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 static 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 static 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 static unsigned int normalized_sysctl_sched_wakeup_granularity = 1000000UL;
87 const_debug unsigned int sysctl_sched_migration_cost = 500000UL;
89 int sched_thermal_decay_shift;
90 static int __init setup_sched_thermal_decay_shift(char *str)
94 if (kstrtoint(str, 0, &_shift))
95 pr_warn("Unable to set scheduler thermal pressure decay shift parameter\n");
97 sched_thermal_decay_shift = clamp(_shift, 0, 10);
100 __setup("sched_thermal_decay_shift=", setup_sched_thermal_decay_shift);
104 * For asym packing, by default the lower numbered CPU has higher priority.
106 int __weak arch_asym_cpu_priority(int cpu)
112 * The margin used when comparing utilization with CPU capacity.
116 #define fits_capacity(cap, max) ((cap) * 1280 < (max) * 1024)
120 #ifdef CONFIG_CFS_BANDWIDTH
122 * Amount of runtime to allocate from global (tg) to local (per-cfs_rq) pool
123 * each time a cfs_rq requests quota.
125 * Note: in the case that the slice exceeds the runtime remaining (either due
126 * to consumption or the quota being specified to be smaller than the slice)
127 * we will always only issue the remaining available time.
129 * (default: 5 msec, units: microseconds)
131 unsigned int sysctl_sched_cfs_bandwidth_slice = 5000UL;
134 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
140 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
146 static inline void update_load_set(struct load_weight *lw, unsigned long w)
153 * Increase the granularity value when there are more CPUs,
154 * because with more CPUs the 'effective latency' as visible
155 * to users decreases. But the relationship is not linear,
156 * so pick a second-best guess by going with the log2 of the
159 * This idea comes from the SD scheduler of Con Kolivas:
161 static unsigned int get_update_sysctl_factor(void)
163 unsigned int cpus = min_t(unsigned int, num_online_cpus(), 8);
166 switch (sysctl_sched_tunable_scaling) {
167 case SCHED_TUNABLESCALING_NONE:
170 case SCHED_TUNABLESCALING_LINEAR:
173 case SCHED_TUNABLESCALING_LOG:
175 factor = 1 + ilog2(cpus);
182 static void update_sysctl(void)
184 unsigned int factor = get_update_sysctl_factor();
186 #define SET_SYSCTL(name) \
187 (sysctl_##name = (factor) * normalized_sysctl_##name)
188 SET_SYSCTL(sched_min_granularity);
189 SET_SYSCTL(sched_latency);
190 SET_SYSCTL(sched_wakeup_granularity);
194 void sched_init_granularity(void)
199 #define WMULT_CONST (~0U)
200 #define WMULT_SHIFT 32
202 static void __update_inv_weight(struct load_weight *lw)
206 if (likely(lw->inv_weight))
209 w = scale_load_down(lw->weight);
211 if (BITS_PER_LONG > 32 && unlikely(w >= WMULT_CONST))
213 else if (unlikely(!w))
214 lw->inv_weight = WMULT_CONST;
216 lw->inv_weight = WMULT_CONST / w;
220 * delta_exec * weight / lw.weight
222 * (delta_exec * (weight * lw->inv_weight)) >> WMULT_SHIFT
224 * Either weight := NICE_0_LOAD and lw \e sched_prio_to_wmult[], in which case
225 * we're guaranteed shift stays positive because inv_weight is guaranteed to
226 * fit 32 bits, and NICE_0_LOAD gives another 10 bits; therefore shift >= 22.
228 * Or, weight =< lw.weight (because lw.weight is the runqueue weight), thus
229 * weight/lw.weight <= 1, and therefore our shift will also be positive.
231 static u64 __calc_delta(u64 delta_exec, unsigned long weight, struct load_weight *lw)
233 u64 fact = scale_load_down(weight);
234 int shift = WMULT_SHIFT;
236 __update_inv_weight(lw);
238 if (unlikely(fact >> 32)) {
245 fact = mul_u32_u32(fact, lw->inv_weight);
252 return mul_u64_u32_shr(delta_exec, fact, shift);
256 const struct sched_class fair_sched_class;
258 /**************************************************************
259 * CFS operations on generic schedulable entities:
262 #ifdef CONFIG_FAIR_GROUP_SCHED
263 static inline struct task_struct *task_of(struct sched_entity *se)
265 SCHED_WARN_ON(!entity_is_task(se));
266 return container_of(se, struct task_struct, se);
269 /* Walk up scheduling entities hierarchy */
270 #define for_each_sched_entity(se) \
271 for (; se; se = se->parent)
273 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
278 /* runqueue on which this entity is (to be) queued */
279 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
284 /* runqueue "owned" by this group */
285 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
290 static inline void cfs_rq_tg_path(struct cfs_rq *cfs_rq, char *path, int len)
295 if (cfs_rq && task_group_is_autogroup(cfs_rq->tg))
296 autogroup_path(cfs_rq->tg, path, len);
297 else if (cfs_rq && cfs_rq->tg->css.cgroup)
298 cgroup_path(cfs_rq->tg->css.cgroup, path, len);
300 strlcpy(path, "(null)", len);
303 static inline bool list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
305 struct rq *rq = rq_of(cfs_rq);
306 int cpu = cpu_of(rq);
309 return rq->tmp_alone_branch == &rq->leaf_cfs_rq_list;
314 * Ensure we either appear before our parent (if already
315 * enqueued) or force our parent to appear after us when it is
316 * enqueued. The fact that we always enqueue bottom-up
317 * reduces this to two cases and a special case for the root
318 * cfs_rq. Furthermore, it also means that we will always reset
319 * tmp_alone_branch either when the branch is connected
320 * to a tree or when we reach the top of the tree
322 if (cfs_rq->tg->parent &&
323 cfs_rq->tg->parent->cfs_rq[cpu]->on_list) {
325 * If parent is already on the list, we add the child
326 * just before. Thanks to circular linked property of
327 * the list, this means to put the child at the tail
328 * of the list that starts by parent.
330 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
331 &(cfs_rq->tg->parent->cfs_rq[cpu]->leaf_cfs_rq_list));
333 * The branch is now connected to its tree so we can
334 * reset tmp_alone_branch to the beginning of the
337 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
341 if (!cfs_rq->tg->parent) {
343 * cfs rq without parent should be put
344 * at the tail of the list.
346 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
347 &rq->leaf_cfs_rq_list);
349 * We have reach the top of a tree so we can reset
350 * tmp_alone_branch to the beginning of the list.
352 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
357 * The parent has not already been added so we want to
358 * make sure that it will be put after us.
359 * tmp_alone_branch points to the begin of the branch
360 * where we will add parent.
362 list_add_rcu(&cfs_rq->leaf_cfs_rq_list, rq->tmp_alone_branch);
364 * update tmp_alone_branch to points to the new begin
367 rq->tmp_alone_branch = &cfs_rq->leaf_cfs_rq_list;
371 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
373 if (cfs_rq->on_list) {
374 struct rq *rq = rq_of(cfs_rq);
377 * With cfs_rq being unthrottled/throttled during an enqueue,
378 * it can happen the tmp_alone_branch points the a leaf that
379 * we finally want to del. In this case, tmp_alone_branch moves
380 * to the prev element but it will point to rq->leaf_cfs_rq_list
381 * at the end of the enqueue.
383 if (rq->tmp_alone_branch == &cfs_rq->leaf_cfs_rq_list)
384 rq->tmp_alone_branch = cfs_rq->leaf_cfs_rq_list.prev;
386 list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
391 static inline void assert_list_leaf_cfs_rq(struct rq *rq)
393 SCHED_WARN_ON(rq->tmp_alone_branch != &rq->leaf_cfs_rq_list);
396 /* Iterate thr' all leaf cfs_rq's on a runqueue */
397 #define for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) \
398 list_for_each_entry_safe(cfs_rq, pos, &rq->leaf_cfs_rq_list, \
401 /* Do the two (enqueued) entities belong to the same group ? */
402 static inline struct cfs_rq *
403 is_same_group(struct sched_entity *se, struct sched_entity *pse)
405 if (se->cfs_rq == pse->cfs_rq)
411 static inline struct sched_entity *parent_entity(struct sched_entity *se)
417 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
419 int se_depth, pse_depth;
422 * preemption test can be made between sibling entities who are in the
423 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
424 * both tasks until we find their ancestors who are siblings of common
428 /* First walk up until both entities are at same depth */
429 se_depth = (*se)->depth;
430 pse_depth = (*pse)->depth;
432 while (se_depth > pse_depth) {
434 *se = parent_entity(*se);
437 while (pse_depth > se_depth) {
439 *pse = parent_entity(*pse);
442 while (!is_same_group(*se, *pse)) {
443 *se = parent_entity(*se);
444 *pse = parent_entity(*pse);
448 #else /* !CONFIG_FAIR_GROUP_SCHED */
450 static inline struct task_struct *task_of(struct sched_entity *se)
452 return container_of(se, struct task_struct, se);
455 #define for_each_sched_entity(se) \
456 for (; se; se = NULL)
458 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
460 return &task_rq(p)->cfs;
463 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
465 struct task_struct *p = task_of(se);
466 struct rq *rq = task_rq(p);
471 /* runqueue "owned" by this group */
472 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
477 static inline void cfs_rq_tg_path(struct cfs_rq *cfs_rq, char *path, int len)
480 strlcpy(path, "(null)", len);
483 static inline bool list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
488 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
492 static inline void assert_list_leaf_cfs_rq(struct rq *rq)
496 #define for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) \
497 for (cfs_rq = &rq->cfs, pos = NULL; cfs_rq; cfs_rq = pos)
499 static inline struct sched_entity *parent_entity(struct sched_entity *se)
505 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
509 #endif /* CONFIG_FAIR_GROUP_SCHED */
511 static __always_inline
512 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec);
514 /**************************************************************
515 * Scheduling class tree data structure manipulation methods:
518 static inline u64 max_vruntime(u64 max_vruntime, u64 vruntime)
520 s64 delta = (s64)(vruntime - max_vruntime);
522 max_vruntime = vruntime;
527 static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime)
529 s64 delta = (s64)(vruntime - min_vruntime);
531 min_vruntime = vruntime;
536 static inline int entity_before(struct sched_entity *a,
537 struct sched_entity *b)
539 return (s64)(a->vruntime - b->vruntime) < 0;
542 static void update_min_vruntime(struct cfs_rq *cfs_rq)
544 struct sched_entity *curr = cfs_rq->curr;
545 struct rb_node *leftmost = rb_first_cached(&cfs_rq->tasks_timeline);
547 u64 vruntime = cfs_rq->min_vruntime;
551 vruntime = curr->vruntime;
556 if (leftmost) { /* non-empty tree */
557 struct sched_entity *se;
558 se = rb_entry(leftmost, struct sched_entity, run_node);
561 vruntime = se->vruntime;
563 vruntime = min_vruntime(vruntime, se->vruntime);
566 /* ensure we never gain time by being placed backwards. */
567 cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime);
570 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
575 * Enqueue an entity into the rb-tree:
577 static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
579 struct rb_node **link = &cfs_rq->tasks_timeline.rb_root.rb_node;
580 struct rb_node *parent = NULL;
581 struct sched_entity *entry;
582 bool leftmost = true;
585 * Find the right place in the rbtree:
589 entry = rb_entry(parent, struct sched_entity, run_node);
591 * We dont care about collisions. Nodes with
592 * the same key stay together.
594 if (entity_before(se, entry)) {
595 link = &parent->rb_left;
597 link = &parent->rb_right;
602 rb_link_node(&se->run_node, parent, link);
603 rb_insert_color_cached(&se->run_node,
604 &cfs_rq->tasks_timeline, leftmost);
607 static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
609 rb_erase_cached(&se->run_node, &cfs_rq->tasks_timeline);
612 struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq)
614 struct rb_node *left = rb_first_cached(&cfs_rq->tasks_timeline);
619 return rb_entry(left, struct sched_entity, run_node);
622 static struct sched_entity *__pick_next_entity(struct sched_entity *se)
624 struct rb_node *next = rb_next(&se->run_node);
629 return rb_entry(next, struct sched_entity, run_node);
632 #ifdef CONFIG_SCHED_DEBUG
633 struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
635 struct rb_node *last = rb_last(&cfs_rq->tasks_timeline.rb_root);
640 return rb_entry(last, struct sched_entity, run_node);
643 /**************************************************************
644 * Scheduling class statistics methods:
647 int sched_proc_update_handler(struct ctl_table *table, int write,
648 void __user *buffer, size_t *lenp,
651 int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
652 unsigned int factor = get_update_sysctl_factor();
657 sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency,
658 sysctl_sched_min_granularity);
660 #define WRT_SYSCTL(name) \
661 (normalized_sysctl_##name = sysctl_##name / (factor))
662 WRT_SYSCTL(sched_min_granularity);
663 WRT_SYSCTL(sched_latency);
664 WRT_SYSCTL(sched_wakeup_granularity);
674 static inline u64 calc_delta_fair(u64 delta, struct sched_entity *se)
676 if (unlikely(se->load.weight != NICE_0_LOAD))
677 delta = __calc_delta(delta, NICE_0_LOAD, &se->load);
683 * The idea is to set a period in which each task runs once.
685 * When there are too many tasks (sched_nr_latency) we have to stretch
686 * this period because otherwise the slices get too small.
688 * p = (nr <= nl) ? l : l*nr/nl
690 static u64 __sched_period(unsigned long nr_running)
692 if (unlikely(nr_running > sched_nr_latency))
693 return nr_running * sysctl_sched_min_granularity;
695 return sysctl_sched_latency;
699 * We calculate the wall-time slice from the period by taking a part
700 * proportional to the weight.
704 static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
706 u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq);
708 for_each_sched_entity(se) {
709 struct load_weight *load;
710 struct load_weight lw;
712 cfs_rq = cfs_rq_of(se);
713 load = &cfs_rq->load;
715 if (unlikely(!se->on_rq)) {
718 update_load_add(&lw, se->load.weight);
721 slice = __calc_delta(slice, se->load.weight, load);
727 * We calculate the vruntime slice of a to-be-inserted task.
731 static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se)
733 return calc_delta_fair(sched_slice(cfs_rq, se), se);
739 static int select_idle_sibling(struct task_struct *p, int prev_cpu, int cpu);
740 static unsigned long task_h_load(struct task_struct *p);
741 static unsigned long capacity_of(int cpu);
743 /* Give new sched_entity start runnable values to heavy its load in infant time */
744 void init_entity_runnable_average(struct sched_entity *se)
746 struct sched_avg *sa = &se->avg;
748 memset(sa, 0, sizeof(*sa));
751 * Tasks are initialized with full load to be seen as heavy tasks until
752 * they get a chance to stabilize to their real load level.
753 * Group entities are initialized with zero load to reflect the fact that
754 * nothing has been attached to the task group yet.
756 if (entity_is_task(se))
757 sa->load_avg = scale_load_down(se->load.weight);
759 /* when this task enqueue'ed, it will contribute to its cfs_rq's load_avg */
762 static void attach_entity_cfs_rq(struct sched_entity *se);
765 * With new tasks being created, their initial util_avgs are extrapolated
766 * based on the cfs_rq's current util_avg:
768 * util_avg = cfs_rq->util_avg / (cfs_rq->load_avg + 1) * se.load.weight
770 * However, in many cases, the above util_avg does not give a desired
771 * value. Moreover, the sum of the util_avgs may be divergent, such
772 * as when the series is a harmonic series.
774 * To solve this problem, we also cap the util_avg of successive tasks to
775 * only 1/2 of the left utilization budget:
777 * util_avg_cap = (cpu_scale - cfs_rq->avg.util_avg) / 2^n
779 * where n denotes the nth task and cpu_scale the CPU capacity.
781 * For example, for a CPU with 1024 of capacity, a simplest series from
782 * the beginning would be like:
784 * task util_avg: 512, 256, 128, 64, 32, 16, 8, ...
785 * cfs_rq util_avg: 512, 768, 896, 960, 992, 1008, 1016, ...
787 * Finally, that extrapolated util_avg is clamped to the cap (util_avg_cap)
788 * if util_avg > util_avg_cap.
790 void post_init_entity_util_avg(struct task_struct *p)
792 struct sched_entity *se = &p->se;
793 struct cfs_rq *cfs_rq = cfs_rq_of(se);
794 struct sched_avg *sa = &se->avg;
795 long cpu_scale = arch_scale_cpu_capacity(cpu_of(rq_of(cfs_rq)));
796 long cap = (long)(cpu_scale - cfs_rq->avg.util_avg) / 2;
799 if (cfs_rq->avg.util_avg != 0) {
800 sa->util_avg = cfs_rq->avg.util_avg * se->load.weight;
801 sa->util_avg /= (cfs_rq->avg.load_avg + 1);
803 if (sa->util_avg > cap)
810 sa->runnable_avg = cpu_scale;
812 if (p->sched_class != &fair_sched_class) {
814 * For !fair tasks do:
816 update_cfs_rq_load_avg(now, cfs_rq);
817 attach_entity_load_avg(cfs_rq, se);
818 switched_from_fair(rq, p);
820 * such that the next switched_to_fair() has the
823 se->avg.last_update_time = cfs_rq_clock_pelt(cfs_rq);
827 attach_entity_cfs_rq(se);
830 #else /* !CONFIG_SMP */
831 void init_entity_runnable_average(struct sched_entity *se)
834 void post_init_entity_util_avg(struct task_struct *p)
837 static void update_tg_load_avg(struct cfs_rq *cfs_rq, int force)
840 #endif /* CONFIG_SMP */
843 * Update the current task's runtime statistics.
845 static void update_curr(struct cfs_rq *cfs_rq)
847 struct sched_entity *curr = cfs_rq->curr;
848 u64 now = rq_clock_task(rq_of(cfs_rq));
854 delta_exec = now - curr->exec_start;
855 if (unlikely((s64)delta_exec <= 0))
858 curr->exec_start = now;
860 schedstat_set(curr->statistics.exec_max,
861 max(delta_exec, curr->statistics.exec_max));
863 curr->sum_exec_runtime += delta_exec;
864 schedstat_add(cfs_rq->exec_clock, delta_exec);
866 curr->vruntime += calc_delta_fair(delta_exec, curr);
867 update_min_vruntime(cfs_rq);
869 if (entity_is_task(curr)) {
870 struct task_struct *curtask = task_of(curr);
872 trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime);
873 cgroup_account_cputime(curtask, delta_exec);
874 account_group_exec_runtime(curtask, delta_exec);
877 account_cfs_rq_runtime(cfs_rq, delta_exec);
880 static void update_curr_fair(struct rq *rq)
882 update_curr(cfs_rq_of(&rq->curr->se));
886 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
888 u64 wait_start, prev_wait_start;
890 if (!schedstat_enabled())
893 wait_start = rq_clock(rq_of(cfs_rq));
894 prev_wait_start = schedstat_val(se->statistics.wait_start);
896 if (entity_is_task(se) && task_on_rq_migrating(task_of(se)) &&
897 likely(wait_start > prev_wait_start))
898 wait_start -= prev_wait_start;
900 __schedstat_set(se->statistics.wait_start, wait_start);
904 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
906 struct task_struct *p;
909 if (!schedstat_enabled())
912 delta = rq_clock(rq_of(cfs_rq)) - schedstat_val(se->statistics.wait_start);
914 if (entity_is_task(se)) {
916 if (task_on_rq_migrating(p)) {
918 * Preserve migrating task's wait time so wait_start
919 * time stamp can be adjusted to accumulate wait time
920 * prior to migration.
922 __schedstat_set(se->statistics.wait_start, delta);
925 trace_sched_stat_wait(p, delta);
928 __schedstat_set(se->statistics.wait_max,
929 max(schedstat_val(se->statistics.wait_max), delta));
930 __schedstat_inc(se->statistics.wait_count);
931 __schedstat_add(se->statistics.wait_sum, delta);
932 __schedstat_set(se->statistics.wait_start, 0);
936 update_stats_enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
938 struct task_struct *tsk = NULL;
939 u64 sleep_start, block_start;
941 if (!schedstat_enabled())
944 sleep_start = schedstat_val(se->statistics.sleep_start);
945 block_start = schedstat_val(se->statistics.block_start);
947 if (entity_is_task(se))
951 u64 delta = rq_clock(rq_of(cfs_rq)) - sleep_start;
956 if (unlikely(delta > schedstat_val(se->statistics.sleep_max)))
957 __schedstat_set(se->statistics.sleep_max, delta);
959 __schedstat_set(se->statistics.sleep_start, 0);
960 __schedstat_add(se->statistics.sum_sleep_runtime, delta);
963 account_scheduler_latency(tsk, delta >> 10, 1);
964 trace_sched_stat_sleep(tsk, delta);
968 u64 delta = rq_clock(rq_of(cfs_rq)) - block_start;
973 if (unlikely(delta > schedstat_val(se->statistics.block_max)))
974 __schedstat_set(se->statistics.block_max, delta);
976 __schedstat_set(se->statistics.block_start, 0);
977 __schedstat_add(se->statistics.sum_sleep_runtime, delta);
980 if (tsk->in_iowait) {
981 __schedstat_add(se->statistics.iowait_sum, delta);
982 __schedstat_inc(se->statistics.iowait_count);
983 trace_sched_stat_iowait(tsk, delta);
986 trace_sched_stat_blocked(tsk, delta);
989 * Blocking time is in units of nanosecs, so shift by
990 * 20 to get a milliseconds-range estimation of the
991 * amount of time that the task spent sleeping:
993 if (unlikely(prof_on == SLEEP_PROFILING)) {
994 profile_hits(SLEEP_PROFILING,
995 (void *)get_wchan(tsk),
998 account_scheduler_latency(tsk, delta >> 10, 0);
1004 * Task is being enqueued - update stats:
1007 update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
1009 if (!schedstat_enabled())
1013 * Are we enqueueing a waiting task? (for current tasks
1014 * a dequeue/enqueue event is a NOP)
1016 if (se != cfs_rq->curr)
1017 update_stats_wait_start(cfs_rq, se);
1019 if (flags & ENQUEUE_WAKEUP)
1020 update_stats_enqueue_sleeper(cfs_rq, se);
1024 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
1027 if (!schedstat_enabled())
1031 * Mark the end of the wait period if dequeueing a
1034 if (se != cfs_rq->curr)
1035 update_stats_wait_end(cfs_rq, se);
1037 if ((flags & DEQUEUE_SLEEP) && entity_is_task(se)) {
1038 struct task_struct *tsk = task_of(se);
1040 if (tsk->state & TASK_INTERRUPTIBLE)
1041 __schedstat_set(se->statistics.sleep_start,
1042 rq_clock(rq_of(cfs_rq)));
1043 if (tsk->state & TASK_UNINTERRUPTIBLE)
1044 __schedstat_set(se->statistics.block_start,
1045 rq_clock(rq_of(cfs_rq)));
1050 * We are picking a new current task - update its stats:
1053 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
1056 * We are starting a new run period:
1058 se->exec_start = rq_clock_task(rq_of(cfs_rq));
1061 /**************************************************
1062 * Scheduling class queueing methods:
1065 #ifdef CONFIG_NUMA_BALANCING
1067 * Approximate time to scan a full NUMA task in ms. The task scan period is
1068 * calculated based on the tasks virtual memory size and
1069 * numa_balancing_scan_size.
1071 unsigned int sysctl_numa_balancing_scan_period_min = 1000;
1072 unsigned int sysctl_numa_balancing_scan_period_max = 60000;
1074 /* Portion of address space to scan in MB */
1075 unsigned int sysctl_numa_balancing_scan_size = 256;
1077 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
1078 unsigned int sysctl_numa_balancing_scan_delay = 1000;
1081 refcount_t refcount;
1083 spinlock_t lock; /* nr_tasks, tasks */
1088 struct rcu_head rcu;
1089 unsigned long total_faults;
1090 unsigned long max_faults_cpu;
1092 * Faults_cpu is used to decide whether memory should move
1093 * towards the CPU. As a consequence, these stats are weighted
1094 * more by CPU use than by memory faults.
1096 unsigned long *faults_cpu;
1097 unsigned long faults[0];
1101 * For functions that can be called in multiple contexts that permit reading
1102 * ->numa_group (see struct task_struct for locking rules).
1104 static struct numa_group *deref_task_numa_group(struct task_struct *p)
1106 return rcu_dereference_check(p->numa_group, p == current ||
1107 (lockdep_is_held(&task_rq(p)->lock) && !READ_ONCE(p->on_cpu)));
1110 static struct numa_group *deref_curr_numa_group(struct task_struct *p)
1112 return rcu_dereference_protected(p->numa_group, p == current);
1115 static inline unsigned long group_faults_priv(struct numa_group *ng);
1116 static inline unsigned long group_faults_shared(struct numa_group *ng);
1118 static unsigned int task_nr_scan_windows(struct task_struct *p)
1120 unsigned long rss = 0;
1121 unsigned long nr_scan_pages;
1124 * Calculations based on RSS as non-present and empty pages are skipped
1125 * by the PTE scanner and NUMA hinting faults should be trapped based
1128 nr_scan_pages = sysctl_numa_balancing_scan_size << (20 - PAGE_SHIFT);
1129 rss = get_mm_rss(p->mm);
1131 rss = nr_scan_pages;
1133 rss = round_up(rss, nr_scan_pages);
1134 return rss / nr_scan_pages;
1137 /* For sanitys sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */
1138 #define MAX_SCAN_WINDOW 2560
1140 static unsigned int task_scan_min(struct task_struct *p)
1142 unsigned int scan_size = READ_ONCE(sysctl_numa_balancing_scan_size);
1143 unsigned int scan, floor;
1144 unsigned int windows = 1;
1146 if (scan_size < MAX_SCAN_WINDOW)
1147 windows = MAX_SCAN_WINDOW / scan_size;
1148 floor = 1000 / windows;
1150 scan = sysctl_numa_balancing_scan_period_min / task_nr_scan_windows(p);
1151 return max_t(unsigned int, floor, scan);
1154 static unsigned int task_scan_start(struct task_struct *p)
1156 unsigned long smin = task_scan_min(p);
1157 unsigned long period = smin;
1158 struct numa_group *ng;
1160 /* Scale the maximum scan period with the amount of shared memory. */
1162 ng = rcu_dereference(p->numa_group);
1164 unsigned long shared = group_faults_shared(ng);
1165 unsigned long private = group_faults_priv(ng);
1167 period *= refcount_read(&ng->refcount);
1168 period *= shared + 1;
1169 period /= private + shared + 1;
1173 return max(smin, period);
1176 static unsigned int task_scan_max(struct task_struct *p)
1178 unsigned long smin = task_scan_min(p);
1180 struct numa_group *ng;
1182 /* Watch for min being lower than max due to floor calculations */
1183 smax = sysctl_numa_balancing_scan_period_max / task_nr_scan_windows(p);
1185 /* Scale the maximum scan period with the amount of shared memory. */
1186 ng = deref_curr_numa_group(p);
1188 unsigned long shared = group_faults_shared(ng);
1189 unsigned long private = group_faults_priv(ng);
1190 unsigned long period = smax;
1192 period *= refcount_read(&ng->refcount);
1193 period *= shared + 1;
1194 period /= private + shared + 1;
1196 smax = max(smax, period);
1199 return max(smin, smax);
1202 static void account_numa_enqueue(struct rq *rq, struct task_struct *p)
1204 rq->nr_numa_running += (p->numa_preferred_nid != NUMA_NO_NODE);
1205 rq->nr_preferred_running += (p->numa_preferred_nid == task_node(p));
1208 static void account_numa_dequeue(struct rq *rq, struct task_struct *p)
1210 rq->nr_numa_running -= (p->numa_preferred_nid != NUMA_NO_NODE);
1211 rq->nr_preferred_running -= (p->numa_preferred_nid == task_node(p));
1214 /* Shared or private faults. */
1215 #define NR_NUMA_HINT_FAULT_TYPES 2
1217 /* Memory and CPU locality */
1218 #define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2)
1220 /* Averaged statistics, and temporary buffers. */
1221 #define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2)
1223 pid_t task_numa_group_id(struct task_struct *p)
1225 struct numa_group *ng;
1229 ng = rcu_dereference(p->numa_group);
1238 * The averaged statistics, shared & private, memory & CPU,
1239 * occupy the first half of the array. The second half of the
1240 * array is for current counters, which are averaged into the
1241 * first set by task_numa_placement.
1243 static inline int task_faults_idx(enum numa_faults_stats s, int nid, int priv)
1245 return NR_NUMA_HINT_FAULT_TYPES * (s * nr_node_ids + nid) + priv;
1248 static inline unsigned long task_faults(struct task_struct *p, int nid)
1250 if (!p->numa_faults)
1253 return p->numa_faults[task_faults_idx(NUMA_MEM, nid, 0)] +
1254 p->numa_faults[task_faults_idx(NUMA_MEM, nid, 1)];
1257 static inline unsigned long group_faults(struct task_struct *p, int nid)
1259 struct numa_group *ng = deref_task_numa_group(p);
1264 return ng->faults[task_faults_idx(NUMA_MEM, nid, 0)] +
1265 ng->faults[task_faults_idx(NUMA_MEM, nid, 1)];
1268 static inline unsigned long group_faults_cpu(struct numa_group *group, int nid)
1270 return group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 0)] +
1271 group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 1)];
1274 static inline unsigned long group_faults_priv(struct numa_group *ng)
1276 unsigned long faults = 0;
1279 for_each_online_node(node) {
1280 faults += ng->faults[task_faults_idx(NUMA_MEM, node, 1)];
1286 static inline unsigned long group_faults_shared(struct numa_group *ng)
1288 unsigned long faults = 0;
1291 for_each_online_node(node) {
1292 faults += ng->faults[task_faults_idx(NUMA_MEM, node, 0)];
1299 * A node triggering more than 1/3 as many NUMA faults as the maximum is
1300 * considered part of a numa group's pseudo-interleaving set. Migrations
1301 * between these nodes are slowed down, to allow things to settle down.
1303 #define ACTIVE_NODE_FRACTION 3
1305 static bool numa_is_active_node(int nid, struct numa_group *ng)
1307 return group_faults_cpu(ng, nid) * ACTIVE_NODE_FRACTION > ng->max_faults_cpu;
1310 /* Handle placement on systems where not all nodes are directly connected. */
1311 static unsigned long score_nearby_nodes(struct task_struct *p, int nid,
1312 int maxdist, bool task)
1314 unsigned long score = 0;
1318 * All nodes are directly connected, and the same distance
1319 * from each other. No need for fancy placement algorithms.
1321 if (sched_numa_topology_type == NUMA_DIRECT)
1325 * This code is called for each node, introducing N^2 complexity,
1326 * which should be ok given the number of nodes rarely exceeds 8.
1328 for_each_online_node(node) {
1329 unsigned long faults;
1330 int dist = node_distance(nid, node);
1333 * The furthest away nodes in the system are not interesting
1334 * for placement; nid was already counted.
1336 if (dist == sched_max_numa_distance || node == nid)
1340 * On systems with a backplane NUMA topology, compare groups
1341 * of nodes, and move tasks towards the group with the most
1342 * memory accesses. When comparing two nodes at distance
1343 * "hoplimit", only nodes closer by than "hoplimit" are part
1344 * of each group. Skip other nodes.
1346 if (sched_numa_topology_type == NUMA_BACKPLANE &&
1350 /* Add up the faults from nearby nodes. */
1352 faults = task_faults(p, node);
1354 faults = group_faults(p, node);
1357 * On systems with a glueless mesh NUMA topology, there are
1358 * no fixed "groups of nodes". Instead, nodes that are not
1359 * directly connected bounce traffic through intermediate
1360 * nodes; a numa_group can occupy any set of nodes.
1361 * The further away a node is, the less the faults count.
1362 * This seems to result in good task placement.
1364 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
1365 faults *= (sched_max_numa_distance - dist);
1366 faults /= (sched_max_numa_distance - LOCAL_DISTANCE);
1376 * These return the fraction of accesses done by a particular task, or
1377 * task group, on a particular numa node. The group weight is given a
1378 * larger multiplier, in order to group tasks together that are almost
1379 * evenly spread out between numa nodes.
1381 static inline unsigned long task_weight(struct task_struct *p, int nid,
1384 unsigned long faults, total_faults;
1386 if (!p->numa_faults)
1389 total_faults = p->total_numa_faults;
1394 faults = task_faults(p, nid);
1395 faults += score_nearby_nodes(p, nid, dist, true);
1397 return 1000 * faults / total_faults;
1400 static inline unsigned long group_weight(struct task_struct *p, int nid,
1403 struct numa_group *ng = deref_task_numa_group(p);
1404 unsigned long faults, total_faults;
1409 total_faults = ng->total_faults;
1414 faults = group_faults(p, nid);
1415 faults += score_nearby_nodes(p, nid, dist, false);
1417 return 1000 * faults / total_faults;
1420 bool should_numa_migrate_memory(struct task_struct *p, struct page * page,
1421 int src_nid, int dst_cpu)
1423 struct numa_group *ng = deref_curr_numa_group(p);
1424 int dst_nid = cpu_to_node(dst_cpu);
1425 int last_cpupid, this_cpupid;
1427 this_cpupid = cpu_pid_to_cpupid(dst_cpu, current->pid);
1428 last_cpupid = page_cpupid_xchg_last(page, this_cpupid);
1431 * Allow first faults or private faults to migrate immediately early in
1432 * the lifetime of a task. The magic number 4 is based on waiting for
1433 * two full passes of the "multi-stage node selection" test that is
1436 if ((p->numa_preferred_nid == NUMA_NO_NODE || p->numa_scan_seq <= 4) &&
1437 (cpupid_pid_unset(last_cpupid) || cpupid_match_pid(p, last_cpupid)))
1441 * Multi-stage node selection is used in conjunction with a periodic
1442 * migration fault to build a temporal task<->page relation. By using
1443 * a two-stage filter we remove short/unlikely relations.
1445 * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate
1446 * a task's usage of a particular page (n_p) per total usage of this
1447 * page (n_t) (in a given time-span) to a probability.
1449 * Our periodic faults will sample this probability and getting the
1450 * same result twice in a row, given these samples are fully
1451 * independent, is then given by P(n)^2, provided our sample period
1452 * is sufficiently short compared to the usage pattern.
1454 * This quadric squishes small probabilities, making it less likely we
1455 * act on an unlikely task<->page relation.
1457 if (!cpupid_pid_unset(last_cpupid) &&
1458 cpupid_to_nid(last_cpupid) != dst_nid)
1461 /* Always allow migrate on private faults */
1462 if (cpupid_match_pid(p, last_cpupid))
1465 /* A shared fault, but p->numa_group has not been set up yet. */
1470 * Destination node is much more heavily used than the source
1471 * node? Allow migration.
1473 if (group_faults_cpu(ng, dst_nid) > group_faults_cpu(ng, src_nid) *
1474 ACTIVE_NODE_FRACTION)
1478 * Distribute memory according to CPU & memory use on each node,
1479 * with 3/4 hysteresis to avoid unnecessary memory migrations:
1481 * faults_cpu(dst) 3 faults_cpu(src)
1482 * --------------- * - > ---------------
1483 * faults_mem(dst) 4 faults_mem(src)
1485 return group_faults_cpu(ng, dst_nid) * group_faults(p, src_nid) * 3 >
1486 group_faults_cpu(ng, src_nid) * group_faults(p, dst_nid) * 4;
1490 * 'numa_type' describes the node at the moment of load balancing.
1493 /* The node has spare capacity that can be used to run more tasks. */
1496 * The node is fully used and the tasks don't compete for more CPU
1497 * cycles. Nevertheless, some tasks might wait before running.
1501 * The node is overloaded and can't provide expected CPU cycles to all
1507 /* Cached statistics for all CPUs within a node */
1511 /* Total compute capacity of CPUs on a node */
1512 unsigned long compute_capacity;
1513 unsigned int nr_running;
1514 unsigned int weight;
1515 enum numa_type node_type;
1519 static inline bool is_core_idle(int cpu)
1521 #ifdef CONFIG_SCHED_SMT
1524 for_each_cpu(sibling, cpu_smt_mask(cpu)) {
1536 struct task_numa_env {
1537 struct task_struct *p;
1539 int src_cpu, src_nid;
1540 int dst_cpu, dst_nid;
1542 struct numa_stats src_stats, dst_stats;
1547 struct task_struct *best_task;
1552 static unsigned long cpu_load(struct rq *rq);
1553 static unsigned long cpu_util(int cpu);
1554 static inline long adjust_numa_imbalance(int imbalance, int src_nr_running);
1557 numa_type numa_classify(unsigned int imbalance_pct,
1558 struct numa_stats *ns)
1560 if ((ns->nr_running > ns->weight) &&
1561 ((ns->compute_capacity * 100) < (ns->util * imbalance_pct)))
1562 return node_overloaded;
1564 if ((ns->nr_running < ns->weight) ||
1565 ((ns->compute_capacity * 100) > (ns->util * imbalance_pct)))
1566 return node_has_spare;
1568 return node_fully_busy;
1571 #ifdef CONFIG_SCHED_SMT
1572 /* Forward declarations of select_idle_sibling helpers */
1573 static inline bool test_idle_cores(int cpu, bool def);
1574 static inline int numa_idle_core(int idle_core, int cpu)
1576 if (!static_branch_likely(&sched_smt_present) ||
1577 idle_core >= 0 || !test_idle_cores(cpu, false))
1581 * Prefer cores instead of packing HT siblings
1582 * and triggering future load balancing.
1584 if (is_core_idle(cpu))
1590 static inline int numa_idle_core(int idle_core, int cpu)
1597 * Gather all necessary information to make NUMA balancing placement
1598 * decisions that are compatible with standard load balancer. This
1599 * borrows code and logic from update_sg_lb_stats but sharing a
1600 * common implementation is impractical.
1602 static void update_numa_stats(struct task_numa_env *env,
1603 struct numa_stats *ns, int nid,
1606 int cpu, idle_core = -1;
1608 memset(ns, 0, sizeof(*ns));
1612 for_each_cpu(cpu, cpumask_of_node(nid)) {
1613 struct rq *rq = cpu_rq(cpu);
1615 ns->load += cpu_load(rq);
1616 ns->util += cpu_util(cpu);
1617 ns->nr_running += rq->cfs.h_nr_running;
1618 ns->compute_capacity += capacity_of(cpu);
1620 if (find_idle && !rq->nr_running && idle_cpu(cpu)) {
1621 if (READ_ONCE(rq->numa_migrate_on) ||
1622 !cpumask_test_cpu(cpu, env->p->cpus_ptr))
1625 if (ns->idle_cpu == -1)
1628 idle_core = numa_idle_core(idle_core, cpu);
1633 ns->weight = cpumask_weight(cpumask_of_node(nid));
1635 ns->node_type = numa_classify(env->imbalance_pct, ns);
1638 ns->idle_cpu = idle_core;
1641 static void task_numa_assign(struct task_numa_env *env,
1642 struct task_struct *p, long imp)
1644 struct rq *rq = cpu_rq(env->dst_cpu);
1646 /* Check if run-queue part of active NUMA balance. */
1647 if (env->best_cpu != env->dst_cpu && xchg(&rq->numa_migrate_on, 1)) {
1649 int start = env->dst_cpu;
1651 /* Find alternative idle CPU. */
1652 for_each_cpu_wrap(cpu, cpumask_of_node(env->dst_nid), start) {
1653 if (cpu == env->best_cpu || !idle_cpu(cpu) ||
1654 !cpumask_test_cpu(cpu, env->p->cpus_ptr)) {
1659 rq = cpu_rq(env->dst_cpu);
1660 if (!xchg(&rq->numa_migrate_on, 1))
1664 /* Failed to find an alternative idle CPU */
1670 * Clear previous best_cpu/rq numa-migrate flag, since task now
1671 * found a better CPU to move/swap.
1673 if (env->best_cpu != -1 && env->best_cpu != env->dst_cpu) {
1674 rq = cpu_rq(env->best_cpu);
1675 WRITE_ONCE(rq->numa_migrate_on, 0);
1679 put_task_struct(env->best_task);
1684 env->best_imp = imp;
1685 env->best_cpu = env->dst_cpu;
1688 static bool load_too_imbalanced(long src_load, long dst_load,
1689 struct task_numa_env *env)
1692 long orig_src_load, orig_dst_load;
1693 long src_capacity, dst_capacity;
1696 * The load is corrected for the CPU capacity available on each node.
1699 * ------------ vs ---------
1700 * src_capacity dst_capacity
1702 src_capacity = env->src_stats.compute_capacity;
1703 dst_capacity = env->dst_stats.compute_capacity;
1705 imb = abs(dst_load * src_capacity - src_load * dst_capacity);
1707 orig_src_load = env->src_stats.load;
1708 orig_dst_load = env->dst_stats.load;
1710 old_imb = abs(orig_dst_load * src_capacity - orig_src_load * dst_capacity);
1712 /* Would this change make things worse? */
1713 return (imb > old_imb);
1717 * Maximum NUMA importance can be 1998 (2*999);
1718 * SMALLIMP @ 30 would be close to 1998/64.
1719 * Used to deter task migration.
1724 * This checks if the overall compute and NUMA accesses of the system would
1725 * be improved if the source tasks was migrated to the target dst_cpu taking
1726 * into account that it might be best if task running on the dst_cpu should
1727 * be exchanged with the source task
1729 static bool task_numa_compare(struct task_numa_env *env,
1730 long taskimp, long groupimp, bool maymove)
1732 struct numa_group *cur_ng, *p_ng = deref_curr_numa_group(env->p);
1733 struct rq *dst_rq = cpu_rq(env->dst_cpu);
1734 long imp = p_ng ? groupimp : taskimp;
1735 struct task_struct *cur;
1736 long src_load, dst_load;
1737 int dist = env->dist;
1740 bool stopsearch = false;
1742 if (READ_ONCE(dst_rq->numa_migrate_on))
1746 cur = rcu_dereference(dst_rq->curr);
1747 if (cur && ((cur->flags & PF_EXITING) || is_idle_task(cur)))
1751 * Because we have preemption enabled we can get migrated around and
1752 * end try selecting ourselves (current == env->p) as a swap candidate.
1754 if (cur == env->p) {
1760 if (maymove && moveimp >= env->best_imp)
1766 /* Skip this swap candidate if cannot move to the source cpu. */
1767 if (!cpumask_test_cpu(env->src_cpu, cur->cpus_ptr))
1771 * Skip this swap candidate if it is not moving to its preferred
1772 * node and the best task is.
1774 if (env->best_task &&
1775 env->best_task->numa_preferred_nid == env->src_nid &&
1776 cur->numa_preferred_nid != env->src_nid) {
1781 * "imp" is the fault differential for the source task between the
1782 * source and destination node. Calculate the total differential for
1783 * the source task and potential destination task. The more negative
1784 * the value is, the more remote accesses that would be expected to
1785 * be incurred if the tasks were swapped.
1787 * If dst and source tasks are in the same NUMA group, or not
1788 * in any group then look only at task weights.
1790 cur_ng = rcu_dereference(cur->numa_group);
1791 if (cur_ng == p_ng) {
1792 imp = taskimp + task_weight(cur, env->src_nid, dist) -
1793 task_weight(cur, env->dst_nid, dist);
1795 * Add some hysteresis to prevent swapping the
1796 * tasks within a group over tiny differences.
1802 * Compare the group weights. If a task is all by itself
1803 * (not part of a group), use the task weight instead.
1806 imp += group_weight(cur, env->src_nid, dist) -
1807 group_weight(cur, env->dst_nid, dist);
1809 imp += task_weight(cur, env->src_nid, dist) -
1810 task_weight(cur, env->dst_nid, dist);
1813 /* Discourage picking a task already on its preferred node */
1814 if (cur->numa_preferred_nid == env->dst_nid)
1818 * Encourage picking a task that moves to its preferred node.
1819 * This potentially makes imp larger than it's maximum of
1820 * 1998 (see SMALLIMP and task_weight for why) but in this
1821 * case, it does not matter.
1823 if (cur->numa_preferred_nid == env->src_nid)
1826 if (maymove && moveimp > imp && moveimp > env->best_imp) {
1833 * Prefer swapping with a task moving to its preferred node over a
1836 if (env->best_task && cur->numa_preferred_nid == env->src_nid &&
1837 env->best_task->numa_preferred_nid != env->src_nid) {
1842 * If the NUMA importance is less than SMALLIMP,
1843 * task migration might only result in ping pong
1844 * of tasks and also hurt performance due to cache
1847 if (imp < SMALLIMP || imp <= env->best_imp + SMALLIMP / 2)
1851 * In the overloaded case, try and keep the load balanced.
1853 load = task_h_load(env->p) - task_h_load(cur);
1857 dst_load = env->dst_stats.load + load;
1858 src_load = env->src_stats.load - load;
1860 if (load_too_imbalanced(src_load, dst_load, env))
1864 /* Evaluate an idle CPU for a task numa move. */
1866 int cpu = env->dst_stats.idle_cpu;
1868 /* Nothing cached so current CPU went idle since the search. */
1873 * If the CPU is no longer truly idle and the previous best CPU
1874 * is, keep using it.
1876 if (!idle_cpu(cpu) && env->best_cpu >= 0 &&
1877 idle_cpu(env->best_cpu)) {
1878 cpu = env->best_cpu;
1884 task_numa_assign(env, cur, imp);
1887 * If a move to idle is allowed because there is capacity or load
1888 * balance improves then stop the search. While a better swap
1889 * candidate may exist, a search is not free.
1891 if (maymove && !cur && env->best_cpu >= 0 && idle_cpu(env->best_cpu))
1895 * If a swap candidate must be identified and the current best task
1896 * moves its preferred node then stop the search.
1898 if (!maymove && env->best_task &&
1899 env->best_task->numa_preferred_nid == env->src_nid) {
1908 static void task_numa_find_cpu(struct task_numa_env *env,
1909 long taskimp, long groupimp)
1911 bool maymove = false;
1915 * If dst node has spare capacity, then check if there is an
1916 * imbalance that would be overruled by the load balancer.
1918 if (env->dst_stats.node_type == node_has_spare) {
1919 unsigned int imbalance;
1920 int src_running, dst_running;
1923 * Would movement cause an imbalance? Note that if src has
1924 * more running tasks that the imbalance is ignored as the
1925 * move improves the imbalance from the perspective of the
1926 * CPU load balancer.
1928 src_running = env->src_stats.nr_running - 1;
1929 dst_running = env->dst_stats.nr_running + 1;
1930 imbalance = max(0, dst_running - src_running);
1931 imbalance = adjust_numa_imbalance(imbalance, src_running);
1933 /* Use idle CPU if there is no imbalance */
1936 if (env->dst_stats.idle_cpu >= 0) {
1937 env->dst_cpu = env->dst_stats.idle_cpu;
1938 task_numa_assign(env, NULL, 0);
1943 long src_load, dst_load, load;
1945 * If the improvement from just moving env->p direction is better
1946 * than swapping tasks around, check if a move is possible.
1948 load = task_h_load(env->p);
1949 dst_load = env->dst_stats.load + load;
1950 src_load = env->src_stats.load - load;
1951 maymove = !load_too_imbalanced(src_load, dst_load, env);
1954 for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) {
1955 /* Skip this CPU if the source task cannot migrate */
1956 if (!cpumask_test_cpu(cpu, env->p->cpus_ptr))
1960 if (task_numa_compare(env, taskimp, groupimp, maymove))
1965 static int task_numa_migrate(struct task_struct *p)
1967 struct task_numa_env env = {
1970 .src_cpu = task_cpu(p),
1971 .src_nid = task_node(p),
1973 .imbalance_pct = 112,
1979 unsigned long taskweight, groupweight;
1980 struct sched_domain *sd;
1981 long taskimp, groupimp;
1982 struct numa_group *ng;
1987 * Pick the lowest SD_NUMA domain, as that would have the smallest
1988 * imbalance and would be the first to start moving tasks about.
1990 * And we want to avoid any moving of tasks about, as that would create
1991 * random movement of tasks -- counter the numa conditions we're trying
1995 sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu));
1997 env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2;
2001 * Cpusets can break the scheduler domain tree into smaller
2002 * balance domains, some of which do not cross NUMA boundaries.
2003 * Tasks that are "trapped" in such domains cannot be migrated
2004 * elsewhere, so there is no point in (re)trying.
2006 if (unlikely(!sd)) {
2007 sched_setnuma(p, task_node(p));
2011 env.dst_nid = p->numa_preferred_nid;
2012 dist = env.dist = node_distance(env.src_nid, env.dst_nid);
2013 taskweight = task_weight(p, env.src_nid, dist);
2014 groupweight = group_weight(p, env.src_nid, dist);
2015 update_numa_stats(&env, &env.src_stats, env.src_nid, false);
2016 taskimp = task_weight(p, env.dst_nid, dist) - taskweight;
2017 groupimp = group_weight(p, env.dst_nid, dist) - groupweight;
2018 update_numa_stats(&env, &env.dst_stats, env.dst_nid, true);
2020 /* Try to find a spot on the preferred nid. */
2021 task_numa_find_cpu(&env, taskimp, groupimp);
2024 * Look at other nodes in these cases:
2025 * - there is no space available on the preferred_nid
2026 * - the task is part of a numa_group that is interleaved across
2027 * multiple NUMA nodes; in order to better consolidate the group,
2028 * we need to check other locations.
2030 ng = deref_curr_numa_group(p);
2031 if (env.best_cpu == -1 || (ng && ng->active_nodes > 1)) {
2032 for_each_online_node(nid) {
2033 if (nid == env.src_nid || nid == p->numa_preferred_nid)
2036 dist = node_distance(env.src_nid, env.dst_nid);
2037 if (sched_numa_topology_type == NUMA_BACKPLANE &&
2039 taskweight = task_weight(p, env.src_nid, dist);
2040 groupweight = group_weight(p, env.src_nid, dist);
2043 /* Only consider nodes where both task and groups benefit */
2044 taskimp = task_weight(p, nid, dist) - taskweight;
2045 groupimp = group_weight(p, nid, dist) - groupweight;
2046 if (taskimp < 0 && groupimp < 0)
2051 update_numa_stats(&env, &env.dst_stats, env.dst_nid, true);
2052 task_numa_find_cpu(&env, taskimp, groupimp);
2057 * If the task is part of a workload that spans multiple NUMA nodes,
2058 * and is migrating into one of the workload's active nodes, remember
2059 * this node as the task's preferred numa node, so the workload can
2061 * A task that migrated to a second choice node will be better off
2062 * trying for a better one later. Do not set the preferred node here.
2065 if (env.best_cpu == -1)
2068 nid = cpu_to_node(env.best_cpu);
2070 if (nid != p->numa_preferred_nid)
2071 sched_setnuma(p, nid);
2074 /* No better CPU than the current one was found. */
2075 if (env.best_cpu == -1) {
2076 trace_sched_stick_numa(p, env.src_cpu, NULL, -1);
2080 best_rq = cpu_rq(env.best_cpu);
2081 if (env.best_task == NULL) {
2082 ret = migrate_task_to(p, env.best_cpu);
2083 WRITE_ONCE(best_rq->numa_migrate_on, 0);
2085 trace_sched_stick_numa(p, env.src_cpu, NULL, env.best_cpu);
2089 ret = migrate_swap(p, env.best_task, env.best_cpu, env.src_cpu);
2090 WRITE_ONCE(best_rq->numa_migrate_on, 0);
2093 trace_sched_stick_numa(p, env.src_cpu, env.best_task, env.best_cpu);
2094 put_task_struct(env.best_task);
2098 /* Attempt to migrate a task to a CPU on the preferred node. */
2099 static void numa_migrate_preferred(struct task_struct *p)
2101 unsigned long interval = HZ;
2103 /* This task has no NUMA fault statistics yet */
2104 if (unlikely(p->numa_preferred_nid == NUMA_NO_NODE || !p->numa_faults))
2107 /* Periodically retry migrating the task to the preferred node */
2108 interval = min(interval, msecs_to_jiffies(p->numa_scan_period) / 16);
2109 p->numa_migrate_retry = jiffies + interval;
2111 /* Success if task is already running on preferred CPU */
2112 if (task_node(p) == p->numa_preferred_nid)
2115 /* Otherwise, try migrate to a CPU on the preferred node */
2116 task_numa_migrate(p);
2120 * Find out how many nodes on the workload is actively running on. Do this by
2121 * tracking the nodes from which NUMA hinting faults are triggered. This can
2122 * be different from the set of nodes where the workload's memory is currently
2125 static void numa_group_count_active_nodes(struct numa_group *numa_group)
2127 unsigned long faults, max_faults = 0;
2128 int nid, active_nodes = 0;
2130 for_each_online_node(nid) {
2131 faults = group_faults_cpu(numa_group, nid);
2132 if (faults > max_faults)
2133 max_faults = faults;
2136 for_each_online_node(nid) {
2137 faults = group_faults_cpu(numa_group, nid);
2138 if (faults * ACTIVE_NODE_FRACTION > max_faults)
2142 numa_group->max_faults_cpu = max_faults;
2143 numa_group->active_nodes = active_nodes;
2147 * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
2148 * increments. The more local the fault statistics are, the higher the scan
2149 * period will be for the next scan window. If local/(local+remote) ratio is
2150 * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS)
2151 * the scan period will decrease. Aim for 70% local accesses.
2153 #define NUMA_PERIOD_SLOTS 10
2154 #define NUMA_PERIOD_THRESHOLD 7
2157 * Increase the scan period (slow down scanning) if the majority of
2158 * our memory is already on our local node, or if the majority of
2159 * the page accesses are shared with other processes.
2160 * Otherwise, decrease the scan period.
2162 static void update_task_scan_period(struct task_struct *p,
2163 unsigned long shared, unsigned long private)
2165 unsigned int period_slot;
2166 int lr_ratio, ps_ratio;
2169 unsigned long remote = p->numa_faults_locality[0];
2170 unsigned long local = p->numa_faults_locality[1];
2173 * If there were no record hinting faults then either the task is
2174 * completely idle or all activity is areas that are not of interest
2175 * to automatic numa balancing. Related to that, if there were failed
2176 * migration then it implies we are migrating too quickly or the local
2177 * node is overloaded. In either case, scan slower
2179 if (local + shared == 0 || p->numa_faults_locality[2]) {
2180 p->numa_scan_period = min(p->numa_scan_period_max,
2181 p->numa_scan_period << 1);
2183 p->mm->numa_next_scan = jiffies +
2184 msecs_to_jiffies(p->numa_scan_period);
2190 * Prepare to scale scan period relative to the current period.
2191 * == NUMA_PERIOD_THRESHOLD scan period stays the same
2192 * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
2193 * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
2195 period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS);
2196 lr_ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote);
2197 ps_ratio = (private * NUMA_PERIOD_SLOTS) / (private + shared);
2199 if (ps_ratio >= NUMA_PERIOD_THRESHOLD) {
2201 * Most memory accesses are local. There is no need to
2202 * do fast NUMA scanning, since memory is already local.
2204 int slot = ps_ratio - NUMA_PERIOD_THRESHOLD;
2207 diff = slot * period_slot;
2208 } else if (lr_ratio >= NUMA_PERIOD_THRESHOLD) {
2210 * Most memory accesses are shared with other tasks.
2211 * There is no point in continuing fast NUMA scanning,
2212 * since other tasks may just move the memory elsewhere.
2214 int slot = lr_ratio - NUMA_PERIOD_THRESHOLD;
2217 diff = slot * period_slot;
2220 * Private memory faults exceed (SLOTS-THRESHOLD)/SLOTS,
2221 * yet they are not on the local NUMA node. Speed up
2222 * NUMA scanning to get the memory moved over.
2224 int ratio = max(lr_ratio, ps_ratio);
2225 diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot;
2228 p->numa_scan_period = clamp(p->numa_scan_period + diff,
2229 task_scan_min(p), task_scan_max(p));
2230 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
2234 * Get the fraction of time the task has been running since the last
2235 * NUMA placement cycle. The scheduler keeps similar statistics, but
2236 * decays those on a 32ms period, which is orders of magnitude off
2237 * from the dozens-of-seconds NUMA balancing period. Use the scheduler
2238 * stats only if the task is so new there are no NUMA statistics yet.
2240 static u64 numa_get_avg_runtime(struct task_struct *p, u64 *period)
2242 u64 runtime, delta, now;
2243 /* Use the start of this time slice to avoid calculations. */
2244 now = p->se.exec_start;
2245 runtime = p->se.sum_exec_runtime;
2247 if (p->last_task_numa_placement) {
2248 delta = runtime - p->last_sum_exec_runtime;
2249 *period = now - p->last_task_numa_placement;
2251 /* Avoid time going backwards, prevent potential divide error: */
2252 if (unlikely((s64)*period < 0))
2255 delta = p->se.avg.load_sum;
2256 *period = LOAD_AVG_MAX;
2259 p->last_sum_exec_runtime = runtime;
2260 p->last_task_numa_placement = now;
2266 * Determine the preferred nid for a task in a numa_group. This needs to
2267 * be done in a way that produces consistent results with group_weight,
2268 * otherwise workloads might not converge.
2270 static int preferred_group_nid(struct task_struct *p, int nid)
2275 /* Direct connections between all NUMA nodes. */
2276 if (sched_numa_topology_type == NUMA_DIRECT)
2280 * On a system with glueless mesh NUMA topology, group_weight
2281 * scores nodes according to the number of NUMA hinting faults on
2282 * both the node itself, and on nearby nodes.
2284 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
2285 unsigned long score, max_score = 0;
2286 int node, max_node = nid;
2288 dist = sched_max_numa_distance;
2290 for_each_online_node(node) {
2291 score = group_weight(p, node, dist);
2292 if (score > max_score) {
2301 * Finding the preferred nid in a system with NUMA backplane
2302 * interconnect topology is more involved. The goal is to locate
2303 * tasks from numa_groups near each other in the system, and
2304 * untangle workloads from different sides of the system. This requires
2305 * searching down the hierarchy of node groups, recursively searching
2306 * inside the highest scoring group of nodes. The nodemask tricks
2307 * keep the complexity of the search down.
2309 nodes = node_online_map;
2310 for (dist = sched_max_numa_distance; dist > LOCAL_DISTANCE; dist--) {
2311 unsigned long max_faults = 0;
2312 nodemask_t max_group = NODE_MASK_NONE;
2315 /* Are there nodes at this distance from each other? */
2316 if (!find_numa_distance(dist))
2319 for_each_node_mask(a, nodes) {
2320 unsigned long faults = 0;
2321 nodemask_t this_group;
2322 nodes_clear(this_group);
2324 /* Sum group's NUMA faults; includes a==b case. */
2325 for_each_node_mask(b, nodes) {
2326 if (node_distance(a, b) < dist) {
2327 faults += group_faults(p, b);
2328 node_set(b, this_group);
2329 node_clear(b, nodes);
2333 /* Remember the top group. */
2334 if (faults > max_faults) {
2335 max_faults = faults;
2336 max_group = this_group;
2338 * subtle: at the smallest distance there is
2339 * just one node left in each "group", the
2340 * winner is the preferred nid.
2345 /* Next round, evaluate the nodes within max_group. */
2353 static void task_numa_placement(struct task_struct *p)
2355 int seq, nid, max_nid = NUMA_NO_NODE;
2356 unsigned long max_faults = 0;
2357 unsigned long fault_types[2] = { 0, 0 };
2358 unsigned long total_faults;
2359 u64 runtime, period;
2360 spinlock_t *group_lock = NULL;
2361 struct numa_group *ng;
2364 * The p->mm->numa_scan_seq field gets updated without
2365 * exclusive access. Use READ_ONCE() here to ensure
2366 * that the field is read in a single access:
2368 seq = READ_ONCE(p->mm->numa_scan_seq);
2369 if (p->numa_scan_seq == seq)
2371 p->numa_scan_seq = seq;
2372 p->numa_scan_period_max = task_scan_max(p);
2374 total_faults = p->numa_faults_locality[0] +
2375 p->numa_faults_locality[1];
2376 runtime = numa_get_avg_runtime(p, &period);
2378 /* If the task is part of a group prevent parallel updates to group stats */
2379 ng = deref_curr_numa_group(p);
2381 group_lock = &ng->lock;
2382 spin_lock_irq(group_lock);
2385 /* Find the node with the highest number of faults */
2386 for_each_online_node(nid) {
2387 /* Keep track of the offsets in numa_faults array */
2388 int mem_idx, membuf_idx, cpu_idx, cpubuf_idx;
2389 unsigned long faults = 0, group_faults = 0;
2392 for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) {
2393 long diff, f_diff, f_weight;
2395 mem_idx = task_faults_idx(NUMA_MEM, nid, priv);
2396 membuf_idx = task_faults_idx(NUMA_MEMBUF, nid, priv);
2397 cpu_idx = task_faults_idx(NUMA_CPU, nid, priv);
2398 cpubuf_idx = task_faults_idx(NUMA_CPUBUF, nid, priv);
2400 /* Decay existing window, copy faults since last scan */
2401 diff = p->numa_faults[membuf_idx] - p->numa_faults[mem_idx] / 2;
2402 fault_types[priv] += p->numa_faults[membuf_idx];
2403 p->numa_faults[membuf_idx] = 0;
2406 * Normalize the faults_from, so all tasks in a group
2407 * count according to CPU use, instead of by the raw
2408 * number of faults. Tasks with little runtime have
2409 * little over-all impact on throughput, and thus their
2410 * faults are less important.
2412 f_weight = div64_u64(runtime << 16, period + 1);
2413 f_weight = (f_weight * p->numa_faults[cpubuf_idx]) /
2415 f_diff = f_weight - p->numa_faults[cpu_idx] / 2;
2416 p->numa_faults[cpubuf_idx] = 0;
2418 p->numa_faults[mem_idx] += diff;
2419 p->numa_faults[cpu_idx] += f_diff;
2420 faults += p->numa_faults[mem_idx];
2421 p->total_numa_faults += diff;
2424 * safe because we can only change our own group
2426 * mem_idx represents the offset for a given
2427 * nid and priv in a specific region because it
2428 * is at the beginning of the numa_faults array.
2430 ng->faults[mem_idx] += diff;
2431 ng->faults_cpu[mem_idx] += f_diff;
2432 ng->total_faults += diff;
2433 group_faults += ng->faults[mem_idx];
2438 if (faults > max_faults) {
2439 max_faults = faults;
2442 } else if (group_faults > max_faults) {
2443 max_faults = group_faults;
2449 numa_group_count_active_nodes(ng);
2450 spin_unlock_irq(group_lock);
2451 max_nid = preferred_group_nid(p, max_nid);
2455 /* Set the new preferred node */
2456 if (max_nid != p->numa_preferred_nid)
2457 sched_setnuma(p, max_nid);
2460 update_task_scan_period(p, fault_types[0], fault_types[1]);
2463 static inline int get_numa_group(struct numa_group *grp)
2465 return refcount_inc_not_zero(&grp->refcount);
2468 static inline void put_numa_group(struct numa_group *grp)
2470 if (refcount_dec_and_test(&grp->refcount))
2471 kfree_rcu(grp, rcu);
2474 static void task_numa_group(struct task_struct *p, int cpupid, int flags,
2477 struct numa_group *grp, *my_grp;
2478 struct task_struct *tsk;
2480 int cpu = cpupid_to_cpu(cpupid);
2483 if (unlikely(!deref_curr_numa_group(p))) {
2484 unsigned int size = sizeof(struct numa_group) +
2485 4*nr_node_ids*sizeof(unsigned long);
2487 grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
2491 refcount_set(&grp->refcount, 1);
2492 grp->active_nodes = 1;
2493 grp->max_faults_cpu = 0;
2494 spin_lock_init(&grp->lock);
2496 /* Second half of the array tracks nids where faults happen */
2497 grp->faults_cpu = grp->faults + NR_NUMA_HINT_FAULT_TYPES *
2500 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2501 grp->faults[i] = p->numa_faults[i];
2503 grp->total_faults = p->total_numa_faults;
2506 rcu_assign_pointer(p->numa_group, grp);
2510 tsk = READ_ONCE(cpu_rq(cpu)->curr);
2512 if (!cpupid_match_pid(tsk, cpupid))
2515 grp = rcu_dereference(tsk->numa_group);
2519 my_grp = deref_curr_numa_group(p);
2524 * Only join the other group if its bigger; if we're the bigger group,
2525 * the other task will join us.
2527 if (my_grp->nr_tasks > grp->nr_tasks)
2531 * Tie-break on the grp address.
2533 if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp)
2536 /* Always join threads in the same process. */
2537 if (tsk->mm == current->mm)
2540 /* Simple filter to avoid false positives due to PID collisions */
2541 if (flags & TNF_SHARED)
2544 /* Update priv based on whether false sharing was detected */
2547 if (join && !get_numa_group(grp))
2555 BUG_ON(irqs_disabled());
2556 double_lock_irq(&my_grp->lock, &grp->lock);
2558 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) {
2559 my_grp->faults[i] -= p->numa_faults[i];
2560 grp->faults[i] += p->numa_faults[i];
2562 my_grp->total_faults -= p->total_numa_faults;
2563 grp->total_faults += p->total_numa_faults;
2568 spin_unlock(&my_grp->lock);
2569 spin_unlock_irq(&grp->lock);
2571 rcu_assign_pointer(p->numa_group, grp);
2573 put_numa_group(my_grp);
2582 * Get rid of NUMA staticstics associated with a task (either current or dead).
2583 * If @final is set, the task is dead and has reached refcount zero, so we can
2584 * safely free all relevant data structures. Otherwise, there might be
2585 * concurrent reads from places like load balancing and procfs, and we should
2586 * reset the data back to default state without freeing ->numa_faults.
2588 void task_numa_free(struct task_struct *p, bool final)
2590 /* safe: p either is current or is being freed by current */
2591 struct numa_group *grp = rcu_dereference_raw(p->numa_group);
2592 unsigned long *numa_faults = p->numa_faults;
2593 unsigned long flags;
2600 spin_lock_irqsave(&grp->lock, flags);
2601 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2602 grp->faults[i] -= p->numa_faults[i];
2603 grp->total_faults -= p->total_numa_faults;
2606 spin_unlock_irqrestore(&grp->lock, flags);
2607 RCU_INIT_POINTER(p->numa_group, NULL);
2608 put_numa_group(grp);
2612 p->numa_faults = NULL;
2615 p->total_numa_faults = 0;
2616 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2622 * Got a PROT_NONE fault for a page on @node.
2624 void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags)
2626 struct task_struct *p = current;
2627 bool migrated = flags & TNF_MIGRATED;
2628 int cpu_node = task_node(current);
2629 int local = !!(flags & TNF_FAULT_LOCAL);
2630 struct numa_group *ng;
2633 if (!static_branch_likely(&sched_numa_balancing))
2636 /* for example, ksmd faulting in a user's mm */
2640 /* Allocate buffer to track faults on a per-node basis */
2641 if (unlikely(!p->numa_faults)) {
2642 int size = sizeof(*p->numa_faults) *
2643 NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids;
2645 p->numa_faults = kzalloc(size, GFP_KERNEL|__GFP_NOWARN);
2646 if (!p->numa_faults)
2649 p->total_numa_faults = 0;
2650 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
2654 * First accesses are treated as private, otherwise consider accesses
2655 * to be private if the accessing pid has not changed
2657 if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) {
2660 priv = cpupid_match_pid(p, last_cpupid);
2661 if (!priv && !(flags & TNF_NO_GROUP))
2662 task_numa_group(p, last_cpupid, flags, &priv);
2666 * If a workload spans multiple NUMA nodes, a shared fault that
2667 * occurs wholly within the set of nodes that the workload is
2668 * actively using should be counted as local. This allows the
2669 * scan rate to slow down when a workload has settled down.
2671 ng = deref_curr_numa_group(p);
2672 if (!priv && !local && ng && ng->active_nodes > 1 &&
2673 numa_is_active_node(cpu_node, ng) &&
2674 numa_is_active_node(mem_node, ng))
2678 * Retry to migrate task to preferred node periodically, in case it
2679 * previously failed, or the scheduler moved us.
2681 if (time_after(jiffies, p->numa_migrate_retry)) {
2682 task_numa_placement(p);
2683 numa_migrate_preferred(p);
2687 p->numa_pages_migrated += pages;
2688 if (flags & TNF_MIGRATE_FAIL)
2689 p->numa_faults_locality[2] += pages;
2691 p->numa_faults[task_faults_idx(NUMA_MEMBUF, mem_node, priv)] += pages;
2692 p->numa_faults[task_faults_idx(NUMA_CPUBUF, cpu_node, priv)] += pages;
2693 p->numa_faults_locality[local] += pages;
2696 static void reset_ptenuma_scan(struct task_struct *p)
2699 * We only did a read acquisition of the mmap sem, so
2700 * p->mm->numa_scan_seq is written to without exclusive access
2701 * and the update is not guaranteed to be atomic. That's not
2702 * much of an issue though, since this is just used for
2703 * statistical sampling. Use READ_ONCE/WRITE_ONCE, which are not
2704 * expensive, to avoid any form of compiler optimizations:
2706 WRITE_ONCE(p->mm->numa_scan_seq, READ_ONCE(p->mm->numa_scan_seq) + 1);
2707 p->mm->numa_scan_offset = 0;
2711 * The expensive part of numa migration is done from task_work context.
2712 * Triggered from task_tick_numa().
2714 static void task_numa_work(struct callback_head *work)
2716 unsigned long migrate, next_scan, now = jiffies;
2717 struct task_struct *p = current;
2718 struct mm_struct *mm = p->mm;
2719 u64 runtime = p->se.sum_exec_runtime;
2720 struct vm_area_struct *vma;
2721 unsigned long start, end;
2722 unsigned long nr_pte_updates = 0;
2723 long pages, virtpages;
2725 SCHED_WARN_ON(p != container_of(work, struct task_struct, numa_work));
2729 * Who cares about NUMA placement when they're dying.
2731 * NOTE: make sure not to dereference p->mm before this check,
2732 * exit_task_work() happens _after_ exit_mm() so we could be called
2733 * without p->mm even though we still had it when we enqueued this
2736 if (p->flags & PF_EXITING)
2739 if (!mm->numa_next_scan) {
2740 mm->numa_next_scan = now +
2741 msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2745 * Enforce maximal scan/migration frequency..
2747 migrate = mm->numa_next_scan;
2748 if (time_before(now, migrate))
2751 if (p->numa_scan_period == 0) {
2752 p->numa_scan_period_max = task_scan_max(p);
2753 p->numa_scan_period = task_scan_start(p);
2756 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
2757 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
2761 * Delay this task enough that another task of this mm will likely win
2762 * the next time around.
2764 p->node_stamp += 2 * TICK_NSEC;
2766 start = mm->numa_scan_offset;
2767 pages = sysctl_numa_balancing_scan_size;
2768 pages <<= 20 - PAGE_SHIFT; /* MB in pages */
2769 virtpages = pages * 8; /* Scan up to this much virtual space */
2774 if (!down_read_trylock(&mm->mmap_sem))
2776 vma = find_vma(mm, start);
2778 reset_ptenuma_scan(p);
2782 for (; vma; vma = vma->vm_next) {
2783 if (!vma_migratable(vma) || !vma_policy_mof(vma) ||
2784 is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_MIXEDMAP)) {
2789 * Shared library pages mapped by multiple processes are not
2790 * migrated as it is expected they are cache replicated. Avoid
2791 * hinting faults in read-only file-backed mappings or the vdso
2792 * as migrating the pages will be of marginal benefit.
2795 (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ)))
2799 * Skip inaccessible VMAs to avoid any confusion between
2800 * PROT_NONE and NUMA hinting ptes
2802 if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)))
2806 start = max(start, vma->vm_start);
2807 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
2808 end = min(end, vma->vm_end);
2809 nr_pte_updates = change_prot_numa(vma, start, end);
2812 * Try to scan sysctl_numa_balancing_size worth of
2813 * hpages that have at least one present PTE that
2814 * is not already pte-numa. If the VMA contains
2815 * areas that are unused or already full of prot_numa
2816 * PTEs, scan up to virtpages, to skip through those
2820 pages -= (end - start) >> PAGE_SHIFT;
2821 virtpages -= (end - start) >> PAGE_SHIFT;
2824 if (pages <= 0 || virtpages <= 0)
2828 } while (end != vma->vm_end);
2833 * It is possible to reach the end of the VMA list but the last few
2834 * VMAs are not guaranteed to the vma_migratable. If they are not, we
2835 * would find the !migratable VMA on the next scan but not reset the
2836 * scanner to the start so check it now.
2839 mm->numa_scan_offset = start;
2841 reset_ptenuma_scan(p);
2842 up_read(&mm->mmap_sem);
2845 * Make sure tasks use at least 32x as much time to run other code
2846 * than they used here, to limit NUMA PTE scanning overhead to 3% max.
2847 * Usually update_task_scan_period slows down scanning enough; on an
2848 * overloaded system we need to limit overhead on a per task basis.
2850 if (unlikely(p->se.sum_exec_runtime != runtime)) {
2851 u64 diff = p->se.sum_exec_runtime - runtime;
2852 p->node_stamp += 32 * diff;
2856 void init_numa_balancing(unsigned long clone_flags, struct task_struct *p)
2859 struct mm_struct *mm = p->mm;
2862 mm_users = atomic_read(&mm->mm_users);
2863 if (mm_users == 1) {
2864 mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2865 mm->numa_scan_seq = 0;
2869 p->numa_scan_seq = mm ? mm->numa_scan_seq : 0;
2870 p->numa_scan_period = sysctl_numa_balancing_scan_delay;
2871 /* Protect against double add, see task_tick_numa and task_numa_work */
2872 p->numa_work.next = &p->numa_work;
2873 p->numa_faults = NULL;
2874 RCU_INIT_POINTER(p->numa_group, NULL);
2875 p->last_task_numa_placement = 0;
2876 p->last_sum_exec_runtime = 0;
2878 init_task_work(&p->numa_work, task_numa_work);
2880 /* New address space, reset the preferred nid */
2881 if (!(clone_flags & CLONE_VM)) {
2882 p->numa_preferred_nid = NUMA_NO_NODE;
2887 * New thread, keep existing numa_preferred_nid which should be copied
2888 * already by arch_dup_task_struct but stagger when scans start.
2893 delay = min_t(unsigned int, task_scan_max(current),
2894 current->numa_scan_period * mm_users * NSEC_PER_MSEC);
2895 delay += 2 * TICK_NSEC;
2896 p->node_stamp = delay;
2901 * Drive the periodic memory faults..
2903 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
2905 struct callback_head *work = &curr->numa_work;
2909 * We don't care about NUMA placement if we don't have memory.
2911 if (!curr->mm || (curr->flags & PF_EXITING) || work->next != work)
2915 * Using runtime rather than walltime has the dual advantage that
2916 * we (mostly) drive the selection from busy threads and that the
2917 * task needs to have done some actual work before we bother with
2920 now = curr->se.sum_exec_runtime;
2921 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
2923 if (now > curr->node_stamp + period) {
2924 if (!curr->node_stamp)
2925 curr->numa_scan_period = task_scan_start(curr);
2926 curr->node_stamp += period;
2928 if (!time_before(jiffies, curr->mm->numa_next_scan))
2929 task_work_add(curr, work, true);
2933 static void update_scan_period(struct task_struct *p, int new_cpu)
2935 int src_nid = cpu_to_node(task_cpu(p));
2936 int dst_nid = cpu_to_node(new_cpu);
2938 if (!static_branch_likely(&sched_numa_balancing))
2941 if (!p->mm || !p->numa_faults || (p->flags & PF_EXITING))
2944 if (src_nid == dst_nid)
2948 * Allow resets if faults have been trapped before one scan
2949 * has completed. This is most likely due to a new task that
2950 * is pulled cross-node due to wakeups or load balancing.
2952 if (p->numa_scan_seq) {
2954 * Avoid scan adjustments if moving to the preferred
2955 * node or if the task was not previously running on
2956 * the preferred node.
2958 if (dst_nid == p->numa_preferred_nid ||
2959 (p->numa_preferred_nid != NUMA_NO_NODE &&
2960 src_nid != p->numa_preferred_nid))
2964 p->numa_scan_period = task_scan_start(p);
2968 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
2972 static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p)
2976 static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p)
2980 static inline void update_scan_period(struct task_struct *p, int new_cpu)
2984 #endif /* CONFIG_NUMA_BALANCING */
2987 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2989 update_load_add(&cfs_rq->load, se->load.weight);
2991 if (entity_is_task(se)) {
2992 struct rq *rq = rq_of(cfs_rq);
2994 account_numa_enqueue(rq, task_of(se));
2995 list_add(&se->group_node, &rq->cfs_tasks);
2998 cfs_rq->nr_running++;
3002 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
3004 update_load_sub(&cfs_rq->load, se->load.weight);
3006 if (entity_is_task(se)) {
3007 account_numa_dequeue(rq_of(cfs_rq), task_of(se));
3008 list_del_init(&se->group_node);
3011 cfs_rq->nr_running--;
3015 * Signed add and clamp on underflow.
3017 * Explicitly do a load-store to ensure the intermediate value never hits
3018 * memory. This allows lockless observations without ever seeing the negative
3021 #define add_positive(_ptr, _val) do { \
3022 typeof(_ptr) ptr = (_ptr); \
3023 typeof(_val) val = (_val); \
3024 typeof(*ptr) res, var = READ_ONCE(*ptr); \
3028 if (val < 0 && res > var) \
3031 WRITE_ONCE(*ptr, res); \
3035 * Unsigned subtract and clamp on underflow.
3037 * Explicitly do a load-store to ensure the intermediate value never hits
3038 * memory. This allows lockless observations without ever seeing the negative
3041 #define sub_positive(_ptr, _val) do { \
3042 typeof(_ptr) ptr = (_ptr); \
3043 typeof(*ptr) val = (_val); \
3044 typeof(*ptr) res, var = READ_ONCE(*ptr); \
3048 WRITE_ONCE(*ptr, res); \
3052 * Remove and clamp on negative, from a local variable.
3054 * A variant of sub_positive(), which does not use explicit load-store
3055 * and is thus optimized for local variable updates.
3057 #define lsub_positive(_ptr, _val) do { \
3058 typeof(_ptr) ptr = (_ptr); \
3059 *ptr -= min_t(typeof(*ptr), *ptr, _val); \
3064 enqueue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3066 cfs_rq->avg.load_avg += se->avg.load_avg;
3067 cfs_rq->avg.load_sum += se_weight(se) * se->avg.load_sum;
3071 dequeue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3073 sub_positive(&cfs_rq->avg.load_avg, se->avg.load_avg);
3074 sub_positive(&cfs_rq->avg.load_sum, se_weight(se) * se->avg.load_sum);
3078 enqueue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
3080 dequeue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
3083 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
3084 unsigned long weight)
3087 /* commit outstanding execution time */
3088 if (cfs_rq->curr == se)
3089 update_curr(cfs_rq);
3090 account_entity_dequeue(cfs_rq, se);
3092 dequeue_load_avg(cfs_rq, se);
3094 update_load_set(&se->load, weight);
3098 u32 divider = LOAD_AVG_MAX - 1024 + se->avg.period_contrib;
3100 se->avg.load_avg = div_u64(se_weight(se) * se->avg.load_sum, divider);
3104 enqueue_load_avg(cfs_rq, se);
3106 account_entity_enqueue(cfs_rq, se);
3110 void reweight_task(struct task_struct *p, int prio)
3112 struct sched_entity *se = &p->se;
3113 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3114 struct load_weight *load = &se->load;
3115 unsigned long weight = scale_load(sched_prio_to_weight[prio]);
3117 reweight_entity(cfs_rq, se, weight);
3118 load->inv_weight = sched_prio_to_wmult[prio];
3121 #ifdef CONFIG_FAIR_GROUP_SCHED
3124 * All this does is approximate the hierarchical proportion which includes that
3125 * global sum we all love to hate.
3127 * That is, the weight of a group entity, is the proportional share of the
3128 * group weight based on the group runqueue weights. That is:
3130 * tg->weight * grq->load.weight
3131 * ge->load.weight = ----------------------------- (1)
3132 * \Sum grq->load.weight
3134 * Now, because computing that sum is prohibitively expensive to compute (been
3135 * there, done that) we approximate it with this average stuff. The average
3136 * moves slower and therefore the approximation is cheaper and more stable.
3138 * So instead of the above, we substitute:
3140 * grq->load.weight -> grq->avg.load_avg (2)
3142 * which yields the following:
3144 * tg->weight * grq->avg.load_avg
3145 * ge->load.weight = ------------------------------ (3)
3148 * Where: tg->load_avg ~= \Sum grq->avg.load_avg
3150 * That is shares_avg, and it is right (given the approximation (2)).
3152 * The problem with it is that because the average is slow -- it was designed
3153 * to be exactly that of course -- this leads to transients in boundary
3154 * conditions. In specific, the case where the group was idle and we start the
3155 * one task. It takes time for our CPU's grq->avg.load_avg to build up,
3156 * yielding bad latency etc..
3158 * Now, in that special case (1) reduces to:
3160 * tg->weight * grq->load.weight
3161 * ge->load.weight = ----------------------------- = tg->weight (4)
3164 * That is, the sum collapses because all other CPUs are idle; the UP scenario.
3166 * So what we do is modify our approximation (3) to approach (4) in the (near)
3171 * tg->weight * grq->load.weight
3172 * --------------------------------------------------- (5)
3173 * tg->load_avg - grq->avg.load_avg + grq->load.weight
3175 * But because grq->load.weight can drop to 0, resulting in a divide by zero,
3176 * we need to use grq->avg.load_avg as its lower bound, which then gives:
3179 * tg->weight * grq->load.weight
3180 * ge->load.weight = ----------------------------- (6)
3185 * tg_load_avg' = tg->load_avg - grq->avg.load_avg +
3186 * max(grq->load.weight, grq->avg.load_avg)
3188 * And that is shares_weight and is icky. In the (near) UP case it approaches
3189 * (4) while in the normal case it approaches (3). It consistently
3190 * overestimates the ge->load.weight and therefore:
3192 * \Sum ge->load.weight >= tg->weight
3196 static long calc_group_shares(struct cfs_rq *cfs_rq)
3198 long tg_weight, tg_shares, load, shares;
3199 struct task_group *tg = cfs_rq->tg;
3201 tg_shares = READ_ONCE(tg->shares);
3203 load = max(scale_load_down(cfs_rq->load.weight), cfs_rq->avg.load_avg);
3205 tg_weight = atomic_long_read(&tg->load_avg);
3207 /* Ensure tg_weight >= load */
3208 tg_weight -= cfs_rq->tg_load_avg_contrib;
3211 shares = (tg_shares * load);
3213 shares /= tg_weight;
3216 * MIN_SHARES has to be unscaled here to support per-CPU partitioning
3217 * of a group with small tg->shares value. It is a floor value which is
3218 * assigned as a minimum load.weight to the sched_entity representing
3219 * the group on a CPU.
3221 * E.g. on 64-bit for a group with tg->shares of scale_load(15)=15*1024
3222 * on an 8-core system with 8 tasks each runnable on one CPU shares has
3223 * to be 15*1024*1/8=1920 instead of scale_load(MIN_SHARES)=2*1024. In
3224 * case no task is runnable on a CPU MIN_SHARES=2 should be returned
3227 return clamp_t(long, shares, MIN_SHARES, tg_shares);
3229 #endif /* CONFIG_SMP */
3231 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
3234 * Recomputes the group entity based on the current state of its group
3237 static void update_cfs_group(struct sched_entity *se)
3239 struct cfs_rq *gcfs_rq = group_cfs_rq(se);
3245 if (throttled_hierarchy(gcfs_rq))
3249 shares = READ_ONCE(gcfs_rq->tg->shares);
3251 if (likely(se->load.weight == shares))
3254 shares = calc_group_shares(gcfs_rq);
3257 reweight_entity(cfs_rq_of(se), se, shares);
3260 #else /* CONFIG_FAIR_GROUP_SCHED */
3261 static inline void update_cfs_group(struct sched_entity *se)
3264 #endif /* CONFIG_FAIR_GROUP_SCHED */
3266 static inline void cfs_rq_util_change(struct cfs_rq *cfs_rq, int flags)
3268 struct rq *rq = rq_of(cfs_rq);
3270 if (&rq->cfs == cfs_rq) {
3272 * There are a few boundary cases this might miss but it should
3273 * get called often enough that that should (hopefully) not be
3276 * It will not get called when we go idle, because the idle
3277 * thread is a different class (!fair), nor will the utilization
3278 * number include things like RT tasks.
3280 * As is, the util number is not freq-invariant (we'd have to
3281 * implement arch_scale_freq_capacity() for that).
3285 cpufreq_update_util(rq, flags);
3290 #ifdef CONFIG_FAIR_GROUP_SCHED
3292 * update_tg_load_avg - update the tg's load avg
3293 * @cfs_rq: the cfs_rq whose avg changed
3294 * @force: update regardless of how small the difference
3296 * This function 'ensures': tg->load_avg := \Sum tg->cfs_rq[]->avg.load.
3297 * However, because tg->load_avg is a global value there are performance
3300 * In order to avoid having to look at the other cfs_rq's, we use a
3301 * differential update where we store the last value we propagated. This in
3302 * turn allows skipping updates if the differential is 'small'.
3304 * Updating tg's load_avg is necessary before update_cfs_share().
3306 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force)
3308 long delta = cfs_rq->avg.load_avg - cfs_rq->tg_load_avg_contrib;
3311 * No need to update load_avg for root_task_group as it is not used.
3313 if (cfs_rq->tg == &root_task_group)
3316 if (force || abs(delta) > cfs_rq->tg_load_avg_contrib / 64) {
3317 atomic_long_add(delta, &cfs_rq->tg->load_avg);
3318 cfs_rq->tg_load_avg_contrib = cfs_rq->avg.load_avg;
3323 * Called within set_task_rq() right before setting a task's CPU. The
3324 * caller only guarantees p->pi_lock is held; no other assumptions,
3325 * including the state of rq->lock, should be made.
3327 void set_task_rq_fair(struct sched_entity *se,
3328 struct cfs_rq *prev, struct cfs_rq *next)
3330 u64 p_last_update_time;
3331 u64 n_last_update_time;
3333 if (!sched_feat(ATTACH_AGE_LOAD))
3337 * We are supposed to update the task to "current" time, then its up to
3338 * date and ready to go to new CPU/cfs_rq. But we have difficulty in
3339 * getting what current time is, so simply throw away the out-of-date
3340 * time. This will result in the wakee task is less decayed, but giving
3341 * the wakee more load sounds not bad.
3343 if (!(se->avg.last_update_time && prev))
3346 #ifndef CONFIG_64BIT
3348 u64 p_last_update_time_copy;
3349 u64 n_last_update_time_copy;
3352 p_last_update_time_copy = prev->load_last_update_time_copy;
3353 n_last_update_time_copy = next->load_last_update_time_copy;
3357 p_last_update_time = prev->avg.last_update_time;
3358 n_last_update_time = next->avg.last_update_time;
3360 } while (p_last_update_time != p_last_update_time_copy ||
3361 n_last_update_time != n_last_update_time_copy);
3364 p_last_update_time = prev->avg.last_update_time;
3365 n_last_update_time = next->avg.last_update_time;
3367 __update_load_avg_blocked_se(p_last_update_time, se);
3368 se->avg.last_update_time = n_last_update_time;
3373 * When on migration a sched_entity joins/leaves the PELT hierarchy, we need to
3374 * propagate its contribution. The key to this propagation is the invariant
3375 * that for each group:
3377 * ge->avg == grq->avg (1)
3379 * _IFF_ we look at the pure running and runnable sums. Because they
3380 * represent the very same entity, just at different points in the hierarchy.
3382 * Per the above update_tg_cfs_util() and update_tg_cfs_runnable() are trivial
3383 * and simply copies the running/runnable sum over (but still wrong, because
3384 * the group entity and group rq do not have their PELT windows aligned).
3386 * However, update_tg_cfs_load() is more complex. So we have:
3388 * ge->avg.load_avg = ge->load.weight * ge->avg.runnable_avg (2)
3390 * And since, like util, the runnable part should be directly transferable,
3391 * the following would _appear_ to be the straight forward approach:
3393 * grq->avg.load_avg = grq->load.weight * grq->avg.runnable_avg (3)
3395 * And per (1) we have:
3397 * ge->avg.runnable_avg == grq->avg.runnable_avg
3401 * ge->load.weight * grq->avg.load_avg
3402 * ge->avg.load_avg = ----------------------------------- (4)
3405 * Except that is wrong!
3407 * Because while for entities historical weight is not important and we
3408 * really only care about our future and therefore can consider a pure
3409 * runnable sum, runqueues can NOT do this.
3411 * We specifically want runqueues to have a load_avg that includes
3412 * historical weights. Those represent the blocked load, the load we expect
3413 * to (shortly) return to us. This only works by keeping the weights as
3414 * integral part of the sum. We therefore cannot decompose as per (3).
3416 * Another reason this doesn't work is that runnable isn't a 0-sum entity.
3417 * Imagine a rq with 2 tasks that each are runnable 2/3 of the time. Then the
3418 * rq itself is runnable anywhere between 2/3 and 1 depending on how the
3419 * runnable section of these tasks overlap (or not). If they were to perfectly
3420 * align the rq as a whole would be runnable 2/3 of the time. If however we
3421 * always have at least 1 runnable task, the rq as a whole is always runnable.
3423 * So we'll have to approximate.. :/
3425 * Given the constraint:
3427 * ge->avg.running_sum <= ge->avg.runnable_sum <= LOAD_AVG_MAX
3429 * We can construct a rule that adds runnable to a rq by assuming minimal
3432 * On removal, we'll assume each task is equally runnable; which yields:
3434 * grq->avg.runnable_sum = grq->avg.load_sum / grq->load.weight
3436 * XXX: only do this for the part of runnable > running ?
3441 update_tg_cfs_util(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq)
3443 long delta = gcfs_rq->avg.util_avg - se->avg.util_avg;
3445 /* Nothing to update */
3450 * The relation between sum and avg is:
3452 * LOAD_AVG_MAX - 1024 + sa->period_contrib
3454 * however, the PELT windows are not aligned between grq and gse.
3457 /* Set new sched_entity's utilization */
3458 se->avg.util_avg = gcfs_rq->avg.util_avg;
3459 se->avg.util_sum = se->avg.util_avg * LOAD_AVG_MAX;
3461 /* Update parent cfs_rq utilization */
3462 add_positive(&cfs_rq->avg.util_avg, delta);
3463 cfs_rq->avg.util_sum = cfs_rq->avg.util_avg * LOAD_AVG_MAX;
3467 update_tg_cfs_runnable(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq)
3469 long delta = gcfs_rq->avg.runnable_avg - se->avg.runnable_avg;
3471 /* Nothing to update */
3476 * The relation between sum and avg is:
3478 * LOAD_AVG_MAX - 1024 + sa->period_contrib
3480 * however, the PELT windows are not aligned between grq and gse.
3483 /* Set new sched_entity's runnable */
3484 se->avg.runnable_avg = gcfs_rq->avg.runnable_avg;
3485 se->avg.runnable_sum = se->avg.runnable_avg * LOAD_AVG_MAX;
3487 /* Update parent cfs_rq runnable */
3488 add_positive(&cfs_rq->avg.runnable_avg, delta);
3489 cfs_rq->avg.runnable_sum = cfs_rq->avg.runnable_avg * LOAD_AVG_MAX;
3493 update_tg_cfs_load(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq)
3495 long delta_avg, running_sum, runnable_sum = gcfs_rq->prop_runnable_sum;
3496 unsigned long load_avg;
3503 gcfs_rq->prop_runnable_sum = 0;
3505 if (runnable_sum >= 0) {
3507 * Add runnable; clip at LOAD_AVG_MAX. Reflects that until
3508 * the CPU is saturated running == runnable.
3510 runnable_sum += se->avg.load_sum;
3511 runnable_sum = min(runnable_sum, (long)LOAD_AVG_MAX);
3514 * Estimate the new unweighted runnable_sum of the gcfs_rq by
3515 * assuming all tasks are equally runnable.
3517 if (scale_load_down(gcfs_rq->load.weight)) {
3518 load_sum = div_s64(gcfs_rq->avg.load_sum,
3519 scale_load_down(gcfs_rq->load.weight));
3522 /* But make sure to not inflate se's runnable */
3523 runnable_sum = min(se->avg.load_sum, load_sum);
3527 * runnable_sum can't be lower than running_sum
3528 * Rescale running sum to be in the same range as runnable sum
3529 * running_sum is in [0 : LOAD_AVG_MAX << SCHED_CAPACITY_SHIFT]
3530 * runnable_sum is in [0 : LOAD_AVG_MAX]
3532 running_sum = se->avg.util_sum >> SCHED_CAPACITY_SHIFT;
3533 runnable_sum = max(runnable_sum, running_sum);
3535 load_sum = (s64)se_weight(se) * runnable_sum;
3536 load_avg = div_s64(load_sum, LOAD_AVG_MAX);
3538 delta_sum = load_sum - (s64)se_weight(se) * se->avg.load_sum;
3539 delta_avg = load_avg - se->avg.load_avg;
3541 se->avg.load_sum = runnable_sum;
3542 se->avg.load_avg = load_avg;
3543 add_positive(&cfs_rq->avg.load_avg, delta_avg);
3544 add_positive(&cfs_rq->avg.load_sum, delta_sum);
3547 static inline void add_tg_cfs_propagate(struct cfs_rq *cfs_rq, long runnable_sum)
3549 cfs_rq->propagate = 1;
3550 cfs_rq->prop_runnable_sum += runnable_sum;
3553 /* Update task and its cfs_rq load average */
3554 static inline int propagate_entity_load_avg(struct sched_entity *se)
3556 struct cfs_rq *cfs_rq, *gcfs_rq;
3558 if (entity_is_task(se))
3561 gcfs_rq = group_cfs_rq(se);
3562 if (!gcfs_rq->propagate)
3565 gcfs_rq->propagate = 0;
3567 cfs_rq = cfs_rq_of(se);
3569 add_tg_cfs_propagate(cfs_rq, gcfs_rq->prop_runnable_sum);
3571 update_tg_cfs_util(cfs_rq, se, gcfs_rq);
3572 update_tg_cfs_runnable(cfs_rq, se, gcfs_rq);
3573 update_tg_cfs_load(cfs_rq, se, gcfs_rq);
3575 trace_pelt_cfs_tp(cfs_rq);
3576 trace_pelt_se_tp(se);
3582 * Check if we need to update the load and the utilization of a blocked
3585 static inline bool skip_blocked_update(struct sched_entity *se)
3587 struct cfs_rq *gcfs_rq = group_cfs_rq(se);
3590 * If sched_entity still have not zero load or utilization, we have to
3593 if (se->avg.load_avg || se->avg.util_avg)
3597 * If there is a pending propagation, we have to update the load and
3598 * the utilization of the sched_entity:
3600 if (gcfs_rq->propagate)
3604 * Otherwise, the load and the utilization of the sched_entity is
3605 * already zero and there is no pending propagation, so it will be a
3606 * waste of time to try to decay it:
3611 #else /* CONFIG_FAIR_GROUP_SCHED */
3613 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq, int force) {}
3615 static inline int propagate_entity_load_avg(struct sched_entity *se)
3620 static inline void add_tg_cfs_propagate(struct cfs_rq *cfs_rq, long runnable_sum) {}
3622 #endif /* CONFIG_FAIR_GROUP_SCHED */
3625 * update_cfs_rq_load_avg - update the cfs_rq's load/util averages
3626 * @now: current time, as per cfs_rq_clock_pelt()
3627 * @cfs_rq: cfs_rq to update
3629 * The cfs_rq avg is the direct sum of all its entities (blocked and runnable)
3630 * avg. The immediate corollary is that all (fair) tasks must be attached, see
3631 * post_init_entity_util_avg().
3633 * cfs_rq->avg is used for task_h_load() and update_cfs_share() for example.
3635 * Returns true if the load decayed or we removed load.
3637 * Since both these conditions indicate a changed cfs_rq->avg.load we should
3638 * call update_tg_load_avg() when this function returns true.
3641 update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq)
3643 unsigned long removed_load = 0, removed_util = 0, removed_runnable = 0;
3644 struct sched_avg *sa = &cfs_rq->avg;
3647 if (cfs_rq->removed.nr) {
3649 u32 divider = LOAD_AVG_MAX - 1024 + sa->period_contrib;
3651 raw_spin_lock(&cfs_rq->removed.lock);
3652 swap(cfs_rq->removed.util_avg, removed_util);
3653 swap(cfs_rq->removed.load_avg, removed_load);
3654 swap(cfs_rq->removed.runnable_avg, removed_runnable);
3655 cfs_rq->removed.nr = 0;
3656 raw_spin_unlock(&cfs_rq->removed.lock);
3659 sub_positive(&sa->load_avg, r);
3660 sub_positive(&sa->load_sum, r * divider);
3663 sub_positive(&sa->util_avg, r);
3664 sub_positive(&sa->util_sum, r * divider);
3666 r = removed_runnable;
3667 sub_positive(&sa->runnable_avg, r);
3668 sub_positive(&sa->runnable_sum, r * divider);
3671 * removed_runnable is the unweighted version of removed_load so we
3672 * can use it to estimate removed_load_sum.
3674 add_tg_cfs_propagate(cfs_rq,
3675 -(long)(removed_runnable * divider) >> SCHED_CAPACITY_SHIFT);
3680 decayed |= __update_load_avg_cfs_rq(now, cfs_rq);
3682 #ifndef CONFIG_64BIT
3684 cfs_rq->load_last_update_time_copy = sa->last_update_time;
3691 * attach_entity_load_avg - attach this entity to its cfs_rq load avg
3692 * @cfs_rq: cfs_rq to attach to
3693 * @se: sched_entity to attach
3695 * Must call update_cfs_rq_load_avg() before this, since we rely on
3696 * cfs_rq->avg.last_update_time being current.
3698 static void attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3700 u32 divider = LOAD_AVG_MAX - 1024 + cfs_rq->avg.period_contrib;
3703 * When we attach the @se to the @cfs_rq, we must align the decay
3704 * window because without that, really weird and wonderful things can
3709 se->avg.last_update_time = cfs_rq->avg.last_update_time;
3710 se->avg.period_contrib = cfs_rq->avg.period_contrib;
3713 * Hell(o) Nasty stuff.. we need to recompute _sum based on the new
3714 * period_contrib. This isn't strictly correct, but since we're
3715 * entirely outside of the PELT hierarchy, nobody cares if we truncate
3718 se->avg.util_sum = se->avg.util_avg * divider;
3720 se->avg.runnable_sum = se->avg.runnable_avg * divider;
3722 se->avg.load_sum = divider;
3723 if (se_weight(se)) {
3725 div_u64(se->avg.load_avg * se->avg.load_sum, se_weight(se));
3728 enqueue_load_avg(cfs_rq, se);
3729 cfs_rq->avg.util_avg += se->avg.util_avg;
3730 cfs_rq->avg.util_sum += se->avg.util_sum;
3731 cfs_rq->avg.runnable_avg += se->avg.runnable_avg;
3732 cfs_rq->avg.runnable_sum += se->avg.runnable_sum;
3734 add_tg_cfs_propagate(cfs_rq, se->avg.load_sum);
3736 cfs_rq_util_change(cfs_rq, 0);
3738 trace_pelt_cfs_tp(cfs_rq);
3742 * detach_entity_load_avg - detach this entity from its cfs_rq load avg
3743 * @cfs_rq: cfs_rq to detach from
3744 * @se: sched_entity to detach
3746 * Must call update_cfs_rq_load_avg() before this, since we rely on
3747 * cfs_rq->avg.last_update_time being current.
3749 static void detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3751 dequeue_load_avg(cfs_rq, se);
3752 sub_positive(&cfs_rq->avg.util_avg, se->avg.util_avg);
3753 sub_positive(&cfs_rq->avg.util_sum, se->avg.util_sum);
3754 sub_positive(&cfs_rq->avg.runnable_avg, se->avg.runnable_avg);
3755 sub_positive(&cfs_rq->avg.runnable_sum, se->avg.runnable_sum);
3757 add_tg_cfs_propagate(cfs_rq, -se->avg.load_sum);
3759 cfs_rq_util_change(cfs_rq, 0);
3761 trace_pelt_cfs_tp(cfs_rq);
3765 * Optional action to be done while updating the load average
3767 #define UPDATE_TG 0x1
3768 #define SKIP_AGE_LOAD 0x2
3769 #define DO_ATTACH 0x4
3771 /* Update task and its cfs_rq load average */
3772 static inline void update_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3774 u64 now = cfs_rq_clock_pelt(cfs_rq);
3778 * Track task load average for carrying it to new CPU after migrated, and
3779 * track group sched_entity load average for task_h_load calc in migration
3781 if (se->avg.last_update_time && !(flags & SKIP_AGE_LOAD))
3782 __update_load_avg_se(now, cfs_rq, se);
3784 decayed = update_cfs_rq_load_avg(now, cfs_rq);
3785 decayed |= propagate_entity_load_avg(se);
3787 if (!se->avg.last_update_time && (flags & DO_ATTACH)) {
3790 * DO_ATTACH means we're here from enqueue_entity().
3791 * !last_update_time means we've passed through
3792 * migrate_task_rq_fair() indicating we migrated.
3794 * IOW we're enqueueing a task on a new CPU.
3796 attach_entity_load_avg(cfs_rq, se);
3797 update_tg_load_avg(cfs_rq, 0);
3799 } else if (decayed) {
3800 cfs_rq_util_change(cfs_rq, 0);
3802 if (flags & UPDATE_TG)
3803 update_tg_load_avg(cfs_rq, 0);
3807 #ifndef CONFIG_64BIT
3808 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
3810 u64 last_update_time_copy;
3811 u64 last_update_time;
3814 last_update_time_copy = cfs_rq->load_last_update_time_copy;
3816 last_update_time = cfs_rq->avg.last_update_time;
3817 } while (last_update_time != last_update_time_copy);
3819 return last_update_time;
3822 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
3824 return cfs_rq->avg.last_update_time;
3829 * Synchronize entity load avg of dequeued entity without locking
3832 static void sync_entity_load_avg(struct sched_entity *se)
3834 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3835 u64 last_update_time;
3837 last_update_time = cfs_rq_last_update_time(cfs_rq);
3838 __update_load_avg_blocked_se(last_update_time, se);
3842 * Task first catches up with cfs_rq, and then subtract
3843 * itself from the cfs_rq (task must be off the queue now).
3845 static void remove_entity_load_avg(struct sched_entity *se)
3847 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3848 unsigned long flags;
3851 * tasks cannot exit without having gone through wake_up_new_task() ->
3852 * post_init_entity_util_avg() which will have added things to the
3853 * cfs_rq, so we can remove unconditionally.
3856 sync_entity_load_avg(se);
3858 raw_spin_lock_irqsave(&cfs_rq->removed.lock, flags);
3859 ++cfs_rq->removed.nr;
3860 cfs_rq->removed.util_avg += se->avg.util_avg;
3861 cfs_rq->removed.load_avg += se->avg.load_avg;
3862 cfs_rq->removed.runnable_avg += se->avg.runnable_avg;
3863 raw_spin_unlock_irqrestore(&cfs_rq->removed.lock, flags);
3866 static inline unsigned long cfs_rq_runnable_avg(struct cfs_rq *cfs_rq)
3868 return cfs_rq->avg.runnable_avg;
3871 static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq)
3873 return cfs_rq->avg.load_avg;
3876 static inline unsigned long task_util(struct task_struct *p)
3878 return READ_ONCE(p->se.avg.util_avg);
3881 static inline unsigned long _task_util_est(struct task_struct *p)
3883 struct util_est ue = READ_ONCE(p->se.avg.util_est);
3885 return (max(ue.ewma, ue.enqueued) | UTIL_AVG_UNCHANGED);
3888 static inline unsigned long task_util_est(struct task_struct *p)
3890 return max(task_util(p), _task_util_est(p));
3893 #ifdef CONFIG_UCLAMP_TASK
3894 static inline unsigned long uclamp_task_util(struct task_struct *p)
3896 return clamp(task_util_est(p),
3897 uclamp_eff_value(p, UCLAMP_MIN),
3898 uclamp_eff_value(p, UCLAMP_MAX));
3901 static inline unsigned long uclamp_task_util(struct task_struct *p)
3903 return task_util_est(p);
3907 static inline void util_est_enqueue(struct cfs_rq *cfs_rq,
3908 struct task_struct *p)
3910 unsigned int enqueued;
3912 if (!sched_feat(UTIL_EST))
3915 /* Update root cfs_rq's estimated utilization */
3916 enqueued = cfs_rq->avg.util_est.enqueued;
3917 enqueued += _task_util_est(p);
3918 WRITE_ONCE(cfs_rq->avg.util_est.enqueued, enqueued);
3922 * Check if a (signed) value is within a specified (unsigned) margin,
3923 * based on the observation that:
3925 * abs(x) < y := (unsigned)(x + y - 1) < (2 * y - 1)
3927 * NOTE: this only works when value + maring < INT_MAX.
3929 static inline bool within_margin(int value, int margin)
3931 return ((unsigned int)(value + margin - 1) < (2 * margin - 1));
3935 util_est_dequeue(struct cfs_rq *cfs_rq, struct task_struct *p, bool task_sleep)
3937 long last_ewma_diff;
3941 if (!sched_feat(UTIL_EST))
3944 /* Update root cfs_rq's estimated utilization */
3945 ue.enqueued = cfs_rq->avg.util_est.enqueued;
3946 ue.enqueued -= min_t(unsigned int, ue.enqueued, _task_util_est(p));
3947 WRITE_ONCE(cfs_rq->avg.util_est.enqueued, ue.enqueued);
3950 * Skip update of task's estimated utilization when the task has not
3951 * yet completed an activation, e.g. being migrated.
3957 * If the PELT values haven't changed since enqueue time,
3958 * skip the util_est update.
3960 ue = p->se.avg.util_est;
3961 if (ue.enqueued & UTIL_AVG_UNCHANGED)
3965 * Reset EWMA on utilization increases, the moving average is used only
3966 * to smooth utilization decreases.
3968 ue.enqueued = (task_util(p) | UTIL_AVG_UNCHANGED);
3969 if (sched_feat(UTIL_EST_FASTUP)) {
3970 if (ue.ewma < ue.enqueued) {
3971 ue.ewma = ue.enqueued;
3977 * Skip update of task's estimated utilization when its EWMA is
3978 * already ~1% close to its last activation value.
3980 last_ewma_diff = ue.enqueued - ue.ewma;
3981 if (within_margin(last_ewma_diff, (SCHED_CAPACITY_SCALE / 100)))
3985 * To avoid overestimation of actual task utilization, skip updates if
3986 * we cannot grant there is idle time in this CPU.
3988 cpu = cpu_of(rq_of(cfs_rq));
3989 if (task_util(p) > capacity_orig_of(cpu))
3993 * Update Task's estimated utilization
3995 * When *p completes an activation we can consolidate another sample
3996 * of the task size. This is done by storing the current PELT value
3997 * as ue.enqueued and by using this value to update the Exponential
3998 * Weighted Moving Average (EWMA):
4000 * ewma(t) = w * task_util(p) + (1-w) * ewma(t-1)
4001 * = w * task_util(p) + ewma(t-1) - w * ewma(t-1)
4002 * = w * (task_util(p) - ewma(t-1)) + ewma(t-1)
4003 * = w * ( last_ewma_diff ) + ewma(t-1)
4004 * = w * (last_ewma_diff + ewma(t-1) / w)
4006 * Where 'w' is the weight of new samples, which is configured to be
4007 * 0.25, thus making w=1/4 ( >>= UTIL_EST_WEIGHT_SHIFT)
4009 ue.ewma <<= UTIL_EST_WEIGHT_SHIFT;
4010 ue.ewma += last_ewma_diff;
4011 ue.ewma >>= UTIL_EST_WEIGHT_SHIFT;
4013 WRITE_ONCE(p->se.avg.util_est, ue);
4016 static inline int task_fits_capacity(struct task_struct *p, long capacity)
4018 return fits_capacity(uclamp_task_util(p), capacity);
4021 static inline void update_misfit_status(struct task_struct *p, struct rq *rq)
4023 if (!static_branch_unlikely(&sched_asym_cpucapacity))
4027 rq->misfit_task_load = 0;
4031 if (task_fits_capacity(p, capacity_of(cpu_of(rq)))) {
4032 rq->misfit_task_load = 0;
4036 rq->misfit_task_load = task_h_load(p);
4039 #else /* CONFIG_SMP */
4041 #define UPDATE_TG 0x0
4042 #define SKIP_AGE_LOAD 0x0
4043 #define DO_ATTACH 0x0
4045 static inline void update_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int not_used1)
4047 cfs_rq_util_change(cfs_rq, 0);
4050 static inline void remove_entity_load_avg(struct sched_entity *se) {}
4053 attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
4055 detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
4057 static inline int idle_balance(struct rq *rq, struct rq_flags *rf)
4063 util_est_enqueue(struct cfs_rq *cfs_rq, struct task_struct *p) {}
4066 util_est_dequeue(struct cfs_rq *cfs_rq, struct task_struct *p,
4068 static inline void update_misfit_status(struct task_struct *p, struct rq *rq) {}
4070 #endif /* CONFIG_SMP */
4072 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
4074 #ifdef CONFIG_SCHED_DEBUG
4075 s64 d = se->vruntime - cfs_rq->min_vruntime;
4080 if (d > 3*sysctl_sched_latency)
4081 schedstat_inc(cfs_rq->nr_spread_over);
4086 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
4088 u64 vruntime = cfs_rq->min_vruntime;
4091 * The 'current' period is already promised to the current tasks,
4092 * however the extra weight of the new task will slow them down a
4093 * little, place the new task so that it fits in the slot that
4094 * stays open at the end.
4096 if (initial && sched_feat(START_DEBIT))
4097 vruntime += sched_vslice(cfs_rq, se);
4099 /* sleeps up to a single latency don't count. */
4101 unsigned long thresh = sysctl_sched_latency;
4104 * Halve their sleep time's effect, to allow
4105 * for a gentler effect of sleepers:
4107 if (sched_feat(GENTLE_FAIR_SLEEPERS))
4113 /* ensure we never gain time by being placed backwards. */
4114 se->vruntime = max_vruntime(se->vruntime, vruntime);
4117 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
4119 static inline void check_schedstat_required(void)
4121 #ifdef CONFIG_SCHEDSTATS
4122 if (schedstat_enabled())
4125 /* Force schedstat enabled if a dependent tracepoint is active */
4126 if (trace_sched_stat_wait_enabled() ||
4127 trace_sched_stat_sleep_enabled() ||
4128 trace_sched_stat_iowait_enabled() ||
4129 trace_sched_stat_blocked_enabled() ||
4130 trace_sched_stat_runtime_enabled()) {
4131 printk_deferred_once("Scheduler tracepoints stat_sleep, stat_iowait, "
4132 "stat_blocked and stat_runtime require the "
4133 "kernel parameter schedstats=enable or "
4134 "kernel.sched_schedstats=1\n");
4139 static inline bool cfs_bandwidth_used(void);
4146 * update_min_vruntime()
4147 * vruntime -= min_vruntime
4151 * update_min_vruntime()
4152 * vruntime += min_vruntime
4154 * this way the vruntime transition between RQs is done when both
4155 * min_vruntime are up-to-date.
4159 * ->migrate_task_rq_fair() (p->state == TASK_WAKING)
4160 * vruntime -= min_vruntime
4164 * update_min_vruntime()
4165 * vruntime += min_vruntime
4167 * this way we don't have the most up-to-date min_vruntime on the originating
4168 * CPU and an up-to-date min_vruntime on the destination CPU.
4172 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
4174 bool renorm = !(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_MIGRATED);
4175 bool curr = cfs_rq->curr == se;
4178 * If we're the current task, we must renormalise before calling
4182 se->vruntime += cfs_rq->min_vruntime;
4184 update_curr(cfs_rq);
4187 * Otherwise, renormalise after, such that we're placed at the current
4188 * moment in time, instead of some random moment in the past. Being
4189 * placed in the past could significantly boost this task to the
4190 * fairness detriment of existing tasks.
4192 if (renorm && !curr)
4193 se->vruntime += cfs_rq->min_vruntime;
4196 * When enqueuing a sched_entity, we must:
4197 * - Update loads to have both entity and cfs_rq synced with now.
4198 * - Add its load to cfs_rq->runnable_avg
4199 * - For group_entity, update its weight to reflect the new share of
4201 * - Add its new weight to cfs_rq->load.weight
4203 update_load_avg(cfs_rq, se, UPDATE_TG | DO_ATTACH);
4204 se_update_runnable(se);
4205 update_cfs_group(se);
4206 account_entity_enqueue(cfs_rq, se);
4208 if (flags & ENQUEUE_WAKEUP)
4209 place_entity(cfs_rq, se, 0);
4211 check_schedstat_required();
4212 update_stats_enqueue(cfs_rq, se, flags);
4213 check_spread(cfs_rq, se);
4215 __enqueue_entity(cfs_rq, se);
4219 * When bandwidth control is enabled, cfs might have been removed
4220 * because of a parent been throttled but cfs->nr_running > 1. Try to
4221 * add it unconditionnally.
4223 if (cfs_rq->nr_running == 1 || cfs_bandwidth_used())
4224 list_add_leaf_cfs_rq(cfs_rq);
4226 if (cfs_rq->nr_running == 1)
4227 check_enqueue_throttle(cfs_rq);
4230 static void __clear_buddies_last(struct sched_entity *se)
4232 for_each_sched_entity(se) {
4233 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4234 if (cfs_rq->last != se)
4237 cfs_rq->last = NULL;
4241 static void __clear_buddies_next(struct sched_entity *se)
4243 for_each_sched_entity(se) {
4244 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4245 if (cfs_rq->next != se)
4248 cfs_rq->next = NULL;
4252 static void __clear_buddies_skip(struct sched_entity *se)
4254 for_each_sched_entity(se) {
4255 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4256 if (cfs_rq->skip != se)
4259 cfs_rq->skip = NULL;
4263 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
4265 if (cfs_rq->last == se)
4266 __clear_buddies_last(se);
4268 if (cfs_rq->next == se)
4269 __clear_buddies_next(se);
4271 if (cfs_rq->skip == se)
4272 __clear_buddies_skip(se);
4275 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
4278 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
4281 * Update run-time statistics of the 'current'.
4283 update_curr(cfs_rq);
4286 * When dequeuing a sched_entity, we must:
4287 * - Update loads to have both entity and cfs_rq synced with now.
4288 * - Subtract its load from the cfs_rq->runnable_avg.
4289 * - Subtract its previous weight from cfs_rq->load.weight.
4290 * - For group entity, update its weight to reflect the new share
4291 * of its group cfs_rq.
4293 update_load_avg(cfs_rq, se, UPDATE_TG);
4294 se_update_runnable(se);
4296 update_stats_dequeue(cfs_rq, se, flags);
4298 clear_buddies(cfs_rq, se);
4300 if (se != cfs_rq->curr)
4301 __dequeue_entity(cfs_rq, se);
4303 account_entity_dequeue(cfs_rq, se);
4306 * Normalize after update_curr(); which will also have moved
4307 * min_vruntime if @se is the one holding it back. But before doing
4308 * update_min_vruntime() again, which will discount @se's position and
4309 * can move min_vruntime forward still more.
4311 if (!(flags & DEQUEUE_SLEEP))
4312 se->vruntime -= cfs_rq->min_vruntime;
4314 /* return excess runtime on last dequeue */
4315 return_cfs_rq_runtime(cfs_rq);
4317 update_cfs_group(se);
4320 * Now advance min_vruntime if @se was the entity holding it back,
4321 * except when: DEQUEUE_SAVE && !DEQUEUE_MOVE, in this case we'll be
4322 * put back on, and if we advance min_vruntime, we'll be placed back
4323 * further than we started -- ie. we'll be penalized.
4325 if ((flags & (DEQUEUE_SAVE | DEQUEUE_MOVE)) != DEQUEUE_SAVE)
4326 update_min_vruntime(cfs_rq);
4330 * Preempt the current task with a newly woken task if needed:
4333 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
4335 unsigned long ideal_runtime, delta_exec;
4336 struct sched_entity *se;
4339 ideal_runtime = sched_slice(cfs_rq, curr);
4340 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
4341 if (delta_exec > ideal_runtime) {
4342 resched_curr(rq_of(cfs_rq));
4344 * The current task ran long enough, ensure it doesn't get
4345 * re-elected due to buddy favours.
4347 clear_buddies(cfs_rq, curr);
4352 * Ensure that a task that missed wakeup preemption by a
4353 * narrow margin doesn't have to wait for a full slice.
4354 * This also mitigates buddy induced latencies under load.
4356 if (delta_exec < sysctl_sched_min_granularity)
4359 se = __pick_first_entity(cfs_rq);
4360 delta = curr->vruntime - se->vruntime;
4365 if (delta > ideal_runtime)
4366 resched_curr(rq_of(cfs_rq));
4370 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
4372 /* 'current' is not kept within the tree. */
4375 * Any task has to be enqueued before it get to execute on
4376 * a CPU. So account for the time it spent waiting on the
4379 update_stats_wait_end(cfs_rq, se);
4380 __dequeue_entity(cfs_rq, se);
4381 update_load_avg(cfs_rq, se, UPDATE_TG);
4384 update_stats_curr_start(cfs_rq, se);
4388 * Track our maximum slice length, if the CPU's load is at
4389 * least twice that of our own weight (i.e. dont track it
4390 * when there are only lesser-weight tasks around):
4392 if (schedstat_enabled() &&
4393 rq_of(cfs_rq)->cfs.load.weight >= 2*se->load.weight) {
4394 schedstat_set(se->statistics.slice_max,
4395 max((u64)schedstat_val(se->statistics.slice_max),
4396 se->sum_exec_runtime - se->prev_sum_exec_runtime));
4399 se->prev_sum_exec_runtime = se->sum_exec_runtime;
4403 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
4406 * Pick the next process, keeping these things in mind, in this order:
4407 * 1) keep things fair between processes/task groups
4408 * 2) pick the "next" process, since someone really wants that to run
4409 * 3) pick the "last" process, for cache locality
4410 * 4) do not run the "skip" process, if something else is available
4412 static struct sched_entity *
4413 pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr)
4415 struct sched_entity *left = __pick_first_entity(cfs_rq);
4416 struct sched_entity *se;
4419 * If curr is set we have to see if its left of the leftmost entity
4420 * still in the tree, provided there was anything in the tree at all.
4422 if (!left || (curr && entity_before(curr, left)))
4425 se = left; /* ideally we run the leftmost entity */
4428 * Avoid running the skip buddy, if running something else can
4429 * be done without getting too unfair.
4431 if (cfs_rq->skip == se) {
4432 struct sched_entity *second;
4435 second = __pick_first_entity(cfs_rq);
4437 second = __pick_next_entity(se);
4438 if (!second || (curr && entity_before(curr, second)))
4442 if (second && wakeup_preempt_entity(second, left) < 1)
4447 * Prefer last buddy, try to return the CPU to a preempted task.
4449 if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
4453 * Someone really wants this to run. If it's not unfair, run it.
4455 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
4458 clear_buddies(cfs_rq, se);
4463 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
4465 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
4468 * If still on the runqueue then deactivate_task()
4469 * was not called and update_curr() has to be done:
4472 update_curr(cfs_rq);
4474 /* throttle cfs_rqs exceeding runtime */
4475 check_cfs_rq_runtime(cfs_rq);
4477 check_spread(cfs_rq, prev);
4480 update_stats_wait_start(cfs_rq, prev);
4481 /* Put 'current' back into the tree. */
4482 __enqueue_entity(cfs_rq, prev);
4483 /* in !on_rq case, update occurred at dequeue */
4484 update_load_avg(cfs_rq, prev, 0);
4486 cfs_rq->curr = NULL;
4490 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
4493 * Update run-time statistics of the 'current'.
4495 update_curr(cfs_rq);
4498 * Ensure that runnable average is periodically updated.
4500 update_load_avg(cfs_rq, curr, UPDATE_TG);
4501 update_cfs_group(curr);
4503 #ifdef CONFIG_SCHED_HRTICK
4505 * queued ticks are scheduled to match the slice, so don't bother
4506 * validating it and just reschedule.
4509 resched_curr(rq_of(cfs_rq));
4513 * don't let the period tick interfere with the hrtick preemption
4515 if (!sched_feat(DOUBLE_TICK) &&
4516 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
4520 if (cfs_rq->nr_running > 1)
4521 check_preempt_tick(cfs_rq, curr);
4525 /**************************************************
4526 * CFS bandwidth control machinery
4529 #ifdef CONFIG_CFS_BANDWIDTH
4531 #ifdef CONFIG_JUMP_LABEL
4532 static struct static_key __cfs_bandwidth_used;
4534 static inline bool cfs_bandwidth_used(void)
4536 return static_key_false(&__cfs_bandwidth_used);
4539 void cfs_bandwidth_usage_inc(void)
4541 static_key_slow_inc_cpuslocked(&__cfs_bandwidth_used);
4544 void cfs_bandwidth_usage_dec(void)
4546 static_key_slow_dec_cpuslocked(&__cfs_bandwidth_used);
4548 #else /* CONFIG_JUMP_LABEL */
4549 static bool cfs_bandwidth_used(void)
4554 void cfs_bandwidth_usage_inc(void) {}
4555 void cfs_bandwidth_usage_dec(void) {}
4556 #endif /* CONFIG_JUMP_LABEL */
4559 * default period for cfs group bandwidth.
4560 * default: 0.1s, units: nanoseconds
4562 static inline u64 default_cfs_period(void)
4564 return 100000000ULL;
4567 static inline u64 sched_cfs_bandwidth_slice(void)
4569 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
4573 * Replenish runtime according to assigned quota. We use sched_clock_cpu
4574 * directly instead of rq->clock to avoid adding additional synchronization
4577 * requires cfs_b->lock
4579 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
4581 if (cfs_b->quota != RUNTIME_INF)
4582 cfs_b->runtime = cfs_b->quota;
4585 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
4587 return &tg->cfs_bandwidth;
4590 /* returns 0 on failure to allocate runtime */
4591 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4593 struct task_group *tg = cfs_rq->tg;
4594 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
4595 u64 amount = 0, min_amount;
4597 /* note: this is a positive sum as runtime_remaining <= 0 */
4598 min_amount = sched_cfs_bandwidth_slice() - cfs_rq->runtime_remaining;
4600 raw_spin_lock(&cfs_b->lock);
4601 if (cfs_b->quota == RUNTIME_INF)
4602 amount = min_amount;
4604 start_cfs_bandwidth(cfs_b);
4606 if (cfs_b->runtime > 0) {
4607 amount = min(cfs_b->runtime, min_amount);
4608 cfs_b->runtime -= amount;
4612 raw_spin_unlock(&cfs_b->lock);
4614 cfs_rq->runtime_remaining += amount;
4616 return cfs_rq->runtime_remaining > 0;
4619 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
4621 /* dock delta_exec before expiring quota (as it could span periods) */
4622 cfs_rq->runtime_remaining -= delta_exec;
4624 if (likely(cfs_rq->runtime_remaining > 0))
4627 if (cfs_rq->throttled)
4630 * if we're unable to extend our runtime we resched so that the active
4631 * hierarchy can be throttled
4633 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
4634 resched_curr(rq_of(cfs_rq));
4637 static __always_inline
4638 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
4640 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
4643 __account_cfs_rq_runtime(cfs_rq, delta_exec);
4646 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
4648 return cfs_bandwidth_used() && cfs_rq->throttled;
4651 /* check whether cfs_rq, or any parent, is throttled */
4652 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
4654 return cfs_bandwidth_used() && cfs_rq->throttle_count;
4658 * Ensure that neither of the group entities corresponding to src_cpu or
4659 * dest_cpu are members of a throttled hierarchy when performing group
4660 * load-balance operations.
4662 static inline int throttled_lb_pair(struct task_group *tg,
4663 int src_cpu, int dest_cpu)
4665 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
4667 src_cfs_rq = tg->cfs_rq[src_cpu];
4668 dest_cfs_rq = tg->cfs_rq[dest_cpu];
4670 return throttled_hierarchy(src_cfs_rq) ||
4671 throttled_hierarchy(dest_cfs_rq);
4674 static int tg_unthrottle_up(struct task_group *tg, void *data)
4676 struct rq *rq = data;
4677 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
4679 cfs_rq->throttle_count--;
4680 if (!cfs_rq->throttle_count) {
4681 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
4682 cfs_rq->throttled_clock_task;
4684 /* Add cfs_rq with already running entity in the list */
4685 if (cfs_rq->nr_running >= 1)
4686 list_add_leaf_cfs_rq(cfs_rq);
4692 static int tg_throttle_down(struct task_group *tg, void *data)
4694 struct rq *rq = data;
4695 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
4697 /* group is entering throttled state, stop time */
4698 if (!cfs_rq->throttle_count) {
4699 cfs_rq->throttled_clock_task = rq_clock_task(rq);
4700 list_del_leaf_cfs_rq(cfs_rq);
4702 cfs_rq->throttle_count++;
4707 static void throttle_cfs_rq(struct cfs_rq *cfs_rq)
4709 struct rq *rq = rq_of(cfs_rq);
4710 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4711 struct sched_entity *se;
4712 long task_delta, idle_task_delta, dequeue = 1;
4715 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
4717 /* freeze hierarchy runnable averages while throttled */
4719 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
4722 task_delta = cfs_rq->h_nr_running;
4723 idle_task_delta = cfs_rq->idle_h_nr_running;
4724 for_each_sched_entity(se) {
4725 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
4726 /* throttled entity or throttle-on-deactivate */
4731 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
4733 update_load_avg(qcfs_rq, se, 0);
4734 se_update_runnable(se);
4737 qcfs_rq->h_nr_running -= task_delta;
4738 qcfs_rq->idle_h_nr_running -= idle_task_delta;
4740 if (qcfs_rq->load.weight)
4745 sub_nr_running(rq, task_delta);
4747 cfs_rq->throttled = 1;
4748 cfs_rq->throttled_clock = rq_clock(rq);
4749 raw_spin_lock(&cfs_b->lock);
4750 empty = list_empty(&cfs_b->throttled_cfs_rq);
4753 * Add to the _head_ of the list, so that an already-started
4754 * distribute_cfs_runtime will not see us. If disribute_cfs_runtime is
4755 * not running add to the tail so that later runqueues don't get starved.
4757 if (cfs_b->distribute_running)
4758 list_add_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
4760 list_add_tail_rcu(&cfs_rq->throttled_list, &cfs_b->throttled_cfs_rq);
4763 * If we're the first throttled task, make sure the bandwidth
4767 start_cfs_bandwidth(cfs_b);
4769 raw_spin_unlock(&cfs_b->lock);
4772 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
4774 struct rq *rq = rq_of(cfs_rq);
4775 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4776 struct sched_entity *se;
4778 long task_delta, idle_task_delta;
4780 se = cfs_rq->tg->se[cpu_of(rq)];
4782 cfs_rq->throttled = 0;
4784 update_rq_clock(rq);
4786 raw_spin_lock(&cfs_b->lock);
4787 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
4788 list_del_rcu(&cfs_rq->throttled_list);
4789 raw_spin_unlock(&cfs_b->lock);
4791 /* update hierarchical throttle state */
4792 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
4794 if (!cfs_rq->load.weight)
4797 task_delta = cfs_rq->h_nr_running;
4798 idle_task_delta = cfs_rq->idle_h_nr_running;
4799 for_each_sched_entity(se) {
4803 cfs_rq = cfs_rq_of(se);
4805 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
4807 update_load_avg(cfs_rq, se, 0);
4808 se_update_runnable(se);
4811 cfs_rq->h_nr_running += task_delta;
4812 cfs_rq->idle_h_nr_running += idle_task_delta;
4814 if (cfs_rq_throttled(cfs_rq))
4819 add_nr_running(rq, task_delta);
4822 * The cfs_rq_throttled() breaks in the above iteration can result in
4823 * incomplete leaf list maintenance, resulting in triggering the
4826 for_each_sched_entity(se) {
4827 cfs_rq = cfs_rq_of(se);
4829 list_add_leaf_cfs_rq(cfs_rq);
4832 assert_list_leaf_cfs_rq(rq);
4834 /* Determine whether we need to wake up potentially idle CPU: */
4835 if (rq->curr == rq->idle && rq->cfs.nr_running)
4839 static u64 distribute_cfs_runtime(struct cfs_bandwidth *cfs_b, u64 remaining)
4841 struct cfs_rq *cfs_rq;
4843 u64 starting_runtime = remaining;
4846 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
4848 struct rq *rq = rq_of(cfs_rq);
4851 rq_lock_irqsave(rq, &rf);
4852 if (!cfs_rq_throttled(cfs_rq))
4855 /* By the above check, this should never be true */
4856 SCHED_WARN_ON(cfs_rq->runtime_remaining > 0);
4858 runtime = -cfs_rq->runtime_remaining + 1;
4859 if (runtime > remaining)
4860 runtime = remaining;
4861 remaining -= runtime;
4863 cfs_rq->runtime_remaining += runtime;
4865 /* we check whether we're throttled above */
4866 if (cfs_rq->runtime_remaining > 0)
4867 unthrottle_cfs_rq(cfs_rq);
4870 rq_unlock_irqrestore(rq, &rf);
4877 return starting_runtime - remaining;
4881 * Responsible for refilling a task_group's bandwidth and unthrottling its
4882 * cfs_rqs as appropriate. If there has been no activity within the last
4883 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
4884 * used to track this state.
4886 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun, unsigned long flags)
4891 /* no need to continue the timer with no bandwidth constraint */
4892 if (cfs_b->quota == RUNTIME_INF)
4893 goto out_deactivate;
4895 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
4896 cfs_b->nr_periods += overrun;
4899 * idle depends on !throttled (for the case of a large deficit), and if
4900 * we're going inactive then everything else can be deferred
4902 if (cfs_b->idle && !throttled)
4903 goto out_deactivate;
4905 __refill_cfs_bandwidth_runtime(cfs_b);
4908 /* mark as potentially idle for the upcoming period */
4913 /* account preceding periods in which throttling occurred */
4914 cfs_b->nr_throttled += overrun;
4917 * This check is repeated as we are holding onto the new bandwidth while
4918 * we unthrottle. This can potentially race with an unthrottled group
4919 * trying to acquire new bandwidth from the global pool. This can result
4920 * in us over-using our runtime if it is all used during this loop, but
4921 * only by limited amounts in that extreme case.
4923 while (throttled && cfs_b->runtime > 0 && !cfs_b->distribute_running) {
4924 runtime = cfs_b->runtime;
4925 cfs_b->distribute_running = 1;
4926 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
4927 /* we can't nest cfs_b->lock while distributing bandwidth */
4928 runtime = distribute_cfs_runtime(cfs_b, runtime);
4929 raw_spin_lock_irqsave(&cfs_b->lock, flags);
4931 cfs_b->distribute_running = 0;
4932 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
4934 lsub_positive(&cfs_b->runtime, runtime);
4938 * While we are ensured activity in the period following an
4939 * unthrottle, this also covers the case in which the new bandwidth is
4940 * insufficient to cover the existing bandwidth deficit. (Forcing the
4941 * timer to remain active while there are any throttled entities.)
4951 /* a cfs_rq won't donate quota below this amount */
4952 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
4953 /* minimum remaining period time to redistribute slack quota */
4954 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
4955 /* how long we wait to gather additional slack before distributing */
4956 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
4959 * Are we near the end of the current quota period?
4961 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
4962 * hrtimer base being cleared by hrtimer_start. In the case of
4963 * migrate_hrtimers, base is never cleared, so we are fine.
4965 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
4967 struct hrtimer *refresh_timer = &cfs_b->period_timer;
4970 /* if the call-back is running a quota refresh is already occurring */
4971 if (hrtimer_callback_running(refresh_timer))
4974 /* is a quota refresh about to occur? */
4975 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
4976 if (remaining < min_expire)
4982 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
4984 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
4986 /* if there's a quota refresh soon don't bother with slack */
4987 if (runtime_refresh_within(cfs_b, min_left))
4990 /* don't push forwards an existing deferred unthrottle */
4991 if (cfs_b->slack_started)
4993 cfs_b->slack_started = true;
4995 hrtimer_start(&cfs_b->slack_timer,
4996 ns_to_ktime(cfs_bandwidth_slack_period),
5000 /* we know any runtime found here is valid as update_curr() precedes return */
5001 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
5003 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
5004 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
5006 if (slack_runtime <= 0)
5009 raw_spin_lock(&cfs_b->lock);
5010 if (cfs_b->quota != RUNTIME_INF) {
5011 cfs_b->runtime += slack_runtime;
5013 /* we are under rq->lock, defer unthrottling using a timer */
5014 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
5015 !list_empty(&cfs_b->throttled_cfs_rq))
5016 start_cfs_slack_bandwidth(cfs_b);
5018 raw_spin_unlock(&cfs_b->lock);
5020 /* even if it's not valid for return we don't want to try again */
5021 cfs_rq->runtime_remaining -= slack_runtime;
5024 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
5026 if (!cfs_bandwidth_used())
5029 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
5032 __return_cfs_rq_runtime(cfs_rq);
5036 * This is done with a timer (instead of inline with bandwidth return) since
5037 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
5039 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
5041 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
5042 unsigned long flags;
5044 /* confirm we're still not at a refresh boundary */
5045 raw_spin_lock_irqsave(&cfs_b->lock, flags);
5046 cfs_b->slack_started = false;
5047 if (cfs_b->distribute_running) {
5048 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
5052 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
5053 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
5057 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice)
5058 runtime = cfs_b->runtime;
5061 cfs_b->distribute_running = 1;
5063 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
5068 runtime = distribute_cfs_runtime(cfs_b, runtime);
5070 raw_spin_lock_irqsave(&cfs_b->lock, flags);
5071 lsub_positive(&cfs_b->runtime, runtime);
5072 cfs_b->distribute_running = 0;
5073 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
5077 * When a group wakes up we want to make sure that its quota is not already
5078 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
5079 * runtime as update_curr() throttling can not not trigger until it's on-rq.
5081 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
5083 if (!cfs_bandwidth_used())
5086 /* an active group must be handled by the update_curr()->put() path */
5087 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
5090 /* ensure the group is not already throttled */
5091 if (cfs_rq_throttled(cfs_rq))
5094 /* update runtime allocation */
5095 account_cfs_rq_runtime(cfs_rq, 0);
5096 if (cfs_rq->runtime_remaining <= 0)
5097 throttle_cfs_rq(cfs_rq);
5100 static void sync_throttle(struct task_group *tg, int cpu)
5102 struct cfs_rq *pcfs_rq, *cfs_rq;
5104 if (!cfs_bandwidth_used())
5110 cfs_rq = tg->cfs_rq[cpu];
5111 pcfs_rq = tg->parent->cfs_rq[cpu];
5113 cfs_rq->throttle_count = pcfs_rq->throttle_count;
5114 cfs_rq->throttled_clock_task = rq_clock_task(cpu_rq(cpu));
5117 /* conditionally throttle active cfs_rq's from put_prev_entity() */
5118 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
5120 if (!cfs_bandwidth_used())
5123 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
5127 * it's possible for a throttled entity to be forced into a running
5128 * state (e.g. set_curr_task), in this case we're finished.
5130 if (cfs_rq_throttled(cfs_rq))
5133 throttle_cfs_rq(cfs_rq);
5137 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
5139 struct cfs_bandwidth *cfs_b =
5140 container_of(timer, struct cfs_bandwidth, slack_timer);
5142 do_sched_cfs_slack_timer(cfs_b);
5144 return HRTIMER_NORESTART;
5147 extern const u64 max_cfs_quota_period;
5149 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
5151 struct cfs_bandwidth *cfs_b =
5152 container_of(timer, struct cfs_bandwidth, period_timer);
5153 unsigned long flags;
5158 raw_spin_lock_irqsave(&cfs_b->lock, flags);
5160 overrun = hrtimer_forward_now(timer, cfs_b->period);
5165 u64 new, old = ktime_to_ns(cfs_b->period);
5168 * Grow period by a factor of 2 to avoid losing precision.
5169 * Precision loss in the quota/period ratio can cause __cfs_schedulable
5173 if (new < max_cfs_quota_period) {
5174 cfs_b->period = ns_to_ktime(new);
5177 pr_warn_ratelimited(
5178 "cfs_period_timer[cpu%d]: period too short, scaling up (new cfs_period_us = %lld, cfs_quota_us = %lld)\n",
5180 div_u64(new, NSEC_PER_USEC),
5181 div_u64(cfs_b->quota, NSEC_PER_USEC));
5183 pr_warn_ratelimited(
5184 "cfs_period_timer[cpu%d]: period too short, but cannot scale up without losing precision (cfs_period_us = %lld, cfs_quota_us = %lld)\n",
5186 div_u64(old, NSEC_PER_USEC),
5187 div_u64(cfs_b->quota, NSEC_PER_USEC));
5190 /* reset count so we don't come right back in here */
5194 idle = do_sched_cfs_period_timer(cfs_b, overrun, flags);
5197 cfs_b->period_active = 0;
5198 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
5200 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
5203 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
5205 raw_spin_lock_init(&cfs_b->lock);
5207 cfs_b->quota = RUNTIME_INF;
5208 cfs_b->period = ns_to_ktime(default_cfs_period());
5210 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
5211 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED);
5212 cfs_b->period_timer.function = sched_cfs_period_timer;
5213 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
5214 cfs_b->slack_timer.function = sched_cfs_slack_timer;
5215 cfs_b->distribute_running = 0;
5216 cfs_b->slack_started = false;
5219 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
5221 cfs_rq->runtime_enabled = 0;
5222 INIT_LIST_HEAD(&cfs_rq->throttled_list);
5225 void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
5227 lockdep_assert_held(&cfs_b->lock);
5229 if (cfs_b->period_active)
5232 cfs_b->period_active = 1;
5233 hrtimer_forward_now(&cfs_b->period_timer, cfs_b->period);
5234 hrtimer_start_expires(&cfs_b->period_timer, HRTIMER_MODE_ABS_PINNED);
5237 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
5239 /* init_cfs_bandwidth() was not called */
5240 if (!cfs_b->throttled_cfs_rq.next)
5243 hrtimer_cancel(&cfs_b->period_timer);
5244 hrtimer_cancel(&cfs_b->slack_timer);
5248 * Both these CPU hotplug callbacks race against unregister_fair_sched_group()
5250 * The race is harmless, since modifying bandwidth settings of unhooked group
5251 * bits doesn't do much.
5254 /* cpu online calback */
5255 static void __maybe_unused update_runtime_enabled(struct rq *rq)
5257 struct task_group *tg;
5259 lockdep_assert_held(&rq->lock);
5262 list_for_each_entry_rcu(tg, &task_groups, list) {
5263 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
5264 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
5266 raw_spin_lock(&cfs_b->lock);
5267 cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF;
5268 raw_spin_unlock(&cfs_b->lock);
5273 /* cpu offline callback */
5274 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
5276 struct task_group *tg;
5278 lockdep_assert_held(&rq->lock);
5281 list_for_each_entry_rcu(tg, &task_groups, list) {
5282 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
5284 if (!cfs_rq->runtime_enabled)
5288 * clock_task is not advancing so we just need to make sure
5289 * there's some valid quota amount
5291 cfs_rq->runtime_remaining = 1;
5293 * Offline rq is schedulable till CPU is completely disabled
5294 * in take_cpu_down(), so we prevent new cfs throttling here.
5296 cfs_rq->runtime_enabled = 0;
5298 if (cfs_rq_throttled(cfs_rq))
5299 unthrottle_cfs_rq(cfs_rq);
5304 #else /* CONFIG_CFS_BANDWIDTH */
5306 static inline bool cfs_bandwidth_used(void)
5311 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {}
5312 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; }
5313 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
5314 static inline void sync_throttle(struct task_group *tg, int cpu) {}
5315 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
5317 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
5322 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
5327 static inline int throttled_lb_pair(struct task_group *tg,
5328 int src_cpu, int dest_cpu)
5333 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
5335 #ifdef CONFIG_FAIR_GROUP_SCHED
5336 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
5339 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
5343 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
5344 static inline void update_runtime_enabled(struct rq *rq) {}
5345 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
5347 #endif /* CONFIG_CFS_BANDWIDTH */
5349 /**************************************************
5350 * CFS operations on tasks:
5353 #ifdef CONFIG_SCHED_HRTICK
5354 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
5356 struct sched_entity *se = &p->se;
5357 struct cfs_rq *cfs_rq = cfs_rq_of(se);
5359 SCHED_WARN_ON(task_rq(p) != rq);
5361 if (rq->cfs.h_nr_running > 1) {
5362 u64 slice = sched_slice(cfs_rq, se);
5363 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
5364 s64 delta = slice - ran;
5371 hrtick_start(rq, delta);
5376 * called from enqueue/dequeue and updates the hrtick when the
5377 * current task is from our class and nr_running is low enough
5380 static void hrtick_update(struct rq *rq)
5382 struct task_struct *curr = rq->curr;
5384 if (!hrtick_enabled(rq) || curr->sched_class != &fair_sched_class)
5387 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
5388 hrtick_start_fair(rq, curr);
5390 #else /* !CONFIG_SCHED_HRTICK */
5392 hrtick_start_fair(struct rq *rq, struct task_struct *p)
5396 static inline void hrtick_update(struct rq *rq)
5402 static inline unsigned long cpu_util(int cpu);
5404 static inline bool cpu_overutilized(int cpu)
5406 return !fits_capacity(cpu_util(cpu), capacity_of(cpu));
5409 static inline void update_overutilized_status(struct rq *rq)
5411 if (!READ_ONCE(rq->rd->overutilized) && cpu_overutilized(rq->cpu)) {
5412 WRITE_ONCE(rq->rd->overutilized, SG_OVERUTILIZED);
5413 trace_sched_overutilized_tp(rq->rd, SG_OVERUTILIZED);
5417 static inline void update_overutilized_status(struct rq *rq) { }
5420 /* Runqueue only has SCHED_IDLE tasks enqueued */
5421 static int sched_idle_rq(struct rq *rq)
5423 return unlikely(rq->nr_running == rq->cfs.idle_h_nr_running &&
5428 static int sched_idle_cpu(int cpu)
5430 return sched_idle_rq(cpu_rq(cpu));
5435 * The enqueue_task method is called before nr_running is
5436 * increased. Here we update the fair scheduling stats and
5437 * then put the task into the rbtree:
5440 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
5442 struct cfs_rq *cfs_rq;
5443 struct sched_entity *se = &p->se;
5444 int idle_h_nr_running = task_has_idle_policy(p);
5447 * The code below (indirectly) updates schedutil which looks at
5448 * the cfs_rq utilization to select a frequency.
5449 * Let's add the task's estimated utilization to the cfs_rq's
5450 * estimated utilization, before we update schedutil.
5452 util_est_enqueue(&rq->cfs, p);
5455 * If in_iowait is set, the code below may not trigger any cpufreq
5456 * utilization updates, so do it here explicitly with the IOWAIT flag
5460 cpufreq_update_util(rq, SCHED_CPUFREQ_IOWAIT);
5462 for_each_sched_entity(se) {
5465 cfs_rq = cfs_rq_of(se);
5466 enqueue_entity(cfs_rq, se, flags);
5468 cfs_rq->h_nr_running++;
5469 cfs_rq->idle_h_nr_running += idle_h_nr_running;
5471 /* end evaluation on encountering a throttled cfs_rq */
5472 if (cfs_rq_throttled(cfs_rq))
5473 goto enqueue_throttle;
5475 flags = ENQUEUE_WAKEUP;
5478 for_each_sched_entity(se) {
5479 cfs_rq = cfs_rq_of(se);
5481 update_load_avg(cfs_rq, se, UPDATE_TG);
5482 se_update_runnable(se);
5483 update_cfs_group(se);
5485 cfs_rq->h_nr_running++;
5486 cfs_rq->idle_h_nr_running += idle_h_nr_running;
5488 /* end evaluation on encountering a throttled cfs_rq */
5489 if (cfs_rq_throttled(cfs_rq))
5490 goto enqueue_throttle;
5495 add_nr_running(rq, 1);
5497 * Since new tasks are assigned an initial util_avg equal to
5498 * half of the spare capacity of their CPU, tiny tasks have the
5499 * ability to cross the overutilized threshold, which will
5500 * result in the load balancer ruining all the task placement
5501 * done by EAS. As a way to mitigate that effect, do not account
5502 * for the first enqueue operation of new tasks during the
5503 * overutilized flag detection.
5505 * A better way of solving this problem would be to wait for
5506 * the PELT signals of tasks to converge before taking them
5507 * into account, but that is not straightforward to implement,
5508 * and the following generally works well enough in practice.
5510 if (flags & ENQUEUE_WAKEUP)
5511 update_overutilized_status(rq);
5515 if (cfs_bandwidth_used()) {
5517 * When bandwidth control is enabled; the cfs_rq_throttled()
5518 * breaks in the above iteration can result in incomplete
5519 * leaf list maintenance, resulting in triggering the assertion
5522 for_each_sched_entity(se) {
5523 cfs_rq = cfs_rq_of(se);
5525 if (list_add_leaf_cfs_rq(cfs_rq))
5530 assert_list_leaf_cfs_rq(rq);
5535 static void set_next_buddy(struct sched_entity *se);
5538 * The dequeue_task method is called before nr_running is
5539 * decreased. We remove the task from the rbtree and
5540 * update the fair scheduling stats:
5542 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
5544 struct cfs_rq *cfs_rq;
5545 struct sched_entity *se = &p->se;
5546 int task_sleep = flags & DEQUEUE_SLEEP;
5547 int idle_h_nr_running = task_has_idle_policy(p);
5548 bool was_sched_idle = sched_idle_rq(rq);
5550 for_each_sched_entity(se) {
5551 cfs_rq = cfs_rq_of(se);
5552 dequeue_entity(cfs_rq, se, flags);
5554 cfs_rq->h_nr_running--;
5555 cfs_rq->idle_h_nr_running -= idle_h_nr_running;
5557 /* end evaluation on encountering a throttled cfs_rq */
5558 if (cfs_rq_throttled(cfs_rq))
5559 goto dequeue_throttle;
5561 /* Don't dequeue parent if it has other entities besides us */
5562 if (cfs_rq->load.weight) {
5563 /* Avoid re-evaluating load for this entity: */
5564 se = parent_entity(se);
5566 * Bias pick_next to pick a task from this cfs_rq, as
5567 * p is sleeping when it is within its sched_slice.
5569 if (task_sleep && se && !throttled_hierarchy(cfs_rq))
5573 flags |= DEQUEUE_SLEEP;
5576 for_each_sched_entity(se) {
5577 cfs_rq = cfs_rq_of(se);
5579 update_load_avg(cfs_rq, se, UPDATE_TG);
5580 se_update_runnable(se);
5581 update_cfs_group(se);
5583 cfs_rq->h_nr_running--;
5584 cfs_rq->idle_h_nr_running -= idle_h_nr_running;
5586 /* end evaluation on encountering a throttled cfs_rq */
5587 if (cfs_rq_throttled(cfs_rq))
5588 goto dequeue_throttle;
5594 sub_nr_running(rq, 1);
5596 /* balance early to pull high priority tasks */
5597 if (unlikely(!was_sched_idle && sched_idle_rq(rq)))
5598 rq->next_balance = jiffies;
5600 util_est_dequeue(&rq->cfs, p, task_sleep);
5606 /* Working cpumask for: load_balance, load_balance_newidle. */
5607 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
5608 DEFINE_PER_CPU(cpumask_var_t, select_idle_mask);
5610 #ifdef CONFIG_NO_HZ_COMMON
5613 cpumask_var_t idle_cpus_mask;
5615 int has_blocked; /* Idle CPUS has blocked load */
5616 unsigned long next_balance; /* in jiffy units */
5617 unsigned long next_blocked; /* Next update of blocked load in jiffies */
5618 } nohz ____cacheline_aligned;
5620 #endif /* CONFIG_NO_HZ_COMMON */
5622 static unsigned long cpu_load(struct rq *rq)
5624 return cfs_rq_load_avg(&rq->cfs);
5628 * cpu_load_without - compute CPU load without any contributions from *p
5629 * @cpu: the CPU which load is requested
5630 * @p: the task which load should be discounted
5632 * The load of a CPU is defined by the load of tasks currently enqueued on that
5633 * CPU as well as tasks which are currently sleeping after an execution on that
5636 * This method returns the load of the specified CPU by discounting the load of
5637 * the specified task, whenever the task is currently contributing to the CPU
5640 static unsigned long cpu_load_without(struct rq *rq, struct task_struct *p)
5642 struct cfs_rq *cfs_rq;
5645 /* Task has no contribution or is new */
5646 if (cpu_of(rq) != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time))
5647 return cpu_load(rq);
5650 load = READ_ONCE(cfs_rq->avg.load_avg);
5652 /* Discount task's util from CPU's util */
5653 lsub_positive(&load, task_h_load(p));
5658 static unsigned long cpu_runnable(struct rq *rq)
5660 return cfs_rq_runnable_avg(&rq->cfs);
5663 static unsigned long cpu_runnable_without(struct rq *rq, struct task_struct *p)
5665 struct cfs_rq *cfs_rq;
5666 unsigned int runnable;
5668 /* Task has no contribution or is new */
5669 if (cpu_of(rq) != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time))
5670 return cpu_runnable(rq);
5673 runnable = READ_ONCE(cfs_rq->avg.runnable_avg);
5675 /* Discount task's runnable from CPU's runnable */
5676 lsub_positive(&runnable, p->se.avg.runnable_avg);
5681 static unsigned long capacity_of(int cpu)
5683 return cpu_rq(cpu)->cpu_capacity;
5686 static void record_wakee(struct task_struct *p)
5689 * Only decay a single time; tasks that have less then 1 wakeup per
5690 * jiffy will not have built up many flips.
5692 if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) {
5693 current->wakee_flips >>= 1;
5694 current->wakee_flip_decay_ts = jiffies;
5697 if (current->last_wakee != p) {
5698 current->last_wakee = p;
5699 current->wakee_flips++;
5704 * Detect M:N waker/wakee relationships via a switching-frequency heuristic.
5706 * A waker of many should wake a different task than the one last awakened
5707 * at a frequency roughly N times higher than one of its wakees.
5709 * In order to determine whether we should let the load spread vs consolidating
5710 * to shared cache, we look for a minimum 'flip' frequency of llc_size in one
5711 * partner, and a factor of lls_size higher frequency in the other.
5713 * With both conditions met, we can be relatively sure that the relationship is
5714 * non-monogamous, with partner count exceeding socket size.
5716 * Waker/wakee being client/server, worker/dispatcher, interrupt source or
5717 * whatever is irrelevant, spread criteria is apparent partner count exceeds
5720 static int wake_wide(struct task_struct *p)
5722 unsigned int master = current->wakee_flips;
5723 unsigned int slave = p->wakee_flips;
5724 int factor = this_cpu_read(sd_llc_size);
5727 swap(master, slave);
5728 if (slave < factor || master < slave * factor)
5734 * The purpose of wake_affine() is to quickly determine on which CPU we can run
5735 * soonest. For the purpose of speed we only consider the waking and previous
5738 * wake_affine_idle() - only considers 'now', it check if the waking CPU is
5739 * cache-affine and is (or will be) idle.
5741 * wake_affine_weight() - considers the weight to reflect the average
5742 * scheduling latency of the CPUs. This seems to work
5743 * for the overloaded case.
5746 wake_affine_idle(int this_cpu, int prev_cpu, int sync)
5749 * If this_cpu is idle, it implies the wakeup is from interrupt
5750 * context. Only allow the move if cache is shared. Otherwise an
5751 * interrupt intensive workload could force all tasks onto one
5752 * node depending on the IO topology or IRQ affinity settings.
5754 * If the prev_cpu is idle and cache affine then avoid a migration.
5755 * There is no guarantee that the cache hot data from an interrupt
5756 * is more important than cache hot data on the prev_cpu and from
5757 * a cpufreq perspective, it's better to have higher utilisation
5760 if (available_idle_cpu(this_cpu) && cpus_share_cache(this_cpu, prev_cpu))
5761 return available_idle_cpu(prev_cpu) ? prev_cpu : this_cpu;
5763 if (sync && cpu_rq(this_cpu)->nr_running == 1)
5766 return nr_cpumask_bits;
5770 wake_affine_weight(struct sched_domain *sd, struct task_struct *p,
5771 int this_cpu, int prev_cpu, int sync)
5773 s64 this_eff_load, prev_eff_load;
5774 unsigned long task_load;
5776 this_eff_load = cpu_load(cpu_rq(this_cpu));
5779 unsigned long current_load = task_h_load(current);
5781 if (current_load > this_eff_load)
5784 this_eff_load -= current_load;
5787 task_load = task_h_load(p);
5789 this_eff_load += task_load;
5790 if (sched_feat(WA_BIAS))
5791 this_eff_load *= 100;
5792 this_eff_load *= capacity_of(prev_cpu);
5794 prev_eff_load = cpu_load(cpu_rq(prev_cpu));
5795 prev_eff_load -= task_load;
5796 if (sched_feat(WA_BIAS))
5797 prev_eff_load *= 100 + (sd->imbalance_pct - 100) / 2;
5798 prev_eff_load *= capacity_of(this_cpu);
5801 * If sync, adjust the weight of prev_eff_load such that if
5802 * prev_eff == this_eff that select_idle_sibling() will consider
5803 * stacking the wakee on top of the waker if no other CPU is
5809 return this_eff_load < prev_eff_load ? this_cpu : nr_cpumask_bits;
5812 static int wake_affine(struct sched_domain *sd, struct task_struct *p,
5813 int this_cpu, int prev_cpu, int sync)
5815 int target = nr_cpumask_bits;
5817 if (sched_feat(WA_IDLE))
5818 target = wake_affine_idle(this_cpu, prev_cpu, sync);
5820 if (sched_feat(WA_WEIGHT) && target == nr_cpumask_bits)
5821 target = wake_affine_weight(sd, p, this_cpu, prev_cpu, sync);
5823 schedstat_inc(p->se.statistics.nr_wakeups_affine_attempts);
5824 if (target == nr_cpumask_bits)
5827 schedstat_inc(sd->ttwu_move_affine);
5828 schedstat_inc(p->se.statistics.nr_wakeups_affine);
5832 static struct sched_group *
5833 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
5834 int this_cpu, int sd_flag);
5837 * find_idlest_group_cpu - find the idlest CPU among the CPUs in the group.
5840 find_idlest_group_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
5842 unsigned long load, min_load = ULONG_MAX;
5843 unsigned int min_exit_latency = UINT_MAX;
5844 u64 latest_idle_timestamp = 0;
5845 int least_loaded_cpu = this_cpu;
5846 int shallowest_idle_cpu = -1;
5849 /* Check if we have any choice: */
5850 if (group->group_weight == 1)
5851 return cpumask_first(sched_group_span(group));
5853 /* Traverse only the allowed CPUs */
5854 for_each_cpu_and(i, sched_group_span(group), p->cpus_ptr) {
5855 if (sched_idle_cpu(i))
5858 if (available_idle_cpu(i)) {
5859 struct rq *rq = cpu_rq(i);
5860 struct cpuidle_state *idle = idle_get_state(rq);
5861 if (idle && idle->exit_latency < min_exit_latency) {
5863 * We give priority to a CPU whose idle state
5864 * has the smallest exit latency irrespective
5865 * of any idle timestamp.
5867 min_exit_latency = idle->exit_latency;
5868 latest_idle_timestamp = rq->idle_stamp;
5869 shallowest_idle_cpu = i;
5870 } else if ((!idle || idle->exit_latency == min_exit_latency) &&
5871 rq->idle_stamp > latest_idle_timestamp) {
5873 * If equal or no active idle state, then
5874 * the most recently idled CPU might have
5877 latest_idle_timestamp = rq->idle_stamp;
5878 shallowest_idle_cpu = i;
5880 } else if (shallowest_idle_cpu == -1) {
5881 load = cpu_load(cpu_rq(i));
5882 if (load < min_load) {
5884 least_loaded_cpu = i;
5889 return shallowest_idle_cpu != -1 ? shallowest_idle_cpu : least_loaded_cpu;
5892 static inline int find_idlest_cpu(struct sched_domain *sd, struct task_struct *p,
5893 int cpu, int prev_cpu, int sd_flag)
5897 if (!cpumask_intersects(sched_domain_span(sd), p->cpus_ptr))
5901 * We need task's util for cpu_util_without, sync it up to
5902 * prev_cpu's last_update_time.
5904 if (!(sd_flag & SD_BALANCE_FORK))
5905 sync_entity_load_avg(&p->se);
5908 struct sched_group *group;
5909 struct sched_domain *tmp;
5912 if (!(sd->flags & sd_flag)) {
5917 group = find_idlest_group(sd, p, cpu, sd_flag);
5923 new_cpu = find_idlest_group_cpu(group, p, cpu);
5924 if (new_cpu == cpu) {
5925 /* Now try balancing at a lower domain level of 'cpu': */
5930 /* Now try balancing at a lower domain level of 'new_cpu': */
5932 weight = sd->span_weight;
5934 for_each_domain(cpu, tmp) {
5935 if (weight <= tmp->span_weight)
5937 if (tmp->flags & sd_flag)
5945 #ifdef CONFIG_SCHED_SMT
5946 DEFINE_STATIC_KEY_FALSE(sched_smt_present);
5947 EXPORT_SYMBOL_GPL(sched_smt_present);
5949 static inline void set_idle_cores(int cpu, int val)
5951 struct sched_domain_shared *sds;
5953 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
5955 WRITE_ONCE(sds->has_idle_cores, val);
5958 static inline bool test_idle_cores(int cpu, bool def)
5960 struct sched_domain_shared *sds;
5962 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
5964 return READ_ONCE(sds->has_idle_cores);
5970 * Scans the local SMT mask to see if the entire core is idle, and records this
5971 * information in sd_llc_shared->has_idle_cores.
5973 * Since SMT siblings share all cache levels, inspecting this limited remote
5974 * state should be fairly cheap.
5976 void __update_idle_core(struct rq *rq)
5978 int core = cpu_of(rq);
5982 if (test_idle_cores(core, true))
5985 for_each_cpu(cpu, cpu_smt_mask(core)) {
5989 if (!available_idle_cpu(cpu))
5993 set_idle_cores(core, 1);
5999 * Scan the entire LLC domain for idle cores; this dynamically switches off if
6000 * there are no idle cores left in the system; tracked through
6001 * sd_llc->shared->has_idle_cores and enabled through update_idle_core() above.
6003 static int select_idle_core(struct task_struct *p, struct sched_domain *sd, int target)
6005 struct cpumask *cpus = this_cpu_cpumask_var_ptr(select_idle_mask);
6008 if (!static_branch_likely(&sched_smt_present))
6011 if (!test_idle_cores(target, false))
6014 cpumask_and(cpus, sched_domain_span(sd), p->cpus_ptr);
6016 for_each_cpu_wrap(core, cpus, target) {
6019 for_each_cpu(cpu, cpu_smt_mask(core)) {
6020 if (!available_idle_cpu(cpu)) {
6025 cpumask_andnot(cpus, cpus, cpu_smt_mask(core));
6032 * Failed to find an idle core; stop looking for one.
6034 set_idle_cores(target, 0);
6040 * Scan the local SMT mask for idle CPUs.
6042 static int select_idle_smt(struct task_struct *p, int target)
6046 if (!static_branch_likely(&sched_smt_present))
6049 for_each_cpu(cpu, cpu_smt_mask(target)) {
6050 if (!cpumask_test_cpu(cpu, p->cpus_ptr))
6052 if (available_idle_cpu(cpu) || sched_idle_cpu(cpu))
6059 #else /* CONFIG_SCHED_SMT */
6061 static inline int select_idle_core(struct task_struct *p, struct sched_domain *sd, int target)
6066 static inline int select_idle_smt(struct task_struct *p, int target)
6071 #endif /* CONFIG_SCHED_SMT */
6074 * Scan the LLC domain for idle CPUs; this is dynamically regulated by
6075 * comparing the average scan cost (tracked in sd->avg_scan_cost) against the
6076 * average idle time for this rq (as found in rq->avg_idle).
6078 static int select_idle_cpu(struct task_struct *p, struct sched_domain *sd, int target)
6080 struct cpumask *cpus = this_cpu_cpumask_var_ptr(select_idle_mask);
6081 struct sched_domain *this_sd;
6082 u64 avg_cost, avg_idle;
6085 int this = smp_processor_id();
6086 int cpu, nr = INT_MAX;
6088 this_sd = rcu_dereference(*this_cpu_ptr(&sd_llc));
6093 * Due to large variance we need a large fuzz factor; hackbench in
6094 * particularly is sensitive here.
6096 avg_idle = this_rq()->avg_idle / 512;
6097 avg_cost = this_sd->avg_scan_cost + 1;
6099 if (sched_feat(SIS_AVG_CPU) && avg_idle < avg_cost)
6102 if (sched_feat(SIS_PROP)) {
6103 u64 span_avg = sd->span_weight * avg_idle;
6104 if (span_avg > 4*avg_cost)
6105 nr = div_u64(span_avg, avg_cost);
6110 time = cpu_clock(this);
6112 cpumask_and(cpus, sched_domain_span(sd), p->cpus_ptr);
6114 for_each_cpu_wrap(cpu, cpus, target) {
6117 if (available_idle_cpu(cpu) || sched_idle_cpu(cpu))
6121 time = cpu_clock(this) - time;
6122 cost = this_sd->avg_scan_cost;
6123 delta = (s64)(time - cost) / 8;
6124 this_sd->avg_scan_cost += delta;
6130 * Scan the asym_capacity domain for idle CPUs; pick the first idle one on which
6131 * the task fits. If no CPU is big enough, but there are idle ones, try to
6132 * maximize capacity.
6135 select_idle_capacity(struct task_struct *p, struct sched_domain *sd, int target)
6137 unsigned long best_cap = 0;
6138 int cpu, best_cpu = -1;
6139 struct cpumask *cpus;
6141 sync_entity_load_avg(&p->se);
6143 cpus = this_cpu_cpumask_var_ptr(select_idle_mask);
6144 cpumask_and(cpus, sched_domain_span(sd), p->cpus_ptr);
6146 for_each_cpu_wrap(cpu, cpus, target) {
6147 unsigned long cpu_cap = capacity_of(cpu);
6149 if (!available_idle_cpu(cpu) && !sched_idle_cpu(cpu))
6151 if (task_fits_capacity(p, cpu_cap))
6154 if (cpu_cap > best_cap) {
6164 * Try and locate an idle core/thread in the LLC cache domain.
6166 static int select_idle_sibling(struct task_struct *p, int prev, int target)
6168 struct sched_domain *sd;
6169 int i, recent_used_cpu;
6172 * For asymmetric CPU capacity systems, our domain of interest is
6173 * sd_asym_cpucapacity rather than sd_llc.
6175 if (static_branch_unlikely(&sched_asym_cpucapacity)) {
6176 sd = rcu_dereference(per_cpu(sd_asym_cpucapacity, target));
6178 * On an asymmetric CPU capacity system where an exclusive
6179 * cpuset defines a symmetric island (i.e. one unique
6180 * capacity_orig value through the cpuset), the key will be set
6181 * but the CPUs within that cpuset will not have a domain with
6182 * SD_ASYM_CPUCAPACITY. These should follow the usual symmetric
6188 i = select_idle_capacity(p, sd, target);
6189 return ((unsigned)i < nr_cpumask_bits) ? i : target;
6193 if (available_idle_cpu(target) || sched_idle_cpu(target))
6197 * If the previous CPU is cache affine and idle, don't be stupid:
6199 if (prev != target && cpus_share_cache(prev, target) &&
6200 (available_idle_cpu(prev) || sched_idle_cpu(prev)))
6204 * Allow a per-cpu kthread to stack with the wakee if the
6205 * kworker thread and the tasks previous CPUs are the same.
6206 * The assumption is that the wakee queued work for the
6207 * per-cpu kthread that is now complete and the wakeup is
6208 * essentially a sync wakeup. An obvious example of this
6209 * pattern is IO completions.
6211 if (is_per_cpu_kthread(current) &&
6212 prev == smp_processor_id() &&
6213 this_rq()->nr_running <= 1) {
6217 /* Check a recently used CPU as a potential idle candidate: */
6218 recent_used_cpu = p->recent_used_cpu;
6219 if (recent_used_cpu != prev &&
6220 recent_used_cpu != target &&
6221 cpus_share_cache(recent_used_cpu, target) &&
6222 (available_idle_cpu(recent_used_cpu) || sched_idle_cpu(recent_used_cpu)) &&
6223 cpumask_test_cpu(p->recent_used_cpu, p->cpus_ptr)) {
6225 * Replace recent_used_cpu with prev as it is a potential
6226 * candidate for the next wake:
6228 p->recent_used_cpu = prev;
6229 return recent_used_cpu;
6232 sd = rcu_dereference(per_cpu(sd_llc, target));
6236 i = select_idle_core(p, sd, target);
6237 if ((unsigned)i < nr_cpumask_bits)
6240 i = select_idle_cpu(p, sd, target);
6241 if ((unsigned)i < nr_cpumask_bits)
6244 i = select_idle_smt(p, target);
6245 if ((unsigned)i < nr_cpumask_bits)
6252 * Amount of capacity of a CPU that is (estimated to be) used by CFS tasks
6253 * @cpu: the CPU to get the utilization of
6255 * The unit of the return value must be the one of capacity so we can compare
6256 * the utilization with the capacity of the CPU that is available for CFS task
6257 * (ie cpu_capacity).
6259 * cfs_rq.avg.util_avg is the sum of running time of runnable tasks plus the
6260 * recent utilization of currently non-runnable tasks on a CPU. It represents
6261 * the amount of utilization of a CPU in the range [0..capacity_orig] where
6262 * capacity_orig is the cpu_capacity available at the highest frequency
6263 * (arch_scale_freq_capacity()).
6264 * The utilization of a CPU converges towards a sum equal to or less than the
6265 * current capacity (capacity_curr <= capacity_orig) of the CPU because it is
6266 * the running time on this CPU scaled by capacity_curr.
6268 * The estimated utilization of a CPU is defined to be the maximum between its
6269 * cfs_rq.avg.util_avg and the sum of the estimated utilization of the tasks
6270 * currently RUNNABLE on that CPU.
6271 * This allows to properly represent the expected utilization of a CPU which
6272 * has just got a big task running since a long sleep period. At the same time
6273 * however it preserves the benefits of the "blocked utilization" in
6274 * describing the potential for other tasks waking up on the same CPU.
6276 * Nevertheless, cfs_rq.avg.util_avg can be higher than capacity_curr or even
6277 * higher than capacity_orig because of unfortunate rounding in
6278 * cfs.avg.util_avg or just after migrating tasks and new task wakeups until
6279 * the average stabilizes with the new running time. We need to check that the
6280 * utilization stays within the range of [0..capacity_orig] and cap it if
6281 * necessary. Without utilization capping, a group could be seen as overloaded
6282 * (CPU0 utilization at 121% + CPU1 utilization at 80%) whereas CPU1 has 20% of
6283 * available capacity. We allow utilization to overshoot capacity_curr (but not
6284 * capacity_orig) as it useful for predicting the capacity required after task
6285 * migrations (scheduler-driven DVFS).
6287 * Return: the (estimated) utilization for the specified CPU
6289 static inline unsigned long cpu_util(int cpu)
6291 struct cfs_rq *cfs_rq;
6294 cfs_rq = &cpu_rq(cpu)->cfs;
6295 util = READ_ONCE(cfs_rq->avg.util_avg);
6297 if (sched_feat(UTIL_EST))
6298 util = max(util, READ_ONCE(cfs_rq->avg.util_est.enqueued));
6300 return min_t(unsigned long, util, capacity_orig_of(cpu));
6304 * cpu_util_without: compute cpu utilization without any contributions from *p
6305 * @cpu: the CPU which utilization is requested
6306 * @p: the task which utilization should be discounted
6308 * The utilization of a CPU is defined by the utilization of tasks currently
6309 * enqueued on that CPU as well as tasks which are currently sleeping after an
6310 * execution on that CPU.
6312 * This method returns the utilization of the specified CPU by discounting the
6313 * utilization of the specified task, whenever the task is currently
6314 * contributing to the CPU utilization.
6316 static unsigned long cpu_util_without(int cpu, struct task_struct *p)
6318 struct cfs_rq *cfs_rq;
6321 /* Task has no contribution or is new */
6322 if (cpu != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time))
6323 return cpu_util(cpu);
6325 cfs_rq = &cpu_rq(cpu)->cfs;
6326 util = READ_ONCE(cfs_rq->avg.util_avg);
6328 /* Discount task's util from CPU's util */
6329 lsub_positive(&util, task_util(p));
6334 * a) if *p is the only task sleeping on this CPU, then:
6335 * cpu_util (== task_util) > util_est (== 0)
6336 * and thus we return:
6337 * cpu_util_without = (cpu_util - task_util) = 0
6339 * b) if other tasks are SLEEPING on this CPU, which is now exiting
6341 * cpu_util >= task_util
6342 * cpu_util > util_est (== 0)
6343 * and thus we discount *p's blocked utilization to return:
6344 * cpu_util_without = (cpu_util - task_util) >= 0
6346 * c) if other tasks are RUNNABLE on that CPU and
6347 * util_est > cpu_util
6348 * then we use util_est since it returns a more restrictive
6349 * estimation of the spare capacity on that CPU, by just
6350 * considering the expected utilization of tasks already
6351 * runnable on that CPU.
6353 * Cases a) and b) are covered by the above code, while case c) is
6354 * covered by the following code when estimated utilization is
6357 if (sched_feat(UTIL_EST)) {
6358 unsigned int estimated =
6359 READ_ONCE(cfs_rq->avg.util_est.enqueued);
6362 * Despite the following checks we still have a small window
6363 * for a possible race, when an execl's select_task_rq_fair()
6364 * races with LB's detach_task():
6367 * p->on_rq = TASK_ON_RQ_MIGRATING;
6368 * ---------------------------------- A
6369 * deactivate_task() \
6370 * dequeue_task() + RaceTime
6371 * util_est_dequeue() /
6372 * ---------------------------------- B
6374 * The additional check on "current == p" it's required to
6375 * properly fix the execl regression and it helps in further
6376 * reducing the chances for the above race.
6378 if (unlikely(task_on_rq_queued(p) || current == p))
6379 lsub_positive(&estimated, _task_util_est(p));
6381 util = max(util, estimated);
6385 * Utilization (estimated) can exceed the CPU capacity, thus let's
6386 * clamp to the maximum CPU capacity to ensure consistency with
6387 * the cpu_util call.
6389 return min_t(unsigned long, util, capacity_orig_of(cpu));
6393 * Predicts what cpu_util(@cpu) would return if @p was migrated (and enqueued)
6396 static unsigned long cpu_util_next(int cpu, struct task_struct *p, int dst_cpu)
6398 struct cfs_rq *cfs_rq = &cpu_rq(cpu)->cfs;
6399 unsigned long util_est, util = READ_ONCE(cfs_rq->avg.util_avg);
6402 * If @p migrates from @cpu to another, remove its contribution. Or,
6403 * if @p migrates from another CPU to @cpu, add its contribution. In
6404 * the other cases, @cpu is not impacted by the migration, so the
6405 * util_avg should already be correct.
6407 if (task_cpu(p) == cpu && dst_cpu != cpu)
6408 sub_positive(&util, task_util(p));
6409 else if (task_cpu(p) != cpu && dst_cpu == cpu)
6410 util += task_util(p);
6412 if (sched_feat(UTIL_EST)) {
6413 util_est = READ_ONCE(cfs_rq->avg.util_est.enqueued);
6416 * During wake-up, the task isn't enqueued yet and doesn't
6417 * appear in the cfs_rq->avg.util_est.enqueued of any rq,
6418 * so just add it (if needed) to "simulate" what will be
6419 * cpu_util() after the task has been enqueued.
6422 util_est += _task_util_est(p);
6424 util = max(util, util_est);
6427 return min(util, capacity_orig_of(cpu));
6431 * compute_energy(): Estimates the energy that @pd would consume if @p was
6432 * migrated to @dst_cpu. compute_energy() predicts what will be the utilization
6433 * landscape of @pd's CPUs after the task migration, and uses the Energy Model
6434 * to compute what would be the energy if we decided to actually migrate that
6438 compute_energy(struct task_struct *p, int dst_cpu, struct perf_domain *pd)
6440 struct cpumask *pd_mask = perf_domain_span(pd);
6441 unsigned long cpu_cap = arch_scale_cpu_capacity(cpumask_first(pd_mask));
6442 unsigned long max_util = 0, sum_util = 0;
6446 * The capacity state of CPUs of the current rd can be driven by CPUs
6447 * of another rd if they belong to the same pd. So, account for the
6448 * utilization of these CPUs too by masking pd with cpu_online_mask
6449 * instead of the rd span.
6451 * If an entire pd is outside of the current rd, it will not appear in
6452 * its pd list and will not be accounted by compute_energy().
6454 for_each_cpu_and(cpu, pd_mask, cpu_online_mask) {
6455 unsigned long cpu_util, util_cfs = cpu_util_next(cpu, p, dst_cpu);
6456 struct task_struct *tsk = cpu == dst_cpu ? p : NULL;
6459 * Busy time computation: utilization clamping is not
6460 * required since the ratio (sum_util / cpu_capacity)
6461 * is already enough to scale the EM reported power
6462 * consumption at the (eventually clamped) cpu_capacity.
6464 sum_util += schedutil_cpu_util(cpu, util_cfs, cpu_cap,
6468 * Performance domain frequency: utilization clamping
6469 * must be considered since it affects the selection
6470 * of the performance domain frequency.
6471 * NOTE: in case RT tasks are running, by default the
6472 * FREQUENCY_UTIL's utilization can be max OPP.
6474 cpu_util = schedutil_cpu_util(cpu, util_cfs, cpu_cap,
6475 FREQUENCY_UTIL, tsk);
6476 max_util = max(max_util, cpu_util);
6479 return em_pd_energy(pd->em_pd, max_util, sum_util);
6483 * find_energy_efficient_cpu(): Find most energy-efficient target CPU for the
6484 * waking task. find_energy_efficient_cpu() looks for the CPU with maximum
6485 * spare capacity in each performance domain and uses it as a potential
6486 * candidate to execute the task. Then, it uses the Energy Model to figure
6487 * out which of the CPU candidates is the most energy-efficient.
6489 * The rationale for this heuristic is as follows. In a performance domain,
6490 * all the most energy efficient CPU candidates (according to the Energy
6491 * Model) are those for which we'll request a low frequency. When there are
6492 * several CPUs for which the frequency request will be the same, we don't
6493 * have enough data to break the tie between them, because the Energy Model
6494 * only includes active power costs. With this model, if we assume that
6495 * frequency requests follow utilization (e.g. using schedutil), the CPU with
6496 * the maximum spare capacity in a performance domain is guaranteed to be among
6497 * the best candidates of the performance domain.
6499 * In practice, it could be preferable from an energy standpoint to pack
6500 * small tasks on a CPU in order to let other CPUs go in deeper idle states,
6501 * but that could also hurt our chances to go cluster idle, and we have no
6502 * ways to tell with the current Energy Model if this is actually a good
6503 * idea or not. So, find_energy_efficient_cpu() basically favors
6504 * cluster-packing, and spreading inside a cluster. That should at least be
6505 * a good thing for latency, and this is consistent with the idea that most
6506 * of the energy savings of EAS come from the asymmetry of the system, and
6507 * not so much from breaking the tie between identical CPUs. That's also the
6508 * reason why EAS is enabled in the topology code only for systems where
6509 * SD_ASYM_CPUCAPACITY is set.
6511 * NOTE: Forkees are not accepted in the energy-aware wake-up path because
6512 * they don't have any useful utilization data yet and it's not possible to
6513 * forecast their impact on energy consumption. Consequently, they will be
6514 * placed by find_idlest_cpu() on the least loaded CPU, which might turn out
6515 * to be energy-inefficient in some use-cases. The alternative would be to
6516 * bias new tasks towards specific types of CPUs first, or to try to infer
6517 * their util_avg from the parent task, but those heuristics could hurt
6518 * other use-cases too. So, until someone finds a better way to solve this,
6519 * let's keep things simple by re-using the existing slow path.
6521 static int find_energy_efficient_cpu(struct task_struct *p, int prev_cpu)
6523 unsigned long prev_delta = ULONG_MAX, best_delta = ULONG_MAX;
6524 struct root_domain *rd = cpu_rq(smp_processor_id())->rd;
6525 unsigned long cpu_cap, util, base_energy = 0;
6526 int cpu, best_energy_cpu = prev_cpu;
6527 struct sched_domain *sd;
6528 struct perf_domain *pd;
6531 pd = rcu_dereference(rd->pd);
6532 if (!pd || READ_ONCE(rd->overutilized))
6536 * Energy-aware wake-up happens on the lowest sched_domain starting
6537 * from sd_asym_cpucapacity spanning over this_cpu and prev_cpu.
6539 sd = rcu_dereference(*this_cpu_ptr(&sd_asym_cpucapacity));
6540 while (sd && !cpumask_test_cpu(prev_cpu, sched_domain_span(sd)))
6545 sync_entity_load_avg(&p->se);
6546 if (!task_util_est(p))
6549 for (; pd; pd = pd->next) {
6550 unsigned long cur_delta, spare_cap, max_spare_cap = 0;
6551 unsigned long base_energy_pd;
6552 int max_spare_cap_cpu = -1;
6554 /* Compute the 'base' energy of the pd, without @p */
6555 base_energy_pd = compute_energy(p, -1, pd);
6556 base_energy += base_energy_pd;
6558 for_each_cpu_and(cpu, perf_domain_span(pd), sched_domain_span(sd)) {
6559 if (!cpumask_test_cpu(cpu, p->cpus_ptr))
6562 util = cpu_util_next(cpu, p, cpu);
6563 cpu_cap = capacity_of(cpu);
6564 spare_cap = cpu_cap - util;
6567 * Skip CPUs that cannot satisfy the capacity request.
6568 * IOW, placing the task there would make the CPU
6569 * overutilized. Take uclamp into account to see how
6570 * much capacity we can get out of the CPU; this is
6571 * aligned with schedutil_cpu_util().
6573 util = uclamp_rq_util_with(cpu_rq(cpu), util, p);
6574 if (!fits_capacity(util, cpu_cap))
6577 /* Always use prev_cpu as a candidate. */
6578 if (cpu == prev_cpu) {
6579 prev_delta = compute_energy(p, prev_cpu, pd);
6580 prev_delta -= base_energy_pd;
6581 best_delta = min(best_delta, prev_delta);
6585 * Find the CPU with the maximum spare capacity in
6586 * the performance domain
6588 if (spare_cap > max_spare_cap) {
6589 max_spare_cap = spare_cap;
6590 max_spare_cap_cpu = cpu;
6594 /* Evaluate the energy impact of using this CPU. */
6595 if (max_spare_cap_cpu >= 0 && max_spare_cap_cpu != prev_cpu) {
6596 cur_delta = compute_energy(p, max_spare_cap_cpu, pd);
6597 cur_delta -= base_energy_pd;
6598 if (cur_delta < best_delta) {
6599 best_delta = cur_delta;
6600 best_energy_cpu = max_spare_cap_cpu;
6608 * Pick the best CPU if prev_cpu cannot be used, or if it saves at
6609 * least 6% of the energy used by prev_cpu.
6611 if (prev_delta == ULONG_MAX)
6612 return best_energy_cpu;
6614 if ((prev_delta - best_delta) > ((prev_delta + base_energy) >> 4))
6615 return best_energy_cpu;
6626 * select_task_rq_fair: Select target runqueue for the waking task in domains
6627 * that have the 'sd_flag' flag set. In practice, this is SD_BALANCE_WAKE,
6628 * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
6630 * Balances load by selecting the idlest CPU in the idlest group, or under
6631 * certain conditions an idle sibling CPU if the domain has SD_WAKE_AFFINE set.
6633 * Returns the target CPU number.
6635 * preempt must be disabled.
6638 select_task_rq_fair(struct task_struct *p, int prev_cpu, int sd_flag, int wake_flags)
6640 struct sched_domain *tmp, *sd = NULL;
6641 int cpu = smp_processor_id();
6642 int new_cpu = prev_cpu;
6643 int want_affine = 0;
6644 int sync = (wake_flags & WF_SYNC) && !(current->flags & PF_EXITING);
6646 if (sd_flag & SD_BALANCE_WAKE) {
6649 if (sched_energy_enabled()) {
6650 new_cpu = find_energy_efficient_cpu(p, prev_cpu);
6656 want_affine = !wake_wide(p) && cpumask_test_cpu(cpu, p->cpus_ptr);
6660 for_each_domain(cpu, tmp) {
6661 if (!(tmp->flags & SD_LOAD_BALANCE))
6665 * If both 'cpu' and 'prev_cpu' are part of this domain,
6666 * cpu is a valid SD_WAKE_AFFINE target.
6668 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
6669 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
6670 if (cpu != prev_cpu)
6671 new_cpu = wake_affine(tmp, p, cpu, prev_cpu, sync);
6673 sd = NULL; /* Prefer wake_affine over balance flags */
6677 if (tmp->flags & sd_flag)
6679 else if (!want_affine)
6685 new_cpu = find_idlest_cpu(sd, p, cpu, prev_cpu, sd_flag);
6686 } else if (sd_flag & SD_BALANCE_WAKE) { /* XXX always ? */
6689 new_cpu = select_idle_sibling(p, prev_cpu, new_cpu);
6692 current->recent_used_cpu = cpu;
6699 static void detach_entity_cfs_rq(struct sched_entity *se);
6702 * Called immediately before a task is migrated to a new CPU; task_cpu(p) and
6703 * cfs_rq_of(p) references at time of call are still valid and identify the
6704 * previous CPU. The caller guarantees p->pi_lock or task_rq(p)->lock is held.
6706 static void migrate_task_rq_fair(struct task_struct *p, int new_cpu)
6709 * As blocked tasks retain absolute vruntime the migration needs to
6710 * deal with this by subtracting the old and adding the new
6711 * min_vruntime -- the latter is done by enqueue_entity() when placing
6712 * the task on the new runqueue.
6714 if (p->state == TASK_WAKING) {
6715 struct sched_entity *se = &p->se;
6716 struct cfs_rq *cfs_rq = cfs_rq_of(se);
6719 #ifndef CONFIG_64BIT
6720 u64 min_vruntime_copy;
6723 min_vruntime_copy = cfs_rq->min_vruntime_copy;
6725 min_vruntime = cfs_rq->min_vruntime;
6726 } while (min_vruntime != min_vruntime_copy);
6728 min_vruntime = cfs_rq->min_vruntime;
6731 se->vruntime -= min_vruntime;
6734 if (p->on_rq == TASK_ON_RQ_MIGRATING) {
6736 * In case of TASK_ON_RQ_MIGRATING we in fact hold the 'old'
6737 * rq->lock and can modify state directly.
6739 lockdep_assert_held(&task_rq(p)->lock);
6740 detach_entity_cfs_rq(&p->se);
6744 * We are supposed to update the task to "current" time, then
6745 * its up to date and ready to go to new CPU/cfs_rq. But we
6746 * have difficulty in getting what current time is, so simply
6747 * throw away the out-of-date time. This will result in the
6748 * wakee task is less decayed, but giving the wakee more load
6751 remove_entity_load_avg(&p->se);
6754 /* Tell new CPU we are migrated */
6755 p->se.avg.last_update_time = 0;
6757 /* We have migrated, no longer consider this task hot */
6758 p->se.exec_start = 0;
6760 update_scan_period(p, new_cpu);
6763 static void task_dead_fair(struct task_struct *p)
6765 remove_entity_load_avg(&p->se);
6769 balance_fair(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
6774 return newidle_balance(rq, rf) != 0;
6776 #endif /* CONFIG_SMP */
6778 static unsigned long wakeup_gran(struct sched_entity *se)
6780 unsigned long gran = sysctl_sched_wakeup_granularity;
6783 * Since its curr running now, convert the gran from real-time
6784 * to virtual-time in his units.
6786 * By using 'se' instead of 'curr' we penalize light tasks, so
6787 * they get preempted easier. That is, if 'se' < 'curr' then
6788 * the resulting gran will be larger, therefore penalizing the
6789 * lighter, if otoh 'se' > 'curr' then the resulting gran will
6790 * be smaller, again penalizing the lighter task.
6792 * This is especially important for buddies when the leftmost
6793 * task is higher priority than the buddy.
6795 return calc_delta_fair(gran, se);
6799 * Should 'se' preempt 'curr'.
6813 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
6815 s64 gran, vdiff = curr->vruntime - se->vruntime;
6820 gran = wakeup_gran(se);
6827 static void set_last_buddy(struct sched_entity *se)
6829 if (entity_is_task(se) && unlikely(task_has_idle_policy(task_of(se))))
6832 for_each_sched_entity(se) {
6833 if (SCHED_WARN_ON(!se->on_rq))
6835 cfs_rq_of(se)->last = se;
6839 static void set_next_buddy(struct sched_entity *se)
6841 if (entity_is_task(se) && unlikely(task_has_idle_policy(task_of(se))))
6844 for_each_sched_entity(se) {
6845 if (SCHED_WARN_ON(!se->on_rq))
6847 cfs_rq_of(se)->next = se;
6851 static void set_skip_buddy(struct sched_entity *se)
6853 for_each_sched_entity(se)
6854 cfs_rq_of(se)->skip = se;
6858 * Preempt the current task with a newly woken task if needed:
6860 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
6862 struct task_struct *curr = rq->curr;
6863 struct sched_entity *se = &curr->se, *pse = &p->se;
6864 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
6865 int scale = cfs_rq->nr_running >= sched_nr_latency;
6866 int next_buddy_marked = 0;
6868 if (unlikely(se == pse))
6872 * This is possible from callers such as attach_tasks(), in which we
6873 * unconditionally check_prempt_curr() after an enqueue (which may have
6874 * lead to a throttle). This both saves work and prevents false
6875 * next-buddy nomination below.
6877 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
6880 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
6881 set_next_buddy(pse);
6882 next_buddy_marked = 1;
6886 * We can come here with TIF_NEED_RESCHED already set from new task
6889 * Note: this also catches the edge-case of curr being in a throttled
6890 * group (e.g. via set_curr_task), since update_curr() (in the
6891 * enqueue of curr) will have resulted in resched being set. This
6892 * prevents us from potentially nominating it as a false LAST_BUDDY
6895 if (test_tsk_need_resched(curr))
6898 /* Idle tasks are by definition preempted by non-idle tasks. */
6899 if (unlikely(task_has_idle_policy(curr)) &&
6900 likely(!task_has_idle_policy(p)))
6904 * Batch and idle tasks do not preempt non-idle tasks (their preemption
6905 * is driven by the tick):
6907 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
6910 find_matching_se(&se, &pse);
6911 update_curr(cfs_rq_of(se));
6913 if (wakeup_preempt_entity(se, pse) == 1) {
6915 * Bias pick_next to pick the sched entity that is
6916 * triggering this preemption.
6918 if (!next_buddy_marked)
6919 set_next_buddy(pse);
6928 * Only set the backward buddy when the current task is still
6929 * on the rq. This can happen when a wakeup gets interleaved
6930 * with schedule on the ->pre_schedule() or idle_balance()
6931 * point, either of which can * drop the rq lock.
6933 * Also, during early boot the idle thread is in the fair class,
6934 * for obvious reasons its a bad idea to schedule back to it.
6936 if (unlikely(!se->on_rq || curr == rq->idle))
6939 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
6943 struct task_struct *
6944 pick_next_task_fair(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
6946 struct cfs_rq *cfs_rq = &rq->cfs;
6947 struct sched_entity *se;
6948 struct task_struct *p;
6952 if (!sched_fair_runnable(rq))
6955 #ifdef CONFIG_FAIR_GROUP_SCHED
6956 if (!prev || prev->sched_class != &fair_sched_class)
6960 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
6961 * likely that a next task is from the same cgroup as the current.
6963 * Therefore attempt to avoid putting and setting the entire cgroup
6964 * hierarchy, only change the part that actually changes.
6968 struct sched_entity *curr = cfs_rq->curr;
6971 * Since we got here without doing put_prev_entity() we also
6972 * have to consider cfs_rq->curr. If it is still a runnable
6973 * entity, update_curr() will update its vruntime, otherwise
6974 * forget we've ever seen it.
6978 update_curr(cfs_rq);
6983 * This call to check_cfs_rq_runtime() will do the
6984 * throttle and dequeue its entity in the parent(s).
6985 * Therefore the nr_running test will indeed
6988 if (unlikely(check_cfs_rq_runtime(cfs_rq))) {
6991 if (!cfs_rq->nr_running)
6998 se = pick_next_entity(cfs_rq, curr);
6999 cfs_rq = group_cfs_rq(se);
7005 * Since we haven't yet done put_prev_entity and if the selected task
7006 * is a different task than we started out with, try and touch the
7007 * least amount of cfs_rqs.
7010 struct sched_entity *pse = &prev->se;
7012 while (!(cfs_rq = is_same_group(se, pse))) {
7013 int se_depth = se->depth;
7014 int pse_depth = pse->depth;
7016 if (se_depth <= pse_depth) {
7017 put_prev_entity(cfs_rq_of(pse), pse);
7018 pse = parent_entity(pse);
7020 if (se_depth >= pse_depth) {
7021 set_next_entity(cfs_rq_of(se), se);
7022 se = parent_entity(se);
7026 put_prev_entity(cfs_rq, pse);
7027 set_next_entity(cfs_rq, se);
7034 put_prev_task(rq, prev);
7037 se = pick_next_entity(cfs_rq, NULL);
7038 set_next_entity(cfs_rq, se);
7039 cfs_rq = group_cfs_rq(se);
7044 done: __maybe_unused;
7047 * Move the next running task to the front of
7048 * the list, so our cfs_tasks list becomes MRU
7051 list_move(&p->se.group_node, &rq->cfs_tasks);
7054 if (hrtick_enabled(rq))
7055 hrtick_start_fair(rq, p);
7057 update_misfit_status(p, rq);
7065 new_tasks = newidle_balance(rq, rf);
7068 * Because newidle_balance() releases (and re-acquires) rq->lock, it is
7069 * possible for any higher priority task to appear. In that case we
7070 * must re-start the pick_next_entity() loop.
7079 * rq is about to be idle, check if we need to update the
7080 * lost_idle_time of clock_pelt
7082 update_idle_rq_clock_pelt(rq);
7087 static struct task_struct *__pick_next_task_fair(struct rq *rq)
7089 return pick_next_task_fair(rq, NULL, NULL);
7093 * Account for a descheduled task:
7095 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
7097 struct sched_entity *se = &prev->se;
7098 struct cfs_rq *cfs_rq;
7100 for_each_sched_entity(se) {
7101 cfs_rq = cfs_rq_of(se);
7102 put_prev_entity(cfs_rq, se);
7107 * sched_yield() is very simple
7109 * The magic of dealing with the ->skip buddy is in pick_next_entity.
7111 static void yield_task_fair(struct rq *rq)
7113 struct task_struct *curr = rq->curr;
7114 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
7115 struct sched_entity *se = &curr->se;
7118 * Are we the only task in the tree?
7120 if (unlikely(rq->nr_running == 1))
7123 clear_buddies(cfs_rq, se);
7125 if (curr->policy != SCHED_BATCH) {
7126 update_rq_clock(rq);
7128 * Update run-time statistics of the 'current'.
7130 update_curr(cfs_rq);
7132 * Tell update_rq_clock() that we've just updated,
7133 * so we don't do microscopic update in schedule()
7134 * and double the fastpath cost.
7136 rq_clock_skip_update(rq);
7142 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
7144 struct sched_entity *se = &p->se;
7146 /* throttled hierarchies are not runnable */
7147 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
7150 /* Tell the scheduler that we'd really like pse to run next. */
7153 yield_task_fair(rq);
7159 /**************************************************
7160 * Fair scheduling class load-balancing methods.
7164 * The purpose of load-balancing is to achieve the same basic fairness the
7165 * per-CPU scheduler provides, namely provide a proportional amount of compute
7166 * time to each task. This is expressed in the following equation:
7168 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
7170 * Where W_i,n is the n-th weight average for CPU i. The instantaneous weight
7171 * W_i,0 is defined as:
7173 * W_i,0 = \Sum_j w_i,j (2)
7175 * Where w_i,j is the weight of the j-th runnable task on CPU i. This weight
7176 * is derived from the nice value as per sched_prio_to_weight[].
7178 * The weight average is an exponential decay average of the instantaneous
7181 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
7183 * C_i is the compute capacity of CPU i, typically it is the
7184 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
7185 * can also include other factors [XXX].
7187 * To achieve this balance we define a measure of imbalance which follows
7188 * directly from (1):
7190 * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4)
7192 * We them move tasks around to minimize the imbalance. In the continuous
7193 * function space it is obvious this converges, in the discrete case we get
7194 * a few fun cases generally called infeasible weight scenarios.
7197 * - infeasible weights;
7198 * - local vs global optima in the discrete case. ]
7203 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
7204 * for all i,j solution, we create a tree of CPUs that follows the hardware
7205 * topology where each level pairs two lower groups (or better). This results
7206 * in O(log n) layers. Furthermore we reduce the number of CPUs going up the
7207 * tree to only the first of the previous level and we decrease the frequency
7208 * of load-balance at each level inv. proportional to the number of CPUs in
7214 * \Sum { --- * --- * 2^i } = O(n) (5)
7216 * `- size of each group
7217 * | | `- number of CPUs doing load-balance
7219 * `- sum over all levels
7221 * Coupled with a limit on how many tasks we can migrate every balance pass,
7222 * this makes (5) the runtime complexity of the balancer.
7224 * An important property here is that each CPU is still (indirectly) connected
7225 * to every other CPU in at most O(log n) steps:
7227 * The adjacency matrix of the resulting graph is given by:
7230 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
7233 * And you'll find that:
7235 * A^(log_2 n)_i,j != 0 for all i,j (7)
7237 * Showing there's indeed a path between every CPU in at most O(log n) steps.
7238 * The task movement gives a factor of O(m), giving a convergence complexity
7241 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
7246 * In order to avoid CPUs going idle while there's still work to do, new idle
7247 * balancing is more aggressive and has the newly idle CPU iterate up the domain
7248 * tree itself instead of relying on other CPUs to bring it work.
7250 * This adds some complexity to both (5) and (8) but it reduces the total idle
7258 * Cgroups make a horror show out of (2), instead of a simple sum we get:
7261 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
7266 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
7268 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on CPU i.
7270 * The big problem is S_k, its a global sum needed to compute a local (W_i)
7273 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
7274 * rewrite all of this once again.]
7277 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
7279 enum fbq_type { regular, remote, all };
7282 * 'group_type' describes the group of CPUs at the moment of load balancing.
7284 * The enum is ordered by pulling priority, with the group with lowest priority
7285 * first so the group_type can simply be compared when selecting the busiest
7286 * group. See update_sd_pick_busiest().
7289 /* The group has spare capacity that can be used to run more tasks. */
7290 group_has_spare = 0,
7292 * The group is fully used and the tasks don't compete for more CPU
7293 * cycles. Nevertheless, some tasks might wait before running.
7297 * SD_ASYM_CPUCAPACITY only: One task doesn't fit with CPU's capacity
7298 * and must be migrated to a more powerful CPU.
7302 * SD_ASYM_PACKING only: One local CPU with higher capacity is available,
7303 * and the task should be migrated to it instead of running on the
7308 * The tasks' affinity constraints previously prevented the scheduler
7309 * from balancing the load across the system.
7313 * The CPU is overloaded and can't provide expected CPU cycles to all
7319 enum migration_type {
7326 #define LBF_ALL_PINNED 0x01
7327 #define LBF_NEED_BREAK 0x02
7328 #define LBF_DST_PINNED 0x04
7329 #define LBF_SOME_PINNED 0x08
7330 #define LBF_NOHZ_STATS 0x10
7331 #define LBF_NOHZ_AGAIN 0x20
7334 struct sched_domain *sd;
7342 struct cpumask *dst_grpmask;
7344 enum cpu_idle_type idle;
7346 /* The set of CPUs under consideration for load-balancing */
7347 struct cpumask *cpus;
7352 unsigned int loop_break;
7353 unsigned int loop_max;
7355 enum fbq_type fbq_type;
7356 enum migration_type migration_type;
7357 struct list_head tasks;
7361 * Is this task likely cache-hot:
7363 static int task_hot(struct task_struct *p, struct lb_env *env)
7367 lockdep_assert_held(&env->src_rq->lock);
7369 if (p->sched_class != &fair_sched_class)
7372 if (unlikely(task_has_idle_policy(p)))
7376 * Buddy candidates are cache hot:
7378 if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running &&
7379 (&p->se == cfs_rq_of(&p->se)->next ||
7380 &p->se == cfs_rq_of(&p->se)->last))
7383 if (sysctl_sched_migration_cost == -1)
7385 if (sysctl_sched_migration_cost == 0)
7388 delta = rq_clock_task(env->src_rq) - p->se.exec_start;
7390 return delta < (s64)sysctl_sched_migration_cost;
7393 #ifdef CONFIG_NUMA_BALANCING
7395 * Returns 1, if task migration degrades locality
7396 * Returns 0, if task migration improves locality i.e migration preferred.
7397 * Returns -1, if task migration is not affected by locality.
7399 static int migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
7401 struct numa_group *numa_group = rcu_dereference(p->numa_group);
7402 unsigned long src_weight, dst_weight;
7403 int src_nid, dst_nid, dist;
7405 if (!static_branch_likely(&sched_numa_balancing))
7408 if (!p->numa_faults || !(env->sd->flags & SD_NUMA))
7411 src_nid = cpu_to_node(env->src_cpu);
7412 dst_nid = cpu_to_node(env->dst_cpu);
7414 if (src_nid == dst_nid)
7417 /* Migrating away from the preferred node is always bad. */
7418 if (src_nid == p->numa_preferred_nid) {
7419 if (env->src_rq->nr_running > env->src_rq->nr_preferred_running)
7425 /* Encourage migration to the preferred node. */
7426 if (dst_nid == p->numa_preferred_nid)
7429 /* Leaving a core idle is often worse than degrading locality. */
7430 if (env->idle == CPU_IDLE)
7433 dist = node_distance(src_nid, dst_nid);
7435 src_weight = group_weight(p, src_nid, dist);
7436 dst_weight = group_weight(p, dst_nid, dist);
7438 src_weight = task_weight(p, src_nid, dist);
7439 dst_weight = task_weight(p, dst_nid, dist);
7442 return dst_weight < src_weight;
7446 static inline int migrate_degrades_locality(struct task_struct *p,
7454 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
7457 int can_migrate_task(struct task_struct *p, struct lb_env *env)
7461 lockdep_assert_held(&env->src_rq->lock);
7464 * We do not migrate tasks that are:
7465 * 1) throttled_lb_pair, or
7466 * 2) cannot be migrated to this CPU due to cpus_ptr, or
7467 * 3) running (obviously), or
7468 * 4) are cache-hot on their current CPU.
7470 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
7473 if (!cpumask_test_cpu(env->dst_cpu, p->cpus_ptr)) {
7476 schedstat_inc(p->se.statistics.nr_failed_migrations_affine);
7478 env->flags |= LBF_SOME_PINNED;
7481 * Remember if this task can be migrated to any other CPU in
7482 * our sched_group. We may want to revisit it if we couldn't
7483 * meet load balance goals by pulling other tasks on src_cpu.
7485 * Avoid computing new_dst_cpu for NEWLY_IDLE or if we have
7486 * already computed one in current iteration.
7488 if (env->idle == CPU_NEWLY_IDLE || (env->flags & LBF_DST_PINNED))
7491 /* Prevent to re-select dst_cpu via env's CPUs: */
7492 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
7493 if (cpumask_test_cpu(cpu, p->cpus_ptr)) {
7494 env->flags |= LBF_DST_PINNED;
7495 env->new_dst_cpu = cpu;
7503 /* Record that we found atleast one task that could run on dst_cpu */
7504 env->flags &= ~LBF_ALL_PINNED;
7506 if (task_running(env->src_rq, p)) {
7507 schedstat_inc(p->se.statistics.nr_failed_migrations_running);
7512 * Aggressive migration if:
7513 * 1) destination numa is preferred
7514 * 2) task is cache cold, or
7515 * 3) too many balance attempts have failed.
7517 tsk_cache_hot = migrate_degrades_locality(p, env);
7518 if (tsk_cache_hot == -1)
7519 tsk_cache_hot = task_hot(p, env);
7521 if (tsk_cache_hot <= 0 ||
7522 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
7523 if (tsk_cache_hot == 1) {
7524 schedstat_inc(env->sd->lb_hot_gained[env->idle]);
7525 schedstat_inc(p->se.statistics.nr_forced_migrations);
7530 schedstat_inc(p->se.statistics.nr_failed_migrations_hot);
7535 * detach_task() -- detach the task for the migration specified in env
7537 static void detach_task(struct task_struct *p, struct lb_env *env)
7539 lockdep_assert_held(&env->src_rq->lock);
7541 deactivate_task(env->src_rq, p, DEQUEUE_NOCLOCK);
7542 set_task_cpu(p, env->dst_cpu);
7546 * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
7547 * part of active balancing operations within "domain".
7549 * Returns a task if successful and NULL otherwise.
7551 static struct task_struct *detach_one_task(struct lb_env *env)
7553 struct task_struct *p;
7555 lockdep_assert_held(&env->src_rq->lock);
7557 list_for_each_entry_reverse(p,
7558 &env->src_rq->cfs_tasks, se.group_node) {
7559 if (!can_migrate_task(p, env))
7562 detach_task(p, env);
7565 * Right now, this is only the second place where
7566 * lb_gained[env->idle] is updated (other is detach_tasks)
7567 * so we can safely collect stats here rather than
7568 * inside detach_tasks().
7570 schedstat_inc(env->sd->lb_gained[env->idle]);
7576 static const unsigned int sched_nr_migrate_break = 32;
7579 * detach_tasks() -- tries to detach up to imbalance load/util/tasks from
7580 * busiest_rq, as part of a balancing operation within domain "sd".
7582 * Returns number of detached tasks if successful and 0 otherwise.
7584 static int detach_tasks(struct lb_env *env)
7586 struct list_head *tasks = &env->src_rq->cfs_tasks;
7587 unsigned long util, load;
7588 struct task_struct *p;
7591 lockdep_assert_held(&env->src_rq->lock);
7593 if (env->imbalance <= 0)
7596 while (!list_empty(tasks)) {
7598 * We don't want to steal all, otherwise we may be treated likewise,
7599 * which could at worst lead to a livelock crash.
7601 if (env->idle != CPU_NOT_IDLE && env->src_rq->nr_running <= 1)
7604 p = list_last_entry(tasks, struct task_struct, se.group_node);
7607 /* We've more or less seen every task there is, call it quits */
7608 if (env->loop > env->loop_max)
7611 /* take a breather every nr_migrate tasks */
7612 if (env->loop > env->loop_break) {
7613 env->loop_break += sched_nr_migrate_break;
7614 env->flags |= LBF_NEED_BREAK;
7618 if (!can_migrate_task(p, env))
7621 switch (env->migration_type) {
7623 load = task_h_load(p);
7625 if (sched_feat(LB_MIN) &&
7626 load < 16 && !env->sd->nr_balance_failed)
7630 * Make sure that we don't migrate too much load.
7631 * Nevertheless, let relax the constraint if
7632 * scheduler fails to find a good waiting task to
7635 if (load/2 > env->imbalance &&
7636 env->sd->nr_balance_failed <= env->sd->cache_nice_tries)
7639 env->imbalance -= load;
7643 util = task_util_est(p);
7645 if (util > env->imbalance)
7648 env->imbalance -= util;
7655 case migrate_misfit:
7656 /* This is not a misfit task */
7657 if (task_fits_capacity(p, capacity_of(env->src_cpu)))
7664 detach_task(p, env);
7665 list_add(&p->se.group_node, &env->tasks);
7669 #ifdef CONFIG_PREEMPTION
7671 * NEWIDLE balancing is a source of latency, so preemptible
7672 * kernels will stop after the first task is detached to minimize
7673 * the critical section.
7675 if (env->idle == CPU_NEWLY_IDLE)
7680 * We only want to steal up to the prescribed amount of
7683 if (env->imbalance <= 0)
7688 list_move(&p->se.group_node, tasks);
7692 * Right now, this is one of only two places we collect this stat
7693 * so we can safely collect detach_one_task() stats here rather
7694 * than inside detach_one_task().
7696 schedstat_add(env->sd->lb_gained[env->idle], detached);
7702 * attach_task() -- attach the task detached by detach_task() to its new rq.
7704 static void attach_task(struct rq *rq, struct task_struct *p)
7706 lockdep_assert_held(&rq->lock);
7708 BUG_ON(task_rq(p) != rq);
7709 activate_task(rq, p, ENQUEUE_NOCLOCK);
7710 check_preempt_curr(rq, p, 0);
7714 * attach_one_task() -- attaches the task returned from detach_one_task() to
7717 static void attach_one_task(struct rq *rq, struct task_struct *p)
7722 update_rq_clock(rq);
7728 * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
7731 static void attach_tasks(struct lb_env *env)
7733 struct list_head *tasks = &env->tasks;
7734 struct task_struct *p;
7737 rq_lock(env->dst_rq, &rf);
7738 update_rq_clock(env->dst_rq);
7740 while (!list_empty(tasks)) {
7741 p = list_first_entry(tasks, struct task_struct, se.group_node);
7742 list_del_init(&p->se.group_node);
7744 attach_task(env->dst_rq, p);
7747 rq_unlock(env->dst_rq, &rf);
7750 #ifdef CONFIG_NO_HZ_COMMON
7751 static inline bool cfs_rq_has_blocked(struct cfs_rq *cfs_rq)
7753 if (cfs_rq->avg.load_avg)
7756 if (cfs_rq->avg.util_avg)
7762 static inline bool others_have_blocked(struct rq *rq)
7764 if (READ_ONCE(rq->avg_rt.util_avg))
7767 if (READ_ONCE(rq->avg_dl.util_avg))
7770 if (thermal_load_avg(rq))
7773 #ifdef CONFIG_HAVE_SCHED_AVG_IRQ
7774 if (READ_ONCE(rq->avg_irq.util_avg))
7781 static inline void update_blocked_load_status(struct rq *rq, bool has_blocked)
7783 rq->last_blocked_load_update_tick = jiffies;
7786 rq->has_blocked_load = 0;
7789 static inline bool cfs_rq_has_blocked(struct cfs_rq *cfs_rq) { return false; }
7790 static inline bool others_have_blocked(struct rq *rq) { return false; }
7791 static inline void update_blocked_load_status(struct rq *rq, bool has_blocked) {}
7794 static bool __update_blocked_others(struct rq *rq, bool *done)
7796 const struct sched_class *curr_class;
7797 u64 now = rq_clock_pelt(rq);
7798 unsigned long thermal_pressure;
7802 * update_load_avg() can call cpufreq_update_util(). Make sure that RT,
7803 * DL and IRQ signals have been updated before updating CFS.
7805 curr_class = rq->curr->sched_class;
7807 thermal_pressure = arch_scale_thermal_pressure(cpu_of(rq));
7809 decayed = update_rt_rq_load_avg(now, rq, curr_class == &rt_sched_class) |
7810 update_dl_rq_load_avg(now, rq, curr_class == &dl_sched_class) |
7811 update_thermal_load_avg(rq_clock_thermal(rq), rq, thermal_pressure) |
7812 update_irq_load_avg(rq, 0);
7814 if (others_have_blocked(rq))
7820 #ifdef CONFIG_FAIR_GROUP_SCHED
7822 static inline bool cfs_rq_is_decayed(struct cfs_rq *cfs_rq)
7824 if (cfs_rq->load.weight)
7827 if (cfs_rq->avg.load_sum)
7830 if (cfs_rq->avg.util_sum)
7833 if (cfs_rq->avg.runnable_sum)
7839 static bool __update_blocked_fair(struct rq *rq, bool *done)
7841 struct cfs_rq *cfs_rq, *pos;
7842 bool decayed = false;
7843 int cpu = cpu_of(rq);
7846 * Iterates the task_group tree in a bottom up fashion, see
7847 * list_add_leaf_cfs_rq() for details.
7849 for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) {
7850 struct sched_entity *se;
7852 if (update_cfs_rq_load_avg(cfs_rq_clock_pelt(cfs_rq), cfs_rq)) {
7853 update_tg_load_avg(cfs_rq, 0);
7855 if (cfs_rq == &rq->cfs)
7859 /* Propagate pending load changes to the parent, if any: */
7860 se = cfs_rq->tg->se[cpu];
7861 if (se && !skip_blocked_update(se))
7862 update_load_avg(cfs_rq_of(se), se, 0);
7865 * There can be a lot of idle CPU cgroups. Don't let fully
7866 * decayed cfs_rqs linger on the list.
7868 if (cfs_rq_is_decayed(cfs_rq))
7869 list_del_leaf_cfs_rq(cfs_rq);
7871 /* Don't need periodic decay once load/util_avg are null */
7872 if (cfs_rq_has_blocked(cfs_rq))
7880 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
7881 * This needs to be done in a top-down fashion because the load of a child
7882 * group is a fraction of its parents load.
7884 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
7886 struct rq *rq = rq_of(cfs_rq);
7887 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
7888 unsigned long now = jiffies;
7891 if (cfs_rq->last_h_load_update == now)
7894 WRITE_ONCE(cfs_rq->h_load_next, NULL);
7895 for_each_sched_entity(se) {
7896 cfs_rq = cfs_rq_of(se);
7897 WRITE_ONCE(cfs_rq->h_load_next, se);
7898 if (cfs_rq->last_h_load_update == now)
7903 cfs_rq->h_load = cfs_rq_load_avg(cfs_rq);
7904 cfs_rq->last_h_load_update = now;
7907 while ((se = READ_ONCE(cfs_rq->h_load_next)) != NULL) {
7908 load = cfs_rq->h_load;
7909 load = div64_ul(load * se->avg.load_avg,
7910 cfs_rq_load_avg(cfs_rq) + 1);
7911 cfs_rq = group_cfs_rq(se);
7912 cfs_rq->h_load = load;
7913 cfs_rq->last_h_load_update = now;
7917 static unsigned long task_h_load(struct task_struct *p)
7919 struct cfs_rq *cfs_rq = task_cfs_rq(p);
7921 update_cfs_rq_h_load(cfs_rq);
7922 return div64_ul(p->se.avg.load_avg * cfs_rq->h_load,
7923 cfs_rq_load_avg(cfs_rq) + 1);
7926 static bool __update_blocked_fair(struct rq *rq, bool *done)
7928 struct cfs_rq *cfs_rq = &rq->cfs;
7931 decayed = update_cfs_rq_load_avg(cfs_rq_clock_pelt(cfs_rq), cfs_rq);
7932 if (cfs_rq_has_blocked(cfs_rq))
7938 static unsigned long task_h_load(struct task_struct *p)
7940 return p->se.avg.load_avg;
7944 static void update_blocked_averages(int cpu)
7946 bool decayed = false, done = true;
7947 struct rq *rq = cpu_rq(cpu);
7950 rq_lock_irqsave(rq, &rf);
7951 update_rq_clock(rq);
7953 decayed |= __update_blocked_others(rq, &done);
7954 decayed |= __update_blocked_fair(rq, &done);
7956 update_blocked_load_status(rq, !done);
7958 cpufreq_update_util(rq, 0);
7959 rq_unlock_irqrestore(rq, &rf);
7962 /********** Helpers for find_busiest_group ************************/
7965 * sg_lb_stats - stats of a sched_group required for load_balancing
7967 struct sg_lb_stats {
7968 unsigned long avg_load; /*Avg load across the CPUs of the group */
7969 unsigned long group_load; /* Total load over the CPUs of the group */
7970 unsigned long group_capacity;
7971 unsigned long group_util; /* Total utilization over the CPUs of the group */
7972 unsigned long group_runnable; /* Total runnable time over the CPUs of the group */
7973 unsigned int sum_nr_running; /* Nr of tasks running in the group */
7974 unsigned int sum_h_nr_running; /* Nr of CFS tasks running in the group */
7975 unsigned int idle_cpus;
7976 unsigned int group_weight;
7977 enum group_type group_type;
7978 unsigned int group_asym_packing; /* Tasks should be moved to preferred CPU */
7979 unsigned long group_misfit_task_load; /* A CPU has a task too big for its capacity */
7980 #ifdef CONFIG_NUMA_BALANCING
7981 unsigned int nr_numa_running;
7982 unsigned int nr_preferred_running;
7987 * sd_lb_stats - Structure to store the statistics of a sched_domain
7988 * during load balancing.
7990 struct sd_lb_stats {
7991 struct sched_group *busiest; /* Busiest group in this sd */
7992 struct sched_group *local; /* Local group in this sd */
7993 unsigned long total_load; /* Total load of all groups in sd */
7994 unsigned long total_capacity; /* Total capacity of all groups in sd */
7995 unsigned long avg_load; /* Average load across all groups in sd */
7996 unsigned int prefer_sibling; /* tasks should go to sibling first */
7998 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
7999 struct sg_lb_stats local_stat; /* Statistics of the local group */
8002 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
8005 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
8006 * local_stat because update_sg_lb_stats() does a full clear/assignment.
8007 * We must however set busiest_stat::group_type and
8008 * busiest_stat::idle_cpus to the worst busiest group because
8009 * update_sd_pick_busiest() reads these before assignment.
8011 *sds = (struct sd_lb_stats){
8015 .total_capacity = 0UL,
8017 .idle_cpus = UINT_MAX,
8018 .group_type = group_has_spare,
8023 static unsigned long scale_rt_capacity(struct sched_domain *sd, int cpu)
8025 struct rq *rq = cpu_rq(cpu);
8026 unsigned long max = arch_scale_cpu_capacity(cpu);
8027 unsigned long used, free;
8030 irq = cpu_util_irq(rq);
8032 if (unlikely(irq >= max))
8036 * avg_rt.util_avg and avg_dl.util_avg track binary signals
8037 * (running and not running) with weights 0 and 1024 respectively.
8038 * avg_thermal.load_avg tracks thermal pressure and the weighted
8039 * average uses the actual delta max capacity(load).
8041 used = READ_ONCE(rq->avg_rt.util_avg);
8042 used += READ_ONCE(rq->avg_dl.util_avg);
8043 used += thermal_load_avg(rq);
8045 if (unlikely(used >= max))
8050 return scale_irq_capacity(free, irq, max);
8053 static void update_cpu_capacity(struct sched_domain *sd, int cpu)
8055 unsigned long capacity = scale_rt_capacity(sd, cpu);
8056 struct sched_group *sdg = sd->groups;
8058 cpu_rq(cpu)->cpu_capacity_orig = arch_scale_cpu_capacity(cpu);
8063 cpu_rq(cpu)->cpu_capacity = capacity;
8064 sdg->sgc->capacity = capacity;
8065 sdg->sgc->min_capacity = capacity;
8066 sdg->sgc->max_capacity = capacity;
8069 void update_group_capacity(struct sched_domain *sd, int cpu)
8071 struct sched_domain *child = sd->child;
8072 struct sched_group *group, *sdg = sd->groups;
8073 unsigned long capacity, min_capacity, max_capacity;
8074 unsigned long interval;
8076 interval = msecs_to_jiffies(sd->balance_interval);
8077 interval = clamp(interval, 1UL, max_load_balance_interval);
8078 sdg->sgc->next_update = jiffies + interval;
8081 update_cpu_capacity(sd, cpu);
8086 min_capacity = ULONG_MAX;
8089 if (child->flags & SD_OVERLAP) {
8091 * SD_OVERLAP domains cannot assume that child groups
8092 * span the current group.
8095 for_each_cpu(cpu, sched_group_span(sdg)) {
8096 unsigned long cpu_cap = capacity_of(cpu);
8098 capacity += cpu_cap;
8099 min_capacity = min(cpu_cap, min_capacity);
8100 max_capacity = max(cpu_cap, max_capacity);
8104 * !SD_OVERLAP domains can assume that child groups
8105 * span the current group.
8108 group = child->groups;
8110 struct sched_group_capacity *sgc = group->sgc;
8112 capacity += sgc->capacity;
8113 min_capacity = min(sgc->min_capacity, min_capacity);
8114 max_capacity = max(sgc->max_capacity, max_capacity);
8115 group = group->next;
8116 } while (group != child->groups);
8119 sdg->sgc->capacity = capacity;
8120 sdg->sgc->min_capacity = min_capacity;
8121 sdg->sgc->max_capacity = max_capacity;
8125 * Check whether the capacity of the rq has been noticeably reduced by side
8126 * activity. The imbalance_pct is used for the threshold.
8127 * Return true is the capacity is reduced
8130 check_cpu_capacity(struct rq *rq, struct sched_domain *sd)
8132 return ((rq->cpu_capacity * sd->imbalance_pct) <
8133 (rq->cpu_capacity_orig * 100));
8137 * Check whether a rq has a misfit task and if it looks like we can actually
8138 * help that task: we can migrate the task to a CPU of higher capacity, or
8139 * the task's current CPU is heavily pressured.
8141 static inline int check_misfit_status(struct rq *rq, struct sched_domain *sd)
8143 return rq->misfit_task_load &&
8144 (rq->cpu_capacity_orig < rq->rd->max_cpu_capacity ||
8145 check_cpu_capacity(rq, sd));
8149 * Group imbalance indicates (and tries to solve) the problem where balancing
8150 * groups is inadequate due to ->cpus_ptr constraints.
8152 * Imagine a situation of two groups of 4 CPUs each and 4 tasks each with a
8153 * cpumask covering 1 CPU of the first group and 3 CPUs of the second group.
8156 * { 0 1 2 3 } { 4 5 6 7 }
8159 * If we were to balance group-wise we'd place two tasks in the first group and
8160 * two tasks in the second group. Clearly this is undesired as it will overload
8161 * cpu 3 and leave one of the CPUs in the second group unused.
8163 * The current solution to this issue is detecting the skew in the first group
8164 * by noticing the lower domain failed to reach balance and had difficulty
8165 * moving tasks due to affinity constraints.
8167 * When this is so detected; this group becomes a candidate for busiest; see
8168 * update_sd_pick_busiest(). And calculate_imbalance() and
8169 * find_busiest_group() avoid some of the usual balance conditions to allow it
8170 * to create an effective group imbalance.
8172 * This is a somewhat tricky proposition since the next run might not find the
8173 * group imbalance and decide the groups need to be balanced again. A most
8174 * subtle and fragile situation.
8177 static inline int sg_imbalanced(struct sched_group *group)
8179 return group->sgc->imbalance;
8183 * group_has_capacity returns true if the group has spare capacity that could
8184 * be used by some tasks.
8185 * We consider that a group has spare capacity if the * number of task is
8186 * smaller than the number of CPUs or if the utilization is lower than the
8187 * available capacity for CFS tasks.
8188 * For the latter, we use a threshold to stabilize the state, to take into
8189 * account the variance of the tasks' load and to return true if the available
8190 * capacity in meaningful for the load balancer.
8191 * As an example, an available capacity of 1% can appear but it doesn't make
8192 * any benefit for the load balance.
8195 group_has_capacity(unsigned int imbalance_pct, struct sg_lb_stats *sgs)
8197 if (sgs->sum_nr_running < sgs->group_weight)
8200 if ((sgs->group_capacity * imbalance_pct) <
8201 (sgs->group_runnable * 100))
8204 if ((sgs->group_capacity * 100) >
8205 (sgs->group_util * imbalance_pct))
8212 * group_is_overloaded returns true if the group has more tasks than it can
8214 * group_is_overloaded is not equals to !group_has_capacity because a group
8215 * with the exact right number of tasks, has no more spare capacity but is not
8216 * overloaded so both group_has_capacity and group_is_overloaded return
8220 group_is_overloaded(unsigned int imbalance_pct, struct sg_lb_stats *sgs)
8222 if (sgs->sum_nr_running <= sgs->group_weight)
8225 if ((sgs->group_capacity * 100) <
8226 (sgs->group_util * imbalance_pct))
8229 if ((sgs->group_capacity * imbalance_pct) <
8230 (sgs->group_runnable * 100))
8237 * group_smaller_min_cpu_capacity: Returns true if sched_group sg has smaller
8238 * per-CPU capacity than sched_group ref.
8241 group_smaller_min_cpu_capacity(struct sched_group *sg, struct sched_group *ref)
8243 return fits_capacity(sg->sgc->min_capacity, ref->sgc->min_capacity);
8247 * group_smaller_max_cpu_capacity: Returns true if sched_group sg has smaller
8248 * per-CPU capacity_orig than sched_group ref.
8251 group_smaller_max_cpu_capacity(struct sched_group *sg, struct sched_group *ref)
8253 return fits_capacity(sg->sgc->max_capacity, ref->sgc->max_capacity);
8257 group_type group_classify(unsigned int imbalance_pct,
8258 struct sched_group *group,
8259 struct sg_lb_stats *sgs)
8261 if (group_is_overloaded(imbalance_pct, sgs))
8262 return group_overloaded;
8264 if (sg_imbalanced(group))
8265 return group_imbalanced;
8267 if (sgs->group_asym_packing)
8268 return group_asym_packing;
8270 if (sgs->group_misfit_task_load)
8271 return group_misfit_task;
8273 if (!group_has_capacity(imbalance_pct, sgs))
8274 return group_fully_busy;
8276 return group_has_spare;
8279 static bool update_nohz_stats(struct rq *rq, bool force)
8281 #ifdef CONFIG_NO_HZ_COMMON
8282 unsigned int cpu = rq->cpu;
8284 if (!rq->has_blocked_load)
8287 if (!cpumask_test_cpu(cpu, nohz.idle_cpus_mask))
8290 if (!force && !time_after(jiffies, rq->last_blocked_load_update_tick))
8293 update_blocked_averages(cpu);
8295 return rq->has_blocked_load;
8302 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
8303 * @env: The load balancing environment.
8304 * @group: sched_group whose statistics are to be updated.
8305 * @sgs: variable to hold the statistics for this group.
8306 * @sg_status: Holds flag indicating the status of the sched_group
8308 static inline void update_sg_lb_stats(struct lb_env *env,
8309 struct sched_group *group,
8310 struct sg_lb_stats *sgs,
8313 int i, nr_running, local_group;
8315 memset(sgs, 0, sizeof(*sgs));
8317 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_span(group));
8319 for_each_cpu_and(i, sched_group_span(group), env->cpus) {
8320 struct rq *rq = cpu_rq(i);
8322 if ((env->flags & LBF_NOHZ_STATS) && update_nohz_stats(rq, false))
8323 env->flags |= LBF_NOHZ_AGAIN;
8325 sgs->group_load += cpu_load(rq);
8326 sgs->group_util += cpu_util(i);
8327 sgs->group_runnable += cpu_runnable(rq);
8328 sgs->sum_h_nr_running += rq->cfs.h_nr_running;
8330 nr_running = rq->nr_running;
8331 sgs->sum_nr_running += nr_running;
8334 *sg_status |= SG_OVERLOAD;
8336 if (cpu_overutilized(i))
8337 *sg_status |= SG_OVERUTILIZED;
8339 #ifdef CONFIG_NUMA_BALANCING
8340 sgs->nr_numa_running += rq->nr_numa_running;
8341 sgs->nr_preferred_running += rq->nr_preferred_running;
8344 * No need to call idle_cpu() if nr_running is not 0
8346 if (!nr_running && idle_cpu(i)) {
8348 /* Idle cpu can't have misfit task */
8355 /* Check for a misfit task on the cpu */
8356 if (env->sd->flags & SD_ASYM_CPUCAPACITY &&
8357 sgs->group_misfit_task_load < rq->misfit_task_load) {
8358 sgs->group_misfit_task_load = rq->misfit_task_load;
8359 *sg_status |= SG_OVERLOAD;
8363 /* Check if dst CPU is idle and preferred to this group */
8364 if (env->sd->flags & SD_ASYM_PACKING &&
8365 env->idle != CPU_NOT_IDLE &&
8366 sgs->sum_h_nr_running &&
8367 sched_asym_prefer(env->dst_cpu, group->asym_prefer_cpu)) {
8368 sgs->group_asym_packing = 1;
8371 sgs->group_capacity = group->sgc->capacity;
8373 sgs->group_weight = group->group_weight;
8375 sgs->group_type = group_classify(env->sd->imbalance_pct, group, sgs);
8377 /* Computing avg_load makes sense only when group is overloaded */
8378 if (sgs->group_type == group_overloaded)
8379 sgs->avg_load = (sgs->group_load * SCHED_CAPACITY_SCALE) /
8380 sgs->group_capacity;
8384 * update_sd_pick_busiest - return 1 on busiest group
8385 * @env: The load balancing environment.
8386 * @sds: sched_domain statistics
8387 * @sg: sched_group candidate to be checked for being the busiest
8388 * @sgs: sched_group statistics
8390 * Determine if @sg is a busier group than the previously selected
8393 * Return: %true if @sg is a busier group than the previously selected
8394 * busiest group. %false otherwise.
8396 static bool update_sd_pick_busiest(struct lb_env *env,
8397 struct sd_lb_stats *sds,
8398 struct sched_group *sg,
8399 struct sg_lb_stats *sgs)
8401 struct sg_lb_stats *busiest = &sds->busiest_stat;
8403 /* Make sure that there is at least one task to pull */
8404 if (!sgs->sum_h_nr_running)
8408 * Don't try to pull misfit tasks we can't help.
8409 * We can use max_capacity here as reduction in capacity on some
8410 * CPUs in the group should either be possible to resolve
8411 * internally or be covered by avg_load imbalance (eventually).
8413 if (sgs->group_type == group_misfit_task &&
8414 (!group_smaller_max_cpu_capacity(sg, sds->local) ||
8415 sds->local_stat.group_type != group_has_spare))
8418 if (sgs->group_type > busiest->group_type)
8421 if (sgs->group_type < busiest->group_type)
8425 * The candidate and the current busiest group are the same type of
8426 * group. Let check which one is the busiest according to the type.
8429 switch (sgs->group_type) {
8430 case group_overloaded:
8431 /* Select the overloaded group with highest avg_load. */
8432 if (sgs->avg_load <= busiest->avg_load)
8436 case group_imbalanced:
8438 * Select the 1st imbalanced group as we don't have any way to
8439 * choose one more than another.
8443 case group_asym_packing:
8444 /* Prefer to move from lowest priority CPU's work */
8445 if (sched_asym_prefer(sg->asym_prefer_cpu, sds->busiest->asym_prefer_cpu))
8449 case group_misfit_task:
8451 * If we have more than one misfit sg go with the biggest
8454 if (sgs->group_misfit_task_load < busiest->group_misfit_task_load)
8458 case group_fully_busy:
8460 * Select the fully busy group with highest avg_load. In
8461 * theory, there is no need to pull task from such kind of
8462 * group because tasks have all compute capacity that they need
8463 * but we can still improve the overall throughput by reducing
8464 * contention when accessing shared HW resources.
8466 * XXX for now avg_load is not computed and always 0 so we
8467 * select the 1st one.
8469 if (sgs->avg_load <= busiest->avg_load)
8473 case group_has_spare:
8475 * Select not overloaded group with lowest number of idle cpus
8476 * and highest number of running tasks. We could also compare
8477 * the spare capacity which is more stable but it can end up
8478 * that the group has less spare capacity but finally more idle
8479 * CPUs which means less opportunity to pull tasks.
8481 if (sgs->idle_cpus > busiest->idle_cpus)
8483 else if ((sgs->idle_cpus == busiest->idle_cpus) &&
8484 (sgs->sum_nr_running <= busiest->sum_nr_running))
8491 * Candidate sg has no more than one task per CPU and has higher
8492 * per-CPU capacity. Migrating tasks to less capable CPUs may harm
8493 * throughput. Maximize throughput, power/energy consequences are not
8496 if ((env->sd->flags & SD_ASYM_CPUCAPACITY) &&
8497 (sgs->group_type <= group_fully_busy) &&
8498 (group_smaller_min_cpu_capacity(sds->local, sg)))
8504 #ifdef CONFIG_NUMA_BALANCING
8505 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
8507 if (sgs->sum_h_nr_running > sgs->nr_numa_running)
8509 if (sgs->sum_h_nr_running > sgs->nr_preferred_running)
8514 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
8516 if (rq->nr_running > rq->nr_numa_running)
8518 if (rq->nr_running > rq->nr_preferred_running)
8523 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
8528 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
8532 #endif /* CONFIG_NUMA_BALANCING */
8538 * task_running_on_cpu - return 1 if @p is running on @cpu.
8541 static unsigned int task_running_on_cpu(int cpu, struct task_struct *p)
8543 /* Task has no contribution or is new */
8544 if (cpu != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time))
8547 if (task_on_rq_queued(p))
8554 * idle_cpu_without - would a given CPU be idle without p ?
8555 * @cpu: the processor on which idleness is tested.
8556 * @p: task which should be ignored.
8558 * Return: 1 if the CPU would be idle. 0 otherwise.
8560 static int idle_cpu_without(int cpu, struct task_struct *p)
8562 struct rq *rq = cpu_rq(cpu);
8564 if (rq->curr != rq->idle && rq->curr != p)
8568 * rq->nr_running can't be used but an updated version without the
8569 * impact of p on cpu must be used instead. The updated nr_running
8570 * be computed and tested before calling idle_cpu_without().
8574 if (!llist_empty(&rq->wake_list))
8582 * update_sg_wakeup_stats - Update sched_group's statistics for wakeup.
8583 * @sd: The sched_domain level to look for idlest group.
8584 * @group: sched_group whose statistics are to be updated.
8585 * @sgs: variable to hold the statistics for this group.
8586 * @p: The task for which we look for the idlest group/CPU.
8588 static inline void update_sg_wakeup_stats(struct sched_domain *sd,
8589 struct sched_group *group,
8590 struct sg_lb_stats *sgs,
8591 struct task_struct *p)
8595 memset(sgs, 0, sizeof(*sgs));
8597 for_each_cpu(i, sched_group_span(group)) {
8598 struct rq *rq = cpu_rq(i);
8601 sgs->group_load += cpu_load_without(rq, p);
8602 sgs->group_util += cpu_util_without(i, p);
8603 sgs->group_runnable += cpu_runnable_without(rq, p);
8604 local = task_running_on_cpu(i, p);
8605 sgs->sum_h_nr_running += rq->cfs.h_nr_running - local;
8607 nr_running = rq->nr_running - local;
8608 sgs->sum_nr_running += nr_running;
8611 * No need to call idle_cpu_without() if nr_running is not 0
8613 if (!nr_running && idle_cpu_without(i, p))
8618 /* Check if task fits in the group */
8619 if (sd->flags & SD_ASYM_CPUCAPACITY &&
8620 !task_fits_capacity(p, group->sgc->max_capacity)) {
8621 sgs->group_misfit_task_load = 1;
8624 sgs->group_capacity = group->sgc->capacity;
8626 sgs->group_weight = group->group_weight;
8628 sgs->group_type = group_classify(sd->imbalance_pct, group, sgs);
8631 * Computing avg_load makes sense only when group is fully busy or
8634 if (sgs->group_type == group_fully_busy ||
8635 sgs->group_type == group_overloaded)
8636 sgs->avg_load = (sgs->group_load * SCHED_CAPACITY_SCALE) /
8637 sgs->group_capacity;
8640 static bool update_pick_idlest(struct sched_group *idlest,
8641 struct sg_lb_stats *idlest_sgs,
8642 struct sched_group *group,
8643 struct sg_lb_stats *sgs)
8645 if (sgs->group_type < idlest_sgs->group_type)
8648 if (sgs->group_type > idlest_sgs->group_type)
8652 * The candidate and the current idlest group are the same type of
8653 * group. Let check which one is the idlest according to the type.
8656 switch (sgs->group_type) {
8657 case group_overloaded:
8658 case group_fully_busy:
8659 /* Select the group with lowest avg_load. */
8660 if (idlest_sgs->avg_load <= sgs->avg_load)
8664 case group_imbalanced:
8665 case group_asym_packing:
8666 /* Those types are not used in the slow wakeup path */
8669 case group_misfit_task:
8670 /* Select group with the highest max capacity */
8671 if (idlest->sgc->max_capacity >= group->sgc->max_capacity)
8675 case group_has_spare:
8676 /* Select group with most idle CPUs */
8677 if (idlest_sgs->idle_cpus >= sgs->idle_cpus)
8686 * find_idlest_group() finds and returns the least busy CPU group within the
8689 * Assumes p is allowed on at least one CPU in sd.
8691 static struct sched_group *
8692 find_idlest_group(struct sched_domain *sd, struct task_struct *p,
8693 int this_cpu, int sd_flag)
8695 struct sched_group *idlest = NULL, *local = NULL, *group = sd->groups;
8696 struct sg_lb_stats local_sgs, tmp_sgs;
8697 struct sg_lb_stats *sgs;
8698 unsigned long imbalance;
8699 struct sg_lb_stats idlest_sgs = {
8700 .avg_load = UINT_MAX,
8701 .group_type = group_overloaded,
8704 imbalance = scale_load_down(NICE_0_LOAD) *
8705 (sd->imbalance_pct-100) / 100;
8710 /* Skip over this group if it has no CPUs allowed */
8711 if (!cpumask_intersects(sched_group_span(group),
8715 local_group = cpumask_test_cpu(this_cpu,
8716 sched_group_span(group));
8725 update_sg_wakeup_stats(sd, group, sgs, p);
8727 if (!local_group && update_pick_idlest(idlest, &idlest_sgs, group, sgs)) {
8732 } while (group = group->next, group != sd->groups);
8735 /* There is no idlest group to push tasks to */
8739 /* The local group has been skipped because of CPU affinity */
8744 * If the local group is idler than the selected idlest group
8745 * don't try and push the task.
8747 if (local_sgs.group_type < idlest_sgs.group_type)
8751 * If the local group is busier than the selected idlest group
8752 * try and push the task.
8754 if (local_sgs.group_type > idlest_sgs.group_type)
8757 switch (local_sgs.group_type) {
8758 case group_overloaded:
8759 case group_fully_busy:
8761 * When comparing groups across NUMA domains, it's possible for
8762 * the local domain to be very lightly loaded relative to the
8763 * remote domains but "imbalance" skews the comparison making
8764 * remote CPUs look much more favourable. When considering
8765 * cross-domain, add imbalance to the load on the remote node
8766 * and consider staying local.
8769 if ((sd->flags & SD_NUMA) &&
8770 ((idlest_sgs.avg_load + imbalance) >= local_sgs.avg_load))
8774 * If the local group is less loaded than the selected
8775 * idlest group don't try and push any tasks.
8777 if (idlest_sgs.avg_load >= (local_sgs.avg_load + imbalance))
8780 if (100 * local_sgs.avg_load <= sd->imbalance_pct * idlest_sgs.avg_load)
8784 case group_imbalanced:
8785 case group_asym_packing:
8786 /* Those type are not used in the slow wakeup path */
8789 case group_misfit_task:
8790 /* Select group with the highest max capacity */
8791 if (local->sgc->max_capacity >= idlest->sgc->max_capacity)
8795 case group_has_spare:
8796 if (sd->flags & SD_NUMA) {
8797 #ifdef CONFIG_NUMA_BALANCING
8800 * If there is spare capacity at NUMA, try to select
8801 * the preferred node
8803 if (cpu_to_node(this_cpu) == p->numa_preferred_nid)
8806 idlest_cpu = cpumask_first(sched_group_span(idlest));
8807 if (cpu_to_node(idlest_cpu) == p->numa_preferred_nid)
8811 * Otherwise, keep the task on this node to stay close
8812 * its wakeup source and improve locality. If there is
8813 * a real need of migration, periodic load balance will
8816 if (local_sgs.idle_cpus)
8821 * Select group with highest number of idle CPUs. We could also
8822 * compare the utilization which is more stable but it can end
8823 * up that the group has less spare capacity but finally more
8824 * idle CPUs which means more opportunity to run task.
8826 if (local_sgs.idle_cpus >= idlest_sgs.idle_cpus)
8835 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
8836 * @env: The load balancing environment.
8837 * @sds: variable to hold the statistics for this sched_domain.
8840 static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
8842 struct sched_domain *child = env->sd->child;
8843 struct sched_group *sg = env->sd->groups;
8844 struct sg_lb_stats *local = &sds->local_stat;
8845 struct sg_lb_stats tmp_sgs;
8848 #ifdef CONFIG_NO_HZ_COMMON
8849 if (env->idle == CPU_NEWLY_IDLE && READ_ONCE(nohz.has_blocked))
8850 env->flags |= LBF_NOHZ_STATS;
8854 struct sg_lb_stats *sgs = &tmp_sgs;
8857 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_span(sg));
8862 if (env->idle != CPU_NEWLY_IDLE ||
8863 time_after_eq(jiffies, sg->sgc->next_update))
8864 update_group_capacity(env->sd, env->dst_cpu);
8867 update_sg_lb_stats(env, sg, sgs, &sg_status);
8873 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
8875 sds->busiest_stat = *sgs;
8879 /* Now, start updating sd_lb_stats */
8880 sds->total_load += sgs->group_load;
8881 sds->total_capacity += sgs->group_capacity;
8884 } while (sg != env->sd->groups);
8886 /* Tag domain that child domain prefers tasks go to siblings first */
8887 sds->prefer_sibling = child && child->flags & SD_PREFER_SIBLING;
8889 #ifdef CONFIG_NO_HZ_COMMON
8890 if ((env->flags & LBF_NOHZ_AGAIN) &&
8891 cpumask_subset(nohz.idle_cpus_mask, sched_domain_span(env->sd))) {
8893 WRITE_ONCE(nohz.next_blocked,
8894 jiffies + msecs_to_jiffies(LOAD_AVG_PERIOD));
8898 if (env->sd->flags & SD_NUMA)
8899 env->fbq_type = fbq_classify_group(&sds->busiest_stat);
8901 if (!env->sd->parent) {
8902 struct root_domain *rd = env->dst_rq->rd;
8904 /* update overload indicator if we are at root domain */
8905 WRITE_ONCE(rd->overload, sg_status & SG_OVERLOAD);
8907 /* Update over-utilization (tipping point, U >= 0) indicator */
8908 WRITE_ONCE(rd->overutilized, sg_status & SG_OVERUTILIZED);
8909 trace_sched_overutilized_tp(rd, sg_status & SG_OVERUTILIZED);
8910 } else if (sg_status & SG_OVERUTILIZED) {
8911 struct root_domain *rd = env->dst_rq->rd;
8913 WRITE_ONCE(rd->overutilized, SG_OVERUTILIZED);
8914 trace_sched_overutilized_tp(rd, SG_OVERUTILIZED);
8918 static inline long adjust_numa_imbalance(int imbalance, int src_nr_running)
8920 unsigned int imbalance_min;
8923 * Allow a small imbalance based on a simple pair of communicating
8924 * tasks that remain local when the source domain is almost idle.
8927 if (src_nr_running <= imbalance_min)
8934 * calculate_imbalance - Calculate the amount of imbalance present within the
8935 * groups of a given sched_domain during load balance.
8936 * @env: load balance environment
8937 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
8939 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
8941 struct sg_lb_stats *local, *busiest;
8943 local = &sds->local_stat;
8944 busiest = &sds->busiest_stat;
8946 if (busiest->group_type == group_misfit_task) {
8947 /* Set imbalance to allow misfit tasks to be balanced. */
8948 env->migration_type = migrate_misfit;
8953 if (busiest->group_type == group_asym_packing) {
8955 * In case of asym capacity, we will try to migrate all load to
8956 * the preferred CPU.
8958 env->migration_type = migrate_task;
8959 env->imbalance = busiest->sum_h_nr_running;
8963 if (busiest->group_type == group_imbalanced) {
8965 * In the group_imb case we cannot rely on group-wide averages
8966 * to ensure CPU-load equilibrium, try to move any task to fix
8967 * the imbalance. The next load balance will take care of
8968 * balancing back the system.
8970 env->migration_type = migrate_task;
8976 * Try to use spare capacity of local group without overloading it or
8979 if (local->group_type == group_has_spare) {
8980 if (busiest->group_type > group_fully_busy) {
8982 * If busiest is overloaded, try to fill spare
8983 * capacity. This might end up creating spare capacity
8984 * in busiest or busiest still being overloaded but
8985 * there is no simple way to directly compute the
8986 * amount of load to migrate in order to balance the
8989 env->migration_type = migrate_util;
8990 env->imbalance = max(local->group_capacity, local->group_util) -
8994 * In some cases, the group's utilization is max or even
8995 * higher than capacity because of migrations but the
8996 * local CPU is (newly) idle. There is at least one
8997 * waiting task in this overloaded busiest group. Let's
9000 if (env->idle != CPU_NOT_IDLE && env->imbalance == 0) {
9001 env->migration_type = migrate_task;
9008 if (busiest->group_weight == 1 || sds->prefer_sibling) {
9009 unsigned int nr_diff = busiest->sum_nr_running;
9011 * When prefer sibling, evenly spread running tasks on
9014 env->migration_type = migrate_task;
9015 lsub_positive(&nr_diff, local->sum_nr_running);
9016 env->imbalance = nr_diff >> 1;
9020 * If there is no overload, we just want to even the number of
9023 env->migration_type = migrate_task;
9024 env->imbalance = max_t(long, 0, (local->idle_cpus -
9025 busiest->idle_cpus) >> 1);
9028 /* Consider allowing a small imbalance between NUMA groups */
9029 if (env->sd->flags & SD_NUMA)
9030 env->imbalance = adjust_numa_imbalance(env->imbalance,
9031 busiest->sum_nr_running);
9037 * Local is fully busy but has to take more load to relieve the
9040 if (local->group_type < group_overloaded) {
9042 * Local will become overloaded so the avg_load metrics are
9046 local->avg_load = (local->group_load * SCHED_CAPACITY_SCALE) /
9047 local->group_capacity;
9049 sds->avg_load = (sds->total_load * SCHED_CAPACITY_SCALE) /
9050 sds->total_capacity;
9054 * Both group are or will become overloaded and we're trying to get all
9055 * the CPUs to the average_load, so we don't want to push ourselves
9056 * above the average load, nor do we wish to reduce the max loaded CPU
9057 * below the average load. At the same time, we also don't want to
9058 * reduce the group load below the group capacity. Thus we look for
9059 * the minimum possible imbalance.
9061 env->migration_type = migrate_load;
9062 env->imbalance = min(
9063 (busiest->avg_load - sds->avg_load) * busiest->group_capacity,
9064 (sds->avg_load - local->avg_load) * local->group_capacity
9065 ) / SCHED_CAPACITY_SCALE;
9068 /******* find_busiest_group() helpers end here *********************/
9071 * Decision matrix according to the local and busiest group type:
9073 * busiest \ local has_spare fully_busy misfit asym imbalanced overloaded
9074 * has_spare nr_idle balanced N/A N/A balanced balanced
9075 * fully_busy nr_idle nr_idle N/A N/A balanced balanced
9076 * misfit_task force N/A N/A N/A force force
9077 * asym_packing force force N/A N/A force force
9078 * imbalanced force force N/A N/A force force
9079 * overloaded force force N/A N/A force avg_load
9081 * N/A : Not Applicable because already filtered while updating
9083 * balanced : The system is balanced for these 2 groups.
9084 * force : Calculate the imbalance as load migration is probably needed.
9085 * avg_load : Only if imbalance is significant enough.
9086 * nr_idle : dst_cpu is not busy and the number of idle CPUs is quite
9087 * different in groups.
9091 * find_busiest_group - Returns the busiest group within the sched_domain
9092 * if there is an imbalance.
9094 * Also calculates the amount of runnable load which should be moved
9095 * to restore balance.
9097 * @env: The load balancing environment.
9099 * Return: - The busiest group if imbalance exists.
9101 static struct sched_group *find_busiest_group(struct lb_env *env)
9103 struct sg_lb_stats *local, *busiest;
9104 struct sd_lb_stats sds;
9106 init_sd_lb_stats(&sds);
9109 * Compute the various statistics relevant for load balancing at
9112 update_sd_lb_stats(env, &sds);
9114 if (sched_energy_enabled()) {
9115 struct root_domain *rd = env->dst_rq->rd;
9117 if (rcu_dereference(rd->pd) && !READ_ONCE(rd->overutilized))
9121 local = &sds.local_stat;
9122 busiest = &sds.busiest_stat;
9124 /* There is no busy sibling group to pull tasks from */
9128 /* Misfit tasks should be dealt with regardless of the avg load */
9129 if (busiest->group_type == group_misfit_task)
9132 /* ASYM feature bypasses nice load balance check */
9133 if (busiest->group_type == group_asym_packing)
9137 * If the busiest group is imbalanced the below checks don't
9138 * work because they assume all things are equal, which typically
9139 * isn't true due to cpus_ptr constraints and the like.
9141 if (busiest->group_type == group_imbalanced)
9145 * If the local group is busier than the selected busiest group
9146 * don't try and pull any tasks.
9148 if (local->group_type > busiest->group_type)
9152 * When groups are overloaded, use the avg_load to ensure fairness
9155 if (local->group_type == group_overloaded) {
9157 * If the local group is more loaded than the selected
9158 * busiest group don't try to pull any tasks.
9160 if (local->avg_load >= busiest->avg_load)
9163 /* XXX broken for overlapping NUMA groups */
9164 sds.avg_load = (sds.total_load * SCHED_CAPACITY_SCALE) /
9168 * Don't pull any tasks if this group is already above the
9169 * domain average load.
9171 if (local->avg_load >= sds.avg_load)
9175 * If the busiest group is more loaded, use imbalance_pct to be
9178 if (100 * busiest->avg_load <=
9179 env->sd->imbalance_pct * local->avg_load)
9183 /* Try to move all excess tasks to child's sibling domain */
9184 if (sds.prefer_sibling && local->group_type == group_has_spare &&
9185 busiest->sum_nr_running > local->sum_nr_running + 1)
9188 if (busiest->group_type != group_overloaded) {
9189 if (env->idle == CPU_NOT_IDLE)
9191 * If the busiest group is not overloaded (and as a
9192 * result the local one too) but this CPU is already
9193 * busy, let another idle CPU try to pull task.
9197 if (busiest->group_weight > 1 &&
9198 local->idle_cpus <= (busiest->idle_cpus + 1))
9200 * If the busiest group is not overloaded
9201 * and there is no imbalance between this and busiest
9202 * group wrt idle CPUs, it is balanced. The imbalance
9203 * becomes significant if the diff is greater than 1
9204 * otherwise we might end up to just move the imbalance
9205 * on another group. Of course this applies only if
9206 * there is more than 1 CPU per group.
9210 if (busiest->sum_h_nr_running == 1)
9212 * busiest doesn't have any tasks waiting to run
9218 /* Looks like there is an imbalance. Compute it */
9219 calculate_imbalance(env, &sds);
9220 return env->imbalance ? sds.busiest : NULL;
9228 * find_busiest_queue - find the busiest runqueue among the CPUs in the group.
9230 static struct rq *find_busiest_queue(struct lb_env *env,
9231 struct sched_group *group)
9233 struct rq *busiest = NULL, *rq;
9234 unsigned long busiest_util = 0, busiest_load = 0, busiest_capacity = 1;
9235 unsigned int busiest_nr = 0;
9238 for_each_cpu_and(i, sched_group_span(group), env->cpus) {
9239 unsigned long capacity, load, util;
9240 unsigned int nr_running;
9244 rt = fbq_classify_rq(rq);
9247 * We classify groups/runqueues into three groups:
9248 * - regular: there are !numa tasks
9249 * - remote: there are numa tasks that run on the 'wrong' node
9250 * - all: there is no distinction
9252 * In order to avoid migrating ideally placed numa tasks,
9253 * ignore those when there's better options.
9255 * If we ignore the actual busiest queue to migrate another
9256 * task, the next balance pass can still reduce the busiest
9257 * queue by moving tasks around inside the node.
9259 * If we cannot move enough load due to this classification
9260 * the next pass will adjust the group classification and
9261 * allow migration of more tasks.
9263 * Both cases only affect the total convergence complexity.
9265 if (rt > env->fbq_type)
9268 capacity = capacity_of(i);
9269 nr_running = rq->cfs.h_nr_running;
9272 * For ASYM_CPUCAPACITY domains, don't pick a CPU that could
9273 * eventually lead to active_balancing high->low capacity.
9274 * Higher per-CPU capacity is considered better than balancing
9277 if (env->sd->flags & SD_ASYM_CPUCAPACITY &&
9278 capacity_of(env->dst_cpu) < capacity &&
9282 switch (env->migration_type) {
9285 * When comparing with load imbalance, use cpu_load()
9286 * which is not scaled with the CPU capacity.
9288 load = cpu_load(rq);
9290 if (nr_running == 1 && load > env->imbalance &&
9291 !check_cpu_capacity(rq, env->sd))
9295 * For the load comparisons with the other CPUs,
9296 * consider the cpu_load() scaled with the CPU
9297 * capacity, so that the load can be moved away
9298 * from the CPU that is potentially running at a
9301 * Thus we're looking for max(load_i / capacity_i),
9302 * crosswise multiplication to rid ourselves of the
9303 * division works out to:
9304 * load_i * capacity_j > load_j * capacity_i;
9305 * where j is our previous maximum.
9307 if (load * busiest_capacity > busiest_load * capacity) {
9308 busiest_load = load;
9309 busiest_capacity = capacity;
9315 util = cpu_util(cpu_of(rq));
9318 * Don't try to pull utilization from a CPU with one
9319 * running task. Whatever its utilization, we will fail
9322 if (nr_running <= 1)
9325 if (busiest_util < util) {
9326 busiest_util = util;
9332 if (busiest_nr < nr_running) {
9333 busiest_nr = nr_running;
9338 case migrate_misfit:
9340 * For ASYM_CPUCAPACITY domains with misfit tasks we
9341 * simply seek the "biggest" misfit task.
9343 if (rq->misfit_task_load > busiest_load) {
9344 busiest_load = rq->misfit_task_load;
9357 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
9358 * so long as it is large enough.
9360 #define MAX_PINNED_INTERVAL 512
9363 asym_active_balance(struct lb_env *env)
9366 * ASYM_PACKING needs to force migrate tasks from busy but
9367 * lower priority CPUs in order to pack all tasks in the
9368 * highest priority CPUs.
9370 return env->idle != CPU_NOT_IDLE && (env->sd->flags & SD_ASYM_PACKING) &&
9371 sched_asym_prefer(env->dst_cpu, env->src_cpu);
9375 voluntary_active_balance(struct lb_env *env)
9377 struct sched_domain *sd = env->sd;
9379 if (asym_active_balance(env))
9383 * The dst_cpu is idle and the src_cpu CPU has only 1 CFS task.
9384 * It's worth migrating the task if the src_cpu's capacity is reduced
9385 * because of other sched_class or IRQs if more capacity stays
9386 * available on dst_cpu.
9388 if ((env->idle != CPU_NOT_IDLE) &&
9389 (env->src_rq->cfs.h_nr_running == 1)) {
9390 if ((check_cpu_capacity(env->src_rq, sd)) &&
9391 (capacity_of(env->src_cpu)*sd->imbalance_pct < capacity_of(env->dst_cpu)*100))
9395 if (env->migration_type == migrate_misfit)
9401 static int need_active_balance(struct lb_env *env)
9403 struct sched_domain *sd = env->sd;
9405 if (voluntary_active_balance(env))
9408 return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
9411 static int active_load_balance_cpu_stop(void *data);
9413 static int should_we_balance(struct lb_env *env)
9415 struct sched_group *sg = env->sd->groups;
9416 int cpu, balance_cpu = -1;
9419 * Ensure the balancing environment is consistent; can happen
9420 * when the softirq triggers 'during' hotplug.
9422 if (!cpumask_test_cpu(env->dst_cpu, env->cpus))
9426 * In the newly idle case, we will allow all the CPUs
9427 * to do the newly idle load balance.
9429 if (env->idle == CPU_NEWLY_IDLE)
9432 /* Try to find first idle CPU */
9433 for_each_cpu_and(cpu, group_balance_mask(sg), env->cpus) {
9441 if (balance_cpu == -1)
9442 balance_cpu = group_balance_cpu(sg);
9445 * First idle CPU or the first CPU(busiest) in this sched group
9446 * is eligible for doing load balancing at this and above domains.
9448 return balance_cpu == env->dst_cpu;
9452 * Check this_cpu to ensure it is balanced within domain. Attempt to move
9453 * tasks if there is an imbalance.
9455 static int load_balance(int this_cpu, struct rq *this_rq,
9456 struct sched_domain *sd, enum cpu_idle_type idle,
9457 int *continue_balancing)
9459 int ld_moved, cur_ld_moved, active_balance = 0;
9460 struct sched_domain *sd_parent = sd->parent;
9461 struct sched_group *group;
9464 struct cpumask *cpus = this_cpu_cpumask_var_ptr(load_balance_mask);
9466 struct lb_env env = {
9468 .dst_cpu = this_cpu,
9470 .dst_grpmask = sched_group_span(sd->groups),
9472 .loop_break = sched_nr_migrate_break,
9475 .tasks = LIST_HEAD_INIT(env.tasks),
9478 cpumask_and(cpus, sched_domain_span(sd), cpu_active_mask);
9480 schedstat_inc(sd->lb_count[idle]);
9483 if (!should_we_balance(&env)) {
9484 *continue_balancing = 0;
9488 group = find_busiest_group(&env);
9490 schedstat_inc(sd->lb_nobusyg[idle]);
9494 busiest = find_busiest_queue(&env, group);
9496 schedstat_inc(sd->lb_nobusyq[idle]);
9500 BUG_ON(busiest == env.dst_rq);
9502 schedstat_add(sd->lb_imbalance[idle], env.imbalance);
9504 env.src_cpu = busiest->cpu;
9505 env.src_rq = busiest;
9508 if (busiest->nr_running > 1) {
9510 * Attempt to move tasks. If find_busiest_group has found
9511 * an imbalance but busiest->nr_running <= 1, the group is
9512 * still unbalanced. ld_moved simply stays zero, so it is
9513 * correctly treated as an imbalance.
9515 env.flags |= LBF_ALL_PINNED;
9516 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
9519 rq_lock_irqsave(busiest, &rf);
9520 update_rq_clock(busiest);
9523 * cur_ld_moved - load moved in current iteration
9524 * ld_moved - cumulative load moved across iterations
9526 cur_ld_moved = detach_tasks(&env);
9529 * We've detached some tasks from busiest_rq. Every
9530 * task is masked "TASK_ON_RQ_MIGRATING", so we can safely
9531 * unlock busiest->lock, and we are able to be sure
9532 * that nobody can manipulate the tasks in parallel.
9533 * See task_rq_lock() family for the details.
9536 rq_unlock(busiest, &rf);
9540 ld_moved += cur_ld_moved;
9543 local_irq_restore(rf.flags);
9545 if (env.flags & LBF_NEED_BREAK) {
9546 env.flags &= ~LBF_NEED_BREAK;
9551 * Revisit (affine) tasks on src_cpu that couldn't be moved to
9552 * us and move them to an alternate dst_cpu in our sched_group
9553 * where they can run. The upper limit on how many times we
9554 * iterate on same src_cpu is dependent on number of CPUs in our
9557 * This changes load balance semantics a bit on who can move
9558 * load to a given_cpu. In addition to the given_cpu itself
9559 * (or a ilb_cpu acting on its behalf where given_cpu is
9560 * nohz-idle), we now have balance_cpu in a position to move
9561 * load to given_cpu. In rare situations, this may cause
9562 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
9563 * _independently_ and at _same_ time to move some load to
9564 * given_cpu) causing exceess load to be moved to given_cpu.
9565 * This however should not happen so much in practice and
9566 * moreover subsequent load balance cycles should correct the
9567 * excess load moved.
9569 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
9571 /* Prevent to re-select dst_cpu via env's CPUs */
9572 __cpumask_clear_cpu(env.dst_cpu, env.cpus);
9574 env.dst_rq = cpu_rq(env.new_dst_cpu);
9575 env.dst_cpu = env.new_dst_cpu;
9576 env.flags &= ~LBF_DST_PINNED;
9578 env.loop_break = sched_nr_migrate_break;
9581 * Go back to "more_balance" rather than "redo" since we
9582 * need to continue with same src_cpu.
9588 * We failed to reach balance because of affinity.
9591 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
9593 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0)
9594 *group_imbalance = 1;
9597 /* All tasks on this runqueue were pinned by CPU affinity */
9598 if (unlikely(env.flags & LBF_ALL_PINNED)) {
9599 __cpumask_clear_cpu(cpu_of(busiest), cpus);
9601 * Attempting to continue load balancing at the current
9602 * sched_domain level only makes sense if there are
9603 * active CPUs remaining as possible busiest CPUs to
9604 * pull load from which are not contained within the
9605 * destination group that is receiving any migrated
9608 if (!cpumask_subset(cpus, env.dst_grpmask)) {
9610 env.loop_break = sched_nr_migrate_break;
9613 goto out_all_pinned;
9618 schedstat_inc(sd->lb_failed[idle]);
9620 * Increment the failure counter only on periodic balance.
9621 * We do not want newidle balance, which can be very
9622 * frequent, pollute the failure counter causing
9623 * excessive cache_hot migrations and active balances.
9625 if (idle != CPU_NEWLY_IDLE)
9626 sd->nr_balance_failed++;
9628 if (need_active_balance(&env)) {
9629 unsigned long flags;
9631 raw_spin_lock_irqsave(&busiest->lock, flags);
9634 * Don't kick the active_load_balance_cpu_stop,
9635 * if the curr task on busiest CPU can't be
9636 * moved to this_cpu:
9638 if (!cpumask_test_cpu(this_cpu, busiest->curr->cpus_ptr)) {
9639 raw_spin_unlock_irqrestore(&busiest->lock,
9641 env.flags |= LBF_ALL_PINNED;
9642 goto out_one_pinned;
9646 * ->active_balance synchronizes accesses to
9647 * ->active_balance_work. Once set, it's cleared
9648 * only after active load balance is finished.
9650 if (!busiest->active_balance) {
9651 busiest->active_balance = 1;
9652 busiest->push_cpu = this_cpu;
9655 raw_spin_unlock_irqrestore(&busiest->lock, flags);
9657 if (active_balance) {
9658 stop_one_cpu_nowait(cpu_of(busiest),
9659 active_load_balance_cpu_stop, busiest,
9660 &busiest->active_balance_work);
9663 /* We've kicked active balancing, force task migration. */
9664 sd->nr_balance_failed = sd->cache_nice_tries+1;
9667 sd->nr_balance_failed = 0;
9669 if (likely(!active_balance) || voluntary_active_balance(&env)) {
9670 /* We were unbalanced, so reset the balancing interval */
9671 sd->balance_interval = sd->min_interval;
9674 * If we've begun active balancing, start to back off. This
9675 * case may not be covered by the all_pinned logic if there
9676 * is only 1 task on the busy runqueue (because we don't call
9679 if (sd->balance_interval < sd->max_interval)
9680 sd->balance_interval *= 2;
9687 * We reach balance although we may have faced some affinity
9688 * constraints. Clear the imbalance flag only if other tasks got
9689 * a chance to move and fix the imbalance.
9691 if (sd_parent && !(env.flags & LBF_ALL_PINNED)) {
9692 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
9694 if (*group_imbalance)
9695 *group_imbalance = 0;
9700 * We reach balance because all tasks are pinned at this level so
9701 * we can't migrate them. Let the imbalance flag set so parent level
9702 * can try to migrate them.
9704 schedstat_inc(sd->lb_balanced[idle]);
9706 sd->nr_balance_failed = 0;
9712 * newidle_balance() disregards balance intervals, so we could
9713 * repeatedly reach this code, which would lead to balance_interval
9714 * skyrocketting in a short amount of time. Skip the balance_interval
9715 * increase logic to avoid that.
9717 if (env.idle == CPU_NEWLY_IDLE)
9720 /* tune up the balancing interval */
9721 if ((env.flags & LBF_ALL_PINNED &&
9722 sd->balance_interval < MAX_PINNED_INTERVAL) ||
9723 sd->balance_interval < sd->max_interval)
9724 sd->balance_interval *= 2;
9729 static inline unsigned long
9730 get_sd_balance_interval(struct sched_domain *sd, int cpu_busy)
9732 unsigned long interval = sd->balance_interval;
9735 interval *= sd->busy_factor;
9737 /* scale ms to jiffies */
9738 interval = msecs_to_jiffies(interval);
9739 interval = clamp(interval, 1UL, max_load_balance_interval);
9745 update_next_balance(struct sched_domain *sd, unsigned long *next_balance)
9747 unsigned long interval, next;
9749 /* used by idle balance, so cpu_busy = 0 */
9750 interval = get_sd_balance_interval(sd, 0);
9751 next = sd->last_balance + interval;
9753 if (time_after(*next_balance, next))
9754 *next_balance = next;
9758 * active_load_balance_cpu_stop is run by the CPU stopper. It pushes
9759 * running tasks off the busiest CPU onto idle CPUs. It requires at
9760 * least 1 task to be running on each physical CPU where possible, and
9761 * avoids physical / logical imbalances.
9763 static int active_load_balance_cpu_stop(void *data)
9765 struct rq *busiest_rq = data;
9766 int busiest_cpu = cpu_of(busiest_rq);
9767 int target_cpu = busiest_rq->push_cpu;
9768 struct rq *target_rq = cpu_rq(target_cpu);
9769 struct sched_domain *sd;
9770 struct task_struct *p = NULL;
9773 rq_lock_irq(busiest_rq, &rf);
9775 * Between queueing the stop-work and running it is a hole in which
9776 * CPUs can become inactive. We should not move tasks from or to
9779 if (!cpu_active(busiest_cpu) || !cpu_active(target_cpu))
9782 /* Make sure the requested CPU hasn't gone down in the meantime: */
9783 if (unlikely(busiest_cpu != smp_processor_id() ||
9784 !busiest_rq->active_balance))
9787 /* Is there any task to move? */
9788 if (busiest_rq->nr_running <= 1)
9792 * This condition is "impossible", if it occurs
9793 * we need to fix it. Originally reported by
9794 * Bjorn Helgaas on a 128-CPU setup.
9796 BUG_ON(busiest_rq == target_rq);
9798 /* Search for an sd spanning us and the target CPU. */
9800 for_each_domain(target_cpu, sd) {
9801 if ((sd->flags & SD_LOAD_BALANCE) &&
9802 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
9807 struct lb_env env = {
9809 .dst_cpu = target_cpu,
9810 .dst_rq = target_rq,
9811 .src_cpu = busiest_rq->cpu,
9812 .src_rq = busiest_rq,
9815 * can_migrate_task() doesn't need to compute new_dst_cpu
9816 * for active balancing. Since we have CPU_IDLE, but no
9817 * @dst_grpmask we need to make that test go away with lying
9820 .flags = LBF_DST_PINNED,
9823 schedstat_inc(sd->alb_count);
9824 update_rq_clock(busiest_rq);
9826 p = detach_one_task(&env);
9828 schedstat_inc(sd->alb_pushed);
9829 /* Active balancing done, reset the failure counter. */
9830 sd->nr_balance_failed = 0;
9832 schedstat_inc(sd->alb_failed);
9837 busiest_rq->active_balance = 0;
9838 rq_unlock(busiest_rq, &rf);
9841 attach_one_task(target_rq, p);
9848 static DEFINE_SPINLOCK(balancing);
9851 * Scale the max load_balance interval with the number of CPUs in the system.
9852 * This trades load-balance latency on larger machines for less cross talk.
9854 void update_max_interval(void)
9856 max_load_balance_interval = HZ*num_online_cpus()/10;
9860 * It checks each scheduling domain to see if it is due to be balanced,
9861 * and initiates a balancing operation if so.
9863 * Balancing parameters are set up in init_sched_domains.
9865 static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle)
9867 int continue_balancing = 1;
9869 int busy = idle != CPU_IDLE && !sched_idle_cpu(cpu);
9870 unsigned long interval;
9871 struct sched_domain *sd;
9872 /* Earliest time when we have to do rebalance again */
9873 unsigned long next_balance = jiffies + 60*HZ;
9874 int update_next_balance = 0;
9875 int need_serialize, need_decay = 0;
9879 for_each_domain(cpu, sd) {
9881 * Decay the newidle max times here because this is a regular
9882 * visit to all the domains. Decay ~1% per second.
9884 if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
9885 sd->max_newidle_lb_cost =
9886 (sd->max_newidle_lb_cost * 253) / 256;
9887 sd->next_decay_max_lb_cost = jiffies + HZ;
9890 max_cost += sd->max_newidle_lb_cost;
9892 if (!(sd->flags & SD_LOAD_BALANCE))
9896 * Stop the load balance at this level. There is another
9897 * CPU in our sched group which is doing load balancing more
9900 if (!continue_balancing) {
9906 interval = get_sd_balance_interval(sd, busy);
9908 need_serialize = sd->flags & SD_SERIALIZE;
9909 if (need_serialize) {
9910 if (!spin_trylock(&balancing))
9914 if (time_after_eq(jiffies, sd->last_balance + interval)) {
9915 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
9917 * The LBF_DST_PINNED logic could have changed
9918 * env->dst_cpu, so we can't know our idle
9919 * state even if we migrated tasks. Update it.
9921 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
9922 busy = idle != CPU_IDLE && !sched_idle_cpu(cpu);
9924 sd->last_balance = jiffies;
9925 interval = get_sd_balance_interval(sd, busy);
9928 spin_unlock(&balancing);
9930 if (time_after(next_balance, sd->last_balance + interval)) {
9931 next_balance = sd->last_balance + interval;
9932 update_next_balance = 1;
9937 * Ensure the rq-wide value also decays but keep it at a
9938 * reasonable floor to avoid funnies with rq->avg_idle.
9940 rq->max_idle_balance_cost =
9941 max((u64)sysctl_sched_migration_cost, max_cost);
9946 * next_balance will be updated only when there is a need.
9947 * When the cpu is attached to null domain for ex, it will not be
9950 if (likely(update_next_balance)) {
9951 rq->next_balance = next_balance;
9953 #ifdef CONFIG_NO_HZ_COMMON
9955 * If this CPU has been elected to perform the nohz idle
9956 * balance. Other idle CPUs have already rebalanced with
9957 * nohz_idle_balance() and nohz.next_balance has been
9958 * updated accordingly. This CPU is now running the idle load
9959 * balance for itself and we need to update the
9960 * nohz.next_balance accordingly.
9962 if ((idle == CPU_IDLE) && time_after(nohz.next_balance, rq->next_balance))
9963 nohz.next_balance = rq->next_balance;
9968 static inline int on_null_domain(struct rq *rq)
9970 return unlikely(!rcu_dereference_sched(rq->sd));
9973 #ifdef CONFIG_NO_HZ_COMMON
9975 * idle load balancing details
9976 * - When one of the busy CPUs notice that there may be an idle rebalancing
9977 * needed, they will kick the idle load balancer, which then does idle
9978 * load balancing for all the idle CPUs.
9979 * - HK_FLAG_MISC CPUs are used for this task, because HK_FLAG_SCHED not set
9983 static inline int find_new_ilb(void)
9987 for_each_cpu_and(ilb, nohz.idle_cpus_mask,
9988 housekeeping_cpumask(HK_FLAG_MISC)) {
9997 * Kick a CPU to do the nohz balancing, if it is time for it. We pick any
9998 * idle CPU in the HK_FLAG_MISC housekeeping set (if there is one).
10000 static void kick_ilb(unsigned int flags)
10004 nohz.next_balance++;
10006 ilb_cpu = find_new_ilb();
10008 if (ilb_cpu >= nr_cpu_ids)
10011 flags = atomic_fetch_or(flags, nohz_flags(ilb_cpu));
10012 if (flags & NOHZ_KICK_MASK)
10016 * Use smp_send_reschedule() instead of resched_cpu().
10017 * This way we generate a sched IPI on the target CPU which
10018 * is idle. And the softirq performing nohz idle load balance
10019 * will be run before returning from the IPI.
10021 smp_send_reschedule(ilb_cpu);
10025 * Current decision point for kicking the idle load balancer in the presence
10026 * of idle CPUs in the system.
10028 static void nohz_balancer_kick(struct rq *rq)
10030 unsigned long now = jiffies;
10031 struct sched_domain_shared *sds;
10032 struct sched_domain *sd;
10033 int nr_busy, i, cpu = rq->cpu;
10034 unsigned int flags = 0;
10036 if (unlikely(rq->idle_balance))
10040 * We may be recently in ticked or tickless idle mode. At the first
10041 * busy tick after returning from idle, we will update the busy stats.
10043 nohz_balance_exit_idle(rq);
10046 * None are in tickless mode and hence no need for NOHZ idle load
10049 if (likely(!atomic_read(&nohz.nr_cpus)))
10052 if (READ_ONCE(nohz.has_blocked) &&
10053 time_after(now, READ_ONCE(nohz.next_blocked)))
10054 flags = NOHZ_STATS_KICK;
10056 if (time_before(now, nohz.next_balance))
10059 if (rq->nr_running >= 2) {
10060 flags = NOHZ_KICK_MASK;
10066 sd = rcu_dereference(rq->sd);
10069 * If there's a CFS task and the current CPU has reduced
10070 * capacity; kick the ILB to see if there's a better CPU to run
10073 if (rq->cfs.h_nr_running >= 1 && check_cpu_capacity(rq, sd)) {
10074 flags = NOHZ_KICK_MASK;
10079 sd = rcu_dereference(per_cpu(sd_asym_packing, cpu));
10082 * When ASYM_PACKING; see if there's a more preferred CPU
10083 * currently idle; in which case, kick the ILB to move tasks
10086 for_each_cpu_and(i, sched_domain_span(sd), nohz.idle_cpus_mask) {
10087 if (sched_asym_prefer(i, cpu)) {
10088 flags = NOHZ_KICK_MASK;
10094 sd = rcu_dereference(per_cpu(sd_asym_cpucapacity, cpu));
10097 * When ASYM_CPUCAPACITY; see if there's a higher capacity CPU
10098 * to run the misfit task on.
10100 if (check_misfit_status(rq, sd)) {
10101 flags = NOHZ_KICK_MASK;
10106 * For asymmetric systems, we do not want to nicely balance
10107 * cache use, instead we want to embrace asymmetry and only
10108 * ensure tasks have enough CPU capacity.
10110 * Skip the LLC logic because it's not relevant in that case.
10115 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
10118 * If there is an imbalance between LLC domains (IOW we could
10119 * increase the overall cache use), we need some less-loaded LLC
10120 * domain to pull some load. Likewise, we may need to spread
10121 * load within the current LLC domain (e.g. packed SMT cores but
10122 * other CPUs are idle). We can't really know from here how busy
10123 * the others are - so just get a nohz balance going if it looks
10124 * like this LLC domain has tasks we could move.
10126 nr_busy = atomic_read(&sds->nr_busy_cpus);
10128 flags = NOHZ_KICK_MASK;
10139 static void set_cpu_sd_state_busy(int cpu)
10141 struct sched_domain *sd;
10144 sd = rcu_dereference(per_cpu(sd_llc, cpu));
10146 if (!sd || !sd->nohz_idle)
10150 atomic_inc(&sd->shared->nr_busy_cpus);
10155 void nohz_balance_exit_idle(struct rq *rq)
10157 SCHED_WARN_ON(rq != this_rq());
10159 if (likely(!rq->nohz_tick_stopped))
10162 rq->nohz_tick_stopped = 0;
10163 cpumask_clear_cpu(rq->cpu, nohz.idle_cpus_mask);
10164 atomic_dec(&nohz.nr_cpus);
10166 set_cpu_sd_state_busy(rq->cpu);
10169 static void set_cpu_sd_state_idle(int cpu)
10171 struct sched_domain *sd;
10174 sd = rcu_dereference(per_cpu(sd_llc, cpu));
10176 if (!sd || sd->nohz_idle)
10180 atomic_dec(&sd->shared->nr_busy_cpus);
10186 * This routine will record that the CPU is going idle with tick stopped.
10187 * This info will be used in performing idle load balancing in the future.
10189 void nohz_balance_enter_idle(int cpu)
10191 struct rq *rq = cpu_rq(cpu);
10193 SCHED_WARN_ON(cpu != smp_processor_id());
10195 /* If this CPU is going down, then nothing needs to be done: */
10196 if (!cpu_active(cpu))
10199 /* Spare idle load balancing on CPUs that don't want to be disturbed: */
10200 if (!housekeeping_cpu(cpu, HK_FLAG_SCHED))
10204 * Can be set safely without rq->lock held
10205 * If a clear happens, it will have evaluated last additions because
10206 * rq->lock is held during the check and the clear
10208 rq->has_blocked_load = 1;
10211 * The tick is still stopped but load could have been added in the
10212 * meantime. We set the nohz.has_blocked flag to trig a check of the
10213 * *_avg. The CPU is already part of nohz.idle_cpus_mask so the clear
10214 * of nohz.has_blocked can only happen after checking the new load
10216 if (rq->nohz_tick_stopped)
10219 /* If we're a completely isolated CPU, we don't play: */
10220 if (on_null_domain(rq))
10223 rq->nohz_tick_stopped = 1;
10225 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
10226 atomic_inc(&nohz.nr_cpus);
10229 * Ensures that if nohz_idle_balance() fails to observe our
10230 * @idle_cpus_mask store, it must observe the @has_blocked
10233 smp_mb__after_atomic();
10235 set_cpu_sd_state_idle(cpu);
10239 * Each time a cpu enter idle, we assume that it has blocked load and
10240 * enable the periodic update of the load of idle cpus
10242 WRITE_ONCE(nohz.has_blocked, 1);
10246 * Internal function that runs load balance for all idle cpus. The load balance
10247 * can be a simple update of blocked load or a complete load balance with
10248 * tasks movement depending of flags.
10249 * The function returns false if the loop has stopped before running
10250 * through all idle CPUs.
10252 static bool _nohz_idle_balance(struct rq *this_rq, unsigned int flags,
10253 enum cpu_idle_type idle)
10255 /* Earliest time when we have to do rebalance again */
10256 unsigned long now = jiffies;
10257 unsigned long next_balance = now + 60*HZ;
10258 bool has_blocked_load = false;
10259 int update_next_balance = 0;
10260 int this_cpu = this_rq->cpu;
10265 SCHED_WARN_ON((flags & NOHZ_KICK_MASK) == NOHZ_BALANCE_KICK);
10268 * We assume there will be no idle load after this update and clear
10269 * the has_blocked flag. If a cpu enters idle in the mean time, it will
10270 * set the has_blocked flag and trig another update of idle load.
10271 * Because a cpu that becomes idle, is added to idle_cpus_mask before
10272 * setting the flag, we are sure to not clear the state and not
10273 * check the load of an idle cpu.
10275 WRITE_ONCE(nohz.has_blocked, 0);
10278 * Ensures that if we miss the CPU, we must see the has_blocked
10279 * store from nohz_balance_enter_idle().
10283 for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
10284 if (balance_cpu == this_cpu || !idle_cpu(balance_cpu))
10288 * If this CPU gets work to do, stop the load balancing
10289 * work being done for other CPUs. Next load
10290 * balancing owner will pick it up.
10292 if (need_resched()) {
10293 has_blocked_load = true;
10297 rq = cpu_rq(balance_cpu);
10299 has_blocked_load |= update_nohz_stats(rq, true);
10302 * If time for next balance is due,
10305 if (time_after_eq(jiffies, rq->next_balance)) {
10306 struct rq_flags rf;
10308 rq_lock_irqsave(rq, &rf);
10309 update_rq_clock(rq);
10310 rq_unlock_irqrestore(rq, &rf);
10312 if (flags & NOHZ_BALANCE_KICK)
10313 rebalance_domains(rq, CPU_IDLE);
10316 if (time_after(next_balance, rq->next_balance)) {
10317 next_balance = rq->next_balance;
10318 update_next_balance = 1;
10322 /* Newly idle CPU doesn't need an update */
10323 if (idle != CPU_NEWLY_IDLE) {
10324 update_blocked_averages(this_cpu);
10325 has_blocked_load |= this_rq->has_blocked_load;
10328 if (flags & NOHZ_BALANCE_KICK)
10329 rebalance_domains(this_rq, CPU_IDLE);
10331 WRITE_ONCE(nohz.next_blocked,
10332 now + msecs_to_jiffies(LOAD_AVG_PERIOD));
10334 /* The full idle balance loop has been done */
10338 /* There is still blocked load, enable periodic update */
10339 if (has_blocked_load)
10340 WRITE_ONCE(nohz.has_blocked, 1);
10343 * next_balance will be updated only when there is a need.
10344 * When the CPU is attached to null domain for ex, it will not be
10347 if (likely(update_next_balance))
10348 nohz.next_balance = next_balance;
10354 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
10355 * rebalancing for all the cpus for whom scheduler ticks are stopped.
10357 static bool nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
10359 int this_cpu = this_rq->cpu;
10360 unsigned int flags;
10362 if (!(atomic_read(nohz_flags(this_cpu)) & NOHZ_KICK_MASK))
10365 if (idle != CPU_IDLE) {
10366 atomic_andnot(NOHZ_KICK_MASK, nohz_flags(this_cpu));
10370 /* could be _relaxed() */
10371 flags = atomic_fetch_andnot(NOHZ_KICK_MASK, nohz_flags(this_cpu));
10372 if (!(flags & NOHZ_KICK_MASK))
10375 _nohz_idle_balance(this_rq, flags, idle);
10380 static void nohz_newidle_balance(struct rq *this_rq)
10382 int this_cpu = this_rq->cpu;
10385 * This CPU doesn't want to be disturbed by scheduler
10388 if (!housekeeping_cpu(this_cpu, HK_FLAG_SCHED))
10391 /* Will wake up very soon. No time for doing anything else*/
10392 if (this_rq->avg_idle < sysctl_sched_migration_cost)
10395 /* Don't need to update blocked load of idle CPUs*/
10396 if (!READ_ONCE(nohz.has_blocked) ||
10397 time_before(jiffies, READ_ONCE(nohz.next_blocked)))
10400 raw_spin_unlock(&this_rq->lock);
10402 * This CPU is going to be idle and blocked load of idle CPUs
10403 * need to be updated. Run the ilb locally as it is a good
10404 * candidate for ilb instead of waking up another idle CPU.
10405 * Kick an normal ilb if we failed to do the update.
10407 if (!_nohz_idle_balance(this_rq, NOHZ_STATS_KICK, CPU_NEWLY_IDLE))
10408 kick_ilb(NOHZ_STATS_KICK);
10409 raw_spin_lock(&this_rq->lock);
10412 #else /* !CONFIG_NO_HZ_COMMON */
10413 static inline void nohz_balancer_kick(struct rq *rq) { }
10415 static inline bool nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
10420 static inline void nohz_newidle_balance(struct rq *this_rq) { }
10421 #endif /* CONFIG_NO_HZ_COMMON */
10424 * idle_balance is called by schedule() if this_cpu is about to become
10425 * idle. Attempts to pull tasks from other CPUs.
10428 * < 0 - we released the lock and there are !fair tasks present
10429 * 0 - failed, no new tasks
10430 * > 0 - success, new (fair) tasks present
10432 int newidle_balance(struct rq *this_rq, struct rq_flags *rf)
10434 unsigned long next_balance = jiffies + HZ;
10435 int this_cpu = this_rq->cpu;
10436 struct sched_domain *sd;
10437 int pulled_task = 0;
10440 update_misfit_status(NULL, this_rq);
10442 * We must set idle_stamp _before_ calling idle_balance(), such that we
10443 * measure the duration of idle_balance() as idle time.
10445 this_rq->idle_stamp = rq_clock(this_rq);
10448 * Do not pull tasks towards !active CPUs...
10450 if (!cpu_active(this_cpu))
10454 * This is OK, because current is on_cpu, which avoids it being picked
10455 * for load-balance and preemption/IRQs are still disabled avoiding
10456 * further scheduler activity on it and we're being very careful to
10457 * re-start the picking loop.
10459 rq_unpin_lock(this_rq, rf);
10461 if (this_rq->avg_idle < sysctl_sched_migration_cost ||
10462 !READ_ONCE(this_rq->rd->overload)) {
10465 sd = rcu_dereference_check_sched_domain(this_rq->sd);
10467 update_next_balance(sd, &next_balance);
10470 nohz_newidle_balance(this_rq);
10475 raw_spin_unlock(&this_rq->lock);
10477 update_blocked_averages(this_cpu);
10479 for_each_domain(this_cpu, sd) {
10480 int continue_balancing = 1;
10481 u64 t0, domain_cost;
10483 if (!(sd->flags & SD_LOAD_BALANCE))
10486 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost) {
10487 update_next_balance(sd, &next_balance);
10491 if (sd->flags & SD_BALANCE_NEWIDLE) {
10492 t0 = sched_clock_cpu(this_cpu);
10494 pulled_task = load_balance(this_cpu, this_rq,
10495 sd, CPU_NEWLY_IDLE,
10496 &continue_balancing);
10498 domain_cost = sched_clock_cpu(this_cpu) - t0;
10499 if (domain_cost > sd->max_newidle_lb_cost)
10500 sd->max_newidle_lb_cost = domain_cost;
10502 curr_cost += domain_cost;
10505 update_next_balance(sd, &next_balance);
10508 * Stop searching for tasks to pull if there are
10509 * now runnable tasks on this rq.
10511 if (pulled_task || this_rq->nr_running > 0)
10516 raw_spin_lock(&this_rq->lock);
10518 if (curr_cost > this_rq->max_idle_balance_cost)
10519 this_rq->max_idle_balance_cost = curr_cost;
10523 * While browsing the domains, we released the rq lock, a task could
10524 * have been enqueued in the meantime. Since we're not going idle,
10525 * pretend we pulled a task.
10527 if (this_rq->cfs.h_nr_running && !pulled_task)
10530 /* Move the next balance forward */
10531 if (time_after(this_rq->next_balance, next_balance))
10532 this_rq->next_balance = next_balance;
10534 /* Is there a task of a high priority class? */
10535 if (this_rq->nr_running != this_rq->cfs.h_nr_running)
10539 this_rq->idle_stamp = 0;
10541 rq_repin_lock(this_rq, rf);
10543 return pulled_task;
10547 * run_rebalance_domains is triggered when needed from the scheduler tick.
10548 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
10550 static __latent_entropy void run_rebalance_domains(struct softirq_action *h)
10552 struct rq *this_rq = this_rq();
10553 enum cpu_idle_type idle = this_rq->idle_balance ?
10554 CPU_IDLE : CPU_NOT_IDLE;
10557 * If this CPU has a pending nohz_balance_kick, then do the
10558 * balancing on behalf of the other idle CPUs whose ticks are
10559 * stopped. Do nohz_idle_balance *before* rebalance_domains to
10560 * give the idle CPUs a chance to load balance. Else we may
10561 * load balance only within the local sched_domain hierarchy
10562 * and abort nohz_idle_balance altogether if we pull some load.
10564 if (nohz_idle_balance(this_rq, idle))
10567 /* normal load balance */
10568 update_blocked_averages(this_rq->cpu);
10569 rebalance_domains(this_rq, idle);
10573 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
10575 void trigger_load_balance(struct rq *rq)
10577 /* Don't need to rebalance while attached to NULL domain */
10578 if (unlikely(on_null_domain(rq)))
10581 if (time_after_eq(jiffies, rq->next_balance))
10582 raise_softirq(SCHED_SOFTIRQ);
10584 nohz_balancer_kick(rq);
10587 static void rq_online_fair(struct rq *rq)
10591 update_runtime_enabled(rq);
10594 static void rq_offline_fair(struct rq *rq)
10598 /* Ensure any throttled groups are reachable by pick_next_task */
10599 unthrottle_offline_cfs_rqs(rq);
10602 #endif /* CONFIG_SMP */
10605 * scheduler tick hitting a task of our scheduling class.
10607 * NOTE: This function can be called remotely by the tick offload that
10608 * goes along full dynticks. Therefore no local assumption can be made
10609 * and everything must be accessed through the @rq and @curr passed in
10612 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
10614 struct cfs_rq *cfs_rq;
10615 struct sched_entity *se = &curr->se;
10617 for_each_sched_entity(se) {
10618 cfs_rq = cfs_rq_of(se);
10619 entity_tick(cfs_rq, se, queued);
10622 if (static_branch_unlikely(&sched_numa_balancing))
10623 task_tick_numa(rq, curr);
10625 update_misfit_status(curr, rq);
10626 update_overutilized_status(task_rq(curr));
10630 * called on fork with the child task as argument from the parent's context
10631 * - child not yet on the tasklist
10632 * - preemption disabled
10634 static void task_fork_fair(struct task_struct *p)
10636 struct cfs_rq *cfs_rq;
10637 struct sched_entity *se = &p->se, *curr;
10638 struct rq *rq = this_rq();
10639 struct rq_flags rf;
10642 update_rq_clock(rq);
10644 cfs_rq = task_cfs_rq(current);
10645 curr = cfs_rq->curr;
10647 update_curr(cfs_rq);
10648 se->vruntime = curr->vruntime;
10650 place_entity(cfs_rq, se, 1);
10652 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
10654 * Upon rescheduling, sched_class::put_prev_task() will place
10655 * 'current' within the tree based on its new key value.
10657 swap(curr->vruntime, se->vruntime);
10661 se->vruntime -= cfs_rq->min_vruntime;
10662 rq_unlock(rq, &rf);
10666 * Priority of the task has changed. Check to see if we preempt
10667 * the current task.
10670 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
10672 if (!task_on_rq_queued(p))
10675 if (rq->cfs.nr_running == 1)
10679 * Reschedule if we are currently running on this runqueue and
10680 * our priority decreased, or if we are not currently running on
10681 * this runqueue and our priority is higher than the current's
10683 if (rq->curr == p) {
10684 if (p->prio > oldprio)
10687 check_preempt_curr(rq, p, 0);
10690 static inline bool vruntime_normalized(struct task_struct *p)
10692 struct sched_entity *se = &p->se;
10695 * In both the TASK_ON_RQ_QUEUED and TASK_ON_RQ_MIGRATING cases,
10696 * the dequeue_entity(.flags=0) will already have normalized the
10703 * When !on_rq, vruntime of the task has usually NOT been normalized.
10704 * But there are some cases where it has already been normalized:
10706 * - A forked child which is waiting for being woken up by
10707 * wake_up_new_task().
10708 * - A task which has been woken up by try_to_wake_up() and
10709 * waiting for actually being woken up by sched_ttwu_pending().
10711 if (!se->sum_exec_runtime ||
10712 (p->state == TASK_WAKING && p->sched_remote_wakeup))
10718 #ifdef CONFIG_FAIR_GROUP_SCHED
10720 * Propagate the changes of the sched_entity across the tg tree to make it
10721 * visible to the root
10723 static void propagate_entity_cfs_rq(struct sched_entity *se)
10725 struct cfs_rq *cfs_rq;
10727 /* Start to propagate at parent */
10730 for_each_sched_entity(se) {
10731 cfs_rq = cfs_rq_of(se);
10733 if (cfs_rq_throttled(cfs_rq))
10736 update_load_avg(cfs_rq, se, UPDATE_TG);
10740 static void propagate_entity_cfs_rq(struct sched_entity *se) { }
10743 static void detach_entity_cfs_rq(struct sched_entity *se)
10745 struct cfs_rq *cfs_rq = cfs_rq_of(se);
10747 /* Catch up with the cfs_rq and remove our load when we leave */
10748 update_load_avg(cfs_rq, se, 0);
10749 detach_entity_load_avg(cfs_rq, se);
10750 update_tg_load_avg(cfs_rq, false);
10751 propagate_entity_cfs_rq(se);
10754 static void attach_entity_cfs_rq(struct sched_entity *se)
10756 struct cfs_rq *cfs_rq = cfs_rq_of(se);
10758 #ifdef CONFIG_FAIR_GROUP_SCHED
10760 * Since the real-depth could have been changed (only FAIR
10761 * class maintain depth value), reset depth properly.
10763 se->depth = se->parent ? se->parent->depth + 1 : 0;
10766 /* Synchronize entity with its cfs_rq */
10767 update_load_avg(cfs_rq, se, sched_feat(ATTACH_AGE_LOAD) ? 0 : SKIP_AGE_LOAD);
10768 attach_entity_load_avg(cfs_rq, se);
10769 update_tg_load_avg(cfs_rq, false);
10770 propagate_entity_cfs_rq(se);
10773 static void detach_task_cfs_rq(struct task_struct *p)
10775 struct sched_entity *se = &p->se;
10776 struct cfs_rq *cfs_rq = cfs_rq_of(se);
10778 if (!vruntime_normalized(p)) {
10780 * Fix up our vruntime so that the current sleep doesn't
10781 * cause 'unlimited' sleep bonus.
10783 place_entity(cfs_rq, se, 0);
10784 se->vruntime -= cfs_rq->min_vruntime;
10787 detach_entity_cfs_rq(se);
10790 static void attach_task_cfs_rq(struct task_struct *p)
10792 struct sched_entity *se = &p->se;
10793 struct cfs_rq *cfs_rq = cfs_rq_of(se);
10795 attach_entity_cfs_rq(se);
10797 if (!vruntime_normalized(p))
10798 se->vruntime += cfs_rq->min_vruntime;
10801 static void switched_from_fair(struct rq *rq, struct task_struct *p)
10803 detach_task_cfs_rq(p);
10806 static void switched_to_fair(struct rq *rq, struct task_struct *p)
10808 attach_task_cfs_rq(p);
10810 if (task_on_rq_queued(p)) {
10812 * We were most likely switched from sched_rt, so
10813 * kick off the schedule if running, otherwise just see
10814 * if we can still preempt the current task.
10819 check_preempt_curr(rq, p, 0);
10823 /* Account for a task changing its policy or group.
10825 * This routine is mostly called to set cfs_rq->curr field when a task
10826 * migrates between groups/classes.
10828 static void set_next_task_fair(struct rq *rq, struct task_struct *p, bool first)
10830 struct sched_entity *se = &p->se;
10833 if (task_on_rq_queued(p)) {
10835 * Move the next running task to the front of the list, so our
10836 * cfs_tasks list becomes MRU one.
10838 list_move(&se->group_node, &rq->cfs_tasks);
10842 for_each_sched_entity(se) {
10843 struct cfs_rq *cfs_rq = cfs_rq_of(se);
10845 set_next_entity(cfs_rq, se);
10846 /* ensure bandwidth has been allocated on our new cfs_rq */
10847 account_cfs_rq_runtime(cfs_rq, 0);
10851 void init_cfs_rq(struct cfs_rq *cfs_rq)
10853 cfs_rq->tasks_timeline = RB_ROOT_CACHED;
10854 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
10855 #ifndef CONFIG_64BIT
10856 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
10859 raw_spin_lock_init(&cfs_rq->removed.lock);
10863 #ifdef CONFIG_FAIR_GROUP_SCHED
10864 static void task_set_group_fair(struct task_struct *p)
10866 struct sched_entity *se = &p->se;
10868 set_task_rq(p, task_cpu(p));
10869 se->depth = se->parent ? se->parent->depth + 1 : 0;
10872 static void task_move_group_fair(struct task_struct *p)
10874 detach_task_cfs_rq(p);
10875 set_task_rq(p, task_cpu(p));
10878 /* Tell se's cfs_rq has been changed -- migrated */
10879 p->se.avg.last_update_time = 0;
10881 attach_task_cfs_rq(p);
10884 static void task_change_group_fair(struct task_struct *p, int type)
10887 case TASK_SET_GROUP:
10888 task_set_group_fair(p);
10891 case TASK_MOVE_GROUP:
10892 task_move_group_fair(p);
10897 void free_fair_sched_group(struct task_group *tg)
10901 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
10903 for_each_possible_cpu(i) {
10905 kfree(tg->cfs_rq[i]);
10914 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
10916 struct sched_entity *se;
10917 struct cfs_rq *cfs_rq;
10920 tg->cfs_rq = kcalloc(nr_cpu_ids, sizeof(cfs_rq), GFP_KERNEL);
10923 tg->se = kcalloc(nr_cpu_ids, sizeof(se), GFP_KERNEL);
10927 tg->shares = NICE_0_LOAD;
10929 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
10931 for_each_possible_cpu(i) {
10932 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
10933 GFP_KERNEL, cpu_to_node(i));
10937 se = kzalloc_node(sizeof(struct sched_entity),
10938 GFP_KERNEL, cpu_to_node(i));
10942 init_cfs_rq(cfs_rq);
10943 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
10944 init_entity_runnable_average(se);
10955 void online_fair_sched_group(struct task_group *tg)
10957 struct sched_entity *se;
10958 struct rq_flags rf;
10962 for_each_possible_cpu(i) {
10965 rq_lock_irq(rq, &rf);
10966 update_rq_clock(rq);
10967 attach_entity_cfs_rq(se);
10968 sync_throttle(tg, i);
10969 rq_unlock_irq(rq, &rf);
10973 void unregister_fair_sched_group(struct task_group *tg)
10975 unsigned long flags;
10979 for_each_possible_cpu(cpu) {
10981 remove_entity_load_avg(tg->se[cpu]);
10984 * Only empty task groups can be destroyed; so we can speculatively
10985 * check on_list without danger of it being re-added.
10987 if (!tg->cfs_rq[cpu]->on_list)
10992 raw_spin_lock_irqsave(&rq->lock, flags);
10993 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
10994 raw_spin_unlock_irqrestore(&rq->lock, flags);
10998 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
10999 struct sched_entity *se, int cpu,
11000 struct sched_entity *parent)
11002 struct rq *rq = cpu_rq(cpu);
11006 init_cfs_rq_runtime(cfs_rq);
11008 tg->cfs_rq[cpu] = cfs_rq;
11011 /* se could be NULL for root_task_group */
11016 se->cfs_rq = &rq->cfs;
11019 se->cfs_rq = parent->my_q;
11020 se->depth = parent->depth + 1;
11024 /* guarantee group entities always have weight */
11025 update_load_set(&se->load, NICE_0_LOAD);
11026 se->parent = parent;
11029 static DEFINE_MUTEX(shares_mutex);
11031 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
11036 * We can't change the weight of the root cgroup.
11041 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
11043 mutex_lock(&shares_mutex);
11044 if (tg->shares == shares)
11047 tg->shares = shares;
11048 for_each_possible_cpu(i) {
11049 struct rq *rq = cpu_rq(i);
11050 struct sched_entity *se = tg->se[i];
11051 struct rq_flags rf;
11053 /* Propagate contribution to hierarchy */
11054 rq_lock_irqsave(rq, &rf);
11055 update_rq_clock(rq);
11056 for_each_sched_entity(se) {
11057 update_load_avg(cfs_rq_of(se), se, UPDATE_TG);
11058 update_cfs_group(se);
11060 rq_unlock_irqrestore(rq, &rf);
11064 mutex_unlock(&shares_mutex);
11067 #else /* CONFIG_FAIR_GROUP_SCHED */
11069 void free_fair_sched_group(struct task_group *tg) { }
11071 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
11076 void online_fair_sched_group(struct task_group *tg) { }
11078 void unregister_fair_sched_group(struct task_group *tg) { }
11080 #endif /* CONFIG_FAIR_GROUP_SCHED */
11083 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
11085 struct sched_entity *se = &task->se;
11086 unsigned int rr_interval = 0;
11089 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
11092 if (rq->cfs.load.weight)
11093 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
11095 return rr_interval;
11099 * All the scheduling class methods:
11101 const struct sched_class fair_sched_class = {
11102 .next = &idle_sched_class,
11103 .enqueue_task = enqueue_task_fair,
11104 .dequeue_task = dequeue_task_fair,
11105 .yield_task = yield_task_fair,
11106 .yield_to_task = yield_to_task_fair,
11108 .check_preempt_curr = check_preempt_wakeup,
11110 .pick_next_task = __pick_next_task_fair,
11111 .put_prev_task = put_prev_task_fair,
11112 .set_next_task = set_next_task_fair,
11115 .balance = balance_fair,
11116 .select_task_rq = select_task_rq_fair,
11117 .migrate_task_rq = migrate_task_rq_fair,
11119 .rq_online = rq_online_fair,
11120 .rq_offline = rq_offline_fair,
11122 .task_dead = task_dead_fair,
11123 .set_cpus_allowed = set_cpus_allowed_common,
11126 .task_tick = task_tick_fair,
11127 .task_fork = task_fork_fair,
11129 .prio_changed = prio_changed_fair,
11130 .switched_from = switched_from_fair,
11131 .switched_to = switched_to_fair,
11133 .get_rr_interval = get_rr_interval_fair,
11135 .update_curr = update_curr_fair,
11137 #ifdef CONFIG_FAIR_GROUP_SCHED
11138 .task_change_group = task_change_group_fair,
11141 #ifdef CONFIG_UCLAMP_TASK
11142 .uclamp_enabled = 1,
11146 #ifdef CONFIG_SCHED_DEBUG
11147 void print_cfs_stats(struct seq_file *m, int cpu)
11149 struct cfs_rq *cfs_rq, *pos;
11152 for_each_leaf_cfs_rq_safe(cpu_rq(cpu), cfs_rq, pos)
11153 print_cfs_rq(m, cpu, cfs_rq);
11157 #ifdef CONFIG_NUMA_BALANCING
11158 void show_numa_stats(struct task_struct *p, struct seq_file *m)
11161 unsigned long tsf = 0, tpf = 0, gsf = 0, gpf = 0;
11162 struct numa_group *ng;
11165 ng = rcu_dereference(p->numa_group);
11166 for_each_online_node(node) {
11167 if (p->numa_faults) {
11168 tsf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 0)];
11169 tpf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 1)];
11172 gsf = ng->faults[task_faults_idx(NUMA_MEM, node, 0)],
11173 gpf = ng->faults[task_faults_idx(NUMA_MEM, node, 1)];
11175 print_numa_stats(m, node, tsf, tpf, gsf, gpf);
11179 #endif /* CONFIG_NUMA_BALANCING */
11180 #endif /* CONFIG_SCHED_DEBUG */
11182 __init void init_sched_fair_class(void)
11185 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
11187 #ifdef CONFIG_NO_HZ_COMMON
11188 nohz.next_balance = jiffies;
11189 nohz.next_blocked = jiffies;
11190 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
11197 * Helper functions to facilitate extracting info from tracepoints.
11200 const struct sched_avg *sched_trace_cfs_rq_avg(struct cfs_rq *cfs_rq)
11203 return cfs_rq ? &cfs_rq->avg : NULL;
11208 EXPORT_SYMBOL_GPL(sched_trace_cfs_rq_avg);
11210 char *sched_trace_cfs_rq_path(struct cfs_rq *cfs_rq, char *str, int len)
11214 strlcpy(str, "(null)", len);
11219 cfs_rq_tg_path(cfs_rq, str, len);
11222 EXPORT_SYMBOL_GPL(sched_trace_cfs_rq_path);
11224 int sched_trace_cfs_rq_cpu(struct cfs_rq *cfs_rq)
11226 return cfs_rq ? cpu_of(rq_of(cfs_rq)) : -1;
11228 EXPORT_SYMBOL_GPL(sched_trace_cfs_rq_cpu);
11230 const struct sched_avg *sched_trace_rq_avg_rt(struct rq *rq)
11233 return rq ? &rq->avg_rt : NULL;
11238 EXPORT_SYMBOL_GPL(sched_trace_rq_avg_rt);
11240 const struct sched_avg *sched_trace_rq_avg_dl(struct rq *rq)
11243 return rq ? &rq->avg_dl : NULL;
11248 EXPORT_SYMBOL_GPL(sched_trace_rq_avg_dl);
11250 const struct sched_avg *sched_trace_rq_avg_irq(struct rq *rq)
11252 #if defined(CONFIG_SMP) && defined(CONFIG_HAVE_SCHED_AVG_IRQ)
11253 return rq ? &rq->avg_irq : NULL;
11258 EXPORT_SYMBOL_GPL(sched_trace_rq_avg_irq);
11260 int sched_trace_rq_cpu(struct rq *rq)
11262 return rq ? cpu_of(rq) : -1;
11264 EXPORT_SYMBOL_GPL(sched_trace_rq_cpu);
11266 const struct cpumask *sched_trace_rd_span(struct root_domain *rd)
11269 return rd ? rd->span : NULL;
11274 EXPORT_SYMBOL_GPL(sched_trace_rd_span);