1 // SPDX-License-Identifier: GPL-2.0
3 * Completely Fair Scheduling (CFS) Class (SCHED_NORMAL/SCHED_BATCH)
5 * Copyright (C) 2007 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
7 * Interactivity improvements by Mike Galbraith
8 * (C) 2007 Mike Galbraith <efault@gmx.de>
10 * Various enhancements by Dmitry Adamushko.
11 * (C) 2007 Dmitry Adamushko <dmitry.adamushko@gmail.com>
13 * Group scheduling enhancements by Srivatsa Vaddagiri
14 * Copyright IBM Corporation, 2007
15 * Author: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
17 * Scaled math optimizations by Thomas Gleixner
18 * Copyright (C) 2007, Thomas Gleixner <tglx@linutronix.de>
20 * Adaptive scheduling granularity, math enhancements by Peter Zijlstra
21 * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra
25 #include <trace/events/sched.h>
28 * Targeted preemption latency for CPU-bound tasks:
30 * NOTE: this latency value is not the same as the concept of
31 * 'timeslice length' - timeslices in CFS are of variable length
32 * and have no persistent notion like in traditional, time-slice
33 * based scheduling concepts.
35 * (to see the precise effective timeslice length of your workload,
36 * run vmstat and monitor the context-switches (cs) field)
38 * (default: 6ms * (1 + ilog(ncpus)), units: nanoseconds)
40 unsigned int sysctl_sched_latency = 6000000ULL;
41 unsigned int normalized_sysctl_sched_latency = 6000000ULL;
44 * The initial- and re-scaling of tunables is configurable
48 * SCHED_TUNABLESCALING_NONE - unscaled, always *1
49 * SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus)
50 * SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus
52 * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus))
54 enum sched_tunable_scaling sysctl_sched_tunable_scaling = SCHED_TUNABLESCALING_LOG;
57 * Minimal preemption granularity for CPU-bound tasks:
59 * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds)
61 unsigned int sysctl_sched_min_granularity = 750000ULL;
62 unsigned int normalized_sysctl_sched_min_granularity = 750000ULL;
65 * This value is kept at sysctl_sched_latency/sysctl_sched_min_granularity
67 static unsigned int sched_nr_latency = 8;
70 * After fork, child runs first. If set to 0 (default) then
71 * parent will (try to) run first.
73 unsigned int sysctl_sched_child_runs_first __read_mostly;
76 * SCHED_OTHER wake-up granularity.
78 * This option delays the preemption effects of decoupled workloads
79 * and reduces their over-scheduling. Synchronous workloads will still
80 * have immediate wakeup/sleep latencies.
82 * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds)
84 unsigned int sysctl_sched_wakeup_granularity = 1000000UL;
85 unsigned int normalized_sysctl_sched_wakeup_granularity = 1000000UL;
87 const_debug unsigned int sysctl_sched_migration_cost = 500000UL;
91 * For asym packing, by default the lower numbered CPU has higher priority.
93 int __weak arch_asym_cpu_priority(int cpu)
99 #ifdef CONFIG_CFS_BANDWIDTH
101 * Amount of runtime to allocate from global (tg) to local (per-cfs_rq) pool
102 * each time a cfs_rq requests quota.
104 * Note: in the case that the slice exceeds the runtime remaining (either due
105 * to consumption or the quota being specified to be smaller than the slice)
106 * we will always only issue the remaining available time.
108 * (default: 5 msec, units: microseconds)
110 unsigned int sysctl_sched_cfs_bandwidth_slice = 5000UL;
114 * The margin used when comparing utilization with CPU capacity:
115 * util * margin < capacity * 1024
119 unsigned int capacity_margin = 1280;
121 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
127 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
133 static inline void update_load_set(struct load_weight *lw, unsigned long w)
140 * Increase the granularity value when there are more CPUs,
141 * because with more CPUs the 'effective latency' as visible
142 * to users decreases. But the relationship is not linear,
143 * so pick a second-best guess by going with the log2 of the
146 * This idea comes from the SD scheduler of Con Kolivas:
148 static unsigned int get_update_sysctl_factor(void)
150 unsigned int cpus = min_t(unsigned int, num_online_cpus(), 8);
153 switch (sysctl_sched_tunable_scaling) {
154 case SCHED_TUNABLESCALING_NONE:
157 case SCHED_TUNABLESCALING_LINEAR:
160 case SCHED_TUNABLESCALING_LOG:
162 factor = 1 + ilog2(cpus);
169 static void update_sysctl(void)
171 unsigned int factor = get_update_sysctl_factor();
173 #define SET_SYSCTL(name) \
174 (sysctl_##name = (factor) * normalized_sysctl_##name)
175 SET_SYSCTL(sched_min_granularity);
176 SET_SYSCTL(sched_latency);
177 SET_SYSCTL(sched_wakeup_granularity);
181 void sched_init_granularity(void)
186 #define WMULT_CONST (~0U)
187 #define WMULT_SHIFT 32
189 static void __update_inv_weight(struct load_weight *lw)
193 if (likely(lw->inv_weight))
196 w = scale_load_down(lw->weight);
198 if (BITS_PER_LONG > 32 && unlikely(w >= WMULT_CONST))
200 else if (unlikely(!w))
201 lw->inv_weight = WMULT_CONST;
203 lw->inv_weight = WMULT_CONST / w;
207 * delta_exec * weight / lw.weight
209 * (delta_exec * (weight * lw->inv_weight)) >> WMULT_SHIFT
211 * Either weight := NICE_0_LOAD and lw \e sched_prio_to_wmult[], in which case
212 * we're guaranteed shift stays positive because inv_weight is guaranteed to
213 * fit 32 bits, and NICE_0_LOAD gives another 10 bits; therefore shift >= 22.
215 * Or, weight =< lw.weight (because lw.weight is the runqueue weight), thus
216 * weight/lw.weight <= 1, and therefore our shift will also be positive.
218 static u64 __calc_delta(u64 delta_exec, unsigned long weight, struct load_weight *lw)
220 u64 fact = scale_load_down(weight);
221 int shift = WMULT_SHIFT;
223 __update_inv_weight(lw);
225 if (unlikely(fact >> 32)) {
232 /* hint to use a 32x32->64 mul */
233 fact = (u64)(u32)fact * lw->inv_weight;
240 return mul_u64_u32_shr(delta_exec, fact, shift);
244 const struct sched_class fair_sched_class;
246 /**************************************************************
247 * CFS operations on generic schedulable entities:
250 #ifdef CONFIG_FAIR_GROUP_SCHED
252 /* cpu runqueue to which this cfs_rq is attached */
253 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
258 static inline struct task_struct *task_of(struct sched_entity *se)
260 SCHED_WARN_ON(!entity_is_task(se));
261 return container_of(se, struct task_struct, se);
264 /* Walk up scheduling entities hierarchy */
265 #define for_each_sched_entity(se) \
266 for (; se; se = se->parent)
268 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
273 /* runqueue on which this entity is (to be) queued */
274 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
279 /* runqueue "owned" by this group */
280 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
285 static inline bool list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
287 struct rq *rq = rq_of(cfs_rq);
288 int cpu = cpu_of(rq);
291 return rq->tmp_alone_branch == &rq->leaf_cfs_rq_list;
296 * Ensure we either appear before our parent (if already
297 * enqueued) or force our parent to appear after us when it is
298 * enqueued. The fact that we always enqueue bottom-up
299 * reduces this to two cases and a special case for the root
300 * cfs_rq. Furthermore, it also means that we will always reset
301 * tmp_alone_branch either when the branch is connected
302 * to a tree or when we reach the top of the tree
304 if (cfs_rq->tg->parent &&
305 cfs_rq->tg->parent->cfs_rq[cpu]->on_list) {
307 * If parent is already on the list, we add the child
308 * just before. Thanks to circular linked property of
309 * the list, this means to put the child at the tail
310 * of the list that starts by parent.
312 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
313 &(cfs_rq->tg->parent->cfs_rq[cpu]->leaf_cfs_rq_list));
315 * The branch is now connected to its tree so we can
316 * reset tmp_alone_branch to the beginning of the
319 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
323 if (!cfs_rq->tg->parent) {
325 * cfs rq without parent should be put
326 * at the tail of the list.
328 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
329 &rq->leaf_cfs_rq_list);
331 * We have reach the top of a tree so we can reset
332 * tmp_alone_branch to the beginning of the list.
334 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
339 * The parent has not already been added so we want to
340 * make sure that it will be put after us.
341 * tmp_alone_branch points to the begin of the branch
342 * where we will add parent.
344 list_add_rcu(&cfs_rq->leaf_cfs_rq_list, rq->tmp_alone_branch);
346 * update tmp_alone_branch to points to the new begin
349 rq->tmp_alone_branch = &cfs_rq->leaf_cfs_rq_list;
353 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
355 if (cfs_rq->on_list) {
356 struct rq *rq = rq_of(cfs_rq);
359 * With cfs_rq being unthrottled/throttled during an enqueue,
360 * it can happen the tmp_alone_branch points the a leaf that
361 * we finally want to del. In this case, tmp_alone_branch moves
362 * to the prev element but it will point to rq->leaf_cfs_rq_list
363 * at the end of the enqueue.
365 if (rq->tmp_alone_branch == &cfs_rq->leaf_cfs_rq_list)
366 rq->tmp_alone_branch = cfs_rq->leaf_cfs_rq_list.prev;
368 list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
373 static inline void assert_list_leaf_cfs_rq(struct rq *rq)
375 SCHED_WARN_ON(rq->tmp_alone_branch != &rq->leaf_cfs_rq_list);
378 /* Iterate thr' all leaf cfs_rq's on a runqueue */
379 #define for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) \
380 list_for_each_entry_safe(cfs_rq, pos, &rq->leaf_cfs_rq_list, \
383 /* Do the two (enqueued) entities belong to the same group ? */
384 static inline struct cfs_rq *
385 is_same_group(struct sched_entity *se, struct sched_entity *pse)
387 if (se->cfs_rq == pse->cfs_rq)
393 static inline struct sched_entity *parent_entity(struct sched_entity *se)
399 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
401 int se_depth, pse_depth;
404 * preemption test can be made between sibling entities who are in the
405 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
406 * both tasks until we find their ancestors who are siblings of common
410 /* First walk up until both entities are at same depth */
411 se_depth = (*se)->depth;
412 pse_depth = (*pse)->depth;
414 while (se_depth > pse_depth) {
416 *se = parent_entity(*se);
419 while (pse_depth > se_depth) {
421 *pse = parent_entity(*pse);
424 while (!is_same_group(*se, *pse)) {
425 *se = parent_entity(*se);
426 *pse = parent_entity(*pse);
430 #else /* !CONFIG_FAIR_GROUP_SCHED */
432 static inline struct task_struct *task_of(struct sched_entity *se)
434 return container_of(se, struct task_struct, se);
437 static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
439 return container_of(cfs_rq, struct rq, cfs);
443 #define for_each_sched_entity(se) \
444 for (; se; se = NULL)
446 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
448 return &task_rq(p)->cfs;
451 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
453 struct task_struct *p = task_of(se);
454 struct rq *rq = task_rq(p);
459 /* runqueue "owned" by this group */
460 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
465 static inline bool list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
470 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
474 static inline void assert_list_leaf_cfs_rq(struct rq *rq)
478 #define for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) \
479 for (cfs_rq = &rq->cfs, pos = NULL; cfs_rq; cfs_rq = pos)
481 static inline struct sched_entity *parent_entity(struct sched_entity *se)
487 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
491 #endif /* CONFIG_FAIR_GROUP_SCHED */
493 static __always_inline
494 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec);
496 /**************************************************************
497 * Scheduling class tree data structure manipulation methods:
500 static inline u64 max_vruntime(u64 max_vruntime, u64 vruntime)
502 s64 delta = (s64)(vruntime - max_vruntime);
504 max_vruntime = vruntime;
509 static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime)
511 s64 delta = (s64)(vruntime - min_vruntime);
513 min_vruntime = vruntime;
518 static inline int entity_before(struct sched_entity *a,
519 struct sched_entity *b)
521 return (s64)(a->vruntime - b->vruntime) < 0;
524 static void update_min_vruntime(struct cfs_rq *cfs_rq)
526 struct sched_entity *curr = cfs_rq->curr;
527 struct rb_node *leftmost = rb_first_cached(&cfs_rq->tasks_timeline);
529 u64 vruntime = cfs_rq->min_vruntime;
533 vruntime = curr->vruntime;
538 if (leftmost) { /* non-empty tree */
539 struct sched_entity *se;
540 se = rb_entry(leftmost, struct sched_entity, run_node);
543 vruntime = se->vruntime;
545 vruntime = min_vruntime(vruntime, se->vruntime);
548 /* ensure we never gain time by being placed backwards. */
549 cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime);
552 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
557 * Enqueue an entity into the rb-tree:
559 static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
561 struct rb_node **link = &cfs_rq->tasks_timeline.rb_root.rb_node;
562 struct rb_node *parent = NULL;
563 struct sched_entity *entry;
564 bool leftmost = true;
567 * Find the right place in the rbtree:
571 entry = rb_entry(parent, struct sched_entity, run_node);
573 * We dont care about collisions. Nodes with
574 * the same key stay together.
576 if (entity_before(se, entry)) {
577 link = &parent->rb_left;
579 link = &parent->rb_right;
584 rb_link_node(&se->run_node, parent, link);
585 rb_insert_color_cached(&se->run_node,
586 &cfs_rq->tasks_timeline, leftmost);
589 static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
591 rb_erase_cached(&se->run_node, &cfs_rq->tasks_timeline);
594 struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq)
596 struct rb_node *left = rb_first_cached(&cfs_rq->tasks_timeline);
601 return rb_entry(left, struct sched_entity, run_node);
604 static struct sched_entity *__pick_next_entity(struct sched_entity *se)
606 struct rb_node *next = rb_next(&se->run_node);
611 return rb_entry(next, struct sched_entity, run_node);
614 #ifdef CONFIG_SCHED_DEBUG
615 struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
617 struct rb_node *last = rb_last(&cfs_rq->tasks_timeline.rb_root);
622 return rb_entry(last, struct sched_entity, run_node);
625 /**************************************************************
626 * Scheduling class statistics methods:
629 int sched_proc_update_handler(struct ctl_table *table, int write,
630 void __user *buffer, size_t *lenp,
633 int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
634 unsigned int factor = get_update_sysctl_factor();
639 sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency,
640 sysctl_sched_min_granularity);
642 #define WRT_SYSCTL(name) \
643 (normalized_sysctl_##name = sysctl_##name / (factor))
644 WRT_SYSCTL(sched_min_granularity);
645 WRT_SYSCTL(sched_latency);
646 WRT_SYSCTL(sched_wakeup_granularity);
656 static inline u64 calc_delta_fair(u64 delta, struct sched_entity *se)
658 if (unlikely(se->load.weight != NICE_0_LOAD))
659 delta = __calc_delta(delta, NICE_0_LOAD, &se->load);
665 * The idea is to set a period in which each task runs once.
667 * When there are too many tasks (sched_nr_latency) we have to stretch
668 * this period because otherwise the slices get too small.
670 * p = (nr <= nl) ? l : l*nr/nl
672 static u64 __sched_period(unsigned long nr_running)
674 if (unlikely(nr_running > sched_nr_latency))
675 return nr_running * sysctl_sched_min_granularity;
677 return sysctl_sched_latency;
681 * We calculate the wall-time slice from the period by taking a part
682 * proportional to the weight.
686 static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
688 u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq);
690 for_each_sched_entity(se) {
691 struct load_weight *load;
692 struct load_weight lw;
694 cfs_rq = cfs_rq_of(se);
695 load = &cfs_rq->load;
697 if (unlikely(!se->on_rq)) {
700 update_load_add(&lw, se->load.weight);
703 slice = __calc_delta(slice, se->load.weight, load);
709 * We calculate the vruntime slice of a to-be-inserted task.
713 static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se)
715 return calc_delta_fair(sched_slice(cfs_rq, se), se);
720 #include "sched-pelt.h"
722 static int select_idle_sibling(struct task_struct *p, int prev_cpu, int cpu);
723 static unsigned long task_h_load(struct task_struct *p);
725 /* Give new sched_entity start runnable values to heavy its load in infant time */
726 void init_entity_runnable_average(struct sched_entity *se)
728 struct sched_avg *sa = &se->avg;
730 memset(sa, 0, sizeof(*sa));
733 * Tasks are intialized with full load to be seen as heavy tasks until
734 * they get a chance to stabilize to their real load level.
735 * Group entities are intialized with zero load to reflect the fact that
736 * nothing has been attached to the task group yet.
738 if (entity_is_task(se))
739 sa->runnable_load_avg = sa->load_avg = scale_load_down(se->load.weight);
741 se->runnable_weight = se->load.weight;
743 /* when this task enqueue'ed, it will contribute to its cfs_rq's load_avg */
746 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq);
747 static void attach_entity_cfs_rq(struct sched_entity *se);
750 * With new tasks being created, their initial util_avgs are extrapolated
751 * based on the cfs_rq's current util_avg:
753 * util_avg = cfs_rq->util_avg / (cfs_rq->load_avg + 1) * se.load.weight
755 * However, in many cases, the above util_avg does not give a desired
756 * value. Moreover, the sum of the util_avgs may be divergent, such
757 * as when the series is a harmonic series.
759 * To solve this problem, we also cap the util_avg of successive tasks to
760 * only 1/2 of the left utilization budget:
762 * util_avg_cap = (cpu_scale - cfs_rq->avg.util_avg) / 2^n
764 * where n denotes the nth task and cpu_scale the CPU capacity.
766 * For example, for a CPU with 1024 of capacity, a simplest series from
767 * the beginning would be like:
769 * task util_avg: 512, 256, 128, 64, 32, 16, 8, ...
770 * cfs_rq util_avg: 512, 768, 896, 960, 992, 1008, 1016, ...
772 * Finally, that extrapolated util_avg is clamped to the cap (util_avg_cap)
773 * if util_avg > util_avg_cap.
775 void post_init_entity_util_avg(struct sched_entity *se)
777 struct cfs_rq *cfs_rq = cfs_rq_of(se);
778 struct sched_avg *sa = &se->avg;
779 long cpu_scale = arch_scale_cpu_capacity(NULL, cpu_of(rq_of(cfs_rq)));
780 long cap = (long)(cpu_scale - cfs_rq->avg.util_avg) / 2;
783 if (cfs_rq->avg.util_avg != 0) {
784 sa->util_avg = cfs_rq->avg.util_avg * se->load.weight;
785 sa->util_avg /= (cfs_rq->avg.load_avg + 1);
787 if (sa->util_avg > cap)
794 if (entity_is_task(se)) {
795 struct task_struct *p = task_of(se);
796 if (p->sched_class != &fair_sched_class) {
798 * For !fair tasks do:
800 update_cfs_rq_load_avg(now, cfs_rq);
801 attach_entity_load_avg(cfs_rq, se, 0);
802 switched_from_fair(rq, p);
804 * such that the next switched_to_fair() has the
807 se->avg.last_update_time = cfs_rq_clock_task(cfs_rq);
812 attach_entity_cfs_rq(se);
815 #else /* !CONFIG_SMP */
816 void init_entity_runnable_average(struct sched_entity *se)
819 void post_init_entity_util_avg(struct sched_entity *se)
822 static void update_tg_load_avg(struct cfs_rq *cfs_rq, int force)
825 #endif /* CONFIG_SMP */
828 * Update the current task's runtime statistics.
830 static void update_curr(struct cfs_rq *cfs_rq)
832 struct sched_entity *curr = cfs_rq->curr;
833 u64 now = rq_clock_task(rq_of(cfs_rq));
839 delta_exec = now - curr->exec_start;
840 if (unlikely((s64)delta_exec <= 0))
843 curr->exec_start = now;
845 schedstat_set(curr->statistics.exec_max,
846 max(delta_exec, curr->statistics.exec_max));
848 curr->sum_exec_runtime += delta_exec;
849 schedstat_add(cfs_rq->exec_clock, delta_exec);
851 curr->vruntime += calc_delta_fair(delta_exec, curr);
852 update_min_vruntime(cfs_rq);
854 if (entity_is_task(curr)) {
855 struct task_struct *curtask = task_of(curr);
857 trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime);
858 cgroup_account_cputime(curtask, delta_exec);
859 account_group_exec_runtime(curtask, delta_exec);
862 account_cfs_rq_runtime(cfs_rq, delta_exec);
865 static void update_curr_fair(struct rq *rq)
867 update_curr(cfs_rq_of(&rq->curr->se));
871 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
873 u64 wait_start, prev_wait_start;
875 if (!schedstat_enabled())
878 wait_start = rq_clock(rq_of(cfs_rq));
879 prev_wait_start = schedstat_val(se->statistics.wait_start);
881 if (entity_is_task(se) && task_on_rq_migrating(task_of(se)) &&
882 likely(wait_start > prev_wait_start))
883 wait_start -= prev_wait_start;
885 __schedstat_set(se->statistics.wait_start, wait_start);
889 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
891 struct task_struct *p;
894 if (!schedstat_enabled())
897 delta = rq_clock(rq_of(cfs_rq)) - schedstat_val(se->statistics.wait_start);
899 if (entity_is_task(se)) {
901 if (task_on_rq_migrating(p)) {
903 * Preserve migrating task's wait time so wait_start
904 * time stamp can be adjusted to accumulate wait time
905 * prior to migration.
907 __schedstat_set(se->statistics.wait_start, delta);
910 trace_sched_stat_wait(p, delta);
913 __schedstat_set(se->statistics.wait_max,
914 max(schedstat_val(se->statistics.wait_max), delta));
915 __schedstat_inc(se->statistics.wait_count);
916 __schedstat_add(se->statistics.wait_sum, delta);
917 __schedstat_set(se->statistics.wait_start, 0);
921 update_stats_enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
923 struct task_struct *tsk = NULL;
924 u64 sleep_start, block_start;
926 if (!schedstat_enabled())
929 sleep_start = schedstat_val(se->statistics.sleep_start);
930 block_start = schedstat_val(se->statistics.block_start);
932 if (entity_is_task(se))
936 u64 delta = rq_clock(rq_of(cfs_rq)) - sleep_start;
941 if (unlikely(delta > schedstat_val(se->statistics.sleep_max)))
942 __schedstat_set(se->statistics.sleep_max, delta);
944 __schedstat_set(se->statistics.sleep_start, 0);
945 __schedstat_add(se->statistics.sum_sleep_runtime, delta);
948 account_scheduler_latency(tsk, delta >> 10, 1);
949 trace_sched_stat_sleep(tsk, delta);
953 u64 delta = rq_clock(rq_of(cfs_rq)) - block_start;
958 if (unlikely(delta > schedstat_val(se->statistics.block_max)))
959 __schedstat_set(se->statistics.block_max, delta);
961 __schedstat_set(se->statistics.block_start, 0);
962 __schedstat_add(se->statistics.sum_sleep_runtime, delta);
965 if (tsk->in_iowait) {
966 __schedstat_add(se->statistics.iowait_sum, delta);
967 __schedstat_inc(se->statistics.iowait_count);
968 trace_sched_stat_iowait(tsk, delta);
971 trace_sched_stat_blocked(tsk, delta);
974 * Blocking time is in units of nanosecs, so shift by
975 * 20 to get a milliseconds-range estimation of the
976 * amount of time that the task spent sleeping:
978 if (unlikely(prof_on == SLEEP_PROFILING)) {
979 profile_hits(SLEEP_PROFILING,
980 (void *)get_wchan(tsk),
983 account_scheduler_latency(tsk, delta >> 10, 0);
989 * Task is being enqueued - update stats:
992 update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
994 if (!schedstat_enabled())
998 * Are we enqueueing a waiting task? (for current tasks
999 * a dequeue/enqueue event is a NOP)
1001 if (se != cfs_rq->curr)
1002 update_stats_wait_start(cfs_rq, se);
1004 if (flags & ENQUEUE_WAKEUP)
1005 update_stats_enqueue_sleeper(cfs_rq, se);
1009 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
1012 if (!schedstat_enabled())
1016 * Mark the end of the wait period if dequeueing a
1019 if (se != cfs_rq->curr)
1020 update_stats_wait_end(cfs_rq, se);
1022 if ((flags & DEQUEUE_SLEEP) && entity_is_task(se)) {
1023 struct task_struct *tsk = task_of(se);
1025 if (tsk->state & TASK_INTERRUPTIBLE)
1026 __schedstat_set(se->statistics.sleep_start,
1027 rq_clock(rq_of(cfs_rq)));
1028 if (tsk->state & TASK_UNINTERRUPTIBLE)
1029 __schedstat_set(se->statistics.block_start,
1030 rq_clock(rq_of(cfs_rq)));
1035 * We are picking a new current task - update its stats:
1038 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
1041 * We are starting a new run period:
1043 se->exec_start = rq_clock_task(rq_of(cfs_rq));
1046 /**************************************************
1047 * Scheduling class queueing methods:
1050 #ifdef CONFIG_NUMA_BALANCING
1052 * Approximate time to scan a full NUMA task in ms. The task scan period is
1053 * calculated based on the tasks virtual memory size and
1054 * numa_balancing_scan_size.
1056 unsigned int sysctl_numa_balancing_scan_period_min = 1000;
1057 unsigned int sysctl_numa_balancing_scan_period_max = 60000;
1059 /* Portion of address space to scan in MB */
1060 unsigned int sysctl_numa_balancing_scan_size = 256;
1062 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
1063 unsigned int sysctl_numa_balancing_scan_delay = 1000;
1068 spinlock_t lock; /* nr_tasks, tasks */
1073 struct rcu_head rcu;
1074 unsigned long total_faults;
1075 unsigned long max_faults_cpu;
1077 * Faults_cpu is used to decide whether memory should move
1078 * towards the CPU. As a consequence, these stats are weighted
1079 * more by CPU use than by memory faults.
1081 unsigned long *faults_cpu;
1082 unsigned long faults[0];
1086 * For functions that can be called in multiple contexts that permit reading
1087 * ->numa_group (see struct task_struct for locking rules).
1089 static struct numa_group *deref_task_numa_group(struct task_struct *p)
1091 return rcu_dereference_check(p->numa_group, p == current ||
1092 (lockdep_is_held(&task_rq(p)->lock) && !READ_ONCE(p->on_cpu)));
1095 static struct numa_group *deref_curr_numa_group(struct task_struct *p)
1097 return rcu_dereference_protected(p->numa_group, p == current);
1100 static inline unsigned long group_faults_priv(struct numa_group *ng);
1101 static inline unsigned long group_faults_shared(struct numa_group *ng);
1103 static unsigned int task_nr_scan_windows(struct task_struct *p)
1105 unsigned long rss = 0;
1106 unsigned long nr_scan_pages;
1109 * Calculations based on RSS as non-present and empty pages are skipped
1110 * by the PTE scanner and NUMA hinting faults should be trapped based
1113 nr_scan_pages = sysctl_numa_balancing_scan_size << (20 - PAGE_SHIFT);
1114 rss = get_mm_rss(p->mm);
1116 rss = nr_scan_pages;
1118 rss = round_up(rss, nr_scan_pages);
1119 return rss / nr_scan_pages;
1122 /* For sanitys sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */
1123 #define MAX_SCAN_WINDOW 2560
1125 static unsigned int task_scan_min(struct task_struct *p)
1127 unsigned int scan_size = READ_ONCE(sysctl_numa_balancing_scan_size);
1128 unsigned int scan, floor;
1129 unsigned int windows = 1;
1131 if (scan_size < MAX_SCAN_WINDOW)
1132 windows = MAX_SCAN_WINDOW / scan_size;
1133 floor = 1000 / windows;
1135 scan = sysctl_numa_balancing_scan_period_min / task_nr_scan_windows(p);
1136 return max_t(unsigned int, floor, scan);
1139 static unsigned int task_scan_start(struct task_struct *p)
1141 unsigned long smin = task_scan_min(p);
1142 unsigned long period = smin;
1143 struct numa_group *ng;
1145 /* Scale the maximum scan period with the amount of shared memory. */
1147 ng = rcu_dereference(p->numa_group);
1149 unsigned long shared = group_faults_shared(ng);
1150 unsigned long private = group_faults_priv(ng);
1152 period *= atomic_read(&ng->refcount);
1153 period *= shared + 1;
1154 period /= private + shared + 1;
1158 return max(smin, period);
1161 static unsigned int task_scan_max(struct task_struct *p)
1163 unsigned long smin = task_scan_min(p);
1165 struct numa_group *ng;
1167 /* Watch for min being lower than max due to floor calculations */
1168 smax = sysctl_numa_balancing_scan_period_max / task_nr_scan_windows(p);
1170 /* Scale the maximum scan period with the amount of shared memory. */
1171 ng = deref_curr_numa_group(p);
1173 unsigned long shared = group_faults_shared(ng);
1174 unsigned long private = group_faults_priv(ng);
1175 unsigned long period = smax;
1177 period *= atomic_read(&ng->refcount);
1178 period *= shared + 1;
1179 period /= private + shared + 1;
1181 smax = max(smax, period);
1184 return max(smin, smax);
1187 void init_numa_balancing(unsigned long clone_flags, struct task_struct *p)
1190 struct mm_struct *mm = p->mm;
1193 mm_users = atomic_read(&mm->mm_users);
1194 if (mm_users == 1) {
1195 mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
1196 mm->numa_scan_seq = 0;
1200 p->numa_scan_seq = mm ? mm->numa_scan_seq : 0;
1201 p->numa_scan_period = sysctl_numa_balancing_scan_delay;
1202 p->numa_work.next = &p->numa_work;
1203 p->numa_faults = NULL;
1204 RCU_INIT_POINTER(p->numa_group, NULL);
1205 p->last_task_numa_placement = 0;
1206 p->last_sum_exec_runtime = 0;
1208 /* New address space, reset the preferred nid */
1209 if (!(clone_flags & CLONE_VM)) {
1210 p->numa_preferred_nid = -1;
1215 * New thread, keep existing numa_preferred_nid which should be copied
1216 * already by arch_dup_task_struct but stagger when scans start.
1221 delay = min_t(unsigned int, task_scan_max(current),
1222 current->numa_scan_period * mm_users * NSEC_PER_MSEC);
1223 delay += 2 * TICK_NSEC;
1224 p->node_stamp = delay;
1228 static void account_numa_enqueue(struct rq *rq, struct task_struct *p)
1230 rq->nr_numa_running += (p->numa_preferred_nid != -1);
1231 rq->nr_preferred_running += (p->numa_preferred_nid == task_node(p));
1234 static void account_numa_dequeue(struct rq *rq, struct task_struct *p)
1236 rq->nr_numa_running -= (p->numa_preferred_nid != -1);
1237 rq->nr_preferred_running -= (p->numa_preferred_nid == task_node(p));
1240 /* Shared or private faults. */
1241 #define NR_NUMA_HINT_FAULT_TYPES 2
1243 /* Memory and CPU locality */
1244 #define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2)
1246 /* Averaged statistics, and temporary buffers. */
1247 #define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2)
1249 pid_t task_numa_group_id(struct task_struct *p)
1251 struct numa_group *ng;
1255 ng = rcu_dereference(p->numa_group);
1264 * The averaged statistics, shared & private, memory & CPU,
1265 * occupy the first half of the array. The second half of the
1266 * array is for current counters, which are averaged into the
1267 * first set by task_numa_placement.
1269 static inline int task_faults_idx(enum numa_faults_stats s, int nid, int priv)
1271 return NR_NUMA_HINT_FAULT_TYPES * (s * nr_node_ids + nid) + priv;
1274 static inline unsigned long task_faults(struct task_struct *p, int nid)
1276 if (!p->numa_faults)
1279 return p->numa_faults[task_faults_idx(NUMA_MEM, nid, 0)] +
1280 p->numa_faults[task_faults_idx(NUMA_MEM, nid, 1)];
1283 static inline unsigned long group_faults(struct task_struct *p, int nid)
1285 struct numa_group *ng = deref_task_numa_group(p);
1290 return ng->faults[task_faults_idx(NUMA_MEM, nid, 0)] +
1291 ng->faults[task_faults_idx(NUMA_MEM, nid, 1)];
1294 static inline unsigned long group_faults_cpu(struct numa_group *group, int nid)
1296 return group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 0)] +
1297 group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 1)];
1300 static inline unsigned long group_faults_priv(struct numa_group *ng)
1302 unsigned long faults = 0;
1305 for_each_online_node(node) {
1306 faults += ng->faults[task_faults_idx(NUMA_MEM, node, 1)];
1312 static inline unsigned long group_faults_shared(struct numa_group *ng)
1314 unsigned long faults = 0;
1317 for_each_online_node(node) {
1318 faults += ng->faults[task_faults_idx(NUMA_MEM, node, 0)];
1325 * A node triggering more than 1/3 as many NUMA faults as the maximum is
1326 * considered part of a numa group's pseudo-interleaving set. Migrations
1327 * between these nodes are slowed down, to allow things to settle down.
1329 #define ACTIVE_NODE_FRACTION 3
1331 static bool numa_is_active_node(int nid, struct numa_group *ng)
1333 return group_faults_cpu(ng, nid) * ACTIVE_NODE_FRACTION > ng->max_faults_cpu;
1336 /* Handle placement on systems where not all nodes are directly connected. */
1337 static unsigned long score_nearby_nodes(struct task_struct *p, int nid,
1338 int maxdist, bool task)
1340 unsigned long score = 0;
1344 * All nodes are directly connected, and the same distance
1345 * from each other. No need for fancy placement algorithms.
1347 if (sched_numa_topology_type == NUMA_DIRECT)
1351 * This code is called for each node, introducing N^2 complexity,
1352 * which should be ok given the number of nodes rarely exceeds 8.
1354 for_each_online_node(node) {
1355 unsigned long faults;
1356 int dist = node_distance(nid, node);
1359 * The furthest away nodes in the system are not interesting
1360 * for placement; nid was already counted.
1362 if (dist == sched_max_numa_distance || node == nid)
1366 * On systems with a backplane NUMA topology, compare groups
1367 * of nodes, and move tasks towards the group with the most
1368 * memory accesses. When comparing two nodes at distance
1369 * "hoplimit", only nodes closer by than "hoplimit" are part
1370 * of each group. Skip other nodes.
1372 if (sched_numa_topology_type == NUMA_BACKPLANE &&
1376 /* Add up the faults from nearby nodes. */
1378 faults = task_faults(p, node);
1380 faults = group_faults(p, node);
1383 * On systems with a glueless mesh NUMA topology, there are
1384 * no fixed "groups of nodes". Instead, nodes that are not
1385 * directly connected bounce traffic through intermediate
1386 * nodes; a numa_group can occupy any set of nodes.
1387 * The further away a node is, the less the faults count.
1388 * This seems to result in good task placement.
1390 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
1391 faults *= (sched_max_numa_distance - dist);
1392 faults /= (sched_max_numa_distance - LOCAL_DISTANCE);
1402 * These return the fraction of accesses done by a particular task, or
1403 * task group, on a particular numa node. The group weight is given a
1404 * larger multiplier, in order to group tasks together that are almost
1405 * evenly spread out between numa nodes.
1407 static inline unsigned long task_weight(struct task_struct *p, int nid,
1410 unsigned long faults, total_faults;
1412 if (!p->numa_faults)
1415 total_faults = p->total_numa_faults;
1420 faults = task_faults(p, nid);
1421 faults += score_nearby_nodes(p, nid, dist, true);
1423 return 1000 * faults / total_faults;
1426 static inline unsigned long group_weight(struct task_struct *p, int nid,
1429 struct numa_group *ng = deref_task_numa_group(p);
1430 unsigned long faults, total_faults;
1435 total_faults = ng->total_faults;
1440 faults = group_faults(p, nid);
1441 faults += score_nearby_nodes(p, nid, dist, false);
1443 return 1000 * faults / total_faults;
1446 bool should_numa_migrate_memory(struct task_struct *p, struct page * page,
1447 int src_nid, int dst_cpu)
1449 struct numa_group *ng = deref_curr_numa_group(p);
1450 int dst_nid = cpu_to_node(dst_cpu);
1451 int last_cpupid, this_cpupid;
1453 this_cpupid = cpu_pid_to_cpupid(dst_cpu, current->pid);
1454 last_cpupid = page_cpupid_xchg_last(page, this_cpupid);
1457 * Allow first faults or private faults to migrate immediately early in
1458 * the lifetime of a task. The magic number 4 is based on waiting for
1459 * two full passes of the "multi-stage node selection" test that is
1462 if ((p->numa_preferred_nid == -1 || p->numa_scan_seq <= 4) &&
1463 (cpupid_pid_unset(last_cpupid) || cpupid_match_pid(p, last_cpupid)))
1467 * Multi-stage node selection is used in conjunction with a periodic
1468 * migration fault to build a temporal task<->page relation. By using
1469 * a two-stage filter we remove short/unlikely relations.
1471 * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate
1472 * a task's usage of a particular page (n_p) per total usage of this
1473 * page (n_t) (in a given time-span) to a probability.
1475 * Our periodic faults will sample this probability and getting the
1476 * same result twice in a row, given these samples are fully
1477 * independent, is then given by P(n)^2, provided our sample period
1478 * is sufficiently short compared to the usage pattern.
1480 * This quadric squishes small probabilities, making it less likely we
1481 * act on an unlikely task<->page relation.
1483 if (!cpupid_pid_unset(last_cpupid) &&
1484 cpupid_to_nid(last_cpupid) != dst_nid)
1487 /* Always allow migrate on private faults */
1488 if (cpupid_match_pid(p, last_cpupid))
1491 /* A shared fault, but p->numa_group has not been set up yet. */
1496 * Destination node is much more heavily used than the source
1497 * node? Allow migration.
1499 if (group_faults_cpu(ng, dst_nid) > group_faults_cpu(ng, src_nid) *
1500 ACTIVE_NODE_FRACTION)
1504 * Distribute memory according to CPU & memory use on each node,
1505 * with 3/4 hysteresis to avoid unnecessary memory migrations:
1507 * faults_cpu(dst) 3 faults_cpu(src)
1508 * --------------- * - > ---------------
1509 * faults_mem(dst) 4 faults_mem(src)
1511 return group_faults_cpu(ng, dst_nid) * group_faults(p, src_nid) * 3 >
1512 group_faults_cpu(ng, src_nid) * group_faults(p, dst_nid) * 4;
1515 static unsigned long weighted_cpuload(struct rq *rq);
1516 static unsigned long source_load(int cpu, int type);
1517 static unsigned long target_load(int cpu, int type);
1518 static unsigned long capacity_of(int cpu);
1520 /* Cached statistics for all CPUs within a node */
1524 /* Total compute capacity of CPUs on a node */
1525 unsigned long compute_capacity;
1527 unsigned int nr_running;
1531 * XXX borrowed from update_sg_lb_stats
1533 static void update_numa_stats(struct numa_stats *ns, int nid)
1535 int smt, cpu, cpus = 0;
1536 unsigned long capacity;
1538 memset(ns, 0, sizeof(*ns));
1539 for_each_cpu(cpu, cpumask_of_node(nid)) {
1540 struct rq *rq = cpu_rq(cpu);
1542 ns->nr_running += rq->nr_running;
1543 ns->load += weighted_cpuload(rq);
1544 ns->compute_capacity += capacity_of(cpu);
1550 * If we raced with hotplug and there are no CPUs left in our mask
1551 * the @ns structure is NULL'ed and task_numa_compare() will
1552 * not find this node attractive.
1554 * We'll detect a huge imbalance and bail there.
1559 /* smt := ceil(cpus / capacity), assumes: 1 < smt_power < 2 */
1560 smt = DIV_ROUND_UP(SCHED_CAPACITY_SCALE * cpus, ns->compute_capacity);
1561 capacity = cpus / smt; /* cores */
1563 capacity = min_t(unsigned, capacity,
1564 DIV_ROUND_CLOSEST(ns->compute_capacity, SCHED_CAPACITY_SCALE));
1567 struct task_numa_env {
1568 struct task_struct *p;
1570 int src_cpu, src_nid;
1571 int dst_cpu, dst_nid;
1573 struct numa_stats src_stats, dst_stats;
1578 struct task_struct *best_task;
1583 static void task_numa_assign(struct task_numa_env *env,
1584 struct task_struct *p, long imp)
1586 struct rq *rq = cpu_rq(env->dst_cpu);
1588 /* Bail out if run-queue part of active NUMA balance. */
1589 if (xchg(&rq->numa_migrate_on, 1))
1593 * Clear previous best_cpu/rq numa-migrate flag, since task now
1594 * found a better CPU to move/swap.
1596 if (env->best_cpu != -1) {
1597 rq = cpu_rq(env->best_cpu);
1598 WRITE_ONCE(rq->numa_migrate_on, 0);
1602 put_task_struct(env->best_task);
1607 env->best_imp = imp;
1608 env->best_cpu = env->dst_cpu;
1611 static bool load_too_imbalanced(long src_load, long dst_load,
1612 struct task_numa_env *env)
1615 long orig_src_load, orig_dst_load;
1616 long src_capacity, dst_capacity;
1619 * The load is corrected for the CPU capacity available on each node.
1622 * ------------ vs ---------
1623 * src_capacity dst_capacity
1625 src_capacity = env->src_stats.compute_capacity;
1626 dst_capacity = env->dst_stats.compute_capacity;
1628 imb = abs(dst_load * src_capacity - src_load * dst_capacity);
1630 orig_src_load = env->src_stats.load;
1631 orig_dst_load = env->dst_stats.load;
1633 old_imb = abs(orig_dst_load * src_capacity - orig_src_load * dst_capacity);
1635 /* Would this change make things worse? */
1636 return (imb > old_imb);
1640 * Maximum NUMA importance can be 1998 (2*999);
1641 * SMALLIMP @ 30 would be close to 1998/64.
1642 * Used to deter task migration.
1647 * This checks if the overall compute and NUMA accesses of the system would
1648 * be improved if the source tasks was migrated to the target dst_cpu taking
1649 * into account that it might be best if task running on the dst_cpu should
1650 * be exchanged with the source task
1652 static void task_numa_compare(struct task_numa_env *env,
1653 long taskimp, long groupimp, bool maymove)
1655 struct numa_group *cur_ng, *p_ng = deref_curr_numa_group(env->p);
1656 struct rq *dst_rq = cpu_rq(env->dst_cpu);
1657 long imp = p_ng ? groupimp : taskimp;
1658 struct task_struct *cur;
1659 long src_load, dst_load;
1660 int dist = env->dist;
1664 if (READ_ONCE(dst_rq->numa_migrate_on))
1668 cur = task_rcu_dereference(&dst_rq->curr);
1669 if (cur && ((cur->flags & PF_EXITING) || is_idle_task(cur)))
1673 * Because we have preemption enabled we can get migrated around and
1674 * end try selecting ourselves (current == env->p) as a swap candidate.
1680 if (maymove && moveimp >= env->best_imp)
1687 * "imp" is the fault differential for the source task between the
1688 * source and destination node. Calculate the total differential for
1689 * the source task and potential destination task. The more negative
1690 * the value is, the more remote accesses that would be expected to
1691 * be incurred if the tasks were swapped.
1693 /* Skip this swap candidate if cannot move to the source cpu */
1694 if (!cpumask_test_cpu(env->src_cpu, &cur->cpus_allowed))
1698 * If dst and source tasks are in the same NUMA group, or not
1699 * in any group then look only at task weights.
1701 cur_ng = rcu_dereference(cur->numa_group);
1702 if (cur_ng == p_ng) {
1703 imp = taskimp + task_weight(cur, env->src_nid, dist) -
1704 task_weight(cur, env->dst_nid, dist);
1706 * Add some hysteresis to prevent swapping the
1707 * tasks within a group over tiny differences.
1713 * Compare the group weights. If a task is all by itself
1714 * (not part of a group), use the task weight instead.
1717 imp += group_weight(cur, env->src_nid, dist) -
1718 group_weight(cur, env->dst_nid, dist);
1720 imp += task_weight(cur, env->src_nid, dist) -
1721 task_weight(cur, env->dst_nid, dist);
1724 if (maymove && moveimp > imp && moveimp > env->best_imp) {
1731 * If the NUMA importance is less than SMALLIMP,
1732 * task migration might only result in ping pong
1733 * of tasks and also hurt performance due to cache
1736 if (imp < SMALLIMP || imp <= env->best_imp + SMALLIMP / 2)
1740 * In the overloaded case, try and keep the load balanced.
1742 load = task_h_load(env->p) - task_h_load(cur);
1746 dst_load = env->dst_stats.load + load;
1747 src_load = env->src_stats.load - load;
1749 if (load_too_imbalanced(src_load, dst_load, env))
1754 * One idle CPU per node is evaluated for a task numa move.
1755 * Call select_idle_sibling to maybe find a better one.
1759 * select_idle_siblings() uses an per-CPU cpumask that
1760 * can be used from IRQ context.
1762 local_irq_disable();
1763 env->dst_cpu = select_idle_sibling(env->p, env->src_cpu,
1768 task_numa_assign(env, cur, imp);
1773 static void task_numa_find_cpu(struct task_numa_env *env,
1774 long taskimp, long groupimp)
1776 long src_load, dst_load, load;
1777 bool maymove = false;
1780 load = task_h_load(env->p);
1781 dst_load = env->dst_stats.load + load;
1782 src_load = env->src_stats.load - load;
1785 * If the improvement from just moving env->p direction is better
1786 * than swapping tasks around, check if a move is possible.
1788 maymove = !load_too_imbalanced(src_load, dst_load, env);
1790 for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) {
1791 /* Skip this CPU if the source task cannot migrate */
1792 if (!cpumask_test_cpu(cpu, &env->p->cpus_allowed))
1796 task_numa_compare(env, taskimp, groupimp, maymove);
1800 static int task_numa_migrate(struct task_struct *p)
1802 struct task_numa_env env = {
1805 .src_cpu = task_cpu(p),
1806 .src_nid = task_node(p),
1808 .imbalance_pct = 112,
1814 unsigned long taskweight, groupweight;
1815 struct sched_domain *sd;
1816 long taskimp, groupimp;
1817 struct numa_group *ng;
1822 * Pick the lowest SD_NUMA domain, as that would have the smallest
1823 * imbalance and would be the first to start moving tasks about.
1825 * And we want to avoid any moving of tasks about, as that would create
1826 * random movement of tasks -- counter the numa conditions we're trying
1830 sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu));
1832 env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2;
1836 * Cpusets can break the scheduler domain tree into smaller
1837 * balance domains, some of which do not cross NUMA boundaries.
1838 * Tasks that are "trapped" in such domains cannot be migrated
1839 * elsewhere, so there is no point in (re)trying.
1841 if (unlikely(!sd)) {
1842 sched_setnuma(p, task_node(p));
1846 env.dst_nid = p->numa_preferred_nid;
1847 dist = env.dist = node_distance(env.src_nid, env.dst_nid);
1848 taskweight = task_weight(p, env.src_nid, dist);
1849 groupweight = group_weight(p, env.src_nid, dist);
1850 update_numa_stats(&env.src_stats, env.src_nid);
1851 taskimp = task_weight(p, env.dst_nid, dist) - taskweight;
1852 groupimp = group_weight(p, env.dst_nid, dist) - groupweight;
1853 update_numa_stats(&env.dst_stats, env.dst_nid);
1855 /* Try to find a spot on the preferred nid. */
1856 task_numa_find_cpu(&env, taskimp, groupimp);
1859 * Look at other nodes in these cases:
1860 * - there is no space available on the preferred_nid
1861 * - the task is part of a numa_group that is interleaved across
1862 * multiple NUMA nodes; in order to better consolidate the group,
1863 * we need to check other locations.
1865 ng = deref_curr_numa_group(p);
1866 if (env.best_cpu == -1 || (ng && ng->active_nodes > 1)) {
1867 for_each_online_node(nid) {
1868 if (nid == env.src_nid || nid == p->numa_preferred_nid)
1871 dist = node_distance(env.src_nid, env.dst_nid);
1872 if (sched_numa_topology_type == NUMA_BACKPLANE &&
1874 taskweight = task_weight(p, env.src_nid, dist);
1875 groupweight = group_weight(p, env.src_nid, dist);
1878 /* Only consider nodes where both task and groups benefit */
1879 taskimp = task_weight(p, nid, dist) - taskweight;
1880 groupimp = group_weight(p, nid, dist) - groupweight;
1881 if (taskimp < 0 && groupimp < 0)
1886 update_numa_stats(&env.dst_stats, env.dst_nid);
1887 task_numa_find_cpu(&env, taskimp, groupimp);
1892 * If the task is part of a workload that spans multiple NUMA nodes,
1893 * and is migrating into one of the workload's active nodes, remember
1894 * this node as the task's preferred numa node, so the workload can
1896 * A task that migrated to a second choice node will be better off
1897 * trying for a better one later. Do not set the preferred node here.
1900 if (env.best_cpu == -1)
1903 nid = cpu_to_node(env.best_cpu);
1905 if (nid != p->numa_preferred_nid)
1906 sched_setnuma(p, nid);
1909 /* No better CPU than the current one was found. */
1910 if (env.best_cpu == -1)
1913 best_rq = cpu_rq(env.best_cpu);
1914 if (env.best_task == NULL) {
1915 ret = migrate_task_to(p, env.best_cpu);
1916 WRITE_ONCE(best_rq->numa_migrate_on, 0);
1918 trace_sched_stick_numa(p, env.src_cpu, env.best_cpu);
1922 ret = migrate_swap(p, env.best_task, env.best_cpu, env.src_cpu);
1923 WRITE_ONCE(best_rq->numa_migrate_on, 0);
1926 trace_sched_stick_numa(p, env.src_cpu, task_cpu(env.best_task));
1927 put_task_struct(env.best_task);
1931 /* Attempt to migrate a task to a CPU on the preferred node. */
1932 static void numa_migrate_preferred(struct task_struct *p)
1934 unsigned long interval = HZ;
1936 /* This task has no NUMA fault statistics yet */
1937 if (unlikely(p->numa_preferred_nid == -1 || !p->numa_faults))
1940 /* Periodically retry migrating the task to the preferred node */
1941 interval = min(interval, msecs_to_jiffies(p->numa_scan_period) / 16);
1942 p->numa_migrate_retry = jiffies + interval;
1944 /* Success if task is already running on preferred CPU */
1945 if (task_node(p) == p->numa_preferred_nid)
1948 /* Otherwise, try migrate to a CPU on the preferred node */
1949 task_numa_migrate(p);
1953 * Find out how many nodes on the workload is actively running on. Do this by
1954 * tracking the nodes from which NUMA hinting faults are triggered. This can
1955 * be different from the set of nodes where the workload's memory is currently
1958 static void numa_group_count_active_nodes(struct numa_group *numa_group)
1960 unsigned long faults, max_faults = 0;
1961 int nid, active_nodes = 0;
1963 for_each_online_node(nid) {
1964 faults = group_faults_cpu(numa_group, nid);
1965 if (faults > max_faults)
1966 max_faults = faults;
1969 for_each_online_node(nid) {
1970 faults = group_faults_cpu(numa_group, nid);
1971 if (faults * ACTIVE_NODE_FRACTION > max_faults)
1975 numa_group->max_faults_cpu = max_faults;
1976 numa_group->active_nodes = active_nodes;
1980 * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
1981 * increments. The more local the fault statistics are, the higher the scan
1982 * period will be for the next scan window. If local/(local+remote) ratio is
1983 * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS)
1984 * the scan period will decrease. Aim for 70% local accesses.
1986 #define NUMA_PERIOD_SLOTS 10
1987 #define NUMA_PERIOD_THRESHOLD 7
1990 * Increase the scan period (slow down scanning) if the majority of
1991 * our memory is already on our local node, or if the majority of
1992 * the page accesses are shared with other processes.
1993 * Otherwise, decrease the scan period.
1995 static void update_task_scan_period(struct task_struct *p,
1996 unsigned long shared, unsigned long private)
1998 unsigned int period_slot;
1999 int lr_ratio, ps_ratio;
2002 unsigned long remote = p->numa_faults_locality[0];
2003 unsigned long local = p->numa_faults_locality[1];
2006 * If there were no record hinting faults then either the task is
2007 * completely idle or all activity is areas that are not of interest
2008 * to automatic numa balancing. Related to that, if there were failed
2009 * migration then it implies we are migrating too quickly or the local
2010 * node is overloaded. In either case, scan slower
2012 if (local + shared == 0 || p->numa_faults_locality[2]) {
2013 p->numa_scan_period = min(p->numa_scan_period_max,
2014 p->numa_scan_period << 1);
2016 p->mm->numa_next_scan = jiffies +
2017 msecs_to_jiffies(p->numa_scan_period);
2023 * Prepare to scale scan period relative to the current period.
2024 * == NUMA_PERIOD_THRESHOLD scan period stays the same
2025 * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
2026 * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
2028 period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS);
2029 lr_ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote);
2030 ps_ratio = (private * NUMA_PERIOD_SLOTS) / (private + shared);
2032 if (ps_ratio >= NUMA_PERIOD_THRESHOLD) {
2034 * Most memory accesses are local. There is no need to
2035 * do fast NUMA scanning, since memory is already local.
2037 int slot = ps_ratio - NUMA_PERIOD_THRESHOLD;
2040 diff = slot * period_slot;
2041 } else if (lr_ratio >= NUMA_PERIOD_THRESHOLD) {
2043 * Most memory accesses are shared with other tasks.
2044 * There is no point in continuing fast NUMA scanning,
2045 * since other tasks may just move the memory elsewhere.
2047 int slot = lr_ratio - NUMA_PERIOD_THRESHOLD;
2050 diff = slot * period_slot;
2053 * Private memory faults exceed (SLOTS-THRESHOLD)/SLOTS,
2054 * yet they are not on the local NUMA node. Speed up
2055 * NUMA scanning to get the memory moved over.
2057 int ratio = max(lr_ratio, ps_ratio);
2058 diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot;
2061 p->numa_scan_period = clamp(p->numa_scan_period + diff,
2062 task_scan_min(p), task_scan_max(p));
2063 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
2067 * Get the fraction of time the task has been running since the last
2068 * NUMA placement cycle. The scheduler keeps similar statistics, but
2069 * decays those on a 32ms period, which is orders of magnitude off
2070 * from the dozens-of-seconds NUMA balancing period. Use the scheduler
2071 * stats only if the task is so new there are no NUMA statistics yet.
2073 static u64 numa_get_avg_runtime(struct task_struct *p, u64 *period)
2075 u64 runtime, delta, now;
2076 /* Use the start of this time slice to avoid calculations. */
2077 now = p->se.exec_start;
2078 runtime = p->se.sum_exec_runtime;
2080 if (p->last_task_numa_placement) {
2081 delta = runtime - p->last_sum_exec_runtime;
2082 *period = now - p->last_task_numa_placement;
2084 /* Avoid time going backwards, prevent potential divide error: */
2085 if (unlikely((s64)*period < 0))
2088 delta = p->se.avg.load_sum;
2089 *period = LOAD_AVG_MAX;
2092 p->last_sum_exec_runtime = runtime;
2093 p->last_task_numa_placement = now;
2099 * Determine the preferred nid for a task in a numa_group. This needs to
2100 * be done in a way that produces consistent results with group_weight,
2101 * otherwise workloads might not converge.
2103 static int preferred_group_nid(struct task_struct *p, int nid)
2108 /* Direct connections between all NUMA nodes. */
2109 if (sched_numa_topology_type == NUMA_DIRECT)
2113 * On a system with glueless mesh NUMA topology, group_weight
2114 * scores nodes according to the number of NUMA hinting faults on
2115 * both the node itself, and on nearby nodes.
2117 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
2118 unsigned long score, max_score = 0;
2119 int node, max_node = nid;
2121 dist = sched_max_numa_distance;
2123 for_each_online_node(node) {
2124 score = group_weight(p, node, dist);
2125 if (score > max_score) {
2134 * Finding the preferred nid in a system with NUMA backplane
2135 * interconnect topology is more involved. The goal is to locate
2136 * tasks from numa_groups near each other in the system, and
2137 * untangle workloads from different sides of the system. This requires
2138 * searching down the hierarchy of node groups, recursively searching
2139 * inside the highest scoring group of nodes. The nodemask tricks
2140 * keep the complexity of the search down.
2142 nodes = node_online_map;
2143 for (dist = sched_max_numa_distance; dist > LOCAL_DISTANCE; dist--) {
2144 unsigned long max_faults = 0;
2145 nodemask_t max_group = NODE_MASK_NONE;
2148 /* Are there nodes at this distance from each other? */
2149 if (!find_numa_distance(dist))
2152 for_each_node_mask(a, nodes) {
2153 unsigned long faults = 0;
2154 nodemask_t this_group;
2155 nodes_clear(this_group);
2157 /* Sum group's NUMA faults; includes a==b case. */
2158 for_each_node_mask(b, nodes) {
2159 if (node_distance(a, b) < dist) {
2160 faults += group_faults(p, b);
2161 node_set(b, this_group);
2162 node_clear(b, nodes);
2166 /* Remember the top group. */
2167 if (faults > max_faults) {
2168 max_faults = faults;
2169 max_group = this_group;
2171 * subtle: at the smallest distance there is
2172 * just one node left in each "group", the
2173 * winner is the preferred nid.
2178 /* Next round, evaluate the nodes within max_group. */
2186 static void task_numa_placement(struct task_struct *p)
2188 int seq, nid, max_nid = -1;
2189 unsigned long max_faults = 0;
2190 unsigned long fault_types[2] = { 0, 0 };
2191 unsigned long total_faults;
2192 u64 runtime, period;
2193 spinlock_t *group_lock = NULL;
2194 struct numa_group *ng;
2197 * The p->mm->numa_scan_seq field gets updated without
2198 * exclusive access. Use READ_ONCE() here to ensure
2199 * that the field is read in a single access:
2201 seq = READ_ONCE(p->mm->numa_scan_seq);
2202 if (p->numa_scan_seq == seq)
2204 p->numa_scan_seq = seq;
2205 p->numa_scan_period_max = task_scan_max(p);
2207 total_faults = p->numa_faults_locality[0] +
2208 p->numa_faults_locality[1];
2209 runtime = numa_get_avg_runtime(p, &period);
2211 /* If the task is part of a group prevent parallel updates to group stats */
2212 ng = deref_curr_numa_group(p);
2214 group_lock = &ng->lock;
2215 spin_lock_irq(group_lock);
2218 /* Find the node with the highest number of faults */
2219 for_each_online_node(nid) {
2220 /* Keep track of the offsets in numa_faults array */
2221 int mem_idx, membuf_idx, cpu_idx, cpubuf_idx;
2222 unsigned long faults = 0, group_faults = 0;
2225 for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) {
2226 long diff, f_diff, f_weight;
2228 mem_idx = task_faults_idx(NUMA_MEM, nid, priv);
2229 membuf_idx = task_faults_idx(NUMA_MEMBUF, nid, priv);
2230 cpu_idx = task_faults_idx(NUMA_CPU, nid, priv);
2231 cpubuf_idx = task_faults_idx(NUMA_CPUBUF, nid, priv);
2233 /* Decay existing window, copy faults since last scan */
2234 diff = p->numa_faults[membuf_idx] - p->numa_faults[mem_idx] / 2;
2235 fault_types[priv] += p->numa_faults[membuf_idx];
2236 p->numa_faults[membuf_idx] = 0;
2239 * Normalize the faults_from, so all tasks in a group
2240 * count according to CPU use, instead of by the raw
2241 * number of faults. Tasks with little runtime have
2242 * little over-all impact on throughput, and thus their
2243 * faults are less important.
2245 f_weight = div64_u64(runtime << 16, period + 1);
2246 f_weight = (f_weight * p->numa_faults[cpubuf_idx]) /
2248 f_diff = f_weight - p->numa_faults[cpu_idx] / 2;
2249 p->numa_faults[cpubuf_idx] = 0;
2251 p->numa_faults[mem_idx] += diff;
2252 p->numa_faults[cpu_idx] += f_diff;
2253 faults += p->numa_faults[mem_idx];
2254 p->total_numa_faults += diff;
2257 * safe because we can only change our own group
2259 * mem_idx represents the offset for a given
2260 * nid and priv in a specific region because it
2261 * is at the beginning of the numa_faults array.
2263 ng->faults[mem_idx] += diff;
2264 ng->faults_cpu[mem_idx] += f_diff;
2265 ng->total_faults += diff;
2266 group_faults += ng->faults[mem_idx];
2271 if (faults > max_faults) {
2272 max_faults = faults;
2275 } else if (group_faults > max_faults) {
2276 max_faults = group_faults;
2282 numa_group_count_active_nodes(ng);
2283 spin_unlock_irq(group_lock);
2284 max_nid = preferred_group_nid(p, max_nid);
2288 /* Set the new preferred node */
2289 if (max_nid != p->numa_preferred_nid)
2290 sched_setnuma(p, max_nid);
2293 update_task_scan_period(p, fault_types[0], fault_types[1]);
2296 static inline int get_numa_group(struct numa_group *grp)
2298 return atomic_inc_not_zero(&grp->refcount);
2301 static inline void put_numa_group(struct numa_group *grp)
2303 if (atomic_dec_and_test(&grp->refcount))
2304 kfree_rcu(grp, rcu);
2307 static void task_numa_group(struct task_struct *p, int cpupid, int flags,
2310 struct numa_group *grp, *my_grp;
2311 struct task_struct *tsk;
2313 int cpu = cpupid_to_cpu(cpupid);
2316 if (unlikely(!deref_curr_numa_group(p))) {
2317 unsigned int size = sizeof(struct numa_group) +
2318 4*nr_node_ids*sizeof(unsigned long);
2320 grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
2324 atomic_set(&grp->refcount, 1);
2325 grp->active_nodes = 1;
2326 grp->max_faults_cpu = 0;
2327 spin_lock_init(&grp->lock);
2329 /* Second half of the array tracks nids where faults happen */
2330 grp->faults_cpu = grp->faults + NR_NUMA_HINT_FAULT_TYPES *
2333 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2334 grp->faults[i] = p->numa_faults[i];
2336 grp->total_faults = p->total_numa_faults;
2339 rcu_assign_pointer(p->numa_group, grp);
2343 tsk = READ_ONCE(cpu_rq(cpu)->curr);
2345 if (!cpupid_match_pid(tsk, cpupid))
2348 grp = rcu_dereference(tsk->numa_group);
2352 my_grp = deref_curr_numa_group(p);
2357 * Only join the other group if its bigger; if we're the bigger group,
2358 * the other task will join us.
2360 if (my_grp->nr_tasks > grp->nr_tasks)
2364 * Tie-break on the grp address.
2366 if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp)
2369 /* Always join threads in the same process. */
2370 if (tsk->mm == current->mm)
2373 /* Simple filter to avoid false positives due to PID collisions */
2374 if (flags & TNF_SHARED)
2377 /* Update priv based on whether false sharing was detected */
2380 if (join && !get_numa_group(grp))
2388 BUG_ON(irqs_disabled());
2389 double_lock_irq(&my_grp->lock, &grp->lock);
2391 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) {
2392 my_grp->faults[i] -= p->numa_faults[i];
2393 grp->faults[i] += p->numa_faults[i];
2395 my_grp->total_faults -= p->total_numa_faults;
2396 grp->total_faults += p->total_numa_faults;
2401 spin_unlock(&my_grp->lock);
2402 spin_unlock_irq(&grp->lock);
2404 rcu_assign_pointer(p->numa_group, grp);
2406 put_numa_group(my_grp);
2415 * Get rid of NUMA staticstics associated with a task (either current or dead).
2416 * If @final is set, the task is dead and has reached refcount zero, so we can
2417 * safely free all relevant data structures. Otherwise, there might be
2418 * concurrent reads from places like load balancing and procfs, and we should
2419 * reset the data back to default state without freeing ->numa_faults.
2421 void task_numa_free(struct task_struct *p, bool final)
2423 /* safe: p either is current or is being freed by current */
2424 struct numa_group *grp = rcu_dereference_raw(p->numa_group);
2425 unsigned long *numa_faults = p->numa_faults;
2426 unsigned long flags;
2433 spin_lock_irqsave(&grp->lock, flags);
2434 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2435 grp->faults[i] -= p->numa_faults[i];
2436 grp->total_faults -= p->total_numa_faults;
2439 spin_unlock_irqrestore(&grp->lock, flags);
2440 RCU_INIT_POINTER(p->numa_group, NULL);
2441 put_numa_group(grp);
2445 p->numa_faults = NULL;
2448 p->total_numa_faults = 0;
2449 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2455 * Got a PROT_NONE fault for a page on @node.
2457 void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags)
2459 struct task_struct *p = current;
2460 bool migrated = flags & TNF_MIGRATED;
2461 int cpu_node = task_node(current);
2462 int local = !!(flags & TNF_FAULT_LOCAL);
2463 struct numa_group *ng;
2466 if (!static_branch_likely(&sched_numa_balancing))
2469 /* for example, ksmd faulting in a user's mm */
2473 /* Allocate buffer to track faults on a per-node basis */
2474 if (unlikely(!p->numa_faults)) {
2475 int size = sizeof(*p->numa_faults) *
2476 NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids;
2478 p->numa_faults = kzalloc(size, GFP_KERNEL|__GFP_NOWARN);
2479 if (!p->numa_faults)
2482 p->total_numa_faults = 0;
2483 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
2487 * First accesses are treated as private, otherwise consider accesses
2488 * to be private if the accessing pid has not changed
2490 if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) {
2493 priv = cpupid_match_pid(p, last_cpupid);
2494 if (!priv && !(flags & TNF_NO_GROUP))
2495 task_numa_group(p, last_cpupid, flags, &priv);
2499 * If a workload spans multiple NUMA nodes, a shared fault that
2500 * occurs wholly within the set of nodes that the workload is
2501 * actively using should be counted as local. This allows the
2502 * scan rate to slow down when a workload has settled down.
2504 ng = deref_curr_numa_group(p);
2505 if (!priv && !local && ng && ng->active_nodes > 1 &&
2506 numa_is_active_node(cpu_node, ng) &&
2507 numa_is_active_node(mem_node, ng))
2511 * Retry task to preferred node migration periodically, in case it
2512 * case it previously failed, or the scheduler moved us.
2514 if (time_after(jiffies, p->numa_migrate_retry)) {
2515 task_numa_placement(p);
2516 numa_migrate_preferred(p);
2520 p->numa_pages_migrated += pages;
2521 if (flags & TNF_MIGRATE_FAIL)
2522 p->numa_faults_locality[2] += pages;
2524 p->numa_faults[task_faults_idx(NUMA_MEMBUF, mem_node, priv)] += pages;
2525 p->numa_faults[task_faults_idx(NUMA_CPUBUF, cpu_node, priv)] += pages;
2526 p->numa_faults_locality[local] += pages;
2529 static void reset_ptenuma_scan(struct task_struct *p)
2532 * We only did a read acquisition of the mmap sem, so
2533 * p->mm->numa_scan_seq is written to without exclusive access
2534 * and the update is not guaranteed to be atomic. That's not
2535 * much of an issue though, since this is just used for
2536 * statistical sampling. Use READ_ONCE/WRITE_ONCE, which are not
2537 * expensive, to avoid any form of compiler optimizations:
2539 WRITE_ONCE(p->mm->numa_scan_seq, READ_ONCE(p->mm->numa_scan_seq) + 1);
2540 p->mm->numa_scan_offset = 0;
2544 * The expensive part of numa migration is done from task_work context.
2545 * Triggered from task_tick_numa().
2547 void task_numa_work(struct callback_head *work)
2549 unsigned long migrate, next_scan, now = jiffies;
2550 struct task_struct *p = current;
2551 struct mm_struct *mm = p->mm;
2552 u64 runtime = p->se.sum_exec_runtime;
2553 struct vm_area_struct *vma;
2554 unsigned long start, end;
2555 unsigned long nr_pte_updates = 0;
2556 long pages, virtpages;
2558 SCHED_WARN_ON(p != container_of(work, struct task_struct, numa_work));
2560 work->next = work; /* protect against double add */
2562 * Who cares about NUMA placement when they're dying.
2564 * NOTE: make sure not to dereference p->mm before this check,
2565 * exit_task_work() happens _after_ exit_mm() so we could be called
2566 * without p->mm even though we still had it when we enqueued this
2569 if (p->flags & PF_EXITING)
2572 if (!mm->numa_next_scan) {
2573 mm->numa_next_scan = now +
2574 msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2578 * Enforce maximal scan/migration frequency..
2580 migrate = mm->numa_next_scan;
2581 if (time_before(now, migrate))
2584 if (p->numa_scan_period == 0) {
2585 p->numa_scan_period_max = task_scan_max(p);
2586 p->numa_scan_period = task_scan_start(p);
2589 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
2590 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
2594 * Delay this task enough that another task of this mm will likely win
2595 * the next time around.
2597 p->node_stamp += 2 * TICK_NSEC;
2599 start = mm->numa_scan_offset;
2600 pages = sysctl_numa_balancing_scan_size;
2601 pages <<= 20 - PAGE_SHIFT; /* MB in pages */
2602 virtpages = pages * 8; /* Scan up to this much virtual space */
2607 if (!down_read_trylock(&mm->mmap_sem))
2609 vma = find_vma(mm, start);
2611 reset_ptenuma_scan(p);
2615 for (; vma; vma = vma->vm_next) {
2616 if (!vma_migratable(vma) || !vma_policy_mof(vma) ||
2617 is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_MIXEDMAP)) {
2622 * Shared library pages mapped by multiple processes are not
2623 * migrated as it is expected they are cache replicated. Avoid
2624 * hinting faults in read-only file-backed mappings or the vdso
2625 * as migrating the pages will be of marginal benefit.
2628 (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ)))
2632 * Skip inaccessible VMAs to avoid any confusion between
2633 * PROT_NONE and NUMA hinting ptes
2635 if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)))
2639 start = max(start, vma->vm_start);
2640 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
2641 end = min(end, vma->vm_end);
2642 nr_pte_updates = change_prot_numa(vma, start, end);
2645 * Try to scan sysctl_numa_balancing_size worth of
2646 * hpages that have at least one present PTE that
2647 * is not already pte-numa. If the VMA contains
2648 * areas that are unused or already full of prot_numa
2649 * PTEs, scan up to virtpages, to skip through those
2653 pages -= (end - start) >> PAGE_SHIFT;
2654 virtpages -= (end - start) >> PAGE_SHIFT;
2657 if (pages <= 0 || virtpages <= 0)
2661 } while (end != vma->vm_end);
2666 * It is possible to reach the end of the VMA list but the last few
2667 * VMAs are not guaranteed to the vma_migratable. If they are not, we
2668 * would find the !migratable VMA on the next scan but not reset the
2669 * scanner to the start so check it now.
2672 mm->numa_scan_offset = start;
2674 reset_ptenuma_scan(p);
2675 up_read(&mm->mmap_sem);
2678 * Make sure tasks use at least 32x as much time to run other code
2679 * than they used here, to limit NUMA PTE scanning overhead to 3% max.
2680 * Usually update_task_scan_period slows down scanning enough; on an
2681 * overloaded system we need to limit overhead on a per task basis.
2683 if (unlikely(p->se.sum_exec_runtime != runtime)) {
2684 u64 diff = p->se.sum_exec_runtime - runtime;
2685 p->node_stamp += 32 * diff;
2690 * Drive the periodic memory faults..
2692 void task_tick_numa(struct rq *rq, struct task_struct *curr)
2694 struct callback_head *work = &curr->numa_work;
2698 * We don't care about NUMA placement if we don't have memory.
2700 if ((curr->flags & (PF_EXITING | PF_KTHREAD)) || work->next != work)
2704 * Using runtime rather than walltime has the dual advantage that
2705 * we (mostly) drive the selection from busy threads and that the
2706 * task needs to have done some actual work before we bother with
2709 now = curr->se.sum_exec_runtime;
2710 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
2712 if (now > curr->node_stamp + period) {
2713 if (!curr->node_stamp)
2714 curr->numa_scan_period = task_scan_start(curr);
2715 curr->node_stamp += period;
2717 if (!time_before(jiffies, curr->mm->numa_next_scan)) {
2718 init_task_work(work, task_numa_work); /* TODO: move this into sched_fork() */
2719 task_work_add(curr, work, true);
2724 static void update_scan_period(struct task_struct *p, int new_cpu)
2726 int src_nid = cpu_to_node(task_cpu(p));
2727 int dst_nid = cpu_to_node(new_cpu);
2729 if (!static_branch_likely(&sched_numa_balancing))
2732 if (!p->mm || !p->numa_faults || (p->flags & PF_EXITING))
2735 if (src_nid == dst_nid)
2739 * Allow resets if faults have been trapped before one scan
2740 * has completed. This is most likely due to a new task that
2741 * is pulled cross-node due to wakeups or load balancing.
2743 if (p->numa_scan_seq) {
2745 * Avoid scan adjustments if moving to the preferred
2746 * node or if the task was not previously running on
2747 * the preferred node.
2749 if (dst_nid == p->numa_preferred_nid ||
2750 (p->numa_preferred_nid != -1 && src_nid != p->numa_preferred_nid))
2754 p->numa_scan_period = task_scan_start(p);
2758 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
2762 static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p)
2766 static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p)
2770 static inline void update_scan_period(struct task_struct *p, int new_cpu)
2774 #endif /* CONFIG_NUMA_BALANCING */
2777 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2779 update_load_add(&cfs_rq->load, se->load.weight);
2780 if (!parent_entity(se))
2781 update_load_add(&rq_of(cfs_rq)->load, se->load.weight);
2783 if (entity_is_task(se)) {
2784 struct rq *rq = rq_of(cfs_rq);
2786 account_numa_enqueue(rq, task_of(se));
2787 list_add(&se->group_node, &rq->cfs_tasks);
2790 cfs_rq->nr_running++;
2794 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2796 update_load_sub(&cfs_rq->load, se->load.weight);
2797 if (!parent_entity(se))
2798 update_load_sub(&rq_of(cfs_rq)->load, se->load.weight);
2800 if (entity_is_task(se)) {
2801 account_numa_dequeue(rq_of(cfs_rq), task_of(se));
2802 list_del_init(&se->group_node);
2805 cfs_rq->nr_running--;
2809 * Signed add and clamp on underflow.
2811 * Explicitly do a load-store to ensure the intermediate value never hits
2812 * memory. This allows lockless observations without ever seeing the negative
2815 #define add_positive(_ptr, _val) do { \
2816 typeof(_ptr) ptr = (_ptr); \
2817 typeof(_val) val = (_val); \
2818 typeof(*ptr) res, var = READ_ONCE(*ptr); \
2822 if (val < 0 && res > var) \
2825 WRITE_ONCE(*ptr, res); \
2829 * Unsigned subtract and clamp on underflow.
2831 * Explicitly do a load-store to ensure the intermediate value never hits
2832 * memory. This allows lockless observations without ever seeing the negative
2835 #define sub_positive(_ptr, _val) do { \
2836 typeof(_ptr) ptr = (_ptr); \
2837 typeof(*ptr) val = (_val); \
2838 typeof(*ptr) res, var = READ_ONCE(*ptr); \
2842 WRITE_ONCE(*ptr, res); \
2847 enqueue_runnable_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2849 cfs_rq->runnable_weight += se->runnable_weight;
2851 cfs_rq->avg.runnable_load_avg += se->avg.runnable_load_avg;
2852 cfs_rq->avg.runnable_load_sum += se_runnable(se) * se->avg.runnable_load_sum;
2856 dequeue_runnable_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2858 cfs_rq->runnable_weight -= se->runnable_weight;
2860 sub_positive(&cfs_rq->avg.runnable_load_avg, se->avg.runnable_load_avg);
2861 sub_positive(&cfs_rq->avg.runnable_load_sum,
2862 se_runnable(se) * se->avg.runnable_load_sum);
2866 enqueue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2868 cfs_rq->avg.load_avg += se->avg.load_avg;
2869 cfs_rq->avg.load_sum += se_weight(se) * se->avg.load_sum;
2873 dequeue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
2875 sub_positive(&cfs_rq->avg.load_avg, se->avg.load_avg);
2876 sub_positive(&cfs_rq->avg.load_sum, se_weight(se) * se->avg.load_sum);
2880 enqueue_runnable_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
2882 dequeue_runnable_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
2884 enqueue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
2886 dequeue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
2889 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
2890 unsigned long weight, unsigned long runnable)
2893 /* commit outstanding execution time */
2894 if (cfs_rq->curr == se)
2895 update_curr(cfs_rq);
2896 account_entity_dequeue(cfs_rq, se);
2897 dequeue_runnable_load_avg(cfs_rq, se);
2899 dequeue_load_avg(cfs_rq, se);
2901 se->runnable_weight = runnable;
2902 update_load_set(&se->load, weight);
2906 u32 divider = LOAD_AVG_MAX - 1024 + se->avg.period_contrib;
2908 se->avg.load_avg = div_u64(se_weight(se) * se->avg.load_sum, divider);
2909 se->avg.runnable_load_avg =
2910 div_u64(se_runnable(se) * se->avg.runnable_load_sum, divider);
2914 enqueue_load_avg(cfs_rq, se);
2916 account_entity_enqueue(cfs_rq, se);
2917 enqueue_runnable_load_avg(cfs_rq, se);
2921 void reweight_task(struct task_struct *p, int prio)
2923 struct sched_entity *se = &p->se;
2924 struct cfs_rq *cfs_rq = cfs_rq_of(se);
2925 struct load_weight *load = &se->load;
2926 unsigned long weight = scale_load(sched_prio_to_weight[prio]);
2928 reweight_entity(cfs_rq, se, weight, weight);
2929 load->inv_weight = sched_prio_to_wmult[prio];
2932 #ifdef CONFIG_FAIR_GROUP_SCHED
2935 * All this does is approximate the hierarchical proportion which includes that
2936 * global sum we all love to hate.
2938 * That is, the weight of a group entity, is the proportional share of the
2939 * group weight based on the group runqueue weights. That is:
2941 * tg->weight * grq->load.weight
2942 * ge->load.weight = ----------------------------- (1)
2943 * \Sum grq->load.weight
2945 * Now, because computing that sum is prohibitively expensive to compute (been
2946 * there, done that) we approximate it with this average stuff. The average
2947 * moves slower and therefore the approximation is cheaper and more stable.
2949 * So instead of the above, we substitute:
2951 * grq->load.weight -> grq->avg.load_avg (2)
2953 * which yields the following:
2955 * tg->weight * grq->avg.load_avg
2956 * ge->load.weight = ------------------------------ (3)
2959 * Where: tg->load_avg ~= \Sum grq->avg.load_avg
2961 * That is shares_avg, and it is right (given the approximation (2)).
2963 * The problem with it is that because the average is slow -- it was designed
2964 * to be exactly that of course -- this leads to transients in boundary
2965 * conditions. In specific, the case where the group was idle and we start the
2966 * one task. It takes time for our CPU's grq->avg.load_avg to build up,
2967 * yielding bad latency etc..
2969 * Now, in that special case (1) reduces to:
2971 * tg->weight * grq->load.weight
2972 * ge->load.weight = ----------------------------- = tg->weight (4)
2975 * That is, the sum collapses because all other CPUs are idle; the UP scenario.
2977 * So what we do is modify our approximation (3) to approach (4) in the (near)
2982 * tg->weight * grq->load.weight
2983 * --------------------------------------------------- (5)
2984 * tg->load_avg - grq->avg.load_avg + grq->load.weight
2986 * But because grq->load.weight can drop to 0, resulting in a divide by zero,
2987 * we need to use grq->avg.load_avg as its lower bound, which then gives:
2990 * tg->weight * grq->load.weight
2991 * ge->load.weight = ----------------------------- (6)
2996 * tg_load_avg' = tg->load_avg - grq->avg.load_avg +
2997 * max(grq->load.weight, grq->avg.load_avg)
2999 * And that is shares_weight and is icky. In the (near) UP case it approaches
3000 * (4) while in the normal case it approaches (3). It consistently
3001 * overestimates the ge->load.weight and therefore:
3003 * \Sum ge->load.weight >= tg->weight
3007 static long calc_group_shares(struct cfs_rq *cfs_rq)
3009 long tg_weight, tg_shares, load, shares;
3010 struct task_group *tg = cfs_rq->tg;
3012 tg_shares = READ_ONCE(tg->shares);
3014 load = max(scale_load_down(cfs_rq->load.weight), cfs_rq->avg.load_avg);
3016 tg_weight = atomic_long_read(&tg->load_avg);
3018 /* Ensure tg_weight >= load */
3019 tg_weight -= cfs_rq->tg_load_avg_contrib;
3022 shares = (tg_shares * load);
3024 shares /= tg_weight;
3027 * MIN_SHARES has to be unscaled here to support per-CPU partitioning
3028 * of a group with small tg->shares value. It is a floor value which is
3029 * assigned as a minimum load.weight to the sched_entity representing
3030 * the group on a CPU.
3032 * E.g. on 64-bit for a group with tg->shares of scale_load(15)=15*1024
3033 * on an 8-core system with 8 tasks each runnable on one CPU shares has
3034 * to be 15*1024*1/8=1920 instead of scale_load(MIN_SHARES)=2*1024. In
3035 * case no task is runnable on a CPU MIN_SHARES=2 should be returned
3038 return clamp_t(long, shares, MIN_SHARES, tg_shares);
3042 * This calculates the effective runnable weight for a group entity based on
3043 * the group entity weight calculated above.
3045 * Because of the above approximation (2), our group entity weight is
3046 * an load_avg based ratio (3). This means that it includes blocked load and
3047 * does not represent the runnable weight.
3049 * Approximate the group entity's runnable weight per ratio from the group
3052 * grq->avg.runnable_load_avg
3053 * ge->runnable_weight = ge->load.weight * -------------------------- (7)
3056 * However, analogous to above, since the avg numbers are slow, this leads to
3057 * transients in the from-idle case. Instead we use:
3059 * ge->runnable_weight = ge->load.weight *
3061 * max(grq->avg.runnable_load_avg, grq->runnable_weight)
3062 * ----------------------------------------------------- (8)
3063 * max(grq->avg.load_avg, grq->load.weight)
3065 * Where these max() serve both to use the 'instant' values to fix the slow
3066 * from-idle and avoid the /0 on to-idle, similar to (6).
3068 static long calc_group_runnable(struct cfs_rq *cfs_rq, long shares)
3070 long runnable, load_avg;
3072 load_avg = max(cfs_rq->avg.load_avg,
3073 scale_load_down(cfs_rq->load.weight));
3075 runnable = max(cfs_rq->avg.runnable_load_avg,
3076 scale_load_down(cfs_rq->runnable_weight));
3080 runnable /= load_avg;
3082 return clamp_t(long, runnable, MIN_SHARES, shares);
3084 #endif /* CONFIG_SMP */
3086 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
3089 * Recomputes the group entity based on the current state of its group
3092 static void update_cfs_group(struct sched_entity *se)
3094 struct cfs_rq *gcfs_rq = group_cfs_rq(se);
3095 long shares, runnable;
3100 if (throttled_hierarchy(gcfs_rq))
3104 runnable = shares = READ_ONCE(gcfs_rq->tg->shares);
3106 if (likely(se->load.weight == shares))
3109 shares = calc_group_shares(gcfs_rq);
3110 runnable = calc_group_runnable(gcfs_rq, shares);
3113 reweight_entity(cfs_rq_of(se), se, shares, runnable);
3116 #else /* CONFIG_FAIR_GROUP_SCHED */
3117 static inline void update_cfs_group(struct sched_entity *se)
3120 #endif /* CONFIG_FAIR_GROUP_SCHED */
3122 static inline void cfs_rq_util_change(struct cfs_rq *cfs_rq, int flags)
3124 struct rq *rq = rq_of(cfs_rq);
3126 if (&rq->cfs == cfs_rq || (flags & SCHED_CPUFREQ_MIGRATION)) {
3128 * There are a few boundary cases this might miss but it should
3129 * get called often enough that that should (hopefully) not be
3132 * It will not get called when we go idle, because the idle
3133 * thread is a different class (!fair), nor will the utilization
3134 * number include things like RT tasks.
3136 * As is, the util number is not freq-invariant (we'd have to
3137 * implement arch_scale_freq_capacity() for that).
3141 cpufreq_update_util(rq, flags);
3146 #ifdef CONFIG_FAIR_GROUP_SCHED
3148 * update_tg_load_avg - update the tg's load avg
3149 * @cfs_rq: the cfs_rq whose avg changed
3150 * @force: update regardless of how small the difference
3152 * This function 'ensures': tg->load_avg := \Sum tg->cfs_rq[]->avg.load.
3153 * However, because tg->load_avg is a global value there are performance
3156 * In order to avoid having to look at the other cfs_rq's, we use a
3157 * differential update where we store the last value we propagated. This in
3158 * turn allows skipping updates if the differential is 'small'.
3160 * Updating tg's load_avg is necessary before update_cfs_share().
3162 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force)
3164 long delta = cfs_rq->avg.load_avg - cfs_rq->tg_load_avg_contrib;
3167 * No need to update load_avg for root_task_group as it is not used.
3169 if (cfs_rq->tg == &root_task_group)
3172 if (force || abs(delta) > cfs_rq->tg_load_avg_contrib / 64) {
3173 atomic_long_add(delta, &cfs_rq->tg->load_avg);
3174 cfs_rq->tg_load_avg_contrib = cfs_rq->avg.load_avg;
3179 * Called within set_task_rq() right before setting a task's CPU. The
3180 * caller only guarantees p->pi_lock is held; no other assumptions,
3181 * including the state of rq->lock, should be made.
3183 void set_task_rq_fair(struct sched_entity *se,
3184 struct cfs_rq *prev, struct cfs_rq *next)
3186 u64 p_last_update_time;
3187 u64 n_last_update_time;
3189 if (!sched_feat(ATTACH_AGE_LOAD))
3193 * We are supposed to update the task to "current" time, then its up to
3194 * date and ready to go to new CPU/cfs_rq. But we have difficulty in
3195 * getting what current time is, so simply throw away the out-of-date
3196 * time. This will result in the wakee task is less decayed, but giving
3197 * the wakee more load sounds not bad.
3199 if (!(se->avg.last_update_time && prev))
3202 #ifndef CONFIG_64BIT
3204 u64 p_last_update_time_copy;
3205 u64 n_last_update_time_copy;
3208 p_last_update_time_copy = prev->load_last_update_time_copy;
3209 n_last_update_time_copy = next->load_last_update_time_copy;
3213 p_last_update_time = prev->avg.last_update_time;
3214 n_last_update_time = next->avg.last_update_time;
3216 } while (p_last_update_time != p_last_update_time_copy ||
3217 n_last_update_time != n_last_update_time_copy);
3220 p_last_update_time = prev->avg.last_update_time;
3221 n_last_update_time = next->avg.last_update_time;
3223 __update_load_avg_blocked_se(p_last_update_time, cpu_of(rq_of(prev)), se);
3224 se->avg.last_update_time = n_last_update_time;
3229 * When on migration a sched_entity joins/leaves the PELT hierarchy, we need to
3230 * propagate its contribution. The key to this propagation is the invariant
3231 * that for each group:
3233 * ge->avg == grq->avg (1)
3235 * _IFF_ we look at the pure running and runnable sums. Because they
3236 * represent the very same entity, just at different points in the hierarchy.
3238 * Per the above update_tg_cfs_util() is trivial and simply copies the running
3239 * sum over (but still wrong, because the group entity and group rq do not have
3240 * their PELT windows aligned).
3242 * However, update_tg_cfs_runnable() is more complex. So we have:
3244 * ge->avg.load_avg = ge->load.weight * ge->avg.runnable_avg (2)
3246 * And since, like util, the runnable part should be directly transferable,
3247 * the following would _appear_ to be the straight forward approach:
3249 * grq->avg.load_avg = grq->load.weight * grq->avg.runnable_avg (3)
3251 * And per (1) we have:
3253 * ge->avg.runnable_avg == grq->avg.runnable_avg
3257 * ge->load.weight * grq->avg.load_avg
3258 * ge->avg.load_avg = ----------------------------------- (4)
3261 * Except that is wrong!
3263 * Because while for entities historical weight is not important and we
3264 * really only care about our future and therefore can consider a pure
3265 * runnable sum, runqueues can NOT do this.
3267 * We specifically want runqueues to have a load_avg that includes
3268 * historical weights. Those represent the blocked load, the load we expect
3269 * to (shortly) return to us. This only works by keeping the weights as
3270 * integral part of the sum. We therefore cannot decompose as per (3).
3272 * Another reason this doesn't work is that runnable isn't a 0-sum entity.
3273 * Imagine a rq with 2 tasks that each are runnable 2/3 of the time. Then the
3274 * rq itself is runnable anywhere between 2/3 and 1 depending on how the
3275 * runnable section of these tasks overlap (or not). If they were to perfectly
3276 * align the rq as a whole would be runnable 2/3 of the time. If however we
3277 * always have at least 1 runnable task, the rq as a whole is always runnable.
3279 * So we'll have to approximate.. :/
3281 * Given the constraint:
3283 * ge->avg.running_sum <= ge->avg.runnable_sum <= LOAD_AVG_MAX
3285 * We can construct a rule that adds runnable to a rq by assuming minimal
3288 * On removal, we'll assume each task is equally runnable; which yields:
3290 * grq->avg.runnable_sum = grq->avg.load_sum / grq->load.weight
3292 * XXX: only do this for the part of runnable > running ?
3297 update_tg_cfs_util(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq)
3299 long delta = gcfs_rq->avg.util_avg - se->avg.util_avg;
3301 /* Nothing to update */
3306 * The relation between sum and avg is:
3308 * LOAD_AVG_MAX - 1024 + sa->period_contrib
3310 * however, the PELT windows are not aligned between grq and gse.
3313 /* Set new sched_entity's utilization */
3314 se->avg.util_avg = gcfs_rq->avg.util_avg;
3315 se->avg.util_sum = se->avg.util_avg * LOAD_AVG_MAX;
3317 /* Update parent cfs_rq utilization */
3318 add_positive(&cfs_rq->avg.util_avg, delta);
3319 cfs_rq->avg.util_sum = cfs_rq->avg.util_avg * LOAD_AVG_MAX;
3323 update_tg_cfs_runnable(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq)
3325 long delta_avg, running_sum, runnable_sum = gcfs_rq->prop_runnable_sum;
3326 unsigned long runnable_load_avg, load_avg;
3327 u64 runnable_load_sum, load_sum = 0;
3333 gcfs_rq->prop_runnable_sum = 0;
3335 if (runnable_sum >= 0) {
3337 * Add runnable; clip at LOAD_AVG_MAX. Reflects that until
3338 * the CPU is saturated running == runnable.
3340 runnable_sum += se->avg.load_sum;
3341 runnable_sum = min(runnable_sum, (long)LOAD_AVG_MAX);
3344 * Estimate the new unweighted runnable_sum of the gcfs_rq by
3345 * assuming all tasks are equally runnable.
3347 if (scale_load_down(gcfs_rq->load.weight)) {
3348 load_sum = div_s64(gcfs_rq->avg.load_sum,
3349 scale_load_down(gcfs_rq->load.weight));
3352 /* But make sure to not inflate se's runnable */
3353 runnable_sum = min(se->avg.load_sum, load_sum);
3357 * runnable_sum can't be lower than running_sum
3358 * As running sum is scale with CPU capacity wehreas the runnable sum
3359 * is not we rescale running_sum 1st
3361 running_sum = se->avg.util_sum /
3362 arch_scale_cpu_capacity(NULL, cpu_of(rq_of(cfs_rq)));
3363 runnable_sum = max(runnable_sum, running_sum);
3365 load_sum = (s64)se_weight(se) * runnable_sum;
3366 load_avg = div_s64(load_sum, LOAD_AVG_MAX);
3368 delta_sum = load_sum - (s64)se_weight(se) * se->avg.load_sum;
3369 delta_avg = load_avg - se->avg.load_avg;
3371 se->avg.load_sum = runnable_sum;
3372 se->avg.load_avg = load_avg;
3373 add_positive(&cfs_rq->avg.load_avg, delta_avg);
3374 add_positive(&cfs_rq->avg.load_sum, delta_sum);
3376 runnable_load_sum = (s64)se_runnable(se) * runnable_sum;
3377 runnable_load_avg = div_s64(runnable_load_sum, LOAD_AVG_MAX);
3378 delta_sum = runnable_load_sum - se_weight(se) * se->avg.runnable_load_sum;
3379 delta_avg = runnable_load_avg - se->avg.runnable_load_avg;
3381 se->avg.runnable_load_sum = runnable_sum;
3382 se->avg.runnable_load_avg = runnable_load_avg;
3385 add_positive(&cfs_rq->avg.runnable_load_avg, delta_avg);
3386 add_positive(&cfs_rq->avg.runnable_load_sum, delta_sum);
3390 static inline void add_tg_cfs_propagate(struct cfs_rq *cfs_rq, long runnable_sum)
3392 cfs_rq->propagate = 1;
3393 cfs_rq->prop_runnable_sum += runnable_sum;
3396 /* Update task and its cfs_rq load average */
3397 static inline int propagate_entity_load_avg(struct sched_entity *se)
3399 struct cfs_rq *cfs_rq, *gcfs_rq;
3401 if (entity_is_task(se))
3404 gcfs_rq = group_cfs_rq(se);
3405 if (!gcfs_rq->propagate)
3408 gcfs_rq->propagate = 0;
3410 cfs_rq = cfs_rq_of(se);
3412 add_tg_cfs_propagate(cfs_rq, gcfs_rq->prop_runnable_sum);
3414 update_tg_cfs_util(cfs_rq, se, gcfs_rq);
3415 update_tg_cfs_runnable(cfs_rq, se, gcfs_rq);
3421 * Check if we need to update the load and the utilization of a blocked
3424 static inline bool skip_blocked_update(struct sched_entity *se)
3426 struct cfs_rq *gcfs_rq = group_cfs_rq(se);
3429 * If sched_entity still have not zero load or utilization, we have to
3432 if (se->avg.load_avg || se->avg.util_avg)
3436 * If there is a pending propagation, we have to update the load and
3437 * the utilization of the sched_entity:
3439 if (gcfs_rq->propagate)
3443 * Otherwise, the load and the utilization of the sched_entity is
3444 * already zero and there is no pending propagation, so it will be a
3445 * waste of time to try to decay it:
3450 #else /* CONFIG_FAIR_GROUP_SCHED */
3452 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force) {}
3454 static inline int propagate_entity_load_avg(struct sched_entity *se)
3459 static inline void add_tg_cfs_propagate(struct cfs_rq *cfs_rq, long runnable_sum) {}
3461 #endif /* CONFIG_FAIR_GROUP_SCHED */
3464 * update_cfs_rq_load_avg - update the cfs_rq's load/util averages
3465 * @now: current time, as per cfs_rq_clock_task()
3466 * @cfs_rq: cfs_rq to update
3468 * The cfs_rq avg is the direct sum of all its entities (blocked and runnable)
3469 * avg. The immediate corollary is that all (fair) tasks must be attached, see
3470 * post_init_entity_util_avg().
3472 * cfs_rq->avg is used for task_h_load() and update_cfs_share() for example.
3474 * Returns true if the load decayed or we removed load.
3476 * Since both these conditions indicate a changed cfs_rq->avg.load we should
3477 * call update_tg_load_avg() when this function returns true.
3480 update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq)
3482 unsigned long removed_load = 0, removed_util = 0, removed_runnable_sum = 0;
3483 struct sched_avg *sa = &cfs_rq->avg;
3486 if (cfs_rq->removed.nr) {
3488 u32 divider = LOAD_AVG_MAX - 1024 + sa->period_contrib;
3490 raw_spin_lock(&cfs_rq->removed.lock);
3491 swap(cfs_rq->removed.util_avg, removed_util);
3492 swap(cfs_rq->removed.load_avg, removed_load);
3493 swap(cfs_rq->removed.runnable_sum, removed_runnable_sum);
3494 cfs_rq->removed.nr = 0;
3495 raw_spin_unlock(&cfs_rq->removed.lock);
3498 sub_positive(&sa->load_avg, r);
3499 sub_positive(&sa->load_sum, r * divider);
3502 sub_positive(&sa->util_avg, r);
3503 sub_positive(&sa->util_sum, r * divider);
3505 add_tg_cfs_propagate(cfs_rq, -(long)removed_runnable_sum);
3510 decayed |= __update_load_avg_cfs_rq(now, cpu_of(rq_of(cfs_rq)), cfs_rq);
3512 #ifndef CONFIG_64BIT
3514 cfs_rq->load_last_update_time_copy = sa->last_update_time;
3518 cfs_rq_util_change(cfs_rq, 0);
3524 * attach_entity_load_avg - attach this entity to its cfs_rq load avg
3525 * @cfs_rq: cfs_rq to attach to
3526 * @se: sched_entity to attach
3527 * @flags: migration hints
3529 * Must call update_cfs_rq_load_avg() before this, since we rely on
3530 * cfs_rq->avg.last_update_time being current.
3532 static void attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3534 u32 divider = LOAD_AVG_MAX - 1024 + cfs_rq->avg.period_contrib;
3537 * When we attach the @se to the @cfs_rq, we must align the decay
3538 * window because without that, really weird and wonderful things can
3543 se->avg.last_update_time = cfs_rq->avg.last_update_time;
3544 se->avg.period_contrib = cfs_rq->avg.period_contrib;
3547 * Hell(o) Nasty stuff.. we need to recompute _sum based on the new
3548 * period_contrib. This isn't strictly correct, but since we're
3549 * entirely outside of the PELT hierarchy, nobody cares if we truncate
3552 se->avg.util_sum = se->avg.util_avg * divider;
3554 se->avg.load_sum = divider;
3555 if (se_weight(se)) {
3557 div_u64(se->avg.load_avg * se->avg.load_sum, se_weight(se));
3560 se->avg.runnable_load_sum = se->avg.load_sum;
3562 enqueue_load_avg(cfs_rq, se);
3563 cfs_rq->avg.util_avg += se->avg.util_avg;
3564 cfs_rq->avg.util_sum += se->avg.util_sum;
3566 add_tg_cfs_propagate(cfs_rq, se->avg.load_sum);
3568 cfs_rq_util_change(cfs_rq, flags);
3572 * detach_entity_load_avg - detach this entity from its cfs_rq load avg
3573 * @cfs_rq: cfs_rq to detach from
3574 * @se: sched_entity to detach
3576 * Must call update_cfs_rq_load_avg() before this, since we rely on
3577 * cfs_rq->avg.last_update_time being current.
3579 static void detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3581 dequeue_load_avg(cfs_rq, se);
3582 sub_positive(&cfs_rq->avg.util_avg, se->avg.util_avg);
3583 sub_positive(&cfs_rq->avg.util_sum, se->avg.util_sum);
3585 add_tg_cfs_propagate(cfs_rq, -se->avg.load_sum);
3587 cfs_rq_util_change(cfs_rq, 0);
3591 * Optional action to be done while updating the load average
3593 #define UPDATE_TG 0x1
3594 #define SKIP_AGE_LOAD 0x2
3595 #define DO_ATTACH 0x4
3597 /* Update task and its cfs_rq load average */
3598 static inline void update_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3600 u64 now = cfs_rq_clock_task(cfs_rq);
3601 struct rq *rq = rq_of(cfs_rq);
3602 int cpu = cpu_of(rq);
3606 * Track task load average for carrying it to new CPU after migrated, and
3607 * track group sched_entity load average for task_h_load calc in migration
3609 if (se->avg.last_update_time && !(flags & SKIP_AGE_LOAD))
3610 __update_load_avg_se(now, cpu, cfs_rq, se);
3612 decayed = update_cfs_rq_load_avg(now, cfs_rq);
3613 decayed |= propagate_entity_load_avg(se);
3615 if (!se->avg.last_update_time && (flags & DO_ATTACH)) {
3618 * DO_ATTACH means we're here from enqueue_entity().
3619 * !last_update_time means we've passed through
3620 * migrate_task_rq_fair() indicating we migrated.
3622 * IOW we're enqueueing a task on a new CPU.
3624 attach_entity_load_avg(cfs_rq, se, SCHED_CPUFREQ_MIGRATION);
3625 update_tg_load_avg(cfs_rq, 0);
3627 } else if (decayed && (flags & UPDATE_TG))
3628 update_tg_load_avg(cfs_rq, 0);
3631 #ifndef CONFIG_64BIT
3632 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
3634 u64 last_update_time_copy;
3635 u64 last_update_time;
3638 last_update_time_copy = cfs_rq->load_last_update_time_copy;
3640 last_update_time = cfs_rq->avg.last_update_time;
3641 } while (last_update_time != last_update_time_copy);
3643 return last_update_time;
3646 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
3648 return cfs_rq->avg.last_update_time;
3653 * Synchronize entity load avg of dequeued entity without locking
3656 void sync_entity_load_avg(struct sched_entity *se)
3658 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3659 u64 last_update_time;
3661 last_update_time = cfs_rq_last_update_time(cfs_rq);
3662 __update_load_avg_blocked_se(last_update_time, cpu_of(rq_of(cfs_rq)), se);
3666 * Task first catches up with cfs_rq, and then subtract
3667 * itself from the cfs_rq (task must be off the queue now).
3669 void remove_entity_load_avg(struct sched_entity *se)
3671 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3672 unsigned long flags;
3675 * tasks cannot exit without having gone through wake_up_new_task() ->
3676 * post_init_entity_util_avg() which will have added things to the
3677 * cfs_rq, so we can remove unconditionally.
3679 * Similarly for groups, they will have passed through
3680 * post_init_entity_util_avg() before unregister_sched_fair_group()
3684 sync_entity_load_avg(se);
3686 raw_spin_lock_irqsave(&cfs_rq->removed.lock, flags);
3687 ++cfs_rq->removed.nr;
3688 cfs_rq->removed.util_avg += se->avg.util_avg;
3689 cfs_rq->removed.load_avg += se->avg.load_avg;
3690 cfs_rq->removed.runnable_sum += se->avg.load_sum; /* == runnable_sum */
3691 raw_spin_unlock_irqrestore(&cfs_rq->removed.lock, flags);
3694 static inline unsigned long cfs_rq_runnable_load_avg(struct cfs_rq *cfs_rq)
3696 return cfs_rq->avg.runnable_load_avg;
3699 static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq)
3701 return cfs_rq->avg.load_avg;
3704 static int idle_balance(struct rq *this_rq, struct rq_flags *rf);
3706 static inline unsigned long task_util(struct task_struct *p)
3708 return READ_ONCE(p->se.avg.util_avg);
3711 static inline unsigned long _task_util_est(struct task_struct *p)
3713 struct util_est ue = READ_ONCE(p->se.avg.util_est);
3715 return max(ue.ewma, ue.enqueued);
3718 static inline unsigned long task_util_est(struct task_struct *p)
3720 return max(task_util(p), _task_util_est(p));
3723 static inline void util_est_enqueue(struct cfs_rq *cfs_rq,
3724 struct task_struct *p)
3726 unsigned int enqueued;
3728 if (!sched_feat(UTIL_EST))
3731 /* Update root cfs_rq's estimated utilization */
3732 enqueued = cfs_rq->avg.util_est.enqueued;
3733 enqueued += (_task_util_est(p) | UTIL_AVG_UNCHANGED);
3734 WRITE_ONCE(cfs_rq->avg.util_est.enqueued, enqueued);
3738 * Check if a (signed) value is within a specified (unsigned) margin,
3739 * based on the observation that:
3741 * abs(x) < y := (unsigned)(x + y - 1) < (2 * y - 1)
3743 * NOTE: this only works when value + maring < INT_MAX.
3745 static inline bool within_margin(int value, int margin)
3747 return ((unsigned int)(value + margin - 1) < (2 * margin - 1));
3751 util_est_dequeue(struct cfs_rq *cfs_rq, struct task_struct *p, bool task_sleep)
3753 long last_ewma_diff;
3756 if (!sched_feat(UTIL_EST))
3759 /* Update root cfs_rq's estimated utilization */
3760 ue.enqueued = cfs_rq->avg.util_est.enqueued;
3761 ue.enqueued -= min_t(unsigned int, ue.enqueued,
3762 (_task_util_est(p) | UTIL_AVG_UNCHANGED));
3763 WRITE_ONCE(cfs_rq->avg.util_est.enqueued, ue.enqueued);
3766 * Skip update of task's estimated utilization when the task has not
3767 * yet completed an activation, e.g. being migrated.
3773 * If the PELT values haven't changed since enqueue time,
3774 * skip the util_est update.
3776 ue = p->se.avg.util_est;
3777 if (ue.enqueued & UTIL_AVG_UNCHANGED)
3781 * Skip update of task's estimated utilization when its EWMA is
3782 * already ~1% close to its last activation value.
3784 ue.enqueued = (task_util(p) | UTIL_AVG_UNCHANGED);
3785 last_ewma_diff = ue.enqueued - ue.ewma;
3786 if (within_margin(last_ewma_diff, (SCHED_CAPACITY_SCALE / 100)))
3790 * Update Task's estimated utilization
3792 * When *p completes an activation we can consolidate another sample
3793 * of the task size. This is done by storing the current PELT value
3794 * as ue.enqueued and by using this value to update the Exponential
3795 * Weighted Moving Average (EWMA):
3797 * ewma(t) = w * task_util(p) + (1-w) * ewma(t-1)
3798 * = w * task_util(p) + ewma(t-1) - w * ewma(t-1)
3799 * = w * (task_util(p) - ewma(t-1)) + ewma(t-1)
3800 * = w * ( last_ewma_diff ) + ewma(t-1)
3801 * = w * (last_ewma_diff + ewma(t-1) / w)
3803 * Where 'w' is the weight of new samples, which is configured to be
3804 * 0.25, thus making w=1/4 ( >>= UTIL_EST_WEIGHT_SHIFT)
3806 ue.ewma <<= UTIL_EST_WEIGHT_SHIFT;
3807 ue.ewma += last_ewma_diff;
3808 ue.ewma >>= UTIL_EST_WEIGHT_SHIFT;
3809 WRITE_ONCE(p->se.avg.util_est, ue);
3812 #else /* CONFIG_SMP */
3814 #define UPDATE_TG 0x0
3815 #define SKIP_AGE_LOAD 0x0
3816 #define DO_ATTACH 0x0
3818 static inline void update_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int not_used1)
3820 cfs_rq_util_change(cfs_rq, 0);
3823 static inline void remove_entity_load_avg(struct sched_entity *se) {}
3826 attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags) {}
3828 detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
3830 static inline int idle_balance(struct rq *rq, struct rq_flags *rf)
3836 util_est_enqueue(struct cfs_rq *cfs_rq, struct task_struct *p) {}
3839 util_est_dequeue(struct cfs_rq *cfs_rq, struct task_struct *p,
3842 #endif /* CONFIG_SMP */
3844 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
3846 #ifdef CONFIG_SCHED_DEBUG
3847 s64 d = se->vruntime - cfs_rq->min_vruntime;
3852 if (d > 3*sysctl_sched_latency)
3853 schedstat_inc(cfs_rq->nr_spread_over);
3857 static inline bool entity_is_long_sleeper(struct sched_entity *se)
3859 struct cfs_rq *cfs_rq;
3862 if (se->exec_start == 0)
3865 cfs_rq = cfs_rq_of(se);
3867 sleep_time = rq_clock_task(rq_of(cfs_rq));
3869 /* Happen while migrating because of clock task divergence */
3870 if (sleep_time <= se->exec_start)
3873 sleep_time -= se->exec_start;
3874 if (sleep_time > ((1ULL << 63) / scale_load_down(NICE_0_LOAD)))
3881 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
3883 u64 vruntime = cfs_rq->min_vruntime;
3886 * The 'current' period is already promised to the current tasks,
3887 * however the extra weight of the new task will slow them down a
3888 * little, place the new task so that it fits in the slot that
3889 * stays open at the end.
3891 if (initial && sched_feat(START_DEBIT))
3892 vruntime += sched_vslice(cfs_rq, se);
3894 /* sleeps up to a single latency don't count. */
3896 unsigned long thresh = sysctl_sched_latency;
3899 * Halve their sleep time's effect, to allow
3900 * for a gentler effect of sleepers:
3902 if (sched_feat(GENTLE_FAIR_SLEEPERS))
3909 * Pull vruntime of the entity being placed to the base level of
3910 * cfs_rq, to prevent boosting it if placed backwards.
3911 * However, min_vruntime can advance much faster than real time, with
3912 * the extreme being when an entity with the minimal weight always runs
3913 * on the cfs_rq. If the waking entity slept for a long time, its
3914 * vruntime difference from min_vruntime may overflow s64 and their
3915 * comparison may get inversed, so ignore the entity's original
3916 * vruntime in that case.
3917 * The maximal vruntime speedup is given by the ratio of normal to
3918 * minimal weight: scale_load_down(NICE_0_LOAD) / MIN_SHARES.
3919 * When placing a migrated waking entity, its exec_start has been set
3920 * from a different rq. In order to take into account a possible
3921 * divergence between new and prev rq's clocks task because of irq and
3922 * stolen time, we take an additional margin.
3923 * So, cutting off on the sleep time of
3924 * 2^63 / scale_load_down(NICE_0_LOAD) ~ 104 days
3927 if (entity_is_long_sleeper(se))
3928 se->vruntime = vruntime;
3930 se->vruntime = max_vruntime(se->vruntime, vruntime);
3933 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
3935 static inline void check_schedstat_required(void)
3937 #ifdef CONFIG_SCHEDSTATS
3938 if (schedstat_enabled())
3941 /* Force schedstat enabled if a dependent tracepoint is active */
3942 if (trace_sched_stat_wait_enabled() ||
3943 trace_sched_stat_sleep_enabled() ||
3944 trace_sched_stat_iowait_enabled() ||
3945 trace_sched_stat_blocked_enabled() ||
3946 trace_sched_stat_runtime_enabled()) {
3947 printk_deferred_once("Scheduler tracepoints stat_sleep, stat_iowait, "
3948 "stat_blocked and stat_runtime require the "
3949 "kernel parameter schedstats=enable or "
3950 "kernel.sched_schedstats=1\n");
3961 * update_min_vruntime()
3962 * vruntime -= min_vruntime
3966 * update_min_vruntime()
3967 * vruntime += min_vruntime
3969 * this way the vruntime transition between RQs is done when both
3970 * min_vruntime are up-to-date.
3974 * ->migrate_task_rq_fair() (p->state == TASK_WAKING)
3975 * vruntime -= min_vruntime
3979 * update_min_vruntime()
3980 * vruntime += min_vruntime
3982 * this way we don't have the most up-to-date min_vruntime on the originating
3983 * CPU and an up-to-date min_vruntime on the destination CPU.
3987 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3989 bool renorm = !(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_MIGRATED);
3990 bool curr = cfs_rq->curr == se;
3993 * If we're the current task, we must renormalise before calling
3997 se->vruntime += cfs_rq->min_vruntime;
3999 update_curr(cfs_rq);
4002 * Otherwise, renormalise after, such that we're placed at the current
4003 * moment in time, instead of some random moment in the past. Being
4004 * placed in the past could significantly boost this task to the
4005 * fairness detriment of existing tasks.
4007 if (renorm && !curr)
4008 se->vruntime += cfs_rq->min_vruntime;
4011 * When enqueuing a sched_entity, we must:
4012 * - Update loads to have both entity and cfs_rq synced with now.
4013 * - Add its load to cfs_rq->runnable_avg
4014 * - For group_entity, update its weight to reflect the new share of
4016 * - Add its new weight to cfs_rq->load.weight
4018 update_load_avg(cfs_rq, se, UPDATE_TG | DO_ATTACH);
4019 update_cfs_group(se);
4020 enqueue_runnable_load_avg(cfs_rq, se);
4021 account_entity_enqueue(cfs_rq, se);
4023 if (flags & ENQUEUE_WAKEUP)
4024 place_entity(cfs_rq, se, 0);
4025 /* Entity has migrated, no longer consider this task hot */
4026 if (flags & ENQUEUE_MIGRATED)
4029 check_schedstat_required();
4030 update_stats_enqueue(cfs_rq, se, flags);
4031 check_spread(cfs_rq, se);
4033 __enqueue_entity(cfs_rq, se);
4036 if (cfs_rq->nr_running == 1) {
4037 list_add_leaf_cfs_rq(cfs_rq);
4038 check_enqueue_throttle(cfs_rq);
4042 static void __clear_buddies_last(struct sched_entity *se)
4044 for_each_sched_entity(se) {
4045 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4046 if (cfs_rq->last != se)
4049 cfs_rq->last = NULL;
4053 static void __clear_buddies_next(struct sched_entity *se)
4055 for_each_sched_entity(se) {
4056 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4057 if (cfs_rq->next != se)
4060 cfs_rq->next = NULL;
4064 static void __clear_buddies_skip(struct sched_entity *se)
4066 for_each_sched_entity(se) {
4067 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4068 if (cfs_rq->skip != se)
4071 cfs_rq->skip = NULL;
4075 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
4077 if (cfs_rq->last == se)
4078 __clear_buddies_last(se);
4080 if (cfs_rq->next == se)
4081 __clear_buddies_next(se);
4083 if (cfs_rq->skip == se)
4084 __clear_buddies_skip(se);
4087 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
4090 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
4093 * Update run-time statistics of the 'current'.
4095 update_curr(cfs_rq);
4098 * When dequeuing a sched_entity, we must:
4099 * - Update loads to have both entity and cfs_rq synced with now.
4100 * - Substract its load from the cfs_rq->runnable_avg.
4101 * - Substract its previous weight from cfs_rq->load.weight.
4102 * - For group entity, update its weight to reflect the new share
4103 * of its group cfs_rq.
4105 update_load_avg(cfs_rq, se, UPDATE_TG);
4106 dequeue_runnable_load_avg(cfs_rq, se);
4108 update_stats_dequeue(cfs_rq, se, flags);
4110 clear_buddies(cfs_rq, se);
4112 if (se != cfs_rq->curr)
4113 __dequeue_entity(cfs_rq, se);
4115 account_entity_dequeue(cfs_rq, se);
4118 * Normalize after update_curr(); which will also have moved
4119 * min_vruntime if @se is the one holding it back. But before doing
4120 * update_min_vruntime() again, which will discount @se's position and
4121 * can move min_vruntime forward still more.
4123 if (!(flags & DEQUEUE_SLEEP))
4124 se->vruntime -= cfs_rq->min_vruntime;
4126 /* return excess runtime on last dequeue */
4127 return_cfs_rq_runtime(cfs_rq);
4129 update_cfs_group(se);
4132 * Now advance min_vruntime if @se was the entity holding it back,
4133 * except when: DEQUEUE_SAVE && !DEQUEUE_MOVE, in this case we'll be
4134 * put back on, and if we advance min_vruntime, we'll be placed back
4135 * further than we started -- ie. we'll be penalized.
4137 if ((flags & (DEQUEUE_SAVE | DEQUEUE_MOVE)) != DEQUEUE_SAVE)
4138 update_min_vruntime(cfs_rq);
4142 * Preempt the current task with a newly woken task if needed:
4145 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
4147 unsigned long ideal_runtime, delta_exec;
4148 struct sched_entity *se;
4151 ideal_runtime = sched_slice(cfs_rq, curr);
4152 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
4153 if (delta_exec > ideal_runtime) {
4154 resched_curr(rq_of(cfs_rq));
4156 * The current task ran long enough, ensure it doesn't get
4157 * re-elected due to buddy favours.
4159 clear_buddies(cfs_rq, curr);
4164 * Ensure that a task that missed wakeup preemption by a
4165 * narrow margin doesn't have to wait for a full slice.
4166 * This also mitigates buddy induced latencies under load.
4168 if (delta_exec < sysctl_sched_min_granularity)
4171 se = __pick_first_entity(cfs_rq);
4172 delta = curr->vruntime - se->vruntime;
4177 if (delta > ideal_runtime)
4178 resched_curr(rq_of(cfs_rq));
4182 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
4184 /* 'current' is not kept within the tree. */
4187 * Any task has to be enqueued before it get to execute on
4188 * a CPU. So account for the time it spent waiting on the
4191 update_stats_wait_end(cfs_rq, se);
4192 __dequeue_entity(cfs_rq, se);
4193 update_load_avg(cfs_rq, se, UPDATE_TG);
4196 update_stats_curr_start(cfs_rq, se);
4200 * Track our maximum slice length, if the CPU's load is at
4201 * least twice that of our own weight (i.e. dont track it
4202 * when there are only lesser-weight tasks around):
4204 if (schedstat_enabled() && rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
4205 schedstat_set(se->statistics.slice_max,
4206 max((u64)schedstat_val(se->statistics.slice_max),
4207 se->sum_exec_runtime - se->prev_sum_exec_runtime));
4210 se->prev_sum_exec_runtime = se->sum_exec_runtime;
4214 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
4217 * Pick the next process, keeping these things in mind, in this order:
4218 * 1) keep things fair between processes/task groups
4219 * 2) pick the "next" process, since someone really wants that to run
4220 * 3) pick the "last" process, for cache locality
4221 * 4) do not run the "skip" process, if something else is available
4223 static struct sched_entity *
4224 pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr)
4226 struct sched_entity *left = __pick_first_entity(cfs_rq);
4227 struct sched_entity *se;
4230 * If curr is set we have to see if its left of the leftmost entity
4231 * still in the tree, provided there was anything in the tree at all.
4233 if (!left || (curr && entity_before(curr, left)))
4236 se = left; /* ideally we run the leftmost entity */
4239 * Avoid running the skip buddy, if running something else can
4240 * be done without getting too unfair.
4242 if (cfs_rq->skip == se) {
4243 struct sched_entity *second;
4246 second = __pick_first_entity(cfs_rq);
4248 second = __pick_next_entity(se);
4249 if (!second || (curr && entity_before(curr, second)))
4253 if (second && wakeup_preempt_entity(second, left) < 1)
4258 * Prefer last buddy, try to return the CPU to a preempted task.
4260 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
4264 * Someone really wants this to run. If it's not unfair, run it.
4266 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
4269 clear_buddies(cfs_rq, se);
4274 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
4276 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
4279 * If still on the runqueue then deactivate_task()
4280 * was not called and update_curr() has to be done:
4283 update_curr(cfs_rq);
4285 /* throttle cfs_rqs exceeding runtime */
4286 check_cfs_rq_runtime(cfs_rq);
4288 check_spread(cfs_rq, prev);
4291 update_stats_wait_start(cfs_rq, prev);
4292 /* Put 'current' back into the tree. */
4293 __enqueue_entity(cfs_rq, prev);
4294 /* in !on_rq case, update occurred at dequeue */
4295 update_load_avg(cfs_rq, prev, 0);
4297 cfs_rq->curr = NULL;
4301 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
4304 * Update run-time statistics of the 'current'.
4306 update_curr(cfs_rq);
4309 * Ensure that runnable average is periodically updated.
4311 update_load_avg(cfs_rq, curr, UPDATE_TG);
4312 update_cfs_group(curr);
4314 #ifdef CONFIG_SCHED_HRTICK
4316 * queued ticks are scheduled to match the slice, so don't bother
4317 * validating it and just reschedule.
4320 resched_curr(rq_of(cfs_rq));
4324 * don't let the period tick interfere with the hrtick preemption
4326 if (!sched_feat(DOUBLE_TICK) &&
4327 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
4331 if (cfs_rq->nr_running > 1)
4332 check_preempt_tick(cfs_rq, curr);
4336 /**************************************************
4337 * CFS bandwidth control machinery
4340 #ifdef CONFIG_CFS_BANDWIDTH
4342 #ifdef CONFIG_JUMP_LABEL
4343 static struct static_key __cfs_bandwidth_used;
4345 static inline bool cfs_bandwidth_used(void)
4347 return static_key_false(&__cfs_bandwidth_used);
4350 void cfs_bandwidth_usage_inc(void)
4352 static_key_slow_inc_cpuslocked(&__cfs_bandwidth_used);
4355 void cfs_bandwidth_usage_dec(void)
4357 static_key_slow_dec_cpuslocked(&__cfs_bandwidth_used);
4359 #else /* CONFIG_JUMP_LABEL */
4360 static bool cfs_bandwidth_used(void)
4365 void cfs_bandwidth_usage_inc(void) {}
4366 void cfs_bandwidth_usage_dec(void) {}
4367 #endif /* CONFIG_JUMP_LABEL */
4370 * default period for cfs group bandwidth.
4371 * default: 0.1s, units: nanoseconds
4373 static inline u64 default_cfs_period(void)
4375 return 100000000ULL;
4378 static inline u64 sched_cfs_bandwidth_slice(void)
4380 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
4384 * Replenish runtime according to assigned quota. We use sched_clock_cpu
4385 * directly instead of rq->clock to avoid adding additional synchronization
4388 * requires cfs_b->lock
4390 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
4392 if (cfs_b->quota != RUNTIME_INF)
4393 cfs_b->runtime = cfs_b->quota;
4396 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
4398 return &tg->cfs_bandwidth;
4401 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
4402 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
4404 if (unlikely(cfs_rq->throttle_count))
4405 return cfs_rq->throttled_clock_task - cfs_rq->throttled_clock_task_time;
4407 return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time;
4410 /* returns 0 on failure to allocate runtime */
4411 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4413 struct task_group *tg = cfs_rq->tg;
4414 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
4415 u64 amount = 0, min_amount;
4417 /* note: this is a positive sum as runtime_remaining <= 0 */
4418 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
4420 raw_spin_lock(&cfs_b->lock);
4421 if (cfs_b->quota == RUNTIME_INF)
4422 amount = min_amount;
4424 start_cfs_bandwidth(cfs_b);
4426 if (cfs_b->runtime > 0) {
4427 amount = min(cfs_b->runtime, min_amount);
4428 cfs_b->runtime -= amount;
4432 raw_spin_unlock(&cfs_b->lock);
4434 cfs_rq->runtime_remaining += amount;
4436 return cfs_rq->runtime_remaining > 0;
4439 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
4441 /* dock delta_exec before expiring quota (as it could span periods) */
4442 cfs_rq->runtime_remaining -= delta_exec;
4444 if (likely(cfs_rq->runtime_remaining > 0))
4447 if (cfs_rq->throttled)
4450 * if we're unable to extend our runtime we resched so that the active
4451 * hierarchy can be throttled
4453 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
4454 resched_curr(rq_of(cfs_rq));
4457 static __always_inline
4458 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
4460 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
4463 __account_cfs_rq_runtime(cfs_rq, delta_exec);
4466 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
4468 return cfs_bandwidth_used() && cfs_rq->throttled;
4471 /* check whether cfs_rq, or any parent, is throttled */
4472 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
4474 return cfs_bandwidth_used() && cfs_rq->throttle_count;
4478 * Ensure that neither of the group entities corresponding to src_cpu or
4479 * dest_cpu are members of a throttled hierarchy when performing group
4480 * load-balance operations.
4482 static inline int throttled_lb_pair(struct task_group *tg,
4483 int src_cpu, int dest_cpu)
4485 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
4487 src_cfs_rq = tg->cfs_rq[src_cpu];
4488 dest_cfs_rq = tg->cfs_rq[dest_cpu];
4490 return throttled_hierarchy(src_cfs_rq) ||
4491 throttled_hierarchy(dest_cfs_rq);
4494 static int tg_unthrottle_up(struct task_group *tg, void *data)
4496 struct rq *rq = data;
4497 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
4499 cfs_rq->throttle_count--;
4500 if (!cfs_rq->throttle_count) {
4501 /* adjust cfs_rq_clock_task() */
4502 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
4503 cfs_rq->throttled_clock_task;
4505 /* Add cfs_rq with already running entity in the list */
4506 if (cfs_rq->nr_running >= 1)
4507 list_add_leaf_cfs_rq(cfs_rq);
4513 static int tg_throttle_down(struct task_group *tg, void *data)
4515 struct rq *rq = data;
4516 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
4518 /* group is entering throttled state, stop time */
4519 if (!cfs_rq->throttle_count) {
4520 cfs_rq->throttled_clock_task = rq_clock_task(rq);
4521 list_del_leaf_cfs_rq(cfs_rq);
4523 cfs_rq->throttle_count++;
4528 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
4530 struct rq *rq = rq_of(cfs_rq);
4531 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4532 struct sched_entity *se;
4533 long task_delta, dequeue = 1;
4536 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
4538 /* freeze hierarchy runnable averages while throttled */
4540 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
4543 task_delta = cfs_rq->h_nr_running;
4544 for_each_sched_entity(se) {
4545 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
4546 /* throttled entity or throttle-on-deactivate */
4551 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
4552 qcfs_rq->h_nr_running -= task_delta;
4554 if (qcfs_rq->load.weight)
4559 sub_nr_running(rq, task_delta);
4561 cfs_rq->throttled = 1;
4562 cfs_rq->throttled_clock = rq_clock(rq);
4563 raw_spin_lock(&cfs_b->lock);
4564 empty = list_empty(&cfs_b->throttled_cfs_rq);
4567 * Add to the _head_ of the list, so that an already-started
4568 * distribute_cfs_runtime will not see us. If disribute_cfs_runtime is
4569 * not running add to the tail so that later runqueues don't get starved.
4571 if (cfs_b->distribute_running)
4572 list_add_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
4574 list_add_tail_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
4577 * If we're the first throttled task, make sure the bandwidth
4581 start_cfs_bandwidth(cfs_b);
4583 raw_spin_unlock(&cfs_b->lock);
4586 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
4588 struct rq *rq = rq_of(cfs_rq);
4589 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4590 struct sched_entity *se;
4594 se = cfs_rq->tg->se[cpu_of(rq)];
4596 cfs_rq->throttled = 0;
4598 update_rq_clock(rq);
4600 raw_spin_lock(&cfs_b->lock);
4601 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
4602 list_del_rcu(&cfs_rq->throttled_list);
4603 raw_spin_unlock(&cfs_b->lock);
4605 /* update hierarchical throttle state */
4606 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
4608 if (!cfs_rq->load.weight)
4611 task_delta = cfs_rq->h_nr_running;
4612 for_each_sched_entity(se) {
4616 cfs_rq = cfs_rq_of(se);
4618 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
4619 cfs_rq->h_nr_running += task_delta;
4621 if (cfs_rq_throttled(cfs_rq))
4625 assert_list_leaf_cfs_rq(rq);
4628 add_nr_running(rq, task_delta);
4630 /* Determine whether we need to wake up potentially idle CPU: */
4631 if (rq->curr == rq->idle && rq->cfs.nr_running)
4635 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b, u64 remaining)
4637 struct cfs_rq *cfs_rq;
4639 u64 starting_runtime = remaining;
4642 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
4644 struct rq *rq = rq_of(cfs_rq);
4648 if (!cfs_rq_throttled(cfs_rq))
4651 /* By the above check, this should never be true */
4652 SCHED_WARN_ON(cfs_rq->runtime_remaining > 0);
4654 runtime = -cfs_rq->runtime_remaining + 1;
4655 if (runtime > remaining)
4656 runtime = remaining;
4657 remaining -= runtime;
4659 cfs_rq->runtime_remaining += runtime;
4661 /* we check whether we're throttled above */
4662 if (cfs_rq->runtime_remaining > 0)
4663 unthrottle_cfs_rq(cfs_rq);
4673 return starting_runtime - remaining;
4677 * Responsible for refilling a task_group's bandwidth and unthrottling its
4678 * cfs_rqs as appropriate. If there has been no activity within the last
4679 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
4680 * used to track this state.
4682 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
4687 /* no need to continue the timer with no bandwidth constraint */
4688 if (cfs_b->quota == RUNTIME_INF)
4689 goto out_deactivate;
4691 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
4692 cfs_b->nr_periods += overrun;
4695 * idle depends on !throttled (for the case of a large deficit), and if
4696 * we're going inactive then everything else can be deferred
4698 if (cfs_b->idle && !throttled)
4699 goto out_deactivate;
4701 __refill_cfs_bandwidth_runtime(cfs_b);
4704 /* mark as potentially idle for the upcoming period */
4709 /* account preceding periods in which throttling occurred */
4710 cfs_b->nr_throttled += overrun;
4713 * This check is repeated as we are holding onto the new bandwidth while
4714 * we unthrottle. This can potentially race with an unthrottled group
4715 * trying to acquire new bandwidth from the global pool. This can result
4716 * in us over-using our runtime if it is all used during this loop, but
4717 * only by limited amounts in that extreme case.
4719 while (throttled && cfs_b->runtime > 0 && !cfs_b->distribute_running) {
4720 runtime = cfs_b->runtime;
4721 cfs_b->distribute_running = 1;
4722 raw_spin_unlock(&cfs_b->lock);
4723 /* we can't nest cfs_b->lock while distributing bandwidth */
4724 runtime = distribute_cfs_runtime(cfs_b, runtime);
4725 raw_spin_lock(&cfs_b->lock);
4727 cfs_b->distribute_running = 0;
4728 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
4730 cfs_b->runtime -= min(runtime, cfs_b->runtime);
4734 * While we are ensured activity in the period following an
4735 * unthrottle, this also covers the case in which the new bandwidth is
4736 * insufficient to cover the existing bandwidth deficit. (Forcing the
4737 * timer to remain active while there are any throttled entities.)
4747 /* a cfs_rq won't donate quota below this amount */
4748 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
4749 /* minimum remaining period time to redistribute slack quota */
4750 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
4751 /* how long we wait to gather additional slack before distributing */
4752 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
4755 * Are we near the end of the current quota period?
4757 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
4758 * hrtimer base being cleared by hrtimer_start. In the case of
4759 * migrate_hrtimers, base is never cleared, so we are fine.
4761 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
4763 struct hrtimer *refresh_timer = &cfs_b->period_timer;
4766 /* if the call-back is running a quota refresh is already occurring */
4767 if (hrtimer_callback_running(refresh_timer))
4770 /* is a quota refresh about to occur? */
4771 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
4772 if (remaining < (s64)min_expire)
4778 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
4780 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
4782 /* if there's a quota refresh soon don't bother with slack */
4783 if (runtime_refresh_within(cfs_b, min_left))
4786 hrtimer_start(&cfs_b->slack_timer,
4787 ns_to_ktime(cfs_bandwidth_slack_period),
4791 /* we know any runtime found here is valid as update_curr() precedes return */
4792 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4794 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4795 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
4797 if (slack_runtime <= 0)
4800 raw_spin_lock(&cfs_b->lock);
4801 if (cfs_b->quota != RUNTIME_INF) {
4802 cfs_b->runtime += slack_runtime;
4804 /* we are under rq->lock, defer unthrottling using a timer */
4805 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
4806 !list_empty(&cfs_b->throttled_cfs_rq))
4807 start_cfs_slack_bandwidth(cfs_b);
4809 raw_spin_unlock(&cfs_b->lock);
4811 /* even if it's not valid for return we don't want to try again */
4812 cfs_rq->runtime_remaining -= slack_runtime;
4815 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4817 if (!cfs_bandwidth_used())
4820 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
4823 __return_cfs_rq_runtime(cfs_rq);
4827 * This is done with a timer (instead of inline with bandwidth return) since
4828 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
4830 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
4832 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
4834 /* confirm we're still not at a refresh boundary */
4835 raw_spin_lock(&cfs_b->lock);
4836 if (cfs_b->distribute_running) {
4837 raw_spin_unlock(&cfs_b->lock);
4841 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
4842 raw_spin_unlock(&cfs_b->lock);
4846 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice)
4847 runtime = cfs_b->runtime;
4850 cfs_b->distribute_running = 1;
4852 raw_spin_unlock(&cfs_b->lock);
4857 runtime = distribute_cfs_runtime(cfs_b, runtime);
4859 raw_spin_lock(&cfs_b->lock);
4860 cfs_b->runtime -= min(runtime, cfs_b->runtime);
4861 cfs_b->distribute_running = 0;
4862 raw_spin_unlock(&cfs_b->lock);
4866 * When a group wakes up we want to make sure that its quota is not already
4867 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
4868 * runtime as update_curr() throttling can not not trigger until it's on-rq.
4870 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
4872 if (!cfs_bandwidth_used())
4875 /* an active group must be handled by the update_curr()->put() path */
4876 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
4879 /* ensure the group is not already throttled */
4880 if (cfs_rq_throttled(cfs_rq))
4883 /* update runtime allocation */
4884 account_cfs_rq_runtime(cfs_rq, 0);
4885 if (cfs_rq->runtime_remaining <= 0)
4886 throttle_cfs_rq(cfs_rq);
4889 static void sync_throttle(struct task_group *tg, int cpu)
4891 struct cfs_rq *pcfs_rq, *cfs_rq;
4893 if (!cfs_bandwidth_used())
4899 cfs_rq = tg->cfs_rq[cpu];
4900 pcfs_rq = tg->parent->cfs_rq[cpu];
4902 cfs_rq->throttle_count = pcfs_rq->throttle_count;
4903 cfs_rq->throttled_clock_task = rq_clock_task(cpu_rq(cpu));
4906 /* conditionally throttle active cfs_rq's from put_prev_entity() */
4907 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4909 if (!cfs_bandwidth_used())
4912 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
4916 * it's possible for a throttled entity to be forced into a running
4917 * state (e.g. set_curr_task), in this case we're finished.
4919 if (cfs_rq_throttled(cfs_rq))
4922 throttle_cfs_rq(cfs_rq);
4926 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
4928 struct cfs_bandwidth *cfs_b =
4929 container_of(timer, struct cfs_bandwidth, slack_timer);
4931 do_sched_cfs_slack_timer(cfs_b);
4933 return HRTIMER_NORESTART;
4936 extern const u64 max_cfs_quota_period;
4938 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
4940 struct cfs_bandwidth *cfs_b =
4941 container_of(timer, struct cfs_bandwidth, period_timer);
4946 raw_spin_lock(&cfs_b->lock);
4948 overrun = hrtimer_forward_now(timer, cfs_b->period);
4953 u64 new, old = ktime_to_ns(cfs_b->period);
4956 * Grow period by a factor of 2 to avoid losing precision.
4957 * Precision loss in the quota/period ratio can cause __cfs_schedulable
4961 if (new < max_cfs_quota_period) {
4962 cfs_b->period = ns_to_ktime(new);
4965 pr_warn_ratelimited(
4966 "cfs_period_timer[cpu%d]: period too short, scaling up (new cfs_period_us = %lld, cfs_quota_us = %lld)\n",
4968 div_u64(new, NSEC_PER_USEC),
4969 div_u64(cfs_b->quota, NSEC_PER_USEC));
4971 pr_warn_ratelimited(
4972 "cfs_period_timer[cpu%d]: period too short, but cannot scale up without losing precision (cfs_period_us = %lld, cfs_quota_us = %lld)\n",
4974 div_u64(old, NSEC_PER_USEC),
4975 div_u64(cfs_b->quota, NSEC_PER_USEC));
4978 /* reset count so we don't come right back in here */
4982 idle = do_sched_cfs_period_timer(cfs_b, overrun);
4985 cfs_b->period_active = 0;
4986 raw_spin_unlock(&cfs_b->lock);
4988 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
4991 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4993 raw_spin_lock_init(&cfs_b->lock);
4995 cfs_b->quota = RUNTIME_INF;
4996 cfs_b->period = ns_to_ktime(default_cfs_period());
4998 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
4999 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED);
5000 cfs_b->period_timer.function = sched_cfs_period_timer;
5001 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
5002 cfs_b->slack_timer.function = sched_cfs_slack_timer;
5003 cfs_b->distribute_running = 0;
5006 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
5008 cfs_rq->runtime_enabled = 0;
5009 INIT_LIST_HEAD(&cfs_rq->throttled_list);
5012 void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
5014 lockdep_assert_held(&cfs_b->lock);
5016 if (cfs_b->period_active)
5019 cfs_b->period_active = 1;
5020 hrtimer_forward_now(&cfs_b->period_timer, cfs_b->period);
5021 hrtimer_start_expires(&cfs_b->period_timer, HRTIMER_MODE_ABS_PINNED);
5024 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
5026 /* init_cfs_bandwidth() was not called */
5027 if (!cfs_b->throttled_cfs_rq.next)
5030 hrtimer_cancel(&cfs_b->period_timer);
5031 hrtimer_cancel(&cfs_b->slack_timer);
5035 * Both these CPU hotplug callbacks race against unregister_fair_sched_group()
5037 * The race is harmless, since modifying bandwidth settings of unhooked group
5038 * bits doesn't do much.
5041 /* cpu online calback */
5042 static void __maybe_unused update_runtime_enabled(struct rq *rq)
5044 struct task_group *tg;
5046 lockdep_assert_held(&rq->lock);
5049 list_for_each_entry_rcu(tg, &task_groups, list) {
5050 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
5051 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
5053 raw_spin_lock(&cfs_b->lock);
5054 cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF;
5055 raw_spin_unlock(&cfs_b->lock);
5060 /* cpu offline callback */
5061 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
5063 struct task_group *tg;
5065 lockdep_assert_held(&rq->lock);
5068 list_for_each_entry_rcu(tg, &task_groups, list) {
5069 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
5071 if (!cfs_rq->runtime_enabled)
5075 * clock_task is not advancing so we just need to make sure
5076 * there's some valid quota amount
5078 cfs_rq->runtime_remaining = 1;
5080 * Offline rq is schedulable till CPU is completely disabled
5081 * in take_cpu_down(), so we prevent new cfs throttling here.
5083 cfs_rq->runtime_enabled = 0;
5085 if (cfs_rq_throttled(cfs_rq))
5086 unthrottle_cfs_rq(cfs_rq);
5091 #else /* CONFIG_CFS_BANDWIDTH */
5093 static inline bool cfs_bandwidth_used(void)
5098 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
5100 return rq_clock_task(rq_of(cfs_rq));
5103 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {}
5104 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; }
5105 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
5106 static inline void sync_throttle(struct task_group *tg, int cpu) {}
5107 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
5109 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
5114 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
5119 static inline int throttled_lb_pair(struct task_group *tg,
5120 int src_cpu, int dest_cpu)
5125 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
5127 #ifdef CONFIG_FAIR_GROUP_SCHED
5128 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
5131 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
5135 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
5136 static inline void update_runtime_enabled(struct rq *rq) {}
5137 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
5139 #endif /* CONFIG_CFS_BANDWIDTH */
5141 /**************************************************
5142 * CFS operations on tasks:
5145 #ifdef CONFIG_SCHED_HRTICK
5146 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
5148 struct sched_entity *se = &p->se;
5149 struct cfs_rq *cfs_rq = cfs_rq_of(se);
5151 SCHED_WARN_ON(task_rq(p) != rq);
5153 if (rq->cfs.h_nr_running > 1) {
5154 u64 slice = sched_slice(cfs_rq, se);
5155 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
5156 s64 delta = slice - ran;
5163 hrtick_start(rq, delta);
5168 * called from enqueue/dequeue and updates the hrtick when the
5169 * current task is from our class and nr_running is low enough
5172 static void hrtick_update(struct rq *rq)
5174 struct task_struct *curr = rq->curr;
5176 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
5179 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
5180 hrtick_start_fair(rq, curr);
5182 #else /* !CONFIG_SCHED_HRTICK */
5184 hrtick_start_fair(struct rq *rq, struct task_struct *p)
5188 static inline void hrtick_update(struct rq *rq)
5194 * The enqueue_task method is called before nr_running is
5195 * increased. Here we update the fair scheduling stats and
5196 * then put the task into the rbtree:
5199 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
5201 struct cfs_rq *cfs_rq;
5202 struct sched_entity *se = &p->se;
5205 * The code below (indirectly) updates schedutil which looks at
5206 * the cfs_rq utilization to select a frequency.
5207 * Let's add the task's estimated utilization to the cfs_rq's
5208 * estimated utilization, before we update schedutil.
5210 util_est_enqueue(&rq->cfs, p);
5213 * If in_iowait is set, the code below may not trigger any cpufreq
5214 * utilization updates, so do it here explicitly with the IOWAIT flag
5218 cpufreq_update_util(rq, SCHED_CPUFREQ_IOWAIT);
5220 for_each_sched_entity(se) {
5223 cfs_rq = cfs_rq_of(se);
5224 enqueue_entity(cfs_rq, se, flags);
5227 * end evaluation on encountering a throttled cfs_rq
5229 * note: in the case of encountering a throttled cfs_rq we will
5230 * post the final h_nr_running increment below.
5232 if (cfs_rq_throttled(cfs_rq))
5234 cfs_rq->h_nr_running++;
5236 flags = ENQUEUE_WAKEUP;
5239 for_each_sched_entity(se) {
5240 cfs_rq = cfs_rq_of(se);
5241 cfs_rq->h_nr_running++;
5243 if (cfs_rq_throttled(cfs_rq))
5246 update_load_avg(cfs_rq, se, UPDATE_TG);
5247 update_cfs_group(se);
5251 add_nr_running(rq, 1);
5253 if (cfs_bandwidth_used()) {
5255 * When bandwidth control is enabled; the cfs_rq_throttled()
5256 * breaks in the above iteration can result in incomplete
5257 * leaf list maintenance, resulting in triggering the assertion
5260 for_each_sched_entity(se) {
5261 cfs_rq = cfs_rq_of(se);
5263 if (list_add_leaf_cfs_rq(cfs_rq))
5268 assert_list_leaf_cfs_rq(rq);
5273 static void set_next_buddy(struct sched_entity *se);
5276 * The dequeue_task method is called before nr_running is
5277 * decreased. We remove the task from the rbtree and
5278 * update the fair scheduling stats:
5280 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
5282 struct cfs_rq *cfs_rq;
5283 struct sched_entity *se = &p->se;
5284 int task_sleep = flags & DEQUEUE_SLEEP;
5286 for_each_sched_entity(se) {
5287 cfs_rq = cfs_rq_of(se);
5288 dequeue_entity(cfs_rq, se, flags);
5291 * end evaluation on encountering a throttled cfs_rq
5293 * note: in the case of encountering a throttled cfs_rq we will
5294 * post the final h_nr_running decrement below.
5296 if (cfs_rq_throttled(cfs_rq))
5298 cfs_rq->h_nr_running--;
5300 /* Don't dequeue parent if it has other entities besides us */
5301 if (cfs_rq->load.weight) {
5302 /* Avoid re-evaluating load for this entity: */
5303 se = parent_entity(se);
5305 * Bias pick_next to pick a task from this cfs_rq, as
5306 * p is sleeping when it is within its sched_slice.
5308 if (task_sleep && se && !throttled_hierarchy(cfs_rq))
5312 flags |= DEQUEUE_SLEEP;
5315 for_each_sched_entity(se) {
5316 cfs_rq = cfs_rq_of(se);
5317 cfs_rq->h_nr_running--;
5319 if (cfs_rq_throttled(cfs_rq))
5322 update_load_avg(cfs_rq, se, UPDATE_TG);
5323 update_cfs_group(se);
5327 sub_nr_running(rq, 1);
5329 util_est_dequeue(&rq->cfs, p, task_sleep);
5335 /* Working cpumask for: load_balance, load_balance_newidle. */
5336 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
5337 DEFINE_PER_CPU(cpumask_var_t, select_idle_mask);
5339 #ifdef CONFIG_NO_HZ_COMMON
5341 * per rq 'load' arrray crap; XXX kill this.
5345 * The exact cpuload calculated at every tick would be:
5347 * load' = (1 - 1/2^i) * load + (1/2^i) * cur_load
5349 * If a CPU misses updates for n ticks (as it was idle) and update gets
5350 * called on the n+1-th tick when CPU may be busy, then we have:
5352 * load_n = (1 - 1/2^i)^n * load_0
5353 * load_n+1 = (1 - 1/2^i) * load_n + (1/2^i) * cur_load
5355 * decay_load_missed() below does efficient calculation of
5357 * load' = (1 - 1/2^i)^n * load
5359 * Because x^(n+m) := x^n * x^m we can decompose any x^n in power-of-2 factors.
5360 * This allows us to precompute the above in said factors, thereby allowing the
5361 * reduction of an arbitrary n in O(log_2 n) steps. (See also
5362 * fixed_power_int())
5364 * The calculation is approximated on a 128 point scale.
5366 #define DEGRADE_SHIFT 7
5368 static const u8 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
5369 static const u8 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
5370 { 0, 0, 0, 0, 0, 0, 0, 0 },
5371 { 64, 32, 8, 0, 0, 0, 0, 0 },
5372 { 96, 72, 40, 12, 1, 0, 0, 0 },
5373 { 112, 98, 75, 43, 15, 1, 0, 0 },
5374 { 120, 112, 98, 76, 45, 16, 2, 0 }
5378 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
5379 * would be when CPU is idle and so we just decay the old load without
5380 * adding any new load.
5382 static unsigned long
5383 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
5387 if (!missed_updates)
5390 if (missed_updates >= degrade_zero_ticks[idx])
5394 return load >> missed_updates;
5396 while (missed_updates) {
5397 if (missed_updates % 2)
5398 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
5400 missed_updates >>= 1;
5407 cpumask_var_t idle_cpus_mask;
5409 int has_blocked; /* Idle CPUS has blocked load */
5410 unsigned long next_balance; /* in jiffy units */
5411 unsigned long next_blocked; /* Next update of blocked load in jiffies */
5412 } nohz ____cacheline_aligned;
5414 #endif /* CONFIG_NO_HZ_COMMON */
5417 * __cpu_load_update - update the rq->cpu_load[] statistics
5418 * @this_rq: The rq to update statistics for
5419 * @this_load: The current load
5420 * @pending_updates: The number of missed updates
5422 * Update rq->cpu_load[] statistics. This function is usually called every
5423 * scheduler tick (TICK_NSEC).
5425 * This function computes a decaying average:
5427 * load[i]' = (1 - 1/2^i) * load[i] + (1/2^i) * load
5429 * Because of NOHZ it might not get called on every tick which gives need for
5430 * the @pending_updates argument.
5432 * load[i]_n = (1 - 1/2^i) * load[i]_n-1 + (1/2^i) * load_n-1
5433 * = A * load[i]_n-1 + B ; A := (1 - 1/2^i), B := (1/2^i) * load
5434 * = A * (A * load[i]_n-2 + B) + B
5435 * = A * (A * (A * load[i]_n-3 + B) + B) + B
5436 * = A^3 * load[i]_n-3 + (A^2 + A + 1) * B
5437 * = A^n * load[i]_0 + (A^(n-1) + A^(n-2) + ... + 1) * B
5438 * = A^n * load[i]_0 + ((1 - A^n) / (1 - A)) * B
5439 * = (1 - 1/2^i)^n * (load[i]_0 - load) + load
5441 * In the above we've assumed load_n := load, which is true for NOHZ_FULL as
5442 * any change in load would have resulted in the tick being turned back on.
5444 * For regular NOHZ, this reduces to:
5446 * load[i]_n = (1 - 1/2^i)^n * load[i]_0
5448 * see decay_load_misses(). For NOHZ_FULL we get to subtract and add the extra
5451 static void cpu_load_update(struct rq *this_rq, unsigned long this_load,
5452 unsigned long pending_updates)
5454 unsigned long __maybe_unused tickless_load = this_rq->cpu_load[0];
5457 this_rq->nr_load_updates++;
5459 /* Update our load: */
5460 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
5461 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
5462 unsigned long old_load, new_load;
5464 /* scale is effectively 1 << i now, and >> i divides by scale */
5466 old_load = this_rq->cpu_load[i];
5467 #ifdef CONFIG_NO_HZ_COMMON
5468 old_load = decay_load_missed(old_load, pending_updates - 1, i);
5469 if (tickless_load) {
5470 old_load -= decay_load_missed(tickless_load, pending_updates - 1, i);
5472 * old_load can never be a negative value because a
5473 * decayed tickless_load cannot be greater than the
5474 * original tickless_load.
5476 old_load += tickless_load;
5479 new_load = this_load;
5481 * Round up the averaging division if load is increasing. This
5482 * prevents us from getting stuck on 9 if the load is 10, for
5485 if (new_load > old_load)
5486 new_load += scale - 1;
5488 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
5492 /* Used instead of source_load when we know the type == 0 */
5493 static unsigned long weighted_cpuload(struct rq *rq)
5495 return cfs_rq_runnable_load_avg(&rq->cfs);
5498 #ifdef CONFIG_NO_HZ_COMMON
5500 * There is no sane way to deal with nohz on smp when using jiffies because the
5501 * CPU doing the jiffies update might drift wrt the CPU doing the jiffy reading
5502 * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}.
5504 * Therefore we need to avoid the delta approach from the regular tick when
5505 * possible since that would seriously skew the load calculation. This is why we
5506 * use cpu_load_update_periodic() for CPUs out of nohz. However we'll rely on
5507 * jiffies deltas for updates happening while in nohz mode (idle ticks, idle
5508 * loop exit, nohz_idle_balance, nohz full exit...)
5510 * This means we might still be one tick off for nohz periods.
5513 static void cpu_load_update_nohz(struct rq *this_rq,
5514 unsigned long curr_jiffies,
5517 unsigned long pending_updates;
5519 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
5520 if (pending_updates) {
5521 this_rq->last_load_update_tick = curr_jiffies;
5523 * In the regular NOHZ case, we were idle, this means load 0.
5524 * In the NOHZ_FULL case, we were non-idle, we should consider
5525 * its weighted load.
5527 cpu_load_update(this_rq, load, pending_updates);
5532 * Called from nohz_idle_balance() to update the load ratings before doing the
5535 static void cpu_load_update_idle(struct rq *this_rq)
5538 * bail if there's load or we're actually up-to-date.
5540 if (weighted_cpuload(this_rq))
5543 cpu_load_update_nohz(this_rq, READ_ONCE(jiffies), 0);
5547 * Record CPU load on nohz entry so we know the tickless load to account
5548 * on nohz exit. cpu_load[0] happens then to be updated more frequently
5549 * than other cpu_load[idx] but it should be fine as cpu_load readers
5550 * shouldn't rely into synchronized cpu_load[*] updates.
5552 void cpu_load_update_nohz_start(void)
5554 struct rq *this_rq = this_rq();
5557 * This is all lockless but should be fine. If weighted_cpuload changes
5558 * concurrently we'll exit nohz. And cpu_load write can race with
5559 * cpu_load_update_idle() but both updater would be writing the same.
5561 this_rq->cpu_load[0] = weighted_cpuload(this_rq);
5565 * Account the tickless load in the end of a nohz frame.
5567 void cpu_load_update_nohz_stop(void)
5569 unsigned long curr_jiffies = READ_ONCE(jiffies);
5570 struct rq *this_rq = this_rq();
5574 if (curr_jiffies == this_rq->last_load_update_tick)
5577 load = weighted_cpuload(this_rq);
5578 rq_lock(this_rq, &rf);
5579 update_rq_clock(this_rq);
5580 cpu_load_update_nohz(this_rq, curr_jiffies, load);
5581 rq_unlock(this_rq, &rf);
5583 #else /* !CONFIG_NO_HZ_COMMON */
5584 static inline void cpu_load_update_nohz(struct rq *this_rq,
5585 unsigned long curr_jiffies,
5586 unsigned long load) { }
5587 #endif /* CONFIG_NO_HZ_COMMON */
5589 static void cpu_load_update_periodic(struct rq *this_rq, unsigned long load)
5591 #ifdef CONFIG_NO_HZ_COMMON
5592 /* See the mess around cpu_load_update_nohz(). */
5593 this_rq->last_load_update_tick = READ_ONCE(jiffies);
5595 cpu_load_update(this_rq, load, 1);
5599 * Called from scheduler_tick()
5601 void cpu_load_update_active(struct rq *this_rq)
5603 unsigned long load = weighted_cpuload(this_rq);
5605 if (tick_nohz_tick_stopped())
5606 cpu_load_update_nohz(this_rq, READ_ONCE(jiffies), load);
5608 cpu_load_update_periodic(this_rq, load);
5612 * Return a low guess at the load of a migration-source CPU weighted
5613 * according to the scheduling class and "nice" value.
5615 * We want to under-estimate the load of migration sources, to
5616 * balance conservatively.
5618 static unsigned long source_load(int cpu, int type)
5620 struct rq *rq = cpu_rq(cpu);
5621 unsigned long total = weighted_cpuload(rq);
5623 if (type == 0 || !sched_feat(LB_BIAS))
5626 return min(rq->cpu_load[type-1], total);
5630 * Return a high guess at the load of a migration-target CPU weighted
5631 * according to the scheduling class and "nice" value.
5633 static unsigned long target_load(int cpu, int type)
5635 struct rq *rq = cpu_rq(cpu);
5636 unsigned long total = weighted_cpuload(rq);
5638 if (type == 0 || !sched_feat(LB_BIAS))
5641 return max(rq->cpu_load[type-1], total);
5644 static unsigned long capacity_of(int cpu)
5646 return cpu_rq(cpu)->cpu_capacity;
5649 static unsigned long capacity_orig_of(int cpu)
5651 return cpu_rq(cpu)->cpu_capacity_orig;
5654 static unsigned long cpu_avg_load_per_task(int cpu)
5656 struct rq *rq = cpu_rq(cpu);
5657 unsigned long nr_running = READ_ONCE(rq->cfs.h_nr_running);
5658 unsigned long load_avg = weighted_cpuload(rq);
5661 return load_avg / nr_running;
5666 static void record_wakee(struct task_struct *p)
5669 * Only decay a single time; tasks that have less then 1 wakeup per
5670 * jiffy will not have built up many flips.
5672 if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) {
5673 current->wakee_flips >>= 1;
5674 current->wakee_flip_decay_ts = jiffies;
5677 if (current->last_wakee != p) {
5678 current->last_wakee = p;
5679 current->wakee_flips++;
5684 * Detect M:N waker/wakee relationships via a switching-frequency heuristic.
5686 * A waker of many should wake a different task than the one last awakened
5687 * at a frequency roughly N times higher than one of its wakees.
5689 * In order to determine whether we should let the load spread vs consolidating
5690 * to shared cache, we look for a minimum 'flip' frequency of llc_size in one
5691 * partner, and a factor of lls_size higher frequency in the other.
5693 * With both conditions met, we can be relatively sure that the relationship is
5694 * non-monogamous, with partner count exceeding socket size.
5696 * Waker/wakee being client/server, worker/dispatcher, interrupt source or
5697 * whatever is irrelevant, spread criteria is apparent partner count exceeds
5700 static int wake_wide(struct task_struct *p)
5702 unsigned int master = current->wakee_flips;
5703 unsigned int slave = p->wakee_flips;
5704 int factor = this_cpu_read(sd_llc_size);
5707 swap(master, slave);
5708 if (slave < factor || master < slave * factor)
5714 * The purpose of wake_affine() is to quickly determine on which CPU we can run
5715 * soonest. For the purpose of speed we only consider the waking and previous
5718 * wake_affine_idle() - only considers 'now', it check if the waking CPU is
5719 * cache-affine and is (or will be) idle.
5721 * wake_affine_weight() - considers the weight to reflect the average
5722 * scheduling latency of the CPUs. This seems to work
5723 * for the overloaded case.
5726 wake_affine_idle(int this_cpu, int prev_cpu, int sync)
5729 * If this_cpu is idle, it implies the wakeup is from interrupt
5730 * context. Only allow the move if cache is shared. Otherwise an
5731 * interrupt intensive workload could force all tasks onto one
5732 * node depending on the IO topology or IRQ affinity settings.
5734 * If the prev_cpu is idle and cache affine then avoid a migration.
5735 * There is no guarantee that the cache hot data from an interrupt
5736 * is more important than cache hot data on the prev_cpu and from
5737 * a cpufreq perspective, it's better to have higher utilisation
5740 if (available_idle_cpu(this_cpu) && cpus_share_cache(this_cpu, prev_cpu))
5741 return available_idle_cpu(prev_cpu) ? prev_cpu : this_cpu;
5743 if (sync && cpu_rq(this_cpu)->nr_running == 1)
5746 return nr_cpumask_bits;
5750 wake_affine_weight(struct sched_domain *sd, struct task_struct *p,
5751 int this_cpu, int prev_cpu, int sync)
5753 s64 this_eff_load, prev_eff_load;
5754 unsigned long task_load;
5756 this_eff_load = target_load(this_cpu, sd->wake_idx);
5759 unsigned long current_load = task_h_load(current);
5761 if (current_load > this_eff_load)
5764 this_eff_load -= current_load;
5767 task_load = task_h_load(p);
5769 this_eff_load += task_load;
5770 if (sched_feat(WA_BIAS))
5771 this_eff_load *= 100;
5772 this_eff_load *= capacity_of(prev_cpu);
5774 prev_eff_load = source_load(prev_cpu, sd->wake_idx);
5775 prev_eff_load -= task_load;
5776 if (sched_feat(WA_BIAS))
5777 prev_eff_load *= 100 + (sd->imbalance_pct - 100) / 2;
5778 prev_eff_load *= capacity_of(this_cpu);
5781 * If sync, adjust the weight of prev_eff_load such that if
5782 * prev_eff == this_eff that select_idle_sibling() will consider
5783 * stacking the wakee on top of the waker if no other CPU is
5789 return this_eff_load < prev_eff_load ? this_cpu : nr_cpumask_bits;
5792 static int wake_affine(struct sched_domain *sd, struct task_struct *p,
5793 int this_cpu, int prev_cpu, int sync)
5795 int target = nr_cpumask_bits;
5797 if (sched_feat(WA_IDLE))
5798 target = wake_affine_idle(this_cpu, prev_cpu, sync);
5800 if (sched_feat(WA_WEIGHT) && target == nr_cpumask_bits)
5801 target = wake_affine_weight(sd, p, this_cpu, prev_cpu, sync);
5803 schedstat_inc(p->se.statistics.nr_wakeups_affine_attempts);
5804 if (target == nr_cpumask_bits)
5807 schedstat_inc(sd->ttwu_move_affine);
5808 schedstat_inc(p->se.statistics.nr_wakeups_affine);
5812 static unsigned long cpu_util_without(int cpu, struct task_struct *p);
5814 static unsigned long capacity_spare_without(int cpu, struct task_struct *p)
5816 return max_t(long, capacity_of(cpu) - cpu_util_without(cpu, p), 0);
5820 * find_idlest_group finds and returns the least busy CPU group within the
5823 * Assumes p is allowed on at least one CPU in sd.
5825 static struct sched_group *
5826 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
5827 int this_cpu, int sd_flag)
5829 struct sched_group *idlest = NULL, *group = sd->groups;
5830 struct sched_group *most_spare_sg = NULL;
5831 unsigned long min_runnable_load = ULONG_MAX;
5832 unsigned long this_runnable_load = ULONG_MAX;
5833 unsigned long min_avg_load = ULONG_MAX, this_avg_load = ULONG_MAX;
5834 unsigned long most_spare = 0, this_spare = 0;
5835 int load_idx = sd->forkexec_idx;
5836 int imbalance_scale = 100 + (sd->imbalance_pct-100)/2;
5837 unsigned long imbalance = scale_load_down(NICE_0_LOAD) *
5838 (sd->imbalance_pct-100) / 100;
5840 if (sd_flag & SD_BALANCE_WAKE)
5841 load_idx = sd->wake_idx;
5844 unsigned long load, avg_load, runnable_load;
5845 unsigned long spare_cap, max_spare_cap;
5849 /* Skip over this group if it has no CPUs allowed */
5850 if (!cpumask_intersects(sched_group_span(group),
5854 local_group = cpumask_test_cpu(this_cpu,
5855 sched_group_span(group));
5858 * Tally up the load of all CPUs in the group and find
5859 * the group containing the CPU with most spare capacity.
5865 for_each_cpu(i, sched_group_span(group)) {
5866 /* Bias balancing toward CPUs of our domain */
5868 load = source_load(i, load_idx);
5870 load = target_load(i, load_idx);
5872 runnable_load += load;
5874 avg_load += cfs_rq_load_avg(&cpu_rq(i)->cfs);
5876 spare_cap = capacity_spare_without(i, p);
5878 if (spare_cap > max_spare_cap)
5879 max_spare_cap = spare_cap;
5882 /* Adjust by relative CPU capacity of the group */
5883 avg_load = (avg_load * SCHED_CAPACITY_SCALE) /
5884 group->sgc->capacity;
5885 runnable_load = (runnable_load * SCHED_CAPACITY_SCALE) /
5886 group->sgc->capacity;
5889 this_runnable_load = runnable_load;
5890 this_avg_load = avg_load;
5891 this_spare = max_spare_cap;
5893 if (min_runnable_load > (runnable_load + imbalance)) {
5895 * The runnable load is significantly smaller
5896 * so we can pick this new CPU:
5898 min_runnable_load = runnable_load;
5899 min_avg_load = avg_load;
5901 } else if ((runnable_load < (min_runnable_load + imbalance)) &&
5902 (100*min_avg_load > imbalance_scale*avg_load)) {
5904 * The runnable loads are close so take the
5905 * blocked load into account through avg_load:
5907 min_avg_load = avg_load;
5911 if (most_spare < max_spare_cap) {
5912 most_spare = max_spare_cap;
5913 most_spare_sg = group;
5916 } while (group = group->next, group != sd->groups);
5919 * The cross-over point between using spare capacity or least load
5920 * is too conservative for high utilization tasks on partially
5921 * utilized systems if we require spare_capacity > task_util(p),
5922 * so we allow for some task stuffing by using
5923 * spare_capacity > task_util(p)/2.
5925 * Spare capacity can't be used for fork because the utilization has
5926 * not been set yet, we must first select a rq to compute the initial
5929 if (sd_flag & SD_BALANCE_FORK)
5932 if (this_spare > task_util(p) / 2 &&
5933 imbalance_scale*this_spare > 100*most_spare)
5936 if (most_spare > task_util(p) / 2)
5937 return most_spare_sg;
5944 * When comparing groups across NUMA domains, it's possible for the
5945 * local domain to be very lightly loaded relative to the remote
5946 * domains but "imbalance" skews the comparison making remote CPUs
5947 * look much more favourable. When considering cross-domain, add
5948 * imbalance to the runnable load on the remote node and consider
5951 if ((sd->flags & SD_NUMA) &&
5952 min_runnable_load + imbalance >= this_runnable_load)
5955 if (min_runnable_load > (this_runnable_load + imbalance))
5958 if ((this_runnable_load < (min_runnable_load + imbalance)) &&
5959 (100*this_avg_load < imbalance_scale*min_avg_load))
5966 * find_idlest_group_cpu - find the idlest CPU among the CPUs in the group.
5969 find_idlest_group_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
5971 unsigned long load, min_load = ULONG_MAX;
5972 unsigned int min_exit_latency = UINT_MAX;
5973 u64 latest_idle_timestamp = 0;
5974 int least_loaded_cpu = this_cpu;
5975 int shallowest_idle_cpu = -1;
5978 /* Check if we have any choice: */
5979 if (group->group_weight == 1)
5980 return cpumask_first(sched_group_span(group));
5982 /* Traverse only the allowed CPUs */
5983 for_each_cpu_and(i, sched_group_span(group), &p->cpus_allowed) {
5984 if (available_idle_cpu(i)) {
5985 struct rq *rq = cpu_rq(i);
5986 struct cpuidle_state *idle = idle_get_state(rq);
5987 if (idle && idle->exit_latency < min_exit_latency) {
5989 * We give priority to a CPU whose idle state
5990 * has the smallest exit latency irrespective
5991 * of any idle timestamp.
5993 min_exit_latency = idle->exit_latency;
5994 latest_idle_timestamp = rq->idle_stamp;
5995 shallowest_idle_cpu = i;
5996 } else if ((!idle || idle->exit_latency == min_exit_latency) &&
5997 rq->idle_stamp > latest_idle_timestamp) {
5999 * If equal or no active idle state, then
6000 * the most recently idled CPU might have
6003 latest_idle_timestamp = rq->idle_stamp;
6004 shallowest_idle_cpu = i;
6006 } else if (shallowest_idle_cpu == -1) {
6007 load = weighted_cpuload(cpu_rq(i));
6008 if (load < min_load) {
6010 least_loaded_cpu = i;
6015 return shallowest_idle_cpu != -1 ? shallowest_idle_cpu : least_loaded_cpu;
6018 static inline int find_idlest_cpu(struct sched_domain *sd, struct task_struct *p,
6019 int cpu, int prev_cpu, int sd_flag)
6023 if (!cpumask_intersects(sched_domain_span(sd), &p->cpus_allowed))
6027 * We need task's util for capacity_spare_without, sync it up to
6028 * prev_cpu's last_update_time.
6030 if (!(sd_flag & SD_BALANCE_FORK))
6031 sync_entity_load_avg(&p->se);
6034 struct sched_group *group;
6035 struct sched_domain *tmp;
6038 if (!(sd->flags & sd_flag)) {
6043 group = find_idlest_group(sd, p, cpu, sd_flag);
6049 new_cpu = find_idlest_group_cpu(group, p, cpu);
6050 if (new_cpu == cpu) {
6051 /* Now try balancing at a lower domain level of 'cpu': */
6056 /* Now try balancing at a lower domain level of 'new_cpu': */
6058 weight = sd->span_weight;
6060 for_each_domain(cpu, tmp) {
6061 if (weight <= tmp->span_weight)
6063 if (tmp->flags & sd_flag)
6071 #ifdef CONFIG_SCHED_SMT
6072 DEFINE_STATIC_KEY_FALSE(sched_smt_present);
6073 EXPORT_SYMBOL_GPL(sched_smt_present);
6075 static inline void set_idle_cores(int cpu, int val)
6077 struct sched_domain_shared *sds;
6079 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
6081 WRITE_ONCE(sds->has_idle_cores, val);
6084 static inline bool test_idle_cores(int cpu, bool def)
6086 struct sched_domain_shared *sds;
6088 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
6090 return READ_ONCE(sds->has_idle_cores);
6096 * Scans the local SMT mask to see if the entire core is idle, and records this
6097 * information in sd_llc_shared->has_idle_cores.
6099 * Since SMT siblings share all cache levels, inspecting this limited remote
6100 * state should be fairly cheap.
6102 void __update_idle_core(struct rq *rq)
6104 int core = cpu_of(rq);
6108 if (test_idle_cores(core, true))
6111 for_each_cpu(cpu, cpu_smt_mask(core)) {
6115 if (!available_idle_cpu(cpu))
6119 set_idle_cores(core, 1);
6125 * Scan the entire LLC domain for idle cores; this dynamically switches off if
6126 * there are no idle cores left in the system; tracked through
6127 * sd_llc->shared->has_idle_cores and enabled through update_idle_core() above.
6129 static int select_idle_core(struct task_struct *p, struct sched_domain *sd, int target)
6131 struct cpumask *cpus = this_cpu_cpumask_var_ptr(select_idle_mask);
6134 if (!static_branch_likely(&sched_smt_present))
6137 if (!test_idle_cores(target, false))
6140 cpumask_and(cpus, sched_domain_span(sd), &p->cpus_allowed);
6142 for_each_cpu_wrap(core, cpus, target) {
6145 for_each_cpu(cpu, cpu_smt_mask(core)) {
6146 cpumask_clear_cpu(cpu, cpus);
6147 if (!available_idle_cpu(cpu))
6156 * Failed to find an idle core; stop looking for one.
6158 set_idle_cores(target, 0);
6164 * Scan the local SMT mask for idle CPUs.
6166 static int select_idle_smt(struct task_struct *p, struct sched_domain *sd, int target)
6170 if (!static_branch_likely(&sched_smt_present))
6173 for_each_cpu(cpu, cpu_smt_mask(target)) {
6174 if (!cpumask_test_cpu(cpu, &p->cpus_allowed))
6176 if (available_idle_cpu(cpu))
6183 #else /* CONFIG_SCHED_SMT */
6185 static inline int select_idle_core(struct task_struct *p, struct sched_domain *sd, int target)
6190 static inline int select_idle_smt(struct task_struct *p, struct sched_domain *sd, int target)
6195 #endif /* CONFIG_SCHED_SMT */
6198 * Scan the LLC domain for idle CPUs; this is dynamically regulated by
6199 * comparing the average scan cost (tracked in sd->avg_scan_cost) against the
6200 * average idle time for this rq (as found in rq->avg_idle).
6202 static int select_idle_cpu(struct task_struct *p, struct sched_domain *sd, int target)
6204 struct cpumask *cpus = this_cpu_cpumask_var_ptr(select_idle_mask);
6205 struct sched_domain *this_sd;
6206 u64 avg_cost, avg_idle;
6209 int cpu, nr = INT_MAX;
6211 this_sd = rcu_dereference(*this_cpu_ptr(&sd_llc));
6216 * Due to large variance we need a large fuzz factor; hackbench in
6217 * particularly is sensitive here.
6219 avg_idle = this_rq()->avg_idle / 512;
6220 avg_cost = this_sd->avg_scan_cost + 1;
6222 if (sched_feat(SIS_AVG_CPU) && avg_idle < avg_cost)
6225 if (sched_feat(SIS_PROP)) {
6226 u64 span_avg = sd->span_weight * avg_idle;
6227 if (span_avg > 4*avg_cost)
6228 nr = div_u64(span_avg, avg_cost);
6233 time = local_clock();
6235 cpumask_and(cpus, sched_domain_span(sd), &p->cpus_allowed);
6237 for_each_cpu_wrap(cpu, cpus, target) {
6240 if (available_idle_cpu(cpu))
6244 time = local_clock() - time;
6245 cost = this_sd->avg_scan_cost;
6246 delta = (s64)(time - cost) / 8;
6247 this_sd->avg_scan_cost += delta;
6253 * Try and locate an idle core/thread in the LLC cache domain.
6255 static int select_idle_sibling(struct task_struct *p, int prev, int target)
6257 struct sched_domain *sd;
6258 int i, recent_used_cpu;
6260 if (available_idle_cpu(target))
6264 * If the previous CPU is cache affine and idle, don't be stupid:
6266 if (prev != target && cpus_share_cache(prev, target) && available_idle_cpu(prev))
6269 /* Check a recently used CPU as a potential idle candidate: */
6270 recent_used_cpu = p->recent_used_cpu;
6271 if (recent_used_cpu != prev &&
6272 recent_used_cpu != target &&
6273 cpus_share_cache(recent_used_cpu, target) &&
6274 available_idle_cpu(recent_used_cpu) &&
6275 cpumask_test_cpu(p->recent_used_cpu, &p->cpus_allowed)) {
6277 * Replace recent_used_cpu with prev as it is a potential
6278 * candidate for the next wake:
6280 p->recent_used_cpu = prev;
6281 return recent_used_cpu;
6284 sd = rcu_dereference(per_cpu(sd_llc, target));
6288 i = select_idle_core(p, sd, target);
6289 if ((unsigned)i < nr_cpumask_bits)
6292 i = select_idle_cpu(p, sd, target);
6293 if ((unsigned)i < nr_cpumask_bits)
6296 i = select_idle_smt(p, sd, target);
6297 if ((unsigned)i < nr_cpumask_bits)
6304 * Amount of capacity of a CPU that is (estimated to be) used by CFS tasks
6305 * @cpu: the CPU to get the utilization of
6307 * The unit of the return value must be the one of capacity so we can compare
6308 * the utilization with the capacity of the CPU that is available for CFS task
6309 * (ie cpu_capacity).
6311 * cfs_rq.avg.util_avg is the sum of running time of runnable tasks plus the
6312 * recent utilization of currently non-runnable tasks on a CPU. It represents
6313 * the amount of utilization of a CPU in the range [0..capacity_orig] where
6314 * capacity_orig is the cpu_capacity available at the highest frequency
6315 * (arch_scale_freq_capacity()).
6316 * The utilization of a CPU converges towards a sum equal to or less than the
6317 * current capacity (capacity_curr <= capacity_orig) of the CPU because it is
6318 * the running time on this CPU scaled by capacity_curr.
6320 * The estimated utilization of a CPU is defined to be the maximum between its
6321 * cfs_rq.avg.util_avg and the sum of the estimated utilization of the tasks
6322 * currently RUNNABLE on that CPU.
6323 * This allows to properly represent the expected utilization of a CPU which
6324 * has just got a big task running since a long sleep period. At the same time
6325 * however it preserves the benefits of the "blocked utilization" in
6326 * describing the potential for other tasks waking up on the same CPU.
6328 * Nevertheless, cfs_rq.avg.util_avg can be higher than capacity_curr or even
6329 * higher than capacity_orig because of unfortunate rounding in
6330 * cfs.avg.util_avg or just after migrating tasks and new task wakeups until
6331 * the average stabilizes with the new running time. We need to check that the
6332 * utilization stays within the range of [0..capacity_orig] and cap it if
6333 * necessary. Without utilization capping, a group could be seen as overloaded
6334 * (CPU0 utilization at 121% + CPU1 utilization at 80%) whereas CPU1 has 20% of
6335 * available capacity. We allow utilization to overshoot capacity_curr (but not
6336 * capacity_orig) as it useful for predicting the capacity required after task
6337 * migrations (scheduler-driven DVFS).
6339 * Return: the (estimated) utilization for the specified CPU
6341 static inline unsigned long cpu_util(int cpu)
6343 struct cfs_rq *cfs_rq;
6346 cfs_rq = &cpu_rq(cpu)->cfs;
6347 util = READ_ONCE(cfs_rq->avg.util_avg);
6349 if (sched_feat(UTIL_EST))
6350 util = max(util, READ_ONCE(cfs_rq->avg.util_est.enqueued));
6352 return min_t(unsigned long, util, capacity_orig_of(cpu));
6356 * cpu_util_without: compute cpu utilization without any contributions from *p
6357 * @cpu: the CPU which utilization is requested
6358 * @p: the task which utilization should be discounted
6360 * The utilization of a CPU is defined by the utilization of tasks currently
6361 * enqueued on that CPU as well as tasks which are currently sleeping after an
6362 * execution on that CPU.
6364 * This method returns the utilization of the specified CPU by discounting the
6365 * utilization of the specified task, whenever the task is currently
6366 * contributing to the CPU utilization.
6368 static unsigned long cpu_util_without(int cpu, struct task_struct *p)
6370 struct cfs_rq *cfs_rq;
6373 /* Task has no contribution or is new */
6374 if (cpu != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time))
6375 return cpu_util(cpu);
6377 cfs_rq = &cpu_rq(cpu)->cfs;
6378 util = READ_ONCE(cfs_rq->avg.util_avg);
6380 /* Discount task's util from CPU's util */
6381 util -= min_t(unsigned int, util, task_util(p));
6386 * a) if *p is the only task sleeping on this CPU, then:
6387 * cpu_util (== task_util) > util_est (== 0)
6388 * and thus we return:
6389 * cpu_util_without = (cpu_util - task_util) = 0
6391 * b) if other tasks are SLEEPING on this CPU, which is now exiting
6393 * cpu_util >= task_util
6394 * cpu_util > util_est (== 0)
6395 * and thus we discount *p's blocked utilization to return:
6396 * cpu_util_without = (cpu_util - task_util) >= 0
6398 * c) if other tasks are RUNNABLE on that CPU and
6399 * util_est > cpu_util
6400 * then we use util_est since it returns a more restrictive
6401 * estimation of the spare capacity on that CPU, by just
6402 * considering the expected utilization of tasks already
6403 * runnable on that CPU.
6405 * Cases a) and b) are covered by the above code, while case c) is
6406 * covered by the following code when estimated utilization is
6409 if (sched_feat(UTIL_EST)) {
6410 unsigned int estimated =
6411 READ_ONCE(cfs_rq->avg.util_est.enqueued);
6414 * Despite the following checks we still have a small window
6415 * for a possible race, when an execl's select_task_rq_fair()
6416 * races with LB's detach_task():
6419 * p->on_rq = TASK_ON_RQ_MIGRATING;
6420 * ---------------------------------- A
6421 * deactivate_task() \
6422 * dequeue_task() + RaceTime
6423 * util_est_dequeue() /
6424 * ---------------------------------- B
6426 * The additional check on "current == p" it's required to
6427 * properly fix the execl regression and it helps in further
6428 * reducing the chances for the above race.
6430 if (unlikely(task_on_rq_queued(p) || current == p)) {
6431 estimated -= min_t(unsigned int, estimated,
6432 (_task_util_est(p) | UTIL_AVG_UNCHANGED));
6434 util = max(util, estimated);
6438 * Utilization (estimated) can exceed the CPU capacity, thus let's
6439 * clamp to the maximum CPU capacity to ensure consistency with
6440 * the cpu_util call.
6442 return min_t(unsigned long, util, capacity_orig_of(cpu));
6446 * Disable WAKE_AFFINE in the case where task @p doesn't fit in the
6447 * capacity of either the waking CPU @cpu or the previous CPU @prev_cpu.
6449 * In that case WAKE_AFFINE doesn't make sense and we'll let
6450 * BALANCE_WAKE sort things out.
6452 static int wake_cap(struct task_struct *p, int cpu, int prev_cpu)
6454 long min_cap, max_cap;
6456 min_cap = min(capacity_orig_of(prev_cpu), capacity_orig_of(cpu));
6457 max_cap = cpu_rq(cpu)->rd->max_cpu_capacity;
6459 /* Minimum capacity is close to max, no need to abort wake_affine */
6460 if (max_cap - min_cap < max_cap >> 3)
6463 /* Bring task utilization in sync with prev_cpu */
6464 sync_entity_load_avg(&p->se);
6466 return min_cap * 1024 < task_util(p) * capacity_margin;
6470 * select_task_rq_fair: Select target runqueue for the waking task in domains
6471 * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE,
6472 * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
6474 * Balances load by selecting the idlest CPU in the idlest group, or under
6475 * certain conditions an idle sibling CPU if the domain has SD_WAKE_AFFINE set.
6477 * Returns the target CPU number.
6479 * preempt must be disabled.
6482 select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags)
6484 struct sched_domain *tmp, *sd = NULL;
6485 int cpu = smp_processor_id();
6486 int new_cpu = prev_cpu;
6487 int want_affine = 0;
6488 int sync = (wake_flags & WF_SYNC) && !(current->flags & PF_EXITING);
6490 if (sd_flag & SD_BALANCE_WAKE) {
6492 want_affine = !wake_wide(p) && !wake_cap(p, cpu, prev_cpu)
6493 && cpumask_test_cpu(cpu, &p->cpus_allowed);
6497 for_each_domain(cpu, tmp) {
6498 if (!(tmp->flags & SD_LOAD_BALANCE))
6502 * If both 'cpu' and 'prev_cpu' are part of this domain,
6503 * cpu is a valid SD_WAKE_AFFINE target.
6505 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
6506 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
6507 if (cpu != prev_cpu)
6508 new_cpu = wake_affine(tmp, p, cpu, prev_cpu, sync);
6510 sd = NULL; /* Prefer wake_affine over balance flags */
6514 if (tmp->flags & sd_flag)
6516 else if (!want_affine)
6522 new_cpu = find_idlest_cpu(sd, p, cpu, prev_cpu, sd_flag);
6523 } else if (sd_flag & SD_BALANCE_WAKE) { /* XXX always ? */
6526 new_cpu = select_idle_sibling(p, prev_cpu, new_cpu);
6529 current->recent_used_cpu = cpu;
6536 static void detach_entity_cfs_rq(struct sched_entity *se);
6539 * Called immediately before a task is migrated to a new CPU; task_cpu(p) and
6540 * cfs_rq_of(p) references at time of call are still valid and identify the
6541 * previous CPU. The caller guarantees p->pi_lock or task_rq(p)->lock is held.
6543 static void migrate_task_rq_fair(struct task_struct *p, int new_cpu)
6546 * As blocked tasks retain absolute vruntime the migration needs to
6547 * deal with this by subtracting the old and adding the new
6548 * min_vruntime -- the latter is done by enqueue_entity() when placing
6549 * the task on the new runqueue.
6551 if (p->state == TASK_WAKING) {
6552 struct sched_entity *se = &p->se;
6553 struct cfs_rq *cfs_rq = cfs_rq_of(se);
6556 #ifndef CONFIG_64BIT
6557 u64 min_vruntime_copy;
6560 min_vruntime_copy = cfs_rq->min_vruntime_copy;
6562 min_vruntime = cfs_rq->min_vruntime;
6563 } while (min_vruntime != min_vruntime_copy);
6565 min_vruntime = cfs_rq->min_vruntime;
6568 se->vruntime -= min_vruntime;
6571 if (p->on_rq == TASK_ON_RQ_MIGRATING) {
6573 * In case of TASK_ON_RQ_MIGRATING we in fact hold the 'old'
6574 * rq->lock and can modify state directly.
6576 lockdep_assert_held(&task_rq(p)->lock);
6577 detach_entity_cfs_rq(&p->se);
6581 * We are supposed to update the task to "current" time, then
6582 * its up to date and ready to go to new CPU/cfs_rq. But we
6583 * have difficulty in getting what current time is, so simply
6584 * throw away the out-of-date time. This will result in the
6585 * wakee task is less decayed, but giving the wakee more load
6588 remove_entity_load_avg(&p->se);
6591 /* Tell new CPU we are migrated */
6592 p->se.avg.last_update_time = 0;
6594 update_scan_period(p, new_cpu);
6597 static void task_dead_fair(struct task_struct *p)
6599 remove_entity_load_avg(&p->se);
6601 #endif /* CONFIG_SMP */
6603 static unsigned long wakeup_gran(struct sched_entity *se)
6605 unsigned long gran = sysctl_sched_wakeup_granularity;
6608 * Since its curr running now, convert the gran from real-time
6609 * to virtual-time in his units.
6611 * By using 'se' instead of 'curr' we penalize light tasks, so
6612 * they get preempted easier. That is, if 'se' < 'curr' then
6613 * the resulting gran will be larger, therefore penalizing the
6614 * lighter, if otoh 'se' > 'curr' then the resulting gran will
6615 * be smaller, again penalizing the lighter task.
6617 * This is especially important for buddies when the leftmost
6618 * task is higher priority than the buddy.
6620 return calc_delta_fair(gran, se);
6624 * Should 'se' preempt 'curr'.
6638 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
6640 s64 gran, vdiff = curr->vruntime - se->vruntime;
6645 gran = wakeup_gran(se);
6652 static void set_last_buddy(struct sched_entity *se)
6654 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
6657 for_each_sched_entity(se) {
6658 if (SCHED_WARN_ON(!se->on_rq))
6660 cfs_rq_of(se)->last = se;
6664 static void set_next_buddy(struct sched_entity *se)
6666 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
6669 for_each_sched_entity(se) {
6670 if (SCHED_WARN_ON(!se->on_rq))
6672 cfs_rq_of(se)->next = se;
6676 static void set_skip_buddy(struct sched_entity *se)
6678 for_each_sched_entity(se)
6679 cfs_rq_of(se)->skip = se;
6683 * Preempt the current task with a newly woken task if needed:
6685 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
6687 struct task_struct *curr = rq->curr;
6688 struct sched_entity *se = &curr->se, *pse = &p->se;
6689 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
6690 int scale = cfs_rq->nr_running >= sched_nr_latency;
6691 int next_buddy_marked = 0;
6693 if (unlikely(se == pse))
6697 * This is possible from callers such as attach_tasks(), in which we
6698 * unconditionally check_prempt_curr() after an enqueue (which may have
6699 * lead to a throttle). This both saves work and prevents false
6700 * next-buddy nomination below.
6702 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
6705 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
6706 set_next_buddy(pse);
6707 next_buddy_marked = 1;
6711 * We can come here with TIF_NEED_RESCHED already set from new task
6714 * Note: this also catches the edge-case of curr being in a throttled
6715 * group (e.g. via set_curr_task), since update_curr() (in the
6716 * enqueue of curr) will have resulted in resched being set. This
6717 * prevents us from potentially nominating it as a false LAST_BUDDY
6720 if (test_tsk_need_resched(curr))
6723 /* Idle tasks are by definition preempted by non-idle tasks. */
6724 if (unlikely(curr->policy == SCHED_IDLE) &&
6725 likely(p->policy != SCHED_IDLE))
6729 * Batch and idle tasks do not preempt non-idle tasks (their preemption
6730 * is driven by the tick):
6732 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
6735 find_matching_se(&se, &pse);
6736 update_curr(cfs_rq_of(se));
6738 if (wakeup_preempt_entity(se, pse) == 1) {
6740 * Bias pick_next to pick the sched entity that is
6741 * triggering this preemption.
6743 if (!next_buddy_marked)
6744 set_next_buddy(pse);
6753 * Only set the backward buddy when the current task is still
6754 * on the rq. This can happen when a wakeup gets interleaved
6755 * with schedule on the ->pre_schedule() or idle_balance()
6756 * point, either of which can * drop the rq lock.
6758 * Also, during early boot the idle thread is in the fair class,
6759 * for obvious reasons its a bad idea to schedule back to it.
6761 if (unlikely(!se->on_rq || curr == rq->idle))
6764 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
6768 static struct task_struct *
6769 pick_next_task_fair(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
6771 struct cfs_rq *cfs_rq = &rq->cfs;
6772 struct sched_entity *se;
6773 struct task_struct *p;
6777 if (!cfs_rq->nr_running)
6780 #ifdef CONFIG_FAIR_GROUP_SCHED
6781 if (prev->sched_class != &fair_sched_class)
6785 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
6786 * likely that a next task is from the same cgroup as the current.
6788 * Therefore attempt to avoid putting and setting the entire cgroup
6789 * hierarchy, only change the part that actually changes.
6793 struct sched_entity *curr = cfs_rq->curr;
6796 * Since we got here without doing put_prev_entity() we also
6797 * have to consider cfs_rq->curr. If it is still a runnable
6798 * entity, update_curr() will update its vruntime, otherwise
6799 * forget we've ever seen it.
6803 update_curr(cfs_rq);
6808 * This call to check_cfs_rq_runtime() will do the
6809 * throttle and dequeue its entity in the parent(s).
6810 * Therefore the nr_running test will indeed
6813 if (unlikely(check_cfs_rq_runtime(cfs_rq))) {
6816 if (!cfs_rq->nr_running)
6823 se = pick_next_entity(cfs_rq, curr);
6824 cfs_rq = group_cfs_rq(se);
6830 * Since we haven't yet done put_prev_entity and if the selected task
6831 * is a different task than we started out with, try and touch the
6832 * least amount of cfs_rqs.
6835 struct sched_entity *pse = &prev->se;
6837 while (!(cfs_rq = is_same_group(se, pse))) {
6838 int se_depth = se->depth;
6839 int pse_depth = pse->depth;
6841 if (se_depth <= pse_depth) {
6842 put_prev_entity(cfs_rq_of(pse), pse);
6843 pse = parent_entity(pse);
6845 if (se_depth >= pse_depth) {
6846 set_next_entity(cfs_rq_of(se), se);
6847 se = parent_entity(se);
6851 put_prev_entity(cfs_rq, pse);
6852 set_next_entity(cfs_rq, se);
6859 put_prev_task(rq, prev);
6862 se = pick_next_entity(cfs_rq, NULL);
6863 set_next_entity(cfs_rq, se);
6864 cfs_rq = group_cfs_rq(se);
6869 done: __maybe_unused;
6872 * Move the next running task to the front of
6873 * the list, so our cfs_tasks list becomes MRU
6876 list_move(&p->se.group_node, &rq->cfs_tasks);
6879 if (hrtick_enabled(rq))
6880 hrtick_start_fair(rq, p);
6885 new_tasks = idle_balance(rq, rf);
6888 * Because idle_balance() releases (and re-acquires) rq->lock, it is
6889 * possible for any higher priority task to appear. In that case we
6890 * must re-start the pick_next_entity() loop.
6902 * Account for a descheduled task:
6904 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
6906 struct sched_entity *se = &prev->se;
6907 struct cfs_rq *cfs_rq;
6909 for_each_sched_entity(se) {
6910 cfs_rq = cfs_rq_of(se);
6911 put_prev_entity(cfs_rq, se);
6916 * sched_yield() is very simple
6918 * The magic of dealing with the ->skip buddy is in pick_next_entity.
6920 static void yield_task_fair(struct rq *rq)
6922 struct task_struct *curr = rq->curr;
6923 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
6924 struct sched_entity *se = &curr->se;
6927 * Are we the only task in the tree?
6929 if (unlikely(rq->nr_running == 1))
6932 clear_buddies(cfs_rq, se);
6934 if (curr->policy != SCHED_BATCH) {
6935 update_rq_clock(rq);
6937 * Update run-time statistics of the 'current'.
6939 update_curr(cfs_rq);
6941 * Tell update_rq_clock() that we've just updated,
6942 * so we don't do microscopic update in schedule()
6943 * and double the fastpath cost.
6945 rq_clock_skip_update(rq);
6951 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
6953 struct sched_entity *se = &p->se;
6955 /* throttled hierarchies are not runnable */
6956 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
6959 /* Tell the scheduler that we'd really like pse to run next. */
6962 yield_task_fair(rq);
6968 /**************************************************
6969 * Fair scheduling class load-balancing methods.
6973 * The purpose of load-balancing is to achieve the same basic fairness the
6974 * per-CPU scheduler provides, namely provide a proportional amount of compute
6975 * time to each task. This is expressed in the following equation:
6977 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
6979 * Where W_i,n is the n-th weight average for CPU i. The instantaneous weight
6980 * W_i,0 is defined as:
6982 * W_i,0 = \Sum_j w_i,j (2)
6984 * Where w_i,j is the weight of the j-th runnable task on CPU i. This weight
6985 * is derived from the nice value as per sched_prio_to_weight[].
6987 * The weight average is an exponential decay average of the instantaneous
6990 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
6992 * C_i is the compute capacity of CPU i, typically it is the
6993 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
6994 * can also include other factors [XXX].
6996 * To achieve this balance we define a measure of imbalance which follows
6997 * directly from (1):
6999 * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4)
7001 * We them move tasks around to minimize the imbalance. In the continuous
7002 * function space it is obvious this converges, in the discrete case we get
7003 * a few fun cases generally called infeasible weight scenarios.
7006 * - infeasible weights;
7007 * - local vs global optima in the discrete case. ]
7012 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
7013 * for all i,j solution, we create a tree of CPUs that follows the hardware
7014 * topology where each level pairs two lower groups (or better). This results
7015 * in O(log n) layers. Furthermore we reduce the number of CPUs going up the
7016 * tree to only the first of the previous level and we decrease the frequency
7017 * of load-balance at each level inv. proportional to the number of CPUs in
7023 * \Sum { --- * --- * 2^i } = O(n) (5)
7025 * `- size of each group
7026 * | | `- number of CPUs doing load-balance
7028 * `- sum over all levels
7030 * Coupled with a limit on how many tasks we can migrate every balance pass,
7031 * this makes (5) the runtime complexity of the balancer.
7033 * An important property here is that each CPU is still (indirectly) connected
7034 * to every other CPU in at most O(log n) steps:
7036 * The adjacency matrix of the resulting graph is given by:
7039 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
7042 * And you'll find that:
7044 * A^(log_2 n)_i,j != 0 for all i,j (7)
7046 * Showing there's indeed a path between every CPU in at most O(log n) steps.
7047 * The task movement gives a factor of O(m), giving a convergence complexity
7050 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
7055 * In order to avoid CPUs going idle while there's still work to do, new idle
7056 * balancing is more aggressive and has the newly idle CPU iterate up the domain
7057 * tree itself instead of relying on other CPUs to bring it work.
7059 * This adds some complexity to both (5) and (8) but it reduces the total idle
7067 * Cgroups make a horror show out of (2), instead of a simple sum we get:
7070 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
7075 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
7077 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on CPU i.
7079 * The big problem is S_k, its a global sum needed to compute a local (W_i)
7082 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
7083 * rewrite all of this once again.]
7086 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
7088 enum fbq_type { regular, remote, all };
7090 #define LBF_ALL_PINNED 0x01
7091 #define LBF_NEED_BREAK 0x02
7092 #define LBF_DST_PINNED 0x04
7093 #define LBF_SOME_PINNED 0x08
7094 #define LBF_NOHZ_STATS 0x10
7095 #define LBF_NOHZ_AGAIN 0x20
7098 struct sched_domain *sd;
7106 struct cpumask *dst_grpmask;
7108 enum cpu_idle_type idle;
7110 /* The set of CPUs under consideration for load-balancing */
7111 struct cpumask *cpus;
7116 unsigned int loop_break;
7117 unsigned int loop_max;
7119 enum fbq_type fbq_type;
7120 struct list_head tasks;
7124 * Is this task likely cache-hot:
7126 static int task_hot(struct task_struct *p, struct lb_env *env)
7130 lockdep_assert_held(&env->src_rq->lock);
7132 if (p->sched_class != &fair_sched_class)
7135 if (unlikely(p->policy == SCHED_IDLE))
7139 * Buddy candidates are cache hot:
7141 if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running &&
7142 (&p->se == cfs_rq_of(&p->se)->next ||
7143 &p->se == cfs_rq_of(&p->se)->last))
7146 if (sysctl_sched_migration_cost == -1)
7148 if (sysctl_sched_migration_cost == 0)
7151 delta = rq_clock_task(env->src_rq) - p->se.exec_start;
7153 return delta < (s64)sysctl_sched_migration_cost;
7156 #ifdef CONFIG_NUMA_BALANCING
7158 * Returns 1, if task migration degrades locality
7159 * Returns 0, if task migration improves locality i.e migration preferred.
7160 * Returns -1, if task migration is not affected by locality.
7162 static int migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
7164 struct numa_group *numa_group = rcu_dereference(p->numa_group);
7165 unsigned long src_weight, dst_weight;
7166 int src_nid, dst_nid, dist;
7168 if (!static_branch_likely(&sched_numa_balancing))
7171 if (!p->numa_faults || !(env->sd->flags & SD_NUMA))
7174 src_nid = cpu_to_node(env->src_cpu);
7175 dst_nid = cpu_to_node(env->dst_cpu);
7177 if (src_nid == dst_nid)
7180 /* Migrating away from the preferred node is always bad. */
7181 if (src_nid == p->numa_preferred_nid) {
7182 if (env->src_rq->nr_running > env->src_rq->nr_preferred_running)
7188 /* Encourage migration to the preferred node. */
7189 if (dst_nid == p->numa_preferred_nid)
7192 /* Leaving a core idle is often worse than degrading locality. */
7193 if (env->idle == CPU_IDLE)
7196 dist = node_distance(src_nid, dst_nid);
7198 src_weight = group_weight(p, src_nid, dist);
7199 dst_weight = group_weight(p, dst_nid, dist);
7201 src_weight = task_weight(p, src_nid, dist);
7202 dst_weight = task_weight(p, dst_nid, dist);
7205 return dst_weight < src_weight;
7209 static inline int migrate_degrades_locality(struct task_struct *p,
7217 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
7220 int can_migrate_task(struct task_struct *p, struct lb_env *env)
7224 lockdep_assert_held(&env->src_rq->lock);
7227 * We do not migrate tasks that are:
7228 * 1) throttled_lb_pair, or
7229 * 2) cannot be migrated to this CPU due to cpus_allowed, or
7230 * 3) running (obviously), or
7231 * 4) are cache-hot on their current CPU.
7233 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
7236 if (!cpumask_test_cpu(env->dst_cpu, &p->cpus_allowed)) {
7239 schedstat_inc(p->se.statistics.nr_failed_migrations_affine);
7241 env->flags |= LBF_SOME_PINNED;
7244 * Remember if this task can be migrated to any other CPU in
7245 * our sched_group. We may want to revisit it if we couldn't
7246 * meet load balance goals by pulling other tasks on src_cpu.
7248 * Avoid computing new_dst_cpu for NEWLY_IDLE or if we have
7249 * already computed one in current iteration.
7251 if (env->idle == CPU_NEWLY_IDLE || (env->flags & LBF_DST_PINNED))
7254 /* Prevent to re-select dst_cpu via env's CPUs: */
7255 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
7256 if (cpumask_test_cpu(cpu, &p->cpus_allowed)) {
7257 env->flags |= LBF_DST_PINNED;
7258 env->new_dst_cpu = cpu;
7266 /* Record that we found atleast one task that could run on dst_cpu */
7267 env->flags &= ~LBF_ALL_PINNED;
7269 if (task_running(env->src_rq, p)) {
7270 schedstat_inc(p->se.statistics.nr_failed_migrations_running);
7275 * Aggressive migration if:
7276 * 1) destination numa is preferred
7277 * 2) task is cache cold, or
7278 * 3) too many balance attempts have failed.
7280 tsk_cache_hot = migrate_degrades_locality(p, env);
7281 if (tsk_cache_hot == -1)
7282 tsk_cache_hot = task_hot(p, env);
7284 if (tsk_cache_hot <= 0 ||
7285 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
7286 if (tsk_cache_hot == 1) {
7287 schedstat_inc(env->sd->lb_hot_gained[env->idle]);
7288 schedstat_inc(p->se.statistics.nr_forced_migrations);
7293 schedstat_inc(p->se.statistics.nr_failed_migrations_hot);
7298 * detach_task() -- detach the task for the migration specified in env
7300 static void detach_task(struct task_struct *p, struct lb_env *env)
7302 lockdep_assert_held(&env->src_rq->lock);
7304 p->on_rq = TASK_ON_RQ_MIGRATING;
7305 deactivate_task(env->src_rq, p, DEQUEUE_NOCLOCK);
7306 set_task_cpu(p, env->dst_cpu);
7310 * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
7311 * part of active balancing operations within "domain".
7313 * Returns a task if successful and NULL otherwise.
7315 static struct task_struct *detach_one_task(struct lb_env *env)
7317 struct task_struct *p;
7319 lockdep_assert_held(&env->src_rq->lock);
7321 list_for_each_entry_reverse(p,
7322 &env->src_rq->cfs_tasks, se.group_node) {
7323 if (!can_migrate_task(p, env))
7326 detach_task(p, env);
7329 * Right now, this is only the second place where
7330 * lb_gained[env->idle] is updated (other is detach_tasks)
7331 * so we can safely collect stats here rather than
7332 * inside detach_tasks().
7334 schedstat_inc(env->sd->lb_gained[env->idle]);
7340 static const unsigned int sched_nr_migrate_break = 32;
7343 * detach_tasks() -- tries to detach up to imbalance weighted load from
7344 * busiest_rq, as part of a balancing operation within domain "sd".
7346 * Returns number of detached tasks if successful and 0 otherwise.
7348 static int detach_tasks(struct lb_env *env)
7350 struct list_head *tasks = &env->src_rq->cfs_tasks;
7351 struct task_struct *p;
7355 lockdep_assert_held(&env->src_rq->lock);
7357 if (env->imbalance <= 0)
7360 while (!list_empty(tasks)) {
7362 * We don't want to steal all, otherwise we may be treated likewise,
7363 * which could at worst lead to a livelock crash.
7365 if (env->idle != CPU_NOT_IDLE && env->src_rq->nr_running <= 1)
7368 p = list_last_entry(tasks, struct task_struct, se.group_node);
7371 /* We've more or less seen every task there is, call it quits */
7372 if (env->loop > env->loop_max)
7375 /* take a breather every nr_migrate tasks */
7376 if (env->loop > env->loop_break) {
7377 env->loop_break += sched_nr_migrate_break;
7378 env->flags |= LBF_NEED_BREAK;
7382 if (!can_migrate_task(p, env))
7386 * Depending of the number of CPUs and tasks and the
7387 * cgroup hierarchy, task_h_load() can return a null
7388 * value. Make sure that env->imbalance decreases
7389 * otherwise detach_tasks() will stop only after
7390 * detaching up to loop_max tasks.
7392 load = max_t(unsigned long, task_h_load(p), 1);
7395 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
7398 if ((load / 2) > env->imbalance)
7401 detach_task(p, env);
7402 list_add(&p->se.group_node, &env->tasks);
7405 env->imbalance -= load;
7407 #ifdef CONFIG_PREEMPT
7409 * NEWIDLE balancing is a source of latency, so preemptible
7410 * kernels will stop after the first task is detached to minimize
7411 * the critical section.
7413 if (env->idle == CPU_NEWLY_IDLE)
7418 * We only want to steal up to the prescribed amount of
7421 if (env->imbalance <= 0)
7426 list_move(&p->se.group_node, tasks);
7430 * Right now, this is one of only two places we collect this stat
7431 * so we can safely collect detach_one_task() stats here rather
7432 * than inside detach_one_task().
7434 schedstat_add(env->sd->lb_gained[env->idle], detached);
7440 * attach_task() -- attach the task detached by detach_task() to its new rq.
7442 static void attach_task(struct rq *rq, struct task_struct *p)
7444 lockdep_assert_held(&rq->lock);
7446 BUG_ON(task_rq(p) != rq);
7447 activate_task(rq, p, ENQUEUE_NOCLOCK);
7448 p->on_rq = TASK_ON_RQ_QUEUED;
7449 check_preempt_curr(rq, p, 0);
7453 * attach_one_task() -- attaches the task returned from detach_one_task() to
7456 static void attach_one_task(struct rq *rq, struct task_struct *p)
7461 update_rq_clock(rq);
7467 * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
7470 static void attach_tasks(struct lb_env *env)
7472 struct list_head *tasks = &env->tasks;
7473 struct task_struct *p;
7476 rq_lock(env->dst_rq, &rf);
7477 update_rq_clock(env->dst_rq);
7479 while (!list_empty(tasks)) {
7480 p = list_first_entry(tasks, struct task_struct, se.group_node);
7481 list_del_init(&p->se.group_node);
7483 attach_task(env->dst_rq, p);
7486 rq_unlock(env->dst_rq, &rf);
7489 static inline bool cfs_rq_has_blocked(struct cfs_rq *cfs_rq)
7491 if (cfs_rq->avg.load_avg)
7494 if (cfs_rq->avg.util_avg)
7500 static inline bool others_have_blocked(struct rq *rq)
7502 if (READ_ONCE(rq->avg_rt.util_avg))
7505 if (READ_ONCE(rq->avg_dl.util_avg))
7508 #ifdef CONFIG_HAVE_SCHED_AVG_IRQ
7509 if (READ_ONCE(rq->avg_irq.util_avg))
7516 #ifdef CONFIG_FAIR_GROUP_SCHED
7518 static inline bool cfs_rq_is_decayed(struct cfs_rq *cfs_rq)
7520 if (cfs_rq->load.weight)
7523 if (cfs_rq->avg.load_sum)
7526 if (cfs_rq->avg.util_sum)
7529 if (cfs_rq->avg.runnable_load_sum)
7535 static void update_blocked_averages(int cpu)
7537 struct rq *rq = cpu_rq(cpu);
7538 struct cfs_rq *cfs_rq, *pos;
7539 const struct sched_class *curr_class;
7543 rq_lock_irqsave(rq, &rf);
7544 update_rq_clock(rq);
7547 * Iterates the task_group tree in a bottom up fashion, see
7548 * list_add_leaf_cfs_rq() for details.
7550 for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) {
7551 struct sched_entity *se;
7553 if (update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq))
7554 update_tg_load_avg(cfs_rq, 0);
7556 /* Propagate pending load changes to the parent, if any: */
7557 se = cfs_rq->tg->se[cpu];
7558 if (se && !skip_blocked_update(se))
7559 update_load_avg(cfs_rq_of(se), se, UPDATE_TG);
7562 * There can be a lot of idle CPU cgroups. Don't let fully
7563 * decayed cfs_rqs linger on the list.
7565 if (cfs_rq_is_decayed(cfs_rq))
7566 list_del_leaf_cfs_rq(cfs_rq);
7568 /* Don't need periodic decay once load/util_avg are null */
7569 if (cfs_rq_has_blocked(cfs_rq))
7573 curr_class = rq->curr->sched_class;
7574 update_rt_rq_load_avg(rq_clock_task(rq), rq, curr_class == &rt_sched_class);
7575 update_dl_rq_load_avg(rq_clock_task(rq), rq, curr_class == &dl_sched_class);
7576 update_irq_load_avg(rq, 0);
7577 /* Don't need periodic decay once load/util_avg are null */
7578 if (others_have_blocked(rq))
7581 #ifdef CONFIG_NO_HZ_COMMON
7582 rq->last_blocked_load_update_tick = jiffies;
7584 rq->has_blocked_load = 0;
7586 rq_unlock_irqrestore(rq, &rf);
7590 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
7591 * This needs to be done in a top-down fashion because the load of a child
7592 * group is a fraction of its parents load.
7594 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
7596 struct rq *rq = rq_of(cfs_rq);
7597 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
7598 unsigned long now = jiffies;
7601 if (cfs_rq->last_h_load_update == now)
7604 WRITE_ONCE(cfs_rq->h_load_next, NULL);
7605 for_each_sched_entity(se) {
7606 cfs_rq = cfs_rq_of(se);
7607 WRITE_ONCE(cfs_rq->h_load_next, se);
7608 if (cfs_rq->last_h_load_update == now)
7613 cfs_rq->h_load = cfs_rq_load_avg(cfs_rq);
7614 cfs_rq->last_h_load_update = now;
7617 while ((se = READ_ONCE(cfs_rq->h_load_next)) != NULL) {
7618 load = cfs_rq->h_load;
7619 load = div64_ul(load * se->avg.load_avg,
7620 cfs_rq_load_avg(cfs_rq) + 1);
7621 cfs_rq = group_cfs_rq(se);
7622 cfs_rq->h_load = load;
7623 cfs_rq->last_h_load_update = now;
7627 static unsigned long task_h_load(struct task_struct *p)
7629 struct cfs_rq *cfs_rq = task_cfs_rq(p);
7631 update_cfs_rq_h_load(cfs_rq);
7632 return div64_ul(p->se.avg.load_avg * cfs_rq->h_load,
7633 cfs_rq_load_avg(cfs_rq) + 1);
7636 static inline void update_blocked_averages(int cpu)
7638 struct rq *rq = cpu_rq(cpu);
7639 struct cfs_rq *cfs_rq = &rq->cfs;
7640 const struct sched_class *curr_class;
7643 rq_lock_irqsave(rq, &rf);
7644 update_rq_clock(rq);
7645 update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq);
7647 curr_class = rq->curr->sched_class;
7648 update_rt_rq_load_avg(rq_clock_task(rq), rq, curr_class == &rt_sched_class);
7649 update_dl_rq_load_avg(rq_clock_task(rq), rq, curr_class == &dl_sched_class);
7650 update_irq_load_avg(rq, 0);
7651 #ifdef CONFIG_NO_HZ_COMMON
7652 rq->last_blocked_load_update_tick = jiffies;
7653 if (!cfs_rq_has_blocked(cfs_rq) && !others_have_blocked(rq))
7654 rq->has_blocked_load = 0;
7656 rq_unlock_irqrestore(rq, &rf);
7659 static unsigned long task_h_load(struct task_struct *p)
7661 return p->se.avg.load_avg;
7665 /********** Helpers for find_busiest_group ************************/
7674 * sg_lb_stats - stats of a sched_group required for load_balancing
7676 struct sg_lb_stats {
7677 unsigned long avg_load; /*Avg load across the CPUs of the group */
7678 unsigned long group_load; /* Total load over the CPUs of the group */
7679 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
7680 unsigned long load_per_task;
7681 unsigned long group_capacity;
7682 unsigned long group_util; /* Total utilization of the group */
7683 unsigned int sum_nr_running; /* Nr tasks running in the group */
7684 unsigned int idle_cpus;
7685 unsigned int group_weight;
7686 enum group_type group_type;
7687 int group_no_capacity;
7688 #ifdef CONFIG_NUMA_BALANCING
7689 unsigned int nr_numa_running;
7690 unsigned int nr_preferred_running;
7695 * sd_lb_stats - Structure to store the statistics of a sched_domain
7696 * during load balancing.
7698 struct sd_lb_stats {
7699 struct sched_group *busiest; /* Busiest group in this sd */
7700 struct sched_group *local; /* Local group in this sd */
7701 unsigned long total_running;
7702 unsigned long total_load; /* Total load of all groups in sd */
7703 unsigned long total_capacity; /* Total capacity of all groups in sd */
7704 unsigned long avg_load; /* Average load across all groups in sd */
7706 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
7707 struct sg_lb_stats local_stat; /* Statistics of the local group */
7710 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
7713 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
7714 * local_stat because update_sg_lb_stats() does a full clear/assignment.
7715 * We must however clear busiest_stat::avg_load because
7716 * update_sd_pick_busiest() reads this before assignment.
7718 *sds = (struct sd_lb_stats){
7721 .total_running = 0UL,
7723 .total_capacity = 0UL,
7726 .sum_nr_running = 0,
7727 .group_type = group_other,
7733 * get_sd_load_idx - Obtain the load index for a given sched domain.
7734 * @sd: The sched_domain whose load_idx is to be obtained.
7735 * @idle: The idle status of the CPU for whose sd load_idx is obtained.
7737 * Return: The load index.
7739 static inline int get_sd_load_idx(struct sched_domain *sd,
7740 enum cpu_idle_type idle)
7746 load_idx = sd->busy_idx;
7749 case CPU_NEWLY_IDLE:
7750 load_idx = sd->newidle_idx;
7753 load_idx = sd->idle_idx;
7760 static unsigned long scale_rt_capacity(struct sched_domain *sd, int cpu)
7762 struct rq *rq = cpu_rq(cpu);
7763 unsigned long max = arch_scale_cpu_capacity(sd, cpu);
7764 unsigned long used, free;
7767 irq = cpu_util_irq(rq);
7769 if (unlikely(irq >= max))
7772 used = READ_ONCE(rq->avg_rt.util_avg);
7773 used += READ_ONCE(rq->avg_dl.util_avg);
7775 if (unlikely(used >= max))
7780 return scale_irq_capacity(free, irq, max);
7783 static void update_cpu_capacity(struct sched_domain *sd, int cpu)
7785 unsigned long capacity = scale_rt_capacity(sd, cpu);
7786 struct sched_group *sdg = sd->groups;
7788 cpu_rq(cpu)->cpu_capacity_orig = arch_scale_cpu_capacity(sd, cpu);
7793 cpu_rq(cpu)->cpu_capacity = capacity;
7794 sdg->sgc->capacity = capacity;
7795 sdg->sgc->min_capacity = capacity;
7798 void update_group_capacity(struct sched_domain *sd, int cpu)
7800 struct sched_domain *child = sd->child;
7801 struct sched_group *group, *sdg = sd->groups;
7802 unsigned long capacity, min_capacity;
7803 unsigned long interval;
7805 interval = msecs_to_jiffies(sd->balance_interval);
7806 interval = clamp(interval, 1UL, max_load_balance_interval);
7807 sdg->sgc->next_update = jiffies + interval;
7810 update_cpu_capacity(sd, cpu);
7815 min_capacity = ULONG_MAX;
7817 if (child->flags & SD_OVERLAP) {
7819 * SD_OVERLAP domains cannot assume that child groups
7820 * span the current group.
7823 for_each_cpu(cpu, sched_group_span(sdg)) {
7824 struct sched_group_capacity *sgc;
7825 struct rq *rq = cpu_rq(cpu);
7828 * build_sched_domains() -> init_sched_groups_capacity()
7829 * gets here before we've attached the domains to the
7832 * Use capacity_of(), which is set irrespective of domains
7833 * in update_cpu_capacity().
7835 * This avoids capacity from being 0 and
7836 * causing divide-by-zero issues on boot.
7838 if (unlikely(!rq->sd)) {
7839 capacity += capacity_of(cpu);
7841 sgc = rq->sd->groups->sgc;
7842 capacity += sgc->capacity;
7845 min_capacity = min(capacity, min_capacity);
7849 * !SD_OVERLAP domains can assume that child groups
7850 * span the current group.
7853 group = child->groups;
7855 struct sched_group_capacity *sgc = group->sgc;
7857 capacity += sgc->capacity;
7858 min_capacity = min(sgc->min_capacity, min_capacity);
7859 group = group->next;
7860 } while (group != child->groups);
7863 sdg->sgc->capacity = capacity;
7864 sdg->sgc->min_capacity = min_capacity;
7868 * Check whether the capacity of the rq has been noticeably reduced by side
7869 * activity. The imbalance_pct is used for the threshold.
7870 * Return true is the capacity is reduced
7873 check_cpu_capacity(struct rq *rq, struct sched_domain *sd)
7875 return ((rq->cpu_capacity * sd->imbalance_pct) <
7876 (rq->cpu_capacity_orig * 100));
7880 * Group imbalance indicates (and tries to solve) the problem where balancing
7881 * groups is inadequate due to ->cpus_allowed constraints.
7883 * Imagine a situation of two groups of 4 CPUs each and 4 tasks each with a
7884 * cpumask covering 1 CPU of the first group and 3 CPUs of the second group.
7887 * { 0 1 2 3 } { 4 5 6 7 }
7890 * If we were to balance group-wise we'd place two tasks in the first group and
7891 * two tasks in the second group. Clearly this is undesired as it will overload
7892 * cpu 3 and leave one of the CPUs in the second group unused.
7894 * The current solution to this issue is detecting the skew in the first group
7895 * by noticing the lower domain failed to reach balance and had difficulty
7896 * moving tasks due to affinity constraints.
7898 * When this is so detected; this group becomes a candidate for busiest; see
7899 * update_sd_pick_busiest(). And calculate_imbalance() and
7900 * find_busiest_group() avoid some of the usual balance conditions to allow it
7901 * to create an effective group imbalance.
7903 * This is a somewhat tricky proposition since the next run might not find the
7904 * group imbalance and decide the groups need to be balanced again. A most
7905 * subtle and fragile situation.
7908 static inline int sg_imbalanced(struct sched_group *group)
7910 return group->sgc->imbalance;
7914 * group_has_capacity returns true if the group has spare capacity that could
7915 * be used by some tasks.
7916 * We consider that a group has spare capacity if the * number of task is
7917 * smaller than the number of CPUs or if the utilization is lower than the
7918 * available capacity for CFS tasks.
7919 * For the latter, we use a threshold to stabilize the state, to take into
7920 * account the variance of the tasks' load and to return true if the available
7921 * capacity in meaningful for the load balancer.
7922 * As an example, an available capacity of 1% can appear but it doesn't make
7923 * any benefit for the load balance.
7926 group_has_capacity(struct lb_env *env, struct sg_lb_stats *sgs)
7928 if (sgs->sum_nr_running < sgs->group_weight)
7931 if ((sgs->group_capacity * 100) >
7932 (sgs->group_util * env->sd->imbalance_pct))
7939 * group_is_overloaded returns true if the group has more tasks than it can
7941 * group_is_overloaded is not equals to !group_has_capacity because a group
7942 * with the exact right number of tasks, has no more spare capacity but is not
7943 * overloaded so both group_has_capacity and group_is_overloaded return
7947 group_is_overloaded(struct lb_env *env, struct sg_lb_stats *sgs)
7949 if (sgs->sum_nr_running <= sgs->group_weight)
7952 if ((sgs->group_capacity * 100) <
7953 (sgs->group_util * env->sd->imbalance_pct))
7960 * group_smaller_cpu_capacity: Returns true if sched_group sg has smaller
7961 * per-CPU capacity than sched_group ref.
7964 group_smaller_cpu_capacity(struct sched_group *sg, struct sched_group *ref)
7966 return sg->sgc->min_capacity * capacity_margin <
7967 ref->sgc->min_capacity * 1024;
7971 group_type group_classify(struct sched_group *group,
7972 struct sg_lb_stats *sgs)
7974 if (sgs->group_no_capacity)
7975 return group_overloaded;
7977 if (sg_imbalanced(group))
7978 return group_imbalanced;
7983 static bool update_nohz_stats(struct rq *rq, bool force)
7985 #ifdef CONFIG_NO_HZ_COMMON
7986 unsigned int cpu = rq->cpu;
7988 if (!rq->has_blocked_load)
7991 if (!cpumask_test_cpu(cpu, nohz.idle_cpus_mask))
7994 if (!force && !time_after(jiffies, rq->last_blocked_load_update_tick))
7997 update_blocked_averages(cpu);
7999 return rq->has_blocked_load;
8006 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
8007 * @env: The load balancing environment.
8008 * @group: sched_group whose statistics are to be updated.
8009 * @load_idx: Load index of sched_domain of this_cpu for load calc.
8010 * @local_group: Does group contain this_cpu.
8011 * @sgs: variable to hold the statistics for this group.
8012 * @overload: Indicate more than one runnable task for any CPU.
8014 static inline void update_sg_lb_stats(struct lb_env *env,
8015 struct sched_group *group, int load_idx,
8016 int local_group, struct sg_lb_stats *sgs,
8022 memset(sgs, 0, sizeof(*sgs));
8024 for_each_cpu_and(i, sched_group_span(group), env->cpus) {
8025 struct rq *rq = cpu_rq(i);
8027 if ((env->flags & LBF_NOHZ_STATS) && update_nohz_stats(rq, false))
8028 env->flags |= LBF_NOHZ_AGAIN;
8030 /* Bias balancing toward CPUs of our domain: */
8032 load = target_load(i, load_idx);
8034 load = source_load(i, load_idx);
8036 sgs->group_load += load;
8037 sgs->group_util += cpu_util(i);
8038 sgs->sum_nr_running += rq->cfs.h_nr_running;
8040 nr_running = rq->nr_running;
8044 #ifdef CONFIG_NUMA_BALANCING
8045 sgs->nr_numa_running += rq->nr_numa_running;
8046 sgs->nr_preferred_running += rq->nr_preferred_running;
8048 sgs->sum_weighted_load += weighted_cpuload(rq);
8050 * No need to call idle_cpu() if nr_running is not 0
8052 if (!nr_running && idle_cpu(i))
8056 /* Adjust by relative CPU capacity of the group */
8057 sgs->group_capacity = group->sgc->capacity;
8058 sgs->avg_load = (sgs->group_load*SCHED_CAPACITY_SCALE) / sgs->group_capacity;
8060 if (sgs->sum_nr_running)
8061 sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
8063 sgs->group_weight = group->group_weight;
8065 sgs->group_no_capacity = group_is_overloaded(env, sgs);
8066 sgs->group_type = group_classify(group, sgs);
8070 * update_sd_pick_busiest - return 1 on busiest group
8071 * @env: The load balancing environment.
8072 * @sds: sched_domain statistics
8073 * @sg: sched_group candidate to be checked for being the busiest
8074 * @sgs: sched_group statistics
8076 * Determine if @sg is a busier group than the previously selected
8079 * Return: %true if @sg is a busier group than the previously selected
8080 * busiest group. %false otherwise.
8082 static bool update_sd_pick_busiest(struct lb_env *env,
8083 struct sd_lb_stats *sds,
8084 struct sched_group *sg,
8085 struct sg_lb_stats *sgs)
8087 struct sg_lb_stats *busiest = &sds->busiest_stat;
8089 if (sgs->group_type > busiest->group_type)
8092 if (sgs->group_type < busiest->group_type)
8095 if (sgs->avg_load <= busiest->avg_load)
8098 if (!(env->sd->flags & SD_ASYM_CPUCAPACITY))
8102 * Candidate sg has no more than one task per CPU and
8103 * has higher per-CPU capacity. Migrating tasks to less
8104 * capable CPUs may harm throughput. Maximize throughput,
8105 * power/energy consequences are not considered.
8107 if (sgs->sum_nr_running <= sgs->group_weight &&
8108 group_smaller_cpu_capacity(sds->local, sg))
8112 /* This is the busiest node in its class. */
8113 if (!(env->sd->flags & SD_ASYM_PACKING))
8116 /* No ASYM_PACKING if target CPU is already busy */
8117 if (env->idle == CPU_NOT_IDLE)
8120 * ASYM_PACKING needs to move all the work to the highest
8121 * prority CPUs in the group, therefore mark all groups
8122 * of lower priority than ourself as busy.
8124 if (sgs->sum_nr_running &&
8125 sched_asym_prefer(env->dst_cpu, sg->asym_prefer_cpu)) {
8129 /* Prefer to move from lowest priority CPU's work */
8130 if (sched_asym_prefer(sds->busiest->asym_prefer_cpu,
8131 sg->asym_prefer_cpu))
8138 #ifdef CONFIG_NUMA_BALANCING
8139 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
8141 if (sgs->sum_nr_running > sgs->nr_numa_running)
8143 if (sgs->sum_nr_running > sgs->nr_preferred_running)
8148 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
8150 if (rq->nr_running > rq->nr_numa_running)
8152 if (rq->nr_running > rq->nr_preferred_running)
8157 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
8162 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
8166 #endif /* CONFIG_NUMA_BALANCING */
8169 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
8170 * @env: The load balancing environment.
8171 * @sds: variable to hold the statistics for this sched_domain.
8173 static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
8175 struct sched_domain *child = env->sd->child;
8176 struct sched_group *sg = env->sd->groups;
8177 struct sg_lb_stats *local = &sds->local_stat;
8178 struct sg_lb_stats tmp_sgs;
8179 int load_idx, prefer_sibling = 0;
8180 bool overload = false;
8182 if (child && child->flags & SD_PREFER_SIBLING)
8185 #ifdef CONFIG_NO_HZ_COMMON
8186 if (env->idle == CPU_NEWLY_IDLE && READ_ONCE(nohz.has_blocked))
8187 env->flags |= LBF_NOHZ_STATS;
8190 load_idx = get_sd_load_idx(env->sd, env->idle);
8193 struct sg_lb_stats *sgs = &tmp_sgs;
8196 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_span(sg));
8201 if (env->idle != CPU_NEWLY_IDLE ||
8202 time_after_eq(jiffies, sg->sgc->next_update))
8203 update_group_capacity(env->sd, env->dst_cpu);
8206 update_sg_lb_stats(env, sg, load_idx, local_group, sgs,
8213 * In case the child domain prefers tasks go to siblings
8214 * first, lower the sg capacity so that we'll try
8215 * and move all the excess tasks away. We lower the capacity
8216 * of a group only if the local group has the capacity to fit
8217 * these excess tasks. The extra check prevents the case where
8218 * you always pull from the heaviest group when it is already
8219 * under-utilized (possible with a large weight task outweighs
8220 * the tasks on the system).
8222 if (prefer_sibling && sds->local &&
8223 group_has_capacity(env, local) &&
8224 (sgs->sum_nr_running > local->sum_nr_running + 1)) {
8225 sgs->group_no_capacity = 1;
8226 sgs->group_type = group_classify(sg, sgs);
8229 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
8231 sds->busiest_stat = *sgs;
8235 /* Now, start updating sd_lb_stats */
8236 sds->total_running += sgs->sum_nr_running;
8237 sds->total_load += sgs->group_load;
8238 sds->total_capacity += sgs->group_capacity;
8241 } while (sg != env->sd->groups);
8243 #ifdef CONFIG_NO_HZ_COMMON
8244 if ((env->flags & LBF_NOHZ_AGAIN) &&
8245 cpumask_subset(nohz.idle_cpus_mask, sched_domain_span(env->sd))) {
8247 WRITE_ONCE(nohz.next_blocked,
8248 jiffies + msecs_to_jiffies(LOAD_AVG_PERIOD));
8252 if (env->sd->flags & SD_NUMA)
8253 env->fbq_type = fbq_classify_group(&sds->busiest_stat);
8255 if (!env->sd->parent) {
8256 /* update overload indicator if we are at root domain */
8257 if (env->dst_rq->rd->overload != overload)
8258 env->dst_rq->rd->overload = overload;
8263 * check_asym_packing - Check to see if the group is packed into the
8266 * This is primarily intended to used at the sibling level. Some
8267 * cores like POWER7 prefer to use lower numbered SMT threads. In the
8268 * case of POWER7, it can move to lower SMT modes only when higher
8269 * threads are idle. When in lower SMT modes, the threads will
8270 * perform better since they share less core resources. Hence when we
8271 * have idle threads, we want them to be the higher ones.
8273 * This packing function is run on idle threads. It checks to see if
8274 * the busiest CPU in this domain (core in the P7 case) has a higher
8275 * CPU number than the packing function is being run on. Here we are
8276 * assuming lower CPU number will be equivalent to lower a SMT thread
8279 * Return: 1 when packing is required and a task should be moved to
8280 * this CPU. The amount of the imbalance is returned in env->imbalance.
8282 * @env: The load balancing environment.
8283 * @sds: Statistics of the sched_domain which is to be packed
8285 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
8289 if (!(env->sd->flags & SD_ASYM_PACKING))
8292 if (env->idle == CPU_NOT_IDLE)
8298 busiest_cpu = sds->busiest->asym_prefer_cpu;
8299 if (sched_asym_prefer(busiest_cpu, env->dst_cpu))
8302 env->imbalance = DIV_ROUND_CLOSEST(
8303 sds->busiest_stat.avg_load * sds->busiest_stat.group_capacity,
8304 SCHED_CAPACITY_SCALE);
8310 * fix_small_imbalance - Calculate the minor imbalance that exists
8311 * amongst the groups of a sched_domain, during
8313 * @env: The load balancing environment.
8314 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
8317 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
8319 unsigned long tmp, capa_now = 0, capa_move = 0;
8320 unsigned int imbn = 2;
8321 unsigned long scaled_busy_load_per_task;
8322 struct sg_lb_stats *local, *busiest;
8324 local = &sds->local_stat;
8325 busiest = &sds->busiest_stat;
8327 if (!local->sum_nr_running)
8328 local->load_per_task = cpu_avg_load_per_task(env->dst_cpu);
8329 else if (busiest->load_per_task > local->load_per_task)
8332 scaled_busy_load_per_task =
8333 (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
8334 busiest->group_capacity;
8336 if (busiest->avg_load + scaled_busy_load_per_task >=
8337 local->avg_load + (scaled_busy_load_per_task * imbn)) {
8338 env->imbalance = busiest->load_per_task;
8343 * OK, we don't have enough imbalance to justify moving tasks,
8344 * however we may be able to increase total CPU capacity used by
8348 capa_now += busiest->group_capacity *
8349 min(busiest->load_per_task, busiest->avg_load);
8350 capa_now += local->group_capacity *
8351 min(local->load_per_task, local->avg_load);
8352 capa_now /= SCHED_CAPACITY_SCALE;
8354 /* Amount of load we'd subtract */
8355 if (busiest->avg_load > scaled_busy_load_per_task) {
8356 capa_move += busiest->group_capacity *
8357 min(busiest->load_per_task,
8358 busiest->avg_load - scaled_busy_load_per_task);
8361 /* Amount of load we'd add */
8362 if (busiest->avg_load * busiest->group_capacity <
8363 busiest->load_per_task * SCHED_CAPACITY_SCALE) {
8364 tmp = (busiest->avg_load * busiest->group_capacity) /
8365 local->group_capacity;
8367 tmp = (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
8368 local->group_capacity;
8370 capa_move += local->group_capacity *
8371 min(local->load_per_task, local->avg_load + tmp);
8372 capa_move /= SCHED_CAPACITY_SCALE;
8374 /* Move if we gain throughput */
8375 if (capa_move > capa_now)
8376 env->imbalance = busiest->load_per_task;
8380 * calculate_imbalance - Calculate the amount of imbalance present within the
8381 * groups of a given sched_domain during load balance.
8382 * @env: load balance environment
8383 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
8385 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
8387 unsigned long max_pull, load_above_capacity = ~0UL;
8388 struct sg_lb_stats *local, *busiest;
8390 local = &sds->local_stat;
8391 busiest = &sds->busiest_stat;
8393 if (busiest->group_type == group_imbalanced) {
8395 * In the group_imb case we cannot rely on group-wide averages
8396 * to ensure CPU-load equilibrium, look at wider averages. XXX
8398 busiest->load_per_task =
8399 min(busiest->load_per_task, sds->avg_load);
8403 * Avg load of busiest sg can be less and avg load of local sg can
8404 * be greater than avg load across all sgs of sd because avg load
8405 * factors in sg capacity and sgs with smaller group_type are
8406 * skipped when updating the busiest sg:
8408 if (busiest->avg_load <= sds->avg_load ||
8409 local->avg_load >= sds->avg_load) {
8411 return fix_small_imbalance(env, sds);
8415 * If there aren't any idle CPUs, avoid creating some.
8417 if (busiest->group_type == group_overloaded &&
8418 local->group_type == group_overloaded) {
8419 load_above_capacity = busiest->sum_nr_running * SCHED_CAPACITY_SCALE;
8420 if (load_above_capacity > busiest->group_capacity) {
8421 load_above_capacity -= busiest->group_capacity;
8422 load_above_capacity *= scale_load_down(NICE_0_LOAD);
8423 load_above_capacity /= busiest->group_capacity;
8425 load_above_capacity = ~0UL;
8429 * We're trying to get all the CPUs to the average_load, so we don't
8430 * want to push ourselves above the average load, nor do we wish to
8431 * reduce the max loaded CPU below the average load. At the same time,
8432 * we also don't want to reduce the group load below the group
8433 * capacity. Thus we look for the minimum possible imbalance.
8435 max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity);
8437 /* How much load to actually move to equalise the imbalance */
8438 env->imbalance = min(
8439 max_pull * busiest->group_capacity,
8440 (sds->avg_load - local->avg_load) * local->group_capacity
8441 ) / SCHED_CAPACITY_SCALE;
8444 * if *imbalance is less than the average load per runnable task
8445 * there is no guarantee that any tasks will be moved so we'll have
8446 * a think about bumping its value to force at least one task to be
8449 if (env->imbalance < busiest->load_per_task)
8450 return fix_small_imbalance(env, sds);
8453 /******* find_busiest_group() helpers end here *********************/
8456 * find_busiest_group - Returns the busiest group within the sched_domain
8457 * if there is an imbalance.
8459 * Also calculates the amount of weighted load which should be moved
8460 * to restore balance.
8462 * @env: The load balancing environment.
8464 * Return: - The busiest group if imbalance exists.
8466 static struct sched_group *find_busiest_group(struct lb_env *env)
8468 struct sg_lb_stats *local, *busiest;
8469 struct sd_lb_stats sds;
8471 init_sd_lb_stats(&sds);
8474 * Compute the various statistics relavent for load balancing at
8477 update_sd_lb_stats(env, &sds);
8478 local = &sds.local_stat;
8479 busiest = &sds.busiest_stat;
8481 /* ASYM feature bypasses nice load balance check */
8482 if (check_asym_packing(env, &sds))
8485 /* There is no busy sibling group to pull tasks from */
8486 if (!sds.busiest || busiest->sum_nr_running == 0)
8489 /* XXX broken for overlapping NUMA groups */
8490 sds.avg_load = (SCHED_CAPACITY_SCALE * sds.total_load)
8491 / sds.total_capacity;
8494 * If the busiest group is imbalanced the below checks don't
8495 * work because they assume all things are equal, which typically
8496 * isn't true due to cpus_allowed constraints and the like.
8498 if (busiest->group_type == group_imbalanced)
8502 * When dst_cpu is idle, prevent SMP nice and/or asymmetric group
8503 * capacities from resulting in underutilization due to avg_load.
8505 if (env->idle != CPU_NOT_IDLE && group_has_capacity(env, local) &&
8506 busiest->group_no_capacity)
8510 * If the local group is busier than the selected busiest group
8511 * don't try and pull any tasks.
8513 if (local->avg_load >= busiest->avg_load)
8517 * Don't pull any tasks if this group is already above the domain
8520 if (local->avg_load >= sds.avg_load)
8523 if (env->idle == CPU_IDLE) {
8525 * This CPU is idle. If the busiest group is not overloaded
8526 * and there is no imbalance between this and busiest group
8527 * wrt idle CPUs, it is balanced. The imbalance becomes
8528 * significant if the diff is greater than 1 otherwise we
8529 * might end up to just move the imbalance on another group
8531 if ((busiest->group_type != group_overloaded) &&
8532 (local->idle_cpus <= (busiest->idle_cpus + 1)))
8536 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
8537 * imbalance_pct to be conservative.
8539 if (100 * busiest->avg_load <=
8540 env->sd->imbalance_pct * local->avg_load)
8545 /* Looks like there is an imbalance. Compute it */
8546 calculate_imbalance(env, &sds);
8547 return env->imbalance ? sds.busiest : NULL;
8555 * find_busiest_queue - find the busiest runqueue among the CPUs in the group.
8557 static struct rq *find_busiest_queue(struct lb_env *env,
8558 struct sched_group *group)
8560 struct rq *busiest = NULL, *rq;
8561 unsigned long busiest_load = 0, busiest_capacity = 1;
8564 for_each_cpu_and(i, sched_group_span(group), env->cpus) {
8565 unsigned long capacity, wl;
8569 rt = fbq_classify_rq(rq);
8572 * We classify groups/runqueues into three groups:
8573 * - regular: there are !numa tasks
8574 * - remote: there are numa tasks that run on the 'wrong' node
8575 * - all: there is no distinction
8577 * In order to avoid migrating ideally placed numa tasks,
8578 * ignore those when there's better options.
8580 * If we ignore the actual busiest queue to migrate another
8581 * task, the next balance pass can still reduce the busiest
8582 * queue by moving tasks around inside the node.
8584 * If we cannot move enough load due to this classification
8585 * the next pass will adjust the group classification and
8586 * allow migration of more tasks.
8588 * Both cases only affect the total convergence complexity.
8590 if (rt > env->fbq_type)
8593 capacity = capacity_of(i);
8595 wl = weighted_cpuload(rq);
8598 * When comparing with imbalance, use weighted_cpuload()
8599 * which is not scaled with the CPU capacity.
8602 if (rq->nr_running == 1 && wl > env->imbalance &&
8603 !check_cpu_capacity(rq, env->sd))
8607 * For the load comparisons with the other CPU's, consider
8608 * the weighted_cpuload() scaled with the CPU capacity, so
8609 * that the load can be moved away from the CPU that is
8610 * potentially running at a lower capacity.
8612 * Thus we're looking for max(wl_i / capacity_i), crosswise
8613 * multiplication to rid ourselves of the division works out
8614 * to: wl_i * capacity_j > wl_j * capacity_i; where j is
8615 * our previous maximum.
8617 if (wl * busiest_capacity > busiest_load * capacity) {
8619 busiest_capacity = capacity;
8628 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
8629 * so long as it is large enough.
8631 #define MAX_PINNED_INTERVAL 512
8633 static int need_active_balance(struct lb_env *env)
8635 struct sched_domain *sd = env->sd;
8637 if (env->idle == CPU_NEWLY_IDLE) {
8640 * ASYM_PACKING needs to force migrate tasks from busy but
8641 * lower priority CPUs in order to pack all tasks in the
8642 * highest priority CPUs.
8644 if ((sd->flags & SD_ASYM_PACKING) &&
8645 sched_asym_prefer(env->dst_cpu, env->src_cpu))
8650 * The dst_cpu is idle and the src_cpu CPU has only 1 CFS task.
8651 * It's worth migrating the task if the src_cpu's capacity is reduced
8652 * because of other sched_class or IRQs if more capacity stays
8653 * available on dst_cpu.
8655 if ((env->idle != CPU_NOT_IDLE) &&
8656 (env->src_rq->cfs.h_nr_running == 1)) {
8657 if ((check_cpu_capacity(env->src_rq, sd)) &&
8658 (capacity_of(env->src_cpu)*sd->imbalance_pct < capacity_of(env->dst_cpu)*100))
8662 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
8665 static int active_load_balance_cpu_stop(void *data);
8667 static int should_we_balance(struct lb_env *env)
8669 struct sched_group *sg = env->sd->groups;
8670 int cpu, balance_cpu = -1;
8673 * Ensure the balancing environment is consistent; can happen
8674 * when the softirq triggers 'during' hotplug.
8676 if (!cpumask_test_cpu(env->dst_cpu, env->cpus))
8680 * In the newly idle case, we will allow all the CPUs
8681 * to do the newly idle load balance.
8683 if (env->idle == CPU_NEWLY_IDLE)
8686 /* Try to find first idle CPU */
8687 for_each_cpu_and(cpu, group_balance_mask(sg), env->cpus) {
8695 if (balance_cpu == -1)
8696 balance_cpu = group_balance_cpu(sg);
8699 * First idle CPU or the first CPU(busiest) in this sched group
8700 * is eligible for doing load balancing at this and above domains.
8702 return balance_cpu == env->dst_cpu;
8706 * Check this_cpu to ensure it is balanced within domain. Attempt to move
8707 * tasks if there is an imbalance.
8709 static int load_balance(int this_cpu, struct rq *this_rq,
8710 struct sched_domain *sd, enum cpu_idle_type idle,
8711 int *continue_balancing)
8713 int ld_moved, cur_ld_moved, active_balance = 0;
8714 struct sched_domain *sd_parent = sd->parent;
8715 struct sched_group *group;
8718 struct cpumask *cpus = this_cpu_cpumask_var_ptr(load_balance_mask);
8720 struct lb_env env = {
8722 .dst_cpu = this_cpu,
8724 .dst_grpmask = sched_group_span(sd->groups),
8726 .loop_break = sched_nr_migrate_break,
8729 .tasks = LIST_HEAD_INIT(env.tasks),
8732 cpumask_and(cpus, sched_domain_span(sd), cpu_active_mask);
8734 schedstat_inc(sd->lb_count[idle]);
8737 if (!should_we_balance(&env)) {
8738 *continue_balancing = 0;
8742 group = find_busiest_group(&env);
8744 schedstat_inc(sd->lb_nobusyg[idle]);
8748 busiest = find_busiest_queue(&env, group);
8750 schedstat_inc(sd->lb_nobusyq[idle]);
8754 BUG_ON(busiest == env.dst_rq);
8756 schedstat_add(sd->lb_imbalance[idle], env.imbalance);
8758 env.src_cpu = busiest->cpu;
8759 env.src_rq = busiest;
8762 if (busiest->nr_running > 1) {
8764 * Attempt to move tasks. If find_busiest_group has found
8765 * an imbalance but busiest->nr_running <= 1, the group is
8766 * still unbalanced. ld_moved simply stays zero, so it is
8767 * correctly treated as an imbalance.
8769 env.flags |= LBF_ALL_PINNED;
8770 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
8773 rq_lock_irqsave(busiest, &rf);
8774 update_rq_clock(busiest);
8777 * cur_ld_moved - load moved in current iteration
8778 * ld_moved - cumulative load moved across iterations
8780 cur_ld_moved = detach_tasks(&env);
8783 * We've detached some tasks from busiest_rq. Every
8784 * task is masked "TASK_ON_RQ_MIGRATING", so we can safely
8785 * unlock busiest->lock, and we are able to be sure
8786 * that nobody can manipulate the tasks in parallel.
8787 * See task_rq_lock() family for the details.
8790 rq_unlock(busiest, &rf);
8794 ld_moved += cur_ld_moved;
8797 local_irq_restore(rf.flags);
8799 if (env.flags & LBF_NEED_BREAK) {
8800 env.flags &= ~LBF_NEED_BREAK;
8805 * Revisit (affine) tasks on src_cpu that couldn't be moved to
8806 * us and move them to an alternate dst_cpu in our sched_group
8807 * where they can run. The upper limit on how many times we
8808 * iterate on same src_cpu is dependent on number of CPUs in our
8811 * This changes load balance semantics a bit on who can move
8812 * load to a given_cpu. In addition to the given_cpu itself
8813 * (or a ilb_cpu acting on its behalf where given_cpu is
8814 * nohz-idle), we now have balance_cpu in a position to move
8815 * load to given_cpu. In rare situations, this may cause
8816 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
8817 * _independently_ and at _same_ time to move some load to
8818 * given_cpu) causing exceess load to be moved to given_cpu.
8819 * This however should not happen so much in practice and
8820 * moreover subsequent load balance cycles should correct the
8821 * excess load moved.
8823 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
8825 /* Prevent to re-select dst_cpu via env's CPUs */
8826 cpumask_clear_cpu(env.dst_cpu, env.cpus);
8828 env.dst_rq = cpu_rq(env.new_dst_cpu);
8829 env.dst_cpu = env.new_dst_cpu;
8830 env.flags &= ~LBF_DST_PINNED;
8832 env.loop_break = sched_nr_migrate_break;
8835 * Go back to "more_balance" rather than "redo" since we
8836 * need to continue with same src_cpu.
8842 * We failed to reach balance because of affinity.
8845 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
8847 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0)
8848 *group_imbalance = 1;
8851 /* All tasks on this runqueue were pinned by CPU affinity */
8852 if (unlikely(env.flags & LBF_ALL_PINNED)) {
8853 cpumask_clear_cpu(cpu_of(busiest), cpus);
8855 * Attempting to continue load balancing at the current
8856 * sched_domain level only makes sense if there are
8857 * active CPUs remaining as possible busiest CPUs to
8858 * pull load from which are not contained within the
8859 * destination group that is receiving any migrated
8862 if (!cpumask_subset(cpus, env.dst_grpmask)) {
8864 env.loop_break = sched_nr_migrate_break;
8867 goto out_all_pinned;
8872 schedstat_inc(sd->lb_failed[idle]);
8874 * Increment the failure counter only on periodic balance.
8875 * We do not want newidle balance, which can be very
8876 * frequent, pollute the failure counter causing
8877 * excessive cache_hot migrations and active balances.
8879 if (idle != CPU_NEWLY_IDLE)
8880 sd->nr_balance_failed++;
8882 if (need_active_balance(&env)) {
8883 unsigned long flags;
8885 raw_spin_lock_irqsave(&busiest->lock, flags);
8888 * Don't kick the active_load_balance_cpu_stop,
8889 * if the curr task on busiest CPU can't be
8890 * moved to this_cpu:
8892 if (!cpumask_test_cpu(this_cpu, &busiest->curr->cpus_allowed)) {
8893 raw_spin_unlock_irqrestore(&busiest->lock,
8895 env.flags |= LBF_ALL_PINNED;
8896 goto out_one_pinned;
8900 * ->active_balance synchronizes accesses to
8901 * ->active_balance_work. Once set, it's cleared
8902 * only after active load balance is finished.
8904 if (!busiest->active_balance) {
8905 busiest->active_balance = 1;
8906 busiest->push_cpu = this_cpu;
8909 raw_spin_unlock_irqrestore(&busiest->lock, flags);
8911 if (active_balance) {
8912 stop_one_cpu_nowait(cpu_of(busiest),
8913 active_load_balance_cpu_stop, busiest,
8914 &busiest->active_balance_work);
8917 /* We've kicked active balancing, force task migration. */
8918 sd->nr_balance_failed = sd->cache_nice_tries+1;
8921 sd->nr_balance_failed = 0;
8923 if (likely(!active_balance)) {
8924 /* We were unbalanced, so reset the balancing interval */
8925 sd->balance_interval = sd->min_interval;
8928 * If we've begun active balancing, start to back off. This
8929 * case may not be covered by the all_pinned logic if there
8930 * is only 1 task on the busy runqueue (because we don't call
8933 if (sd->balance_interval < sd->max_interval)
8934 sd->balance_interval *= 2;
8941 * We reach balance although we may have faced some affinity
8942 * constraints. Clear the imbalance flag only if other tasks got
8943 * a chance to move and fix the imbalance.
8945 if (sd_parent && !(env.flags & LBF_ALL_PINNED)) {
8946 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
8948 if (*group_imbalance)
8949 *group_imbalance = 0;
8954 * We reach balance because all tasks are pinned at this level so
8955 * we can't migrate them. Let the imbalance flag set so parent level
8956 * can try to migrate them.
8958 schedstat_inc(sd->lb_balanced[idle]);
8960 sd->nr_balance_failed = 0;
8966 * idle_balance() disregards balance intervals, so we could repeatedly
8967 * reach this code, which would lead to balance_interval skyrocketting
8968 * in a short amount of time. Skip the balance_interval increase logic
8971 if (env.idle == CPU_NEWLY_IDLE)
8974 /* tune up the balancing interval */
8975 if (((env.flags & LBF_ALL_PINNED) &&
8976 sd->balance_interval < MAX_PINNED_INTERVAL) ||
8977 (sd->balance_interval < sd->max_interval))
8978 sd->balance_interval *= 2;
8983 static inline unsigned long
8984 get_sd_balance_interval(struct sched_domain *sd, int cpu_busy)
8986 unsigned long interval = sd->balance_interval;
8989 interval *= sd->busy_factor;
8991 /* scale ms to jiffies */
8992 interval = msecs_to_jiffies(interval);
8993 interval = clamp(interval, 1UL, max_load_balance_interval);
8999 update_next_balance(struct sched_domain *sd, unsigned long *next_balance)
9001 unsigned long interval, next;
9003 /* used by idle balance, so cpu_busy = 0 */
9004 interval = get_sd_balance_interval(sd, 0);
9005 next = sd->last_balance + interval;
9007 if (time_after(*next_balance, next))
9008 *next_balance = next;
9012 * active_load_balance_cpu_stop is run by the CPU stopper. It pushes
9013 * running tasks off the busiest CPU onto idle CPUs. It requires at
9014 * least 1 task to be running on each physical CPU where possible, and
9015 * avoids physical / logical imbalances.
9017 static int active_load_balance_cpu_stop(void *data)
9019 struct rq *busiest_rq = data;
9020 int busiest_cpu = cpu_of(busiest_rq);
9021 int target_cpu = busiest_rq->push_cpu;
9022 struct rq *target_rq = cpu_rq(target_cpu);
9023 struct sched_domain *sd;
9024 struct task_struct *p = NULL;
9027 rq_lock_irq(busiest_rq, &rf);
9029 * Between queueing the stop-work and running it is a hole in which
9030 * CPUs can become inactive. We should not move tasks from or to
9033 if (!cpu_active(busiest_cpu) || !cpu_active(target_cpu))
9036 /* Make sure the requested CPU hasn't gone down in the meantime: */
9037 if (unlikely(busiest_cpu != smp_processor_id() ||
9038 !busiest_rq->active_balance))
9041 /* Is there any task to move? */
9042 if (busiest_rq->nr_running <= 1)
9046 * This condition is "impossible", if it occurs
9047 * we need to fix it. Originally reported by
9048 * Bjorn Helgaas on a 128-CPU setup.
9050 BUG_ON(busiest_rq == target_rq);
9052 /* Search for an sd spanning us and the target CPU. */
9054 for_each_domain(target_cpu, sd) {
9055 if ((sd->flags & SD_LOAD_BALANCE) &&
9056 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
9061 struct lb_env env = {
9063 .dst_cpu = target_cpu,
9064 .dst_rq = target_rq,
9065 .src_cpu = busiest_rq->cpu,
9066 .src_rq = busiest_rq,
9069 * can_migrate_task() doesn't need to compute new_dst_cpu
9070 * for active balancing. Since we have CPU_IDLE, but no
9071 * @dst_grpmask we need to make that test go away with lying
9074 .flags = LBF_DST_PINNED,
9077 schedstat_inc(sd->alb_count);
9078 update_rq_clock(busiest_rq);
9080 p = detach_one_task(&env);
9082 schedstat_inc(sd->alb_pushed);
9083 /* Active balancing done, reset the failure counter. */
9084 sd->nr_balance_failed = 0;
9086 schedstat_inc(sd->alb_failed);
9091 busiest_rq->active_balance = 0;
9092 rq_unlock(busiest_rq, &rf);
9095 attach_one_task(target_rq, p);
9102 static DEFINE_SPINLOCK(balancing);
9105 * Scale the max load_balance interval with the number of CPUs in the system.
9106 * This trades load-balance latency on larger machines for less cross talk.
9108 void update_max_interval(void)
9110 max_load_balance_interval = HZ*num_online_cpus()/10;
9114 * It checks each scheduling domain to see if it is due to be balanced,
9115 * and initiates a balancing operation if so.
9117 * Balancing parameters are set up in init_sched_domains.
9119 static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle)
9121 int continue_balancing = 1;
9123 unsigned long interval;
9124 struct sched_domain *sd;
9125 /* Earliest time when we have to do rebalance again */
9126 unsigned long next_balance = jiffies + 60*HZ;
9127 int update_next_balance = 0;
9128 int need_serialize, need_decay = 0;
9132 for_each_domain(cpu, sd) {
9134 * Decay the newidle max times here because this is a regular
9135 * visit to all the domains. Decay ~1% per second.
9137 if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
9138 sd->max_newidle_lb_cost =
9139 (sd->max_newidle_lb_cost * 253) / 256;
9140 sd->next_decay_max_lb_cost = jiffies + HZ;
9143 max_cost += sd->max_newidle_lb_cost;
9145 if (!(sd->flags & SD_LOAD_BALANCE))
9149 * Stop the load balance at this level. There is another
9150 * CPU in our sched group which is doing load balancing more
9153 if (!continue_balancing) {
9159 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
9161 need_serialize = sd->flags & SD_SERIALIZE;
9162 if (need_serialize) {
9163 if (!spin_trylock(&balancing))
9167 if (time_after_eq(jiffies, sd->last_balance + interval)) {
9168 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
9170 * The LBF_DST_PINNED logic could have changed
9171 * env->dst_cpu, so we can't know our idle
9172 * state even if we migrated tasks. Update it.
9174 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
9176 sd->last_balance = jiffies;
9177 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
9180 spin_unlock(&balancing);
9182 if (time_after(next_balance, sd->last_balance + interval)) {
9183 next_balance = sd->last_balance + interval;
9184 update_next_balance = 1;
9189 * Ensure the rq-wide value also decays but keep it at a
9190 * reasonable floor to avoid funnies with rq->avg_idle.
9192 rq->max_idle_balance_cost =
9193 max((u64)sysctl_sched_migration_cost, max_cost);
9198 * next_balance will be updated only when there is a need.
9199 * When the cpu is attached to null domain for ex, it will not be
9202 if (likely(update_next_balance)) {
9203 rq->next_balance = next_balance;
9205 #ifdef CONFIG_NO_HZ_COMMON
9207 * If this CPU has been elected to perform the nohz idle
9208 * balance. Other idle CPUs have already rebalanced with
9209 * nohz_idle_balance() and nohz.next_balance has been
9210 * updated accordingly. This CPU is now running the idle load
9211 * balance for itself and we need to update the
9212 * nohz.next_balance accordingly.
9214 if ((idle == CPU_IDLE) && time_after(nohz.next_balance, rq->next_balance))
9215 nohz.next_balance = rq->next_balance;
9220 static inline int on_null_domain(struct rq *rq)
9222 return unlikely(!rcu_dereference_sched(rq->sd));
9225 #ifdef CONFIG_NO_HZ_COMMON
9227 * idle load balancing details
9228 * - When one of the busy CPUs notice that there may be an idle rebalancing
9229 * needed, they will kick the idle load balancer, which then does idle
9230 * load balancing for all the idle CPUs.
9231 * - HK_FLAG_MISC CPUs are used for this task, because HK_FLAG_SCHED not set
9235 static inline int find_new_ilb(void)
9239 for_each_cpu_and(ilb, nohz.idle_cpus_mask,
9240 housekeeping_cpumask(HK_FLAG_MISC)) {
9249 * Kick a CPU to do the nohz balancing, if it is time for it. We pick any
9250 * idle CPU in the HK_FLAG_MISC housekeeping set (if there is one).
9252 static void kick_ilb(unsigned int flags)
9257 * Increase nohz.next_balance only when if full ilb is triggered but
9258 * not if we only update stats.
9260 if (flags & NOHZ_BALANCE_KICK)
9261 nohz.next_balance = jiffies+1;
9263 ilb_cpu = find_new_ilb();
9265 if (ilb_cpu >= nr_cpu_ids)
9268 flags = atomic_fetch_or(flags, nohz_flags(ilb_cpu));
9269 if (flags & NOHZ_KICK_MASK)
9273 * Use smp_send_reschedule() instead of resched_cpu().
9274 * This way we generate a sched IPI on the target CPU which
9275 * is idle. And the softirq performing nohz idle load balance
9276 * will be run before returning from the IPI.
9278 smp_send_reschedule(ilb_cpu);
9282 * Current heuristic for kicking the idle load balancer in the presence
9283 * of an idle cpu in the system.
9284 * - This rq has more than one task.
9285 * - This rq has at least one CFS task and the capacity of the CPU is
9286 * significantly reduced because of RT tasks or IRQs.
9287 * - At parent of LLC scheduler domain level, this cpu's scheduler group has
9288 * multiple busy cpu.
9289 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
9290 * domain span are idle.
9292 static void nohz_balancer_kick(struct rq *rq)
9294 unsigned long now = jiffies;
9295 struct sched_domain_shared *sds;
9296 struct sched_domain *sd;
9297 int nr_busy, i, cpu = rq->cpu;
9298 unsigned int flags = 0;
9300 if (unlikely(rq->idle_balance))
9304 * We may be recently in ticked or tickless idle mode. At the first
9305 * busy tick after returning from idle, we will update the busy stats.
9307 nohz_balance_exit_idle(rq);
9310 * None are in tickless mode and hence no need for NOHZ idle load
9313 if (likely(!atomic_read(&nohz.nr_cpus)))
9316 if (READ_ONCE(nohz.has_blocked) &&
9317 time_after(now, READ_ONCE(nohz.next_blocked)))
9318 flags = NOHZ_STATS_KICK;
9320 if (time_before(now, nohz.next_balance))
9323 if (rq->nr_running >= 2) {
9324 flags = NOHZ_KICK_MASK;
9329 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
9332 * XXX: write a coherent comment on why we do this.
9333 * See also: http://lkml.kernel.org/r/20111202010832.602203411@sbsiddha-desk.sc.intel.com
9335 nr_busy = atomic_read(&sds->nr_busy_cpus);
9337 flags = NOHZ_KICK_MASK;
9343 sd = rcu_dereference(rq->sd);
9345 if ((rq->cfs.h_nr_running >= 1) &&
9346 check_cpu_capacity(rq, sd)) {
9347 flags = NOHZ_KICK_MASK;
9352 sd = rcu_dereference(per_cpu(sd_asym, cpu));
9354 for_each_cpu(i, sched_domain_span(sd)) {
9356 !cpumask_test_cpu(i, nohz.idle_cpus_mask))
9359 if (sched_asym_prefer(i, cpu)) {
9360 flags = NOHZ_KICK_MASK;
9372 static void set_cpu_sd_state_busy(int cpu)
9374 struct sched_domain *sd;
9377 sd = rcu_dereference(per_cpu(sd_llc, cpu));
9379 if (!sd || !sd->nohz_idle)
9383 atomic_inc(&sd->shared->nr_busy_cpus);
9388 void nohz_balance_exit_idle(struct rq *rq)
9390 SCHED_WARN_ON(rq != this_rq());
9392 if (likely(!rq->nohz_tick_stopped))
9395 rq->nohz_tick_stopped = 0;
9396 cpumask_clear_cpu(rq->cpu, nohz.idle_cpus_mask);
9397 atomic_dec(&nohz.nr_cpus);
9399 set_cpu_sd_state_busy(rq->cpu);
9402 static void set_cpu_sd_state_idle(int cpu)
9404 struct sched_domain *sd;
9407 sd = rcu_dereference(per_cpu(sd_llc, cpu));
9409 if (!sd || sd->nohz_idle)
9413 atomic_dec(&sd->shared->nr_busy_cpus);
9419 * This routine will record that the CPU is going idle with tick stopped.
9420 * This info will be used in performing idle load balancing in the future.
9422 void nohz_balance_enter_idle(int cpu)
9424 struct rq *rq = cpu_rq(cpu);
9426 SCHED_WARN_ON(cpu != smp_processor_id());
9428 /* If this CPU is going down, then nothing needs to be done: */
9429 if (!cpu_active(cpu))
9432 /* Spare idle load balancing on CPUs that don't want to be disturbed: */
9433 if (!housekeeping_cpu(cpu, HK_FLAG_SCHED))
9437 * Can be set safely without rq->lock held
9438 * If a clear happens, it will have evaluated last additions because
9439 * rq->lock is held during the check and the clear
9441 rq->has_blocked_load = 1;
9444 * The tick is still stopped but load could have been added in the
9445 * meantime. We set the nohz.has_blocked flag to trig a check of the
9446 * *_avg. The CPU is already part of nohz.idle_cpus_mask so the clear
9447 * of nohz.has_blocked can only happen after checking the new load
9449 if (rq->nohz_tick_stopped)
9452 /* If we're a completely isolated CPU, we don't play: */
9453 if (on_null_domain(rq))
9456 rq->nohz_tick_stopped = 1;
9458 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
9459 atomic_inc(&nohz.nr_cpus);
9462 * Ensures that if nohz_idle_balance() fails to observe our
9463 * @idle_cpus_mask store, it must observe the @has_blocked
9466 smp_mb__after_atomic();
9468 set_cpu_sd_state_idle(cpu);
9472 * Each time a cpu enter idle, we assume that it has blocked load and
9473 * enable the periodic update of the load of idle cpus
9475 WRITE_ONCE(nohz.has_blocked, 1);
9479 * Internal function that runs load balance for all idle cpus. The load balance
9480 * can be a simple update of blocked load or a complete load balance with
9481 * tasks movement depending of flags.
9482 * The function returns false if the loop has stopped before running
9483 * through all idle CPUs.
9485 static bool _nohz_idle_balance(struct rq *this_rq, unsigned int flags,
9486 enum cpu_idle_type idle)
9488 /* Earliest time when we have to do rebalance again */
9489 unsigned long now = jiffies;
9490 unsigned long next_balance = now + 60*HZ;
9491 bool has_blocked_load = false;
9492 int update_next_balance = 0;
9493 int this_cpu = this_rq->cpu;
9498 SCHED_WARN_ON((flags & NOHZ_KICK_MASK) == NOHZ_BALANCE_KICK);
9501 * We assume there will be no idle load after this update and clear
9502 * the has_blocked flag. If a cpu enters idle in the mean time, it will
9503 * set the has_blocked flag and trig another update of idle load.
9504 * Because a cpu that becomes idle, is added to idle_cpus_mask before
9505 * setting the flag, we are sure to not clear the state and not
9506 * check the load of an idle cpu.
9508 WRITE_ONCE(nohz.has_blocked, 0);
9511 * Ensures that if we miss the CPU, we must see the has_blocked
9512 * store from nohz_balance_enter_idle().
9516 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
9517 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
9521 * If this CPU gets work to do, stop the load balancing
9522 * work being done for other CPUs. Next load
9523 * balancing owner will pick it up.
9525 if (need_resched()) {
9526 has_blocked_load = true;
9530 rq = cpu_rq(balance_cpu);
9532 has_blocked_load |= update_nohz_stats(rq, true);
9535 * If time for next balance is due,
9538 if (time_after_eq(jiffies, rq->next_balance)) {
9541 rq_lock_irqsave(rq, &rf);
9542 update_rq_clock(rq);
9543 cpu_load_update_idle(rq);
9544 rq_unlock_irqrestore(rq, &rf);
9546 if (flags & NOHZ_BALANCE_KICK)
9547 rebalance_domains(rq, CPU_IDLE);
9550 if (time_after(next_balance, rq->next_balance)) {
9551 next_balance = rq->next_balance;
9552 update_next_balance = 1;
9557 * next_balance will be updated only when there is a need.
9558 * When the CPU is attached to null domain for ex, it will not be
9561 if (likely(update_next_balance))
9562 nohz.next_balance = next_balance;
9564 /* Newly idle CPU doesn't need an update */
9565 if (idle != CPU_NEWLY_IDLE) {
9566 update_blocked_averages(this_cpu);
9567 has_blocked_load |= this_rq->has_blocked_load;
9570 if (flags & NOHZ_BALANCE_KICK)
9571 rebalance_domains(this_rq, CPU_IDLE);
9573 WRITE_ONCE(nohz.next_blocked,
9574 now + msecs_to_jiffies(LOAD_AVG_PERIOD));
9576 /* The full idle balance loop has been done */
9580 /* There is still blocked load, enable periodic update */
9581 if (has_blocked_load)
9582 WRITE_ONCE(nohz.has_blocked, 1);
9588 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
9589 * rebalancing for all the cpus for whom scheduler ticks are stopped.
9591 static bool nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
9593 int this_cpu = this_rq->cpu;
9596 if (!(atomic_read(nohz_flags(this_cpu)) & NOHZ_KICK_MASK))
9599 if (idle != CPU_IDLE) {
9600 atomic_andnot(NOHZ_KICK_MASK, nohz_flags(this_cpu));
9605 * barrier, pairs with nohz_balance_enter_idle(), ensures ...
9607 flags = atomic_fetch_andnot(NOHZ_KICK_MASK, nohz_flags(this_cpu));
9608 if (!(flags & NOHZ_KICK_MASK))
9611 _nohz_idle_balance(this_rq, flags, idle);
9616 static void nohz_newidle_balance(struct rq *this_rq)
9618 int this_cpu = this_rq->cpu;
9621 * This CPU doesn't want to be disturbed by scheduler
9624 if (!housekeeping_cpu(this_cpu, HK_FLAG_SCHED))
9627 /* Will wake up very soon. No time for doing anything else*/
9628 if (this_rq->avg_idle < sysctl_sched_migration_cost)
9631 /* Don't need to update blocked load of idle CPUs*/
9632 if (!READ_ONCE(nohz.has_blocked) ||
9633 time_before(jiffies, READ_ONCE(nohz.next_blocked)))
9636 raw_spin_unlock(&this_rq->lock);
9638 * This CPU is going to be idle and blocked load of idle CPUs
9639 * need to be updated. Run the ilb locally as it is a good
9640 * candidate for ilb instead of waking up another idle CPU.
9641 * Kick an normal ilb if we failed to do the update.
9643 if (!_nohz_idle_balance(this_rq, NOHZ_STATS_KICK, CPU_NEWLY_IDLE))
9644 kick_ilb(NOHZ_STATS_KICK);
9645 raw_spin_lock(&this_rq->lock);
9648 #else /* !CONFIG_NO_HZ_COMMON */
9649 static inline void nohz_balancer_kick(struct rq *rq) { }
9651 static inline bool nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
9656 static inline void nohz_newidle_balance(struct rq *this_rq) { }
9657 #endif /* CONFIG_NO_HZ_COMMON */
9660 * idle_balance is called by schedule() if this_cpu is about to become
9661 * idle. Attempts to pull tasks from other CPUs.
9663 static int idle_balance(struct rq *this_rq, struct rq_flags *rf)
9665 unsigned long next_balance = jiffies + HZ;
9666 int this_cpu = this_rq->cpu;
9667 struct sched_domain *sd;
9668 int pulled_task = 0;
9672 * We must set idle_stamp _before_ calling idle_balance(), such that we
9673 * measure the duration of idle_balance() as idle time.
9675 this_rq->idle_stamp = rq_clock(this_rq);
9678 * Do not pull tasks towards !active CPUs...
9680 if (!cpu_active(this_cpu))
9684 * This is OK, because current is on_cpu, which avoids it being picked
9685 * for load-balance and preemption/IRQs are still disabled avoiding
9686 * further scheduler activity on it and we're being very careful to
9687 * re-start the picking loop.
9689 rq_unpin_lock(this_rq, rf);
9691 if (this_rq->avg_idle < sysctl_sched_migration_cost ||
9692 !this_rq->rd->overload) {
9695 sd = rcu_dereference_check_sched_domain(this_rq->sd);
9697 update_next_balance(sd, &next_balance);
9700 nohz_newidle_balance(this_rq);
9705 raw_spin_unlock(&this_rq->lock);
9707 update_blocked_averages(this_cpu);
9709 for_each_domain(this_cpu, sd) {
9710 int continue_balancing = 1;
9711 u64 t0, domain_cost;
9713 if (!(sd->flags & SD_LOAD_BALANCE))
9716 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost) {
9717 update_next_balance(sd, &next_balance);
9721 if (sd->flags & SD_BALANCE_NEWIDLE) {
9722 t0 = sched_clock_cpu(this_cpu);
9724 pulled_task = load_balance(this_cpu, this_rq,
9726 &continue_balancing);
9728 domain_cost = sched_clock_cpu(this_cpu) - t0;
9729 if (domain_cost > sd->max_newidle_lb_cost)
9730 sd->max_newidle_lb_cost = domain_cost;
9732 curr_cost += domain_cost;
9735 update_next_balance(sd, &next_balance);
9738 * Stop searching for tasks to pull if there are
9739 * now runnable tasks on this rq.
9741 if (pulled_task || this_rq->nr_running > 0)
9746 raw_spin_lock(&this_rq->lock);
9748 if (curr_cost > this_rq->max_idle_balance_cost)
9749 this_rq->max_idle_balance_cost = curr_cost;
9753 * While browsing the domains, we released the rq lock, a task could
9754 * have been enqueued in the meantime. Since we're not going idle,
9755 * pretend we pulled a task.
9757 if (this_rq->cfs.h_nr_running && !pulled_task)
9760 /* Move the next balance forward */
9761 if (time_after(this_rq->next_balance, next_balance))
9762 this_rq->next_balance = next_balance;
9764 /* Is there a task of a high priority class? */
9765 if (this_rq->nr_running != this_rq->cfs.h_nr_running)
9769 this_rq->idle_stamp = 0;
9771 rq_repin_lock(this_rq, rf);
9777 * run_rebalance_domains is triggered when needed from the scheduler tick.
9778 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
9780 static __latent_entropy void run_rebalance_domains(struct softirq_action *h)
9782 struct rq *this_rq = this_rq();
9783 enum cpu_idle_type idle = this_rq->idle_balance ?
9784 CPU_IDLE : CPU_NOT_IDLE;
9787 * If this CPU has a pending nohz_balance_kick, then do the
9788 * balancing on behalf of the other idle CPUs whose ticks are
9789 * stopped. Do nohz_idle_balance *before* rebalance_domains to
9790 * give the idle CPUs a chance to load balance. Else we may
9791 * load balance only within the local sched_domain hierarchy
9792 * and abort nohz_idle_balance altogether if we pull some load.
9794 if (nohz_idle_balance(this_rq, idle))
9797 /* normal load balance */
9798 update_blocked_averages(this_rq->cpu);
9799 rebalance_domains(this_rq, idle);
9803 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
9805 void trigger_load_balance(struct rq *rq)
9807 /* Don't need to rebalance while attached to NULL domain */
9808 if (unlikely(on_null_domain(rq)))
9811 if (time_after_eq(jiffies, rq->next_balance))
9812 raise_softirq(SCHED_SOFTIRQ);
9814 nohz_balancer_kick(rq);
9817 static void rq_online_fair(struct rq *rq)
9821 update_runtime_enabled(rq);
9824 static void rq_offline_fair(struct rq *rq)
9828 /* Ensure any throttled groups are reachable by pick_next_task */
9829 unthrottle_offline_cfs_rqs(rq);
9832 #endif /* CONFIG_SMP */
9835 * scheduler tick hitting a task of our scheduling class.
9837 * NOTE: This function can be called remotely by the tick offload that
9838 * goes along full dynticks. Therefore no local assumption can be made
9839 * and everything must be accessed through the @rq and @curr passed in
9842 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
9844 struct cfs_rq *cfs_rq;
9845 struct sched_entity *se = &curr->se;
9847 for_each_sched_entity(se) {
9848 cfs_rq = cfs_rq_of(se);
9849 entity_tick(cfs_rq, se, queued);
9852 if (static_branch_unlikely(&sched_numa_balancing))
9853 task_tick_numa(rq, curr);
9857 * called on fork with the child task as argument from the parent's context
9858 * - child not yet on the tasklist
9859 * - preemption disabled
9861 static void task_fork_fair(struct task_struct *p)
9863 struct cfs_rq *cfs_rq;
9864 struct sched_entity *se = &p->se, *curr;
9865 struct rq *rq = this_rq();
9869 update_rq_clock(rq);
9871 cfs_rq = task_cfs_rq(current);
9872 curr = cfs_rq->curr;
9874 update_curr(cfs_rq);
9875 se->vruntime = curr->vruntime;
9877 place_entity(cfs_rq, se, 1);
9879 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
9881 * Upon rescheduling, sched_class::put_prev_task() will place
9882 * 'current' within the tree based on its new key value.
9884 swap(curr->vruntime, se->vruntime);
9888 se->vruntime -= cfs_rq->min_vruntime;
9893 * Priority of the task has changed. Check to see if we preempt
9897 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
9899 if (!task_on_rq_queued(p))
9903 * Reschedule if we are currently running on this runqueue and
9904 * our priority decreased, or if we are not currently running on
9905 * this runqueue and our priority is higher than the current's
9907 if (rq->curr == p) {
9908 if (p->prio > oldprio)
9911 check_preempt_curr(rq, p, 0);
9914 static inline bool vruntime_normalized(struct task_struct *p)
9916 struct sched_entity *se = &p->se;
9919 * In both the TASK_ON_RQ_QUEUED and TASK_ON_RQ_MIGRATING cases,
9920 * the dequeue_entity(.flags=0) will already have normalized the
9927 * When !on_rq, vruntime of the task has usually NOT been normalized.
9928 * But there are some cases where it has already been normalized:
9930 * - A forked child which is waiting for being woken up by
9931 * wake_up_new_task().
9932 * - A task which has been woken up by try_to_wake_up() and
9933 * waiting for actually being woken up by sched_ttwu_pending().
9935 if (!se->sum_exec_runtime ||
9936 (p->state == TASK_WAKING && p->sched_remote_wakeup))
9942 #ifdef CONFIG_FAIR_GROUP_SCHED
9944 * Propagate the changes of the sched_entity across the tg tree to make it
9945 * visible to the root
9947 static void propagate_entity_cfs_rq(struct sched_entity *se)
9949 struct cfs_rq *cfs_rq;
9951 list_add_leaf_cfs_rq(cfs_rq_of(se));
9953 /* Start to propagate at parent */
9956 for_each_sched_entity(se) {
9957 cfs_rq = cfs_rq_of(se);
9959 if (!cfs_rq_throttled(cfs_rq)){
9960 update_load_avg(cfs_rq, se, UPDATE_TG);
9961 list_add_leaf_cfs_rq(cfs_rq);
9965 if (list_add_leaf_cfs_rq(cfs_rq))
9970 static void propagate_entity_cfs_rq(struct sched_entity *se) { }
9973 static void detach_entity_cfs_rq(struct sched_entity *se)
9975 struct cfs_rq *cfs_rq = cfs_rq_of(se);
9977 /* Catch up with the cfs_rq and remove our load when we leave */
9978 update_load_avg(cfs_rq, se, 0);
9979 detach_entity_load_avg(cfs_rq, se);
9980 update_tg_load_avg(cfs_rq, false);
9981 propagate_entity_cfs_rq(se);
9984 static void attach_entity_cfs_rq(struct sched_entity *se)
9986 struct cfs_rq *cfs_rq = cfs_rq_of(se);
9988 #ifdef CONFIG_FAIR_GROUP_SCHED
9990 * Since the real-depth could have been changed (only FAIR
9991 * class maintain depth value), reset depth properly.
9993 se->depth = se->parent ? se->parent->depth + 1 : 0;
9996 /* Synchronize entity with its cfs_rq */
9997 update_load_avg(cfs_rq, se, sched_feat(ATTACH_AGE_LOAD) ? 0 : SKIP_AGE_LOAD);
9998 attach_entity_load_avg(cfs_rq, se, 0);
9999 update_tg_load_avg(cfs_rq, false);
10000 propagate_entity_cfs_rq(se);
10003 static void detach_task_cfs_rq(struct task_struct *p)
10005 struct sched_entity *se = &p->se;
10006 struct cfs_rq *cfs_rq = cfs_rq_of(se);
10008 if (!vruntime_normalized(p)) {
10010 * Fix up our vruntime so that the current sleep doesn't
10011 * cause 'unlimited' sleep bonus.
10013 place_entity(cfs_rq, se, 0);
10014 se->vruntime -= cfs_rq->min_vruntime;
10017 detach_entity_cfs_rq(se);
10020 static void attach_task_cfs_rq(struct task_struct *p)
10022 struct sched_entity *se = &p->se;
10023 struct cfs_rq *cfs_rq = cfs_rq_of(se);
10025 attach_entity_cfs_rq(se);
10027 if (!vruntime_normalized(p))
10028 se->vruntime += cfs_rq->min_vruntime;
10031 static void switched_from_fair(struct rq *rq, struct task_struct *p)
10033 detach_task_cfs_rq(p);
10036 static void switched_to_fair(struct rq *rq, struct task_struct *p)
10038 attach_task_cfs_rq(p);
10040 if (task_on_rq_queued(p)) {
10042 * We were most likely switched from sched_rt, so
10043 * kick off the schedule if running, otherwise just see
10044 * if we can still preempt the current task.
10049 check_preempt_curr(rq, p, 0);
10053 /* Account for a task changing its policy or group.
10055 * This routine is mostly called to set cfs_rq->curr field when a task
10056 * migrates between groups/classes.
10058 static void set_curr_task_fair(struct rq *rq)
10060 struct sched_entity *se = &rq->curr->se;
10062 for_each_sched_entity(se) {
10063 struct cfs_rq *cfs_rq = cfs_rq_of(se);
10065 set_next_entity(cfs_rq, se);
10066 /* ensure bandwidth has been allocated on our new cfs_rq */
10067 account_cfs_rq_runtime(cfs_rq, 0);
10071 void init_cfs_rq(struct cfs_rq *cfs_rq)
10073 cfs_rq->tasks_timeline = RB_ROOT_CACHED;
10074 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
10075 #ifndef CONFIG_64BIT
10076 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
10079 raw_spin_lock_init(&cfs_rq->removed.lock);
10083 #ifdef CONFIG_FAIR_GROUP_SCHED
10084 static void task_set_group_fair(struct task_struct *p)
10086 struct sched_entity *se = &p->se;
10088 set_task_rq(p, task_cpu(p));
10089 se->depth = se->parent ? se->parent->depth + 1 : 0;
10092 static void task_move_group_fair(struct task_struct *p)
10094 detach_task_cfs_rq(p);
10095 set_task_rq(p, task_cpu(p));
10098 /* Tell se's cfs_rq has been changed -- migrated */
10099 p->se.avg.last_update_time = 0;
10101 attach_task_cfs_rq(p);
10104 static void task_change_group_fair(struct task_struct *p, int type)
10107 case TASK_SET_GROUP:
10108 task_set_group_fair(p);
10111 case TASK_MOVE_GROUP:
10112 task_move_group_fair(p);
10117 void free_fair_sched_group(struct task_group *tg)
10121 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
10123 for_each_possible_cpu(i) {
10125 kfree(tg->cfs_rq[i]);
10134 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
10136 struct sched_entity *se;
10137 struct cfs_rq *cfs_rq;
10140 tg->cfs_rq = kcalloc(nr_cpu_ids, sizeof(cfs_rq), GFP_KERNEL);
10143 tg->se = kcalloc(nr_cpu_ids, sizeof(se), GFP_KERNEL);
10147 tg->shares = NICE_0_LOAD;
10149 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
10151 for_each_possible_cpu(i) {
10152 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
10153 GFP_KERNEL, cpu_to_node(i));
10157 se = kzalloc_node(sizeof(struct sched_entity),
10158 GFP_KERNEL, cpu_to_node(i));
10162 init_cfs_rq(cfs_rq);
10163 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
10164 init_entity_runnable_average(se);
10175 void online_fair_sched_group(struct task_group *tg)
10177 struct sched_entity *se;
10178 struct rq_flags rf;
10182 for_each_possible_cpu(i) {
10185 rq_lock_irq(rq, &rf);
10186 update_rq_clock(rq);
10187 attach_entity_cfs_rq(se);
10188 sync_throttle(tg, i);
10189 rq_unlock_irq(rq, &rf);
10193 void unregister_fair_sched_group(struct task_group *tg)
10195 unsigned long flags;
10199 for_each_possible_cpu(cpu) {
10201 remove_entity_load_avg(tg->se[cpu]);
10204 * Only empty task groups can be destroyed; so we can speculatively
10205 * check on_list without danger of it being re-added.
10207 if (!tg->cfs_rq[cpu]->on_list)
10212 raw_spin_lock_irqsave(&rq->lock, flags);
10213 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
10214 raw_spin_unlock_irqrestore(&rq->lock, flags);
10218 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
10219 struct sched_entity *se, int cpu,
10220 struct sched_entity *parent)
10222 struct rq *rq = cpu_rq(cpu);
10226 init_cfs_rq_runtime(cfs_rq);
10228 tg->cfs_rq[cpu] = cfs_rq;
10231 /* se could be NULL for root_task_group */
10236 se->cfs_rq = &rq->cfs;
10239 se->cfs_rq = parent->my_q;
10240 se->depth = parent->depth + 1;
10244 /* guarantee group entities always have weight */
10245 update_load_set(&se->load, NICE_0_LOAD);
10246 se->parent = parent;
10249 static DEFINE_MUTEX(shares_mutex);
10251 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
10256 * We can't change the weight of the root cgroup.
10261 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
10263 mutex_lock(&shares_mutex);
10264 if (tg->shares == shares)
10267 tg->shares = shares;
10268 for_each_possible_cpu(i) {
10269 struct rq *rq = cpu_rq(i);
10270 struct sched_entity *se = tg->se[i];
10271 struct rq_flags rf;
10273 /* Propagate contribution to hierarchy */
10274 rq_lock_irqsave(rq, &rf);
10275 update_rq_clock(rq);
10276 for_each_sched_entity(se) {
10277 update_load_avg(cfs_rq_of(se), se, UPDATE_TG);
10278 update_cfs_group(se);
10280 rq_unlock_irqrestore(rq, &rf);
10284 mutex_unlock(&shares_mutex);
10287 #else /* CONFIG_FAIR_GROUP_SCHED */
10289 void free_fair_sched_group(struct task_group *tg) { }
10291 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
10296 void online_fair_sched_group(struct task_group *tg) { }
10298 void unregister_fair_sched_group(struct task_group *tg) { }
10300 #endif /* CONFIG_FAIR_GROUP_SCHED */
10303 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
10305 struct sched_entity *se = &task->se;
10306 unsigned int rr_interval = 0;
10309 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
10312 if (rq->cfs.load.weight)
10313 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
10315 return rr_interval;
10319 * All the scheduling class methods:
10321 const struct sched_class fair_sched_class = {
10322 .next = &idle_sched_class,
10323 .enqueue_task = enqueue_task_fair,
10324 .dequeue_task = dequeue_task_fair,
10325 .yield_task = yield_task_fair,
10326 .yield_to_task = yield_to_task_fair,
10328 .check_preempt_curr = check_preempt_wakeup,
10330 .pick_next_task = pick_next_task_fair,
10331 .put_prev_task = put_prev_task_fair,
10334 .select_task_rq = select_task_rq_fair,
10335 .migrate_task_rq = migrate_task_rq_fair,
10337 .rq_online = rq_online_fair,
10338 .rq_offline = rq_offline_fair,
10340 .task_dead = task_dead_fair,
10341 .set_cpus_allowed = set_cpus_allowed_common,
10344 .set_curr_task = set_curr_task_fair,
10345 .task_tick = task_tick_fair,
10346 .task_fork = task_fork_fair,
10348 .prio_changed = prio_changed_fair,
10349 .switched_from = switched_from_fair,
10350 .switched_to = switched_to_fair,
10352 .get_rr_interval = get_rr_interval_fair,
10354 .update_curr = update_curr_fair,
10356 #ifdef CONFIG_FAIR_GROUP_SCHED
10357 .task_change_group = task_change_group_fair,
10361 #ifdef CONFIG_SCHED_DEBUG
10362 void print_cfs_stats(struct seq_file *m, int cpu)
10364 struct cfs_rq *cfs_rq, *pos;
10367 for_each_leaf_cfs_rq_safe(cpu_rq(cpu), cfs_rq, pos)
10368 print_cfs_rq(m, cpu, cfs_rq);
10372 #ifdef CONFIG_NUMA_BALANCING
10373 void show_numa_stats(struct task_struct *p, struct seq_file *m)
10376 unsigned long tsf = 0, tpf = 0, gsf = 0, gpf = 0;
10377 struct numa_group *ng;
10380 ng = rcu_dereference(p->numa_group);
10381 for_each_online_node(node) {
10382 if (p->numa_faults) {
10383 tsf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 0)];
10384 tpf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 1)];
10387 gsf = ng->faults[task_faults_idx(NUMA_MEM, node, 0)],
10388 gpf = ng->faults[task_faults_idx(NUMA_MEM, node, 1)];
10390 print_numa_stats(m, node, tsf, tpf, gsf, gpf);
10394 #endif /* CONFIG_NUMA_BALANCING */
10395 #endif /* CONFIG_SCHED_DEBUG */
10397 __init void init_sched_fair_class(void)
10400 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
10402 #ifdef CONFIG_NO_HZ_COMMON
10403 nohz.next_balance = jiffies;
10404 nohz.next_blocked = jiffies;
10405 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);