodes where both task and groups benefit */
1796 taskimp = task_weight(p, nid, dist) - taskweight;
1797 groupimp = group_weight(p, nid, dist) - groupweight;
1798 if (taskimp < 0 && groupimp < 0)
1803 update_numa_stats(&env.dst_stats, env.dst_nid);
1804 if (numa_has_capacity(&env))
1805 task_numa_find_cpu(&env, taskimp, groupimp);
1810 * If the task is part of a workload that spans multiple NUMA nodes,
1811 * and is migrating into one of the workload's active nodes, remember
1812 * this node as the task's preferred numa node, so the workload can
1814 * A task that migrated to a second choice node will be better off
1815 * trying for a better one later. Do not set the preferred node here.
1817 if (p->numa_group) {
1818 struct numa_group *ng = p->numa_group;
1820 if (env.best_cpu == -1)
1825 if (ng->active_nodes > 1 && numa_is_active_node(env.dst_nid, ng))
1826 sched_setnuma(p, env.dst_nid);
1829 /* No better CPU than the current one was found. */
1830 if (env.best_cpu == -1)
1834 * Reset the scan period if the task is being rescheduled on an
1835 * alternative node to recheck if the tasks is now properly placed.
1837 p->numa_scan_period = task_scan_start(p);
1839 if (env.best_task == NULL) {
1840 ret = migrate_task_to(p, env.best_cpu);
1842 trace_sched_stick_numa(p, env.src_cpu, env.best_cpu);
1846 ret = migrate_swap(p, env.best_task);
1848 trace_sched_stick_numa(p, env.src_cpu, task_cpu(env.best_task));
1849 put_task_struct(env.best_task);
1853 /* Attempt to migrate a task to a CPU on the preferred node. */
1854 static void numa_migrate_preferred(struct task_struct *p)
1856 unsigned long interval = HZ;
1858 /* This task has no NUMA fault statistics yet */
1859 if (unlikely(p->numa_preferred_nid == -1 || !p->numa_faults))
1862 /* Periodically retry migrating the task to the preferred node */
1863 interval = min(interval, msecs_to_jiffies(p->numa_scan_period) / 16);
1864 p->numa_migrate_retry = jiffies + interval;
1866 /* Success if task is already running on preferred CPU */
1867 if (task_node(p) == p->numa_preferred_nid)
1870 /* Otherwise, try migrate to a CPU on the preferred node */
1871 task_numa_migrate(p);
1875 * Find out how many nodes on the workload is actively running on. Do this by
1876 * tracking the nodes from which NUMA hinting faults are triggered. This can
1877 * be different from the set of nodes where the workload's memory is currently
1880 static void numa_group_count_active_nodes(struct numa_group *numa_group)
1882 unsigned long faults, max_faults = 0;
1883 int nid, active_nodes = 0;
1885 for_each_online_node(nid) {
1886 faults = group_faults_cpu(numa_group, nid);
1887 if (faults > max_faults)
1888 max_faults = faults;
1891 for_each_online_node(nid) {
1892 faults = group_faults_cpu(numa_group, nid);
1893 if (faults * ACTIVE_NODE_FRACTION > max_faults)
1897 numa_group->max_faults_cpu = max_faults;
1898 numa_group->active_nodes = active_nodes;
1902 * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
1903 * increments. The more local the fault statistics are, the higher the scan
1904 * period will be for the next scan window. If local/(local+remote) ratio is
1905 * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS)
1906 * the scan period will decrease. Aim for 70% local accesses.
1908 #define NUMA_PERIOD_SLOTS 10
1909 #define NUMA_PERIOD_THRESHOLD 7
1912 * Increase the scan period (slow down scanning) if the majority of
1913 * our memory is already on our local node, or if the majority of
1914 * the page accesses are shared with other processes.
1915 * Otherwise, decrease the scan period.
1917 static void update_task_scan_period(struct task_struct *p,
1918 unsigned long shared, unsigned long private)
1920 unsigned int period_slot;
1921 int lr_ratio, ps_ratio;
1924 unsigned long remote = p->numa_faults_locality[0];
1925 unsigned long local = p->numa_faults_locality[1];
1928 * If there were no record hinting faults then either the task is
1929 * completely idle or all activity is areas that are not of interest
1930 * to automatic numa balancing. Related to that, if there were failed
1931 * migration then it implies we are migrating too quickly or the local
1932 * node is overloaded. In either case, scan slower
1934 if (local + shared == 0 || p->numa_faults_locality[2]) {
1935 p->numa_scan_period = min(p->numa_scan_period_max,
1936 p->numa_scan_period << 1);
1938 p->mm->numa_next_scan = jiffies +
1939 msecs_to_jiffies(p->numa_scan_period);
1945 * Prepare to scale scan period relative to the current period.
1946 * == NUMA_PERIOD_THRESHOLD scan period stays the same
1947 * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
1948 * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
1950 period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS);
1951 lr_ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote);
1952 ps_ratio = (private * NUMA_PERIOD_SLOTS) / (private + shared);
1954 if (ps_ratio >= NUMA_PERIOD_THRESHOLD) {
1956 * Most memory accesses are local. There is no need to
1957 * do fast NUMA scanning, since memory is already local.
1959 int slot = ps_ratio - NUMA_PERIOD_THRESHOLD;
1962 diff = slot * period_slot;
1963 } else if (lr_ratio >= NUMA_PERIOD_THRESHOLD) {
1965 * Most memory accesses are shared with other tasks.
1966 * There is no point in continuing fast NUMA scanning,
1967 * since other tasks may just move the memory elsewhere.
1969 int slot = lr_ratio - NUMA_PERIOD_THRESHOLD;
1972 diff = slot * period_slot;
1975 * Private memory faults exceed (SLOTS-THRESHOLD)/SLOTS,
1976 * yet they are not on the local NUMA node. Speed up
1977 * NUMA scanning to get the memory moved over.
1979 int ratio = max(lr_ratio, ps_ratio);
1980 diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot;
1983 p->numa_scan_period = clamp(p->numa_scan_period + diff,
1984 task_scan_min(p), task_scan_max(p));
1985 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
1989 * Get the fraction of time the task has been running since the last
1990 * NUMA placement cycle. The scheduler keeps similar statistics, but
1991 * decays those on a 32ms period, which is orders of magnitude off
1992 * from the dozens-of-seconds NUMA balancing period. Use the scheduler
1993 * stats only if the task is so new there are no NUMA statistics yet.
1995 static u64 numa_get_avg_runtime(struct task_struct *p, u64 *period)
1997 u64 runtime, delta, now;
1998 /* Use the start of this time slice to avoid calculations. */
1999 now = p->se.exec_start;
2000 runtime = p->se.sum_exec_runtime;
2002 if (p->last_task_numa_placement) {
2003 delta = runtime - p->last_sum_exec_runtime;
2004 *period = now - p->last_task_numa_placement;
2006 delta = p->se.avg.load_sum;
2007 *period = LOAD_AVG_MAX;
2010 p->last_sum_exec_runtime = runtime;
2011 p->last_task_numa_placement = now;
2017 * Determine the preferred nid for a task in a numa_group. This needs to
2018 * be done in a way that produces consistent results with group_weight,
2019 * otherwise workloads might not converge.
2021 static int preferred_group_nid(struct task_struct *p, int nid)
2026 /* Direct connections between all NUMA nodes. */
2027 if (sched_numa_topology_type == NUMA_DIRECT)
2031 * On a system with glueless mesh NUMA topology, group_weight
2032 * scores nodes according to the number of NUMA hinting faults on
2033 * both the node itself, and on nearby nodes.
2035 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
2036 unsigned long score, max_score = 0;
2037 int node, max_node = nid;
2039 dist = sched_max_numa_distance;
2041 for_each_online_node(node) {
2042 score = group_weight(p, node, dist);
2043 if (score > max_score) {
2052 * Finding the preferred nid in a system with NUMA backplane
2053 * interconnect topology is more involved. The goal is to locate
2054 * tasks from numa_groups near each other in the system, and
2055 * untangle workloads from different sides of the system. This requires
2056 * searching down the hierarchy of node groups, recursively searching
2057 * inside the highest scoring group of nodes. The nodemask tricks
2058 * keep the complexity of the search down.
2060 nodes = node_online_map;
2061 for (dist = sched_max_numa_distance; dist > LOCAL_DISTANCE; dist--) {
2062 unsigned long max_faults = 0;
2063 nodemask_t max_group = NODE_MASK_NONE;
2066 /* Are there nodes at this distance from each other? */
2067 if (!find_numa_distance(dist))
2070 for_each_node_mask(a, nodes) {
2071 unsigned long faults = 0;
2072 nodemask_t this_group;
2073 nodes_clear(this_group);
2075 /* Sum group's NUMA faults; includes a==b case. */
2076 for_each_node_mask(b, nodes) {
2077 if (node_distance(a, b) < dist) {
2078 faults += group_faults(p, b);
2079 node_set(b, this_group);
2080 node_clear(b, nodes);
2084 /* Remember the top group. */
2085 if (faults > max_faults) {
2086 max_faults = faults;
2087 max_group = this_group;
2089 * subtle: at the smallest distance there is
2090 * just one node left in each "group", the
2091 * winner is the preferred nid.
2096 /* Next round, evaluate the nodes within max_group. */
2104 static void task_numa_placement(struct task_struct *p)
2106 int seq, nid, max_nid = -1, max_group_nid = -1;
2107 unsigned long max_faults = 0, max_group_faults = 0;
2108 unsigned long fault_types[2] = { 0, 0 };
2109 unsigned long total_faults;
2110 u64 runtime, period;
2111 spinlock_t *group_lock = NULL;
2114 * The p->mm->numa_scan_seq field gets updated without
2115 * exclusive access. Use READ_ONCE() here to ensure
2116 * that the field is read in a single access:
2118 seq = READ_ONCE(p->mm->numa_scan_seq);
2119 if (p->numa_scan_seq == seq)
2121 p->numa_scan_seq = seq;
2122 p->numa_scan_period_max = task_scan_max(p);
2124 total_faults = p->numa_faults_locality[0] +
2125 p->numa_faults_locality[1];
2126 runtime = numa_get_avg_runtime(p, &period);
2128 /* If the task is part of a group prevent parallel updates to group stats */
2129 if (p->numa_group) {
2130 group_lock = &p->numa_group->lock;
2131 spin_lock_irq(group_lock);
2134 /* Find the node with the highest number of faults */
2135 for_each_online_node(nid) {
2136 /* Keep track of the offsets in numa_faults array */
2137 int mem_idx, membuf_idx, cpu_idx, cpubuf_idx;
2138 unsigned long faults = 0, group_faults = 0;
2141 for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) {
2142 long diff, f_diff, f_weight;
2144 mem_idx = task_faults_idx(NUMA_MEM, nid, priv);
2145 membuf_idx = task_faults_idx(NUMA_MEMBUF, nid, priv);
2146 cpu_idx = task_faults_idx(NUMA_CPU, nid, priv);
2147 cpubuf_idx = task_faults_idx(NUMA_CPUBUF, nid, priv);
2149 /* Decay existing window, copy faults since last scan */
2150 diff = p->numa_faults[membuf_idx] - p->numa_faults[mem_idx] / 2;
2151 fault_types[priv] += p->numa_faults[membuf_idx];
2152 p->numa_faults[membuf_idx] = 0;
2155 * Normalize the faults_from, so all tasks in a group
2156 * count according to CPU use, instead of by the raw
2157 * number of faults. Tasks with little runtime have
2158 * little over-all impact on throughput, and thus their
2159 * faults are less important.
2161 f_weight = div64_u64(runtime << 16, period + 1);
2162 f_weight = (f_weight * p->numa_faults[cpubuf_idx]) /
2164 f_diff = f_weight - p->numa_faults[cpu_idx] / 2;
2165 p->numa_faults[cpubuf_idx] = 0;
2167 p->numa_faults[mem_idx] += diff;
2168 p->numa_faults[cpu_idx] += f_diff;
2169 faults += p->numa_faults[mem_idx];
2170 p->total_numa_faults += diff;
2171 if (p->numa_group) {
2173 * safe because we can only change our own group
2175 * mem_idx represents the offset for a given
2176 * nid and priv in a specific region because it
2177 * is at the beginning of the numa_faults array.
2179 p->numa_group->faults[mem_idx] += diff;
2180 p->numa_group->faults_cpu[mem_idx] += f_diff;
2181 p->numa_group->total_faults += diff;
2182 group_faults += p->numa_group->faults[mem_idx];
2186 if (faults > max_faults) {
2187 max_faults = faults;
2191 if (group_faults > max_group_faults) {
2192 max_group_faults = group_faults;
2193 max_group_nid = nid;
2197 update_task_scan_period(p, fault_types[0], fault_types[1]);
2199 if (p->numa_group) {
2200 numa_group_count_active_nodes(p->numa_group);
2201 spin_unlock_irq(group_lock);
2202 max_nid = preferred_group_nid(p, max_group_nid);
2206 /* Set the new preferred node */
2207 if (max_nid != p->numa_preferred_nid)
2208 sched_setnuma(p, max_nid);
2210 if (task_node(p) != p->numa_preferred_nid)
2211 numa_migrate_preferred(p);
2215 static inline int get_numa_group(struct numa_group *grp)
2217 return atomic_inc_not_zero(&grp->refcount);
2220 static inline void put_numa_group(struct numa_group *grp)
2222 if (atomic_dec_and_test(&grp->refcount))
2223 kfree_rcu(grp, rcu);
2226 static void task_numa_group(struct task_struct *p, int cpupid, int flags,
2229 struct numa_group *grp, *my_grp;
2230 struct task_struct *tsk;
2232 int cpu = cpupid_to_cpu(cpupid);
2235 if (unlikely(!p->numa_group)) {
2236 unsigned int size = sizeof(struct numa_group) +
2237 4*nr_node_ids*sizeof(unsigned long);
2239 grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
2243 atomic_set(&grp->refcount, 1);
2244 grp->active_nodes = 1;
2245 grp->max_faults_cpu = 0;
2246 spin_lock_init(&grp->lock);
2248 /* Second half of the array tracks nids where faults happen */
2249 grp->faults_cpu = grp->faults + NR_NUMA_HINT_FAULT_TYPES *
2252 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2253 grp->faults[i] = p->numa_faults[i];
2255 grp->total_faults = p->total_numa_faults;
2258 rcu_assign_pointer(p->numa_group, grp);
2262 tsk = READ_ONCE(cpu_rq(cpu)->curr);
2264 if (!cpupid_match_pid(tsk, cpupid))
2267 grp = rcu_dereference(tsk->numa_group);
2271 my_grp = p->numa_group;
2276 * Only join the other group if its bigger; if we're the bigger group,
2277 * the other task will join us.
2279 if (my_grp->nr_tasks > grp->nr_tasks)
2283 * Tie-break on the grp address.
2285 if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp)
2288 /* Always join threads in the same process. */
2289 if (tsk->mm == current->mm)
2292 /* Simple filter to avoid false positives due to PID collisions */
2293 if (flags & TNF_SHARED)
2296 /* Update priv based on whether false sharing was detected */
2299 if (join && !get_numa_group(grp))
2307 BUG_ON(irqs_disabled());
2308 double_lock_irq(&my_grp->lock, &grp->lock);
2310 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) {
2311 my_grp->faults[i] -= p->numa_faults[i];
2312 grp->faults[i] += p->numa_faults[i];
2314 my_grp->total_faults -= p->total_numa_faults;
2315 grp->total_faults += p->total_numa_faults;
2320 spin_unlock(&my_grp->lock);
2321 spin_unlock_irq(&grp->lock);
2323 rcu_assign_pointer(p->numa_group, grp);
2325 put_numa_group(my_grp);
2333 void task_numa_free(struct task_struct *p)
2335 struct numa_group *grp = p->numa_group;
2336 void *numa_faults = p->numa_faults;
2337 unsigned long flags;
2341 spin_lock_irqsave(&grp->lock, flags);
2342 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2343 grp->faults[i] -= p->numa_faults[i];
2344 grp->total_faults -= p->total_numa_faults;
2347 spin_unlock_irqrestore(&grp->lock, flags);
2348 RCU_INIT_POINTER(p->numa_group, NULL);
2349 put_numa_group(grp);
2352 p->numa_faults = NULL;
2357 * Got a PROT_NONE fault for a page on @node.
2359 void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags)
2361 struct task_struct *p = current;
2362 bool migrated = flags & TNF_MIGRATED;
2363 int cpu_node = task_node(current);
2364 int local = !!(flags & TNF_FAULT_LOCAL);
2365 struct numa_group *ng;
2368 if (!static_branch_likely(&sched_numa_balancing))
2371 /* for example, ksmd faulting in a user's mm */
2375 /* Allocate buffer to track faults on a per-node basis */
2376 if (unlikely(!p->numa_faults)) {
2377 int size = sizeof(*p->numa_faults) *
2378 NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids;
2380 p->numa_faults = kzalloc(size, GFP_KERNEL|__GFP_NOWARN);
2381 if (!p->numa_faults)
2384 p->total_numa_faults = 0;
2385 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
2389 * First accesses are treated as private, otherwise consider accesses
2390 * to be private if the accessing pid has not changed
2392 if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) {
2395 priv = cpupid_match_pid(p, last_cpupid);
2396 if (!priv && !(flags & TNF_NO_GROUP))
2397 task_numa_group(p, last_cpupid, flags, &priv);
2401 * If a workload spans multiple NUMA nodes, a shared fault that
2402 * occurs wholly within the set of nodes that the workload is
2403 * actively using should be counted as local. This allows the
2404 * scan rate to slow down when a workload has settled down.
2407 if (!priv && !local && ng && ng->active_nodes > 1 &&
2408 numa_is_active_node(cpu_node, ng) &&
2409 numa_is_active_node(mem_node, ng))
2412 task_numa_placement(p);
2415 * Retry task to preferred node migration periodically, in case it
2416 * case it previously failed, or the scheduler moved us.
2418 if (time_after(jiffies, p->numa_migrate_retry))
2419 numa_migrate_preferred(p);
2422 p->numa_pages_migrated += pages;
2423 if (flags & TNF_MIGRATE_FAIL)
2424 p->numa_faults_locality[2] += pages;
2426 p->numa_faults[task_faults_idx(NUMA_MEMBUF, mem_node, priv)] += pages;
2427 p->numa_faults[task_faults_idx(NUMA_CPUBUF, cpu_node, priv)] += pages;
2428 p->numa_faults_locality[local] += pages;
2431 static void reset_ptenuma_scan(struct task_struct *p)
2434 * We only did a read acquisition of the mmap sem, so
2435 * p->mm->numa_scan_seq is written to without exclusive access
2436 * and the update is not guaranteed to be atomic. That's not
2437 * much of an issue though, since this is just used for
2438 * statistical sampling. Use READ_ONCE/WRITE_ONCE, which are not
2439 * expensive, to avoid any form of compiler optimizations:
2441 WRITE_ONCE(p->mm->numa_scan_seq, READ_ONCE(p->mm->numa_scan_seq) + 1);
2442 p->mm->numa_scan_offset = 0;
2446 * The expensive part of numa migration is done from task_work context.
2447 * Triggered from task_tick_numa().
2449 void task_numa_work(struct callback_head *work)
2451 unsigned long migrate, next_scan, now = jiffies;
2452 struct task_struct *p = current;
2453 struct mm_struct *mm = p->mm;
2454 u64 runtime = p->se.sum_exec_runtime;
2455 struct vm_area_struct *vma;
2456 unsigned long start, end;
2457 unsigned long nr_pte_updates = 0;
2458 long pages, virtpages;
2460 SCHED_WARN_ON(p != container_of(work, struct task_struct, numa_work));
2462 work->next = work; /* protect against double add */
2464 * Who cares about NUMA placement when they're dying.
2466 * NOTE: make sure not to dereference p->mm before this check,
2467 * exit_task_work() happens _after_ exit_mm() so we could be called
2468 * without p->mm even though we still had it when we enqueued this
2471 if (p->flags & PF_EXITING)
2474 if (!mm->numa_next_scan) {
2475 mm->numa_next_scan = now +
2476 msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2480 * Enforce maximal scan/migration frequency..
2482 migrate = mm->numa_next_scan;
2483 if (time_before(now, migrate))
2486 if (p->numa_scan_period == 0) {
2487 p->numa_scan_period_max = task_scan_max(p);
2488 p->numa_scan_period = task_scan_start(p);
2491 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
2492 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
2496 * Delay this task enough that another task of this mm will likely win
2497 * the next time around.
2499 p->node_stamp += 2 * TICK_NSEC;
2501 start = mm->numa_scan_offset;
2502 pages = sysctl_numa_balancing_scan_size;
2503 pages <<= 20 - PAGE_SHIFT; /* MB in pages */
2504 virtpages = pages * 8; /* Scan up to this much virtual space */
2509 if (!down_read_trylock(&mm->mmap_sem))
2511 vma = find_vma(mm, start);
2513 reset_ptenuma_scan(p);
2517 for (; vma; vma = vma->vm_next) {
2518 if (!vma_migratable(vma) || !vma_policy_mof(vma) ||
2519 is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_MIXEDMAP)) {
2524 * Shared library pages mapped by multiple processes are not
2525 * migrated as it is expected they are cache replicated. Avoid
2526 * hinting faults in read-only file-backed mappings or the vdso
2527 * as migrating the pages will be of marginal benefit.
2530 (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ)))
2534 * Skip inaccessible VMAs to avoid any confusion between
2535 * PROT_NONE and NUMA hinting ptes
2537 if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)))
2541 start = max(start, vma->vm_start);
2542 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
2543 end = min(end, vma->vm_end);
2544 nr_pte_updates = change_prot_numa(vma, start, end);
2547 * Try to scan sysctl_numa_balancing_size worth of
2548 * hpages that have at least one present PTE that
2549 * is not already pte-numa. If the VMA contains
2550 * areas that are unused or already full of prot_numa
2551 * PTEs, scan up to virtpages, to skip through those
2555 pages -= (end - start) >> PAGE_SHIFT;
2556 virtpages -= (end - start) >> PAGE_SHIFT;
2559 if (pages <= 0 || virtpages <= 0)
2563 } while (end != vma->vm_end);
2568 * It is possible to reach the end of the VMA list but the last few
2569 * VMAs are not guaranteed to the vma_migratable. If they are not, we
2570 * would find the !migratable VMA on the next scan but not reset the
2571 * scanner to the start so check it now.
2574 mm->numa_scan_offset = start;
2576 reset_ptenuma_scan(p);
2577 up_read(&mm->mmap_sem);
2580 * Make sure tasks use at least 32x as much time to run other code
2581 * than they used here, to limit NUMA PTE scanning overhead to 3% max.
2582 * Usually update_task_scan_period slows down scanning enough; on an
2583 * overloaded system we need to limit overhead on a per task basis.
2585 if (unlikely(p->se.sum_exec_runtime != runtime)) {
2586 u64 diff = p->se.sum_exec_runtime - runtime;
2587 p->node_stamp += 32 * diff;
2592 * Drive the periodic memory faults..
2594 void task_tick_numa(struct rq *rq, struct task_struct *curr)
2596 struct callback_head *work = &curr->numa_work;
2600 * We don't care about NUMA placement if we don't have memory.
2602 if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
2606 * Using runtime rather than walltime has the dual advantage that
2607 * we (mostly) drive the selection from busy threads and that the
2608 * task needs to have done some actual work before we bother with
2611 now = curr->se.sum_exec_runtime;
2612 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
2614 if (now > curr->node_stamp + period) {
2615 if (!curr->node_stamp)
2616 curr->numa_scan_period = task_scan_start(curr);
2617 curr->node_stamp += period;
2619 if (!time_before(jiffies, curr->mm->numa_next_scan)) {
2620 init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */
2621 task_work_add(curr, work, true);
2627 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
2631 static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p)
2635 static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p)
2639 #endif /* CONFIG_NUMA_BALANCING */
2642 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2644 update_load_add(&cfs_rq->load, se->load.weight);
2645 if (!parent_entity(se))
2646 update_load_add(&rq_of(cfs_rq)->load, se->load.weight);
2648 if (entity_is_task(se)) {
2649 struct rq *rq = rq_of(cfs_rq);
2651 account_numa_enqueue(rq, task_of(se));
2652 list_add(&se->group_node, &rq->cfs_tasks);
2655 cfs_rq->nr_running++;
2659 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2661 update_load_sub(&cfs_rq->load, se->load.weight);
2662 if (!parent_entity(se))
2663 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
2665 if (entity_is_task(se)) {
2666 account_numa_dequeue(rq_of(cfs_rq), task_of(se));
2667 list_del_init(&se->group_node);
2670 cfs_rq->nr_running--;
2674 * Signed add and clamp on underflow.
2676 * Explicitly do a load-store to ensure the intermediate value never hits
2677 * memory. This allows lockless observations without ever seeing the negative
2680 #define add_positive(_ptr, _val) do { \
2681 typeof(_ptr) ptr = (_ptr); \
2682 typeof(_val) val = (_val); \
2683 typeof(*ptr) res, var = READ_ONCE(*ptr); \
2687 if (val < 0 && res > var) \
2690 WRITE_ONCE(*ptr, res); \
2694 * Unsigned subtract and clamp on underflow.
2696 * Explicitly do a load-store to ensure the intermediate value never hits
2697 * memory. This allows lockless observations without ever seeing the negative
2700 #define sub_positive(_ptr, _val) do { \
2701 typeof(_ptr) ptr = (_ptr); \
2702 typeof(*ptr) val = (_val); \
2703 typeof(*ptr) res, var = READ_ONCE(*ptr); \
2707 WRITE_ONCE(*ptr, res); \
2712 * XXX we want to get rid of these helpers and use the full load resolution.
2714 static inline long se_weight(struct sched_entity *se)
2716 return scale_load_down(se->load.weight);
2719 static inline long se_runnable(struct sched_entity *se)
2721 return scale_load_down(se->runnable_weight);
2725 enqueue_runnable_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2727 cfs_rq->runnable_weight += se->runnable_weight;
2729 cfs_rq->avg.runnable_load_avg += se->avg.runnable_load_avg;
2730 cfs_rq->avg.runnable_load_sum += se_runnable(se) * se->avg.runnable_load_sum;
2734 dequeue_runnable_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2736 cfs_rq->runnable_weight -= se->runnable_weight;
2738 sub_positive(&cfs_rq->avg.runnable_load_avg, se->avg.runnable_load_avg);
2739 sub_positive(&cfs_rq->avg.runnable_load_sum,
2740 se_runnable(se) * se->avg.runnable_load_sum);
2744 enqueue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2746 cfs_rq->avg.load_avg += se->avg.load_avg;
2747 cfs_rq->avg.load_sum += se_weight(se) * se->avg.load_sum;
2751 dequeue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2753 sub_positive(&cfs_rq->avg.load_avg, se->avg.load_avg);
2754 sub_positive(&cfs_rq->avg.load_sum, se_weight(se) * se->avg.load_sum);
2758 enqueue_runnable_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
2760 dequeue_runnable_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
2762 enqueue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
2764 dequeue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
2767 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
2768 unsigned long weight, unsigned long runnable)
2771 /* commit outstanding execution time */
2772 if (cfs_rq->curr == se)
2773 update_curr(cfs_rq);
2774 account_entity_dequeue(cfs_rq, se);
2775 dequeue_runnable_load_avg(cfs_rq, se);
2777 dequeue_load_avg(cfs_rq, se);
2779 se->runnable_weight = runnable;
2780 update_load_set(&se->load, weight);
2784 u32 divider = LOAD_AVG_MAX - 1024 + se->avg.period_contrib;
2786 se->avg.load_avg = div_u64(se_weight(se) * se->avg.load_sum, divider);
2787 se->avg.runnable_load_avg =
2788 div_u64(se_runnable(se) * se->avg.runnable_load_sum, divider);
2792 enqueue_load_avg(cfs_rq, se);
2794 account_entity_enqueue(cfs_rq, se);
2795 enqueue_runnable_load_avg(cfs_rq, se);
2799 void reweight_task(struct task_struct *p, int prio)
2801 struct sched_entity *se = &p->se;
2802 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2803 struct load_weight *load = &se->load;
2804 unsigned long weight = scale_load(sched_prio_to_weight[prio]);
2806 reweight_entity(cfs_rq, se, weight, weight);
2807 load->inv_weight = sched_prio_to_wmult[prio];
2810 #ifdef CONFIG_FAIR_GROUP_SCHED
2813 * All this does is approximate the hierarchical proportion which includes that
2814 * global sum we all love to hate.
2816 * That is, the weight of a group entity, is the proportional share of the
2817 * group weight based on the group runqueue weights. That is:
2819 * tg->weight * grq->load.weight
2820 * ge->load.weight = ----------------------------- (1)
2821 * \Sum grq->load.weight
2823 * Now, because computing that sum is prohibitively expensive to compute (been
2824 * there, done that) we approximate it with this average stuff. The average
2825 * moves slower and therefore the approximation is cheaper and more stable.
2827 * So instead of the above, we substitute:
2829 * grq->load.weight -> grq->avg.load_avg (2)
2831 * which yields the following:
2833 * tg->weight * grq->avg.load_avg
2834 * ge->load.weight = ------------------------------ (3)
2837 * Where: tg->load_avg ~= \Sum grq->avg.load_avg
2839 * That is shares_avg, and it is right (given the approximation (2)).
2841 * The problem with it is that because the average is slow -- it was designed
2842 * to be exactly that of course -- this leads to transients in boundary
2843 * conditions. In specific, the case where the group was idle and we start the
2844 * one task. It takes time for our CPU's grq->avg.load_avg to build up,
2845 * yielding bad latency etc..
2847 * Now, in that special case (1) reduces to:
2849 * tg->weight * grq->load.weight
2850 * ge->load.weight = ----------------------------- = tg->weight (4)
2853 * That is, the sum collapses because all other CPUs are idle; the UP scenario.
2855 * So what we do is modify our approximation (3) to approach (4) in the (near)
2860 * tg->weight * grq->load.weight
2861 * --------------------------------------------------- (5)
2862 * tg->load_avg - grq->avg.load_avg + grq->load.weight
2864 * But because grq->load.weight can drop to 0, resulting in a divide by zero,
2865 * we need to use grq->avg.load_avg as its lower bound, which then gives:
2868 * tg->weight * grq->load.weight
2869 * ge->load.weight = ----------------------------- (6)
2874 * tg_load_avg' = tg->load_avg - grq->avg.load_avg +
2875 * max(grq->load.weight, grq->avg.load_avg)
2877 * And that is shares_weight and is icky. In the (near) UP case it approaches
2878 * (4) while in the normal case it approaches (3). It consistently
2879 * overestimates the ge->load.weight and therefore:
2881 * \Sum ge->load.weight >= tg->weight
2885 static long calc_group_shares(struct cfs_rq *cfs_rq)
2887 long tg_weight, tg_shares, load, shares;
2888 struct task_group *tg = cfs_rq->tg;
2890 tg_shares = READ_ONCE(tg->shares);
2892 load = max(scale_load_down(cfs_rq->load.weight), cfs_rq->avg.load_avg);
2894 tg_weight = atomic_long_read(&tg->load_avg);
2896 /* Ensure tg_weight >= load */
2897 tg_weight -= cfs_rq->tg_load_avg_contrib;
2900 shares = (tg_shares * load);
2902 shares /= tg_weight;
2905 * MIN_SHARES has to be unscaled here to support per-CPU partitioning
2906 * of a group with small tg->shares value. It is a floor value which is
2907 * assigned as a minimum load.weight to the sched_entity representing
2908 * the group on a CPU.
2910 * E.g. on 64-bit for a group with tg->shares of scale_load(15)=15*1024
2911 * on an 8-core system with 8 tasks each runnable on one CPU shares has
2912 * to be 15*1024*1/8=1920 instead of scale_load(MIN_SHARES)=2*1024. In
2913 * case no task is runnable on a CPU MIN_SHARES=2 should be returned
2916 return clamp_t(long, shares, MIN_SHARES, tg_shares);
2920 * This calculates the effective runnable weight for a group entity based on
2921 * the group entity weight calculated above.
2923 * Because of the above approximation (2), our group entity weight is
2924 * an load_avg based ratio (3). This means that it includes blocked load and
2925 * does not represent the runnable weight.
2927 * Approximate the group entity's runnable weight per ratio from the group
2930 * grq->avg.runnable_load_avg
2931 * ge->runnable_weight = ge->load.weight * -------------------------- (7)
2934 * However, analogous to above, since the avg numbers are slow, this leads to
2935 * transients in the from-idle case. Instead we use:
2937 * ge->runnable_weight = ge->load.weight *
2939 * max(grq->avg.runnable_load_avg, grq->runnable_weight)
2940 * ----------------------------------------------------- (8)
2941 * max(grq->avg.load_avg, grq->load.weight)
2943 * Where these max() serve both to use the 'instant' values to fix the slow
2944 * from-idle and avoid the /0 on to-idle, similar to (6).
2946 static long calc_group_runnable(struct cfs_rq *cfs_rq, long shares)
2948 long runnable, load_avg;
2950 load_avg = max(cfs_rq->avg.load_avg,
2951 scale_load_down(cfs_rq->load.weight));
2953 runnable = max(cfs_rq->avg.runnable_load_avg,
2954 scale_load_down(cfs_rq->runnable_weight));
2958 runnable /= load_avg;
2960 return clamp_t(long, runnable, MIN_SHARES, shares);
2962 #endif /* CONFIG_SMP */
2964 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
2967 * Recomputes the group entity based on the current state of its group
2970 static void update_cfs_group(struct sched_entity *se)
2972 struct cfs_rq *gcfs_rq = group_cfs_rq(se);
2973 long shares, runnable;
2978 if (throttled_hierarchy(gcfs_rq))
2982 runnable = shares = READ_ONCE(gcfs_rq->tg->shares);
2984 if (likely(se->load.weight == shares))
2987 shares = calc_group_shares(gcfs_rq);
2988 runnable = calc_group_runnable(gcfs_rq, shares);
2991 reweight_entity(cfs_rq_of(se), se, shares, runnable);
2994 #else /* CONFIG_FAIR_GROUP_SCHED */
2995 static inline void update_cfs_group(struct sched_entity *se)
2998 #endif /* CONFIG_FAIR_GROUP_SCHED */
3000 static inline void cfs_rq_util_change(struct cfs_rq *cfs_rq, int flags)
3002 struct rq *rq = rq_of(cfs_rq);
3004 if (&rq->cfs == cfs_rq || (flags & SCHED_CPUFREQ_MIGRATION)) {
3006 * There are a few boundary cases this might miss but it should
3007 * get called often enough that that should (hopefully) not be
3010 * It will not get called when we go idle, because the idle
3011 * thread is a different class (!fair), nor will the utilization
3012 * number include things like RT tasks.
3014 * As is, the util number is not freq-invariant (we'd have to
3015 * implement arch_scale_freq_capacity() for that).
3019 cpufreq_update_util(rq, flags);
3026 * val * y^n, where y^32 ~= 0.5 (~1 scheduling period)
3028 static u64 decay_load(u64 val, u64 n)
3030 unsigned int local_n;
3032 if (unlikely(n > LOAD_AVG_PERIOD * 63))
3035 /* after bounds checking we can collapse to 32-bit */
3039 * As y^PERIOD = 1/2, we can combine
3040 * y^n = 1/2^(n/PERIOD) * y^(n%PERIOD)
3041 * With a look-up table which covers y^n (n<PERIOD)
3043 * To achieve constant time decay_load.
3045 if (unlikely(local_n >= LOAD_AVG_PERIOD)) {
3046 val >>= local_n / LOAD_AVG_PERIOD;
3047 local_n %= LOAD_AVG_PERIOD;
3050 val = mul_u64_u32_shr(val, runnable_avg_yN_inv[local_n], 32);
3054 static u32 __accumulate_pelt_segments(u64 periods, u32 d1, u32 d3)
3056 u32 c1, c2, c3 = d3; /* y^0 == 1 */
3061 c1 = decay_load((u64)d1, periods);
3065 * c2 = 1024 \Sum y^n
3069 * = 1024 ( \Sum y^n - \Sum y^n - y^0 )
3072 c2 = LOAD_AVG_MAX - decay_load(LOAD_AVG_MAX, periods) - 1024;
3074 return c1 + c2 + c3;
3078 * Accumulate the three separate parts of the sum; d1 the remainder
3079 * of the last (incomplete) period, d2 the span of full periods and d3
3080 * the remainder of the (incomplete) current period.
3085 * |<->|<----------------->|<--->|
3086 * ... |---x---|------| ... |------|-----x (now)
3089 * u' = (u + d1) y^p + 1024 \Sum y^n + d3 y^0
3092 * = u y^p + (Step 1)
3095 * d1 y^p + 1024 \Sum y^n + d3 y^0 (Step 2)
3098 static __always_inline u32
3099 accumulate_sum(u64 delta, int cpu, struct sched_avg *sa,
3100 unsigned long load, unsigned long runnable, int running)
3102 unsigned long scale_freq, scale_cpu;
3103 u32 contrib = (u32)delta; /* p == 0 -> delta < 1024 */
3106 scale_freq = arch_scale_freq_capacity(cpu);
3107 scale_cpu = arch_scale_cpu_capacity(NULL, cpu);
3109 delta += sa->period_contrib;
3110 periods = delta / 1024; /* A period is 1024us (~1ms) */
3113 * Step 1: decay old *_sum if we crossed period boundaries.
3116 sa->load_sum = decay_load(sa->load_sum, periods);
3117 sa->runnable_load_sum =
3118 decay_load(sa->runnable_load_sum, periods);
3119 sa->util_sum = decay_load((u64)(sa->util_sum), periods);
3125 contrib = __accumulate_pelt_segments(periods,
3126 1024 - sa->period_contrib, delta);
3128 sa->period_contrib = delta;
3130 contrib = cap_scale(contrib, scale_freq);
3132 sa->load_sum += load * contrib;
3134 sa->runnable_load_sum += runnable * contrib;
3136 sa->util_sum += contrib * scale_cpu;
3142 * We can represent the historical contribution to runnable average as the
3143 * coefficients of a geometric series. To do this we sub-divide our runnable
3144 * history into segments of approximately 1ms (1024us); label the segment that
3145 * occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
3147 * [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
3149 * (now) (~1ms ago) (~2ms ago)
3151 * Let u_i denote the fraction of p_i that the entity was runnable.
3153 * We then designate the fractions u_i as our co-efficients, yielding the
3154 * following representation of historical load:
3155 * u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
3157 * We choose y based on the with of a reasonably scheduling period, fixing:
3160 * This means that the contribution to load ~32ms ago (u_32) will be weighted
3161 * approximately half as much as the contribution to load within the last ms
3164 * When a period "rolls over" and we have new u_0`, multiplying the previous
3165 * sum again by y is sufficient to update:
3166 * load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
3167 * = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
3169 static __always_inline int
3170 ___update_load_sum(u64 now, int cpu, struct sched_avg *sa,
3171 unsigned long load, unsigned long runnable, int running)
3175 delta = now - sa->last_update_time;
3177 * This should only happen when time goes backwards, which it
3178 * unfortunately does during sched clock init when we swap over to TSC.
3180 if ((s64)delta < 0) {
3181 sa->last_update_time = now;
3186 * Use 1024ns as the unit of measurement since it's a reasonable
3187 * approximation of 1us and fast to compute.
3193 sa->last_update_time += delta << 10;
3196 * running is a subset of runnable (weight) so running can't be set if
3197 * runnable is clear. But there are some corner cases where the current
3198 * se has been already dequeued but cfs_rq->curr still points to it.
3199 * This means that weight will be 0 but not running for a sched_entity
3200 * but also for a cfs_rq if the latter becomes idle. As an example,
3201 * this happens during idle_balance() which calls
3202 * update_blocked_averages()
3205 runnable = running = 0;
3208 * Now we know we crossed measurement unit boundaries. The *_avg
3209 * accrues by two steps:
3211 * Step 1: accumulate *_sum since last_update_time. If we haven't
3212 * crossed period boundaries, finish.
3214 if (!accumulate_sum(delta, cpu, sa, load, runnable, running))
3220 static __always_inline void
3221 ___update_load_avg(struct sched_avg *sa, unsigned long load, unsigned long runnable)
3223 u32 divider = LOAD_AVG_MAX - 1024 + sa->period_contrib;
3226 * Step 2: update *_avg.
3228 sa->load_avg = div_u64(load * sa->load_sum, divider);
3229 sa->runnable_load_avg = div_u64(runnable * sa->runnable_load_sum, divider);
3230 sa->util_avg = sa->util_sum / divider;
3234 * When a task is dequeued, its estimated utilization should not be update if
3235 * its util_avg has not been updated at least once.
3236 * This flag is used to synchronize util_avg updates with util_est updates.
3237 * We map this information into the LSB bit of the utilization saved at
3238 * dequeue time (i.e. util_est.dequeued).
3240 #define UTIL_AVG_UNCHANGED 0x1
3242 static inline void cfs_se_util_change(struct sched_avg *avg)
3244 unsigned int enqueued;
3246 if (!sched_feat(UTIL_EST))
3249 /* Avoid store if the flag has been already set */
3250 enqueued = avg->util_est.enqueued;
3251 if (!(enqueued & UTIL_AVG_UNCHANGED))
3254 /* Reset flag to report util_avg has been updated */
3255 enqueued &= ~UTIL_AVG_UNCHANGED;
3256 WRITE_ONCE(avg->util_est.enqueued, enqueued);
3263 * se_runnable() == se_weight()
3265 * group: [ see update_cfs_group() ]
3266 * se_weight() = tg->weight * grq->load_avg / tg->load_avg
3267 * se_runnable() = se_weight(se) * grq->runnable_load_avg / grq->load_avg
3269 * load_sum := runnable_sum
3270 * load_avg = se_weight(se) * runnable_avg
3272 * runnable_load_sum := runnable_sum
3273 * runnable_load_avg = se_runnable(se) * runnable_avg
3275 * XXX collapse load_sum and runnable_load_sum
3279 * load_sum = \Sum se_weight(se) * se->avg.load_sum
3280 * load_avg = \Sum se->avg.load_avg
3282 * runnable_load_sum = \Sum se_runnable(se) * se->avg.runnable_load_sum
3283 * runnable_load_avg = \Sum se->avg.runable_load_avg
3287 __update_load_avg_blocked_se(u64 now, int cpu, struct sched_entity *se)
3289 if (entity_is_task(se))
3290 se->runnable_weight = se->load.weight;
3292 if (___update_load_sum(now, cpu, &se->avg, 0, 0, 0)) {
3293 ___update_load_avg(&se->avg, se_weight(se), se_runnable(se));
3301 __update_load_avg_se(u64 now, int cpu, struct cfs_rq *cfs_rq, struct sched_entity *se)
3303 if (entity_is_task(se))
3304 se->runnable_weight = se->load.weight;
3306 if (___update_load_sum(now, cpu, &se->avg, !!se->on_rq, !!se->on_rq,
3307 cfs_rq->curr == se)) {
3309 ___update_load_avg(&se->avg, se_weight(se), se_runnable(se));
3310 cfs_se_util_change(&se->avg);
3318 __update_load_avg_cfs_rq(u64 now, int cpu, struct cfs_rq *cfs_rq)
3320 if (___update_load_sum(now, cpu, &cfs_rq->avg,
3321 scale_load_down(cfs_rq->load.weight),
3322 scale_load_down(cfs_rq->runnable_weight),
3323 cfs_rq->curr != NULL)) {
3325 ___update_load_avg(&cfs_rq->avg, 1, 1);
3332 #ifdef CONFIG_FAIR_GROUP_SCHED
3334 * update_tg_load_avg - update the tg's load avg
3335 * @cfs_rq: the cfs_rq whose avg changed
3336 * @force: update regardless of how small the difference
3338 * This function 'ensures': tg->load_avg := \Sum tg->cfs_rq[]->avg.load.
3339 * However, because tg->load_avg is a global value there are performance
3342 * In order to avoid having to look at the other cfs_rq's, we use a
3343 * differential update where we store the last value we propagated. This in
3344 * turn allows skipping updates if the differential is 'small'.
3346 * Updating tg's load_avg is necessary before update_cfs_share().
3348 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force)
3350 long delta = cfs_rq->avg.load_avg - cfs_rq->tg_load_avg_contrib;
3353 * No need to update load_avg for root_task_group as it is not used.
3355 if (cfs_rq->tg == &root_task_group)
3358 if (force || abs(delta) > cfs_rq->tg_load_avg_contrib / 64) {
3359 atomic_long_add(delta, &cfs_rq->tg->load_avg);
3360 cfs_rq->tg_load_avg_contrib = cfs_rq->avg.load_avg;
3365 * Called within set_task_rq() right before setting a task's CPU. The
3366 * caller only guarantees p->pi_lock is held; no other assumptions,
3367 * including the state of rq->lock, should be made.
3369 void set_task_rq_fair(struct sched_entity *se,
3370 struct cfs_rq *prev, struct cfs_rq *next)
3372 u64 p_last_update_time;
3373 u64 n_last_update_time;
3375 if (!sched_feat(ATTACH_AGE_LOAD))
3379 * We are supposed to update the task to "current" time, then its up to
3380 * date and ready to go to new CPU/cfs_rq. But we have difficulty in
3381 * getting what current time is, so simply throw away the out-of-date
3382 * time. This will result in the wakee task is less decayed, but giving
3383 * the wakee more load sounds not bad.
3385 if (!(se->avg.last_update_time && prev))
3388 #ifndef CONFIG_64BIT
3390 u64 p_last_update_time_copy;
3391 u64 n_last_update_time_copy;
3394 p_last_update_time_copy = prev->load_last_update_time_copy;
3395 n_last_update_time_copy = next->load_last_update_time_copy;
3399 p_last_update_time = prev->avg.last_update_time;
3400 n_last_update_time = next->avg.last_update_time;
3402 } while (p_last_update_time != p_last_update_time_copy ||
3403 n_last_update_time != n_last_update_time_copy);
3406 p_last_update_time = prev->avg.last_update_time;
3407 n_last_update_time = next->avg.last_update_time;
3409 __update_load_avg_blocked_se(p_last_update_time, cpu_of(rq_of(prev)), se);
3410 se->avg.last_update_time = n_last_update_time;
3415 * When on migration a sched_entity joins/leaves the PELT hierarchy, we need to
3416 * propagate its contribution. The key to this propagation is the invariant
3417 * that for each group:
3419 * ge->avg == grq->avg (1)
3421 * _IFF_ we look at the pure running and runnable sums. Because they
3422 * represent the very same entity, just at different points in the hierarchy.
3424 * Per the above update_tg_cfs_util() is trivial and simply copies the running
3425 * sum over (but still wrong, because the group entity and group rq do not have
3426 * their PELT windows aligned).
3428 * However, update_tg_cfs_runnable() is more complex. So we have:
3430 * ge->avg.load_avg = ge->load.weight * ge->avg.runnable_avg (2)
3432 * And since, like util, the runnable part should be directly transferable,
3433 * the following would _appear_ to be the straight forward approach:
3435 * grq->avg.load_avg = grq->load.weight * grq->avg.runnable_avg (3)
3437 * And per (1) we have:
3439 * ge->avg.runnable_avg == grq->avg.runnable_avg
3443 * ge->load.weight * grq->avg.load_avg
3444 * ge->avg.load_avg = ----------------------------------- (4)
3447 * Except that is wrong!
3449 * Because while for entities historical weight is not important and we
3450 * really only care about our future and therefore can consider a pure
3451 * runnable sum, runqueues can NOT do this.
3453 * We specifically want runqueues to have a load_avg that includes
3454 * historical weights. Those represent the blocked load, the load we expect
3455 * to (shortly) return to us. This only works by keeping the weights as
3456 * integral part of the sum. We therefore cannot decompose as per (3).
3458 * Another reason this doesn't work is that runnable isn't a 0-sum entity.
3459 * Imagine a rq with 2 tasks that each are runnable 2/3 of the time. Then the
3460 * rq itself is runnable anywhere between 2/3 and 1 depending on how the
3461 * runnable section of these tasks overlap (or not). If they were to perfectly
3462 * align the rq as a whole would be runnable 2/3 of the time. If however we
3463 * always have at least 1 runnable task, the rq as a whole is always runnable.
3465 * So we'll have to approximate.. :/
3467 * Given the constraint:
3469 * ge->avg.running_sum <= ge->avg.runnable_sum <= LOAD_AVG_MAX
3471 * We can construct a rule that adds runnable to a rq by assuming minimal
3474 * On removal, we'll assume each task is equally runnable; which yields:
3476 * grq->avg.runnable_sum = grq->avg.load_sum / grq->load.weight
3478 * XXX: only do this for the part of runnable > running ?
3483 update_tg_cfs_util(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq)
3485 long delta = gcfs_rq->avg.util_avg - se->avg.util_avg;
3487 /* Nothing to update */
3492 * The relation between sum and avg is:
3494 * LOAD_AVG_MAX - 1024 + sa->period_contrib
3496 * however, the PELT windows are not aligned between grq and gse.
3499 /* Set new sched_entity's utilization */
3500 se->avg.util_avg = gcfs_rq->avg.util_avg;
3501 se->avg.util_sum = se->avg.util_avg * LOAD_AVG_MAX;
3503 /* Update parent cfs_rq utilization */
3504 add_positive(&cfs_rq->avg.util_avg, delta);
3505 cfs_rq->avg.util_sum = cfs_rq->avg.util_avg * LOAD_AVG_MAX;
3509 update_tg_cfs_runnable(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq)
3511 long delta_avg, running_sum, runnable_sum = gcfs_rq->prop_runnable_sum;
3512 unsigned long runnable_load_avg, load_avg;
3513 u64 runnable_load_sum, load_sum = 0;
3519 gcfs_rq->prop_runnable_sum = 0;
3521 if (runnable_sum >= 0) {
3523 * Add runnable; clip at LOAD_AVG_MAX. Reflects that until
3524 * the CPU is saturated running == runnable.
3526 runnable_sum += se->avg.load_sum;
3527 runnable_sum = min(runnable_sum, (long)LOAD_AVG_MAX);
3530 * Estimate the new unweighted runnable_sum of the gcfs_rq by
3531 * assuming all tasks are equally runnable.
3533 if (scale_load_down(gcfs_rq->load.weight)) {
3534 load_sum = div_s64(gcfs_rq->avg.load_sum,
3535 scale_load_down(gcfs_rq->load.weight));
3538 /* But make sure to not inflate se's runnable */
3539 runnable_sum = min(se->avg.load_sum, load_sum);
3543 * runnable_sum can't be lower than running_sum
3544 * As running sum is scale with CPU capacity wehreas the runnable sum
3545 * is not we rescale running_sum 1st
3547 running_sum = se->avg.util_sum /
3548 arch_scale_cpu_capacity(NULL, cpu_of(rq_of(cfs_rq)));
3549 runnable_sum = max(runnable_sum, running_sum);
3551 load_sum = (s64)se_weight(se) * runnable_sum;
3552 load_avg = div_s64(load_sum, LOAD_AVG_MAX);
3554 delta_sum = load_sum - (s64)se_weight(se) * se->avg.load_sum;
3555 delta_avg = load_avg - se->avg.load_avg;
3557 se->avg.load_sum = runnable_sum;
3558 se->avg.load_avg = load_avg;
3559 add_positive(&cfs_rq->avg.load_avg, delta_avg);
3560 add_positive(&cfs_rq->avg.load_sum, delta_sum);
3562 runnable_load_sum = (s64)se_runnable(se) * runnable_sum;
3563 runnable_load_avg = div_s64(runnable_load_sum, LOAD_AVG_MAX);
3564 delta_sum = runnable_load_sum - se_weight(se) * se->avg.runnable_load_sum;
3565 delta_avg = runnable_load_avg - se->avg.runnable_load_avg;
3567 se->avg.runnable_load_sum = runnable_sum;
3568 se->avg.runnable_load_avg = runnable_load_avg;
3571 add_positive(&cfs_rq->avg.runnable_load_avg, delta_avg);
3572 add_positive(&cfs_rq->avg.runnable_load_sum, delta_sum);
3576 static inline void add_tg_cfs_propagate(struct cfs_rq *cfs_rq, long runnable_sum)
3578 cfs_rq->propagate = 1;
3579 cfs_rq->prop_runnable_sum += runnable_sum;
3582 /* Update task and its cfs_rq load average */
3583 static inline int propagate_entity_load_avg(struct sched_entity *se)
3585 struct cfs_rq *cfs_rq, *gcfs_rq;
3587 if (entity_is_task(se))
3590 gcfs_rq = group_cfs_rq(se);
3591 if (!gcfs_rq->propagate)
3594 gcfs_rq->propagate = 0;
3596 cfs_rq = cfs_rq_of(se);
3598 add_tg_cfs_propagate(cfs_rq, gcfs_rq->prop_runnable_sum);
3600 update_tg_cfs_util(cfs_rq, se, gcfs_rq);
3601 update_tg_cfs_runnable(cfs_rq, se, gcfs_rq);
3607 * Check if we need to update the load and the utilization of a blocked
3610 static inline bool skip_blocked_update(struct sched_entity *se)
3612 struct cfs_rq *gcfs_rq = group_cfs_rq(se);
3615 * If sched_entity still have not zero load or utilization, we have to
3618 if (se->avg.load_avg || se->avg.util_avg)
3622 * If there is a pending propagation, we have to update the load and
3623 * the utilization of the sched_entity:
3625 if (gcfs_rq->propagate)
3629 * Otherwise, the load and the utilization of the sched_entity is
3630 * already zero and there is no pending propagation, so it will be a
3631 * waste of time to try to decay it:
3636 #else /* CONFIG_FAIR_GROUP_SCHED */
3638 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force) {}
3640 static inline int propagate_entity_load_avg(struct sched_entity *se)
3645 static inline void add_tg_cfs_propagate(struct cfs_rq *cfs_rq, long runnable_sum) {}
3647 #endif /* CONFIG_FAIR_GROUP_SCHED */
3650 * update_cfs_rq_load_avg - update the cfs_rq's load/util averages
3651 * @now: current time, as per cfs_rq_clock_task()
3652 * @cfs_rq: cfs_rq to update
3654 * The cfs_rq avg is the direct sum of all its entities (blocked and runnable)
3655 * avg. The immediate corollary is that all (fair) tasks must be attached, see
3656 * post_init_entity_util_avg().
3658 * cfs_rq->avg is used for task_h_load() and update_cfs_share() for example.
3660 * Returns true if the load decayed or we removed load.
3662 * Since both these conditions indicate a changed cfs_rq->avg.load we should
3663 * call update_tg_load_avg() when this function returns true.
3666 update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq)
3668 unsigned long removed_load = 0, removed_util = 0, removed_runnable_sum = 0;
3669 struct sched_avg *sa = &cfs_rq->avg;
3672 if (cfs_rq->removed.nr) {
3674 u32 divider = LOAD_AVG_MAX - 1024 + sa->period_contrib;
3676 raw_spin_lock(&cfs_rq->removed.lock);
3677 swap(cfs_rq->removed.util_avg, removed_util);
3678 swap(cfs_rq->removed.load_avg, removed_load);
3679 swap(cfs_rq->removed.runnable_sum, removed_runnable_sum);
3680 cfs_rq->removed.nr = 0;
3681 raw_spin_unlock(&cfs_rq->removed.lock);
3684 sub_positive(&sa->load_avg, r);
3685 sub_positive(&sa->load_sum, r * divider);
3688 sub_positive(&sa->util_avg, r);
3689 sub_positive(&sa->util_sum, r * divider);
3691 add_tg_cfs_propagate(cfs_rq, -(long)removed_runnable_sum);
3696 decayed |= __update_load_avg_cfs_rq(now, cpu_of(rq_of(cfs_rq)), cfs_rq);
3698 #ifndef CONFIG_64BIT
3700 cfs_rq->load_last_update_time_copy = sa->last_update_time;
3704 cfs_rq_util_change(cfs_rq, 0);
3710 * attach_entity_load_avg - attach this entity to its cfs_rq load avg
3711 * @cfs_rq: cfs_rq to attach to
3712 * @se: sched_entity to attach
3714 * Must call update_cfs_rq_load_avg() before this, since we rely on
3715 * cfs_rq->avg.last_update_time being current.
3717 static void attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3719 u32 divider = LOAD_AVG_MAX - 1024 + cfs_rq->avg.period_contrib;
3722 * When we attach the @se to the @cfs_rq, we must align the decay
3723 * window because without that, really weird and wonderful things can
3728 se->avg.last_update_time = cfs_rq->avg.last_update_time;
3729 se->avg.period_contrib = cfs_rq->avg.period_contrib;
3732 * Hell(o) Nasty stuff.. we need to recompute _sum based on the new
3733 * period_contrib. This isn't strictly correct, but since we're
3734 * entirely outside of the PELT hierarchy, nobody cares if we truncate
3737 se->avg.util_sum = se->avg.util_avg * divider;
3739 se->avg.load_sum = divider;
3740 if (se_weight(se)) {
3742 div_u64(se->avg.load_avg * se->avg.load_sum, se_weight(se));
3745 se->avg.runnable_load_sum = se->avg.load_sum;
3747 enqueue_load_avg(cfs_rq, se);
3748 cfs_rq->avg.util_avg += se->avg.util_avg;
3749 cfs_rq->avg.util_sum += se->avg.util_sum;
3751 add_tg_cfs_propagate(cfs_rq, se->avg.load_sum);
3753 cfs_rq_util_change(cfs_rq, flags);
3757 * detach_entity_load_avg - detach this entity from its cfs_rq load avg
3758 * @cfs_rq: cfs_rq to detach from
3759 * @se: sched_entity to detach
3761 * Must call update_cfs_rq_load_avg() before this, since we rely on
3762 * cfs_rq->avg.last_update_time being current.
3764 static void detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3766 dequeue_load_avg(cfs_rq, se);
3767 sub_positive(&cfs_rq->avg.util_avg, se->avg.util_avg);
3768 sub_positive(&cfs_rq->avg.util_sum, se->avg.util_sum);
3770 add_tg_cfs_propagate(cfs_rq, -se->avg.load_sum);
3772 cfs_rq_util_change(cfs_rq, 0);
3776 * Optional action to be done while updating the load average
3778 #define UPDATE_TG 0x1
3779 #define SKIP_AGE_LOAD 0x2
3780 #define DO_ATTACH 0x4
3782 /* Update task and its cfs_rq load average */
3783 static inline void update_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3785 u64 now = cfs_rq_clock_task(cfs_rq);
3786 struct rq *rq = rq_of(cfs_rq);
3787 int cpu = cpu_of(rq);
3791 * Track task load average for carrying it to new CPU after migrated, and
3792 * track group sched_entity load average for task_h_load calc in migration
3794 if (se->avg.last_update_time && !(flags & SKIP_AGE_LOAD))
3795 __update_load_avg_se(now, cpu, cfs_rq, se);
3797 decayed = update_cfs_rq_load_avg(now, cfs_rq);
3798 decayed |= propagate_entity_load_avg(se);
3800 if (!se->avg.last_update_time && (flags & DO_ATTACH)) {
3803 * DO_ATTACH means we're here from enqueue_entity().
3804 * !last_update_time means we've passed through
3805 * migrate_task_rq_fair() indicating we migrated.
3807 * IOW we're enqueueing a task on a new CPU.
3809 attach_entity_load_avg(cfs_rq, se, SCHED_CPUFREQ_MIGRATION);
3810 update_tg_load_avg(cfs_rq, 0);
3812 } else if (decayed && (flags & UPDATE_TG))
3813 update_tg_load_avg(cfs_rq, 0);
3816 #ifndef CONFIG_64BIT
3817 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
3819 u64 last_update_time_copy;
3820 u64 last_update_time;
3823 last_update_time_copy = cfs_rq->load_last_update_time_copy;
3825 last_update_time = cfs_rq->avg.last_update_time;
3826 } while (last_update_time != last_update_time_copy);
3828 return last_update_time;
3831 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
3833 return cfs_rq->avg.last_update_time;
3838 * Synchronize entity load avg of dequeued entity without locking
3841 void sync_entity_load_avg(struct sched_entity *se)
3843 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3844 u64 last_update_time;
3846 last_update_time = cfs_rq_last_update_time(cfs_rq);
3847 __update_load_avg_blocked_se(last_update_time, cpu_of(rq_of(cfs_rq)), se);
3851 * Task first catches up with cfs_rq, and then subtract
3852 * itself from the cfs_rq (task must be off the queue now).
3854 void remove_entity_load_avg(struct sched_entity *se)
3856 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3857 unsigned long flags;
3860 * tasks cannot exit without having gone through wake_up_new_task() ->
3861 * post_init_entity_util_avg() which will have added things to the
3862 * cfs_rq, so we can remove unconditionally.
3864 * Similarly for groups, they will have passed through
3865 * post_init_entity_util_avg() before unregister_sched_fair_group()
3869 sync_entity_load_avg(se);
3871 raw_spin_lock_irqsave(&cfs_rq->removed.lock, flags);
3872 ++cfs_rq->removed.nr;
3873 cfs_rq->removed.util_avg += se->avg.util_avg;
3874 cfs_rq->removed.load_avg += se->avg.load_avg;
3875 cfs_rq->removed.runnable_sum += se->avg.load_sum; /* == runnable_sum */
3876 raw_spin_unlock_irqrestore(&cfs_rq->removed.lock, flags);
3879 static inline unsigned long cfs_rq_runnable_load_avg(struct cfs_rq *cfs_rq)
3881 return cfs_rq->avg.runnable_load_avg;
3884 static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq)
3886 return cfs_rq->avg.load_avg;
3889 static int idle_balance(struct rq *this_rq, struct rq_flags *rf);
3891 static inline unsigned long task_util(struct task_struct *p)
3893 return READ_ONCE(p->se.avg.util_avg);
3896 static inline unsigned long _task_util_est(struct task_struct *p)
3898 struct util_est ue = READ_ONCE(p->se.avg.util_est);
3900 return max(ue.ewma, ue.enqueued);
3903 static inline unsigned long task_util_est(struct task_struct *p)
3905 return max(task_util(p), _task_util_est(p));
3908 static inline void util_est_enqueue(struct cfs_rq *cfs_rq,
3909 struct task_struct *p)
3911 unsigned int enqueued;
3913 if (!sched_feat(UTIL_EST))
3916 /* Update root cfs_rq's estimated utilization */
3917 enqueued = cfs_rq->avg.util_est.enqueued;
3918 enqueued += (_task_util_est(p) | UTIL_AVG_UNCHANGED);
3919 WRITE_ONCE(cfs_rq->avg.util_est.enqueued, enqueued);
3923 * Check if a (signed) value is within a specified (unsigned) margin,
3924 * based on the observation that:
3926 * abs(x) < y := (unsigned)(x + y - 1) < (2 * y - 1)
3928 * NOTE: this only works when value + maring < INT_MAX.
3930 static inline bool within_margin(int value, int margin)
3932 return ((unsigned int)(value + margin - 1) < (2 * margin - 1));
3936 util_est_dequeue(struct cfs_rq *cfs_rq, struct task_struct *p, bool task_sleep)
3938 long last_ewma_diff;
3941 if (!sched_feat(UTIL_EST))
3945 * Update root cfs_rq's estimated utilization
3947 * If *p is the last task then the root cfs_rq's estimated utilization
3948 * of a CPU is 0 by definition.
3951 if (cfs_rq->nr_running) {
3952 ue.enqueued = cfs_rq->avg.util_est.enqueued;
3953 ue.enqueued -= min_t(unsigned int, ue.enqueued,
3954 (_task_util_est(p) | UTIL_AVG_UNCHANGED));
3956 WRITE_ONCE(cfs_rq->avg.util_est.enqueued, ue.enqueued);
3959 * Skip update of task's estimated utilization when the task has not
3960 * yet completed an activation, e.g. being migrated.
3966 * If the PELT values haven't changed since enqueue time,
3967 * skip the util_est update.
3969 ue = p->se.avg.util_est;
3970 if (ue.enqueued & UTIL_AVG_UNCHANGED)
3974 * Skip update of task's estimated utilization when its EWMA is
3975 * already ~1% close to its last activation value.
3977 ue.enqueued = (task_util(p) | UTIL_AVG_UNCHANGED);
3978 last_ewma_diff = ue.enqueued - ue.ewma;
3979 if (within_margin(last_ewma_diff, (SCHED_CAPACITY_SCALE / 100)))
3983 * Update Task's estimated utilization
3985 * When *p completes an activation we can consolidate another sample
3986 * of the task size. This is done by storing the current PELT value
3987 * as ue.enqueued and by using this value to update the Exponential
3988 * Weighted Moving Average (EWMA):
3990 * ewma(t) = w * task_util(p) + (1-w) * ewma(t-1)
3991 * = w * task_util(p) + ewma(t-1) - w * ewma(t-1)
3992 * = w * (task_util(p) - ewma(t-1)) + ewma(t-1)
3993 * = w * ( last_ewma_diff ) + ewma(t-1)
3994 * = w * (last_ewma_diff + ewma(t-1) / w)
3996 * Where 'w' is the weight of new samples, which is configured to be
3997 * 0.25, thus making w=1/4 ( >>= UTIL_EST_WEIGHT_SHIFT)
3999 ue.ewma <<= UTIL_EST_WEIGHT_SHIFT;
4000 ue.ewma += last_ewma_diff;
4001 ue.ewma >>= UTIL_EST_WEIGHT_SHIFT;
4002 WRITE_ONCE(p->se.avg.util_est, ue);
4005 #else /* CONFIG_SMP */
4008 update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq)
4013 #define UPDATE_TG 0x0
4014 #define SKIP_AGE_LOAD 0x0
4015 #define DO_ATTACH 0x0
4017 static inline void update_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int not_used1)
4019 cfs_rq_util_change(cfs_rq, 0);
4022 static inline void remove_entity_load_avg(struct sched_entity *se) {}
4025 attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) {}
4027 detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
4029 static inline int idle_balance(struct rq *rq, struct rq_flags *rf)
4035 util_est_enqueue(struct cfs_rq *cfs_rq, struct task_struct *p) {}
4038 util_est_dequeue(struct cfs_rq *cfs_rq, struct task_struct *p,
4041 #endif /* CONFIG_SMP */
4043 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
4045 #ifdef CONFIG_SCHED_DEBUG
4046 s64 d = se->vruntime - cfs_rq->min_vruntime;
4051 if (d > 3*sysctl_sched_latency)
4052 schedstat_inc(cfs_rq->nr_spread_over);
4057 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
4059 u64 vruntime = cfs_rq->min_vruntime;
4062 * The 'current' period is already promised to the current tasks,
4063 * however the extra weight of the new task will slow them down a
4064 * little, place the new task so that it fits in the slot that
4065 * stays open at the end.
4067 if (initial && sched_feat(START_DEBIT))
4068 vruntime += sched_vslice(cfs_rq, se);
4070 /* sleeps up to a single latency don't count. */
4072 unsigned long thresh = sysctl_sched_latency;
4075 * Halve their sleep time's effect, to allow
4076 * for a gentler effect of sleepers:
4078 if (sched_feat(GENTLE_FAIR_SLEEPERS))
4084 /* ensure we never gain time by being placed backwards. */
4085 se->vruntime = max_vruntime(se->vruntime, vruntime);
4088 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
4090 static inline void check_schedstat_required(void)
4092 #ifdef CONFIG_SCHEDSTATS
4093 if (schedstat_enabled())
4096 /* Force schedstat enabled if a dependent tracepoint is active */
4097 if (trace_sched_stat_wait_enabled() ||
4098 trace_sched_stat_sleep_enabled() ||
4099 trace_sched_stat_iowait_enabled() ||
4100 trace_sched_stat_blocked_enabled() ||
4101 trace_sched_stat_runtime_enabled()) {
4102 printk_deferred_once("Scheduler tracepoints stat_sleep, stat_iowait, "
4103 "stat_blocked and stat_runtime require the "
4104 "kernel parameter schedstats=enable or "
4105 "kernel.sched_schedstats=1\n");
4116 * update_min_vruntime()
4117 * vruntime -= min_vruntime
4121 * update_min_vruntime()
4122 * vruntime += min_vruntime
4124 * this way the vruntime transition between RQs is done when both
4125 * min_vruntime are up-to-date.
4129 * ->migrate_task_rq_fair() (p->state == TASK_WAKING)
4130 * vruntime -= min_vruntime
4134 * update_min_vruntime()
4135 * vruntime += min_vruntime
4137 * this way we don't have the most up-to-date min_vruntime on the originating
4138 * CPU and an up-to-date min_vruntime on the destination CPU.
4142 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
4144 bool renorm = !(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_MIGRATED);
4145 bool curr = cfs_rq->curr == se;
4148 * If we're the current task, we must renormalise before calling
4152 se->vruntime += cfs_rq->min_vruntime;
4154 update_curr(cfs_rq);
4157 * Otherwise, renormalise after, such that we're placed at the current
4158 * moment in time, instead of some random moment in the past. Being
4159 * placed in the past could significantly boost this task to the
4160 * fairness detriment of existing tasks.
4162 if (renorm && !curr)
4163 se->vruntime += cfs_rq->min_vruntime;
4166 * When enqueuing a sched_entity, we must:
4167 * - Update loads to have both entity and cfs_rq synced with now.
4168 * - Add its load to cfs_rq->runnable_avg
4169 * - For group_entity, update its weight to reflect the new share of
4171 * - Add its new weight to cfs_rq->load.weight
4173 update_load_avg(cfs_rq, se, UPDATE_TG | DO_ATTACH);
4174 update_cfs_group(se);
4175 enqueue_runnable_load_avg(cfs_rq, se);
4176 account_entity_enqueue(cfs_rq, se);
4178 if (flags & ENQUEUE_WAKEUP)
4179 place_entity(cfs_rq, se, 0);
4181 check_schedstat_required();
4182 update_stats_enqueue(cfs_rq, se, flags);
4183 check_spread(cfs_rq, se);
4185 __enqueue_entity(cfs_rq, se);
4188 if (cfs_rq->nr_running == 1) {
4189 list_add_leaf_cfs_rq(cfs_rq);
4190 check_enqueue_throttle(cfs_rq);
4194 static void __clear_buddies_last(struct sched_entity *se)
4196 for_each_sched_entity(se) {
4197 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4198 if (cfs_rq->last != se)
4201 cfs_rq->last = NULL;
4205 static void __clear_buddies_next(struct sched_entity *se)
4207 for_each_sched_entity(se) {
4208 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4209 if (cfs_rq->next != se)
4212 cfs_rq->next = NULL;
4216 static void __clear_buddies_skip(struct sched_entity *se)
4218 for_each_sched_entity(se) {
4219 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4220 if (cfs_rq->skip != se)
4223 cfs_rq->skip = NULL;
4227 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
4229 if (cfs_rq->last == se)
4230 __clear_buddies_last(se);
4232 if (cfs_rq->next == se)
4233 __clear_buddies_next(se);
4235 if (cfs_rq->skip == se)
4236 __clear_buddies_skip(se);
4239 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
4242 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
4245 * Update run-time statistics of the 'current'.
4247 update_curr(cfs_rq);
4250 * When dequeuing a sched_entity, we must:
4251 * - Update loads to have both entity and cfs_rq synced with now.
4252 * - Substract its load from the cfs_rq->runnable_avg.
4253 * - Substract its previous weight from cfs_rq->load.weight.
4254 * - For group entity, update its weight to reflect the new share
4255 * of its group cfs_rq.
4257 update_load_avg(cfs_rq, se, UPDATE_TG);
4258 dequeue_runnable_load_avg(cfs_rq, se);
4260 update_stats_dequeue(cfs_rq, se, flags);
4262 clear_buddies(cfs_rq, se);
4264 if (se != cfs_rq->curr)
4265 __dequeue_entity(cfs_rq, se);
4267 account_entity_dequeue(cfs_rq, se);
4270 * Normalize after update_curr(); which will also have moved
4271 * min_vruntime if @se is the one holding it back. But before doing
4272 * update_min_vruntime() again, which will discount @se's position and
4273 * can move min_vruntime forward still more.
4275 if (!(flags & DEQUEUE_SLEEP))
4276 se->vruntime -= cfs_rq->min_vruntime;
4278 /* return excess runtime on last dequeue */
4279 return_cfs_rq_runtime(cfs_rq);
4281 update_cfs_group(se);
4284 * Now advance min_vruntime if @se was the entity holding it back,
4285 * except when: DEQUEUE_SAVE && !DEQUEUE_MOVE, in this case we'll be
4286 * put back on, and if we advance min_vruntime, we'll be placed back
4287 * further than we started -- ie. we'll be penalized.
4289 if ((flags & (DEQUEUE_SAVE | DEQUEUE_MOVE)) == DEQUEUE_SAVE)
4290 update_min_vruntime(cfs_rq);
4294 * Preempt the current task with a newly woken task if needed:
4297 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
4299 unsigned long ideal_runtime, delta_exec;
4300 struct sched_entity *se;
4303 ideal_runtime = sched_slice(cfs_rq, curr);
4304 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
4305 if (delta_exec > ideal_runtime) {
4306 resched_curr(rq_of(cfs_rq));
4308 * The current task ran long enough, ensure it doesn't get
4309 * re-elected due to buddy favours.
4311 clear_buddies(cfs_rq, curr);
4316 * Ensure that a task that missed wakeup preemption by a
4317 * narrow margin doesn't have to wait for a full slice.
4318 * This also mitigates buddy induced latencies under load.
4320 if (delta_exec < sysctl_sched_min_granularity)
4323 se = __pick_first_entity(cfs_rq);
4324 delta = curr->vruntime - se->vruntime;
4329 if (delta > ideal_runtime)
4330 resched_curr(rq_of(cfs_rq));
4334 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
4336 /* 'current' is not kept within the tree. */
4339 * Any task has to be enqueued before it get to execute on
4340 * a CPU. So account for the time it spent waiting on the
4343 update_stats_wait_end(cfs_rq, se);
4344 __dequeue_entity(cfs_rq, se);
4345 update_load_avg(cfs_rq, se, UPDATE_TG);
4348 update_stats_curr_start(cfs_rq, se);
4352 * Track our maximum slice length, if the CPU's load is at
4353 * least twice that of our own weight (i.e. dont track it
4354 * when there are only lesser-weight tasks around):
4356 if (schedstat_enabled() && rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
4357 schedstat_set(se->statistics.slice_max,
4358 max((u64)schedstat_val(se->statistics.slice_max),
4359 se->sum_exec_runtime - se->prev_sum_exec_runtime));
4362 se->prev_sum_exec_runtime = se->sum_exec_runtime;
4366 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
4369 * Pick the next process, keeping these things in mind, in this order:
4370 * 1) keep things fair between processes/task groups
4371 * 2) pick the "next" process, since someone really wants that to run
4372 * 3) pick the "last" process, for cache locality
4373 * 4) do not run the "skip" process, if something else is available
4375 static struct sched_entity *
4376 pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr)
4378 struct sched_entity *left = __pick_first_entity(cfs_rq);
4379 struct sched_entity *se;
4382 * If curr is set we have to see if its left of the leftmost entity
4383 * still in the tree, provided there was anything in the tree at all.
4385 if (!left || (curr && entity_before(curr, left)))
4388 se = left; /* ideally we run the leftmost entity */
4391 * Avoid running the skip buddy, if running something else can
4392 * be done without getting too unfair.
4394 if (cfs_rq->skip == se) {
4395 struct sched_entity *second;
4398 second = __pick_first_entity(cfs_rq);
4400 second = __pick_next_entity(se);
4401 if (!second || (curr && entity_before(curr, second)))
4405 if (second && wakeup_preempt_entity(second, left) < 1)
4410 * Prefer last buddy, try to return the CPU to a preempted task.
4412 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
4416 * Someone really wants this to run. If it's not unfair, run it.
4418 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
4421 clear_buddies(cfs_rq, se);
4426 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
4428 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
4431 * If still on the runqueue then deactivate_task()
4432 * was not called and update_curr() has to be done:
4435 update_curr(cfs_rq);
4437 /* throttle cfs_rqs exceeding runtime */
4438 check_cfs_rq_runtime(cfs_rq);
4440 check_spread(cfs_rq, prev);
4443 update_stats_wait_start(cfs_rq, prev);
4444 /* Put 'current' back into the tree. */
4445 __enqueue_entity(cfs_rq, prev);
4446 /* in !on_rq case, update occurred at dequeue */
4447 update_load_avg(cfs_rq, prev, 0);
4449 cfs_rq->curr = NULL;
4453 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
4456 * Update run-time statistics of the 'current'.
4458 update_curr(cfs_rq);
4461 * Ensure that runnable average is periodically updated.
4463 update_load_avg(cfs_rq, curr, UPDATE_TG);
4464 update_cfs_group(curr);
4466 #ifdef CONFIG_SCHED_HRTICK
4468 * queued ticks are scheduled to match the slice, so don't bother
4469 * validating it and just reschedule.
4472 resched_curr(rq_of(cfs_rq));
4476 * don't let the period tick interfere with the hrtick preemption
4478 if (!sched_feat(DOUBLE_TICK) &&
4479 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
4483 if (cfs_rq->nr_running > 1)
4484 check_preempt_tick(cfs_rq, curr);
4488 /**************************************************
4489 * CFS bandwidth control machinery
4492 #ifdef CONFIG_CFS_BANDWIDTH
4494 #ifdef HAVE_JUMP_LABEL
4495 static struct static_key __cfs_bandwidth_used;
4497 static inline bool cfs_bandwidth_used(void)
4499 return static_key_false(&__cfs_bandwidth_used);
4502 void cfs_bandwidth_usage_inc(void)
4504 static_key_slow_inc_cpuslocked(&__cfs_bandwidth_used);
4507 void cfs_bandwidth_usage_dec(void)
4509 static_key_slow_dec_cpuslocked(&__cfs_bandwidth_used);
4511 #else /* HAVE_JUMP_LABEL */
4512 static bool cfs_bandwidth_used(void)
4517 void cfs_bandwidth_usage_inc(void) {}
4518 void cfs_bandwidth_usage_dec(void) {}
4519 #endif /* HAVE_JUMP_LABEL */
4522 * default period for cfs group bandwidth.
4523 * default: 0.1s, units: nanoseconds
4525 static inline u64 default_cfs_period(void)
4527 return 100000000ULL;
4530 static inline u64 sched_cfs_bandwidth_slice(void)
4532 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
4536 * Replenish runtime according to assigned quota and update expiration time.
4537 * We use sched_clock_cpu directly instead of rq->clock to avoid adding
4538 * additional synchronization around rq->lock.
4540 * requires cfs_b->lock
4542 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
4546 if (cfs_b->quota == RUNTIME_INF)
4549 now = sched_clock_cpu(smp_processor_id());
4550 cfs_b->runtime = cfs_b->quota;
4551 cfs_b->runtime_expires = now + ktime_to_ns(cfs_b->period);
4554 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
4556 return &tg->cfs_bandwidth;
4559 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
4560 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
4562 if (unlikely(cfs_rq->throttle_count))
4563 return cfs_rq->throttled_clock_task - cfs_rq->throttled_clock_task_time;
4565 return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time;
4568 /* returns 0 on failure to allocate runtime */
4569 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4571 struct task_group *tg = cfs_rq->tg;
4572 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
4573 u64 amount = 0, min_amount, expires;
4575 /* note: this is a positive sum as runtime_remaining <= 0 */
4576 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
4578 raw_spin_lock(&cfs_b->lock);
4579 if (cfs_b->quota == RUNTIME_INF)
4580 amount = min_amount;
4582 start_cfs_bandwidth(cfs_b);
4584 if (cfs_b->runtime > 0) {
4585 amount = min(cfs_b->runtime, min_amount);
4586 cfs_b->runtime -= amount;
4590 expires = cfs_b->runtime_expires;
4591 raw_spin_unlock(&cfs_b->lock);
4593 cfs_rq->runtime_remaining += amount;
4595 * we may have advanced our local expiration to account for allowed
4596 * spread between our sched_clock and the one on which runtime was
4599 if ((s64)(expires - cfs_rq->runtime_expires) > 0)
4600 cfs_rq->runtime_expires = expires;
4602 return cfs_rq->runtime_remaining > 0;
4606 * Note: This depends on the synchronization provided by sched_clock and the
4607 * fact that rq->clock snapshots this value.
4609 static void expire_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4611 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4613 /* if the deadline is ahead of our clock, nothing to do */
4614 if (likely((s64)(rq_clock(rq_of(cfs_rq)) - cfs_rq->runtime_expires) < 0))
4617 if (cfs_rq->runtime_remaining < 0)
4621 * If the local deadline has passed we have to consider the
4622 * possibility that our sched_clock is 'fast' and the global deadline
4623 * has not truly expired.
4625 * Fortunately we can check determine whether this the case by checking
4626 * whether the global deadline has advanced. It is valid to compare
4627 * cfs_b->runtime_expires without any locks since we only care about
4628 * exact equality, so a partial write will still work.
4631 if (cfs_rq->runtime_expires != cfs_b->runtime_expires) {
4632 /* extend local deadline, drift is bounded above by 2 ticks */
4633 cfs_rq->runtime_expires += TICK_NSEC;
4635 /* global deadline is ahead, expiration has passed */
4636 cfs_rq->runtime_remaining = 0;
4640 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
4642 /* dock delta_exec before expiring quota (as it could span periods) */
4643 cfs_rq->runtime_remaining -= delta_exec;
4644 expire_cfs_rq_runtime(cfs_rq);
4646 if (likely(cfs_rq->runtime_remaining > 0))
4650 * if we're unable to extend our runtime we resched so that the active
4651 * hierarchy can be throttled
4653 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
4654 resched_curr(rq_of(cfs_rq));
4657 static __always_inline
4658 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
4660 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
4663 __account_cfs_rq_runtime(cfs_rq, delta_exec);
4666 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
4668 return cfs_bandwidth_used() && cfs_rq->throttled;
4671 /* check whether cfs_rq, or any parent, is throttled */
4672 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
4674 return cfs_bandwidth_used() && cfs_rq->throttle_count;
4678 * Ensure that neither of the group entities corresponding to src_cpu or
4679 * dest_cpu are members of a throttled hierarchy when performing group
4680 * load-balance operations.
4682 static inline int throttled_lb_pair(struct task_group *tg,
4683 int src_cpu, int dest_cpu)
4685 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
4687 src_cfs_rq = tg->cfs_rq[src_cpu];
4688 dest_cfs_rq = tg->cfs_rq[dest_cpu];
4690 return throttled_hierarchy(src_cfs_rq) ||
4691 throttled_hierarchy(dest_cfs_rq);
4694 /* updated child weight may affect parent so we have to do this bottom up */
4695 static int tg_unthrottle_up(struct task_group *tg, void *data)
4697 struct rq *rq = data;
4698 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
4700 cfs_rq->throttle_count--;
4701 if (!cfs_rq->throttle_count) {
4702 /* adjust cfs_rq_clock_task() */
4703 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
4704 cfs_rq->throttled_clock_task;
4710 static int tg_throttle_down(struct task_group *tg, void *data)
4712 struct rq *rq = data;
4713 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
4715 /* group is entering throttled state, stop time */
4716 if (!cfs_rq->throttle_count)
4717 cfs_rq->throttled_clock_task = rq_clock_task(rq);
4718 cfs_rq->throttle_count++;
4723 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
4725 struct rq *rq = rq_of(cfs_rq);
4726 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4727 struct sched_entity *se;
4728 long task_delta, dequeue = 1;
4731 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
4733 /* freeze hierarchy runnable averages while throttled */
4735 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
4738 task_delta = cfs_rq->h_nr_running;
4739 for_each_sched_entity(se) {
4740 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
4741 /* throttled entity or throttle-on-deactivate */
4746 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
4747 qcfs_rq->h_nr_running -= task_delta;
4749 if (qcfs_rq->load.weight)
4754 sub_nr_running(rq, task_delta);
4756 cfs_rq->throttled = 1;
4757 cfs_rq->throttled_clock = rq_clock(rq);
4758 raw_spin_lock(&cfs_b->lock);
4759 empty = list_empty(&cfs_b->throttled_cfs_rq);
4762 * Add to the _head_ of the list, so that an already-started
4763 * distribute_cfs_runtime will not see us
4765 list_add_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
4768 * If we're the first throttled task, make sure the bandwidth
4772 start_cfs_bandwidth(cfs_b);
4774 raw_spin_unlock(&cfs_b->lock);
4777 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
4779 struct rq *rq = rq_of(cfs_rq);
4780 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4781 struct sched_entity *se;
4785 se = cfs_rq->tg->se[cpu_of(rq)];
4787 cfs_rq->throttled = 0;
4789 update_rq_clock(rq);
4791 raw_spin_lock(&cfs_b->lock);
4792 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
4793 list_del_rcu(&cfs_rq->throttled_list);
4794 raw_spin_unlock(&cfs_b->lock);
4796 /* update hierarchical throttle state */
4797 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
4799 if (!cfs_rq->load.weight)
4802 task_delta = cfs_rq->h_nr_running;
4803 for_each_sched_entity(se) {
4807 cfs_rq = cfs_rq_of(se);
4809 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
4810 cfs_rq->h_nr_running += task_delta;
4812 if (cfs_rq_throttled(cfs_rq))
4817 add_nr_running(rq, task_delta);
4819 /* Determine whether we need to wake up potentially idle CPU: */
4820 if (rq->curr == rq->idle && rq->cfs.nr_running)
4824 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b,
4825 u64 remaining, u64 expires)
4827 struct cfs_rq *cfs_rq;
4829 u64 starting_runtime = remaining;
4832 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
4834 struct rq *rq = rq_of(cfs_rq);
4838 if (!cfs_rq_throttled(cfs_rq))
4841 runtime = -cfs_rq->runtime_remaining + 1;
4842 if (runtime > remaining)
4843 runtime = remaining;
4844 remaining -= runtime;
4846 cfs_rq->runtime_remaining += runtime;
4847 cfs_rq->runtime_expires = expires;
4849 /* we check whether we're throttled above */
4850 if (cfs_rq->runtime_remaining > 0)
4851 unthrottle_cfs_rq(cfs_rq);
4861 return starting_runtime - remaining;
4865 * Responsible for refilling a task_group's bandwidth and unthrottling its
4866 * cfs_rqs as appropriate. If there has been no activity within the last
4867 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
4868 * used to track this state.
4870 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
4872 u64 runtime, runtime_expires;
4875 /* no need to continue the timer with no bandwidth constraint */
4876 if (cfs_b->quota == RUNTIME_INF)
4877 goto out_deactivate;
4879 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
4880 cfs_b->nr_periods += overrun;
4883 * idle depends on !throttled (for the case of a large deficit), and if
4884 * we're going inactive then everything else can be deferred
4886 if (cfs_b->idle && !throttled)
4887 goto out_deactivate;
4889 __refill_cfs_bandwidth_runtime(cfs_b);
4892 /* mark as potentially idle for the upcoming period */
4897 /* account preceding periods in which throttling occurred */
4898 cfs_b->nr_throttled += overrun;
4900 runtime_expires = cfs_b->runtime_expires;
4903 * This check is repeated as we are holding onto the new bandwidth while
4904 * we unthrottle. This can potentially race with an unthrottled group
4905 * trying to acquire new bandwidth from the global pool. This can result
4906 * in us over-using our runtime if it is all used during this loop, but
4907 * only by limited amounts in that extreme case.
4909 while (throttled && cfs_b->runtime > 0) {
4910 runtime = cfs_b->runtime;
4911 raw_spin_unlock(&cfs_b->lock);
4912 /* we can't nest cfs_b->lock while distributing bandwidth */
4913 runtime = distribute_cfs_runtime(cfs_b, runtime,
4915 raw_spin_lock(&cfs_b->lock);
4917 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
4919 cfs_b->runtime -= min(runtime, cfs_b->runtime);
4923 * While we are ensured activity in the period following an
4924 * unthrottle, this also covers the case in which the new bandwidth is
4925 * insufficient to cover the existing bandwidth deficit. (Forcing the
4926 * timer to remain active while there are any throttled entities.)
4936 /* a cfs_rq won't donate quota below this amount */
4937 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
4938 /* minimum remaining period time to redistribute slack quota */
4939 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
4940 /* how long we wait to gather additional slack before distributing */
4941 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
4944 * Are we near the end of the current quota period?
4946 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
4947 * hrtimer base being cleared by hrtimer_start. In the case of
4948 * migrate_hrtimers, base is never cleared, so we are fine.
4950 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
4952 struct hrtimer *refresh_timer = &cfs_b->period_timer;
4955 /* if the call-back is running a quota refresh is already occurring */
4956 if (hrtimer_callback_running(refresh_timer))
4959 /* is a quota refresh about to occur? */
4960 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
4961 if (remaining < min_expire)
4967 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
4969 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
4971 /* if there's a quota refresh soon don't bother with slack */
4972 if (runtime_refresh_within(cfs_b, min_left))
4975 hrtimer_start(&cfs_b->slack_timer,
4976 ns_to_ktime(cfs_bandwidth_slack_period),
4980 /* we know any runtime found here is valid as update_curr() precedes return */
4981 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4983 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4984 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
4986 if (slack_runtime <= 0)
4989 raw_spin_lock(&cfs_b->lock);
4990 if (cfs_b->quota != RUNTIME_INF &&
4991 cfs_rq->runtime_expires == cfs_b->runtime_expires) {
4992 cfs_b->runtime += slack_runtime;
4994 /* we are under rq->lock, defer unthrottling using a timer */
4995 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
4996 !list_empty(&cfs_b->throttled_cfs_rq))
4997 start_cfs_slack_bandwidth(cfs_b);
4999 raw_spin_unlock(&cfs_b->lock);
5001 /* even if it's not valid for return we don't want to try again */
5002 cfs_rq->runtime_remaining -= slack_runtime;
5005 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
5007 if (!cfs_bandwidth_used())
5010 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
5013 __return_cfs_rq_runtime(cfs_rq);
5017 * This is done with a timer (instead of inline with bandwidth return) since
5018 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
5020 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
5022 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
5025 /* confirm we're still not at a refresh boundary */
5026 raw_spin_lock(&cfs_b->lock);
5027 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
5028 raw_spin_unlock(&cfs_b->lock);
5032 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice)
5033 runtime = cfs_b->runtime;
5035 expires = cfs_b->runtime_expires;
5036 raw_spin_unlock(&cfs_b->lock);
5041 runtime = distribute_cfs_runtime(cfs_b, runtime, expires);
5043 raw_spin_lock(&cfs_b->lock);
5044 if (expires == cfs_b->runtime_expires)
5045 cfs_b->runtime -= min(runtime, cfs_b->runtime);
5046 raw_spin_unlock(&cfs_b->lock);
5050 * When a group wakes up we want to make sure that its quota is not already
5051 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
5052 * runtime as update_curr() throttling can not not trigger until it's on-rq.
5054 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
5056 if (!cfs_bandwidth_used())
5059 /* an active group must be handled by the update_curr()->put() path */
5060 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
5063 /* ensure the group is not already throttled */
5064 if (cfs_rq_throttled(cfs_rq))
5067 /* update runtime allocation */
5068 account_cfs_rq_runtime(cfs_rq, 0);
5069 if (cfs_rq->runtime_remaining <= 0)
5070 throttle_cfs_rq(cfs_rq);
5073 static void sync_throttle(struct task_group *tg, int cpu)
5075 struct cfs_rq *pcfs_rq, *cfs_rq;
5077 if (!cfs_bandwidth_used())
5083 cfs_rq = tg->cfs_rq[cpu];
5084 pcfs_rq = tg->parent->cfs_rq[cpu];
5086 cfs_rq->throttle_count = pcfs_rq->throttle_count;
5087 cfs_rq->throttled_clock_task = rq_clock_task(cpu_rq(cpu));
5090 /* conditionally throttle active cfs_rq's from put_prev_entity() */
5091 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
5093 if (!cfs_bandwidth_used())
5096 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
5100 * it's possible for a throttled entity to be forced into a running
5101 * state (e.g. set_curr_task), in this case we're finished.
5103 if (cfs_rq_throttled(cfs_rq))
5106 throttle_cfs_rq(cfs_rq);
5110 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
5112 struct cfs_bandwidth *cfs_b =
5113 container_of(timer, struct cfs_bandwidth, slack_timer);
5115 do_sched_cfs_slack_timer(cfs_b);
5117 return HRTIMER_NORESTART;
5120 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
5122 struct cfs_bandwidth *cfs_b =
5123 container_of(timer, struct cfs_bandwidth, period_timer);
5127 raw_spin_lock(&cfs_b->lock);
5129 overrun = hrtimer_forward_now(timer, cfs_b->period);
5133 idle = do_sched_cfs_period_timer(cfs_b, overrun);
5136 cfs_b->period_active = 0;
5137 raw_spin_unlock(&cfs_b->lock);
5139 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
5142 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
5144 raw_spin_lock_init(&cfs_b->lock);
5146 cfs_b->quota = RUNTIME_INF;
5147 cfs_b->period = ns_to_ktime(default_cfs_period());
5149 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
5150 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED);
5151 cfs_b->period_timer.function = sched_cfs_period_timer;
5152 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
5153 cfs_b->slack_timer.function = sched_cfs_slack_timer;
5156 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
5158 cfs_rq->runtime_enabled = 0;
5159 INIT_LIST_HEAD(&cfs_rq->throttled_list);
5162 void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
5164 lockdep_assert_held(&cfs_b->lock);
5166 if (!cfs_b->period_active) {
5167 cfs_b->period_active = 1;
5168 hrtimer_forward_now(&cfs_b->period_timer, cfs_b->period);
5169 hrtimer_start_expires(&cfs_b->period_timer, HRTIMER_MODE_ABS_PINNED);
5173 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
5175 /* init_cfs_bandwidth() was not called */
5176 if (!cfs_b->throttled_cfs_rq.next)
5179 hrtimer_cancel(&cfs_b->period_timer);
5180 hrtimer_cancel(&cfs_b->slack_timer);
5184 * Both these CPU hotplug callbacks race against unregister_fair_sched_group()
5186 * The race is harmless, since modifying bandwidth settings of unhooked group
5187 * bits doesn't do much.
5190 /* cpu online calback */
5191 static void __maybe_unused update_runtime_enabled(struct rq *rq)
5193 struct task_group *tg;
5195 lockdep_assert_held(&rq->lock);
5198 list_for_each_entry_rcu(tg, &task_groups, list) {
5199 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
5200 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
5202 raw_spin_lock(&cfs_b->lock);
5203 cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF;
5204 raw_spin_unlock(&cfs_b->lock);
5209 /* cpu offline callback */
5210 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
5212 struct task_group *tg;
5214 lockdep_assert_held(&rq->lock);
5217 list_for_each_entry_rcu(tg, &task_groups, list) {
5218 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
5220 if (!cfs_rq->runtime_enabled)
5224 * clock_task is not advancing so we just need to make sure
5225 * there's some valid quota amount
5227 cfs_rq->runtime_remaining = 1;
5229 * Offline rq is schedulable till CPU is completely disabled
5230 * in take_cpu_down(), so we prevent new cfs throttling here.
5232 cfs_rq->runtime_enabled = 0;
5234 if (cfs_rq_throttled(cfs_rq))
5235 unthrottle_cfs_rq(cfs_rq);
5240 #else /* CONFIG_CFS_BANDWIDTH */
5241 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
5243 return rq_clock_task(rq_of(cfs_rq));
5246 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {}
5247 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; }
5248 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
5249 static inline void sync_throttle(struct task_group *tg, int cpu) {}
5250 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
5252 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
5257 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
5262 static inline int throttled_lb_pair(struct task_group *tg,
5263 int src_cpu, int dest_cpu)
5268 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
5270 #ifdef CONFIG_FAIR_GROUP_SCHED
5271 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
5274 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
5278 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
5279 static inline void update_runtime_enabled(struct rq *rq) {}
5280 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
5282 #endif /* CONFIG_CFS_BANDWIDTH */
5284 /**************************************************
5285 * CFS operations on tasks:
5288 #ifdef CONFIG_SCHED_HRTICK
5289 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
5291 struct sched_entity *se = &p->se;
5292 struct cfs_rq *cfs_rq = cfs_rq_of(se);
5294 SCHED_WARN_ON(task_rq(p) != rq);
5296 if (rq->cfs.h_nr_running > 1) {
5297 u64 slice = sched_slice(cfs_rq, se);
5298 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
5299 s64 delta = slice - ran;
5306 hrtick_start(rq, delta);
5311 * called from enqueue/dequeue and updates the hrtick when the
5312 * current task is from our class and nr_running is low enough
5315 static void hrtick_update(struct rq *rq)
5317 struct task_struct *curr = rq->curr;
5319 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
5322 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
5323 hrtick_start_fair(rq, curr);
5325 #else /* !CONFIG_SCHED_HRTICK */
5327 hrtick_start_fair(struct rq *rq, struct task_struct *p)
5331 static inline void hrtick_update(struct rq *rq)
5337 * The enqueue_task method is called before nr_running is
5338 * increased. Here we update the fair scheduling stats and
5339 * then put the task into the rbtree:
5342 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
5344 struct cfs_rq *cfs_rq;
5345 struct sched_entity *se = &p->se;
5348 * If in_iowait is set, the code below may not trigger any cpufreq
5349 * utilization updates, so do it here explicitly with the IOWAIT flag
5353 cpufreq_update_util(rq, SCHED_CPUFREQ_IOWAIT);
5355 for_each_sched_entity(se) {
5358 cfs_rq = cfs_rq_of(se);
5359 enqueue_entity(cfs_rq, se, flags);
5362 * end evaluation on encountering a throttled cfs_rq
5364 * note: in the case of encountering a throttled cfs_rq we will
5365 * post the final h_nr_running increment below.
5367 if (cfs_rq_throttled(cfs_rq))
5369 cfs_rq->h_nr_running++;
5371 flags = ENQUEUE_WAKEUP;
5374 for_each_sched_entity(se) {
5375 cfs_rq = cfs_rq_of(se);
5376 cfs_rq->h_nr_running++;
5378 if (cfs_rq_throttled(cfs_rq))
5381 update_load_avg(cfs_rq, se, UPDATE_TG);
5382 update_cfs_group(se);
5386 add_nr_running(rq, 1);
5388 util_est_enqueue(&rq->cfs, p);
5392 static void set_next_buddy(struct sched_entity *se);
5395 * The dequeue_task method is called before nr_running is
5396 * decreased. We remove the task from the rbtree and
5397 * update the fair scheduling stats:
5399 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
5401 struct cfs_rq *cfs_rq;
5402 struct sched_entity *se = &p->se;
5403 int task_sleep = flags & DEQUEUE_SLEEP;
5405 for_each_sched_entity(se) {
5406 cfs_rq = cfs_rq_of(se);
5407 dequeue_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 decrement below.
5415 if (cfs_rq_throttled(cfs_rq))
5417 cfs_rq->h_nr_running--;
5419 /* Don't dequeue parent if it has other entities besides us */
5420 if (cfs_rq->load.weight) {
5421 /* Avoid re-evaluating load for this entity: */
5422 se = parent_entity(se);
5424 * Bias pick_next to pick a task from this cfs_rq, as
5425 * p is sleeping when it is within its sched_slice.
5427 if (task_sleep && se && !throttled_hierarchy(cfs_rq))
5431 flags |= DEQUEUE_SLEEP;
5434 for_each_sched_entity(se) {
5435 cfs_rq = cfs_rq_of(se);
5436 cfs_rq->h_nr_running--;
5438 if (cfs_rq_throttled(cfs_rq))
5441 update_load_avg(cfs_rq, se, UPDATE_TG);
5442 update_cfs_group(se);
5446 sub_nr_running(rq, 1);
5448 util_est_dequeue(&rq->cfs, p, task_sleep);
5454 /* Working cpumask for: load_balance, load_balance_newidle. */
5455 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
5456 DEFINE_PER_CPU(cpumask_var_t, select_idle_mask);
5458 #ifdef CONFIG_NO_HZ_COMMON
5460 * per rq 'load' arrray crap; XXX kill this.
5464 * The exact cpuload calculated at every tick would be:
5466 * load' = (1 - 1/2^i) * load + (1/2^i) * cur_load
5468 * If a CPU misses updates for n ticks (as it was idle) and update gets
5469 * called on the n+1-th tick when CPU may be busy, then we have:
5471 * load_n = (1 - 1/2^i)^n * load_0
5472 * load_n+1 = (1 - 1/2^i) * load_n + (1/2^i) * cur_load
5474 * decay_load_missed() below does efficient calculation of
5476 * load' = (1 - 1/2^i)^n * load
5478 * Because x^(n+m) := x^n * x^m we can decompose any x^n in power-of-2 factors.
5479 * This allows us to precompute the above in said factors, thereby allowing the
5480 * reduction of an arbitrary n in O(log_2 n) steps. (See also
5481 * fixed_power_int())
5483 * The calculation is approximated on a 128 point scale.
5485 #define DEGRADE_SHIFT 7
5487 static const u8 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
5488 static const u8 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
5489 { 0, 0, 0, 0, 0, 0, 0, 0 },
5490 { 64, 32, 8, 0, 0, 0, 0, 0 },
5491 { 96, 72, 40, 12, 1, 0, 0, 0 },
5492 { 112, 98, 75, 43, 15, 1, 0, 0 },
5493 { 120, 112, 98, 76, 45, 16, 2, 0 }
5497 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
5498 * would be when CPU is idle and so we just decay the old load without
5499 * adding any new load.
5501 static unsigned long
5502 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
5506 if (!missed_updates)
5509 if (missed_updates >= degrade_zero_ticks[idx])
5513 return load >> missed_updates;
5515 while (missed_updates) {
5516 if (missed_updates % 2)
5517 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
5519 missed_updates >>= 1;
5526 cpumask_var_t idle_cpus_mask;
5528 int has_blocked; /* Idle CPUS has blocked load */
5529 unsigned long next_balance; /* in jiffy units */
5530 unsigned long next_blocked; /* Next update of blocked load in jiffies */
5531 } nohz ____cacheline_aligned;
5533 #endif /* CONFIG_NO_HZ_COMMON */
5536 * __cpu_load_update - update the rq->cpu_load[] statistics
5537 * @this_rq: The rq to update statistics for
5538 * @this_load: The current load
5539 * @pending_updates: The number of missed updates
5541 * Update rq->cpu_load[] statistics. This function is usually called every
5542 * scheduler tick (TICK_NSEC).
5544 * This function computes a decaying average:
5546 * load[i]' = (1 - 1/2^i) * load[i] + (1/2^i) * load
5548 * Because of NOHZ it might not get called on every tick which gives need for
5549 * the @pending_updates argument.
5551 * load[i]_n = (1 - 1/2^i) * load[i]_n-1 + (1/2^i) * load_n-1
5552 * = A * load[i]_n-1 + B ; A := (1 - 1/2^i), B := (1/2^i) * load
5553 * = A * (A * load[i]_n-2 + B) + B
5554 * = A * (A * (A * load[i]_n-3 + B) + B) + B
5555 * = A^3 * load[i]_n-3 + (A^2 + A + 1) * B
5556 * = A^n * load[i]_0 + (A^(n-1) + A^(n-2) + ... + 1) * B
5557 * = A^n * load[i]_0 + ((1 - A^n) / (1 - A)) * B
5558 * = (1 - 1/2^i)^n * (load[i]_0 - load) + load
5560 * In the above we've assumed load_n := load, which is true for NOHZ_FULL as
5561 * any change in load would have resulted in the tick being turned back on.
5563 * For regular NOHZ, this reduces to:
5565 * load[i]_n = (1 - 1/2^i)^n * load[i]_0
5567 * see decay_load_misses(). For NOHZ_FULL we get to subtract and add the extra
5570 static void cpu_load_update(struct rq *this_rq, unsigned long this_load,
5571 unsigned long pending_updates)
5573 unsigned long __maybe_unused tickless_load = this_rq->cpu_load[0];
5576 this_rq->nr_load_updates++;
5578 /* Update our load: */
5579 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
5580 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
5581 unsigned long old_load, new_load;
5583 /* scale is effectively 1 << i now, and >> i divides by scale */
5585 old_load = this_rq->cpu_load[i];
5586 #ifdef CONFIG_NO_HZ_COMMON
5587 old_load = decay_load_missed(old_load, pending_updates - 1, i);
5588 if (tickless_load) {
5589 old_load -= decay_load_missed(tickless_load, pending_updates - 1, i);
5591 * old_load can never be a negative value because a
5592 * decayed tickless_load cannot be greater than the
5593 * original tickless_load.
5595 old_load += tickless_load;
5598 new_load = this_load;
5600 * Round up the averaging division if load is increasing. This
5601 * prevents us from getting stuck on 9 if the load is 10, for
5604 if (new_load > old_load)
5605 new_load += scale - 1;
5607 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
5610 sched_avg_update(this_rq);
5613 /* Used instead of source_load when we know the type == 0 */
5614 static unsigned long weighted_cpuload(struct rq *rq)
5616 return cfs_rq_runnable_load_avg(&rq->cfs);
5619 #ifdef CONFIG_NO_HZ_COMMON
5621 * There is no sane way to deal with nohz on smp when using jiffies because the
5622 * CPU doing the jiffies update might drift wrt the CPU doing the jiffy reading
5623 * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}.
5625 * Therefore we need to avoid the delta approach from the regular tick when
5626 * possible since that would seriously skew the load calculation. This is why we
5627 * use cpu_load_update_periodic() for CPUs out of nohz. However we'll rely on
5628 * jiffies deltas for updates happening while in nohz mode (idle ticks, idle
5629 * loop exit, nohz_idle_balance, nohz full exit...)
5631 * This means we might still be one tick off for nohz periods.
5634 static void cpu_load_update_nohz(struct rq *this_rq,
5635 unsigned long curr_jiffies,
5638 unsigned long pending_updates;
5640 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
5641 if (pending_updates) {
5642 this_rq->last_load_update_tick = curr_jiffies;
5644 * In the regular NOHZ case, we were idle, this means load 0.
5645 * In the NOHZ_FULL case, we were non-idle, we should consider
5646 * its weighted load.
5648 cpu_load_update(this_rq, load, pending_updates);
5653 * Called from nohz_idle_balance() to update the load ratings before doing the
5656 static void cpu_load_update_idle(struct rq *this_rq)
5659 * bail if there's load or we're actually up-to-date.
5661 if (weighted_cpuload(this_rq))
5664 cpu_load_update_nohz(this_rq, READ_ONCE(jiffies), 0);
5668 * Record CPU load on nohz entry so we know the tickless load to account
5669 * on nohz exit. cpu_load[0] happens then to be updated more frequently
5670 * than other cpu_load[idx] but it should be fine as cpu_load readers
5671 * shouldn't rely into synchronized cpu_load[*] updates.
5673 void cpu_load_update_nohz_start(void)
5675 struct rq *this_rq = this_rq();
5678 * This is all lockless but should be fine. If weighted_cpuload changes
5679 * concurrently we'll exit nohz. And cpu_load write can race with
5680 * cpu_load_update_idle() but both updater would be writing the same.
5682 this_rq->cpu_load[0] = weighted_cpuload(this_rq);
5686 * Account the tickless load in the end of a nohz frame.
5688 void cpu_load_update_nohz_stop(void)
5690 unsigned long curr_jiffies = READ_ONCE(jiffies);
5691 struct rq *this_rq = this_rq();
5695 if (curr_jiffies == this_rq->last_load_update_tick)
5698 load = weighted_cpuload(this_rq);
5699 rq_lock(this_rq, &rf);
5700 update_rq_clock(this_rq);
5701 cpu_load_update_nohz(this_rq, curr_jiffies, load);
5702 rq_unlock(this_rq, &rf);
5704 #else /* !CONFIG_NO_HZ_COMMON */
5705 static inline void cpu_load_update_nohz(struct rq *this_rq,
5706 unsigned long curr_jiffies,
5707 unsigned long load) { }
5708 #endif /* CONFIG_NO_HZ_COMMON */
5710 static void cpu_load_update_periodic(struct rq *this_rq, unsigned long load)
5712 #ifdef CONFIG_NO_HZ_COMMON
5713 /* See the mess around cpu_load_update_nohz(). */
5714 this_rq->last_load_update_tick = READ_ONCE(jiffies);
5716 cpu_load_update(this_rq, load, 1);
5720 * Called from scheduler_tick()
5722 void cpu_load_update_active(struct rq *this_rq)
5724 unsigned long load = weighted_cpuload(this_rq);
5726 if (tick_nohz_tick_stopped())
5727 cpu_load_update_nohz(this_rq, READ_ONCE(jiffies), load);
5729 cpu_load_update_periodic(this_rq, load);
5733 * Return a low guess at the load of a migration-source CPU weighted
5734 * according to the scheduling class and "nice" value.
5736 * We want to under-estimate the load of migration sources, to
5737 * balance conservatively.
5739 static unsigned long source_load(int cpu, int type)
5741 struct rq *rq = cpu_rq(cpu);
5742 unsigned long total = weighted_cpuload(rq);
5744 if (type == 0 || !sched_feat(LB_BIAS))
5747 return min(rq->cpu_load[type-1], total);
5751 * Return a high guess at the load of a migration-target CPU weighted
5752 * according to the scheduling class and "nice" value.
5754 static unsigned long target_load(int cpu, int type)
5756 struct rq *rq = cpu_rq(cpu);
5757 unsigned long total = weighted_cpuload(rq);
5759 if (type == 0 || !sched_feat(LB_BIAS))
5762 return max(rq->cpu_load[type-1], total);
5765 static unsigned long capacity_of(int cpu)
5767 return cpu_rq(cpu)->cpu_capacity;
5770 static unsigned long capacity_orig_of(int cpu)
5772 return cpu_rq(cpu)->cpu_capacity_orig;
5775 static unsigned long cpu_avg_load_per_task(int cpu)
5777 struct rq *rq = cpu_rq(cpu);
5778 unsigned long nr_running = READ_ONCE(rq->cfs.h_nr_running);
5779 unsigned long load_avg = weighted_cpuload(rq);
5782 return load_avg / nr_running;
5787 static void record_wakee(struct task_struct *p)
5790 * Only decay a single time; tasks that have less then 1 wakeup per
5791 * jiffy will not have built up many flips.
5793 if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) {
5794 current->wakee_flips >>= 1;
5795 current->wakee_flip_decay_ts = jiffies;
5798 if (current->last_wakee != p) {
5799 current->last_wakee = p;
5800 current->wakee_flips++;
5805 * Detect M:N waker/wakee relationships via a switching-frequency heuristic.
5807 * A waker of many should wake a different task than the one last awakened
5808 * at a frequency roughly N times higher than one of its wakees.
5810 * In order to determine whether we should let the load spread vs consolidating
5811 * to shared cache, we look for a minimum 'flip' frequency of llc_size in one
5812 * partner, and a factor of lls_size higher frequency in the other.
5814 * With both conditions met, we can be relatively sure that the relationship is
5815 * non-monogamous, with partner count exceeding socket size.
5817 * Waker/wakee being client/server, worker/dispatcher, interrupt source or
5818 * whatever is irrelevant, spread criteria is apparent partner count exceeds
5821 static int wake_wide(struct task_struct *p)
5823 unsigned int master = current->wakee_flips;
5824 unsigned int slave = p->wakee_flips;
5825 int factor = this_cpu_read(sd_llc_size);
5828 swap(master, slave);
5829 if (slave < factor || master < slave * factor)
5835 * The purpose of wake_affine() is to quickly determine on which CPU we can run
5836 * soonest. For the purpose of speed we only consider the waking and previous
5839 * wake_affine_idle() - only considers 'now', it check if the waking CPU is
5840 * cache-affine and is (or will be) idle.
5842 * wake_affine_weight() - considers the weight to reflect the average
5843 * scheduling latency of the CPUs. This seems to work
5844 * for the overloaded case.
5847 wake_affine_idle(int this_cpu, int prev_cpu, int sync)
5850 * If this_cpu is idle, it implies the wakeup is from interrupt
5851 * context. Only allow the move if cache is shared. Otherwise an
5852 * interrupt intensive workload could force all tasks onto one
5853 * node depending on the IO topology or IRQ affinity settings.
5855 * If the prev_cpu is idle and cache affine then avoid a migration.
5856 * There is no guarantee that the cache hot data from an interrupt
5857 * is more important than cache hot data on the prev_cpu and from
5858 * a cpufreq perspective, it's better to have higher utilisation
5861 if (idle_cpu(this_cpu) && cpus_share_cache(this_cpu, prev_cpu))
5862 return idle_cpu(prev_cpu) ? prev_cpu : this_cpu;
5864 if (sync && cpu_rq(this_cpu)->nr_running == 1)
5867 return nr_cpumask_bits;
5871 wake_affine_weight(struct sched_domain *sd, struct task_struct *p,
5872 int this_cpu, int prev_cpu, int sync)
5874 s64 this_eff_load, prev_eff_load;
5875 unsigned long task_load;
5877 this_eff_load = target_load(this_cpu, sd->wake_idx);
5880 unsigned long current_load = task_h_load(current);
5882 if (current_load > this_eff_load)
5885 this_eff_load -= current_load;
5888 task_load = task_h_load(p);
5890 this_eff_load += task_load;
5891 if (sched_feat(WA_BIAS))
5892 this_eff_load *= 100;
5893 this_eff_load *= capacity_of(prev_cpu);
5895 prev_eff_load = source_load(prev_cpu, sd->wake_idx);
5896 prev_eff_load -= task_load;
5897 if (sched_feat(WA_BIAS))
5898 prev_eff_load *= 100 + (sd->imbalance_pct - 100) / 2;
5899 prev_eff_load *= capacity_of(this_cpu);
5902 * If sync, adjust the weight of prev_eff_load such that if
5903 * prev_eff == this_eff that select_idle_sibling() will consider
5904 * stacking the wakee on top of the waker if no other CPU is
5910 return this_eff_load < prev_eff_load ? this_cpu : nr_cpumask_bits;
5913 static int wake_affine(struct sched_domain *sd, struct task_struct *p,
5914 int this_cpu, int prev_cpu, int sync)
5916 int target = nr_cpumask_bits;
5918 if (sched_feat(WA_IDLE))
5919 target = wake_affine_idle(this_cpu, prev_cpu, sync);
5921 if (sched_feat(WA_WEIGHT) && target == nr_cpumask_bits)
5922 target = wake_affine_weight(sd, p, this_cpu, prev_cpu, sync);
5924 schedstat_inc(p->se.statistics.nr_wakeups_affine_attempts);
5925 if (target == nr_cpumask_bits)
5928 schedstat_inc(sd->ttwu_move_affine);
5929 schedstat_inc(p->se.statistics.nr_wakeups_affine);
5933 static unsigned long cpu_util_wake(int cpu, struct task_struct *p);
5935 static unsigned long capacity_spare_wake(int cpu, struct task_struct *p)
5937 return max_t(long, capacity_of(cpu) - cpu_util_wake(cpu, p), 0);
5941 * find_idlest_group finds and returns the least busy CPU group within the
5944 * Assumes p is allowed on at least one CPU in sd.
5946 static struct sched_group *
5947 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
5948 int this_cpu, int sd_flag)
5950 struct sched_group *idlest = NULL, *group = sd->groups;
5951 struct sched_group *most_spare_sg = NULL;
5952 unsigned long min_runnable_load = ULONG_MAX;
5953 unsigned long this_runnable_load = ULONG_MAX;
5954 unsigned long min_avg_load = ULONG_MAX, this_avg_load = ULONG_MAX;
5955 unsigned long most_spare = 0, this_spare = 0;
5956 int load_idx = sd->forkexec_idx;
5957 int imbalance_scale = 100 + (sd->imbalance_pct-100)/2;
5958 unsigned long imbalance = scale_load_down(NICE_0_LOAD) *
5959 (sd->imbalance_pct-100) / 100;
5961 if (sd_flag & SD_BALANCE_WAKE)
5962 load_idx = sd->wake_idx;
5965 unsigned long load, avg_load, runnable_load;
5966 unsigned long spare_cap, max_spare_cap;
5970 /* Skip over this group if it has no CPUs allowed */
5971 if (!cpumask_intersects(sched_group_span(group),
5975 local_group = cpumask_test_cpu(this_cpu,
5976 sched_group_span(group));
5979 * Tally up the load of all CPUs in the group and find
5980 * the group containing the CPU with most spare capacity.
5986 for_each_cpu(i, sched_group_span(group)) {
5987 /* Bias balancing toward CPUs of our domain */
5989 load = source_load(i, load_idx);
5991 load = target_load(i, load_idx);
5993 runnable_load += load;
5995 avg_load += cfs_rq_load_avg(&cpu_rq(i)->cfs);
5997 spare_cap = capacity_spare_wake(i, p);
5999 if (spare_cap > max_spare_cap)
6000 max_spare_cap = spare_cap;
6003 /* Adjust by relative CPU capacity of the group */
6004 avg_load = (avg_load * SCHED_CAPACITY_SCALE) /
6005 group->sgc->capacity;
6006 runnable_load = (runnable_load * SCHED_CAPACITY_SCALE) /
6007 group->sgc->capacity;
6010 this_runnable_load = runnable_load;
6011 this_avg_load = avg_load;
6012 this_spare = max_spare_cap;
6014 if (min_runnable_load > (runnable_load + imbalance)) {
6016 * The runnable load is significantly smaller
6017 * so we can pick this new CPU:
6019 min_runnable_load = runnable_load;
6020 min_avg_load = avg_load;
6022 } else if ((runnable_load < (min_runnable_load + imbalance)) &&
6023 (100*min_avg_load > imbalance_scale*avg_load)) {
6025 * The runnable loads are close so take the
6026 * blocked load into account through avg_load:
6028 min_avg_load = avg_load;
6032 if (most_spare < max_spare_cap) {
6033 most_spare = max_spare_cap;
6034 most_spare_sg = group;
6037 } while (group = group->next, group != sd->groups);
6040 * The cross-over point between using spare capacity or least load
6041 * is too conservative for high utilization tasks on partially
6042 * utilized systems if we require spare_capacity > task_util(p),
6043 * so we allow for some task stuffing by using
6044 * spare_capacity > task_util(p)/2.
6046 * Spare capacity can't be used for fork because the utilization has
6047 * not been set yet, we must first select a rq to compute the initial
6050 if (sd_flag & SD_BALANCE_FORK)
6053 if (this_spare > task_util(p) / 2 &&
6054 imbalance_scale*this_spare > 100*most_spare)
6057 if (most_spare > task_util(p) / 2)
6058 return most_spare_sg;
6065 * When comparing groups across NUMA domains, it's possible for the
6066 * local domain to be very lightly loaded relative to the remote
6067 * domains but "imbalance" skews the comparison making remote CPUs
6068 * look much more favourable. When considering cross-domain, add
6069 * imbalance to the runnable load on the remote node and consider
6072 if ((sd->flags & SD_NUMA) &&
6073 min_runnable_load + imbalance >= this_runnable_load)
6076 if (min_runnable_load > (this_runnable_load + imbalance))
6079 if ((this_runnable_load < (min_runnable_load + imbalance)) &&
6080 (100*this_avg_load < imbalance_scale*min_avg_load))
6087 * find_idlest_group_cpu - find the idlest CPU among the CPUs in the group.
6090 find_idlest_group_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
6092 unsigned long load, min_load = ULONG_MAX;
6093 unsigned int min_exit_latency = UINT_MAX;
6094 u64 latest_idle_timestamp = 0;
6095 int least_loaded_cpu = this_cpu;
6096 int shallowest_idle_cpu = -1;
6099 /* Check if we have any choice: */
6100 if (group->group_weight == 1)
6101 return cpumask_first(sched_group_span(group));
6103 /* Traverse only the allowed CPUs */
6104 for_each_cpu_and(i, sched_group_span(group), &p->cpus_allowed) {
6106 struct rq *rq = cpu_rq(i);
6107 struct cpuidle_state *idle = idle_get_state(rq);
6108 if (idle && idle->exit_latency < min_exit_latency) {
6110 * We give priority to a CPU whose idle state
6111 * has the smallest exit latency irrespective
6112 * of any idle timestamp.
6114 min_exit_latency = idle->exit_latency;
6115 latest_idle_timestamp = rq->idle_stamp;
6116 shallowest_idle_cpu = i;
6117 } else if ((!idle || idle->exit_latency == min_exit_latency) &&
6118 rq->idle_stamp > latest_idle_timestamp) {
6120 * If equal or no active idle state, then
6121 * the most recently idled CPU might have
6124 latest_idle_timestamp = rq->idle_stamp;
6125 shallowest_idle_cpu = i;
6127 } else if (shallowest_idle_cpu == -1) {
6128 load = weighted_cpuload(cpu_rq(i));
6129 if (load < min_load) {
6131 least_loaded_cpu = i;
6136 return shallowest_idle_cpu != -1 ? shallowest_idle_cpu : least_loaded_cpu;
6139 static inline int find_idlest_cpu(struct sched_domain *sd, struct task_struct *p,
6140 int cpu, int prev_cpu, int sd_flag)
6144 if (!cpumask_intersects(sched_domain_span(sd), &p->cpus_allowed))
6148 struct sched_group *group;
6149 struct sched_domain *tmp;
6152 if (!(sd->flags & sd_flag)) {
6157 group = find_idlest_group(sd, p, cpu, sd_flag);
6163 new_cpu = find_idlest_group_cpu(group, p, cpu);
6164 if (new_cpu == cpu) {
6165 /* Now try balancing at a lower domain level of 'cpu': */
6170 /* Now try balancing at a lower domain level of 'new_cpu': */
6172 weight = sd->span_weight;
6174 for_each_domain(cpu, tmp) {
6175 if (weight <= tmp->span_weight)
6177 if (tmp->flags & sd_flag)
6185 #ifdef CONFIG_SCHED_SMT
6187 static inline void set_idle_cores(int cpu, int val)
6189 struct sched_domain_shared *sds;
6191 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
6193 WRITE_ONCE(sds->has_idle_cores, val);
6196 static inline bool test_idle_cores(int cpu, bool def)
6198 struct sched_domain_shared *sds;
6200 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
6202 return READ_ONCE(sds->has_idle_cores);
6208 * Scans the local SMT mask to see if the entire core is idle, and records this
6209 * information in sd_llc_shared->has_idle_cores.
6211 * Since SMT siblings share all cache levels, inspecting this limited remote
6212 * state should be fairly cheap.
6214 void __update_idle_core(struct rq *rq)
6216 int core = cpu_of(rq);
6220 if (test_idle_cores(core, true))
6223 for_each_cpu(cpu, cpu_smt_mask(core)) {
6231 set_idle_cores(core, 1);
6237 * Scan the entire LLC domain for idle cores; this dynamically switches off if
6238 * there are no idle cores left in the system; tracked through
6239 * sd_llc->shared->has_idle_cores and enabled through update_idle_core() above.
6241 static int select_idle_core(struct task_struct *p, struct sched_domain *sd, int target)
6243 struct cpumask *cpus = this_cpu_cpumask_var_ptr(select_idle_mask);
6246 if (!static_branch_likely(&sched_smt_present))
6249 if (!test_idle_cores(target, false))
6252 cpumask_and(cpus, sched_domain_span(sd), &p->cpus_allowed);
6254 for_each_cpu_wrap(core, cpus, target) {
6257 for_each_cpu(cpu, cpu_smt_mask(core)) {
6258 cpumask_clear_cpu(cpu, cpus);
6268 * Failed to find an idle core; stop looking for one.
6270 set_idle_cores(target, 0);
6276 * Scan the local SMT mask for idle CPUs.
6278 static int select_idle_smt(struct task_struct *p, struct sched_domain *sd, int target)
6282 if (!static_branch_likely(&sched_smt_present))
6285 for_each_cpu(cpu, cpu_smt_mask(target)) {
6286 if (!cpumask_test_cpu(cpu, &p->cpus_allowed))
6295 #else /* CONFIG_SCHED_SMT */
6297 static inline int select_idle_core(struct task_struct *p, struct sched_domain *sd, int target)
6302 static inline int select_idle_smt(struct task_struct *p, struct sched_domain *sd, int target)
6307 #endif /* CONFIG_SCHED_SMT */
6310 * Scan the LLC domain for idle CPUs; this is dynamically regulated by
6311 * comparing the average scan cost (tracked in sd->avg_scan_cost) against the
6312 * average idle time for this rq (as found in rq->avg_idle).
6314 static int select_idle_cpu(struct task_struct *p, struct sched_domain *sd, int target)
6316 struct sched_domain *this_sd;
6317 u64 avg_cost, avg_idle;
6320 int cpu, nr = INT_MAX;
6322 this_sd = rcu_dereference(*this_cpu_ptr(&sd_llc));
6327 * Due to large variance we need a large fuzz factor; hackbench in
6328 * particularly is sensitive here.
6330 avg_idle = this_rq()->avg_idle / 512;
6331 avg_cost = this_sd->avg_scan_cost + 1;
6333 if (sched_feat(SIS_AVG_CPU) && avg_idle < avg_cost)
6336 if (sched_feat(SIS_PROP)) {
6337 u64 span_avg = sd->span_weight * avg_idle;
6338 if (span_avg > 4*avg_cost)
6339 nr = div_u64(span_avg, avg_cost);
6344 time = local_clock();
6346 for_each_cpu_wrap(cpu, sched_domain_span(sd), target) {
6349 if (!cpumask_test_cpu(cpu, &p->cpus_allowed))
6355 time = local_clock() - time;
6356 cost = this_sd->avg_scan_cost;
6357 delta = (s64)(time - cost) / 8;
6358 this_sd->avg_scan_cost += delta;
6364 * Try and locate an idle core/thread in the LLC cache domain.
6366 static int select_idle_sibling(struct task_struct *p, int prev, int target)
6368 struct sched_domain *sd;
6369 int i, recent_used_cpu;
6371 if (idle_cpu(target))
6375 * If the previous CPU is cache affine and idle, don't be stupid:
6377 if (prev != target && cpus_share_cache(prev, target) && idle_cpu(prev))
6380 /* Check a recently used CPU as a potential idle candidate: */
6381 recent_used_cpu = p->recent_used_cpu;
6382 if (recent_used_cpu != prev &&
6383 recent_used_cpu != target &&
6384 cpus_share_cache(recent_used_cpu, target) &&
6385 idle_cpu(recent_used_cpu) &&
6386 cpumask_test_cpu(p->recent_used_cpu, &p->cpus_allowed)) {
6388 * Replace recent_used_cpu with prev as it is a potential
6389 * candidate for the next wake:
6391 p->recent_used_cpu = prev;
6392 return recent_used_cpu;
6395 sd = rcu_dereference(per_cpu(sd_llc, target));
6399 i = select_idle_core(p, sd, target);
6400 if ((unsigned)i < nr_cpumask_bits)
6403 i = select_idle_cpu(p, sd, target);
6404 if ((unsigned)i < nr_cpumask_bits)
6407 i = select_idle_smt(p, sd, target);
6408 if ((unsigned)i < nr_cpumask_bits)
6415 * Amount of capacity of a CPU that is (estimated to be) used by CFS tasks
6416 * @cpu: the CPU to get the utilization of
6418 * The unit of the return value must be the one of capacity so we can compare
6419 * the utilization with the capacity of the CPU that is available for CFS task
6420 * (ie cpu_capacity).
6422 * cfs_rq.avg.util_avg is the sum of running time of runnable tasks plus the
6423 * recent utilization of currently non-runnable tasks on a CPU. It represents
6424 * the amount of utilization of a CPU in the range [0..capacity_orig] where
6425 * capacity_orig is the cpu_capacity available at the highest frequency
6426 * (arch_scale_freq_capacity()).
6427 * The utilization of a CPU converges towards a sum equal to or less than the
6428 * current capacity (capacity_curr <= capacity_orig) of the CPU because it is
6429 * the running time on this CPU scaled by capacity_curr.
6431 * The estimated utilization of a CPU is defined to be the maximum between its
6432 * cfs_rq.avg.util_avg and the sum of the estimated utilization of the tasks
6433 * currently RUNNABLE on that CPU.
6434 * This allows to properly represent the expected utilization of a CPU which
6435 * has just got a big task running since a long sleep period. At the same time
6436 * however it preserves the benefits of the "blocked utilization" in
6437 * describing the potential for other tasks waking up on the same CPU.
6439 * Nevertheless, cfs_rq.avg.util_avg can be higher than capacity_curr or even
6440 * higher than capacity_orig because of unfortunate rounding in
6441 * cfs.avg.util_avg or just after migrating tasks and new task wakeups until
6442 * the average stabilizes with the new running time. We need to check that the
6443 * utilization stays within the range of [0..capacity_orig] and cap it if
6444 * necessary. Without utilization capping, a group could be seen as overloaded
6445 * (CPU0 utilization at 121% + CPU1 utilization at 80%) whereas CPU1 has 20% of
6446 * available capacity. We allow utilization to overshoot capacity_curr (but not
6447 * capacity_orig) as it useful for predicting the capacity required after task
6448 * migrations (scheduler-driven DVFS).
6450 * Return: the (estimated) utilization for the specified CPU
6452 static inline unsigned long cpu_util(int cpu)
6454 struct cfs_rq *cfs_rq;
6457 cfs_rq = &cpu_rq(cpu)->cfs;
6458 util = READ_ONCE(cfs_rq->avg.util_avg);
6460 if (sched_feat(UTIL_EST))
6461 util = max(util, READ_ONCE(cfs_rq->avg.util_est.enqueued));
6463 return min_t(unsigned long, util, capacity_orig_of(cpu));
6467 * cpu_util_wake: Compute CPU utilization with any contributions from
6468 * the waking task p removed.
6470 static unsigned long cpu_util_wake(int cpu, struct task_struct *p)
6472 struct cfs_rq *cfs_rq;
6475 /* Task has no contribution or is new */
6476 if (cpu != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time))
6477 return cpu_util(cpu);
6479 cfs_rq = &cpu_rq(cpu)->cfs;
6480 util = READ_ONCE(cfs_rq->avg.util_avg);
6482 /* Discount task's blocked util from CPU's util */
6483 util -= min_t(unsigned int, util, task_util(p));
6488 * a) if *p is the only task sleeping on this CPU, then:
6489 * cpu_util (== task_util) > util_est (== 0)
6490 * and thus we return:
6491 * cpu_util_wake = (cpu_util - task_util) = 0
6493 * b) if other tasks are SLEEPING on this CPU, which is now exiting
6495 * cpu_util >= task_util
6496 * cpu_util > util_est (== 0)
6497 * and thus we discount *p's blocked utilization to return:
6498 * cpu_util_wake = (cpu_util - task_util) >= 0
6500 * c) if other tasks are RUNNABLE on that CPU and
6501 * util_est > cpu_util
6502 * then we use util_est since it returns a more restrictive
6503 * estimation of the spare capacity on that CPU, by just
6504 * considering the expected utilization of tasks already
6505 * runnable on that CPU.
6507 * Cases a) and b) are covered by the above code, while case c) is
6508 * covered by the following code when estimated utilization is
6511 if (sched_feat(UTIL_EST))
6512 util = max(util, READ_ONCE(cfs_rq->avg.util_est.enqueued));
6515 * Utilization (estimated) can exceed the CPU capacity, thus let's
6516 * clamp to the maximum CPU capacity to ensure consistency with
6517 * the cpu_util call.
6519 return min_t(unsigned long, util, capacity_orig_of(cpu));
6523 * Disable WAKE_AFFINE in the case where task @p doesn't fit in the
6524 * capacity of either the waking CPU @cpu or the previous CPU @prev_cpu.
6526 * In that case WAKE_AFFINE doesn't make sense and we'll let
6527 * BALANCE_WAKE sort things out.
6529 static int wake_cap(struct task_struct *p, int cpu, int prev_cpu)
6531 long min_cap, max_cap;
6533 min_cap = min(capacity_orig_of(prev_cpu), capacity_orig_of(cpu));
6534 max_cap = cpu_rq(cpu)->rd->max_cpu_capacity;
6536 /* Minimum capacity is close to max, no need to abort wake_affine */
6537 if (max_cap - min_cap < max_cap >> 3)
6540 /* Bring task utilization in sync with prev_cpu */
6541 sync_entity_load_avg(&p->se);
6543 return min_cap * 1024 < task_util(p) * capacity_margin;
6547 * select_task_rq_fair: Select target runqueue for the waking task in domains
6548 * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE,
6549 * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
6551 * Balances load by selecting the idlest CPU in the idlest group, or under
6552 * certain conditions an idle sibling CPU if the domain has SD_WAKE_AFFINE set.
6554 * Returns the target CPU number.
6556 * preempt must be disabled.
6559 select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags)
6561 struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
6562 int cpu = smp_processor_id();
6563 int new_cpu = prev_cpu;
6564 int want_affine = 0;
6565 int sync = (wake_flags & WF_SYNC) && !(current->flags & PF_EXITING);
6567 if (sd_flag & SD_BALANCE_WAKE) {
6569 want_affine = !wake_wide(p) && !wake_cap(p, cpu, prev_cpu)
6570 && cpumask_test_cpu(cpu, &p->cpus_allowed);
6574 for_each_domain(cpu, tmp) {
6575 if (!(tmp->flags & SD_LOAD_BALANCE))
6579 * If both 'cpu' and 'prev_cpu' are part of this domain,
6580 * cpu is a valid SD_WAKE_AFFINE target.
6582 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
6583 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
6588 if (tmp->flags & sd_flag)
6590 else if (!want_affine)
6595 sd = NULL; /* Prefer wake_affine over balance flags */
6596 if (cpu == prev_cpu)
6599 new_cpu = wake_affine(affine_sd, p, cpu, prev_cpu, sync);
6602 if (sd && !(sd_flag & SD_BALANCE_FORK)) {
6604 * We're going to need the task's util for capacity_spare_wake
6605 * in find_idlest_group. Sync it up to prev_cpu's
6608 sync_entity_load_avg(&p->se);
6613 if (sd_flag & SD_BALANCE_WAKE) { /* XXX always ? */
6614 new_cpu = select_idle_sibling(p, prev_cpu, new_cpu);
6617 current->recent_used_cpu = cpu;
6620 new_cpu = find_idlest_cpu(sd, p, cpu, prev_cpu, sd_flag);
6627 static void detach_entity_cfs_rq(struct sched_entity *se);
6630 * Called immediately before a task is migrated to a new CPU; task_cpu(p) and
6631 * cfs_rq_of(p) references at time of call are still valid and identify the
6632 * previous CPU. The caller guarantees p->pi_lock or task_rq(p)->lock is held.
6634 static void migrate_task_rq_fair(struct task_struct *p)
6637 * As blocked tasks retain absolute vruntime the migration needs to
6638 * deal with this by subtracting the old and adding the new
6639 * min_vruntime -- the latter is done by enqueue_entity() when placing
6640 * the task on the new runqueue.
6642 if (p->state == TASK_WAKING) {
6643 struct sched_entity *se = &p->se;
6644 struct cfs_rq *cfs_rq = cfs_rq_of(se);
6647 #ifndef CONFIG_64BIT
6648 u64 min_vruntime_copy;
6651 min_vruntime_copy = cfs_rq->min_vruntime_copy;
6653 min_vruntime = cfs_rq->min_vruntime;
6654 } while (min_vruntime != min_vruntime_copy);
6656 min_vruntime = cfs_rq->min_vruntime;
6659 se->vruntime -= min_vruntime;
6662 if (p->on_rq == TASK_ON_RQ_MIGRATING) {
6664 * In case of TASK_ON_RQ_MIGRATING we in fact hold the 'old'
6665 * rq->lock and can modify state directly.
6667 lockdep_assert_held(&task_rq(p)->lock);
6668 detach_entity_cfs_rq(&p->se);
6672 * We are supposed to update the task to "current" time, then
6673 * its up to date and ready to go to new CPU/cfs_rq. But we
6674 * have difficulty in getting what current time is, so simply
6675 * throw away the out-of-date time. This will result in the
6676 * wakee task is less decayed, but giving the wakee more load
6679 remove_entity_load_avg(&p->se);
6682 /* Tell new CPU we are migrated */
6683 p->se.avg.last_update_time = 0;
6685 /* We have migrated, no longer consider this task hot */
6686 p->se.exec_start = 0;
6689 static void task_dead_fair(struct task_struct *p)
6691 remove_entity_load_avg(&p->se);
6693 #endif /* CONFIG_SMP */
6695 static unsigned long wakeup_gran(struct sched_entity *se)
6697 unsigned long gran = sysctl_sched_wakeup_granularity;
6700 * Since its curr running now, convert the gran from real-time
6701 * to virtual-time in his units.
6703 * By using 'se' instead of 'curr' we penalize light tasks, so
6704 * they get preempted easier. That is, if 'se' < 'curr' then
6705 * the resulting gran will be larger, therefore penalizing the
6706 * lighter, if otoh 'se' > 'curr' then the resulting gran will
6707 * be smaller, again penalizing the lighter task.
6709 * This is especially important for buddies when the leftmost
6710 * task is higher priority than the buddy.
6712 return calc_delta_fair(gran, se);
6716 * Should 'se' preempt 'curr'.
6730 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
6732 s64 gran, vdiff = curr->vruntime - se->vruntime;
6737 gran = wakeup_gran(se);
6744 static void set_last_buddy(struct sched_entity *se)
6746 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
6749 for_each_sched_entity(se) {
6750 if (SCHED_WARN_ON(!se->on_rq))
6752 cfs_rq_of(se)->last = se;
6756 static void set_next_buddy(struct sched_entity *se)
6758 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
6761 for_each_sched_entity(se) {
6762 if (SCHED_WARN_ON(!se->on_rq))
6764 cfs_rq_of(se)->next = se;
6768 static void set_skip_buddy(struct sched_entity *se)
6770 for_each_sched_entity(se)
6771 cfs_rq_of(se)->skip = se;
6775 * Preempt the current task with a newly woken task if needed:
6777 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
6779 struct task_struct *curr = rq->curr;
6780 struct sched_entity *se = &curr->se, *pse = &p->se;
6781 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
6782 int scale = cfs_rq->nr_running >= sched_nr_latency;
6783 int next_buddy_marked = 0;
6785 if (unlikely(se == pse))
6789 * This is possible from callers such as attach_tasks(), in which we
6790 * unconditionally check_prempt_curr() after an enqueue (which may have
6791 * lead to a throttle). This both saves work and prevents false
6792 * next-buddy nomination below.
6794 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
6797 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
6798 set_next_buddy(pse);
6799 next_buddy_marked = 1;
6803 * We can come here with TIF_NEED_RESCHED already set from new task
6806 * Note: this also catches the edge-case of curr being in a throttled
6807 * group (e.g. via set_curr_task), since update_curr() (in the
6808 * enqueue of curr) will have resulted in resched being set. This
6809 * prevents us from potentially nominating it as a false LAST_BUDDY
6812 if (test_tsk_need_resched(curr))
6815 /* Idle tasks are by definition preempted by non-idle tasks. */
6816 if (unlikely(curr->policy == SCHED_IDLE) &&
6817 likely(p->policy != SCHED_IDLE))
6821 * Batch and idle tasks do not preempt non-idle tasks (their preemption
6822 * is driven by the tick):
6824 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
6827 find_matching_se(&se, &pse);
6828 update_curr(cfs_rq_of(se));
6830 if (wakeup_preempt_entity(se, pse) == 1) {
6832 * Bias pick_next to pick the sched entity that is
6833 * triggering this preemption.
6835 if (!next_buddy_marked)
6836 set_next_buddy(pse);
6845 * Only set the backward buddy when the current task is still
6846 * on the rq. This can happen when a wakeup gets interleaved
6847 * with schedule on the ->pre_schedule() or idle_balance()
6848 * point, either of which can * drop the rq lock.
6850 * Also, during early boot the idle thread is in the fair class,
6851 * for obvious reasons its a bad idea to schedule back to it.
6853 if (unlikely(!se->on_rq || curr == rq->idle))
6856 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
6860 static struct task_struct *
6861 pick_next_task_fair(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
6863 struct cfs_rq *cfs_rq = &rq->cfs;
6864 struct sched_entity *se;
6865 struct task_struct *p;
6869 if (!cfs_rq->nr_running)
6872 #ifdef CONFIG_FAIR_GROUP_SCHED
6873 if (prev->sched_class != &fair_sched_class)
6877 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
6878 * likely that a next task is from the same cgroup as the current.
6880 * Therefore attempt to avoid putting and setting the entire cgroup
6881 * hierarchy, only change the part that actually changes.
6885 struct sched_entity *curr = cfs_rq->curr;
6888 * Since we got here without doing put_prev_entity() we also
6889 * have to consider cfs_rq->curr. If it is still a runnable
6890 * entity, update_curr() will update its vruntime, otherwise
6891 * forget we've ever seen it.
6895 update_curr(cfs_rq);
6900 * This call to check_cfs_rq_runtime() will do the
6901 * throttle and dequeue its entity in the parent(s).
6902 * Therefore the nr_running test will indeed
6905 if (unlikely(check_cfs_rq_runtime(cfs_rq))) {
6908 if (!cfs_rq->nr_running)
6915 se = pick_next_entity(cfs_rq, curr);
6916 cfs_rq = group_cfs_rq(se);
6922 * Since we haven't yet done put_prev_entity and if the selected task
6923 * is a different task than we started out with, try and touch the
6924 * least amount of cfs_rqs.
6927 struct sched_entity *pse = &prev->se;
6929 while (!(cfs_rq = is_same_group(se, pse))) {
6930 int se_depth = se->depth;
6931 int pse_depth = pse->depth;
6933 if (se_depth <= pse_depth) {
6934 put_prev_entity(cfs_rq_of(pse), pse);
6935 pse = parent_entity(pse);
6937 if (se_depth >= pse_depth) {
6938 set_next_entity(cfs_rq_of(se), se);
6939 se = parent_entity(se);
6943 put_prev_entity(cfs_rq, pse);
6944 set_next_entity(cfs_rq, se);
6951 put_prev_task(rq, prev);
6954 se = pick_next_entity(cfs_rq, NULL);
6955 set_next_entity(cfs_rq, se);
6956 cfs_rq = group_cfs_rq(se);
6961 done: __maybe_unused;
6964 * Move the next running task to the front of
6965 * the list, so our cfs_tasks list becomes MRU
6968 list_move(&p->se.group_node, &rq->cfs_tasks);
6971 if (hrtick_enabled(rq))
6972 hrtick_start_fair(rq, p);
6977 new_tasks = idle_balance(rq, rf);
6980 * Because idle_balance() releases (and re-acquires) rq->lock, it is
6981 * possible for any higher priority task to appear. In that case we
6982 * must re-start the pick_next_entity() loop.
6994 * Account for a descheduled task:
6996 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
6998 struct sched_entity *se = &prev->se;
6999 struct cfs_rq *cfs_rq;
7001 for_each_sched_entity(se) {
7002 cfs_rq = cfs_rq_of(se);
7003 put_prev_entity(cfs_rq, se);
7008 * sched_yield() is very simple
7010 * The magic of dealing with the ->skip buddy is in pick_next_entity.
7012 static void yield_task_fair(struct rq *rq)
7014 struct task_struct *curr = rq->curr;
7015 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
7016 struct sched_entity *se = &curr->se;
7019 * Are we the only task in the tree?
7021 if (unlikely(rq->nr_running == 1))
7024 clear_buddies(cfs_rq, se);
7026 if (curr->policy != SCHED_BATCH) {
7027 update_rq_clock(rq);
7029 * Update run-time statistics of the 'current'.
7031 update_curr(cfs_rq);
7033 * Tell update_rq_clock() that we've just updated,
7034 * so we don't do microscopic update in schedule()
7035 * and double the fastpath cost.
7037 rq_clock_skip_update(rq);
7043 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
7045 struct sched_entity *se = &p->se;
7047 /* throttled hierarchies are not runnable */
7048 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
7051 /* Tell the scheduler that we'd really like pse to run next. */
7054 yield_task_fair(rq);
7060 /**************************************************
7061 * Fair scheduling class load-balancing methods.
7065 * The purpose of load-balancing is to achieve the same basic fairness the
7066 * per-CPU scheduler provides, namely provide a proportional amount of compute
7067 * time to each task. This is expressed in the following equation:
7069 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
7071 * Where W_i,n is the n-th weight average for CPU i. The instantaneous weight
7072 * W_i,0 is defined as:
7074 * W_i,0 = \Sum_j w_i,j (2)
7076 * Where w_i,j is the weight of the j-th runnable task on CPU i. This weight
7077 * is derived from the nice value as per sched_prio_to_weight[].
7079 * The weight average is an exponential decay average of the instantaneous
7082 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
7084 * C_i is the compute capacity of CPU i, typically it is the
7085 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
7086 * can also include other factors [XXX].
7088 * To achieve this balance we define a measure of imbalance which follows
7089 * directly from (1):
7091 * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4)
7093 * We them move tasks around to minimize the imbalance. In the continuous
7094 * function space it is obvious this converges, in the discrete case we get
7095 * a few fun cases generally called infeasible weight scenarios.
7098 * - infeasible weights;
7099 * - local vs global optima in the discrete case. ]
7104 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
7105 * for all i,j solution, we create a tree of CPUs that follows the hardware
7106 * topology where each level pairs two lower groups (or better). This results
7107 * in O(log n) layers. Furthermore we reduce the number of CPUs going up the
7108 * tree to only the first of the previous level and we decrease the frequency
7109 * of load-balance at each level inv. proportional to the number of CPUs in
7115 * \Sum { --- * --- * 2^i } = O(n) (5)
7117 * `- size of each group
7118 * | | `- number of CPUs doing load-balance
7120 * `- sum over all levels
7122 * Coupled with a limit on how many tasks we can migrate every balance pass,
7123 * this makes (5) the runtime complexity of the balancer.
7125 * An important property here is that each CPU is still (indirectly) connected
7126 * to every other CPU in at most O(log n) steps:
7128 * The adjacency matrix of the resulting graph is given by:
7131 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
7134 * And you'll find that:
7136 * A^(log_2 n)_i,j != 0 for all i,j (7)
7138 * Showing there's indeed a path between every CPU in at most O(log n) steps.
7139 * The task movement gives a factor of O(m), giving a convergence complexity
7142 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
7147 * In order to avoid CPUs going idle while there's still work to do, new idle
7148 * balancing is more aggressive and has the newly idle CPU iterate up the domain
7149 * tree itself instead of relying on other CPUs to bring it work.
7151 * This adds some complexity to both (5) and (8) but it reduces the total idle
7159 * Cgroups make a horror show out of (2), instead of a simple sum we get:
7162 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
7167 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
7169 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on CPU i.
7171 * The big problem is S_k, its a global sum needed to compute a local (W_i)
7174 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
7175 * rewrite all of this once again.]
7178 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
7180 enum fbq_type { regular, remote, all };
7182 #define LBF_ALL_PINNED 0x01
7183 #define LBF_NEED_BREAK 0x02
7184 #define LBF_DST_PINNED 0x04
7185 #define LBF_SOME_PINNED 0x08
7186 #define LBF_NOHZ_STATS 0x10
7187 #define LBF_NOHZ_AGAIN 0x20
7190 struct sched_domain *sd;
7198 struct cpumask *dst_grpmask;
7200 enum cpu_idle_type idle;
7202 /* The set of CPUs under consideration for load-balancing */
7203 struct cpumask *cpus;
7208 unsigned int loop_break;
7209 unsigned int loop_max;
7211 enum fbq_type fbq_type;
7212 struct list_head tasks;
7216 * Is this task likely cache-hot:
7218 static int task_hot(struct task_struct *p, struct lb_env *env)
7222 lockdep_assert_held(&env->src_rq->lock);
7224 if (p->sched_class != &fair_sched_class)
7227 if (unlikely(p->policy == SCHED_IDLE))
7231 * Buddy candidates are cache hot:
7233 if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running &&
7234 (&p->se == cfs_rq_of(&p->se)->next ||
7235 &p->se == cfs_rq_of(&p->se)->last))
7238 if (sysctl_sched_migration_cost == -1)
7240 if (sysctl_sched_migration_cost == 0)
7243 delta = rq_clock_task(env->src_rq) - p->se.exec_start;
7245 return delta < (s64)sysctl_sched_migration_cost;
7248 #ifdef CONFIG_NUMA_BALANCING
7250 * Returns 1, if task migration degrades locality
7251 * Returns 0, if task migration improves locality i.e migration preferred.
7252 * Returns -1, if task migration is not affected by locality.
7254 static int migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
7256 struct numa_group *numa_group = rcu_dereference(p->numa_group);
7257 unsigned long src_faults, dst_faults;
7258 int src_nid, dst_nid;
7260 if (!static_branch_likely(&sched_numa_balancing))
7263 if (!p->numa_faults || !(env->sd->flags & SD_NUMA))
7266 src_nid = cpu_to_node(env->src_cpu);
7267 dst_nid = cpu_to_node(env->dst_cpu);
7269 if (src_nid == dst_nid)
7272 /* Migrating away from the preferred node is always bad. */
7273 if (src_nid == p->numa_preferred_nid) {
7274 if (env->src_rq->nr_running > env->src_rq->nr_preferred_running)
7280 /* Encourage migration to the preferred node. */
7281 if (dst_nid == p->numa_preferred_nid)
7284 /* Leaving a core idle is often worse than degrading locality. */
7285 if (env->idle != CPU_NOT_IDLE)
7289 src_faults = group_faults(p, src_nid);
7290 dst_faults = group_faults(p, dst_nid);
7292 src_faults = task_faults(p, src_nid);
7293 dst_faults = task_faults(p, dst_nid);
7296 return dst_faults < src_faults;
7300 static inline int migrate_degrades_locality(struct task_struct *p,
7308 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
7311 int can_migrate_task(struct task_struct *p, struct lb_env *env)
7315 lockdep_assert_held(&env->src_rq->lock);
7318 * We do not migrate tasks that are:
7319 * 1) throttled_lb_pair, or
7320 * 2) cannot be migrated to this CPU due to cpus_allowed, or
7321 * 3) running (obviously), or
7322 * 4) are cache-hot on their current CPU.
7324 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
7327 if (!cpumask_test_cpu(env->dst_cpu, &p->cpus_allowed)) {
7330 schedstat_inc(p->se.statistics.nr_failed_migrations_affine);
7332 env->flags |= LBF_SOME_PINNED;
7335 * Remember if this task can be migrated to any other CPU in
7336 * our sched_group. We may want to revisit it if we couldn't
7337 * meet load balance goals by pulling other tasks on src_cpu.
7339 * Avoid computing new_dst_cpu for NEWLY_IDLE or if we have
7340 * already computed one in current iteration.
7342 if (env->idle == CPU_NEWLY_IDLE || (env->flags & LBF_DST_PINNED))
7345 /* Prevent to re-select dst_cpu via env's CPUs: */
7346 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
7347 if (cpumask_test_cpu(cpu, &p->cpus_allowed)) {
7348 env->flags |= LBF_DST_PINNED;
7349 env->new_dst_cpu = cpu;
7357 /* Record that we found atleast one task that could run on dst_cpu */
7358 env->flags &= ~LBF_ALL_PINNED;
7360 if (task_running(env->src_rq, p)) {
7361 schedstat_inc(p->se.statistics.nr_failed_migrations_running);
7366 * Aggressive migration if:
7367 * 1) destination numa is preferred
7368 * 2) task is cache cold, or
7369 * 3) too many balance attempts have failed.
7371 tsk_cache_hot = migrate_degrades_locality(p, env);
7372 if (tsk_cache_hot == -1)
7373 tsk_cache_hot = task_hot(p, env);
7375 if (tsk_cache_hot <= 0 ||
7376 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
7377 if (tsk_cache_hot == 1) {
7378 schedstat_inc(env->sd->lb_hot_gained[env->idle]);
7379 schedstat_inc(p->se.statistics.nr_forced_migrations);
7384 schedstat_inc(p->se.statistics.nr_failed_migrations_hot);
7389 * detach_task() -- detach the task for the migration specified in env
7391 static void detach_task(struct task_struct *p, struct lb_env *env)
7393 lockdep_assert_held(&env->src_rq->lock);
7395 p->on_rq = TASK_ON_RQ_MIGRATING;
7396 deactivate_task(env->src_rq, p, DEQUEUE_NOCLOCK);
7397 set_task_cpu(p, env->dst_cpu);
7401 * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
7402 * part of active balancing operations within "domain".
7404 * Returns a task if successful and NULL otherwise.
7406 static struct task_struct *detach_one_task(struct lb_env *env)
7408 struct task_struct *p;
7410 lockdep_assert_held(&env->src_rq->lock);
7412 list_for_each_entry_reverse(p,
7413 &env->src_rq->cfs_tasks, se.group_node) {
7414 if (!can_migrate_task(p, env))
7417 detach_task(p, env);
7420 * Right now, this is only the second place where
7421 * lb_gained[env->idle] is updated (other is detach_tasks)
7422 * so we can safely collect stats here rather than
7423 * inside detach_tasks().
7425 schedstat_inc(env->sd->lb_gained[env->idle]);
7431 static const unsigned int sched_nr_migrate_break = 32;
7434 * detach_tasks() -- tries to detach up to imbalance weighted load from
7435 * busiest_rq, as part of a balancing operation within domain "sd".
7437 * Returns number of detached tasks if successful and 0 otherwise.
7439 static int detach_tasks(struct lb_env *env)
7441 struct list_head *tasks = &env->src_rq->cfs_tasks;
7442 struct task_struct *p;
7446 lockdep_assert_held(&env->src_rq->lock);
7448 if (env->imbalance <= 0)
7451 while (!list_empty(tasks)) {
7453 * We don't want to steal all, otherwise we may be treated likewise,
7454 * which could at worst lead to a livelock crash.
7456 if (env->idle != CPU_NOT_IDLE && env->src_rq->nr_running <= 1)
7459 p = list_last_entry(tasks, struct task_struct, se.group_node);
7462 /* We've more or less seen every task there is, call it quits */
7463 if (env->loop > env->loop_max)
7466 /* take a breather every nr_migrate tasks */
7467 if (env->loop > env->loop_break) {
7468 env->loop_break += sched_nr_migrate_break;
7469 env->flags |= LBF_NEED_BREAK;
7473 if (!can_migrate_task(p, env))
7476 load = task_h_load(p);
7478 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
7481 if ((load / 2) > env->imbalance)
7484 detach_task(p, env);
7485 list_add(&p->se.group_node, &env->tasks);
7488 env->imbalance -= load;
7490 #ifdef CONFIG_PREEMPT
7492 * NEWIDLE balancing is a source of latency, so preemptible
7493 * kernels will stop after the first task is detached to minimize
7494 * the critical section.
7496 if (env->idle == CPU_NEWLY_IDLE)
7501 * We only want to steal up to the prescribed amount of
7504 if (env->imbalance <= 0)
7509 list_move(&p->se.group_node, tasks);
7513 * Right now, this is one of only two places we collect this stat
7514 * so we can safely collect detach_one_task() stats here rather
7515 * than inside detach_one_task().
7517 schedstat_add(env->sd->lb_gained[env->idle], detached);
7523 * attach_task() -- attach the task detached by detach_task() to its new rq.
7525 static void attach_task(struct rq *rq, struct task_struct *p)
7527 lockdep_assert_held(&rq->lock);
7529 BUG_ON(task_rq(p) != rq);
7530 activate_task(rq, p, ENQUEUE_NOCLOCK);
7531 p->on_rq = TASK_ON_RQ_QUEUED;
7532 check_preempt_curr(rq, p, 0);
7536 * attach_one_task() -- attaches the task returned from detach_one_task() to
7539 static void attach_one_task(struct rq *rq, struct task_struct *p)
7544 update_rq_clock(rq);
7550 * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
7553 static void attach_tasks(struct lb_env *env)
7555 struct list_head *tasks = &env->tasks;
7556 struct task_struct *p;
7559 rq_lock(env->dst_rq, &rf);
7560 update_rq_clock(env->dst_rq);
7562 while (!list_empty(tasks)) {
7563 p = list_first_entry(tasks, struct task_struct, se.group_node);
7564 list_del_init(&p->se.group_node);
7566 attach_task(env->dst_rq, p);
7569 rq_unlock(env->dst_rq, &rf);
7572 static inline bool cfs_rq_has_blocked(struct cfs_rq *cfs_rq)
7574 if (cfs_rq->avg.load_avg)
7577 if (cfs_rq->avg.util_avg)
7583 #ifdef CONFIG_FAIR_GROUP_SCHED
7585 static inline bool cfs_rq_is_decayed(struct cfs_rq *cfs_rq)
7587 if (cfs_rq->load.weight)
7590 if (cfs_rq->avg.load_sum)
7593 if (cfs_rq->avg.util_sum)
7596 if (cfs_rq->avg.runnable_load_sum)
7602 static void update_blocked_averages(int cpu)
7604 struct rq *rq = cpu_rq(cpu);
7605 struct cfs_rq *cfs_rq, *pos;
7609 rq_lock_irqsave(rq, &rf);
7610 update_rq_clock(rq);
7613 * Iterates the task_group tree in a bottom up fashion, see
7614 * list_add_leaf_cfs_rq() for details.
7616 for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) {
7617 struct sched_entity *se;
7619 /* throttled entities do not contribute to load */
7620 if (throttled_hierarchy(cfs_rq))
7623 if (update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq))
7624 update_tg_load_avg(cfs_rq, 0);
7626 /* Propagate pending load changes to the parent, if any: */
7627 se = cfs_rq->tg->se[cpu];
7628 if (se && !skip_blocked_update(se))
7629 update_load_avg(cfs_rq_of(se), se, 0);
7632 * There can be a lot of idle CPU cgroups. Don't let fully
7633 * decayed cfs_rqs linger on the list.
7635 if (cfs_rq_is_decayed(cfs_rq))
7636 list_del_leaf_cfs_rq(cfs_rq);
7638 /* Don't need periodic decay once load/util_avg are null */
7639 if (cfs_rq_has_blocked(cfs_rq))
7643 #ifdef CONFIG_NO_HZ_COMMON
7644 rq->last_blocked_load_update_tick = jiffies;
7646 rq->has_blocked_load = 0;
7648 rq_unlock_irqrestore(rq, &rf);
7652 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
7653 * This needs to be done in a top-down fashion because the load of a child
7654 * group is a fraction of its parents load.
7656 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
7658 struct rq *rq = rq_of(cfs_rq);
7659 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
7660 unsigned long now = jiffies;
7663 if (cfs_rq->last_h_load_update == now)
7666 cfs_rq->h_load_next = NULL;
7667 for_each_sched_entity(se) {
7668 cfs_rq = cfs_rq_of(se);
7669 cfs_rq->h_load_next = se;
7670 if (cfs_rq->last_h_load_update == now)
7675 cfs_rq->h_load = cfs_rq_load_avg(cfs_rq);
7676 cfs_rq->last_h_load_update = now;
7679 while ((se = cfs_rq->h_load_next) != NULL) {
7680 load = cfs_rq->h_load;
7681 load = div64_ul(load * se->avg.load_avg,
7682 cfs_rq_load_avg(cfs_rq) + 1);
7683 cfs_rq = group_cfs_rq(se);
7684 cfs_rq->h_load = load;
7685 cfs_rq->last_h_load_update = now;
7689 static unsigned long task_h_load(struct task_struct *p)
7691 struct cfs_rq *cfs_rq = task_cfs_rq(p);
7693 update_cfs_rq_h_load(cfs_rq);
7694 return div64_ul(p->se.avg.load_avg * cfs_rq->h_load,
7695 cfs_rq_load_avg(cfs_rq) + 1);
7698 static inline void update_blocked_averages(int cpu)
7700 struct rq *rq = cpu_rq(cpu);
7701 struct cfs_rq *cfs_rq = &rq->cfs;
7704 rq_lock_irqsave(rq, &rf);
7705 update_rq_clock(rq);
7706 update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq);
7707 #ifdef CONFIG_NO_HZ_COMMON
7708 rq->last_blocked_load_update_tick = jiffies;
7709 if (!cfs_rq_has_blocked(cfs_rq))
7710 rq->has_blocked_load = 0;
7712 rq_unlock_irqrestore(rq, &rf);
7715 static unsigned long task_h_load(struct task_struct *p)
7717 return p->se.avg.load_avg;
7721 /********** Helpers for find_busiest_group ************************/
7730 * sg_lb_stats - stats of a sched_group required for load_balancing
7732 struct sg_lb_stats {
7733 unsigned long avg_load; /*Avg load across the CPUs of the group */
7734 unsigned long group_load; /* Total load over the CPUs of the group */
7735 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
7736 unsigned long load_per_task;
7737 unsigned long group_capacity;
7738 unsigned long group_util; /* Total utilization of the group */
7739 unsigned int sum_nr_running; /* Nr tasks running in the group */
7740 unsigned int idle_cpus;
7741 unsigned int group_weight;
7742 enum group_type group_type;
7743 int group_no_capacity;
7744 #ifdef CONFIG_NUMA_BALANCING
7745 unsigned int nr_numa_running;
7746 unsigned int nr_preferred_running;
7751 * sd_lb_stats - Structure to store the statistics of a sched_domain
7752 * during load balancing.
7754 struct sd_lb_stats {
7755 struct sched_group *busiest; /* Busiest group in this sd */
7756 struct sched_group *local; /* Local group in this sd */
7757 unsigned long total_running;
7758 unsigned long total_load; /* Total load of all groups in sd */
7759 unsigned long total_capacity; /* Total capacity of all groups in sd */
7760 unsigned long avg_load; /* Average load across all groups in sd */
7762 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
7763 struct sg_lb_stats local_stat; /* Statistics of the local group */
7766 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
7769 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
7770 * local_stat because update_sg_lb_stats() does a full clear/assignment.
7771 * We must however clear busiest_stat::avg_load because
7772 * update_sd_pick_busiest() reads this before assignment.
7774 *sds = (struct sd_lb_stats){
7777 .total_running = 0UL,
7779 .total_capacity = 0UL,
7782 .sum_nr_running = 0,
7783 .group_type = group_other,
7789 * get_sd_load_idx - Obtain the load index for a given sched domain.
7790 * @sd: The sched_domain whose load_idx is to be obtained.
7791 * @idle: The idle status of the CPU for whose sd load_idx is obtained.
7793 * Return: The load index.
7795 static inline int get_sd_load_idx(struct sched_domain *sd,
7796 enum cpu_idle_type idle)
7802 load_idx = sd->busy_idx;
7805 case CPU_NEWLY_IDLE:
7806 load_idx = sd->newidle_idx;
7809 load_idx = sd->idle_idx;
7816 static unsigned long scale_rt_capacity(int cpu)
7818 struct rq *rq = cpu_rq(cpu);
7819 u64 total, used, age_stamp, avg;
7823 * Since we're reading these variables without serialization make sure
7824 * we read them once before doing sanity checks on them.
7826 age_stamp = READ_ONCE(rq->age_stamp);
7827 avg = READ_ONCE(rq->rt_avg);
7828 delta = __rq_clock_broken(rq) - age_stamp;
7830 if (unlikely(delta < 0))
7833 total = sched_avg_period() + delta;
7835 used = div_u64(avg, total);
7837 if (likely(used < SCHED_CAPACITY_SCALE))
7838 return SCHED_CAPACITY_SCALE - used;
7843 static void update_cpu_capacity(struct sched_domain *sd, int cpu)
7845 unsigned long capacity = arch_scale_cpu_capacity(sd, cpu);
7846 struct sched_group *sdg = sd->groups;
7848 cpu_rq(cpu)->cpu_capacity_orig = capacity;
7850 capacity *= scale_rt_capacity(cpu);
7851 capacity >>= SCHED_CAPACITY_SHIFT;
7856 cpu_rq(cpu)->cpu_capacity = capacity;
7857 sdg->sgc->capacity = capacity;
7858 sdg->sgc->min_capacity = capacity;
7861 void update_group_capacity(struct sched_domain *sd, int cpu)
7863 struct sched_domain *child = sd->child;
7864 struct sched_group *group, *sdg = sd->groups;
7865 unsigned long capacity, min_capacity;
7866 unsigned long interval;
7868 interval = msecs_to_jiffies(sd->balance_interval);
7869 interval = clamp(interval, 1UL, max_load_balance_interval);
7870 sdg->sgc->next_update = jiffies + interval;
7873 update_cpu_capacity(sd, cpu);
7878 min_capacity = ULONG_MAX;
7880 if (child->flags & SD_OVERLAP) {
7882 * SD_OVERLAP domains cannot assume that child groups
7883 * span the current group.
7886 for_each_cpu(cpu, sched_group_span(sdg)) {
7887 struct sched_group_capacity *sgc;
7888 struct rq *rq = cpu_rq(cpu);
7891 * build_sched_domains() -> init_sched_groups_capacity()
7892 * gets here before we've attached the domains to the
7895 * Use capacity_of(), which is set irrespective of domains
7896 * in update_cpu_capacity().
7898 * This avoids capacity from being 0 and
7899 * causing divide-by-zero issues on boot.
7901 if (unlikely(!rq->sd)) {
7902 capacity += capacity_of(cpu);
7904 sgc = rq->sd->groups->sgc;
7905 capacity += sgc->capacity;
7908 min_capacity = min(capacity, min_capacity);
7912 * !SD_OVERLAP domains can assume that child groups
7913 * span the current group.
7916 group = child->groups;
7918 struct sched_group_capacity *sgc = group->sgc;
7920 capacity += sgc->capacity;
7921 min_capacity = min(sgc->min_capacity, min_capacity);
7922 group = group->next;
7923 } while (group != child->groups);
7926 sdg->sgc->capacity = capacity;
7927 sdg->sgc->min_capacity = min_capacity;
7931 * Check whether the capacity of the rq has been noticeably reduced by side
7932 * activity. The imbalance_pct is used for the threshold.
7933 * Return true is the capacity is reduced
7936 check_cpu_capacity(struct rq *rq, struct sched_domain *sd)
7938 return ((rq->cpu_capacity * sd->imbalance_pct) <
7939 (rq->cpu_capacity_orig * 100));
7943 * Group imbalance indicates (and tries to solve) the problem where balancing
7944 * groups is inadequate due to ->cpus_allowed constraints.
7946 * Imagine a situation of two groups of 4 CPUs each and 4 tasks each with a
7947 * cpumask covering 1 CPU of the first group and 3 CPUs of the second group.
7950 * { 0 1 2 3 } { 4 5 6 7 }
7953 * If we were to balance group-wise we'd place two tasks in the first group and
7954 * two tasks in the second group. Clearly this is undesired as it will overload
7955 * cpu 3 and leave one of the CPUs in the second group unused.
7957 * The current solution to this issue is detecting the skew in the first group
7958 * by noticing the lower domain failed to reach balance and had difficulty
7959 * moving tasks due to affinity constraints.
7961 * When this is so detected; this group becomes a candidate for busiest; see
7962 * update_sd_pick_busiest(). And calculate_imbalance() and
7963 * find_busiest_group() avoid some of the usual balance conditions to allow it
7964 * to create an effective group imbalance.
7966 * This is a somewhat tricky proposition since the next run might not find the
7967 * group imbalance and decide the groups need to be balanced again. A most
7968 * subtle and fragile situation.
7971 static inline int sg_imbalanced(struct sched_group *group)
7973 return group->sgc->imbalance;
7977 * group_has_capacity returns true if the group has spare capacity that could
7978 * be used by some tasks.
7979 * We consider that a group has spare capacity if the * number of task is
7980 * smaller than the number of CPUs or if the utilization is lower than the
7981 * available capacity for CFS tasks.
7982 * For the latter, we use a threshold to stabilize the state, to take into
7983 * account the variance of the tasks' load and to return true if the available
7984 * capacity in meaningful for the load balancer.
7985 * As an example, an available capacity of 1% can appear but it doesn't make
7986 * any benefit for the load balance.
7989 group_has_capacity(struct lb_env *env, struct sg_lb_stats *sgs)
7991 if (sgs->sum_nr_running < sgs->group_weight)
7994 if ((sgs->group_capacity * 100) >
7995 (sgs->group_util * env->sd->imbalance_pct))
8002 * group_is_overloaded returns true if the group has more tasks than it can
8004 * group_is_overloaded is not equals to !group_has_capacity because a group
8005 * with the exact right number of tasks, has no more spare capacity but is not
8006 * overloaded so both group_has_capacity and group_is_overloaded return
8010 group_is_overloaded(struct lb_env *env, struct sg_lb_stats *sgs)
8012 if (sgs->sum_nr_running <= sgs->group_weight)
8015 if ((sgs->group_capacity * 100) <
8016 (sgs->group_util * env->sd->imbalance_pct))
8023 * group_smaller_cpu_capacity: Returns true if sched_group sg has smaller
8024 * per-CPU capacity than sched_group ref.
8027 group_smaller_cpu_capacity(struct sched_group *sg, struct sched_group *ref)
8029 return sg->sgc->min_capacity * capacity_margin <
8030 ref->sgc->min_capacity * 1024;
8034 group_type group_classify(struct sched_group *group,
8035 struct sg_lb_stats *sgs)
8037 if (sgs->group_no_capacity)
8038 return group_overloaded;
8040 if (sg_imbalanced(group))
8041 return group_imbalanced;
8046 static bool update_nohz_stats(struct rq *rq, bool force)
8048 #ifdef CONFIG_NO_HZ_COMMON
8049 unsigned int cpu = rq->cpu;
8051 if (!rq->has_blocked_load)
8054 if (!cpumask_test_cpu(cpu, nohz.idle_cpus_mask))
8057 if (!force && !time_after(jiffies, rq->last_blocked_load_update_tick))
8060 update_blocked_averages(cpu);
8062 return rq->has_blocked_load;
8069 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
8070 * @env: The load balancing environment.
8071 * @group: sched_group whose statistics are to be updated.
8072 * @load_idx: Load index of sched_domain of this_cpu for load calc.
8073 * @local_group: Does group contain this_cpu.
8074 * @sgs: variable to hold the statistics for this group.
8075 * @overload: Indicate more than one runnable task for any CPU.
8077 static inline void update_sg_lb_stats(struct lb_env *env,
8078 struct sched_group *group, int load_idx,
8079 int local_group, struct sg_lb_stats *sgs,
8085 memset(sgs, 0, sizeof(*sgs));
8087 for_each_cpu_and(i, sched_group_span(group), env->cpus) {
8088 struct rq *rq = cpu_rq(i);
8090 if ((env->flags & LBF_NOHZ_STATS) && update_nohz_stats(rq, false))
8091 env->flags |= LBF_NOHZ_AGAIN;
8093 /* Bias balancing toward CPUs of our domain: */
8095 load = target_load(i, load_idx);
8097 load = source_load(i, load_idx);
8099 sgs->group_load += load;
8100 sgs->group_util += cpu_util(i);
8101 sgs->sum_nr_running += rq->cfs.h_nr_running;
8103 nr_running = rq->nr_running;
8107 #ifdef CONFIG_NUMA_BALANCING
8108 sgs->nr_numa_running += rq->nr_numa_running;
8109 sgs->nr_preferred_running += rq->nr_preferred_running;
8111 sgs->sum_weighted_load += weighted_cpuload(rq);
8113 * No need to call idle_cpu() if nr_running is not 0
8115 if (!nr_running && idle_cpu(i))
8119 /* Adjust by relative CPU capacity of the group */
8120 sgs->group_capacity = group->sgc->capacity;
8121 sgs->avg_load = (sgs->group_load*SCHED_CAPACITY_SCALE) / sgs->group_capacity;
8123 if (sgs->sum_nr_running)
8124 sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
8126 sgs->group_weight = group->group_weight;
8128 sgs->group_no_capacity = group_is_overloaded(env, sgs);
8129 sgs->group_type = group_classify(group, sgs);
8133 * update_sd_pick_busiest - return 1 on busiest group
8134 * @env: The load balancing environment.
8135 * @sds: sched_domain statistics
8136 * @sg: sched_group candidate to be checked for being the busiest
8137 * @sgs: sched_group statistics
8139 * Determine if @sg is a busier group than the previously selected
8142 * Return: %true if @sg is a busier group than the previously selected
8143 * busiest group. %false otherwise.
8145 static bool update_sd_pick_busiest(struct lb_env *env,
8146 struct sd_lb_stats *sds,
8147 struct sched_group *sg,
8148 struct sg_lb_stats *sgs)
8150 struct sg_lb_stats *busiest = &sds->busiest_stat;
8152 if (sgs->group_type > busiest->group_type)
8155 if (sgs->group_type < busiest->group_type)
8158 if (sgs->avg_load <= busiest->avg_load)
8161 if (!(env->sd->flags & SD_ASYM_CPUCAPACITY))
8165 * Candidate sg has no more than one task per CPU and
8166 * has higher per-CPU capacity. Migrating tasks to less
8167 * capable CPUs may harm throughput. Maximize throughput,
8168 * power/energy consequences are not considered.
8170 if (sgs->sum_nr_running <= sgs->group_weight &&
8171 group_smaller_cpu_capacity(sds->local, sg))
8175 /* This is the busiest node in its class. */
8176 if (!(env->sd->flags & SD_ASYM_PACKING))
8179 /* No ASYM_PACKING if target CPU is already busy */
8180 if (env->idle == CPU_NOT_IDLE)
8183 * ASYM_PACKING needs to move all the work to the highest
8184 * prority CPUs in the group, therefore mark all groups
8185 * of lower priority than ourself as busy.
8187 if (sgs->sum_nr_running &&
8188 sched_asym_prefer(env->dst_cpu, sg->asym_prefer_cpu)) {
8192 /* Prefer to move from lowest priority CPU's work */
8193 if (sched_asym_prefer(sds->busiest->asym_prefer_cpu,
8194 sg->asym_prefer_cpu))
8201 #ifdef CONFIG_NUMA_BALANCING
8202 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
8204 if (sgs->sum_nr_running > sgs->nr_numa_running)
8206 if (sgs->sum_nr_running > sgs->nr_preferred_running)
8211 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
8213 if (rq->nr_running > rq->nr_numa_running)
8215 if (rq->nr_running > rq->nr_preferred_running)
8220 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
8225 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
8229 #endif /* CONFIG_NUMA_BALANCING */
8232 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
8233 * @env: The load balancing environment.
8234 * @sds: variable to hold the statistics for this sched_domain.
8236 static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
8238 struct sched_domain *child = env->sd->child;
8239 struct sched_group *sg = env->sd->groups;
8240 struct sg_lb_stats *local = &sds->local_stat;
8241 struct sg_lb_stats tmp_sgs;
8242 int load_idx, prefer_sibling = 0;
8243 bool overload = false;
8245 if (child && child->flags & SD_PREFER_SIBLING)
8248 #ifdef CONFIG_NO_HZ_COMMON
8249 if (env->idle == CPU_NEWLY_IDLE && READ_ONCE(nohz.has_blocked))
8250 env->flags |= LBF_NOHZ_STATS;
8253 load_idx = get_sd_load_idx(env->sd, env->idle);
8256 struct sg_lb_stats *sgs = &tmp_sgs;
8259 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_span(sg));
8264 if (env->idle != CPU_NEWLY_IDLE ||
8265 time_after_eq(jiffies, sg->sgc->next_update))
8266 update_group_capacity(env->sd, env->dst_cpu);
8269 update_sg_lb_stats(env, sg, load_idx, local_group, sgs,
8276 * In case the child domain prefers tasks go to siblings
8277 * first, lower the sg capacity so that we'll try
8278 * and move all the excess tasks away. We lower the capacity
8279 * of a group only if the local group has the capacity to fit
8280 * these excess tasks. The extra check prevents the case where
8281 * you always pull from the heaviest group when it is already
8282 * under-utilized (possible with a large weight task outweighs
8283 * the tasks on the system).
8285 if (prefer_sibling && sds->local &&
8286 group_has_capacity(env, local) &&
8287 (sgs->sum_nr_running > local->sum_nr_running + 1)) {
8288 sgs->group_no_capacity = 1;
8289 sgs->group_type = group_classify(sg, sgs);
8292 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
8294 sds->busiest_stat = *sgs;
8298 /* Now, start updating sd_lb_stats */
8299 sds->total_running += sgs->sum_nr_running;
8300 sds->total_load += sgs->group_load;
8301 sds->total_capacity += sgs->group_capacity;
8304 } while (sg != env->sd->groups);
8306 #ifdef CONFIG_NO_HZ_COMMON
8307 if ((env->flags & LBF_NOHZ_AGAIN) &&
8308 cpumask_subset(nohz.idle_cpus_mask, sched_domain_span(env->sd))) {
8310 WRITE_ONCE(nohz.next_blocked,
8311 jiffies + msecs_to_jiffies(LOAD_AVG_PERIOD));
8315 if (env->sd->flags & SD_NUMA)
8316 env->fbq_type = fbq_classify_group(&sds->busiest_stat);
8318 if (!env->sd->parent) {
8319 /* update overload indicator if we are at root domain */
8320 if (env->dst_rq->rd->overload != overload)
8321 env->dst_rq->rd->overload = overload;
8326 * check_asym_packing - Check to see if the group is packed into the
8329 * This is primarily intended to used at the sibling level. Some
8330 * cores like POWER7 prefer to use lower numbered SMT threads. In the
8331 * case of POWER7, it can move to lower SMT modes only when higher
8332 * threads are idle. When in lower SMT modes, the threads will
8333 * perform better since they share less core resources. Hence when we
8334 * have idle threads, we want them to be the higher ones.
8336 * This packing function is run on idle threads. It checks to see if
8337 * the busiest CPU in this domain (core in the P7 case) has a higher
8338 * CPU number than the packing function is being run on. Here we are
8339 * assuming lower CPU number will be equivalent to lower a SMT thread
8342 * Return: 1 when packing is required and a task should be moved to
8343 * this CPU. The amount of the imbalance is returned in env->imbalance.
8345 * @env: The load balancing environment.
8346 * @sds: Statistics of the sched_domain which is to be packed
8348 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
8352 if (!(env->sd->flags & SD_ASYM_PACKING))
8355 if (env->idle == CPU_NOT_IDLE)
8361 busiest_cpu = sds->busiest->asym_prefer_cpu;
8362 if (sched_asym_prefer(busiest_cpu, env->dst_cpu))
8365 env->imbalance = DIV_ROUND_CLOSEST(
8366 sds->busiest_stat.avg_load * sds->busiest_stat.group_capacity,
8367 SCHED_CAPACITY_SCALE);
8373 * fix_small_imbalance - Calculate the minor imbalance that exists
8374 * amongst the groups of a sched_domain, during
8376 * @env: The load balancing environment.
8377 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
8380 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
8382 unsigned long tmp, capa_now = 0, capa_move = 0;
8383 unsigned int imbn = 2;
8384 unsigned long scaled_busy_load_per_task;
8385 struct sg_lb_stats *local, *busiest;
8387 local = &sds->local_stat;
8388 busiest = &sds->busiest_stat;
8390 if (!local->sum_nr_running)
8391 local->load_per_task = cpu_avg_load_per_task(env->dst_cpu);
8392 else if (busiest->load_per_task > local->load_per_task)
8395 scaled_busy_load_per_task =
8396 (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
8397 busiest->group_capacity;
8399 if (busiest->avg_load + scaled_busy_load_per_task >=
8400 local->avg_load + (scaled_busy_load_per_task * imbn)) {
8401 env->imbalance = busiest->load_per_task;
8406 * OK, we don't have enough imbalance to justify moving tasks,
8407 * however we may be able to increase total CPU capacity used by
8411 capa_now += busiest->group_capacity *
8412 min(busiest->load_per_task, busiest->avg_load);
8413 capa_now += local->group_capacity *
8414 min(local->load_per_task, local->avg_load);
8415 capa_now /= SCHED_CAPACITY_SCALE;
8417 /* Amount of load we'd subtract */
8418 if (busiest->avg_load > scaled_busy_load_per_task) {
8419 capa_move += busiest->group_capacity *
8420 min(busiest->load_per_task,
8421 busiest->avg_load - scaled_busy_load_per_task);
8424 /* Amount of load we'd add */
8425 if (busiest->avg_load * busiest->group_capacity <
8426 busiest->load_per_task * SCHED_CAPACITY_SCALE) {
8427 tmp = (busiest->avg_load * busiest->group_capacity) /
8428 local->group_capacity;
8430 tmp = (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
8431 local->group_capacity;
8433 capa_move += local->group_capacity *
8434 min(local->load_per_task, local->avg_load + tmp);
8435 capa_move /= SCHED_CAPACITY_SCALE;
8437 /* Move if we gain throughput */
8438 if (capa_move > capa_now)
8439 env->imbalance = busiest->load_per_task;
8443 * calculate_imbalance - Calculate the amount of imbalance present within the
8444 * groups of a given sched_domain during load balance.
8445 * @env: load balance environment
8446 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
8448 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
8450 unsigned long max_pull, load_above_capacity = ~0UL;
8451 struct sg_lb_stats *local, *busiest;
8453 local = &sds->local_stat;
8454 busiest = &sds->busiest_stat;
8456 if (busiest->group_type == group_imbalanced) {
8458 * In the group_imb case we cannot rely on group-wide averages
8459 * to ensure CPU-load equilibrium, look at wider averages. XXX
8461 busiest->load_per_task =
8462 min(busiest->load_per_task, sds->avg_load);
8466 * Avg load of busiest sg can be less and avg load of local sg can
8467 * be greater than avg load across all sgs of sd because avg load
8468 * factors in sg capacity and sgs with smaller group_type are
8469 * skipped when updating the busiest sg:
8471 if (busiest->avg_load <= sds->avg_load ||
8472 local->avg_load >= sds->avg_load) {
8474 return fix_small_imbalance(env, sds);
8478 * If there aren't any idle CPUs, avoid creating some.
8480 if (busiest->group_type == group_overloaded &&
8481 local->group_type == group_overloaded) {
8482 load_above_capacity = busiest->sum_nr_running * SCHED_CAPACITY_SCALE;
8483 if (load_above_capacity > busiest->group_capacity) {
8484 load_above_capacity -= busiest->group_capacity;
8485 load_above_capacity *= scale_load_down(NICE_0_LOAD);
8486 load_above_capacity /= busiest->group_capacity;
8488 load_above_capacity = ~0UL;
8492 * We're trying to get all the CPUs to the average_load, so we don't
8493 * want to push ourselves above the average load, nor do we wish to
8494 * reduce the max loaded CPU below the average load. At the same time,
8495 * we also don't want to reduce the group load below the group
8496 * capacity. Thus we look for the minimum possible imbalance.
8498 max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity);
8500 /* How much load to actually move to equalise the imbalance */
8501 env->imbalance = min(
8502 max_pull * busiest->group_capacity,
8503 (sds->avg_load - local->avg_load) * local->group_capacity
8504 ) / SCHED_CAPACITY_SCALE;
8507 * if *imbalance is less than the average load per runnable task
8508 * there is no guarantee that any tasks will be moved so we'll have
8509 * a think about bumping its value to force at least one task to be
8512 if (env->imbalance < busiest->load_per_task)
8513 return fix_small_imbalance(env, sds);
8516 /******* find_busiest_group() helpers end here *********************/
8519 * find_busiest_group - Returns the busiest group within the sched_domain
8520 * if there is an imbalance.
8522 * Also calculates the amount of weighted load which should be moved
8523 * to restore balance.
8525 * @env: The load balancing environment.
8527 * Return: - The busiest group if imbalance exists.
8529 static struct sched_group *find_busiest_group(struct lb_env *env)
8531 struct sg_lb_stats *local, *busiest;
8532 struct sd_lb_stats sds;
8534 init_sd_lb_stats(&sds);
8537 * Compute the various statistics relavent for load balancing at
8540 update_sd_lb_stats(env, &sds);
8541 local = &sds.local_stat;
8542 busiest = &sds.busiest_stat;
8544 /* ASYM feature bypasses nice load balance check */
8545 if (check_asym_packing(env, &sds))
8548 /* There is no busy sibling group to pull tasks from */
8549 if (!sds.busiest || busiest->sum_nr_running == 0)
8552 /* XXX broken for overlapping NUMA groups */
8553 sds.avg_load = (SCHED_CAPACITY_SCALE * sds.total_load)
8554 / sds.total_capacity;
8557 * If the busiest group is imbalanced the below checks don't
8558 * work because they assume all things are equal, which typically
8559 * isn't true due to cpus_allowed constraints and the like.
8561 if (busiest->group_type == group_imbalanced)
8565 * When dst_cpu is idle, prevent SMP nice and/or asymmetric group
8566 * capacities from resulting in underutilization due to avg_load.
8568 if (env->idle != CPU_NOT_IDLE && group_has_capacity(env, local) &&
8569 busiest->group_no_capacity)
8573 * If the local group is busier than the selected busiest group
8574 * don't try and pull any tasks.
8576 if (local->avg_load >= busiest->avg_load)
8580 * Don't pull any tasks if this group is already above the domain
8583 if (local->avg_load >= sds.avg_load)
8586 if (env->idle == CPU_IDLE) {
8588 * This CPU is idle. If the busiest group is not overloaded
8589 * and there is no imbalance between this and busiest group
8590 * wrt idle CPUs, it is balanced. The imbalance becomes
8591 * significant if the diff is greater than 1 otherwise we
8592 * might end up to just move the imbalance on another group
8594 if ((busiest->group_type != group_overloaded) &&
8595 (local->idle_cpus <= (busiest->idle_cpus + 1)))
8599 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
8600 * imbalance_pct to be conservative.
8602 if (100 * busiest->avg_load <=
8603 env->sd->imbalance_pct * local->avg_load)
8608 /* Looks like there is an imbalance. Compute it */
8609 calculate_imbalance(env, &sds);
8618 * find_busiest_queue - find the busiest runqueue among the CPUs in the group.
8620 static struct rq *find_busiest_queue(struct lb_env *env,
8621 struct sched_group *group)
8623 struct rq *busiest = NULL, *rq;
8624 unsigned long busiest_load = 0, busiest_capacity = 1;
8627 for_each_cpu_and(i, sched_group_span(group), env->cpus) {
8628 unsigned long capacity, wl;
8632 rt = fbq_classify_rq(rq);
8635 * We classify groups/runqueues into three groups:
8636 * - regular: there are !numa tasks
8637 * - remote: there are numa tasks that run on the 'wrong' node
8638 * - all: there is no distinction
8640 * In order to avoid migrating ideally placed numa tasks,
8641 * ignore those when there's better options.
8643 * If we ignore the actual busiest queue to migrate another
8644 * task, the next balance pass can still reduce the busiest
8645 * queue by moving tasks around inside the node.
8647 * If we cannot move enough load due to this classification
8648 * the next pass will adjust the group classification and
8649 * allow migration of more tasks.
8651 * Both cases only affect the total convergence complexity.
8653 if (rt > env->fbq_type)
8656 capacity = capacity_of(i);
8658 wl = weighted_cpuload(rq);
8661 * When comparing with imbalance, use weighted_cpuload()
8662 * which is not scaled with the CPU capacity.
8665 if (rq->nr_running == 1 && wl > env->imbalance &&
8666 !check_cpu_capacity(rq, env->sd))
8670 * For the load comparisons with the other CPU's, consider
8671 * the weighted_cpuload() scaled with the CPU capacity, so
8672 * that the load can be moved away from the CPU that is
8673 * potentially running at a lower capacity.
8675 * Thus we're looking for max(wl_i / capacity_i), crosswise
8676 * multiplication to rid ourselves of the division works out
8677 * to: wl_i * capacity_j > wl_j * capacity_i; where j is
8678 * our previous maximum.
8680 if (wl * busiest_capacity > busiest_load * capacity) {
8682 busiest_capacity = capacity;
8691 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
8692 * so long as it is large enough.
8694 #define MAX_PINNED_INTERVAL 512
8696 static int need_active_balance(struct lb_env *env)
8698 struct sched_domain *sd = env->sd;
8700 if (env->idle == CPU_NEWLY_IDLE) {
8703 * ASYM_PACKING needs to force migrate tasks from busy but
8704 * lower priority CPUs in order to pack all tasks in the
8705 * highest priority CPUs.
8707 if ((sd->flags & SD_ASYM_PACKING) &&
8708 sched_asym_prefer(env->dst_cpu, env->src_cpu))
8713 * The dst_cpu is idle and the src_cpu CPU has only 1 CFS task.
8714 * It's worth migrating the task if the src_cpu's capacity is reduced
8715 * because of other sched_class or IRQs if more capacity stays
8716 * available on dst_cpu.
8718 if ((env->idle != CPU_NOT_IDLE) &&
8719 (env->src_rq->cfs.h_nr_running == 1)) {
8720 if ((check_cpu_capacity(env->src_rq, sd)) &&
8721 (capacity_of(env->src_cpu)*sd->imbalance_pct < capacity_of(env->dst_cpu)*100))
8725 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
8728 static int active_load_balance_cpu_stop(void *data);
8730 static int should_we_balance(struct lb_env *env)
8732 struct sched_group *sg = env->sd->groups;
8733 int cpu, balance_cpu = -1;
8736 * Ensure the balancing environment is consistent; can happen
8737 * when the softirq triggers 'during' hotplug.
8739 if (!cpumask_test_cpu(env->dst_cpu, env->cpus))
8743 * In the newly idle case, we will allow all the CPUs
8744 * to do the newly idle load balance.
8746 if (env->idle == CPU_NEWLY_IDLE)
8749 /* Try to find first idle CPU */
8750 for_each_cpu_and(cpu, group_balance_mask(sg), env->cpus) {
8758 if (balance_cpu == -1)
8759 balance_cpu = group_balance_cpu(sg);
8762 * First idle CPU or the first CPU(busiest) in this sched group
8763 * is eligible for doing load balancing at this and above domains.
8765 return balance_cpu == env->dst_cpu;
8769 * Check this_cpu to ensure it is balanced within domain. Attempt to move
8770 * tasks if there is an imbalance.
8772 static int load_balance(int this_cpu, struct rq *this_rq,
8773 struct sched_domain *sd, enum cpu_idle_type idle,
8774 int *continue_balancing)
8776 int ld_moved, cur_ld_moved, active_balance = 0;
8777 struct sched_domain *sd_parent = sd->parent;
8778 struct sched_group *group;
8781 struct cpumask *cpus = this_cpu_cpumask_var_ptr(load_balance_mask);
8783 struct lb_env env = {
8785 .dst_cpu = this_cpu,
8787 .dst_grpmask = sched_group_span(sd->groups),
8789 .loop_break = sched_nr_migrate_break,
8792 .tasks = LIST_HEAD_INIT(env.tasks),
8795 cpumask_and(cpus, sched_domain_span(sd), cpu_active_mask);
8797 schedstat_inc(sd->lb_count[idle]);
8800 if (!should_we_balance(&env)) {
8801 *continue_balancing = 0;
8805 group = find_busiest_group(&env);
8807 schedstat_inc(sd->lb_nobusyg[idle]);
8811 busiest = find_busiest_queue(&env, group);
8813 schedstat_inc(sd->lb_nobusyq[idle]);
8817 BUG_ON(busiest == env.dst_rq);
8819 schedstat_add(sd->lb_imbalance[idle], env.imbalance);
8821 env.src_cpu = busiest->cpu;
8822 env.src_rq = busiest;
8825 if (busiest->nr_running > 1) {
8827 * Attempt to move tasks. If find_busiest_group has found
8828 * an imbalance but busiest->nr_running <= 1, the group is
8829 * still unbalanced. ld_moved simply stays zero, so it is
8830 * correctly treated as an imbalance.
8832 env.flags |= LBF_ALL_PINNED;
8833 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
8836 rq_lock_irqsave(busiest, &rf);
8837 update_rq_clock(busiest);
8840 * cur_ld_moved - load moved in current iteration
8841 * ld_moved - cumulative load moved across iterations
8843 cur_ld_moved = detach_tasks(&env);
8846 * We've detached some tasks from busiest_rq. Every
8847 * task is masked "TASK_ON_RQ_MIGRATING", so we can safely
8848 * unlock busiest->lock, and we are able to be sure
8849 * that nobody can manipulate the tasks in parallel.
8850 * See task_rq_lock() family for the details.
8853 rq_unlock(busiest, &rf);
8857 ld_moved += cur_ld_moved;
8860 local_irq_restore(rf.flags);
8862 if (env.flags & LBF_NEED_BREAK) {
8863 env.flags &= ~LBF_NEED_BREAK;
8868 * Revisit (affine) tasks on src_cpu that couldn't be moved to
8869 * us and move them to an alternate dst_cpu in our sched_group
8870 * where they can run. The upper limit on how many times we
8871 * iterate on same src_cpu is dependent on number of CPUs in our
8874 * This changes load balance semantics a bit on who can move
8875 * load to a given_cpu. In addition to the given_cpu itself
8876 * (or a ilb_cpu acting on its behalf where given_cpu is
8877 * nohz-idle), we now have balance_cpu in a position to move
8878 * load to given_cpu. In rare situations, this may cause
8879 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
8880 * _independently_ and at _same_ time to move some load to
8881 * given_cpu) causing exceess load to be moved to given_cpu.
8882 * This however should not happen so much in practice and
8883 * moreover subsequent load balance cycles should correct the
8884 * excess load moved.
8886 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
8888 /* Prevent to re-select dst_cpu via env's CPUs */
8889 cpumask_clear_cpu(env.dst_cpu, env.cpus);
8891 env.dst_rq = cpu_rq(env.new_dst_cpu);
8892 env.dst_cpu = env.new_dst_cpu;
8893 env.flags &= ~LBF_DST_PINNED;
8895 env.loop_break = sched_nr_migrate_break;
8898 * Go back to "more_balance" rather than "redo" since we
8899 * need to continue with same src_cpu.
8905 * We failed to reach balance because of affinity.
8908 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
8910 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0)
8911 *group_imbalance = 1;
8914 /* All tasks on this runqueue were pinned by CPU affinity */
8915 if (unlikely(env.flags & LBF_ALL_PINNED)) {
8916 cpumask_clear_cpu(cpu_of(busiest), cpus);
8918 * Attempting to continue load balancing at the current
8919 * sched_domain level only makes sense if there are
8920 * active CPUs remaining as possible busiest CPUs to
8921 * pull load from which are not contained within the
8922 * destination group that is receiving any migrated
8925 if (!cpumask_subset(cpus, env.dst_grpmask)) {
8927 env.loop_break = sched_nr_migrate_break;
8930 goto out_all_pinned;
8935 schedstat_inc(sd->lb_failed[idle]);
8937 * Increment the failure counter only on periodic balance.
8938 * We do not want newidle balance, which can be very
8939 * frequent, pollute the failure counter causing
8940 * excessive cache_hot migrations and active balances.
8942 if (idle != CPU_NEWLY_IDLE)
8943 sd->nr_balance_failed++;
8945 if (need_active_balance(&env)) {
8946 unsigned long flags;
8948 raw_spin_lock_irqsave(&busiest->lock, flags);
8951 * Don't kick the active_load_balance_cpu_stop,
8952 * if the curr task on busiest CPU can't be
8953 * moved to this_cpu:
8955 if (!cpumask_test_cpu(this_cpu, &busiest->curr->cpus_allowed)) {
8956 raw_spin_unlock_irqrestore(&busiest->lock,
8958 env.flags |= LBF_ALL_PINNED;
8959 goto out_one_pinned;
8963 * ->active_balance synchronizes accesses to
8964 * ->active_balance_work. Once set, it's cleared
8965 * only after active load balance is finished.
8967 if (!busiest->active_balance) {
8968 busiest->active_balance = 1;
8969 busiest->push_cpu = this_cpu;
8972 raw_spin_unlock_irqrestore(&busiest->lock, flags);
8974 if (active_balance) {
8975 stop_one_cpu_nowait(cpu_of(busiest),
8976 active_load_balance_cpu_stop, busiest,
8977 &busiest->active_balance_work);
8980 /* We've kicked active balancing, force task migration. */
8981 sd->nr_balance_failed = sd->cache_nice_tries+1;
8984 sd->nr_balance_failed = 0;
8986 if (likely(!active_balance)) {
8987 /* We were unbalanced, so reset the balancing interval */
8988 sd->balance_interval = sd->min_interval;
8991 * If we've begun active balancing, start to back off. This
8992 * case may not be covered by the all_pinned logic if there
8993 * is only 1 task on the busy runqueue (because we don't call
8996 if (sd->balance_interval < sd->max_interval)
8997 sd->balance_interval *= 2;
9004 * We reach balance although we may have faced some affinity
9005 * constraints. Clear the imbalance flag if it was set.
9008 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
9010 if (*group_imbalance)
9011 *group_imbalance = 0;
9016 * We reach balance because all tasks are pinned at this level so
9017 * we can't migrate them. Let the imbalance flag set so parent level
9018 * can try to migrate them.
9020 schedstat_inc(sd->lb_balanced[idle]);
9022 sd->nr_balance_failed = 0;
9025 /* tune up the balancing interval */
9026 if (((env.flags & LBF_ALL_PINNED) &&
9027 sd->balance_interval < MAX_PINNED_INTERVAL) ||
9028 (sd->balance_interval < sd->max_interval))
9029 sd->balance_interval *= 2;
9036 static inline unsigned long
9037 get_sd_balance_interval(struct sched_domain *sd, int cpu_busy)
9039 unsigned long interval = sd->balance_interval;
9042 interval *= sd->busy_factor;
9044 /* scale ms to jiffies */
9045 interval = msecs_to_jiffies(interval);
9046 interval = clamp(interval, 1UL, max_load_balance_interval);
9052 update_next_balance(struct sched_domain *sd, unsigned long *next_balance)
9054 unsigned long interval, next;
9056 /* used by idle balance, so cpu_busy = 0 */
9057 interval = get_sd_balance_interval(sd, 0);
9058 next = sd->last_balance + interval;
9060 if (time_after(*next_balance, next))
9061 *next_balance = next;
9065 * active_load_balance_cpu_stop is run by the CPU stopper. It pushes
9066 * running tasks off the busiest CPU onto idle CPUs. It requires at
9067 * least 1 task to be running on each physical CPU where possible, and
9068 * avoids physical / logical imbalances.
9070 static int active_load_balance_cpu_stop(void *data)
9072 struct rq *busiest_rq = data;
9073 int busiest_cpu = cpu_of(busiest_rq);
9074 int target_cpu = busiest_rq->push_cpu;
9075 struct rq *target_rq = cpu_rq(target_cpu);
9076 struct sched_domain *sd;
9077 struct task_struct *p = NULL;
9080 rq_lock_irq(busiest_rq, &rf);
9082 * Between queueing the stop-work and running it is a hole in which
9083 * CPUs can become inactive. We should not move tasks from or to
9086 if (!cpu_active(busiest_cpu) || !cpu_active(target_cpu))
9089 /* Make sure the requested CPU hasn't gone down in the meantime: */
9090 if (unlikely(busiest_cpu != smp_processor_id() ||
9091 !busiest_rq->active_balance))
9094 /* Is there any task to move? */
9095 if (busiest_rq->nr_running <= 1)
9099 * This condition is "impossible", if it occurs
9100 * we need to fix it. Originally reported by
9101 * Bjorn Helgaas on a 128-CPU setup.
9103 BUG_ON(busiest_rq == target_rq);
9105 /* Search for an sd spanning us and the target CPU. */
9107 for_each_domain(target_cpu, sd) {
9108 if ((sd->flags & SD_LOAD_BALANCE) &&
9109 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
9114 struct lb_env env = {
9116 .dst_cpu = target_cpu,
9117 .dst_rq = target_rq,
9118 .src_cpu = busiest_rq->cpu,
9119 .src_rq = busiest_rq,
9122 * can_migrate_task() doesn't need to compute new_dst_cpu
9123 * for active balancing. Since we have CPU_IDLE, but no
9124 * @dst_grpmask we need to make that test go away with lying
9127 .flags = LBF_DST_PINNED,
9130 schedstat_inc(sd->alb_count);
9131 update_rq_clock(busiest_rq);
9133 p = detach_one_task(&env);
9135 schedstat_inc(sd->alb_pushed);
9136 /* Active balancing done, reset the failure counter. */
9137 sd->nr_balance_failed = 0;
9139 schedstat_inc(sd->alb_failed);
9144 busiest_rq->active_balance = 0;
9145 rq_unlock(busiest_rq, &rf);
9148 attach_one_task(target_rq, p);
9155 static DEFINE_SPINLOCK(balancing);
9158 * Scale the max load_balance interval with the number of CPUs in the system.
9159 * This trades load-balance latency on larger machines for less cross talk.
9161 void update_max_interval(void)
9163 max_load_balance_interval = HZ*num_online_cpus()/10;
9167 * It checks each scheduling domain to see if it is due to be balanced,
9168 * and initiates a balancing operation if so.
9170 * Balancing parameters are set up in init_sched_domains.
9172 static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle)
9174 int continue_balancing = 1;
9176 unsigned long interval;
9177 struct sched_domain *sd;
9178 /* Earliest time when we have to do rebalance again */
9179 unsigned long next_balance = jiffies + 60*HZ;
9180 int update_next_balance = 0;
9181 int need_serialize, need_decay = 0;
9185 for_each_domain(cpu, sd) {
9187 * Decay the newidle max times here because this is a regular
9188 * visit to all the domains. Decay ~1% per second.
9190 if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
9191 sd->max_newidle_lb_cost =
9192 (sd->max_newidle_lb_cost * 253) / 256;
9193 sd->next_decay_max_lb_cost = jiffies + HZ;
9196 max_cost += sd->max_newidle_lb_cost;
9198 if (!(sd->flags & SD_LOAD_BALANCE))
9202 * Stop the load balance at this level. There is another
9203 * CPU in our sched group which is doing load balancing more
9206 if (!continue_balancing) {
9212 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
9214 need_serialize = sd->flags & SD_SERIALIZE;
9215 if (need_serialize) {
9216 if (!spin_trylock(&balancing))
9220 if (time_after_eq(jiffies, sd->last_balance + interval)) {
9221 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
9223 * The LBF_DST_PINNED logic could have changed
9224 * env->dst_cpu, so we can't know our idle
9225 * state even if we migrated tasks. Update it.
9227 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
9229 sd->last_balance = jiffies;
9230 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
9233 spin_unlock(&balancing);
9235 if (time_after(next_balance, sd->last_balance + interval)) {
9236 next_balance = sd->last_balance + interval;
9237 update_next_balance = 1;
9242 * Ensure the rq-wide value also decays but keep it at a
9243 * reasonable floor to avoid funnies with rq->avg_idle.
9245 rq->max_idle_balance_cost =
9246 max((u64)sysctl_sched_migration_cost, max_cost);
9251 * next_balance will be updated only when there is a need.
9252 * When the cpu is attached to null domain for ex, it will not be
9255 if (likely(update_next_balance)) {
9256 rq->next_balance = next_balance;
9258 #ifdef CONFIG_NO_HZ_COMMON
9260 * If this CPU has been elected to perform the nohz idle
9261 * balance. Other idle CPUs have already rebalanced with
9262 * nohz_idle_balance() and nohz.next_balance has been
9263 * updated accordingly. This CPU is now running the idle load
9264 * balance for itself and we need to update the
9265 * nohz.next_balance accordingly.
9267 if ((idle == CPU_IDLE) && time_after(nohz.next_balance, rq->next_balance))
9268 nohz.next_balance = rq->next_balance;
9273 static inline int on_null_domain(struct rq *rq)
9275 return unlikely(!rcu_dereference_sched(rq->sd));
9278 #ifdef CONFIG_NO_HZ_COMMON
9280 * idle load balancing details
9281 * - When one of the busy CPUs notice that there may be an idle rebalancing
9282 * needed, they will kick the idle load balancer, which then does idle
9283 * load balancing for all the idle CPUs.
9286 static inline int find_new_ilb(void)
9288 int ilb = cpumask_first(nohz.idle_cpus_mask);
9290 if (ilb < nr_cpu_ids && idle_cpu(ilb))
9297 * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
9298 * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
9299 * CPU (if there is one).
9301 static void kick_ilb(unsigned int flags)
9305 nohz.next_balance++;
9307 ilb_cpu = find_new_ilb();
9309 if (ilb_cpu >= nr_cpu_ids)
9312 flags = atomic_fetch_or(flags, nohz_flags(ilb_cpu));
9313 if (flags & NOHZ_KICK_MASK)
9317 * Use smp_send_reschedule() instead of resched_cpu().
9318 * This way we generate a sched IPI on the target CPU which
9319 * is idle. And the softirq performing nohz idle load balance
9320 * will be run before returning from the IPI.
9322 smp_send_reschedule(ilb_cpu);
9326 * Current heuristic for kicking the idle load balancer in the presence
9327 * of an idle cpu in the system.
9328 * - This rq has more than one task.
9329 * - This rq has at least one CFS task and the capacity of the CPU is
9330 * significantly reduced because of RT tasks or IRQs.
9331 * - At parent of LLC scheduler domain level, this cpu's scheduler group has
9332 * multiple busy cpu.
9333 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
9334 * domain span are idle.
9336 static void nohz_balancer_kick(struct rq *rq)
9338 unsigned long now = jiffies;
9339 struct sched_domain_shared *sds;
9340 struct sched_domain *sd;
9341 int nr_busy, i, cpu = rq->cpu;
9342 unsigned int flags = 0;
9344 if (unlikely(rq->idle_balance))
9348 * We may be recently in ticked or tickless idle mode. At the first
9349 * busy tick after returning from idle, we will update the busy stats.
9351 nohz_balance_exit_idle(rq);
9354 * None are in tickless mode and hence no need for NOHZ idle load
9357 if (likely(!atomic_read(&nohz.nr_cpus)))
9360 if (READ_ONCE(nohz.has_blocked) &&
9361 time_after(now, READ_ONCE(nohz.next_blocked)))
9362 flags = NOHZ_STATS_KICK;
9364 if (time_before(now, nohz.next_balance))
9367 if (rq->nr_running >= 2) {
9368 flags = NOHZ_KICK_MASK;
9373 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
9376 * XXX: write a coherent comment on why we do this.
9377 * See also: http://lkml.kernel.org/r/20111202010832.602203411@sbsiddha-desk.sc.intel.com
9379 nr_busy = atomic_read(&sds->nr_busy_cpus);
9381 flags = NOHZ_KICK_MASK;
9387 sd = rcu_dereference(rq->sd);
9389 if ((rq->cfs.h_nr_running >= 1) &&
9390 check_cpu_capacity(rq, sd)) {
9391 flags = NOHZ_KICK_MASK;
9396 sd = rcu_dereference(per_cpu(sd_asym, cpu));
9398 for_each_cpu(i, sched_domain_span(sd)) {
9400 !cpumask_test_cpu(i, nohz.idle_cpus_mask))
9403 if (sched_asym_prefer(i, cpu)) {
9404 flags = NOHZ_KICK_MASK;
9416 static void set_cpu_sd_state_busy(int cpu)
9418 struct sched_domain *sd;
9421 sd = rcu_dereference(per_cpu(sd_llc, cpu));
9423 if (!sd || !sd->nohz_idle)
9427 atomic_inc(&sd->shared->nr_busy_cpus);
9432 void nohz_balance_exit_idle(struct rq *rq)
9434 SCHED_WARN_ON(rq != this_rq());
9436 if (likely(!rq->nohz_tick_stopped))
9439 rq->nohz_tick_stopped = 0;
9440 cpumask_clear_cpu(rq->cpu, nohz.idle_cpus_mask);
9441 atomic_dec(&nohz.nr_cpus);
9443 set_cpu_sd_state_busy(rq->cpu);
9446 static void set_cpu_sd_state_idle(int cpu)
9448 struct sched_domain *sd;
9451 sd = rcu_dereference(per_cpu(sd_llc, cpu));
9453 if (!sd || sd->nohz_idle)
9457 atomic_dec(&sd->shared->nr_busy_cpus);
9463 * This routine will record that the CPU is going idle with tick stopped.
9464 * This info will be used in performing idle load balancing in the future.
9466 void nohz_balance_enter_idle(int cpu)
9468 struct rq *rq = cpu_rq(cpu);
9470 SCHED_WARN_ON(cpu != smp_processor_id());
9472 /* If this CPU is going down, then nothing needs to be done: */
9473 if (!cpu_active(cpu))
9476 /* Spare idle load balancing on CPUs that don't want to be disturbed: */
9477 if (!housekeeping_cpu(cpu, HK_FLAG_SCHED))
9481 * Can be set safely without rq->lock held
9482 * If a clear happens, it will have evaluated last additions because
9483 * rq->lock is held during the check and the clear
9485 rq->has_blocked_load = 1;
9488 * The tick is still stopped but load could have been added in the
9489 * meantime. We set the nohz.has_blocked flag to trig a check of the
9490 * *_avg. The CPU is already part of nohz.idle_cpus_mask so the clear
9491 * of nohz.has_blocked can only happen after checking the new load
9493 if (rq->nohz_tick_stopped)
9496 /* If we're a completely isolated CPU, we don't play: */
9497 if (on_null_domain(rq))
9500 rq->nohz_tick_stopped = 1;
9502 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
9503 atomic_inc(&nohz.nr_cpus);
9506 * Ensures that if nohz_idle_balance() fails to observe our
9507 * @idle_cpus_mask store, it must observe the @has_blocked
9510 smp_mb__after_atomic();
9512 set_cpu_sd_state_idle(cpu);
9516 * Each time a cpu enter idle, we assume that it has blocked load and
9517 * enable the periodic update of the load of idle cpus
9519 WRITE_ONCE(nohz.has_blocked, 1);
9523 * Internal function that runs load balance for all idle cpus. The load balance
9524 * can be a simple update of blocked load or a complete load balance with
9525 * tasks movement depending of flags.
9526 * The function returns false if the loop has stopped before running
9527 * through all idle CPUs.
9529 static bool _nohz_idle_balance(struct rq *this_rq, unsigned int flags,
9530 enum cpu_idle_type idle)
9532 /* Earliest time when we have to do rebalance again */
9533 unsigned long now = jiffies;
9534 unsigned long next_balance = now + 60*HZ;
9535 bool has_blocked_load = false;
9536 int update_next_balance = 0;
9537 int this_cpu = this_rq->cpu;
9542 SCHED_WARN_ON((flags & NOHZ_KICK_MASK) == NOHZ_BALANCE_KICK);
9545 * We assume there will be no idle load after this update and clear
9546 * the has_blocked flag. If a cpu enters idle in the mean time, it will
9547 * set the has_blocked flag and trig another update of idle load.
9548 * Because a cpu that becomes idle, is added to idle_cpus_mask before
9549 * setting the flag, we are sure to not clear the state and not
9550 * check the load of an idle cpu.
9552 WRITE_ONCE(nohz.has_blocked, 0);
9555 * Ensures that if we miss the CPU, we must see the has_blocked
9556 * store from nohz_balance_enter_idle().
9560 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
9561 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
9565 * If this CPU gets work to do, stop the load balancing
9566 * work being done for other CPUs. Next load
9567 * balancing owner will pick it up.
9569 if (need_resched()) {
9570 has_blocked_load = true;
9574 rq = cpu_rq(balance_cpu);
9576 has_blocked_load |= update_nohz_stats(rq, true);
9579 * If time for next balance is due,
9582 if (time_after_eq(jiffies, rq->next_balance)) {
9585 rq_lock_irqsave(rq, &rf);
9586 update_rq_clock(rq);
9587 cpu_load_update_idle(rq);
9588 rq_unlock_irqrestore(rq, &rf);
9590 if (flags & NOHZ_BALANCE_KICK)
9591 rebalance_domains(rq, CPU_IDLE);
9594 if (time_after(next_balance, rq->next_balance)) {
9595 next_balance = rq->next_balance;
9596 update_next_balance = 1;
9600 /* Newly idle CPU doesn't need an update */
9601 if (idle != CPU_NEWLY_IDLE) {
9602 update_blocked_averages(this_cpu);
9603 has_blocked_load |= this_rq->has_blocked_load;
9606 if (flags & NOHZ_BALANCE_KICK)
9607 rebalance_domains(this_rq, CPU_IDLE);
9609 WRITE_ONCE(nohz.next_blocked,
9610 now + msecs_to_jiffies(LOAD_AVG_PERIOD));
9612 /* The full idle balance loop has been done */
9616 /* There is still blocked load, enable periodic update */
9617 if (has_blocked_load)
9618 WRITE_ONCE(nohz.has_blocked, 1);
9621 * next_balance will be updated only when there is a need.
9622 * When the CPU is attached to null domain for ex, it will not be
9625 if (likely(update_next_balance))
9626 nohz.next_balance = next_balance;
9632 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
9633 * rebalancing for all the cpus for whom scheduler ticks are stopped.
9635 static bool nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
9637 int this_cpu = this_rq->cpu;
9640 if (!(atomic_read(nohz_flags(this_cpu)) & NOHZ_KICK_MASK))
9643 if (idle != CPU_IDLE) {
9644 atomic_andnot(NOHZ_KICK_MASK, nohz_flags(this_cpu));
9649 * barrier, pairs with nohz_balance_enter_idle(), ensures ...
9651 flags = atomic_fetch_andnot(NOHZ_KICK_MASK, nohz_flags(this_cpu));
9652 if (!(flags & NOHZ_KICK_MASK))
9655 _nohz_idle_balance(this_rq, flags, idle);
9660 static void nohz_newidle_balance(struct rq *this_rq)
9662 int this_cpu = this_rq->cpu;
9665 * This CPU doesn't want to be disturbed by scheduler
9668 if (!housekeeping_cpu(this_cpu, HK_FLAG_SCHED))
9671 /* Will wake up very soon. No time for doing anything else*/
9672 if (this_rq->avg_idle < sysctl_sched_migration_cost)
9675 /* Don't need to update blocked load of idle CPUs*/
9676 if (!READ_ONCE(nohz.has_blocked) ||
9677 time_before(jiffies, READ_ONCE(nohz.next_blocked)))
9680 raw_spin_unlock(&this_rq->lock);
9682 * This CPU is going to be idle and blocked load of idle CPUs
9683 * need to be updated. Run the ilb locally as it is a good
9684 * candidate for ilb instead of waking up another idle CPU.
9685 * Kick an normal ilb if we failed to do the update.
9687 if (!_nohz_idle_balance(this_rq, NOHZ_STATS_KICK, CPU_NEWLY_IDLE))
9688 kick_ilb(NOHZ_STATS_KICK);
9689 raw_spin_lock(&this_rq->lock);
9692 #else /* !CONFIG_NO_HZ_COMMON */
9693 static inline void nohz_balancer_kick(struct rq *rq) { }
9695 static inline bool nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
9700 static inline void nohz_newidle_balance(struct rq *this_rq) { }
9701 #endif /* CONFIG_NO_HZ_COMMON */
9704 * idle_balance is called by schedule() if this_cpu is about to become
9705 * idle. Attempts to pull tasks from other CPUs.
9707 static int idle_balance(struct rq *this_rq, struct rq_flags *rf)
9709 unsigned long next_balance = jiffies + HZ;
9710 int this_cpu = this_rq->cpu;
9711 struct sched_domain *sd;
9712 int pulled_task = 0;
9716 * We must set idle_stamp _before_ calling idle_balance(), such that we
9717 * measure the duration of idle_balance() as idle time.
9719 this_rq->idle_stamp = rq_clock(this_rq);
9722 * Do not pull tasks towards !active CPUs...
9724 if (!cpu_active(this_cpu))
9728 * This is OK, because current is on_cpu, which avoids it being picked
9729 * for load-balance and preemption/IRQs are still disabled avoiding
9730 * further scheduler activity on it and we're being very careful to
9731 * re-start the picking loop.
9733 rq_unpin_lock(this_rq, rf);
9735 if (this_rq->avg_idle < sysctl_sched_migration_cost ||
9736 !this_rq->rd->overload) {
9739 sd = rcu_dereference_check_sched_domain(this_rq->sd);
9741 update_next_balance(sd, &next_balance);
9744 nohz_newidle_balance(this_rq);
9749 raw_spin_unlock(&this_rq->lock);
9751 update_blocked_averages(this_cpu);
9753 for_each_domain(this_cpu, sd) {
9754 int continue_balancing = 1;
9755 u64 t0, domain_cost;
9757 if (!(sd->flags & SD_LOAD_BALANCE))
9760 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost) {
9761 update_next_balance(sd, &next_balance);
9765 if (sd->flags & SD_BALANCE_NEWIDLE) {
9766 t0 = sched_clock_cpu(this_cpu);
9768 pulled_task = load_balance(this_cpu, this_rq,
9770 &continue_balancing);
9772 domain_cost = sched_clock_cpu(this_cpu) - t0;
9773 if (domain_cost > sd->max_newidle_lb_cost)
9774 sd->max_newidle_lb_cost = domain_cost;
9776 curr_cost += domain_cost;
9779 update_next_balance(sd, &next_balance);
9782 * Stop searching for tasks to pull if there are
9783 * now runnable tasks on this rq.
9785 if (pulled_task || this_rq->nr_running > 0)
9790 raw_spin_lock(&this_rq->lock);
9792 if (curr_cost > this_rq->max_idle_balance_cost)
9793 this_rq->max_idle_balance_cost = curr_cost;
9797 * While browsing the domains, we released the rq lock, a task could
9798 * have been enqueued in the meantime. Since we're not going idle,
9799 * pretend we pulled a task.
9801 if (this_rq->cfs.h_nr_running && !pulled_task)
9804 /* Move the next balance forward */
9805 if (time_after(this_rq->next_balance, next_balance))
9806 this_rq->next_balance = next_balance;
9808 /* Is there a task of a high priority class? */
9809 if (this_rq->nr_running != this_rq->cfs.h_nr_running)
9813 this_rq->idle_stamp = 0;
9815 rq_repin_lock(this_rq, rf);
9821 * run_rebalance_domains is triggered when needed from the scheduler tick.
9822 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
9824 static __latent_entropy void run_rebalance_domains(struct softirq_action *h)
9826 struct rq *this_rq = this_rq();
9827 enum cpu_idle_type idle = this_rq->idle_balance ?
9828 CPU_IDLE : CPU_NOT_IDLE;
9831 * If this CPU has a pending nohz_balance_kick, then do the
9832 * balancing on behalf of the other idle CPUs whose ticks are
9833 * stopped. Do nohz_idle_balance *before* rebalance_domains to
9834 * give the idle CPUs a chance to load balance. Else we may
9835 * load balance only within the local sched_domain hierarchy
9836 * and abort nohz_idle_balance altogether if we pull some load.
9838 if (nohz_idle_balance(this_rq, idle))
9841 /* normal load balance */
9842 update_blocked_averages(this_rq->cpu);
9843 rebalance_domains(this_rq, idle);
9847 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
9849 void trigger_load_balance(struct rq *rq)
9851 /* Don't need to rebalance while attached to NULL domain */
9852 if (unlikely(on_null_domain(rq)))
9855 if (time_after_eq(jiffies, rq->next_balance))
9856 raise_softirq(SCHED_SOFTIRQ);
9858 nohz_balancer_kick(rq);
9861 static void rq_online_fair(struct rq *rq)
9865 update_runtime_enabled(rq);
9868 static void rq_offline_fair(struct rq *rq)
9872 /* Ensure any throttled groups are reachable by pick_next_task */
9873 unthrottle_offline_cfs_rqs(rq);
9876 #endif /* CONFIG_SMP */
9879 * scheduler tick hitting a task of our scheduling class.
9881 * NOTE: This function can be called remotely by the tick offload that
9882 * goes along full dynticks. Therefore no local assumption can be made
9883 * and everything must be accessed through the @rq and @curr passed in
9886 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
9888 struct cfs_rq *cfs_rq;
9889 struct sched_entity *se = &curr->se;
9891 for_each_sched_entity(se) {
9892 cfs_rq = cfs_rq_of(se);
9893 entity_tick(cfs_rq, se, queued);
9896 if (static_branch_unlikely(&sched_numa_balancing))
9897 task_tick_numa(rq, curr);
9901 * called on fork with the child task as argument from the parent's context
9902 * - child not yet on the tasklist
9903 * - preemption disabled
9905 static void task_fork_fair(struct task_struct *p)
9907 struct cfs_rq *cfs_rq;
9908 struct sched_entity *se = &p->se, *curr;
9909 struct rq *rq = this_rq();
9913 update_rq_clock(rq);
9915 cfs_rq = task_cfs_rq(current);
9916 curr = cfs_rq->curr;
9918 update_curr(cfs_rq);
9919 se->vruntime = curr->vruntime;
9921 place_entity(cfs_rq, se, 1);
9923 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
9925 * Upon rescheduling, sched_class::put_prev_task() will place
9926 * 'current' within the tree based on its new key value.
9928 swap(curr->vruntime, se->vruntime);
9932 se->vruntime -= cfs_rq->min_vruntime;
9937 * Priority of the task has changed. Check to see if we preempt
9941 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
9943 if (!task_on_rq_queued(p))
9947 * Reschedule if we are currently running on this runqueue and
9948 * our priority decreased, or if we are not currently running on
9949 * this runqueue and our priority is higher than the current's
9951 if (rq->curr == p) {
9952 if (p->prio > oldprio)
9955 check_preempt_curr(rq, p, 0);
9958 static inline bool vruntime_normalized(struct task_struct *p)
9960 struct sched_entity *se = &p->se;
9963 * In both the TASK_ON_RQ_QUEUED and TASK_ON_RQ_MIGRATING cases,
9964 * the dequeue_entity(.flags=0) will already have normalized the
9971 * When !on_rq, vruntime of the task has usually NOT been normalized.
9972 * But there are some cases where it has already been normalized:
9974 * - A forked child which is waiting for being woken up by
9975 * wake_up_new_task().
9976 * - A task which has been woken up by try_to_wake_up() and
9977 * waiting for actually being woken up by sched_ttwu_pending().
9979 if (!se->sum_exec_runtime || p->state == TASK_WAKING)
9985 #ifdef CONFIG_FAIR_GROUP_SCHED
9987 * Propagate the changes of the sched_entity across the tg tree to make it
9988 * visible to the root
9990 static void propagate_entity_cfs_rq(struct sched_entity *se)
9992 struct cfs_rq *cfs_rq;
9994 /* Start to propagate at parent */
9997 for_each_sched_entity(se) {
9998 cfs_rq = cfs_rq_of(se);
10000 if (cfs_rq_throttled(cfs_rq))
10003 update_load_avg(cfs_rq, se, UPDATE_TG);
10007 static void propagate_entity_cfs_rq(struct sched_entity *se) { }
10010 static void detach_entity_cfs_rq(struct sched_entity *se)
10012 struct cfs_rq *cfs_rq = cfs_rq_of(se);
10014 /* Catch up with the cfs_rq and remove our load when we leave */
10015 update_load_avg(cfs_rq, se, 0);
10016 detach_entity_load_avg(cfs_rq, se);
10017 update_tg_load_avg(cfs_rq, false);
10018 propagate_entity_cfs_rq(se);
10021 static void attach_entity_cfs_rq(struct sched_entity *se)
10023 struct cfs_rq *cfs_rq = cfs_rq_of(se);
10025 #ifdef CONFIG_FAIR_GROUP_SCHED
10027 * Since the real-depth could have been changed (only FAIR
10028 * class maintain depth value), reset depth properly.
10030 se->depth = se->parent ? se->parent->depth + 1 : 0;
10033 /* Synchronize entity with its cfs_rq */
10034 update_load_avg(cfs_rq, se, sched_feat(ATTACH_AGE_LOAD) ? 0 : SKIP_AGE_LOAD);
10035 attach_entity_load_avg(cfs_rq, se, 0);
10036 update_tg_load_avg(cfs_rq, false);
10037 propagate_entity_cfs_rq(se);
10040 static void detach_task_cfs_rq(struct task_struct *p)
10042 struct sched_entity *se = &p->se;
10043 struct cfs_rq *cfs_rq = cfs_rq_of(se);
10045 if (!vruntime_normalized(p)) {
10047 * Fix up our vruntime so that the current sleep doesn't
10048 * cause 'unlimited' sleep bonus.
10050 place_entity(cfs_rq, se, 0);
10051 se->vruntime -= cfs_rq->min_vruntime;
10054 detach_entity_cfs_rq(se);
10057 static void attach_task_cfs_rq(struct task_struct *p)
10059 struct sched_entity *se = &p->se;
10060 struct cfs_rq *cfs_rq = cfs_rq_of(se);
10062 attach_entity_cfs_rq(se);
10064 if (!vruntime_normalized(p))
10065 se->vruntime += cfs_rq->min_vruntime;
10068 static void switched_from_fair(struct rq *rq, struct task_struct *p)
10070 detach_task_cfs_rq(p);
10073 static void switched_to_fair(struct rq *rq, struct task_struct *p)
10075 attach_task_cfs_rq(p);
10077 if (task_on_rq_queued(p)) {
10079 * We were most likely switched from sched_rt, so
10080 * kick off the schedule if running, otherwise just see
10081 * if we can still preempt the current task.
10086 check_preempt_curr(rq, p, 0);
10090 /* Account for a task changing its policy or group.
10092 * This routine is mostly called to set cfs_rq->curr field when a task
10093 * migrates between groups/classes.
10095 static void set_curr_task_fair(struct rq *rq)
10097 struct sched_entity *se = &rq->curr->se;
10099 for_each_sched_entity(se) {
10100 struct cfs_rq *cfs_rq = cfs_rq_of(se);
10102 set_next_entity(cfs_rq, se);
10103 /* ensure bandwidth has been allocated on our new cfs_rq */
10104 account_cfs_rq_runtime(cfs_rq, 0);
10108 void init_cfs_rq(struct cfs_rq *cfs_rq)
10110 cfs_rq->tasks_timeline = RB_ROOT_CACHED;
10111 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
10112 #ifndef CONFIG_64BIT
10113 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
10116 raw_spin_lock_init(&cfs_rq->removed.lock);
10120 #ifdef CONFIG_FAIR_GROUP_SCHED
10121 static void task_set_group_fair(struct task_struct *p)
10123 struct sched_entity *se = &p->se;
10125 set_task_rq(p, task_cpu(p));
10126 se->depth = se->parent ? se->parent->depth + 1 : 0;
10129 static void task_move_group_fair(struct task_struct *p)
10131 detach_task_cfs_rq(p);
10132 set_task_rq(p, task_cpu(p));
10135 /* Tell se's cfs_rq has been changed -- migrated */
10136 p->se.avg.last_update_time = 0;
10138 attach_task_cfs_rq(p);
10141 static void task_change_group_fair(struct task_struct *p, int type)
10144 case TASK_SET_GROUP:
10145 task_set_group_fair(p);
10148 case TASK_MOVE_GROUP:
10149 task_move_group_fair(p);
10154 void free_fair_sched_group(struct task_group *tg)
10158 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
10160 for_each_possible_cpu(i) {
10162 kfree(tg->cfs_rq[i]);
10171 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
10173 struct sched_entity *se;
10174 struct cfs_rq *cfs_rq;
10177 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
10180 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
10184 tg->shares = NICE_0_LOAD;
10186 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
10188 for_each_possible_cpu(i) {
10189 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
10190 GFP_KERNEL, cpu_to_node(i));
10194 se = kzalloc_node(sizeof(struct sched_entity),
10195 GFP_KERNEL, cpu_to_node(i));
10199 init_cfs_rq(cfs_rq);
10200 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
10201 init_entity_runnable_average(se);
10212 void online_fair_sched_group(struct task_group *tg)
10214 struct sched_entity *se;
10218 for_each_possible_cpu(i) {
10222 raw_spin_lock_irq(&rq->lock);
10223 update_rq_clock(rq);
10224 attach_entity_cfs_rq(se);
10225 sync_throttle(tg, i);
10226 raw_spin_unlock_irq(&rq->lock);
10230 void unregister_fair_sched_group(struct task_group *tg)
10232 unsigned long flags;
10236 for_each_possible_cpu(cpu) {
10238 remove_entity_load_avg(tg->se[cpu]);
10241 * Only empty task groups can be destroyed; so we can speculatively
10242 * check on_list without danger of it being re-added.
10244 if (!tg->cfs_rq[cpu]->on_list)
10249 raw_spin_lock_irqsave(&rq->lock, flags);
10250 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
10251 raw_spin_unlock_irqrestore(&rq->lock, flags);
10255 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
10256 struct sched_entity *se, int cpu,
10257 struct sched_entity *parent)
10259 struct rq *rq = cpu_rq(cpu);
10263 init_cfs_rq_runtime(cfs_rq);
10265 tg->cfs_rq[cpu] = cfs_rq;
10268 /* se could be NULL for root_task_group */
10273 se->cfs_rq = &rq->cfs;
10276 se->cfs_rq = parent->my_q;
10277 se->depth = parent->depth + 1;
10281 /* guarantee group entities always have weight */
10282 update_load_set(&se->load, NICE_0_LOAD);
10283 se->parent = parent;
10286 static DEFINE_MUTEX(shares_mutex);
10288 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
10293 * We can't change the weight of the root cgroup.
10298 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
10300 mutex_lock(&shares_mutex);
10301 if (tg->shares == shares)
10304 tg->shares = shares;
10305 for_each_possible_cpu(i) {
10306 struct rq *rq = cpu_rq(i);
10307 struct sched_entity *se = tg->se[i];
10308 struct rq_flags rf;
10310 /* Propagate contribution to hierarchy */
10311 rq_lock_irqsave(rq, &rf);
10312 update_rq_clock(rq);
10313 for_each_sched_entity(se) {
10314 update_load_avg(cfs_rq_of(se), se, UPDATE_TG);
10315 update_cfs_group(se);
10317 rq_unlock_irqrestore(rq, &rf);
10321 mutex_unlock(&shares_mutex);
10324 #else /* CONFIG_FAIR_GROUP_SCHED */
10326 void free_fair_sched_group(struct task_group *tg) { }
10328 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
10333 void online_fair_sched_group(struct task_group *tg) { }
10335 void unregister_fair_sched_group(struct task_group *tg) { }
10337 #endif /* CONFIG_FAIR_GROUP_SCHED */
10340 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
10342 struct sched_entity *se = &task->se;
10343 unsigned int rr_interval = 0;
10346 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
10349 if (rq->cfs.load.weight)
10350 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
10352 return rr_interval;
10356 * All the scheduling class methods:
10358 const struct sched_class fair_sched_class = {
10359 .next = &idle_sched_class,
10360 .enqueue_task = enqueue_task_fair,
10361 .dequeue_task = dequeue_task_fair,
10362 .yield_task = yield_task_fair,
10363 .yield_to_task = yield_to_task_fair,
10365 .check_preempt_curr = check_preempt_wakeup,
10367 .pick_next_task = pick_next_task_fair,
10368 .put_prev_task = put_prev_task_fair,
10371 .select_task_rq = select_task_rq_fair,
10372 .migrate_task_rq = migrate_task_rq_fair,
10374 .rq_online = rq_online_fair,
10375 .rq_offline = rq_offline_fair,
10377 .task_dead = task_dead_fair,
10378 .set_cpus_allowed = set_cpus_allowed_common,
10381 .set_curr_task = set_curr_task_fair,
10382 .task_tick = task_tick_fair,
10383 .task_fork = task_fork_fair,
10385 .prio_changed = prio_changed_fair,
10386 .switched_from = switched_from_fair,
10387 .switched_to = switched_to_fair,
10389 .get_rr_interval = get_rr_interval_fair,
10391 .update_curr = update_curr_fair,
10393 #ifdef CONFIG_FAIR_GROUP_SCHED
10394 .task_change_group = task_change_group_fair,
10398 #ifdef CONFIG_SCHED_DEBUG
10399 void print_cfs_stats(struct seq_file *m, int cpu)
10401 struct cfs_rq *cfs_rq, *pos;
10404 for_each_leaf_cfs_rq_safe(cpu_rq(cpu), cfs_rq, pos)
10405 print_cfs_rq(m, cpu, cfs_rq);
10409 #ifdef CONFIG_NUMA_BALANCING
10410 void show_numa_stats(struct task_struct *p, struct seq_file *m)
10413 unsigned long tsf = 0, tpf = 0, gsf = 0, gpf = 0;
10415 for_each_online_node(node) {
10416 if (p->numa_faults) {
10417 tsf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 0)];
10418 tpf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 1)];
10420 if (p->numa_group) {
10421 gsf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 0)],
10422 gpf = p->numa_group->faults[task_faults_idx(NUMA_MEM, node, 1)];
10424 print_numa_stats(m, node, tsf, tpf, gsf, gpf);
10427 #endif /* CONFIG_NUMA_BALANCING */
10428 #endif /* CONFIG_SCHED_DEBUG */
10430 __init void init_sched_fair_class(void)
10433 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
10435 #ifdef CONFIG_NO_HZ_COMMON
10436 nohz.next_balance = jiffies;
10437 nohz.next_blocked = jiffies;
10438 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);