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);
3858 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
3860 u64 vruntime = cfs_rq->min_vruntime;
3863 * The 'current' period is already promised to the current tasks,
3864 * however the extra weight of the new task will slow them down a
3865 * little, place the new task so that it fits in the slot that
3866 * stays open at the end.
3868 if (initial && sched_feat(START_DEBIT))
3869 vruntime += sched_vslice(cfs_rq, se);
3871 /* sleeps up to a single latency don't count. */
3873 unsigned long thresh = sysctl_sched_latency;
3876 * Halve their sleep time's effect, to allow
3877 * for a gentler effect of sleepers:
3879 if (sched_feat(GENTLE_FAIR_SLEEPERS))
3885 /* ensure we never gain time by being placed backwards. */
3886 se->vruntime = max_vruntime(se->vruntime, vruntime);
3889 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
3891 static inline void check_schedstat_required(void)
3893 #ifdef CONFIG_SCHEDSTATS
3894 if (schedstat_enabled())
3897 /* Force schedstat enabled if a dependent tracepoint is active */
3898 if (trace_sched_stat_wait_enabled() ||
3899 trace_sched_stat_sleep_enabled() ||
3900 trace_sched_stat_iowait_enabled() ||
3901 trace_sched_stat_blocked_enabled() ||
3902 trace_sched_stat_runtime_enabled()) {
3903 printk_deferred_once("Scheduler tracepoints stat_sleep, stat_iowait, "
3904 "stat_blocked and stat_runtime require the "
3905 "kernel parameter schedstats=enable or "
3906 "kernel.sched_schedstats=1\n");
3917 * update_min_vruntime()
3918 * vruntime -= min_vruntime
3922 * update_min_vruntime()
3923 * vruntime += min_vruntime
3925 * this way the vruntime transition between RQs is done when both
3926 * min_vruntime are up-to-date.
3930 * ->migrate_task_rq_fair() (p->state == TASK_WAKING)
3931 * vruntime -= min_vruntime
3935 * update_min_vruntime()
3936 * vruntime += min_vruntime
3938 * this way we don't have the most up-to-date min_vruntime on the originating
3939 * CPU and an up-to-date min_vruntime on the destination CPU.
3943 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3945 bool renorm = !(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_MIGRATED);
3946 bool curr = cfs_rq->curr == se;
3949 * If we're the current task, we must renormalise before calling
3953 se->vruntime += cfs_rq->min_vruntime;
3955 update_curr(cfs_rq);
3958 * Otherwise, renormalise after, such that we're placed at the current
3959 * moment in time, instead of some random moment in the past. Being
3960 * placed in the past could significantly boost this task to the
3961 * fairness detriment of existing tasks.
3963 if (renorm && !curr)
3964 se->vruntime += cfs_rq->min_vruntime;
3967 * When enqueuing a sched_entity, we must:
3968 * - Update loads to have both entity and cfs_rq synced with now.
3969 * - Add its load to cfs_rq->runnable_avg
3970 * - For group_entity, update its weight to reflect the new share of
3972 * - Add its new weight to cfs_rq->load.weight
3974 update_load_avg(cfs_rq, se, UPDATE_TG | DO_ATTACH);
3975 update_cfs_group(se);
3976 enqueue_runnable_load_avg(cfs_rq, se);
3977 account_entity_enqueue(cfs_rq, se);
3979 if (flags & ENQUEUE_WAKEUP)
3980 place_entity(cfs_rq, se, 0);
3982 check_schedstat_required();
3983 update_stats_enqueue(cfs_rq, se, flags);
3984 check_spread(cfs_rq, se);
3986 __enqueue_entity(cfs_rq, se);
3989 if (cfs_rq->nr_running == 1) {
3990 list_add_leaf_cfs_rq(cfs_rq);
3991 check_enqueue_throttle(cfs_rq);
3995 static void __clear_buddies_last(struct sched_entity *se)
3997 for_each_sched_entity(se) {
3998 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3999 if (cfs_rq->last != se)
4002 cfs_rq->last = NULL;
4006 static void __clear_buddies_next(struct sched_entity *se)
4008 for_each_sched_entity(se) {
4009 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4010 if (cfs_rq->next != se)
4013 cfs_rq->next = NULL;
4017 static void __clear_buddies_skip(struct sched_entity *se)
4019 for_each_sched_entity(se) {
4020 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4021 if (cfs_rq->skip != se)
4024 cfs_rq->skip = NULL;
4028 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
4030 if (cfs_rq->last == se)
4031 __clear_buddies_last(se);
4033 if (cfs_rq->next == se)
4034 __clear_buddies_next(se);
4036 if (cfs_rq->skip == se)
4037 __clear_buddies_skip(se);
4040 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
4043 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
4046 * Update run-time statistics of the 'current'.
4048 update_curr(cfs_rq);
4051 * When dequeuing a sched_entity, we must:
4052 * - Update loads to have both entity and cfs_rq synced with now.
4053 * - Substract its load from the cfs_rq->runnable_avg.
4054 * - Substract its previous weight from cfs_rq->load.weight.
4055 * - For group entity, update its weight to reflect the new share
4056 * of its group cfs_rq.
4058 update_load_avg(cfs_rq, se, UPDATE_TG);
4059 dequeue_runnable_load_avg(cfs_rq, se);
4061 update_stats_dequeue(cfs_rq, se, flags);
4063 clear_buddies(cfs_rq, se);
4065 if (se != cfs_rq->curr)
4066 __dequeue_entity(cfs_rq, se);
4068 account_entity_dequeue(cfs_rq, se);
4071 * Normalize after update_curr(); which will also have moved
4072 * min_vruntime if @se is the one holding it back. But before doing
4073 * update_min_vruntime() again, which will discount @se's position and
4074 * can move min_vruntime forward still more.
4076 if (!(flags & DEQUEUE_SLEEP))
4077 se->vruntime -= cfs_rq->min_vruntime;
4079 /* return excess runtime on last dequeue */
4080 return_cfs_rq_runtime(cfs_rq);
4082 update_cfs_group(se);
4085 * Now advance min_vruntime if @se was the entity holding it back,
4086 * except when: DEQUEUE_SAVE && !DEQUEUE_MOVE, in this case we'll be
4087 * put back on, and if we advance min_vruntime, we'll be placed back
4088 * further than we started -- ie. we'll be penalized.
4090 if ((flags & (DEQUEUE_SAVE | DEQUEUE_MOVE)) != DEQUEUE_SAVE)
4091 update_min_vruntime(cfs_rq);
4095 * Preempt the current task with a newly woken task if needed:
4098 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
4100 unsigned long ideal_runtime, delta_exec;
4101 struct sched_entity *se;
4104 ideal_runtime = sched_slice(cfs_rq, curr);
4105 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
4106 if (delta_exec > ideal_runtime) {
4107 resched_curr(rq_of(cfs_rq));
4109 * The current task ran long enough, ensure it doesn't get
4110 * re-elected due to buddy favours.
4112 clear_buddies(cfs_rq, curr);
4117 * Ensure that a task that missed wakeup preemption by a
4118 * narrow margin doesn't have to wait for a full slice.
4119 * This also mitigates buddy induced latencies under load.
4121 if (delta_exec < sysctl_sched_min_granularity)
4124 se = __pick_first_entity(cfs_rq);
4125 delta = curr->vruntime - se->vruntime;
4130 if (delta > ideal_runtime)
4131 resched_curr(rq_of(cfs_rq));
4135 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
4137 /* 'current' is not kept within the tree. */
4140 * Any task has to be enqueued before it get to execute on
4141 * a CPU. So account for the time it spent waiting on the
4144 update_stats_wait_end(cfs_rq, se);
4145 __dequeue_entity(cfs_rq, se);
4146 update_load_avg(cfs_rq, se, UPDATE_TG);
4149 update_stats_curr_start(cfs_rq, se);
4153 * Track our maximum slice length, if the CPU's load is at
4154 * least twice that of our own weight (i.e. dont track it
4155 * when there are only lesser-weight tasks around):
4157 if (schedstat_enabled() && rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
4158 schedstat_set(se->statistics.slice_max,
4159 max((u64)schedstat_val(se->statistics.slice_max),
4160 se->sum_exec_runtime - se->prev_sum_exec_runtime));
4163 se->prev_sum_exec_runtime = se->sum_exec_runtime;
4167 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
4170 * Pick the next process, keeping these things in mind, in this order:
4171 * 1) keep things fair between processes/task groups
4172 * 2) pick the "next" process, since someone really wants that to run
4173 * 3) pick the "last" process, for cache locality
4174 * 4) do not run the "skip" process, if something else is available
4176 static struct sched_entity *
4177 pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr)
4179 struct sched_entity *left = __pick_first_entity(cfs_rq);
4180 struct sched_entity *se;
4183 * If curr is set we have to see if its left of the leftmost entity
4184 * still in the tree, provided there was anything in the tree at all.
4186 if (!left || (curr && entity_before(curr, left)))
4189 se = left; /* ideally we run the leftmost entity */
4192 * Avoid running the skip buddy, if running something else can
4193 * be done without getting too unfair.
4195 if (cfs_rq->skip == se) {
4196 struct sched_entity *second;
4199 second = __pick_first_entity(cfs_rq);
4201 second = __pick_next_entity(se);
4202 if (!second || (curr && entity_before(curr, second)))
4206 if (second && wakeup_preempt_entity(second, left) < 1)
4211 * Prefer last buddy, try to return the CPU to a preempted task.
4213 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
4217 * Someone really wants this to run. If it's not unfair, run it.
4219 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
4222 clear_buddies(cfs_rq, se);
4227 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
4229 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
4232 * If still on the runqueue then deactivate_task()
4233 * was not called and update_curr() has to be done:
4236 update_curr(cfs_rq);
4238 /* throttle cfs_rqs exceeding runtime */
4239 check_cfs_rq_runtime(cfs_rq);
4241 check_spread(cfs_rq, prev);
4244 update_stats_wait_start(cfs_rq, prev);
4245 /* Put 'current' back into the tree. */
4246 __enqueue_entity(cfs_rq, prev);
4247 /* in !on_rq case, update occurred at dequeue */
4248 update_load_avg(cfs_rq, prev, 0);
4250 cfs_rq->curr = NULL;
4254 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
4257 * Update run-time statistics of the 'current'.
4259 update_curr(cfs_rq);
4262 * Ensure that runnable average is periodically updated.
4264 update_load_avg(cfs_rq, curr, UPDATE_TG);
4265 update_cfs_group(curr);
4267 #ifdef CONFIG_SCHED_HRTICK
4269 * queued ticks are scheduled to match the slice, so don't bother
4270 * validating it and just reschedule.
4273 resched_curr(rq_of(cfs_rq));
4277 * don't let the period tick interfere with the hrtick preemption
4279 if (!sched_feat(DOUBLE_TICK) &&
4280 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
4284 if (cfs_rq->nr_running > 1)
4285 check_preempt_tick(cfs_rq, curr);
4289 /**************************************************
4290 * CFS bandwidth control machinery
4293 #ifdef CONFIG_CFS_BANDWIDTH
4295 #ifdef CONFIG_JUMP_LABEL
4296 static struct static_key __cfs_bandwidth_used;
4298 static inline bool cfs_bandwidth_used(void)
4300 return static_key_false(&__cfs_bandwidth_used);
4303 void cfs_bandwidth_usage_inc(void)
4305 static_key_slow_inc_cpuslocked(&__cfs_bandwidth_used);
4308 void cfs_bandwidth_usage_dec(void)
4310 static_key_slow_dec_cpuslocked(&__cfs_bandwidth_used);
4312 #else /* CONFIG_JUMP_LABEL */
4313 static bool cfs_bandwidth_used(void)
4318 void cfs_bandwidth_usage_inc(void) {}
4319 void cfs_bandwidth_usage_dec(void) {}
4320 #endif /* CONFIG_JUMP_LABEL */
4323 * default period for cfs group bandwidth.
4324 * default: 0.1s, units: nanoseconds
4326 static inline u64 default_cfs_period(void)
4328 return 100000000ULL;
4331 static inline u64 sched_cfs_bandwidth_slice(void)
4333 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
4337 * Replenish runtime according to assigned quota. We use sched_clock_cpu
4338 * directly instead of rq->clock to avoid adding additional synchronization
4341 * requires cfs_b->lock
4343 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
4345 if (cfs_b->quota != RUNTIME_INF)
4346 cfs_b->runtime = cfs_b->quota;
4349 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
4351 return &tg->cfs_bandwidth;
4354 /* rq->task_clock normalized against any time this cfs_rq has spent throttled */
4355 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
4357 if (unlikely(cfs_rq->throttle_count))
4358 return cfs_rq->throttled_clock_task - cfs_rq->throttled_clock_task_time;
4360 return rq_clock_task(rq_of(cfs_rq)) - cfs_rq->throttled_clock_task_time;
4363 /* returns 0 on failure to allocate runtime */
4364 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4366 struct task_group *tg = cfs_rq->tg;
4367 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
4368 u64 amount = 0, min_amount;
4370 /* note: this is a positive sum as runtime_remaining <= 0 */
4371 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
4373 raw_spin_lock(&cfs_b->lock);
4374 if (cfs_b->quota == RUNTIME_INF)
4375 amount = min_amount;
4377 start_cfs_bandwidth(cfs_b);
4379 if (cfs_b->runtime > 0) {
4380 amount = min(cfs_b->runtime, min_amount);
4381 cfs_b->runtime -= amount;
4385 raw_spin_unlock(&cfs_b->lock);
4387 cfs_rq->runtime_remaining += amount;
4389 return cfs_rq->runtime_remaining > 0;
4392 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
4394 /* dock delta_exec before expiring quota (as it could span periods) */
4395 cfs_rq->runtime_remaining -= delta_exec;
4397 if (likely(cfs_rq->runtime_remaining > 0))
4400 if (cfs_rq->throttled)
4403 * if we're unable to extend our runtime we resched so that the active
4404 * hierarchy can be throttled
4406 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
4407 resched_curr(rq_of(cfs_rq));
4410 static __always_inline
4411 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
4413 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
4416 __account_cfs_rq_runtime(cfs_rq, delta_exec);
4419 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
4421 return cfs_bandwidth_used() && cfs_rq->throttled;
4424 /* check whether cfs_rq, or any parent, is throttled */
4425 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
4427 return cfs_bandwidth_used() && cfs_rq->throttle_count;
4431 * Ensure that neither of the group entities corresponding to src_cpu or
4432 * dest_cpu are members of a throttled hierarchy when performing group
4433 * load-balance operations.
4435 static inline int throttled_lb_pair(struct task_group *tg,
4436 int src_cpu, int dest_cpu)
4438 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
4440 src_cfs_rq = tg->cfs_rq[src_cpu];
4441 dest_cfs_rq = tg->cfs_rq[dest_cpu];
4443 return throttled_hierarchy(src_cfs_rq) ||
4444 throttled_hierarchy(dest_cfs_rq);
4447 static int tg_unthrottle_up(struct task_group *tg, void *data)
4449 struct rq *rq = data;
4450 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
4452 cfs_rq->throttle_count--;
4453 if (!cfs_rq->throttle_count) {
4454 /* adjust cfs_rq_clock_task() */
4455 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
4456 cfs_rq->throttled_clock_task;
4458 /* Add cfs_rq with already running entity in the list */
4459 if (cfs_rq->nr_running >= 1)
4460 list_add_leaf_cfs_rq(cfs_rq);
4466 static int tg_throttle_down(struct task_group *tg, void *data)
4468 struct rq *rq = data;
4469 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
4471 /* group is entering throttled state, stop time */
4472 if (!cfs_rq->throttle_count) {
4473 cfs_rq->throttled_clock_task = rq_clock_task(rq);
4474 list_del_leaf_cfs_rq(cfs_rq);
4476 cfs_rq->throttle_count++;
4481 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
4483 struct rq *rq = rq_of(cfs_rq);
4484 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4485 struct sched_entity *se;
4486 long task_delta, dequeue = 1;
4489 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
4491 /* freeze hierarchy runnable averages while throttled */
4493 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
4496 task_delta = cfs_rq->h_nr_running;
4497 for_each_sched_entity(se) {
4498 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
4499 /* throttled entity or throttle-on-deactivate */
4504 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
4505 qcfs_rq->h_nr_running -= task_delta;
4507 if (qcfs_rq->load.weight)
4512 sub_nr_running(rq, task_delta);
4514 cfs_rq->throttled = 1;
4515 cfs_rq->throttled_clock = rq_clock(rq);
4516 raw_spin_lock(&cfs_b->lock);
4517 empty = list_empty(&cfs_b->throttled_cfs_rq);
4520 * Add to the _head_ of the list, so that an already-started
4521 * distribute_cfs_runtime will not see us. If disribute_cfs_runtime is
4522 * not running add to the tail so that later runqueues don't get starved.
4524 if (cfs_b->distribute_running)
4525 list_add_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
4527 list_add_tail_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
4530 * If we're the first throttled task, make sure the bandwidth
4534 start_cfs_bandwidth(cfs_b);
4536 raw_spin_unlock(&cfs_b->lock);
4539 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
4541 struct rq *rq = rq_of(cfs_rq);
4542 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4543 struct sched_entity *se;
4547 se = cfs_rq->tg->se[cpu_of(rq)];
4549 cfs_rq->throttled = 0;
4551 update_rq_clock(rq);
4553 raw_spin_lock(&cfs_b->lock);
4554 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
4555 list_del_rcu(&cfs_rq->throttled_list);
4556 raw_spin_unlock(&cfs_b->lock);
4558 /* update hierarchical throttle state */
4559 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
4561 if (!cfs_rq->load.weight)
4564 task_delta = cfs_rq->h_nr_running;
4565 for_each_sched_entity(se) {
4569 cfs_rq = cfs_rq_of(se);
4571 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
4572 cfs_rq->h_nr_running += task_delta;
4574 if (cfs_rq_throttled(cfs_rq))
4578 assert_list_leaf_cfs_rq(rq);
4581 add_nr_running(rq, task_delta);
4583 /* Determine whether we need to wake up potentially idle CPU: */
4584 if (rq->curr == rq->idle && rq->cfs.nr_running)
4588 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b, u64 remaining)
4590 struct cfs_rq *cfs_rq;
4592 u64 starting_runtime = remaining;
4595 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
4597 struct rq *rq = rq_of(cfs_rq);
4601 if (!cfs_rq_throttled(cfs_rq))
4604 /* By the above check, this should never be true */
4605 SCHED_WARN_ON(cfs_rq->runtime_remaining > 0);
4607 runtime = -cfs_rq->runtime_remaining + 1;
4608 if (runtime > remaining)
4609 runtime = remaining;
4610 remaining -= runtime;
4612 cfs_rq->runtime_remaining += runtime;
4614 /* we check whether we're throttled above */
4615 if (cfs_rq->runtime_remaining > 0)
4616 unthrottle_cfs_rq(cfs_rq);
4626 return starting_runtime - remaining;
4630 * Responsible for refilling a task_group's bandwidth and unthrottling its
4631 * cfs_rqs as appropriate. If there has been no activity within the last
4632 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
4633 * used to track this state.
4635 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun)
4640 /* no need to continue the timer with no bandwidth constraint */
4641 if (cfs_b->quota == RUNTIME_INF)
4642 goto out_deactivate;
4644 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
4645 cfs_b->nr_periods += overrun;
4648 * idle depends on !throttled (for the case of a large deficit), and if
4649 * we're going inactive then everything else can be deferred
4651 if (cfs_b->idle && !throttled)
4652 goto out_deactivate;
4654 __refill_cfs_bandwidth_runtime(cfs_b);
4657 /* mark as potentially idle for the upcoming period */
4662 /* account preceding periods in which throttling occurred */
4663 cfs_b->nr_throttled += overrun;
4666 * This check is repeated as we are holding onto the new bandwidth while
4667 * we unthrottle. This can potentially race with an unthrottled group
4668 * trying to acquire new bandwidth from the global pool. This can result
4669 * in us over-using our runtime if it is all used during this loop, but
4670 * only by limited amounts in that extreme case.
4672 while (throttled && cfs_b->runtime > 0 && !cfs_b->distribute_running) {
4673 runtime = cfs_b->runtime;
4674 cfs_b->distribute_running = 1;
4675 raw_spin_unlock(&cfs_b->lock);
4676 /* we can't nest cfs_b->lock while distributing bandwidth */
4677 runtime = distribute_cfs_runtime(cfs_b, runtime);
4678 raw_spin_lock(&cfs_b->lock);
4680 cfs_b->distribute_running = 0;
4681 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
4683 cfs_b->runtime -= min(runtime, cfs_b->runtime);
4687 * While we are ensured activity in the period following an
4688 * unthrottle, this also covers the case in which the new bandwidth is
4689 * insufficient to cover the existing bandwidth deficit. (Forcing the
4690 * timer to remain active while there are any throttled entities.)
4700 /* a cfs_rq won't donate quota below this amount */
4701 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
4702 /* minimum remaining period time to redistribute slack quota */
4703 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
4704 /* how long we wait to gather additional slack before distributing */
4705 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
4708 * Are we near the end of the current quota period?
4710 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
4711 * hrtimer base being cleared by hrtimer_start. In the case of
4712 * migrate_hrtimers, base is never cleared, so we are fine.
4714 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
4716 struct hrtimer *refresh_timer = &cfs_b->period_timer;
4719 /* if the call-back is running a quota refresh is already occurring */
4720 if (hrtimer_callback_running(refresh_timer))
4723 /* is a quota refresh about to occur? */
4724 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
4725 if (remaining < (s64)min_expire)
4731 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
4733 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
4735 /* if there's a quota refresh soon don't bother with slack */
4736 if (runtime_refresh_within(cfs_b, min_left))
4739 hrtimer_start(&cfs_b->slack_timer,
4740 ns_to_ktime(cfs_bandwidth_slack_period),
4744 /* we know any runtime found here is valid as update_curr() precedes return */
4745 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4747 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4748 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
4750 if (slack_runtime <= 0)
4753 raw_spin_lock(&cfs_b->lock);
4754 if (cfs_b->quota != RUNTIME_INF) {
4755 cfs_b->runtime += slack_runtime;
4757 /* we are under rq->lock, defer unthrottling using a timer */
4758 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
4759 !list_empty(&cfs_b->throttled_cfs_rq))
4760 start_cfs_slack_bandwidth(cfs_b);
4762 raw_spin_unlock(&cfs_b->lock);
4764 /* even if it's not valid for return we don't want to try again */
4765 cfs_rq->runtime_remaining -= slack_runtime;
4768 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4770 if (!cfs_bandwidth_used())
4773 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
4776 __return_cfs_rq_runtime(cfs_rq);
4780 * This is done with a timer (instead of inline with bandwidth return) since
4781 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
4783 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
4785 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
4787 /* confirm we're still not at a refresh boundary */
4788 raw_spin_lock(&cfs_b->lock);
4789 if (cfs_b->distribute_running) {
4790 raw_spin_unlock(&cfs_b->lock);
4794 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
4795 raw_spin_unlock(&cfs_b->lock);
4799 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice)
4800 runtime = cfs_b->runtime;
4803 cfs_b->distribute_running = 1;
4805 raw_spin_unlock(&cfs_b->lock);
4810 runtime = distribute_cfs_runtime(cfs_b, runtime);
4812 raw_spin_lock(&cfs_b->lock);
4813 cfs_b->runtime -= min(runtime, cfs_b->runtime);
4814 cfs_b->distribute_running = 0;
4815 raw_spin_unlock(&cfs_b->lock);
4819 * When a group wakes up we want to make sure that its quota is not already
4820 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
4821 * runtime as update_curr() throttling can not not trigger until it's on-rq.
4823 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
4825 if (!cfs_bandwidth_used())
4828 /* an active group must be handled by the update_curr()->put() path */
4829 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
4832 /* ensure the group is not already throttled */
4833 if (cfs_rq_throttled(cfs_rq))
4836 /* update runtime allocation */
4837 account_cfs_rq_runtime(cfs_rq, 0);
4838 if (cfs_rq->runtime_remaining <= 0)
4839 throttle_cfs_rq(cfs_rq);
4842 static void sync_throttle(struct task_group *tg, int cpu)
4844 struct cfs_rq *pcfs_rq, *cfs_rq;
4846 if (!cfs_bandwidth_used())
4852 cfs_rq = tg->cfs_rq[cpu];
4853 pcfs_rq = tg->parent->cfs_rq[cpu];
4855 cfs_rq->throttle_count = pcfs_rq->throttle_count;
4856 cfs_rq->throttled_clock_task = rq_clock_task(cpu_rq(cpu));
4859 /* conditionally throttle active cfs_rq's from put_prev_entity() */
4860 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4862 if (!cfs_bandwidth_used())
4865 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
4869 * it's possible for a throttled entity to be forced into a running
4870 * state (e.g. set_curr_task), in this case we're finished.
4872 if (cfs_rq_throttled(cfs_rq))
4875 throttle_cfs_rq(cfs_rq);
4879 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
4881 struct cfs_bandwidth *cfs_b =
4882 container_of(timer, struct cfs_bandwidth, slack_timer);
4884 do_sched_cfs_slack_timer(cfs_b);
4886 return HRTIMER_NORESTART;
4889 extern const u64 max_cfs_quota_period;
4891 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
4893 struct cfs_bandwidth *cfs_b =
4894 container_of(timer, struct cfs_bandwidth, period_timer);
4899 raw_spin_lock(&cfs_b->lock);
4901 overrun = hrtimer_forward_now(timer, cfs_b->period);
4906 u64 new, old = ktime_to_ns(cfs_b->period);
4909 * Grow period by a factor of 2 to avoid losing precision.
4910 * Precision loss in the quota/period ratio can cause __cfs_schedulable
4914 if (new < max_cfs_quota_period) {
4915 cfs_b->period = ns_to_ktime(new);
4918 pr_warn_ratelimited(
4919 "cfs_period_timer[cpu%d]: period too short, scaling up (new cfs_period_us = %lld, cfs_quota_us = %lld)\n",
4921 div_u64(new, NSEC_PER_USEC),
4922 div_u64(cfs_b->quota, NSEC_PER_USEC));
4924 pr_warn_ratelimited(
4925 "cfs_period_timer[cpu%d]: period too short, but cannot scale up without losing precision (cfs_period_us = %lld, cfs_quota_us = %lld)\n",
4927 div_u64(old, NSEC_PER_USEC),
4928 div_u64(cfs_b->quota, NSEC_PER_USEC));
4931 /* reset count so we don't come right back in here */
4935 idle = do_sched_cfs_period_timer(cfs_b, overrun);
4938 cfs_b->period_active = 0;
4939 raw_spin_unlock(&cfs_b->lock);
4941 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
4944 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4946 raw_spin_lock_init(&cfs_b->lock);
4948 cfs_b->quota = RUNTIME_INF;
4949 cfs_b->period = ns_to_ktime(default_cfs_period());
4951 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
4952 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED);
4953 cfs_b->period_timer.function = sched_cfs_period_timer;
4954 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
4955 cfs_b->slack_timer.function = sched_cfs_slack_timer;
4956 cfs_b->distribute_running = 0;
4959 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4961 cfs_rq->runtime_enabled = 0;
4962 INIT_LIST_HEAD(&cfs_rq->throttled_list);
4965 void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4967 lockdep_assert_held(&cfs_b->lock);
4969 if (cfs_b->period_active)
4972 cfs_b->period_active = 1;
4973 hrtimer_forward_now(&cfs_b->period_timer, cfs_b->period);
4974 hrtimer_start_expires(&cfs_b->period_timer, HRTIMER_MODE_ABS_PINNED);
4977 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
4979 /* init_cfs_bandwidth() was not called */
4980 if (!cfs_b->throttled_cfs_rq.next)
4983 hrtimer_cancel(&cfs_b->period_timer);
4984 hrtimer_cancel(&cfs_b->slack_timer);
4988 * Both these CPU hotplug callbacks race against unregister_fair_sched_group()
4990 * The race is harmless, since modifying bandwidth settings of unhooked group
4991 * bits doesn't do much.
4994 /* cpu online calback */
4995 static void __maybe_unused update_runtime_enabled(struct rq *rq)
4997 struct task_group *tg;
4999 lockdep_assert_held(&rq->lock);
5002 list_for_each_entry_rcu(tg, &task_groups, list) {
5003 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
5004 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
5006 raw_spin_lock(&cfs_b->lock);
5007 cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF;
5008 raw_spin_unlock(&cfs_b->lock);
5013 /* cpu offline callback */
5014 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
5016 struct task_group *tg;
5018 lockdep_assert_held(&rq->lock);
5021 list_for_each_entry_rcu(tg, &task_groups, list) {
5022 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
5024 if (!cfs_rq->runtime_enabled)
5028 * clock_task is not advancing so we just need to make sure
5029 * there's some valid quota amount
5031 cfs_rq->runtime_remaining = 1;
5033 * Offline rq is schedulable till CPU is completely disabled
5034 * in take_cpu_down(), so we prevent new cfs throttling here.
5036 cfs_rq->runtime_enabled = 0;
5038 if (cfs_rq_throttled(cfs_rq))
5039 unthrottle_cfs_rq(cfs_rq);
5044 #else /* CONFIG_CFS_BANDWIDTH */
5046 static inline bool cfs_bandwidth_used(void)
5051 static inline u64 cfs_rq_clock_task(struct cfs_rq *cfs_rq)
5053 return rq_clock_task(rq_of(cfs_rq));
5056 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {}
5057 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; }
5058 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
5059 static inline void sync_throttle(struct task_group *tg, int cpu) {}
5060 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
5062 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
5067 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
5072 static inline int throttled_lb_pair(struct task_group *tg,
5073 int src_cpu, int dest_cpu)
5078 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
5080 #ifdef CONFIG_FAIR_GROUP_SCHED
5081 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
5084 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
5088 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
5089 static inline void update_runtime_enabled(struct rq *rq) {}
5090 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
5092 #endif /* CONFIG_CFS_BANDWIDTH */
5094 /**************************************************
5095 * CFS operations on tasks:
5098 #ifdef CONFIG_SCHED_HRTICK
5099 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
5101 struct sched_entity *se = &p->se;
5102 struct cfs_rq *cfs_rq = cfs_rq_of(se);
5104 SCHED_WARN_ON(task_rq(p) != rq);
5106 if (rq->cfs.h_nr_running > 1) {
5107 u64 slice = sched_slice(cfs_rq, se);
5108 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
5109 s64 delta = slice - ran;
5116 hrtick_start(rq, delta);
5121 * called from enqueue/dequeue and updates the hrtick when the
5122 * current task is from our class and nr_running is low enough
5125 static void hrtick_update(struct rq *rq)
5127 struct task_struct *curr = rq->curr;
5129 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
5132 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
5133 hrtick_start_fair(rq, curr);
5135 #else /* !CONFIG_SCHED_HRTICK */
5137 hrtick_start_fair(struct rq *rq, struct task_struct *p)
5141 static inline void hrtick_update(struct rq *rq)
5147 * The enqueue_task method is called before nr_running is
5148 * increased. Here we update the fair scheduling stats and
5149 * then put the task into the rbtree:
5152 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
5154 struct cfs_rq *cfs_rq;
5155 struct sched_entity *se = &p->se;
5158 * The code below (indirectly) updates schedutil which looks at
5159 * the cfs_rq utilization to select a frequency.
5160 * Let's add the task's estimated utilization to the cfs_rq's
5161 * estimated utilization, before we update schedutil.
5163 util_est_enqueue(&rq->cfs, p);
5166 * If in_iowait is set, the code below may not trigger any cpufreq
5167 * utilization updates, so do it here explicitly with the IOWAIT flag
5171 cpufreq_update_util(rq, SCHED_CPUFREQ_IOWAIT);
5173 for_each_sched_entity(se) {
5176 cfs_rq = cfs_rq_of(se);
5177 enqueue_entity(cfs_rq, se, flags);
5180 * end evaluation on encountering a throttled cfs_rq
5182 * note: in the case of encountering a throttled cfs_rq we will
5183 * post the final h_nr_running increment below.
5185 if (cfs_rq_throttled(cfs_rq))
5187 cfs_rq->h_nr_running++;
5189 flags = ENQUEUE_WAKEUP;
5192 for_each_sched_entity(se) {
5193 cfs_rq = cfs_rq_of(se);
5194 cfs_rq->h_nr_running++;
5196 if (cfs_rq_throttled(cfs_rq))
5199 update_load_avg(cfs_rq, se, UPDATE_TG);
5200 update_cfs_group(se);
5204 add_nr_running(rq, 1);
5206 if (cfs_bandwidth_used()) {
5208 * When bandwidth control is enabled; the cfs_rq_throttled()
5209 * breaks in the above iteration can result in incomplete
5210 * leaf list maintenance, resulting in triggering the assertion
5213 for_each_sched_entity(se) {
5214 cfs_rq = cfs_rq_of(se);
5216 if (list_add_leaf_cfs_rq(cfs_rq))
5221 assert_list_leaf_cfs_rq(rq);
5226 static void set_next_buddy(struct sched_entity *se);
5229 * The dequeue_task method is called before nr_running is
5230 * decreased. We remove the task from the rbtree and
5231 * update the fair scheduling stats:
5233 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
5235 struct cfs_rq *cfs_rq;
5236 struct sched_entity *se = &p->se;
5237 int task_sleep = flags & DEQUEUE_SLEEP;
5239 for_each_sched_entity(se) {
5240 cfs_rq = cfs_rq_of(se);
5241 dequeue_entity(cfs_rq, se, flags);
5244 * end evaluation on encountering a throttled cfs_rq
5246 * note: in the case of encountering a throttled cfs_rq we will
5247 * post the final h_nr_running decrement below.
5249 if (cfs_rq_throttled(cfs_rq))
5251 cfs_rq->h_nr_running--;
5253 /* Don't dequeue parent if it has other entities besides us */
5254 if (cfs_rq->load.weight) {
5255 /* Avoid re-evaluating load for this entity: */
5256 se = parent_entity(se);
5258 * Bias pick_next to pick a task from this cfs_rq, as
5259 * p is sleeping when it is within its sched_slice.
5261 if (task_sleep && se && !throttled_hierarchy(cfs_rq))
5265 flags |= DEQUEUE_SLEEP;
5268 for_each_sched_entity(se) {
5269 cfs_rq = cfs_rq_of(se);
5270 cfs_rq->h_nr_running--;
5272 if (cfs_rq_throttled(cfs_rq))
5275 update_load_avg(cfs_rq, se, UPDATE_TG);
5276 update_cfs_group(se);
5280 sub_nr_running(rq, 1);
5282 util_est_dequeue(&rq->cfs, p, task_sleep);
5288 /* Working cpumask for: load_balance, load_balance_newidle. */
5289 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
5290 DEFINE_PER_CPU(cpumask_var_t, select_idle_mask);
5292 #ifdef CONFIG_NO_HZ_COMMON
5294 * per rq 'load' arrray crap; XXX kill this.
5298 * The exact cpuload calculated at every tick would be:
5300 * load' = (1 - 1/2^i) * load + (1/2^i) * cur_load
5302 * If a CPU misses updates for n ticks (as it was idle) and update gets
5303 * called on the n+1-th tick when CPU may be busy, then we have:
5305 * load_n = (1 - 1/2^i)^n * load_0
5306 * load_n+1 = (1 - 1/2^i) * load_n + (1/2^i) * cur_load
5308 * decay_load_missed() below does efficient calculation of
5310 * load' = (1 - 1/2^i)^n * load
5312 * Because x^(n+m) := x^n * x^m we can decompose any x^n in power-of-2 factors.
5313 * This allows us to precompute the above in said factors, thereby allowing the
5314 * reduction of an arbitrary n in O(log_2 n) steps. (See also
5315 * fixed_power_int())
5317 * The calculation is approximated on a 128 point scale.
5319 #define DEGRADE_SHIFT 7
5321 static const u8 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
5322 static const u8 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
5323 { 0, 0, 0, 0, 0, 0, 0, 0 },
5324 { 64, 32, 8, 0, 0, 0, 0, 0 },
5325 { 96, 72, 40, 12, 1, 0, 0, 0 },
5326 { 112, 98, 75, 43, 15, 1, 0, 0 },
5327 { 120, 112, 98, 76, 45, 16, 2, 0 }
5331 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
5332 * would be when CPU is idle and so we just decay the old load without
5333 * adding any new load.
5335 static unsigned long
5336 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
5340 if (!missed_updates)
5343 if (missed_updates >= degrade_zero_ticks[idx])
5347 return load >> missed_updates;
5349 while (missed_updates) {
5350 if (missed_updates % 2)
5351 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
5353 missed_updates >>= 1;
5360 cpumask_var_t idle_cpus_mask;
5362 int has_blocked; /* Idle CPUS has blocked load */
5363 unsigned long next_balance; /* in jiffy units */
5364 unsigned long next_blocked; /* Next update of blocked load in jiffies */
5365 } nohz ____cacheline_aligned;
5367 #endif /* CONFIG_NO_HZ_COMMON */
5370 * __cpu_load_update - update the rq->cpu_load[] statistics
5371 * @this_rq: The rq to update statistics for
5372 * @this_load: The current load
5373 * @pending_updates: The number of missed updates
5375 * Update rq->cpu_load[] statistics. This function is usually called every
5376 * scheduler tick (TICK_NSEC).
5378 * This function computes a decaying average:
5380 * load[i]' = (1 - 1/2^i) * load[i] + (1/2^i) * load
5382 * Because of NOHZ it might not get called on every tick which gives need for
5383 * the @pending_updates argument.
5385 * load[i]_n = (1 - 1/2^i) * load[i]_n-1 + (1/2^i) * load_n-1
5386 * = A * load[i]_n-1 + B ; A := (1 - 1/2^i), B := (1/2^i) * load
5387 * = A * (A * load[i]_n-2 + B) + B
5388 * = A * (A * (A * load[i]_n-3 + B) + B) + B
5389 * = A^3 * load[i]_n-3 + (A^2 + A + 1) * B
5390 * = A^n * load[i]_0 + (A^(n-1) + A^(n-2) + ... + 1) * B
5391 * = A^n * load[i]_0 + ((1 - A^n) / (1 - A)) * B
5392 * = (1 - 1/2^i)^n * (load[i]_0 - load) + load
5394 * In the above we've assumed load_n := load, which is true for NOHZ_FULL as
5395 * any change in load would have resulted in the tick being turned back on.
5397 * For regular NOHZ, this reduces to:
5399 * load[i]_n = (1 - 1/2^i)^n * load[i]_0
5401 * see decay_load_misses(). For NOHZ_FULL we get to subtract and add the extra
5404 static void cpu_load_update(struct rq *this_rq, unsigned long this_load,
5405 unsigned long pending_updates)
5407 unsigned long __maybe_unused tickless_load = this_rq->cpu_load[0];
5410 this_rq->nr_load_updates++;
5412 /* Update our load: */
5413 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
5414 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
5415 unsigned long old_load, new_load;
5417 /* scale is effectively 1 << i now, and >> i divides by scale */
5419 old_load = this_rq->cpu_load[i];
5420 #ifdef CONFIG_NO_HZ_COMMON
5421 old_load = decay_load_missed(old_load, pending_updates - 1, i);
5422 if (tickless_load) {
5423 old_load -= decay_load_missed(tickless_load, pending_updates - 1, i);
5425 * old_load can never be a negative value because a
5426 * decayed tickless_load cannot be greater than the
5427 * original tickless_load.
5429 old_load += tickless_load;
5432 new_load = this_load;
5434 * Round up the averaging division if load is increasing. This
5435 * prevents us from getting stuck on 9 if the load is 10, for
5438 if (new_load > old_load)
5439 new_load += scale - 1;
5441 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
5445 /* Used instead of source_load when we know the type == 0 */
5446 static unsigned long weighted_cpuload(struct rq *rq)
5448 return cfs_rq_runnable_load_avg(&rq->cfs);
5451 #ifdef CONFIG_NO_HZ_COMMON
5453 * There is no sane way to deal with nohz on smp when using jiffies because the
5454 * CPU doing the jiffies update might drift wrt the CPU doing the jiffy reading
5455 * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}.
5457 * Therefore we need to avoid the delta approach from the regular tick when
5458 * possible since that would seriously skew the load calculation. This is why we
5459 * use cpu_load_update_periodic() for CPUs out of nohz. However we'll rely on
5460 * jiffies deltas for updates happening while in nohz mode (idle ticks, idle
5461 * loop exit, nohz_idle_balance, nohz full exit...)
5463 * This means we might still be one tick off for nohz periods.
5466 static void cpu_load_update_nohz(struct rq *this_rq,
5467 unsigned long curr_jiffies,
5470 unsigned long pending_updates;
5472 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
5473 if (pending_updates) {
5474 this_rq->last_load_update_tick = curr_jiffies;
5476 * In the regular NOHZ case, we were idle, this means load 0.
5477 * In the NOHZ_FULL case, we were non-idle, we should consider
5478 * its weighted load.
5480 cpu_load_update(this_rq, load, pending_updates);
5485 * Called from nohz_idle_balance() to update the load ratings before doing the
5488 static void cpu_load_update_idle(struct rq *this_rq)
5491 * bail if there's load or we're actually up-to-date.
5493 if (weighted_cpuload(this_rq))
5496 cpu_load_update_nohz(this_rq, READ_ONCE(jiffies), 0);
5500 * Record CPU load on nohz entry so we know the tickless load to account
5501 * on nohz exit. cpu_load[0] happens then to be updated more frequently
5502 * than other cpu_load[idx] but it should be fine as cpu_load readers
5503 * shouldn't rely into synchronized cpu_load[*] updates.
5505 void cpu_load_update_nohz_start(void)
5507 struct rq *this_rq = this_rq();
5510 * This is all lockless but should be fine. If weighted_cpuload changes
5511 * concurrently we'll exit nohz. And cpu_load write can race with
5512 * cpu_load_update_idle() but both updater would be writing the same.
5514 this_rq->cpu_load[0] = weighted_cpuload(this_rq);
5518 * Account the tickless load in the end of a nohz frame.
5520 void cpu_load_update_nohz_stop(void)
5522 unsigned long curr_jiffies = READ_ONCE(jiffies);
5523 struct rq *this_rq = this_rq();
5527 if (curr_jiffies == this_rq->last_load_update_tick)
5530 load = weighted_cpuload(this_rq);
5531 rq_lock(this_rq, &rf);
5532 update_rq_clock(this_rq);
5533 cpu_load_update_nohz(this_rq, curr_jiffies, load);
5534 rq_unlock(this_rq, &rf);
5536 #else /* !CONFIG_NO_HZ_COMMON */
5537 static inline void cpu_load_update_nohz(struct rq *this_rq,
5538 unsigned long curr_jiffies,
5539 unsigned long load) { }
5540 #endif /* CONFIG_NO_HZ_COMMON */
5542 static void cpu_load_update_periodic(struct rq *this_rq, unsigned long load)
5544 #ifdef CONFIG_NO_HZ_COMMON
5545 /* See the mess around cpu_load_update_nohz(). */
5546 this_rq->last_load_update_tick = READ_ONCE(jiffies);
5548 cpu_load_update(this_rq, load, 1);
5552 * Called from scheduler_tick()
5554 void cpu_load_update_active(struct rq *this_rq)
5556 unsigned long load = weighted_cpuload(this_rq);
5558 if (tick_nohz_tick_stopped())
5559 cpu_load_update_nohz(this_rq, READ_ONCE(jiffies), load);
5561 cpu_load_update_periodic(this_rq, load);
5565 * Return a low guess at the load of a migration-source CPU weighted
5566 * according to the scheduling class and "nice" value.
5568 * We want to under-estimate the load of migration sources, to
5569 * balance conservatively.
5571 static unsigned long source_load(int cpu, int type)
5573 struct rq *rq = cpu_rq(cpu);
5574 unsigned long total = weighted_cpuload(rq);
5576 if (type == 0 || !sched_feat(LB_BIAS))
5579 return min(rq->cpu_load[type-1], total);
5583 * Return a high guess at the load of a migration-target CPU weighted
5584 * according to the scheduling class and "nice" value.
5586 static unsigned long target_load(int cpu, int type)
5588 struct rq *rq = cpu_rq(cpu);
5589 unsigned long total = weighted_cpuload(rq);
5591 if (type == 0 || !sched_feat(LB_BIAS))
5594 return max(rq->cpu_load[type-1], total);
5597 static unsigned long capacity_of(int cpu)
5599 return cpu_rq(cpu)->cpu_capacity;
5602 static unsigned long capacity_orig_of(int cpu)
5604 return cpu_rq(cpu)->cpu_capacity_orig;
5607 static unsigned long cpu_avg_load_per_task(int cpu)
5609 struct rq *rq = cpu_rq(cpu);
5610 unsigned long nr_running = READ_ONCE(rq->cfs.h_nr_running);
5611 unsigned long load_avg = weighted_cpuload(rq);
5614 return load_avg / nr_running;
5619 static void record_wakee(struct task_struct *p)
5622 * Only decay a single time; tasks that have less then 1 wakeup per
5623 * jiffy will not have built up many flips.
5625 if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) {
5626 current->wakee_flips >>= 1;
5627 current->wakee_flip_decay_ts = jiffies;
5630 if (current->last_wakee != p) {
5631 current->last_wakee = p;
5632 current->wakee_flips++;
5637 * Detect M:N waker/wakee relationships via a switching-frequency heuristic.
5639 * A waker of many should wake a different task than the one last awakened
5640 * at a frequency roughly N times higher than one of its wakees.
5642 * In order to determine whether we should let the load spread vs consolidating
5643 * to shared cache, we look for a minimum 'flip' frequency of llc_size in one
5644 * partner, and a factor of lls_size higher frequency in the other.
5646 * With both conditions met, we can be relatively sure that the relationship is
5647 * non-monogamous, with partner count exceeding socket size.
5649 * Waker/wakee being client/server, worker/dispatcher, interrupt source or
5650 * whatever is irrelevant, spread criteria is apparent partner count exceeds
5653 static int wake_wide(struct task_struct *p)
5655 unsigned int master = current->wakee_flips;
5656 unsigned int slave = p->wakee_flips;
5657 int factor = this_cpu_read(sd_llc_size);
5660 swap(master, slave);
5661 if (slave < factor || master < slave * factor)
5667 * The purpose of wake_affine() is to quickly determine on which CPU we can run
5668 * soonest. For the purpose of speed we only consider the waking and previous
5671 * wake_affine_idle() - only considers 'now', it check if the waking CPU is
5672 * cache-affine and is (or will be) idle.
5674 * wake_affine_weight() - considers the weight to reflect the average
5675 * scheduling latency of the CPUs. This seems to work
5676 * for the overloaded case.
5679 wake_affine_idle(int this_cpu, int prev_cpu, int sync)
5682 * If this_cpu is idle, it implies the wakeup is from interrupt
5683 * context. Only allow the move if cache is shared. Otherwise an
5684 * interrupt intensive workload could force all tasks onto one
5685 * node depending on the IO topology or IRQ affinity settings.
5687 * If the prev_cpu is idle and cache affine then avoid a migration.
5688 * There is no guarantee that the cache hot data from an interrupt
5689 * is more important than cache hot data on the prev_cpu and from
5690 * a cpufreq perspective, it's better to have higher utilisation
5693 if (available_idle_cpu(this_cpu) && cpus_share_cache(this_cpu, prev_cpu))
5694 return available_idle_cpu(prev_cpu) ? prev_cpu : this_cpu;
5696 if (sync && cpu_rq(this_cpu)->nr_running == 1)
5699 return nr_cpumask_bits;
5703 wake_affine_weight(struct sched_domain *sd, struct task_struct *p,
5704 int this_cpu, int prev_cpu, int sync)
5706 s64 this_eff_load, prev_eff_load;
5707 unsigned long task_load;
5709 this_eff_load = target_load(this_cpu, sd->wake_idx);
5712 unsigned long current_load = task_h_load(current);
5714 if (current_load > this_eff_load)
5717 this_eff_load -= current_load;
5720 task_load = task_h_load(p);
5722 this_eff_load += task_load;
5723 if (sched_feat(WA_BIAS))
5724 this_eff_load *= 100;
5725 this_eff_load *= capacity_of(prev_cpu);
5727 prev_eff_load = source_load(prev_cpu, sd->wake_idx);
5728 prev_eff_load -= task_load;
5729 if (sched_feat(WA_BIAS))
5730 prev_eff_load *= 100 + (sd->imbalance_pct - 100) / 2;
5731 prev_eff_load *= capacity_of(this_cpu);
5734 * If sync, adjust the weight of prev_eff_load such that if
5735 * prev_eff == this_eff that select_idle_sibling() will consider
5736 * stacking the wakee on top of the waker if no other CPU is
5742 return this_eff_load < prev_eff_load ? this_cpu : nr_cpumask_bits;
5745 static int wake_affine(struct sched_domain *sd, struct task_struct *p,
5746 int this_cpu, int prev_cpu, int sync)
5748 int target = nr_cpumask_bits;
5750 if (sched_feat(WA_IDLE))
5751 target = wake_affine_idle(this_cpu, prev_cpu, sync);
5753 if (sched_feat(WA_WEIGHT) && target == nr_cpumask_bits)
5754 target = wake_affine_weight(sd, p, this_cpu, prev_cpu, sync);
5756 schedstat_inc(p->se.statistics.nr_wakeups_affine_attempts);
5757 if (target == nr_cpumask_bits)
5760 schedstat_inc(sd->ttwu_move_affine);
5761 schedstat_inc(p->se.statistics.nr_wakeups_affine);
5765 static unsigned long cpu_util_without(int cpu, struct task_struct *p);
5767 static unsigned long capacity_spare_without(int cpu, struct task_struct *p)
5769 return max_t(long, capacity_of(cpu) - cpu_util_without(cpu, p), 0);
5773 * find_idlest_group finds and returns the least busy CPU group within the
5776 * Assumes p is allowed on at least one CPU in sd.
5778 static struct sched_group *
5779 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
5780 int this_cpu, int sd_flag)
5782 struct sched_group *idlest = NULL, *group = sd->groups;
5783 struct sched_group *most_spare_sg = NULL;
5784 unsigned long min_runnable_load = ULONG_MAX;
5785 unsigned long this_runnable_load = ULONG_MAX;
5786 unsigned long min_avg_load = ULONG_MAX, this_avg_load = ULONG_MAX;
5787 unsigned long most_spare = 0, this_spare = 0;
5788 int load_idx = sd->forkexec_idx;
5789 int imbalance_scale = 100 + (sd->imbalance_pct-100)/2;
5790 unsigned long imbalance = scale_load_down(NICE_0_LOAD) *
5791 (sd->imbalance_pct-100) / 100;
5793 if (sd_flag & SD_BALANCE_WAKE)
5794 load_idx = sd->wake_idx;
5797 unsigned long load, avg_load, runnable_load;
5798 unsigned long spare_cap, max_spare_cap;
5802 /* Skip over this group if it has no CPUs allowed */
5803 if (!cpumask_intersects(sched_group_span(group),
5807 local_group = cpumask_test_cpu(this_cpu,
5808 sched_group_span(group));
5811 * Tally up the load of all CPUs in the group and find
5812 * the group containing the CPU with most spare capacity.
5818 for_each_cpu(i, sched_group_span(group)) {
5819 /* Bias balancing toward CPUs of our domain */
5821 load = source_load(i, load_idx);
5823 load = target_load(i, load_idx);
5825 runnable_load += load;
5827 avg_load += cfs_rq_load_avg(&cpu_rq(i)->cfs);
5829 spare_cap = capacity_spare_without(i, p);
5831 if (spare_cap > max_spare_cap)
5832 max_spare_cap = spare_cap;
5835 /* Adjust by relative CPU capacity of the group */
5836 avg_load = (avg_load * SCHED_CAPACITY_SCALE) /
5837 group->sgc->capacity;
5838 runnable_load = (runnable_load * SCHED_CAPACITY_SCALE) /
5839 group->sgc->capacity;
5842 this_runnable_load = runnable_load;
5843 this_avg_load = avg_load;
5844 this_spare = max_spare_cap;
5846 if (min_runnable_load > (runnable_load + imbalance)) {
5848 * The runnable load is significantly smaller
5849 * so we can pick this new CPU:
5851 min_runnable_load = runnable_load;
5852 min_avg_load = avg_load;
5854 } else if ((runnable_load < (min_runnable_load + imbalance)) &&
5855 (100*min_avg_load > imbalance_scale*avg_load)) {
5857 * The runnable loads are close so take the
5858 * blocked load into account through avg_load:
5860 min_avg_load = avg_load;
5864 if (most_spare < max_spare_cap) {
5865 most_spare = max_spare_cap;
5866 most_spare_sg = group;
5869 } while (group = group->next, group != sd->groups);
5872 * The cross-over point between using spare capacity or least load
5873 * is too conservative for high utilization tasks on partially
5874 * utilized systems if we require spare_capacity > task_util(p),
5875 * so we allow for some task stuffing by using
5876 * spare_capacity > task_util(p)/2.
5878 * Spare capacity can't be used for fork because the utilization has
5879 * not been set yet, we must first select a rq to compute the initial
5882 if (sd_flag & SD_BALANCE_FORK)
5885 if (this_spare > task_util(p) / 2 &&
5886 imbalance_scale*this_spare > 100*most_spare)
5889 if (most_spare > task_util(p) / 2)
5890 return most_spare_sg;
5897 * When comparing groups across NUMA domains, it's possible for the
5898 * local domain to be very lightly loaded relative to the remote
5899 * domains but "imbalance" skews the comparison making remote CPUs
5900 * look much more favourable. When considering cross-domain, add
5901 * imbalance to the runnable load on the remote node and consider
5904 if ((sd->flags & SD_NUMA) &&
5905 min_runnable_load + imbalance >= this_runnable_load)
5908 if (min_runnable_load > (this_runnable_load + imbalance))
5911 if ((this_runnable_load < (min_runnable_load + imbalance)) &&
5912 (100*this_avg_load < imbalance_scale*min_avg_load))
5919 * find_idlest_group_cpu - find the idlest CPU among the CPUs in the group.
5922 find_idlest_group_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
5924 unsigned long load, min_load = ULONG_MAX;
5925 unsigned int min_exit_latency = UINT_MAX;
5926 u64 latest_idle_timestamp = 0;
5927 int least_loaded_cpu = this_cpu;
5928 int shallowest_idle_cpu = -1;
5931 /* Check if we have any choice: */
5932 if (group->group_weight == 1)
5933 return cpumask_first(sched_group_span(group));
5935 /* Traverse only the allowed CPUs */
5936 for_each_cpu_and(i, sched_group_span(group), &p->cpus_allowed) {
5937 if (available_idle_cpu(i)) {
5938 struct rq *rq = cpu_rq(i);
5939 struct cpuidle_state *idle = idle_get_state(rq);
5940 if (idle && idle->exit_latency < min_exit_latency) {
5942 * We give priority to a CPU whose idle state
5943 * has the smallest exit latency irrespective
5944 * of any idle timestamp.
5946 min_exit_latency = idle->exit_latency;
5947 latest_idle_timestamp = rq->idle_stamp;
5948 shallowest_idle_cpu = i;
5949 } else if ((!idle || idle->exit_latency == min_exit_latency) &&
5950 rq->idle_stamp > latest_idle_timestamp) {
5952 * If equal or no active idle state, then
5953 * the most recently idled CPU might have
5956 latest_idle_timestamp = rq->idle_stamp;
5957 shallowest_idle_cpu = i;
5959 } else if (shallowest_idle_cpu == -1) {
5960 load = weighted_cpuload(cpu_rq(i));
5961 if (load < min_load) {
5963 least_loaded_cpu = i;
5968 return shallowest_idle_cpu != -1 ? shallowest_idle_cpu : least_loaded_cpu;
5971 static inline int find_idlest_cpu(struct sched_domain *sd, struct task_struct *p,
5972 int cpu, int prev_cpu, int sd_flag)
5976 if (!cpumask_intersects(sched_domain_span(sd), &p->cpus_allowed))
5980 * We need task's util for capacity_spare_without, sync it up to
5981 * prev_cpu's last_update_time.
5983 if (!(sd_flag & SD_BALANCE_FORK))
5984 sync_entity_load_avg(&p->se);
5987 struct sched_group *group;
5988 struct sched_domain *tmp;
5991 if (!(sd->flags & sd_flag)) {
5996 group = find_idlest_group(sd, p, cpu, sd_flag);
6002 new_cpu = find_idlest_group_cpu(group, p, cpu);
6003 if (new_cpu == cpu) {
6004 /* Now try balancing at a lower domain level of 'cpu': */
6009 /* Now try balancing at a lower domain level of 'new_cpu': */
6011 weight = sd->span_weight;
6013 for_each_domain(cpu, tmp) {
6014 if (weight <= tmp->span_weight)
6016 if (tmp->flags & sd_flag)
6024 #ifdef CONFIG_SCHED_SMT
6025 DEFINE_STATIC_KEY_FALSE(sched_smt_present);
6026 EXPORT_SYMBOL_GPL(sched_smt_present);
6028 static inline void set_idle_cores(int cpu, int val)
6030 struct sched_domain_shared *sds;
6032 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
6034 WRITE_ONCE(sds->has_idle_cores, val);
6037 static inline bool test_idle_cores(int cpu, bool def)
6039 struct sched_domain_shared *sds;
6041 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
6043 return READ_ONCE(sds->has_idle_cores);
6049 * Scans the local SMT mask to see if the entire core is idle, and records this
6050 * information in sd_llc_shared->has_idle_cores.
6052 * Since SMT siblings share all cache levels, inspecting this limited remote
6053 * state should be fairly cheap.
6055 void __update_idle_core(struct rq *rq)
6057 int core = cpu_of(rq);
6061 if (test_idle_cores(core, true))
6064 for_each_cpu(cpu, cpu_smt_mask(core)) {
6068 if (!available_idle_cpu(cpu))
6072 set_idle_cores(core, 1);
6078 * Scan the entire LLC domain for idle cores; this dynamically switches off if
6079 * there are no idle cores left in the system; tracked through
6080 * sd_llc->shared->has_idle_cores and enabled through update_idle_core() above.
6082 static int select_idle_core(struct task_struct *p, struct sched_domain *sd, int target)
6084 struct cpumask *cpus = this_cpu_cpumask_var_ptr(select_idle_mask);
6087 if (!static_branch_likely(&sched_smt_present))
6090 if (!test_idle_cores(target, false))
6093 cpumask_and(cpus, sched_domain_span(sd), &p->cpus_allowed);
6095 for_each_cpu_wrap(core, cpus, target) {
6098 for_each_cpu(cpu, cpu_smt_mask(core)) {
6099 cpumask_clear_cpu(cpu, cpus);
6100 if (!available_idle_cpu(cpu))
6109 * Failed to find an idle core; stop looking for one.
6111 set_idle_cores(target, 0);
6117 * Scan the local SMT mask for idle CPUs.
6119 static int select_idle_smt(struct task_struct *p, struct sched_domain *sd, int target)
6123 if (!static_branch_likely(&sched_smt_present))
6126 for_each_cpu(cpu, cpu_smt_mask(target)) {
6127 if (!cpumask_test_cpu(cpu, &p->cpus_allowed))
6129 if (available_idle_cpu(cpu))
6136 #else /* CONFIG_SCHED_SMT */
6138 static inline int select_idle_core(struct task_struct *p, struct sched_domain *sd, int target)
6143 static inline int select_idle_smt(struct task_struct *p, struct sched_domain *sd, int target)
6148 #endif /* CONFIG_SCHED_SMT */
6151 * Scan the LLC domain for idle CPUs; this is dynamically regulated by
6152 * comparing the average scan cost (tracked in sd->avg_scan_cost) against the
6153 * average idle time for this rq (as found in rq->avg_idle).
6155 static int select_idle_cpu(struct task_struct *p, struct sched_domain *sd, int target)
6157 struct cpumask *cpus = this_cpu_cpumask_var_ptr(select_idle_mask);
6158 struct sched_domain *this_sd;
6159 u64 avg_cost, avg_idle;
6162 int cpu, nr = INT_MAX;
6164 this_sd = rcu_dereference(*this_cpu_ptr(&sd_llc));
6169 * Due to large variance we need a large fuzz factor; hackbench in
6170 * particularly is sensitive here.
6172 avg_idle = this_rq()->avg_idle / 512;
6173 avg_cost = this_sd->avg_scan_cost + 1;
6175 if (sched_feat(SIS_AVG_CPU) && avg_idle < avg_cost)
6178 if (sched_feat(SIS_PROP)) {
6179 u64 span_avg = sd->span_weight * avg_idle;
6180 if (span_avg > 4*avg_cost)
6181 nr = div_u64(span_avg, avg_cost);
6186 time = local_clock();
6188 cpumask_and(cpus, sched_domain_span(sd), &p->cpus_allowed);
6190 for_each_cpu_wrap(cpu, cpus, target) {
6193 if (available_idle_cpu(cpu))
6197 time = local_clock() - time;
6198 cost = this_sd->avg_scan_cost;
6199 delta = (s64)(time - cost) / 8;
6200 this_sd->avg_scan_cost += delta;
6206 * Try and locate an idle core/thread in the LLC cache domain.
6208 static int select_idle_sibling(struct task_struct *p, int prev, int target)
6210 struct sched_domain *sd;
6211 int i, recent_used_cpu;
6213 if (available_idle_cpu(target))
6217 * If the previous CPU is cache affine and idle, don't be stupid:
6219 if (prev != target && cpus_share_cache(prev, target) && available_idle_cpu(prev))
6222 /* Check a recently used CPU as a potential idle candidate: */
6223 recent_used_cpu = p->recent_used_cpu;
6224 if (recent_used_cpu != prev &&
6225 recent_used_cpu != target &&
6226 cpus_share_cache(recent_used_cpu, target) &&
6227 available_idle_cpu(recent_used_cpu) &&
6228 cpumask_test_cpu(p->recent_used_cpu, &p->cpus_allowed)) {
6230 * Replace recent_used_cpu with prev as it is a potential
6231 * candidate for the next wake:
6233 p->recent_used_cpu = prev;
6234 return recent_used_cpu;
6237 sd = rcu_dereference(per_cpu(sd_llc, target));
6241 i = select_idle_core(p, sd, target);
6242 if ((unsigned)i < nr_cpumask_bits)
6245 i = select_idle_cpu(p, sd, target);
6246 if ((unsigned)i < nr_cpumask_bits)
6249 i = select_idle_smt(p, sd, target);
6250 if ((unsigned)i < nr_cpumask_bits)
6257 * Amount of capacity of a CPU that is (estimated to be) used by CFS tasks
6258 * @cpu: the CPU to get the utilization of
6260 * The unit of the return value must be the one of capacity so we can compare
6261 * the utilization with the capacity of the CPU that is available for CFS task
6262 * (ie cpu_capacity).
6264 * cfs_rq.avg.util_avg is the sum of running time of runnable tasks plus the
6265 * recent utilization of currently non-runnable tasks on a CPU. It represents
6266 * the amount of utilization of a CPU in the range [0..capacity_orig] where
6267 * capacity_orig is the cpu_capacity available at the highest frequency
6268 * (arch_scale_freq_capacity()).
6269 * The utilization of a CPU converges towards a sum equal to or less than the
6270 * current capacity (capacity_curr <= capacity_orig) of the CPU because it is
6271 * the running time on this CPU scaled by capacity_curr.
6273 * The estimated utilization of a CPU is defined to be the maximum between its
6274 * cfs_rq.avg.util_avg and the sum of the estimated utilization of the tasks
6275 * currently RUNNABLE on that CPU.
6276 * This allows to properly represent the expected utilization of a CPU which
6277 * has just got a big task running since a long sleep period. At the same time
6278 * however it preserves the benefits of the "blocked utilization" in
6279 * describing the potential for other tasks waking up on the same CPU.
6281 * Nevertheless, cfs_rq.avg.util_avg can be higher than capacity_curr or even
6282 * higher than capacity_orig because of unfortunate rounding in
6283 * cfs.avg.util_avg or just after migrating tasks and new task wakeups until
6284 * the average stabilizes with the new running time. We need to check that the
6285 * utilization stays within the range of [0..capacity_orig] and cap it if
6286 * necessary. Without utilization capping, a group could be seen as overloaded
6287 * (CPU0 utilization at 121% + CPU1 utilization at 80%) whereas CPU1 has 20% of
6288 * available capacity. We allow utilization to overshoot capacity_curr (but not
6289 * capacity_orig) as it useful for predicting the capacity required after task
6290 * migrations (scheduler-driven DVFS).
6292 * Return: the (estimated) utilization for the specified CPU
6294 static inline unsigned long cpu_util(int cpu)
6296 struct cfs_rq *cfs_rq;
6299 cfs_rq = &cpu_rq(cpu)->cfs;
6300 util = READ_ONCE(cfs_rq->avg.util_avg);
6302 if (sched_feat(UTIL_EST))
6303 util = max(util, READ_ONCE(cfs_rq->avg.util_est.enqueued));
6305 return min_t(unsigned long, util, capacity_orig_of(cpu));
6309 * cpu_util_without: compute cpu utilization without any contributions from *p
6310 * @cpu: the CPU which utilization is requested
6311 * @p: the task which utilization should be discounted
6313 * The utilization of a CPU is defined by the utilization of tasks currently
6314 * enqueued on that CPU as well as tasks which are currently sleeping after an
6315 * execution on that CPU.
6317 * This method returns the utilization of the specified CPU by discounting the
6318 * utilization of the specified task, whenever the task is currently
6319 * contributing to the CPU utilization.
6321 static unsigned long cpu_util_without(int cpu, struct task_struct *p)
6323 struct cfs_rq *cfs_rq;
6326 /* Task has no contribution or is new */
6327 if (cpu != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time))
6328 return cpu_util(cpu);
6330 cfs_rq = &cpu_rq(cpu)->cfs;
6331 util = READ_ONCE(cfs_rq->avg.util_avg);
6333 /* Discount task's util from CPU's util */
6334 util -= min_t(unsigned int, util, task_util(p));
6339 * a) if *p is the only task sleeping on this CPU, then:
6340 * cpu_util (== task_util) > util_est (== 0)
6341 * and thus we return:
6342 * cpu_util_without = (cpu_util - task_util) = 0
6344 * b) if other tasks are SLEEPING on this CPU, which is now exiting
6346 * cpu_util >= task_util
6347 * cpu_util > util_est (== 0)
6348 * and thus we discount *p's blocked utilization to return:
6349 * cpu_util_without = (cpu_util - task_util) >= 0
6351 * c) if other tasks are RUNNABLE on that CPU and
6352 * util_est > cpu_util
6353 * then we use util_est since it returns a more restrictive
6354 * estimation of the spare capacity on that CPU, by just
6355 * considering the expected utilization of tasks already
6356 * runnable on that CPU.
6358 * Cases a) and b) are covered by the above code, while case c) is
6359 * covered by the following code when estimated utilization is
6362 if (sched_feat(UTIL_EST)) {
6363 unsigned int estimated =
6364 READ_ONCE(cfs_rq->avg.util_est.enqueued);
6367 * Despite the following checks we still have a small window
6368 * for a possible race, when an execl's select_task_rq_fair()
6369 * races with LB's detach_task():
6372 * p->on_rq = TASK_ON_RQ_MIGRATING;
6373 * ---------------------------------- A
6374 * deactivate_task() \
6375 * dequeue_task() + RaceTime
6376 * util_est_dequeue() /
6377 * ---------------------------------- B
6379 * The additional check on "current == p" it's required to
6380 * properly fix the execl regression and it helps in further
6381 * reducing the chances for the above race.
6383 if (unlikely(task_on_rq_queued(p) || current == p)) {
6384 estimated -= min_t(unsigned int, estimated,
6385 (_task_util_est(p) | UTIL_AVG_UNCHANGED));
6387 util = max(util, estimated);
6391 * Utilization (estimated) can exceed the CPU capacity, thus let's
6392 * clamp to the maximum CPU capacity to ensure consistency with
6393 * the cpu_util call.
6395 return min_t(unsigned long, util, capacity_orig_of(cpu));
6399 * Disable WAKE_AFFINE in the case where task @p doesn't fit in the
6400 * capacity of either the waking CPU @cpu or the previous CPU @prev_cpu.
6402 * In that case WAKE_AFFINE doesn't make sense and we'll let
6403 * BALANCE_WAKE sort things out.
6405 static int wake_cap(struct task_struct *p, int cpu, int prev_cpu)
6407 long min_cap, max_cap;
6409 min_cap = min(capacity_orig_of(prev_cpu), capacity_orig_of(cpu));
6410 max_cap = cpu_rq(cpu)->rd->max_cpu_capacity;
6412 /* Minimum capacity is close to max, no need to abort wake_affine */
6413 if (max_cap - min_cap < max_cap >> 3)
6416 /* Bring task utilization in sync with prev_cpu */
6417 sync_entity_load_avg(&p->se);
6419 return min_cap * 1024 < task_util(p) * capacity_margin;
6423 * select_task_rq_fair: Select target runqueue for the waking task in domains
6424 * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE,
6425 * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
6427 * Balances load by selecting the idlest CPU in the idlest group, or under
6428 * certain conditions an idle sibling CPU if the domain has SD_WAKE_AFFINE set.
6430 * Returns the target CPU number.
6432 * preempt must be disabled.
6435 select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags)
6437 struct sched_domain *tmp, *sd = NULL;
6438 int cpu = smp_processor_id();
6439 int new_cpu = prev_cpu;
6440 int want_affine = 0;
6441 int sync = (wake_flags & WF_SYNC) && !(current->flags & PF_EXITING);
6443 if (sd_flag & SD_BALANCE_WAKE) {
6445 want_affine = !wake_wide(p) && !wake_cap(p, cpu, prev_cpu)
6446 && cpumask_test_cpu(cpu, &p->cpus_allowed);
6450 for_each_domain(cpu, tmp) {
6451 if (!(tmp->flags & SD_LOAD_BALANCE))
6455 * If both 'cpu' and 'prev_cpu' are part of this domain,
6456 * cpu is a valid SD_WAKE_AFFINE target.
6458 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
6459 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
6460 if (cpu != prev_cpu)
6461 new_cpu = wake_affine(tmp, p, cpu, prev_cpu, sync);
6463 sd = NULL; /* Prefer wake_affine over balance flags */
6467 if (tmp->flags & sd_flag)
6469 else if (!want_affine)
6475 new_cpu = find_idlest_cpu(sd, p, cpu, prev_cpu, sd_flag);
6476 } else if (sd_flag & SD_BALANCE_WAKE) { /* XXX always ? */
6479 new_cpu = select_idle_sibling(p, prev_cpu, new_cpu);
6482 current->recent_used_cpu = cpu;
6489 static void detach_entity_cfs_rq(struct sched_entity *se);
6492 * Called immediately before a task is migrated to a new CPU; task_cpu(p) and
6493 * cfs_rq_of(p) references at time of call are still valid and identify the
6494 * previous CPU. The caller guarantees p->pi_lock or task_rq(p)->lock is held.
6496 static void migrate_task_rq_fair(struct task_struct *p, int new_cpu)
6499 * As blocked tasks retain absolute vruntime the migration needs to
6500 * deal with this by subtracting the old and adding the new
6501 * min_vruntime -- the latter is done by enqueue_entity() when placing
6502 * the task on the new runqueue.
6504 if (p->state == TASK_WAKING) {
6505 struct sched_entity *se = &p->se;
6506 struct cfs_rq *cfs_rq = cfs_rq_of(se);
6509 #ifndef CONFIG_64BIT
6510 u64 min_vruntime_copy;
6513 min_vruntime_copy = cfs_rq->min_vruntime_copy;
6515 min_vruntime = cfs_rq->min_vruntime;
6516 } while (min_vruntime != min_vruntime_copy);
6518 min_vruntime = cfs_rq->min_vruntime;
6521 se->vruntime -= min_vruntime;
6524 if (p->on_rq == TASK_ON_RQ_MIGRATING) {
6526 * In case of TASK_ON_RQ_MIGRATING we in fact hold the 'old'
6527 * rq->lock and can modify state directly.
6529 lockdep_assert_held(&task_rq(p)->lock);
6530 detach_entity_cfs_rq(&p->se);
6534 * We are supposed to update the task to "current" time, then
6535 * its up to date and ready to go to new CPU/cfs_rq. But we
6536 * have difficulty in getting what current time is, so simply
6537 * throw away the out-of-date time. This will result in the
6538 * wakee task is less decayed, but giving the wakee more load
6541 remove_entity_load_avg(&p->se);
6544 /* Tell new CPU we are migrated */
6545 p->se.avg.last_update_time = 0;
6547 /* We have migrated, no longer consider this task hot */
6548 p->se.exec_start = 0;
6550 update_scan_period(p, new_cpu);
6553 static void task_dead_fair(struct task_struct *p)
6555 remove_entity_load_avg(&p->se);
6557 #endif /* CONFIG_SMP */
6559 static unsigned long wakeup_gran(struct sched_entity *se)
6561 unsigned long gran = sysctl_sched_wakeup_granularity;
6564 * Since its curr running now, convert the gran from real-time
6565 * to virtual-time in his units.
6567 * By using 'se' instead of 'curr' we penalize light tasks, so
6568 * they get preempted easier. That is, if 'se' < 'curr' then
6569 * the resulting gran will be larger, therefore penalizing the
6570 * lighter, if otoh 'se' > 'curr' then the resulting gran will
6571 * be smaller, again penalizing the lighter task.
6573 * This is especially important for buddies when the leftmost
6574 * task is higher priority than the buddy.
6576 return calc_delta_fair(gran, se);
6580 * Should 'se' preempt 'curr'.
6594 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
6596 s64 gran, vdiff = curr->vruntime - se->vruntime;
6601 gran = wakeup_gran(se);
6608 static void set_last_buddy(struct sched_entity *se)
6610 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
6613 for_each_sched_entity(se) {
6614 if (SCHED_WARN_ON(!se->on_rq))
6616 cfs_rq_of(se)->last = se;
6620 static void set_next_buddy(struct sched_entity *se)
6622 if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
6625 for_each_sched_entity(se) {
6626 if (SCHED_WARN_ON(!se->on_rq))
6628 cfs_rq_of(se)->next = se;
6632 static void set_skip_buddy(struct sched_entity *se)
6634 for_each_sched_entity(se)
6635 cfs_rq_of(se)->skip = se;
6639 * Preempt the current task with a newly woken task if needed:
6641 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
6643 struct task_struct *curr = rq->curr;
6644 struct sched_entity *se = &curr->se, *pse = &p->se;
6645 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
6646 int scale = cfs_rq->nr_running >= sched_nr_latency;
6647 int next_buddy_marked = 0;
6649 if (unlikely(se == pse))
6653 * This is possible from callers such as attach_tasks(), in which we
6654 * unconditionally check_prempt_curr() after an enqueue (which may have
6655 * lead to a throttle). This both saves work and prevents false
6656 * next-buddy nomination below.
6658 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
6661 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
6662 set_next_buddy(pse);
6663 next_buddy_marked = 1;
6667 * We can come here with TIF_NEED_RESCHED already set from new task
6670 * Note: this also catches the edge-case of curr being in a throttled
6671 * group (e.g. via set_curr_task), since update_curr() (in the
6672 * enqueue of curr) will have resulted in resched being set. This
6673 * prevents us from potentially nominating it as a false LAST_BUDDY
6676 if (test_tsk_need_resched(curr))
6679 /* Idle tasks are by definition preempted by non-idle tasks. */
6680 if (unlikely(curr->policy == SCHED_IDLE) &&
6681 likely(p->policy != SCHED_IDLE))
6685 * Batch and idle tasks do not preempt non-idle tasks (their preemption
6686 * is driven by the tick):
6688 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
6691 find_matching_se(&se, &pse);
6692 update_curr(cfs_rq_of(se));
6694 if (wakeup_preempt_entity(se, pse) == 1) {
6696 * Bias pick_next to pick the sched entity that is
6697 * triggering this preemption.
6699 if (!next_buddy_marked)
6700 set_next_buddy(pse);
6709 * Only set the backward buddy when the current task is still
6710 * on the rq. This can happen when a wakeup gets interleaved
6711 * with schedule on the ->pre_schedule() or idle_balance()
6712 * point, either of which can * drop the rq lock.
6714 * Also, during early boot the idle thread is in the fair class,
6715 * for obvious reasons its a bad idea to schedule back to it.
6717 if (unlikely(!se->on_rq || curr == rq->idle))
6720 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
6724 static struct task_struct *
6725 pick_next_task_fair(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
6727 struct cfs_rq *cfs_rq = &rq->cfs;
6728 struct sched_entity *se;
6729 struct task_struct *p;
6733 if (!cfs_rq->nr_running)
6736 #ifdef CONFIG_FAIR_GROUP_SCHED
6737 if (prev->sched_class != &fair_sched_class)
6741 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
6742 * likely that a next task is from the same cgroup as the current.
6744 * Therefore attempt to avoid putting and setting the entire cgroup
6745 * hierarchy, only change the part that actually changes.
6749 struct sched_entity *curr = cfs_rq->curr;
6752 * Since we got here without doing put_prev_entity() we also
6753 * have to consider cfs_rq->curr. If it is still a runnable
6754 * entity, update_curr() will update its vruntime, otherwise
6755 * forget we've ever seen it.
6759 update_curr(cfs_rq);
6764 * This call to check_cfs_rq_runtime() will do the
6765 * throttle and dequeue its entity in the parent(s).
6766 * Therefore the nr_running test will indeed
6769 if (unlikely(check_cfs_rq_runtime(cfs_rq))) {
6772 if (!cfs_rq->nr_running)
6779 se = pick_next_entity(cfs_rq, curr);
6780 cfs_rq = group_cfs_rq(se);
6786 * Since we haven't yet done put_prev_entity and if the selected task
6787 * is a different task than we started out with, try and touch the
6788 * least amount of cfs_rqs.
6791 struct sched_entity *pse = &prev->se;
6793 while (!(cfs_rq = is_same_group(se, pse))) {
6794 int se_depth = se->depth;
6795 int pse_depth = pse->depth;
6797 if (se_depth <= pse_depth) {
6798 put_prev_entity(cfs_rq_of(pse), pse);
6799 pse = parent_entity(pse);
6801 if (se_depth >= pse_depth) {
6802 set_next_entity(cfs_rq_of(se), se);
6803 se = parent_entity(se);
6807 put_prev_entity(cfs_rq, pse);
6808 set_next_entity(cfs_rq, se);
6815 put_prev_task(rq, prev);
6818 se = pick_next_entity(cfs_rq, NULL);
6819 set_next_entity(cfs_rq, se);
6820 cfs_rq = group_cfs_rq(se);
6825 done: __maybe_unused;
6828 * Move the next running task to the front of
6829 * the list, so our cfs_tasks list becomes MRU
6832 list_move(&p->se.group_node, &rq->cfs_tasks);
6835 if (hrtick_enabled(rq))
6836 hrtick_start_fair(rq, p);
6841 new_tasks = idle_balance(rq, rf);
6844 * Because idle_balance() releases (and re-acquires) rq->lock, it is
6845 * possible for any higher priority task to appear. In that case we
6846 * must re-start the pick_next_entity() loop.
6858 * Account for a descheduled task:
6860 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
6862 struct sched_entity *se = &prev->se;
6863 struct cfs_rq *cfs_rq;
6865 for_each_sched_entity(se) {
6866 cfs_rq = cfs_rq_of(se);
6867 put_prev_entity(cfs_rq, se);
6872 * sched_yield() is very simple
6874 * The magic of dealing with the ->skip buddy is in pick_next_entity.
6876 static void yield_task_fair(struct rq *rq)
6878 struct task_struct *curr = rq->curr;
6879 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
6880 struct sched_entity *se = &curr->se;
6883 * Are we the only task in the tree?
6885 if (unlikely(rq->nr_running == 1))
6888 clear_buddies(cfs_rq, se);
6890 if (curr->policy != SCHED_BATCH) {
6891 update_rq_clock(rq);
6893 * Update run-time statistics of the 'current'.
6895 update_curr(cfs_rq);
6897 * Tell update_rq_clock() that we've just updated,
6898 * so we don't do microscopic update in schedule()
6899 * and double the fastpath cost.
6901 rq_clock_skip_update(rq);
6907 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
6909 struct sched_entity *se = &p->se;
6911 /* throttled hierarchies are not runnable */
6912 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
6915 /* Tell the scheduler that we'd really like pse to run next. */
6918 yield_task_fair(rq);
6924 /**************************************************
6925 * Fair scheduling class load-balancing methods.
6929 * The purpose of load-balancing is to achieve the same basic fairness the
6930 * per-CPU scheduler provides, namely provide a proportional amount of compute
6931 * time to each task. This is expressed in the following equation:
6933 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
6935 * Where W_i,n is the n-th weight average for CPU i. The instantaneous weight
6936 * W_i,0 is defined as:
6938 * W_i,0 = \Sum_j w_i,j (2)
6940 * Where w_i,j is the weight of the j-th runnable task on CPU i. This weight
6941 * is derived from the nice value as per sched_prio_to_weight[].
6943 * The weight average is an exponential decay average of the instantaneous
6946 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
6948 * C_i is the compute capacity of CPU i, typically it is the
6949 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
6950 * can also include other factors [XXX].
6952 * To achieve this balance we define a measure of imbalance which follows
6953 * directly from (1):
6955 * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4)
6957 * We them move tasks around to minimize the imbalance. In the continuous
6958 * function space it is obvious this converges, in the discrete case we get
6959 * a few fun cases generally called infeasible weight scenarios.
6962 * - infeasible weights;
6963 * - local vs global optima in the discrete case. ]
6968 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
6969 * for all i,j solution, we create a tree of CPUs that follows the hardware
6970 * topology where each level pairs two lower groups (or better). This results
6971 * in O(log n) layers. Furthermore we reduce the number of CPUs going up the
6972 * tree to only the first of the previous level and we decrease the frequency
6973 * of load-balance at each level inv. proportional to the number of CPUs in
6979 * \Sum { --- * --- * 2^i } = O(n) (5)
6981 * `- size of each group
6982 * | | `- number of CPUs doing load-balance
6984 * `- sum over all levels
6986 * Coupled with a limit on how many tasks we can migrate every balance pass,
6987 * this makes (5) the runtime complexity of the balancer.
6989 * An important property here is that each CPU is still (indirectly) connected
6990 * to every other CPU in at most O(log n) steps:
6992 * The adjacency matrix of the resulting graph is given by:
6995 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
6998 * And you'll find that:
7000 * A^(log_2 n)_i,j != 0 for all i,j (7)
7002 * Showing there's indeed a path between every CPU in at most O(log n) steps.
7003 * The task movement gives a factor of O(m), giving a convergence complexity
7006 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
7011 * In order to avoid CPUs going idle while there's still work to do, new idle
7012 * balancing is more aggressive and has the newly idle CPU iterate up the domain
7013 * tree itself instead of relying on other CPUs to bring it work.
7015 * This adds some complexity to both (5) and (8) but it reduces the total idle
7023 * Cgroups make a horror show out of (2), instead of a simple sum we get:
7026 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
7031 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
7033 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on CPU i.
7035 * The big problem is S_k, its a global sum needed to compute a local (W_i)
7038 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
7039 * rewrite all of this once again.]
7042 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
7044 enum fbq_type { regular, remote, all };
7046 #define LBF_ALL_PINNED 0x01
7047 #define LBF_NEED_BREAK 0x02
7048 #define LBF_DST_PINNED 0x04
7049 #define LBF_SOME_PINNED 0x08
7050 #define LBF_NOHZ_STATS 0x10
7051 #define LBF_NOHZ_AGAIN 0x20
7054 struct sched_domain *sd;
7062 struct cpumask *dst_grpmask;
7064 enum cpu_idle_type idle;
7066 /* The set of CPUs under consideration for load-balancing */
7067 struct cpumask *cpus;
7072 unsigned int loop_break;
7073 unsigned int loop_max;
7075 enum fbq_type fbq_type;
7076 struct list_head tasks;
7080 * Is this task likely cache-hot:
7082 static int task_hot(struct task_struct *p, struct lb_env *env)
7086 lockdep_assert_held(&env->src_rq->lock);
7088 if (p->sched_class != &fair_sched_class)
7091 if (unlikely(p->policy == SCHED_IDLE))
7095 * Buddy candidates are cache hot:
7097 if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running &&
7098 (&p->se == cfs_rq_of(&p->se)->next ||
7099 &p->se == cfs_rq_of(&p->se)->last))
7102 if (sysctl_sched_migration_cost == -1)
7104 if (sysctl_sched_migration_cost == 0)
7107 delta = rq_clock_task(env->src_rq) - p->se.exec_start;
7109 return delta < (s64)sysctl_sched_migration_cost;
7112 #ifdef CONFIG_NUMA_BALANCING
7114 * Returns 1, if task migration degrades locality
7115 * Returns 0, if task migration improves locality i.e migration preferred.
7116 * Returns -1, if task migration is not affected by locality.
7118 static int migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
7120 struct numa_group *numa_group = rcu_dereference(p->numa_group);
7121 unsigned long src_weight, dst_weight;
7122 int src_nid, dst_nid, dist;
7124 if (!static_branch_likely(&sched_numa_balancing))
7127 if (!p->numa_faults || !(env->sd->flags & SD_NUMA))
7130 src_nid = cpu_to_node(env->src_cpu);
7131 dst_nid = cpu_to_node(env->dst_cpu);
7133 if (src_nid == dst_nid)
7136 /* Migrating away from the preferred node is always bad. */
7137 if (src_nid == p->numa_preferred_nid) {
7138 if (env->src_rq->nr_running > env->src_rq->nr_preferred_running)
7144 /* Encourage migration to the preferred node. */
7145 if (dst_nid == p->numa_preferred_nid)
7148 /* Leaving a core idle is often worse than degrading locality. */
7149 if (env->idle == CPU_IDLE)
7152 dist = node_distance(src_nid, dst_nid);
7154 src_weight = group_weight(p, src_nid, dist);
7155 dst_weight = group_weight(p, dst_nid, dist);
7157 src_weight = task_weight(p, src_nid, dist);
7158 dst_weight = task_weight(p, dst_nid, dist);
7161 return dst_weight < src_weight;
7165 static inline int migrate_degrades_locality(struct task_struct *p,
7173 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
7176 int can_migrate_task(struct task_struct *p, struct lb_env *env)
7180 lockdep_assert_held(&env->src_rq->lock);
7183 * We do not migrate tasks that are:
7184 * 1) throttled_lb_pair, or
7185 * 2) cannot be migrated to this CPU due to cpus_allowed, or
7186 * 3) running (obviously), or
7187 * 4) are cache-hot on their current CPU.
7189 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
7192 if (!cpumask_test_cpu(env->dst_cpu, &p->cpus_allowed)) {
7195 schedstat_inc(p->se.statistics.nr_failed_migrations_affine);
7197 env->flags |= LBF_SOME_PINNED;
7200 * Remember if this task can be migrated to any other CPU in
7201 * our sched_group. We may want to revisit it if we couldn't
7202 * meet load balance goals by pulling other tasks on src_cpu.
7204 * Avoid computing new_dst_cpu for NEWLY_IDLE or if we have
7205 * already computed one in current iteration.
7207 if (env->idle == CPU_NEWLY_IDLE || (env->flags & LBF_DST_PINNED))
7210 /* Prevent to re-select dst_cpu via env's CPUs: */
7211 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
7212 if (cpumask_test_cpu(cpu, &p->cpus_allowed)) {
7213 env->flags |= LBF_DST_PINNED;
7214 env->new_dst_cpu = cpu;
7222 /* Record that we found atleast one task that could run on dst_cpu */
7223 env->flags &= ~LBF_ALL_PINNED;
7225 if (task_running(env->src_rq, p)) {
7226 schedstat_inc(p->se.statistics.nr_failed_migrations_running);
7231 * Aggressive migration if:
7232 * 1) destination numa is preferred
7233 * 2) task is cache cold, or
7234 * 3) too many balance attempts have failed.
7236 tsk_cache_hot = migrate_degrades_locality(p, env);
7237 if (tsk_cache_hot == -1)
7238 tsk_cache_hot = task_hot(p, env);
7240 if (tsk_cache_hot <= 0 ||
7241 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
7242 if (tsk_cache_hot == 1) {
7243 schedstat_inc(env->sd->lb_hot_gained[env->idle]);
7244 schedstat_inc(p->se.statistics.nr_forced_migrations);
7249 schedstat_inc(p->se.statistics.nr_failed_migrations_hot);
7254 * detach_task() -- detach the task for the migration specified in env
7256 static void detach_task(struct task_struct *p, struct lb_env *env)
7258 lockdep_assert_held(&env->src_rq->lock);
7260 p->on_rq = TASK_ON_RQ_MIGRATING;
7261 deactivate_task(env->src_rq, p, DEQUEUE_NOCLOCK);
7262 set_task_cpu(p, env->dst_cpu);
7266 * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
7267 * part of active balancing operations within "domain".
7269 * Returns a task if successful and NULL otherwise.
7271 static struct task_struct *detach_one_task(struct lb_env *env)
7273 struct task_struct *p;
7275 lockdep_assert_held(&env->src_rq->lock);
7277 list_for_each_entry_reverse(p,
7278 &env->src_rq->cfs_tasks, se.group_node) {
7279 if (!can_migrate_task(p, env))
7282 detach_task(p, env);
7285 * Right now, this is only the second place where
7286 * lb_gained[env->idle] is updated (other is detach_tasks)
7287 * so we can safely collect stats here rather than
7288 * inside detach_tasks().
7290 schedstat_inc(env->sd->lb_gained[env->idle]);
7296 static const unsigned int sched_nr_migrate_break = 32;
7299 * detach_tasks() -- tries to detach up to imbalance weighted load from
7300 * busiest_rq, as part of a balancing operation within domain "sd".
7302 * Returns number of detached tasks if successful and 0 otherwise.
7304 static int detach_tasks(struct lb_env *env)
7306 struct list_head *tasks = &env->src_rq->cfs_tasks;
7307 struct task_struct *p;
7311 lockdep_assert_held(&env->src_rq->lock);
7313 if (env->imbalance <= 0)
7316 while (!list_empty(tasks)) {
7318 * We don't want to steal all, otherwise we may be treated likewise,
7319 * which could at worst lead to a livelock crash.
7321 if (env->idle != CPU_NOT_IDLE && env->src_rq->nr_running <= 1)
7324 p = list_last_entry(tasks, struct task_struct, se.group_node);
7327 /* We've more or less seen every task there is, call it quits */
7328 if (env->loop > env->loop_max)
7331 /* take a breather every nr_migrate tasks */
7332 if (env->loop > env->loop_break) {
7333 env->loop_break += sched_nr_migrate_break;
7334 env->flags |= LBF_NEED_BREAK;
7338 if (!can_migrate_task(p, env))
7342 * Depending of the number of CPUs and tasks and the
7343 * cgroup hierarchy, task_h_load() can return a null
7344 * value. Make sure that env->imbalance decreases
7345 * otherwise detach_tasks() will stop only after
7346 * detaching up to loop_max tasks.
7348 load = max_t(unsigned long, task_h_load(p), 1);
7351 if (sched_feat(LB_MIN) && load < 16 && !env->sd->nr_balance_failed)
7354 if ((load / 2) > env->imbalance)
7357 detach_task(p, env);
7358 list_add(&p->se.group_node, &env->tasks);
7361 env->imbalance -= load;
7363 #ifdef CONFIG_PREEMPT
7365 * NEWIDLE balancing is a source of latency, so preemptible
7366 * kernels will stop after the first task is detached to minimize
7367 * the critical section.
7369 if (env->idle == CPU_NEWLY_IDLE)
7374 * We only want to steal up to the prescribed amount of
7377 if (env->imbalance <= 0)
7382 list_move(&p->se.group_node, tasks);
7386 * Right now, this is one of only two places we collect this stat
7387 * so we can safely collect detach_one_task() stats here rather
7388 * than inside detach_one_task().
7390 schedstat_add(env->sd->lb_gained[env->idle], detached);
7396 * attach_task() -- attach the task detached by detach_task() to its new rq.
7398 static void attach_task(struct rq *rq, struct task_struct *p)
7400 lockdep_assert_held(&rq->lock);
7402 BUG_ON(task_rq(p) != rq);
7403 activate_task(rq, p, ENQUEUE_NOCLOCK);
7404 p->on_rq = TASK_ON_RQ_QUEUED;
7405 check_preempt_curr(rq, p, 0);
7409 * attach_one_task() -- attaches the task returned from detach_one_task() to
7412 static void attach_one_task(struct rq *rq, struct task_struct *p)
7417 update_rq_clock(rq);
7423 * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
7426 static void attach_tasks(struct lb_env *env)
7428 struct list_head *tasks = &env->tasks;
7429 struct task_struct *p;
7432 rq_lock(env->dst_rq, &rf);
7433 update_rq_clock(env->dst_rq);
7435 while (!list_empty(tasks)) {
7436 p = list_first_entry(tasks, struct task_struct, se.group_node);
7437 list_del_init(&p->se.group_node);
7439 attach_task(env->dst_rq, p);
7442 rq_unlock(env->dst_rq, &rf);
7445 static inline bool cfs_rq_has_blocked(struct cfs_rq *cfs_rq)
7447 if (cfs_rq->avg.load_avg)
7450 if (cfs_rq->avg.util_avg)
7456 static inline bool others_have_blocked(struct rq *rq)
7458 if (READ_ONCE(rq->avg_rt.util_avg))
7461 if (READ_ONCE(rq->avg_dl.util_avg))
7464 #ifdef CONFIG_HAVE_SCHED_AVG_IRQ
7465 if (READ_ONCE(rq->avg_irq.util_avg))
7472 #ifdef CONFIG_FAIR_GROUP_SCHED
7474 static inline bool cfs_rq_is_decayed(struct cfs_rq *cfs_rq)
7476 if (cfs_rq->load.weight)
7479 if (cfs_rq->avg.load_sum)
7482 if (cfs_rq->avg.util_sum)
7485 if (cfs_rq->avg.runnable_load_sum)
7491 static void update_blocked_averages(int cpu)
7493 struct rq *rq = cpu_rq(cpu);
7494 struct cfs_rq *cfs_rq, *pos;
7495 const struct sched_class *curr_class;
7499 rq_lock_irqsave(rq, &rf);
7500 update_rq_clock(rq);
7503 * Iterates the task_group tree in a bottom up fashion, see
7504 * list_add_leaf_cfs_rq() for details.
7506 for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) {
7507 struct sched_entity *se;
7509 if (update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq))
7510 update_tg_load_avg(cfs_rq, 0);
7512 /* Propagate pending load changes to the parent, if any: */
7513 se = cfs_rq->tg->se[cpu];
7514 if (se && !skip_blocked_update(se))
7515 update_load_avg(cfs_rq_of(se), se, UPDATE_TG);
7518 * There can be a lot of idle CPU cgroups. Don't let fully
7519 * decayed cfs_rqs linger on the list.
7521 if (cfs_rq_is_decayed(cfs_rq))
7522 list_del_leaf_cfs_rq(cfs_rq);
7524 /* Don't need periodic decay once load/util_avg are null */
7525 if (cfs_rq_has_blocked(cfs_rq))
7529 curr_class = rq->curr->sched_class;
7530 update_rt_rq_load_avg(rq_clock_task(rq), rq, curr_class == &rt_sched_class);
7531 update_dl_rq_load_avg(rq_clock_task(rq), rq, curr_class == &dl_sched_class);
7532 update_irq_load_avg(rq, 0);
7533 /* Don't need periodic decay once load/util_avg are null */
7534 if (others_have_blocked(rq))
7537 #ifdef CONFIG_NO_HZ_COMMON
7538 rq->last_blocked_load_update_tick = jiffies;
7540 rq->has_blocked_load = 0;
7542 rq_unlock_irqrestore(rq, &rf);
7546 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
7547 * This needs to be done in a top-down fashion because the load of a child
7548 * group is a fraction of its parents load.
7550 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
7552 struct rq *rq = rq_of(cfs_rq);
7553 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
7554 unsigned long now = jiffies;
7557 if (cfs_rq->last_h_load_update == now)
7560 WRITE_ONCE(cfs_rq->h_load_next, NULL);
7561 for_each_sched_entity(se) {
7562 cfs_rq = cfs_rq_of(se);
7563 WRITE_ONCE(cfs_rq->h_load_next, se);
7564 if (cfs_rq->last_h_load_update == now)
7569 cfs_rq->h_load = cfs_rq_load_avg(cfs_rq);
7570 cfs_rq->last_h_load_update = now;
7573 while ((se = READ_ONCE(cfs_rq->h_load_next)) != NULL) {
7574 load = cfs_rq->h_load;
7575 load = div64_ul(load * se->avg.load_avg,
7576 cfs_rq_load_avg(cfs_rq) + 1);
7577 cfs_rq = group_cfs_rq(se);
7578 cfs_rq->h_load = load;
7579 cfs_rq->last_h_load_update = now;
7583 static unsigned long task_h_load(struct task_struct *p)
7585 struct cfs_rq *cfs_rq = task_cfs_rq(p);
7587 update_cfs_rq_h_load(cfs_rq);
7588 return div64_ul(p->se.avg.load_avg * cfs_rq->h_load,
7589 cfs_rq_load_avg(cfs_rq) + 1);
7592 static inline void update_blocked_averages(int cpu)
7594 struct rq *rq = cpu_rq(cpu);
7595 struct cfs_rq *cfs_rq = &rq->cfs;
7596 const struct sched_class *curr_class;
7599 rq_lock_irqsave(rq, &rf);
7600 update_rq_clock(rq);
7601 update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq);
7603 curr_class = rq->curr->sched_class;
7604 update_rt_rq_load_avg(rq_clock_task(rq), rq, curr_class == &rt_sched_class);
7605 update_dl_rq_load_avg(rq_clock_task(rq), rq, curr_class == &dl_sched_class);
7606 update_irq_load_avg(rq, 0);
7607 #ifdef CONFIG_NO_HZ_COMMON
7608 rq->last_blocked_load_update_tick = jiffies;
7609 if (!cfs_rq_has_blocked(cfs_rq) && !others_have_blocked(rq))
7610 rq->has_blocked_load = 0;
7612 rq_unlock_irqrestore(rq, &rf);
7615 static unsigned long task_h_load(struct task_struct *p)
7617 return p->se.avg.load_avg;
7621 /********** Helpers for find_busiest_group ************************/
7630 * sg_lb_stats - stats of a sched_group required for load_balancing
7632 struct sg_lb_stats {
7633 unsigned long avg_load; /*Avg load across the CPUs of the group */
7634 unsigned long group_load; /* Total load over the CPUs of the group */
7635 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
7636 unsigned long load_per_task;
7637 unsigned long group_capacity;
7638 unsigned long group_util; /* Total utilization of the group */
7639 unsigned int sum_nr_running; /* Nr tasks running in the group */
7640 unsigned int idle_cpus;
7641 unsigned int group_weight;
7642 enum group_type group_type;
7643 int group_no_capacity;
7644 #ifdef CONFIG_NUMA_BALANCING
7645 unsigned int nr_numa_running;
7646 unsigned int nr_preferred_running;
7651 * sd_lb_stats - Structure to store the statistics of a sched_domain
7652 * during load balancing.
7654 struct sd_lb_stats {
7655 struct sched_group *busiest; /* Busiest group in this sd */
7656 struct sched_group *local; /* Local group in this sd */
7657 unsigned long total_running;
7658 unsigned long total_load; /* Total load of all groups in sd */
7659 unsigned long total_capacity; /* Total capacity of all groups in sd */
7660 unsigned long avg_load; /* Average load across all groups in sd */
7662 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
7663 struct sg_lb_stats local_stat; /* Statistics of the local group */
7666 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
7669 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
7670 * local_stat because update_sg_lb_stats() does a full clear/assignment.
7671 * We must however clear busiest_stat::avg_load because
7672 * update_sd_pick_busiest() reads this before assignment.
7674 *sds = (struct sd_lb_stats){
7677 .total_running = 0UL,
7679 .total_capacity = 0UL,
7682 .sum_nr_running = 0,
7683 .group_type = group_other,
7689 * get_sd_load_idx - Obtain the load index for a given sched domain.
7690 * @sd: The sched_domain whose load_idx is to be obtained.
7691 * @idle: The idle status of the CPU for whose sd load_idx is obtained.
7693 * Return: The load index.
7695 static inline int get_sd_load_idx(struct sched_domain *sd,
7696 enum cpu_idle_type idle)
7702 load_idx = sd->busy_idx;
7705 case CPU_NEWLY_IDLE:
7706 load_idx = sd->newidle_idx;
7709 load_idx = sd->idle_idx;
7716 static unsigned long scale_rt_capacity(struct sched_domain *sd, int cpu)
7718 struct rq *rq = cpu_rq(cpu);
7719 unsigned long max = arch_scale_cpu_capacity(sd, cpu);
7720 unsigned long used, free;
7723 irq = cpu_util_irq(rq);
7725 if (unlikely(irq >= max))
7728 used = READ_ONCE(rq->avg_rt.util_avg);
7729 used += READ_ONCE(rq->avg_dl.util_avg);
7731 if (unlikely(used >= max))
7736 return scale_irq_capacity(free, irq, max);
7739 static void update_cpu_capacity(struct sched_domain *sd, int cpu)
7741 unsigned long capacity = scale_rt_capacity(sd, cpu);
7742 struct sched_group *sdg = sd->groups;
7744 cpu_rq(cpu)->cpu_capacity_orig = arch_scale_cpu_capacity(sd, cpu);
7749 cpu_rq(cpu)->cpu_capacity = capacity;
7750 sdg->sgc->capacity = capacity;
7751 sdg->sgc->min_capacity = capacity;
7754 void update_group_capacity(struct sched_domain *sd, int cpu)
7756 struct sched_domain *child = sd->child;
7757 struct sched_group *group, *sdg = sd->groups;
7758 unsigned long capacity, min_capacity;
7759 unsigned long interval;
7761 interval = msecs_to_jiffies(sd->balance_interval);
7762 interval = clamp(interval, 1UL, max_load_balance_interval);
7763 sdg->sgc->next_update = jiffies + interval;
7766 update_cpu_capacity(sd, cpu);
7771 min_capacity = ULONG_MAX;
7773 if (child->flags & SD_OVERLAP) {
7775 * SD_OVERLAP domains cannot assume that child groups
7776 * span the current group.
7779 for_each_cpu(cpu, sched_group_span(sdg)) {
7780 struct sched_group_capacity *sgc;
7781 struct rq *rq = cpu_rq(cpu);
7784 * build_sched_domains() -> init_sched_groups_capacity()
7785 * gets here before we've attached the domains to the
7788 * Use capacity_of(), which is set irrespective of domains
7789 * in update_cpu_capacity().
7791 * This avoids capacity from being 0 and
7792 * causing divide-by-zero issues on boot.
7794 if (unlikely(!rq->sd)) {
7795 capacity += capacity_of(cpu);
7797 sgc = rq->sd->groups->sgc;
7798 capacity += sgc->capacity;
7801 min_capacity = min(capacity, min_capacity);
7805 * !SD_OVERLAP domains can assume that child groups
7806 * span the current group.
7809 group = child->groups;
7811 struct sched_group_capacity *sgc = group->sgc;
7813 capacity += sgc->capacity;
7814 min_capacity = min(sgc->min_capacity, min_capacity);
7815 group = group->next;
7816 } while (group != child->groups);
7819 sdg->sgc->capacity = capacity;
7820 sdg->sgc->min_capacity = min_capacity;
7824 * Check whether the capacity of the rq has been noticeably reduced by side
7825 * activity. The imbalance_pct is used for the threshold.
7826 * Return true is the capacity is reduced
7829 check_cpu_capacity(struct rq *rq, struct sched_domain *sd)
7831 return ((rq->cpu_capacity * sd->imbalance_pct) <
7832 (rq->cpu_capacity_orig * 100));
7836 * Group imbalance indicates (and tries to solve) the problem where balancing
7837 * groups is inadequate due to ->cpus_allowed constraints.
7839 * Imagine a situation of two groups of 4 CPUs each and 4 tasks each with a
7840 * cpumask covering 1 CPU of the first group and 3 CPUs of the second group.
7843 * { 0 1 2 3 } { 4 5 6 7 }
7846 * If we were to balance group-wise we'd place two tasks in the first group and
7847 * two tasks in the second group. Clearly this is undesired as it will overload
7848 * cpu 3 and leave one of the CPUs in the second group unused.
7850 * The current solution to this issue is detecting the skew in the first group
7851 * by noticing the lower domain failed to reach balance and had difficulty
7852 * moving tasks due to affinity constraints.
7854 * When this is so detected; this group becomes a candidate for busiest; see
7855 * update_sd_pick_busiest(). And calculate_imbalance() and
7856 * find_busiest_group() avoid some of the usual balance conditions to allow it
7857 * to create an effective group imbalance.
7859 * This is a somewhat tricky proposition since the next run might not find the
7860 * group imbalance and decide the groups need to be balanced again. A most
7861 * subtle and fragile situation.
7864 static inline int sg_imbalanced(struct sched_group *group)
7866 return group->sgc->imbalance;
7870 * group_has_capacity returns true if the group has spare capacity that could
7871 * be used by some tasks.
7872 * We consider that a group has spare capacity if the * number of task is
7873 * smaller than the number of CPUs or if the utilization is lower than the
7874 * available capacity for CFS tasks.
7875 * For the latter, we use a threshold to stabilize the state, to take into
7876 * account the variance of the tasks' load and to return true if the available
7877 * capacity in meaningful for the load balancer.
7878 * As an example, an available capacity of 1% can appear but it doesn't make
7879 * any benefit for the load balance.
7882 group_has_capacity(struct lb_env *env, struct sg_lb_stats *sgs)
7884 if (sgs->sum_nr_running < sgs->group_weight)
7887 if ((sgs->group_capacity * 100) >
7888 (sgs->group_util * env->sd->imbalance_pct))
7895 * group_is_overloaded returns true if the group has more tasks than it can
7897 * group_is_overloaded is not equals to !group_has_capacity because a group
7898 * with the exact right number of tasks, has no more spare capacity but is not
7899 * overloaded so both group_has_capacity and group_is_overloaded return
7903 group_is_overloaded(struct lb_env *env, struct sg_lb_stats *sgs)
7905 if (sgs->sum_nr_running <= sgs->group_weight)
7908 if ((sgs->group_capacity * 100) <
7909 (sgs->group_util * env->sd->imbalance_pct))
7916 * group_smaller_cpu_capacity: Returns true if sched_group sg has smaller
7917 * per-CPU capacity than sched_group ref.
7920 group_smaller_cpu_capacity(struct sched_group *sg, struct sched_group *ref)
7922 return sg->sgc->min_capacity * capacity_margin <
7923 ref->sgc->min_capacity * 1024;
7927 group_type group_classify(struct sched_group *group,
7928 struct sg_lb_stats *sgs)
7930 if (sgs->group_no_capacity)
7931 return group_overloaded;
7933 if (sg_imbalanced(group))
7934 return group_imbalanced;
7939 static bool update_nohz_stats(struct rq *rq, bool force)
7941 #ifdef CONFIG_NO_HZ_COMMON
7942 unsigned int cpu = rq->cpu;
7944 if (!rq->has_blocked_load)
7947 if (!cpumask_test_cpu(cpu, nohz.idle_cpus_mask))
7950 if (!force && !time_after(jiffies, rq->last_blocked_load_update_tick))
7953 update_blocked_averages(cpu);
7955 return rq->has_blocked_load;
7962 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
7963 * @env: The load balancing environment.
7964 * @group: sched_group whose statistics are to be updated.
7965 * @load_idx: Load index of sched_domain of this_cpu for load calc.
7966 * @local_group: Does group contain this_cpu.
7967 * @sgs: variable to hold the statistics for this group.
7968 * @overload: Indicate more than one runnable task for any CPU.
7970 static inline void update_sg_lb_stats(struct lb_env *env,
7971 struct sched_group *group, int load_idx,
7972 int local_group, struct sg_lb_stats *sgs,
7978 memset(sgs, 0, sizeof(*sgs));
7980 for_each_cpu_and(i, sched_group_span(group), env->cpus) {
7981 struct rq *rq = cpu_rq(i);
7983 if ((env->flags & LBF_NOHZ_STATS) && update_nohz_stats(rq, false))
7984 env->flags |= LBF_NOHZ_AGAIN;
7986 /* Bias balancing toward CPUs of our domain: */
7988 load = target_load(i, load_idx);
7990 load = source_load(i, load_idx);
7992 sgs->group_load += load;
7993 sgs->group_util += cpu_util(i);
7994 sgs->sum_nr_running += rq->cfs.h_nr_running;
7996 nr_running = rq->nr_running;
8000 #ifdef CONFIG_NUMA_BALANCING
8001 sgs->nr_numa_running += rq->nr_numa_running;
8002 sgs->nr_preferred_running += rq->nr_preferred_running;
8004 sgs->sum_weighted_load += weighted_cpuload(rq);
8006 * No need to call idle_cpu() if nr_running is not 0
8008 if (!nr_running && idle_cpu(i))
8012 /* Adjust by relative CPU capacity of the group */
8013 sgs->group_capacity = group->sgc->capacity;
8014 sgs->avg_load = (sgs->group_load*SCHED_CAPACITY_SCALE) / sgs->group_capacity;
8016 if (sgs->sum_nr_running)
8017 sgs->load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
8019 sgs->group_weight = group->group_weight;
8021 sgs->group_no_capacity = group_is_overloaded(env, sgs);
8022 sgs->group_type = group_classify(group, sgs);
8026 * update_sd_pick_busiest - return 1 on busiest group
8027 * @env: The load balancing environment.
8028 * @sds: sched_domain statistics
8029 * @sg: sched_group candidate to be checked for being the busiest
8030 * @sgs: sched_group statistics
8032 * Determine if @sg is a busier group than the previously selected
8035 * Return: %true if @sg is a busier group than the previously selected
8036 * busiest group. %false otherwise.
8038 static bool update_sd_pick_busiest(struct lb_env *env,
8039 struct sd_lb_stats *sds,
8040 struct sched_group *sg,
8041 struct sg_lb_stats *sgs)
8043 struct sg_lb_stats *busiest = &sds->busiest_stat;
8045 if (sgs->group_type > busiest->group_type)
8048 if (sgs->group_type < busiest->group_type)
8051 if (sgs->avg_load <= busiest->avg_load)
8054 if (!(env->sd->flags & SD_ASYM_CPUCAPACITY))
8058 * Candidate sg has no more than one task per CPU and
8059 * has higher per-CPU capacity. Migrating tasks to less
8060 * capable CPUs may harm throughput. Maximize throughput,
8061 * power/energy consequences are not considered.
8063 if (sgs->sum_nr_running <= sgs->group_weight &&
8064 group_smaller_cpu_capacity(sds->local, sg))
8068 /* This is the busiest node in its class. */
8069 if (!(env->sd->flags & SD_ASYM_PACKING))
8072 /* No ASYM_PACKING if target CPU is already busy */
8073 if (env->idle == CPU_NOT_IDLE)
8076 * ASYM_PACKING needs to move all the work to the highest
8077 * prority CPUs in the group, therefore mark all groups
8078 * of lower priority than ourself as busy.
8080 if (sgs->sum_nr_running &&
8081 sched_asym_prefer(env->dst_cpu, sg->asym_prefer_cpu)) {
8085 /* Prefer to move from lowest priority CPU's work */
8086 if (sched_asym_prefer(sds->busiest->asym_prefer_cpu,
8087 sg->asym_prefer_cpu))
8094 #ifdef CONFIG_NUMA_BALANCING
8095 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
8097 if (sgs->sum_nr_running > sgs->nr_numa_running)
8099 if (sgs->sum_nr_running > sgs->nr_preferred_running)
8104 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
8106 if (rq->nr_running > rq->nr_numa_running)
8108 if (rq->nr_running > rq->nr_preferred_running)
8113 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
8118 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
8122 #endif /* CONFIG_NUMA_BALANCING */
8125 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
8126 * @env: The load balancing environment.
8127 * @sds: variable to hold the statistics for this sched_domain.
8129 static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
8131 struct sched_domain *child = env->sd->child;
8132 struct sched_group *sg = env->sd->groups;
8133 struct sg_lb_stats *local = &sds->local_stat;
8134 struct sg_lb_stats tmp_sgs;
8135 int load_idx, prefer_sibling = 0;
8136 bool overload = false;
8138 if (child && child->flags & SD_PREFER_SIBLING)
8141 #ifdef CONFIG_NO_HZ_COMMON
8142 if (env->idle == CPU_NEWLY_IDLE && READ_ONCE(nohz.has_blocked))
8143 env->flags |= LBF_NOHZ_STATS;
8146 load_idx = get_sd_load_idx(env->sd, env->idle);
8149 struct sg_lb_stats *sgs = &tmp_sgs;
8152 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_span(sg));
8157 if (env->idle != CPU_NEWLY_IDLE ||
8158 time_after_eq(jiffies, sg->sgc->next_update))
8159 update_group_capacity(env->sd, env->dst_cpu);
8162 update_sg_lb_stats(env, sg, load_idx, local_group, sgs,
8169 * In case the child domain prefers tasks go to siblings
8170 * first, lower the sg capacity so that we'll try
8171 * and move all the excess tasks away. We lower the capacity
8172 * of a group only if the local group has the capacity to fit
8173 * these excess tasks. The extra check prevents the case where
8174 * you always pull from the heaviest group when it is already
8175 * under-utilized (possible with a large weight task outweighs
8176 * the tasks on the system).
8178 if (prefer_sibling && sds->local &&
8179 group_has_capacity(env, local) &&
8180 (sgs->sum_nr_running > local->sum_nr_running + 1)) {
8181 sgs->group_no_capacity = 1;
8182 sgs->group_type = group_classify(sg, sgs);
8185 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
8187 sds->busiest_stat = *sgs;
8191 /* Now, start updating sd_lb_stats */
8192 sds->total_running += sgs->sum_nr_running;
8193 sds->total_load += sgs->group_load;
8194 sds->total_capacity += sgs->group_capacity;
8197 } while (sg != env->sd->groups);
8199 #ifdef CONFIG_NO_HZ_COMMON
8200 if ((env->flags & LBF_NOHZ_AGAIN) &&
8201 cpumask_subset(nohz.idle_cpus_mask, sched_domain_span(env->sd))) {
8203 WRITE_ONCE(nohz.next_blocked,
8204 jiffies + msecs_to_jiffies(LOAD_AVG_PERIOD));
8208 if (env->sd->flags & SD_NUMA)
8209 env->fbq_type = fbq_classify_group(&sds->busiest_stat);
8211 if (!env->sd->parent) {
8212 /* update overload indicator if we are at root domain */
8213 if (env->dst_rq->rd->overload != overload)
8214 env->dst_rq->rd->overload = overload;
8219 * check_asym_packing - Check to see if the group is packed into the
8222 * This is primarily intended to used at the sibling level. Some
8223 * cores like POWER7 prefer to use lower numbered SMT threads. In the
8224 * case of POWER7, it can move to lower SMT modes only when higher
8225 * threads are idle. When in lower SMT modes, the threads will
8226 * perform better since they share less core resources. Hence when we
8227 * have idle threads, we want them to be the higher ones.
8229 * This packing function is run on idle threads. It checks to see if
8230 * the busiest CPU in this domain (core in the P7 case) has a higher
8231 * CPU number than the packing function is being run on. Here we are
8232 * assuming lower CPU number will be equivalent to lower a SMT thread
8235 * Return: 1 when packing is required and a task should be moved to
8236 * this CPU. The amount of the imbalance is returned in env->imbalance.
8238 * @env: The load balancing environment.
8239 * @sds: Statistics of the sched_domain which is to be packed
8241 static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
8245 if (!(env->sd->flags & SD_ASYM_PACKING))
8248 if (env->idle == CPU_NOT_IDLE)
8254 busiest_cpu = sds->busiest->asym_prefer_cpu;
8255 if (sched_asym_prefer(busiest_cpu, env->dst_cpu))
8258 env->imbalance = DIV_ROUND_CLOSEST(
8259 sds->busiest_stat.avg_load * sds->busiest_stat.group_capacity,
8260 SCHED_CAPACITY_SCALE);
8266 * fix_small_imbalance - Calculate the minor imbalance that exists
8267 * amongst the groups of a sched_domain, during
8269 * @env: The load balancing environment.
8270 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
8273 void fix_small_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
8275 unsigned long tmp, capa_now = 0, capa_move = 0;
8276 unsigned int imbn = 2;
8277 unsigned long scaled_busy_load_per_task;
8278 struct sg_lb_stats *local, *busiest;
8280 local = &sds->local_stat;
8281 busiest = &sds->busiest_stat;
8283 if (!local->sum_nr_running)
8284 local->load_per_task = cpu_avg_load_per_task(env->dst_cpu);
8285 else if (busiest->load_per_task > local->load_per_task)
8288 scaled_busy_load_per_task =
8289 (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
8290 busiest->group_capacity;
8292 if (busiest->avg_load + scaled_busy_load_per_task >=
8293 local->avg_load + (scaled_busy_load_per_task * imbn)) {
8294 env->imbalance = busiest->load_per_task;
8299 * OK, we don't have enough imbalance to justify moving tasks,
8300 * however we may be able to increase total CPU capacity used by
8304 capa_now += busiest->group_capacity *
8305 min(busiest->load_per_task, busiest->avg_load);
8306 capa_now += local->group_capacity *
8307 min(local->load_per_task, local->avg_load);
8308 capa_now /= SCHED_CAPACITY_SCALE;
8310 /* Amount of load we'd subtract */
8311 if (busiest->avg_load > scaled_busy_load_per_task) {
8312 capa_move += busiest->group_capacity *
8313 min(busiest->load_per_task,
8314 busiest->avg_load - scaled_busy_load_per_task);
8317 /* Amount of load we'd add */
8318 if (busiest->avg_load * busiest->group_capacity <
8319 busiest->load_per_task * SCHED_CAPACITY_SCALE) {
8320 tmp = (busiest->avg_load * busiest->group_capacity) /
8321 local->group_capacity;
8323 tmp = (busiest->load_per_task * SCHED_CAPACITY_SCALE) /
8324 local->group_capacity;
8326 capa_move += local->group_capacity *
8327 min(local->load_per_task, local->avg_load + tmp);
8328 capa_move /= SCHED_CAPACITY_SCALE;
8330 /* Move if we gain throughput */
8331 if (capa_move > capa_now)
8332 env->imbalance = busiest->load_per_task;
8336 * calculate_imbalance - Calculate the amount of imbalance present within the
8337 * groups of a given sched_domain during load balance.
8338 * @env: load balance environment
8339 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
8341 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
8343 unsigned long max_pull, load_above_capacity = ~0UL;
8344 struct sg_lb_stats *local, *busiest;
8346 local = &sds->local_stat;
8347 busiest = &sds->busiest_stat;
8349 if (busiest->group_type == group_imbalanced) {
8351 * In the group_imb case we cannot rely on group-wide averages
8352 * to ensure CPU-load equilibrium, look at wider averages. XXX
8354 busiest->load_per_task =
8355 min(busiest->load_per_task, sds->avg_load);
8359 * Avg load of busiest sg can be less and avg load of local sg can
8360 * be greater than avg load across all sgs of sd because avg load
8361 * factors in sg capacity and sgs with smaller group_type are
8362 * skipped when updating the busiest sg:
8364 if (busiest->avg_load <= sds->avg_load ||
8365 local->avg_load >= sds->avg_load) {
8367 return fix_small_imbalance(env, sds);
8371 * If there aren't any idle CPUs, avoid creating some.
8373 if (busiest->group_type == group_overloaded &&
8374 local->group_type == group_overloaded) {
8375 load_above_capacity = busiest->sum_nr_running * SCHED_CAPACITY_SCALE;
8376 if (load_above_capacity > busiest->group_capacity) {
8377 load_above_capacity -= busiest->group_capacity;
8378 load_above_capacity *= scale_load_down(NICE_0_LOAD);
8379 load_above_capacity /= busiest->group_capacity;
8381 load_above_capacity = ~0UL;
8385 * We're trying to get all the CPUs to the average_load, so we don't
8386 * want to push ourselves above the average load, nor do we wish to
8387 * reduce the max loaded CPU below the average load. At the same time,
8388 * we also don't want to reduce the group load below the group
8389 * capacity. Thus we look for the minimum possible imbalance.
8391 max_pull = min(busiest->avg_load - sds->avg_load, load_above_capacity);
8393 /* How much load to actually move to equalise the imbalance */
8394 env->imbalance = min(
8395 max_pull * busiest->group_capacity,
8396 (sds->avg_load - local->avg_load) * local->group_capacity
8397 ) / SCHED_CAPACITY_SCALE;
8400 * if *imbalance is less than the average load per runnable task
8401 * there is no guarantee that any tasks will be moved so we'll have
8402 * a think about bumping its value to force at least one task to be
8405 if (env->imbalance < busiest->load_per_task)
8406 return fix_small_imbalance(env, sds);
8409 /******* find_busiest_group() helpers end here *********************/
8412 * find_busiest_group - Returns the busiest group within the sched_domain
8413 * if there is an imbalance.
8415 * Also calculates the amount of weighted load which should be moved
8416 * to restore balance.
8418 * @env: The load balancing environment.
8420 * Return: - The busiest group if imbalance exists.
8422 static struct sched_group *find_busiest_group(struct lb_env *env)
8424 struct sg_lb_stats *local, *busiest;
8425 struct sd_lb_stats sds;
8427 init_sd_lb_stats(&sds);
8430 * Compute the various statistics relavent for load balancing at
8433 update_sd_lb_stats(env, &sds);
8434 local = &sds.local_stat;
8435 busiest = &sds.busiest_stat;
8437 /* ASYM feature bypasses nice load balance check */
8438 if (check_asym_packing(env, &sds))
8441 /* There is no busy sibling group to pull tasks from */
8442 if (!sds.busiest || busiest->sum_nr_running == 0)
8445 /* XXX broken for overlapping NUMA groups */
8446 sds.avg_load = (SCHED_CAPACITY_SCALE * sds.total_load)
8447 / sds.total_capacity;
8450 * If the busiest group is imbalanced the below checks don't
8451 * work because they assume all things are equal, which typically
8452 * isn't true due to cpus_allowed constraints and the like.
8454 if (busiest->group_type == group_imbalanced)
8458 * When dst_cpu is idle, prevent SMP nice and/or asymmetric group
8459 * capacities from resulting in underutilization due to avg_load.
8461 if (env->idle != CPU_NOT_IDLE && group_has_capacity(env, local) &&
8462 busiest->group_no_capacity)
8466 * If the local group is busier than the selected busiest group
8467 * don't try and pull any tasks.
8469 if (local->avg_load >= busiest->avg_load)
8473 * Don't pull any tasks if this group is already above the domain
8476 if (local->avg_load >= sds.avg_load)
8479 if (env->idle == CPU_IDLE) {
8481 * This CPU is idle. If the busiest group is not overloaded
8482 * and there is no imbalance between this and busiest group
8483 * wrt idle CPUs, it is balanced. The imbalance becomes
8484 * significant if the diff is greater than 1 otherwise we
8485 * might end up to just move the imbalance on another group
8487 if ((busiest->group_type != group_overloaded) &&
8488 (local->idle_cpus <= (busiest->idle_cpus + 1)))
8492 * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
8493 * imbalance_pct to be conservative.
8495 if (100 * busiest->avg_load <=
8496 env->sd->imbalance_pct * local->avg_load)
8501 /* Looks like there is an imbalance. Compute it */
8502 calculate_imbalance(env, &sds);
8503 return env->imbalance ? sds.busiest : NULL;
8511 * find_busiest_queue - find the busiest runqueue among the CPUs in the group.
8513 static struct rq *find_busiest_queue(struct lb_env *env,
8514 struct sched_group *group)
8516 struct rq *busiest = NULL, *rq;
8517 unsigned long busiest_load = 0, busiest_capacity = 1;
8520 for_each_cpu_and(i, sched_group_span(group), env->cpus) {
8521 unsigned long capacity, wl;
8525 rt = fbq_classify_rq(rq);
8528 * We classify groups/runqueues into three groups:
8529 * - regular: there are !numa tasks
8530 * - remote: there are numa tasks that run on the 'wrong' node
8531 * - all: there is no distinction
8533 * In order to avoid migrating ideally placed numa tasks,
8534 * ignore those when there's better options.
8536 * If we ignore the actual busiest queue to migrate another
8537 * task, the next balance pass can still reduce the busiest
8538 * queue by moving tasks around inside the node.
8540 * If we cannot move enough load due to this classification
8541 * the next pass will adjust the group classification and
8542 * allow migration of more tasks.
8544 * Both cases only affect the total convergence complexity.
8546 if (rt > env->fbq_type)
8549 capacity = capacity_of(i);
8551 wl = weighted_cpuload(rq);
8554 * When comparing with imbalance, use weighted_cpuload()
8555 * which is not scaled with the CPU capacity.
8558 if (rq->nr_running == 1 && wl > env->imbalance &&
8559 !check_cpu_capacity(rq, env->sd))
8563 * For the load comparisons with the other CPU's, consider
8564 * the weighted_cpuload() scaled with the CPU capacity, so
8565 * that the load can be moved away from the CPU that is
8566 * potentially running at a lower capacity.
8568 * Thus we're looking for max(wl_i / capacity_i), crosswise
8569 * multiplication to rid ourselves of the division works out
8570 * to: wl_i * capacity_j > wl_j * capacity_i; where j is
8571 * our previous maximum.
8573 if (wl * busiest_capacity > busiest_load * capacity) {
8575 busiest_capacity = capacity;
8584 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
8585 * so long as it is large enough.
8587 #define MAX_PINNED_INTERVAL 512
8589 static int need_active_balance(struct lb_env *env)
8591 struct sched_domain *sd = env->sd;
8593 if (env->idle == CPU_NEWLY_IDLE) {
8596 * ASYM_PACKING needs to force migrate tasks from busy but
8597 * lower priority CPUs in order to pack all tasks in the
8598 * highest priority CPUs.
8600 if ((sd->flags & SD_ASYM_PACKING) &&
8601 sched_asym_prefer(env->dst_cpu, env->src_cpu))
8606 * The dst_cpu is idle and the src_cpu CPU has only 1 CFS task.
8607 * It's worth migrating the task if the src_cpu's capacity is reduced
8608 * because of other sched_class or IRQs if more capacity stays
8609 * available on dst_cpu.
8611 if ((env->idle != CPU_NOT_IDLE) &&
8612 (env->src_rq->cfs.h_nr_running == 1)) {
8613 if ((check_cpu_capacity(env->src_rq, sd)) &&
8614 (capacity_of(env->src_cpu)*sd->imbalance_pct < capacity_of(env->dst_cpu)*100))
8618 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
8621 static int active_load_balance_cpu_stop(void *data);
8623 static int should_we_balance(struct lb_env *env)
8625 struct sched_group *sg = env->sd->groups;
8626 int cpu, balance_cpu = -1;
8629 * Ensure the balancing environment is consistent; can happen
8630 * when the softirq triggers 'during' hotplug.
8632 if (!cpumask_test_cpu(env->dst_cpu, env->cpus))
8636 * In the newly idle case, we will allow all the CPUs
8637 * to do the newly idle load balance.
8639 if (env->idle == CPU_NEWLY_IDLE)
8642 /* Try to find first idle CPU */
8643 for_each_cpu_and(cpu, group_balance_mask(sg), env->cpus) {
8651 if (balance_cpu == -1)
8652 balance_cpu = group_balance_cpu(sg);
8655 * First idle CPU or the first CPU(busiest) in this sched group
8656 * is eligible for doing load balancing at this and above domains.
8658 return balance_cpu == env->dst_cpu;
8662 * Check this_cpu to ensure it is balanced within domain. Attempt to move
8663 * tasks if there is an imbalance.
8665 static int load_balance(int this_cpu, struct rq *this_rq,
8666 struct sched_domain *sd, enum cpu_idle_type idle,
8667 int *continue_balancing)
8669 int ld_moved, cur_ld_moved, active_balance = 0;
8670 struct sched_domain *sd_parent = sd->parent;
8671 struct sched_group *group;
8674 struct cpumask *cpus = this_cpu_cpumask_var_ptr(load_balance_mask);
8676 struct lb_env env = {
8678 .dst_cpu = this_cpu,
8680 .dst_grpmask = sched_group_span(sd->groups),
8682 .loop_break = sched_nr_migrate_break,
8685 .tasks = LIST_HEAD_INIT(env.tasks),
8688 cpumask_and(cpus, sched_domain_span(sd), cpu_active_mask);
8690 schedstat_inc(sd->lb_count[idle]);
8693 if (!should_we_balance(&env)) {
8694 *continue_balancing = 0;
8698 group = find_busiest_group(&env);
8700 schedstat_inc(sd->lb_nobusyg[idle]);
8704 busiest = find_busiest_queue(&env, group);
8706 schedstat_inc(sd->lb_nobusyq[idle]);
8710 BUG_ON(busiest == env.dst_rq);
8712 schedstat_add(sd->lb_imbalance[idle], env.imbalance);
8714 env.src_cpu = busiest->cpu;
8715 env.src_rq = busiest;
8718 if (busiest->nr_running > 1) {
8720 * Attempt to move tasks. If find_busiest_group has found
8721 * an imbalance but busiest->nr_running <= 1, the group is
8722 * still unbalanced. ld_moved simply stays zero, so it is
8723 * correctly treated as an imbalance.
8725 env.flags |= LBF_ALL_PINNED;
8726 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
8729 rq_lock_irqsave(busiest, &rf);
8730 update_rq_clock(busiest);
8733 * cur_ld_moved - load moved in current iteration
8734 * ld_moved - cumulative load moved across iterations
8736 cur_ld_moved = detach_tasks(&env);
8739 * We've detached some tasks from busiest_rq. Every
8740 * task is masked "TASK_ON_RQ_MIGRATING", so we can safely
8741 * unlock busiest->lock, and we are able to be sure
8742 * that nobody can manipulate the tasks in parallel.
8743 * See task_rq_lock() family for the details.
8746 rq_unlock(busiest, &rf);
8750 ld_moved += cur_ld_moved;
8753 local_irq_restore(rf.flags);
8755 if (env.flags & LBF_NEED_BREAK) {
8756 env.flags &= ~LBF_NEED_BREAK;
8761 * Revisit (affine) tasks on src_cpu that couldn't be moved to
8762 * us and move them to an alternate dst_cpu in our sched_group
8763 * where they can run. The upper limit on how many times we
8764 * iterate on same src_cpu is dependent on number of CPUs in our
8767 * This changes load balance semantics a bit on who can move
8768 * load to a given_cpu. In addition to the given_cpu itself
8769 * (or a ilb_cpu acting on its behalf where given_cpu is
8770 * nohz-idle), we now have balance_cpu in a position to move
8771 * load to given_cpu. In rare situations, this may cause
8772 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
8773 * _independently_ and at _same_ time to move some load to
8774 * given_cpu) causing exceess load to be moved to given_cpu.
8775 * This however should not happen so much in practice and
8776 * moreover subsequent load balance cycles should correct the
8777 * excess load moved.
8779 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
8781 /* Prevent to re-select dst_cpu via env's CPUs */
8782 cpumask_clear_cpu(env.dst_cpu, env.cpus);
8784 env.dst_rq = cpu_rq(env.new_dst_cpu);
8785 env.dst_cpu = env.new_dst_cpu;
8786 env.flags &= ~LBF_DST_PINNED;
8788 env.loop_break = sched_nr_migrate_break;
8791 * Go back to "more_balance" rather than "redo" since we
8792 * need to continue with same src_cpu.
8798 * We failed to reach balance because of affinity.
8801 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
8803 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0)
8804 *group_imbalance = 1;
8807 /* All tasks on this runqueue were pinned by CPU affinity */
8808 if (unlikely(env.flags & LBF_ALL_PINNED)) {
8809 cpumask_clear_cpu(cpu_of(busiest), cpus);
8811 * Attempting to continue load balancing at the current
8812 * sched_domain level only makes sense if there are
8813 * active CPUs remaining as possible busiest CPUs to
8814 * pull load from which are not contained within the
8815 * destination group that is receiving any migrated
8818 if (!cpumask_subset(cpus, env.dst_grpmask)) {
8820 env.loop_break = sched_nr_migrate_break;
8823 goto out_all_pinned;
8828 schedstat_inc(sd->lb_failed[idle]);
8830 * Increment the failure counter only on periodic balance.
8831 * We do not want newidle balance, which can be very
8832 * frequent, pollute the failure counter causing
8833 * excessive cache_hot migrations and active balances.
8835 if (idle != CPU_NEWLY_IDLE)
8836 sd->nr_balance_failed++;
8838 if (need_active_balance(&env)) {
8839 unsigned long flags;
8841 raw_spin_lock_irqsave(&busiest->lock, flags);
8844 * Don't kick the active_load_balance_cpu_stop,
8845 * if the curr task on busiest CPU can't be
8846 * moved to this_cpu:
8848 if (!cpumask_test_cpu(this_cpu, &busiest->curr->cpus_allowed)) {
8849 raw_spin_unlock_irqrestore(&busiest->lock,
8851 env.flags |= LBF_ALL_PINNED;
8852 goto out_one_pinned;
8856 * ->active_balance synchronizes accesses to
8857 * ->active_balance_work. Once set, it's cleared
8858 * only after active load balance is finished.
8860 if (!busiest->active_balance) {
8861 busiest->active_balance = 1;
8862 busiest->push_cpu = this_cpu;
8865 raw_spin_unlock_irqrestore(&busiest->lock, flags);
8867 if (active_balance) {
8868 stop_one_cpu_nowait(cpu_of(busiest),
8869 active_load_balance_cpu_stop, busiest,
8870 &busiest->active_balance_work);
8873 /* We've kicked active balancing, force task migration. */
8874 sd->nr_balance_failed = sd->cache_nice_tries+1;
8877 sd->nr_balance_failed = 0;
8879 if (likely(!active_balance)) {
8880 /* We were unbalanced, so reset the balancing interval */
8881 sd->balance_interval = sd->min_interval;
8884 * If we've begun active balancing, start to back off. This
8885 * case may not be covered by the all_pinned logic if there
8886 * is only 1 task on the busy runqueue (because we don't call
8889 if (sd->balance_interval < sd->max_interval)
8890 sd->balance_interval *= 2;
8897 * We reach balance although we may have faced some affinity
8898 * constraints. Clear the imbalance flag only if other tasks got
8899 * a chance to move and fix the imbalance.
8901 if (sd_parent && !(env.flags & LBF_ALL_PINNED)) {
8902 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
8904 if (*group_imbalance)
8905 *group_imbalance = 0;
8910 * We reach balance because all tasks are pinned at this level so
8911 * we can't migrate them. Let the imbalance flag set so parent level
8912 * can try to migrate them.
8914 schedstat_inc(sd->lb_balanced[idle]);
8916 sd->nr_balance_failed = 0;
8922 * idle_balance() disregards balance intervals, so we could repeatedly
8923 * reach this code, which would lead to balance_interval skyrocketting
8924 * in a short amount of time. Skip the balance_interval increase logic
8927 if (env.idle == CPU_NEWLY_IDLE)
8930 /* tune up the balancing interval */
8931 if (((env.flags & LBF_ALL_PINNED) &&
8932 sd->balance_interval < MAX_PINNED_INTERVAL) ||
8933 (sd->balance_interval < sd->max_interval))
8934 sd->balance_interval *= 2;
8939 static inline unsigned long
8940 get_sd_balance_interval(struct sched_domain *sd, int cpu_busy)
8942 unsigned long interval = sd->balance_interval;
8945 interval *= sd->busy_factor;
8947 /* scale ms to jiffies */
8948 interval = msecs_to_jiffies(interval);
8949 interval = clamp(interval, 1UL, max_load_balance_interval);
8955 update_next_balance(struct sched_domain *sd, unsigned long *next_balance)
8957 unsigned long interval, next;
8959 /* used by idle balance, so cpu_busy = 0 */
8960 interval = get_sd_balance_interval(sd, 0);
8961 next = sd->last_balance + interval;
8963 if (time_after(*next_balance, next))
8964 *next_balance = next;
8968 * active_load_balance_cpu_stop is run by the CPU stopper. It pushes
8969 * running tasks off the busiest CPU onto idle CPUs. It requires at
8970 * least 1 task to be running on each physical CPU where possible, and
8971 * avoids physical / logical imbalances.
8973 static int active_load_balance_cpu_stop(void *data)
8975 struct rq *busiest_rq = data;
8976 int busiest_cpu = cpu_of(busiest_rq);
8977 int target_cpu = busiest_rq->push_cpu;
8978 struct rq *target_rq = cpu_rq(target_cpu);
8979 struct sched_domain *sd;
8980 struct task_struct *p = NULL;
8983 rq_lock_irq(busiest_rq, &rf);
8985 * Between queueing the stop-work and running it is a hole in which
8986 * CPUs can become inactive. We should not move tasks from or to
8989 if (!cpu_active(busiest_cpu) || !cpu_active(target_cpu))
8992 /* Make sure the requested CPU hasn't gone down in the meantime: */
8993 if (unlikely(busiest_cpu != smp_processor_id() ||
8994 !busiest_rq->active_balance))
8997 /* Is there any task to move? */
8998 if (busiest_rq->nr_running <= 1)
9002 * This condition is "impossible", if it occurs
9003 * we need to fix it. Originally reported by
9004 * Bjorn Helgaas on a 128-CPU setup.
9006 BUG_ON(busiest_rq == target_rq);
9008 /* Search for an sd spanning us and the target CPU. */
9010 for_each_domain(target_cpu, sd) {
9011 if ((sd->flags & SD_LOAD_BALANCE) &&
9012 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
9017 struct lb_env env = {
9019 .dst_cpu = target_cpu,
9020 .dst_rq = target_rq,
9021 .src_cpu = busiest_rq->cpu,
9022 .src_rq = busiest_rq,
9025 * can_migrate_task() doesn't need to compute new_dst_cpu
9026 * for active balancing. Since we have CPU_IDLE, but no
9027 * @dst_grpmask we need to make that test go away with lying
9030 .flags = LBF_DST_PINNED,
9033 schedstat_inc(sd->alb_count);
9034 update_rq_clock(busiest_rq);
9036 p = detach_one_task(&env);
9038 schedstat_inc(sd->alb_pushed);
9039 /* Active balancing done, reset the failure counter. */
9040 sd->nr_balance_failed = 0;
9042 schedstat_inc(sd->alb_failed);
9047 busiest_rq->active_balance = 0;
9048 rq_unlock(busiest_rq, &rf);
9051 attach_one_task(target_rq, p);
9058 static DEFINE_SPINLOCK(balancing);
9061 * Scale the max load_balance interval with the number of CPUs in the system.
9062 * This trades load-balance latency on larger machines for less cross talk.
9064 void update_max_interval(void)
9066 max_load_balance_interval = HZ*num_online_cpus()/10;
9070 * It checks each scheduling domain to see if it is due to be balanced,
9071 * and initiates a balancing operation if so.
9073 * Balancing parameters are set up in init_sched_domains.
9075 static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle)
9077 int continue_balancing = 1;
9079 unsigned long interval;
9080 struct sched_domain *sd;
9081 /* Earliest time when we have to do rebalance again */
9082 unsigned long next_balance = jiffies + 60*HZ;
9083 int update_next_balance = 0;
9084 int need_serialize, need_decay = 0;
9088 for_each_domain(cpu, sd) {
9090 * Decay the newidle max times here because this is a regular
9091 * visit to all the domains. Decay ~1% per second.
9093 if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
9094 sd->max_newidle_lb_cost =
9095 (sd->max_newidle_lb_cost * 253) / 256;
9096 sd->next_decay_max_lb_cost = jiffies + HZ;
9099 max_cost += sd->max_newidle_lb_cost;
9101 if (!(sd->flags & SD_LOAD_BALANCE))
9105 * Stop the load balance at this level. There is another
9106 * CPU in our sched group which is doing load balancing more
9109 if (!continue_balancing) {
9115 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
9117 need_serialize = sd->flags & SD_SERIALIZE;
9118 if (need_serialize) {
9119 if (!spin_trylock(&balancing))
9123 if (time_after_eq(jiffies, sd->last_balance + interval)) {
9124 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
9126 * The LBF_DST_PINNED logic could have changed
9127 * env->dst_cpu, so we can't know our idle
9128 * state even if we migrated tasks. Update it.
9130 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
9132 sd->last_balance = jiffies;
9133 interval = get_sd_balance_interval(sd, idle != CPU_IDLE);
9136 spin_unlock(&balancing);
9138 if (time_after(next_balance, sd->last_balance + interval)) {
9139 next_balance = sd->last_balance + interval;
9140 update_next_balance = 1;
9145 * Ensure the rq-wide value also decays but keep it at a
9146 * reasonable floor to avoid funnies with rq->avg_idle.
9148 rq->max_idle_balance_cost =
9149 max((u64)sysctl_sched_migration_cost, max_cost);
9154 * next_balance will be updated only when there is a need.
9155 * When the cpu is attached to null domain for ex, it will not be
9158 if (likely(update_next_balance)) {
9159 rq->next_balance = next_balance;
9161 #ifdef CONFIG_NO_HZ_COMMON
9163 * If this CPU has been elected to perform the nohz idle
9164 * balance. Other idle CPUs have already rebalanced with
9165 * nohz_idle_balance() and nohz.next_balance has been
9166 * updated accordingly. This CPU is now running the idle load
9167 * balance for itself and we need to update the
9168 * nohz.next_balance accordingly.
9170 if ((idle == CPU_IDLE) && time_after(nohz.next_balance, rq->next_balance))
9171 nohz.next_balance = rq->next_balance;
9176 static inline int on_null_domain(struct rq *rq)
9178 return unlikely(!rcu_dereference_sched(rq->sd));
9181 #ifdef CONFIG_NO_HZ_COMMON
9183 * idle load balancing details
9184 * - When one of the busy CPUs notice that there may be an idle rebalancing
9185 * needed, they will kick the idle load balancer, which then does idle
9186 * load balancing for all the idle CPUs.
9187 * - HK_FLAG_MISC CPUs are used for this task, because HK_FLAG_SCHED not set
9191 static inline int find_new_ilb(void)
9195 for_each_cpu_and(ilb, nohz.idle_cpus_mask,
9196 housekeeping_cpumask(HK_FLAG_MISC)) {
9205 * Kick a CPU to do the nohz balancing, if it is time for it. We pick any
9206 * idle CPU in the HK_FLAG_MISC housekeeping set (if there is one).
9208 static void kick_ilb(unsigned int flags)
9213 * Increase nohz.next_balance only when if full ilb is triggered but
9214 * not if we only update stats.
9216 if (flags & NOHZ_BALANCE_KICK)
9217 nohz.next_balance = jiffies+1;
9219 ilb_cpu = find_new_ilb();
9221 if (ilb_cpu >= nr_cpu_ids)
9224 flags = atomic_fetch_or(flags, nohz_flags(ilb_cpu));
9225 if (flags & NOHZ_KICK_MASK)
9229 * Use smp_send_reschedule() instead of resched_cpu().
9230 * This way we generate a sched IPI on the target CPU which
9231 * is idle. And the softirq performing nohz idle load balance
9232 * will be run before returning from the IPI.
9234 smp_send_reschedule(ilb_cpu);
9238 * Current heuristic for kicking the idle load balancer in the presence
9239 * of an idle cpu in the system.
9240 * - This rq has more than one task.
9241 * - This rq has at least one CFS task and the capacity of the CPU is
9242 * significantly reduced because of RT tasks or IRQs.
9243 * - At parent of LLC scheduler domain level, this cpu's scheduler group has
9244 * multiple busy cpu.
9245 * - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
9246 * domain span are idle.
9248 static void nohz_balancer_kick(struct rq *rq)
9250 unsigned long now = jiffies;
9251 struct sched_domain_shared *sds;
9252 struct sched_domain *sd;
9253 int nr_busy, i, cpu = rq->cpu;
9254 unsigned int flags = 0;
9256 if (unlikely(rq->idle_balance))
9260 * We may be recently in ticked or tickless idle mode. At the first
9261 * busy tick after returning from idle, we will update the busy stats.
9263 nohz_balance_exit_idle(rq);
9266 * None are in tickless mode and hence no need for NOHZ idle load
9269 if (likely(!atomic_read(&nohz.nr_cpus)))
9272 if (READ_ONCE(nohz.has_blocked) &&
9273 time_after(now, READ_ONCE(nohz.next_blocked)))
9274 flags = NOHZ_STATS_KICK;
9276 if (time_before(now, nohz.next_balance))
9279 if (rq->nr_running >= 2) {
9280 flags = NOHZ_KICK_MASK;
9285 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
9288 * XXX: write a coherent comment on why we do this.
9289 * See also: http://lkml.kernel.org/r/20111202010832.602203411@sbsiddha-desk.sc.intel.com
9291 nr_busy = atomic_read(&sds->nr_busy_cpus);
9293 flags = NOHZ_KICK_MASK;
9299 sd = rcu_dereference(rq->sd);
9301 if ((rq->cfs.h_nr_running >= 1) &&
9302 check_cpu_capacity(rq, sd)) {
9303 flags = NOHZ_KICK_MASK;
9308 sd = rcu_dereference(per_cpu(sd_asym, cpu));
9310 for_each_cpu(i, sched_domain_span(sd)) {
9312 !cpumask_test_cpu(i, nohz.idle_cpus_mask))
9315 if (sched_asym_prefer(i, cpu)) {
9316 flags = NOHZ_KICK_MASK;
9328 static void set_cpu_sd_state_busy(int cpu)
9330 struct sched_domain *sd;
9333 sd = rcu_dereference(per_cpu(sd_llc, cpu));
9335 if (!sd || !sd->nohz_idle)
9339 atomic_inc(&sd->shared->nr_busy_cpus);
9344 void nohz_balance_exit_idle(struct rq *rq)
9346 SCHED_WARN_ON(rq != this_rq());
9348 if (likely(!rq->nohz_tick_stopped))
9351 rq->nohz_tick_stopped = 0;
9352 cpumask_clear_cpu(rq->cpu, nohz.idle_cpus_mask);
9353 atomic_dec(&nohz.nr_cpus);
9355 set_cpu_sd_state_busy(rq->cpu);
9358 static void set_cpu_sd_state_idle(int cpu)
9360 struct sched_domain *sd;
9363 sd = rcu_dereference(per_cpu(sd_llc, cpu));
9365 if (!sd || sd->nohz_idle)
9369 atomic_dec(&sd->shared->nr_busy_cpus);
9375 * This routine will record that the CPU is going idle with tick stopped.
9376 * This info will be used in performing idle load balancing in the future.
9378 void nohz_balance_enter_idle(int cpu)
9380 struct rq *rq = cpu_rq(cpu);
9382 SCHED_WARN_ON(cpu != smp_processor_id());
9384 /* If this CPU is going down, then nothing needs to be done: */
9385 if (!cpu_active(cpu))
9388 /* Spare idle load balancing on CPUs that don't want to be disturbed: */
9389 if (!housekeeping_cpu(cpu, HK_FLAG_SCHED))
9393 * Can be set safely without rq->lock held
9394 * If a clear happens, it will have evaluated last additions because
9395 * rq->lock is held during the check and the clear
9397 rq->has_blocked_load = 1;
9400 * The tick is still stopped but load could have been added in the
9401 * meantime. We set the nohz.has_blocked flag to trig a check of the
9402 * *_avg. The CPU is already part of nohz.idle_cpus_mask so the clear
9403 * of nohz.has_blocked can only happen after checking the new load
9405 if (rq->nohz_tick_stopped)
9408 /* If we're a completely isolated CPU, we don't play: */
9409 if (on_null_domain(rq))
9412 rq->nohz_tick_stopped = 1;
9414 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
9415 atomic_inc(&nohz.nr_cpus);
9418 * Ensures that if nohz_idle_balance() fails to observe our
9419 * @idle_cpus_mask store, it must observe the @has_blocked
9422 smp_mb__after_atomic();
9424 set_cpu_sd_state_idle(cpu);
9428 * Each time a cpu enter idle, we assume that it has blocked load and
9429 * enable the periodic update of the load of idle cpus
9431 WRITE_ONCE(nohz.has_blocked, 1);
9435 * Internal function that runs load balance for all idle cpus. The load balance
9436 * can be a simple update of blocked load or a complete load balance with
9437 * tasks movement depending of flags.
9438 * The function returns false if the loop has stopped before running
9439 * through all idle CPUs.
9441 static bool _nohz_idle_balance(struct rq *this_rq, unsigned int flags,
9442 enum cpu_idle_type idle)
9444 /* Earliest time when we have to do rebalance again */
9445 unsigned long now = jiffies;
9446 unsigned long next_balance = now + 60*HZ;
9447 bool has_blocked_load = false;
9448 int update_next_balance = 0;
9449 int this_cpu = this_rq->cpu;
9454 SCHED_WARN_ON((flags & NOHZ_KICK_MASK) == NOHZ_BALANCE_KICK);
9457 * We assume there will be no idle load after this update and clear
9458 * the has_blocked flag. If a cpu enters idle in the mean time, it will
9459 * set the has_blocked flag and trig another update of idle load.
9460 * Because a cpu that becomes idle, is added to idle_cpus_mask before
9461 * setting the flag, we are sure to not clear the state and not
9462 * check the load of an idle cpu.
9464 WRITE_ONCE(nohz.has_blocked, 0);
9467 * Ensures that if we miss the CPU, we must see the has_blocked
9468 * store from nohz_balance_enter_idle().
9472 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
9473 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
9477 * If this CPU gets work to do, stop the load balancing
9478 * work being done for other CPUs. Next load
9479 * balancing owner will pick it up.
9481 if (need_resched()) {
9482 has_blocked_load = true;
9486 rq = cpu_rq(balance_cpu);
9488 has_blocked_load |= update_nohz_stats(rq, true);
9491 * If time for next balance is due,
9494 if (time_after_eq(jiffies, rq->next_balance)) {
9497 rq_lock_irqsave(rq, &rf);
9498 update_rq_clock(rq);
9499 cpu_load_update_idle(rq);
9500 rq_unlock_irqrestore(rq, &rf);
9502 if (flags & NOHZ_BALANCE_KICK)
9503 rebalance_domains(rq, CPU_IDLE);
9506 if (time_after(next_balance, rq->next_balance)) {
9507 next_balance = rq->next_balance;
9508 update_next_balance = 1;
9513 * next_balance will be updated only when there is a need.
9514 * When the CPU is attached to null domain for ex, it will not be
9517 if (likely(update_next_balance))
9518 nohz.next_balance = next_balance;
9520 /* Newly idle CPU doesn't need an update */
9521 if (idle != CPU_NEWLY_IDLE) {
9522 update_blocked_averages(this_cpu);
9523 has_blocked_load |= this_rq->has_blocked_load;
9526 if (flags & NOHZ_BALANCE_KICK)
9527 rebalance_domains(this_rq, CPU_IDLE);
9529 WRITE_ONCE(nohz.next_blocked,
9530 now + msecs_to_jiffies(LOAD_AVG_PERIOD));
9532 /* The full idle balance loop has been done */
9536 /* There is still blocked load, enable periodic update */
9537 if (has_blocked_load)
9538 WRITE_ONCE(nohz.has_blocked, 1);
9544 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
9545 * rebalancing for all the cpus for whom scheduler ticks are stopped.
9547 static bool nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
9549 int this_cpu = this_rq->cpu;
9552 if (!(atomic_read(nohz_flags(this_cpu)) & NOHZ_KICK_MASK))
9555 if (idle != CPU_IDLE) {
9556 atomic_andnot(NOHZ_KICK_MASK, nohz_flags(this_cpu));
9561 * barrier, pairs with nohz_balance_enter_idle(), ensures ...
9563 flags = atomic_fetch_andnot(NOHZ_KICK_MASK, nohz_flags(this_cpu));
9564 if (!(flags & NOHZ_KICK_MASK))
9567 _nohz_idle_balance(this_rq, flags, idle);
9572 static void nohz_newidle_balance(struct rq *this_rq)
9574 int this_cpu = this_rq->cpu;
9577 * This CPU doesn't want to be disturbed by scheduler
9580 if (!housekeeping_cpu(this_cpu, HK_FLAG_SCHED))
9583 /* Will wake up very soon. No time for doing anything else*/
9584 if (this_rq->avg_idle < sysctl_sched_migration_cost)
9587 /* Don't need to update blocked load of idle CPUs*/
9588 if (!READ_ONCE(nohz.has_blocked) ||
9589 time_before(jiffies, READ_ONCE(nohz.next_blocked)))
9592 raw_spin_unlock(&this_rq->lock);
9594 * This CPU is going to be idle and blocked load of idle CPUs
9595 * need to be updated. Run the ilb locally as it is a good
9596 * candidate for ilb instead of waking up another idle CPU.
9597 * Kick an normal ilb if we failed to do the update.
9599 if (!_nohz_idle_balance(this_rq, NOHZ_STATS_KICK, CPU_NEWLY_IDLE))
9600 kick_ilb(NOHZ_STATS_KICK);
9601 raw_spin_lock(&this_rq->lock);
9604 #else /* !CONFIG_NO_HZ_COMMON */
9605 static inline void nohz_balancer_kick(struct rq *rq) { }
9607 static inline bool nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
9612 static inline void nohz_newidle_balance(struct rq *this_rq) { }
9613 #endif /* CONFIG_NO_HZ_COMMON */
9616 * idle_balance is called by schedule() if this_cpu is about to become
9617 * idle. Attempts to pull tasks from other CPUs.
9619 static int idle_balance(struct rq *this_rq, struct rq_flags *rf)
9621 unsigned long next_balance = jiffies + HZ;
9622 int this_cpu = this_rq->cpu;
9623 struct sched_domain *sd;
9624 int pulled_task = 0;
9628 * We must set idle_stamp _before_ calling idle_balance(), such that we
9629 * measure the duration of idle_balance() as idle time.
9631 this_rq->idle_stamp = rq_clock(this_rq);
9634 * Do not pull tasks towards !active CPUs...
9636 if (!cpu_active(this_cpu))
9640 * This is OK, because current is on_cpu, which avoids it being picked
9641 * for load-balance and preemption/IRQs are still disabled avoiding
9642 * further scheduler activity on it and we're being very careful to
9643 * re-start the picking loop.
9645 rq_unpin_lock(this_rq, rf);
9647 if (this_rq->avg_idle < sysctl_sched_migration_cost ||
9648 !this_rq->rd->overload) {
9651 sd = rcu_dereference_check_sched_domain(this_rq->sd);
9653 update_next_balance(sd, &next_balance);
9656 nohz_newidle_balance(this_rq);
9661 raw_spin_unlock(&this_rq->lock);
9663 update_blocked_averages(this_cpu);
9665 for_each_domain(this_cpu, sd) {
9666 int continue_balancing = 1;
9667 u64 t0, domain_cost;
9669 if (!(sd->flags & SD_LOAD_BALANCE))
9672 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost) {
9673 update_next_balance(sd, &next_balance);
9677 if (sd->flags & SD_BALANCE_NEWIDLE) {
9678 t0 = sched_clock_cpu(this_cpu);
9680 pulled_task = load_balance(this_cpu, this_rq,
9682 &continue_balancing);
9684 domain_cost = sched_clock_cpu(this_cpu) - t0;
9685 if (domain_cost > sd->max_newidle_lb_cost)
9686 sd->max_newidle_lb_cost = domain_cost;
9688 curr_cost += domain_cost;
9691 update_next_balance(sd, &next_balance);
9694 * Stop searching for tasks to pull if there are
9695 * now runnable tasks on this rq.
9697 if (pulled_task || this_rq->nr_running > 0)
9702 raw_spin_lock(&this_rq->lock);
9704 if (curr_cost > this_rq->max_idle_balance_cost)
9705 this_rq->max_idle_balance_cost = curr_cost;
9709 * While browsing the domains, we released the rq lock, a task could
9710 * have been enqueued in the meantime. Since we're not going idle,
9711 * pretend we pulled a task.
9713 if (this_rq->cfs.h_nr_running && !pulled_task)
9716 /* Move the next balance forward */
9717 if (time_after(this_rq->next_balance, next_balance))
9718 this_rq->next_balance = next_balance;
9720 /* Is there a task of a high priority class? */
9721 if (this_rq->nr_running != this_rq->cfs.h_nr_running)
9725 this_rq->idle_stamp = 0;
9727 rq_repin_lock(this_rq, rf);
9733 * run_rebalance_domains is triggered when needed from the scheduler tick.
9734 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
9736 static __latent_entropy void run_rebalance_domains(struct softirq_action *h)
9738 struct rq *this_rq = this_rq();
9739 enum cpu_idle_type idle = this_rq->idle_balance ?
9740 CPU_IDLE : CPU_NOT_IDLE;
9743 * If this CPU has a pending nohz_balance_kick, then do the
9744 * balancing on behalf of the other idle CPUs whose ticks are
9745 * stopped. Do nohz_idle_balance *before* rebalance_domains to
9746 * give the idle CPUs a chance to load balance. Else we may
9747 * load balance only within the local sched_domain hierarchy
9748 * and abort nohz_idle_balance altogether if we pull some load.
9750 if (nohz_idle_balance(this_rq, idle))
9753 /* normal load balance */
9754 update_blocked_averages(this_rq->cpu);
9755 rebalance_domains(this_rq, idle);
9759 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
9761 void trigger_load_balance(struct rq *rq)
9763 /* Don't need to rebalance while attached to NULL domain */
9764 if (unlikely(on_null_domain(rq)))
9767 if (time_after_eq(jiffies, rq->next_balance))
9768 raise_softirq(SCHED_SOFTIRQ);
9770 nohz_balancer_kick(rq);
9773 static void rq_online_fair(struct rq *rq)
9777 update_runtime_enabled(rq);
9780 static void rq_offline_fair(struct rq *rq)
9784 /* Ensure any throttled groups are reachable by pick_next_task */
9785 unthrottle_offline_cfs_rqs(rq);
9788 #endif /* CONFIG_SMP */
9791 * scheduler tick hitting a task of our scheduling class.
9793 * NOTE: This function can be called remotely by the tick offload that
9794 * goes along full dynticks. Therefore no local assumption can be made
9795 * and everything must be accessed through the @rq and @curr passed in
9798 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
9800 struct cfs_rq *cfs_rq;
9801 struct sched_entity *se = &curr->se;
9803 for_each_sched_entity(se) {
9804 cfs_rq = cfs_rq_of(se);
9805 entity_tick(cfs_rq, se, queued);
9808 if (static_branch_unlikely(&sched_numa_balancing))
9809 task_tick_numa(rq, curr);
9813 * called on fork with the child task as argument from the parent's context
9814 * - child not yet on the tasklist
9815 * - preemption disabled
9817 static void task_fork_fair(struct task_struct *p)
9819 struct cfs_rq *cfs_rq;
9820 struct sched_entity *se = &p->se, *curr;
9821 struct rq *rq = this_rq();
9825 update_rq_clock(rq);
9827 cfs_rq = task_cfs_rq(current);
9828 curr = cfs_rq->curr;
9830 update_curr(cfs_rq);
9831 se->vruntime = curr->vruntime;
9833 place_entity(cfs_rq, se, 1);
9835 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
9837 * Upon rescheduling, sched_class::put_prev_task() will place
9838 * 'current' within the tree based on its new key value.
9840 swap(curr->vruntime, se->vruntime);
9844 se->vruntime -= cfs_rq->min_vruntime;
9849 * Priority of the task has changed. Check to see if we preempt
9853 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
9855 if (!task_on_rq_queued(p))
9859 * Reschedule if we are currently running on this runqueue and
9860 * our priority decreased, or if we are not currently running on
9861 * this runqueue and our priority is higher than the current's
9863 if (rq->curr == p) {
9864 if (p->prio > oldprio)
9867 check_preempt_curr(rq, p, 0);
9870 static inline bool vruntime_normalized(struct task_struct *p)
9872 struct sched_entity *se = &p->se;
9875 * In both the TASK_ON_RQ_QUEUED and TASK_ON_RQ_MIGRATING cases,
9876 * the dequeue_entity(.flags=0) will already have normalized the
9883 * When !on_rq, vruntime of the task has usually NOT been normalized.
9884 * But there are some cases where it has already been normalized:
9886 * - A forked child which is waiting for being woken up by
9887 * wake_up_new_task().
9888 * - A task which has been woken up by try_to_wake_up() and
9889 * waiting for actually being woken up by sched_ttwu_pending().
9891 if (!se->sum_exec_runtime ||
9892 (p->state == TASK_WAKING && p->sched_remote_wakeup))
9898 #ifdef CONFIG_FAIR_GROUP_SCHED
9900 * Propagate the changes of the sched_entity across the tg tree to make it
9901 * visible to the root
9903 static void propagate_entity_cfs_rq(struct sched_entity *se)
9905 struct cfs_rq *cfs_rq;
9907 list_add_leaf_cfs_rq(cfs_rq_of(se));
9909 /* Start to propagate at parent */
9912 for_each_sched_entity(se) {
9913 cfs_rq = cfs_rq_of(se);
9915 if (!cfs_rq_throttled(cfs_rq)){
9916 update_load_avg(cfs_rq, se, UPDATE_TG);
9917 list_add_leaf_cfs_rq(cfs_rq);
9921 if (list_add_leaf_cfs_rq(cfs_rq))
9926 static void propagate_entity_cfs_rq(struct sched_entity *se) { }
9929 static void detach_entity_cfs_rq(struct sched_entity *se)
9931 struct cfs_rq *cfs_rq = cfs_rq_of(se);
9933 /* Catch up with the cfs_rq and remove our load when we leave */
9934 update_load_avg(cfs_rq, se, 0);
9935 detach_entity_load_avg(cfs_rq, se);
9936 update_tg_load_avg(cfs_rq, false);
9937 propagate_entity_cfs_rq(se);
9940 static void attach_entity_cfs_rq(struct sched_entity *se)
9942 struct cfs_rq *cfs_rq = cfs_rq_of(se);
9944 #ifdef CONFIG_FAIR_GROUP_SCHED
9946 * Since the real-depth could have been changed (only FAIR
9947 * class maintain depth value), reset depth properly.
9949 se->depth = se->parent ? se->parent->depth + 1 : 0;
9952 /* Synchronize entity with its cfs_rq */
9953 update_load_avg(cfs_rq, se, sched_feat(ATTACH_AGE_LOAD) ? 0 : SKIP_AGE_LOAD);
9954 attach_entity_load_avg(cfs_rq, se, 0);
9955 update_tg_load_avg(cfs_rq, false);
9956 propagate_entity_cfs_rq(se);
9959 static void detach_task_cfs_rq(struct task_struct *p)
9961 struct sched_entity *se = &p->se;
9962 struct cfs_rq *cfs_rq = cfs_rq_of(se);
9964 if (!vruntime_normalized(p)) {
9966 * Fix up our vruntime so that the current sleep doesn't
9967 * cause 'unlimited' sleep bonus.
9969 place_entity(cfs_rq, se, 0);
9970 se->vruntime -= cfs_rq->min_vruntime;
9973 detach_entity_cfs_rq(se);
9976 static void attach_task_cfs_rq(struct task_struct *p)
9978 struct sched_entity *se = &p->se;
9979 struct cfs_rq *cfs_rq = cfs_rq_of(se);
9981 attach_entity_cfs_rq(se);
9983 if (!vruntime_normalized(p))
9984 se->vruntime += cfs_rq->min_vruntime;
9987 static void switched_from_fair(struct rq *rq, struct task_struct *p)
9989 detach_task_cfs_rq(p);
9992 static void switched_to_fair(struct rq *rq, struct task_struct *p)
9994 attach_task_cfs_rq(p);
9996 if (task_on_rq_queued(p)) {
9998 * We were most likely switched from sched_rt, so
9999 * kick off the schedule if running, otherwise just see
10000 * if we can still preempt the current task.
10005 check_preempt_curr(rq, p, 0);
10009 /* Account for a task changing its policy or group.
10011 * This routine is mostly called to set cfs_rq->curr field when a task
10012 * migrates between groups/classes.
10014 static void set_curr_task_fair(struct rq *rq)
10016 struct sched_entity *se = &rq->curr->se;
10018 for_each_sched_entity(se) {
10019 struct cfs_rq *cfs_rq = cfs_rq_of(se);
10021 set_next_entity(cfs_rq, se);
10022 /* ensure bandwidth has been allocated on our new cfs_rq */
10023 account_cfs_rq_runtime(cfs_rq, 0);
10027 void init_cfs_rq(struct cfs_rq *cfs_rq)
10029 cfs_rq->tasks_timeline = RB_ROOT_CACHED;
10030 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
10031 #ifndef CONFIG_64BIT
10032 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
10035 raw_spin_lock_init(&cfs_rq->removed.lock);
10039 #ifdef CONFIG_FAIR_GROUP_SCHED
10040 static void task_set_group_fair(struct task_struct *p)
10042 struct sched_entity *se = &p->se;
10044 set_task_rq(p, task_cpu(p));
10045 se->depth = se->parent ? se->parent->depth + 1 : 0;
10048 static void task_move_group_fair(struct task_struct *p)
10050 detach_task_cfs_rq(p);
10051 set_task_rq(p, task_cpu(p));
10054 /* Tell se's cfs_rq has been changed -- migrated */
10055 p->se.avg.last_update_time = 0;
10057 attach_task_cfs_rq(p);
10060 static void task_change_group_fair(struct task_struct *p, int type)
10063 case TASK_SET_GROUP:
10064 task_set_group_fair(p);
10067 case TASK_MOVE_GROUP:
10068 task_move_group_fair(p);
10073 void free_fair_sched_group(struct task_group *tg)
10077 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
10079 for_each_possible_cpu(i) {
10081 kfree(tg->cfs_rq[i]);
10090 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
10092 struct sched_entity *se;
10093 struct cfs_rq *cfs_rq;
10096 tg->cfs_rq = kcalloc(nr_cpu_ids, sizeof(cfs_rq), GFP_KERNEL);
10099 tg->se = kcalloc(nr_cpu_ids, sizeof(se), GFP_KERNEL);
10103 tg->shares = NICE_0_LOAD;
10105 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
10107 for_each_possible_cpu(i) {
10108 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
10109 GFP_KERNEL, cpu_to_node(i));
10113 se = kzalloc_node(sizeof(struct sched_entity),
10114 GFP_KERNEL, cpu_to_node(i));
10118 init_cfs_rq(cfs_rq);
10119 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
10120 init_entity_runnable_average(se);
10131 void online_fair_sched_group(struct task_group *tg)
10133 struct sched_entity *se;
10134 struct rq_flags rf;
10138 for_each_possible_cpu(i) {
10141 rq_lock_irq(rq, &rf);
10142 update_rq_clock(rq);
10143 attach_entity_cfs_rq(se);
10144 sync_throttle(tg, i);
10145 rq_unlock_irq(rq, &rf);
10149 void unregister_fair_sched_group(struct task_group *tg)
10151 unsigned long flags;
10155 for_each_possible_cpu(cpu) {
10157 remove_entity_load_avg(tg->se[cpu]);
10160 * Only empty task groups can be destroyed; so we can speculatively
10161 * check on_list without danger of it being re-added.
10163 if (!tg->cfs_rq[cpu]->on_list)
10168 raw_spin_lock_irqsave(&rq->lock, flags);
10169 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
10170 raw_spin_unlock_irqrestore(&rq->lock, flags);
10174 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
10175 struct sched_entity *se, int cpu,
10176 struct sched_entity *parent)
10178 struct rq *rq = cpu_rq(cpu);
10182 init_cfs_rq_runtime(cfs_rq);
10184 tg->cfs_rq[cpu] = cfs_rq;
10187 /* se could be NULL for root_task_group */
10192 se->cfs_rq = &rq->cfs;
10195 se->cfs_rq = parent->my_q;
10196 se->depth = parent->depth + 1;
10200 /* guarantee group entities always have weight */
10201 update_load_set(&se->load, NICE_0_LOAD);
10202 se->parent = parent;
10205 static DEFINE_MUTEX(shares_mutex);
10207 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
10212 * We can't change the weight of the root cgroup.
10217 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
10219 mutex_lock(&shares_mutex);
10220 if (tg->shares == shares)
10223 tg->shares = shares;
10224 for_each_possible_cpu(i) {
10225 struct rq *rq = cpu_rq(i);
10226 struct sched_entity *se = tg->se[i];
10227 struct rq_flags rf;
10229 /* Propagate contribution to hierarchy */
10230 rq_lock_irqsave(rq, &rf);
10231 update_rq_clock(rq);
10232 for_each_sched_entity(se) {
10233 update_load_avg(cfs_rq_of(se), se, UPDATE_TG);
10234 update_cfs_group(se);
10236 rq_unlock_irqrestore(rq, &rf);
10240 mutex_unlock(&shares_mutex);
10243 #else /* CONFIG_FAIR_GROUP_SCHED */
10245 void free_fair_sched_group(struct task_group *tg) { }
10247 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
10252 void online_fair_sched_group(struct task_group *tg) { }
10254 void unregister_fair_sched_group(struct task_group *tg) { }
10256 #endif /* CONFIG_FAIR_GROUP_SCHED */
10259 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
10261 struct sched_entity *se = &task->se;
10262 unsigned int rr_interval = 0;
10265 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
10268 if (rq->cfs.load.weight)
10269 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
10271 return rr_interval;
10275 * All the scheduling class methods:
10277 const struct sched_class fair_sched_class = {
10278 .next = &idle_sched_class,
10279 .enqueue_task = enqueue_task_fair,
10280 .dequeue_task = dequeue_task_fair,
10281 .yield_task = yield_task_fair,
10282 .yield_to_task = yield_to_task_fair,
10284 .check_preempt_curr = check_preempt_wakeup,
10286 .pick_next_task = pick_next_task_fair,
10287 .put_prev_task = put_prev_task_fair,
10290 .select_task_rq = select_task_rq_fair,
10291 .migrate_task_rq = migrate_task_rq_fair,
10293 .rq_online = rq_online_fair,
10294 .rq_offline = rq_offline_fair,
10296 .task_dead = task_dead_fair,
10297 .set_cpus_allowed = set_cpus_allowed_common,
10300 .set_curr_task = set_curr_task_fair,
10301 .task_tick = task_tick_fair,
10302 .task_fork = task_fork_fair,
10304 .prio_changed = prio_changed_fair,
10305 .switched_from = switched_from_fair,
10306 .switched_to = switched_to_fair,
10308 .get_rr_interval = get_rr_interval_fair,
10310 .update_curr = update_curr_fair,
10312 #ifdef CONFIG_FAIR_GROUP_SCHED
10313 .task_change_group = task_change_group_fair,
10317 #ifdef CONFIG_SCHED_DEBUG
10318 void print_cfs_stats(struct seq_file *m, int cpu)
10320 struct cfs_rq *cfs_rq, *pos;
10323 for_each_leaf_cfs_rq_safe(cpu_rq(cpu), cfs_rq, pos)
10324 print_cfs_rq(m, cpu, cfs_rq);
10328 #ifdef CONFIG_NUMA_BALANCING
10329 void show_numa_stats(struct task_struct *p, struct seq_file *m)
10332 unsigned long tsf = 0, tpf = 0, gsf = 0, gpf = 0;
10333 struct numa_group *ng;
10336 ng = rcu_dereference(p->numa_group);
10337 for_each_online_node(node) {
10338 if (p->numa_faults) {
10339 tsf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 0)];
10340 tpf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 1)];
10343 gsf = ng->faults[task_faults_idx(NUMA_MEM, node, 0)],
10344 gpf = ng->faults[task_faults_idx(NUMA_MEM, node, 1)];
10346 print_numa_stats(m, node, tsf, tpf, gsf, gpf);
10350 #endif /* CONFIG_NUMA_BALANCING */
10351 #endif /* CONFIG_SCHED_DEBUG */
10353 __init void init_sched_fair_class(void)
10356 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
10358 #ifdef CONFIG_NO_HZ_COMMON
10359 nohz.next_balance = jiffies;
10360 nohz.next_blocked = jiffies;
10361 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);