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
26 * Targeted preemption latency for CPU-bound tasks:
28 * NOTE: this latency value is not the same as the concept of
29 * 'timeslice length' - timeslices in CFS are of variable length
30 * and have no persistent notion like in traditional, time-slice
31 * based scheduling concepts.
33 * (to see the precise effective timeslice length of your workload,
34 * run vmstat and monitor the context-switches (cs) field)
36 * (default: 6ms * (1 + ilog(ncpus)), units: nanoseconds)
38 unsigned int sysctl_sched_latency = 6000000ULL;
39 static unsigned int normalized_sysctl_sched_latency = 6000000ULL;
42 * The initial- and re-scaling of tunables is configurable
46 * SCHED_TUNABLESCALING_NONE - unscaled, always *1
47 * SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus)
48 * SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus
50 * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus))
52 unsigned int sysctl_sched_tunable_scaling = SCHED_TUNABLESCALING_LOG;
55 * Minimal preemption granularity for CPU-bound tasks:
57 * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds)
59 unsigned int sysctl_sched_min_granularity = 750000ULL;
60 static unsigned int normalized_sysctl_sched_min_granularity = 750000ULL;
63 * This value is kept at sysctl_sched_latency/sysctl_sched_min_granularity
65 static unsigned int sched_nr_latency = 8;
68 * After fork, child runs first. If set to 0 (default) then
69 * parent will (try to) run first.
71 unsigned int sysctl_sched_child_runs_first __read_mostly;
74 * SCHED_OTHER wake-up granularity.
76 * This option delays the preemption effects of decoupled workloads
77 * and reduces their over-scheduling. Synchronous workloads will still
78 * have immediate wakeup/sleep latencies.
80 * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds)
82 unsigned int sysctl_sched_wakeup_granularity = 1000000UL;
83 static unsigned int normalized_sysctl_sched_wakeup_granularity = 1000000UL;
85 const_debug unsigned int sysctl_sched_migration_cost = 500000UL;
87 int sched_thermal_decay_shift;
88 static int __init setup_sched_thermal_decay_shift(char *str)
92 if (kstrtoint(str, 0, &_shift))
93 pr_warn("Unable to set scheduler thermal pressure decay shift parameter\n");
95 sched_thermal_decay_shift = clamp(_shift, 0, 10);
98 __setup("sched_thermal_decay_shift=", setup_sched_thermal_decay_shift);
102 * For asym packing, by default the lower numbered CPU has higher priority.
104 int __weak arch_asym_cpu_priority(int cpu)
110 * The margin used when comparing utilization with CPU capacity.
114 #define fits_capacity(cap, max) ((cap) * 1280 < (max) * 1024)
117 * The margin used when comparing CPU capacities.
118 * is 'cap1' noticeably greater than 'cap2'
122 #define capacity_greater(cap1, cap2) ((cap1) * 1024 > (cap2) * 1078)
125 #ifdef CONFIG_CFS_BANDWIDTH
127 * Amount of runtime to allocate from global (tg) to local (per-cfs_rq) pool
128 * each time a cfs_rq requests quota.
130 * Note: in the case that the slice exceeds the runtime remaining (either due
131 * to consumption or the quota being specified to be smaller than the slice)
132 * we will always only issue the remaining available time.
134 * (default: 5 msec, units: microseconds)
136 unsigned int sysctl_sched_cfs_bandwidth_slice = 5000UL;
139 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
145 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
151 static inline void update_load_set(struct load_weight *lw, unsigned long w)
158 * Increase the granularity value when there are more CPUs,
159 * because with more CPUs the 'effective latency' as visible
160 * to users decreases. But the relationship is not linear,
161 * so pick a second-best guess by going with the log2 of the
164 * This idea comes from the SD scheduler of Con Kolivas:
166 static unsigned int get_update_sysctl_factor(void)
168 unsigned int cpus = min_t(unsigned int, num_online_cpus(), 8);
171 switch (sysctl_sched_tunable_scaling) {
172 case SCHED_TUNABLESCALING_NONE:
175 case SCHED_TUNABLESCALING_LINEAR:
178 case SCHED_TUNABLESCALING_LOG:
180 factor = 1 + ilog2(cpus);
187 static void update_sysctl(void)
189 unsigned int factor = get_update_sysctl_factor();
191 #define SET_SYSCTL(name) \
192 (sysctl_##name = (factor) * normalized_sysctl_##name)
193 SET_SYSCTL(sched_min_granularity);
194 SET_SYSCTL(sched_latency);
195 SET_SYSCTL(sched_wakeup_granularity);
199 void __init sched_init_granularity(void)
204 #define WMULT_CONST (~0U)
205 #define WMULT_SHIFT 32
207 static void __update_inv_weight(struct load_weight *lw)
211 if (likely(lw->inv_weight))
214 w = scale_load_down(lw->weight);
216 if (BITS_PER_LONG > 32 && unlikely(w >= WMULT_CONST))
218 else if (unlikely(!w))
219 lw->inv_weight = WMULT_CONST;
221 lw->inv_weight = WMULT_CONST / w;
225 * delta_exec * weight / lw.weight
227 * (delta_exec * (weight * lw->inv_weight)) >> WMULT_SHIFT
229 * Either weight := NICE_0_LOAD and lw \e sched_prio_to_wmult[], in which case
230 * we're guaranteed shift stays positive because inv_weight is guaranteed to
231 * fit 32 bits, and NICE_0_LOAD gives another 10 bits; therefore shift >= 22.
233 * Or, weight =< lw.weight (because lw.weight is the runqueue weight), thus
234 * weight/lw.weight <= 1, and therefore our shift will also be positive.
236 static u64 __calc_delta(u64 delta_exec, unsigned long weight, struct load_weight *lw)
238 u64 fact = scale_load_down(weight);
239 u32 fact_hi = (u32)(fact >> 32);
240 int shift = WMULT_SHIFT;
243 __update_inv_weight(lw);
245 if (unlikely(fact_hi)) {
251 fact = mul_u32_u32(fact, lw->inv_weight);
253 fact_hi = (u32)(fact >> 32);
260 return mul_u64_u32_shr(delta_exec, fact, shift);
264 const struct sched_class fair_sched_class;
266 /**************************************************************
267 * CFS operations on generic schedulable entities:
270 #ifdef CONFIG_FAIR_GROUP_SCHED
271 static inline struct task_struct *task_of(struct sched_entity *se)
273 SCHED_WARN_ON(!entity_is_task(se));
274 return container_of(se, struct task_struct, se);
277 /* Walk up scheduling entities hierarchy */
278 #define for_each_sched_entity(se) \
279 for (; se; se = se->parent)
281 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
286 /* runqueue on which this entity is (to be) queued */
287 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
292 /* runqueue "owned" by this group */
293 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
298 static inline void cfs_rq_tg_path(struct cfs_rq *cfs_rq, char *path, int len)
303 if (cfs_rq && task_group_is_autogroup(cfs_rq->tg))
304 autogroup_path(cfs_rq->tg, path, len);
305 else if (cfs_rq && cfs_rq->tg->css.cgroup)
306 cgroup_path(cfs_rq->tg->css.cgroup, path, len);
308 strlcpy(path, "(null)", len);
311 static inline bool list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
313 struct rq *rq = rq_of(cfs_rq);
314 int cpu = cpu_of(rq);
317 return rq->tmp_alone_branch == &rq->leaf_cfs_rq_list;
322 * Ensure we either appear before our parent (if already
323 * enqueued) or force our parent to appear after us when it is
324 * enqueued. The fact that we always enqueue bottom-up
325 * reduces this to two cases and a special case for the root
326 * cfs_rq. Furthermore, it also means that we will always reset
327 * tmp_alone_branch either when the branch is connected
328 * to a tree or when we reach the top of the tree
330 if (cfs_rq->tg->parent &&
331 cfs_rq->tg->parent->cfs_rq[cpu]->on_list) {
333 * If parent is already on the list, we add the child
334 * just before. Thanks to circular linked property of
335 * the list, this means to put the child at the tail
336 * of the list that starts by parent.
338 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
339 &(cfs_rq->tg->parent->cfs_rq[cpu]->leaf_cfs_rq_list));
341 * The branch is now connected to its tree so we can
342 * reset tmp_alone_branch to the beginning of the
345 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
349 if (!cfs_rq->tg->parent) {
351 * cfs rq without parent should be put
352 * at the tail of the list.
354 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
355 &rq->leaf_cfs_rq_list);
357 * We have reach the top of a tree so we can reset
358 * tmp_alone_branch to the beginning of the list.
360 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
365 * The parent has not already been added so we want to
366 * make sure that it will be put after us.
367 * tmp_alone_branch points to the begin of the branch
368 * where we will add parent.
370 list_add_rcu(&cfs_rq->leaf_cfs_rq_list, rq->tmp_alone_branch);
372 * update tmp_alone_branch to points to the new begin
375 rq->tmp_alone_branch = &cfs_rq->leaf_cfs_rq_list;
379 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
381 if (cfs_rq->on_list) {
382 struct rq *rq = rq_of(cfs_rq);
385 * With cfs_rq being unthrottled/throttled during an enqueue,
386 * it can happen the tmp_alone_branch points the a leaf that
387 * we finally want to del. In this case, tmp_alone_branch moves
388 * to the prev element but it will point to rq->leaf_cfs_rq_list
389 * at the end of the enqueue.
391 if (rq->tmp_alone_branch == &cfs_rq->leaf_cfs_rq_list)
392 rq->tmp_alone_branch = cfs_rq->leaf_cfs_rq_list.prev;
394 list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
399 static inline void assert_list_leaf_cfs_rq(struct rq *rq)
401 SCHED_WARN_ON(rq->tmp_alone_branch != &rq->leaf_cfs_rq_list);
404 /* Iterate thr' all leaf cfs_rq's on a runqueue */
405 #define for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) \
406 list_for_each_entry_safe(cfs_rq, pos, &rq->leaf_cfs_rq_list, \
409 /* Do the two (enqueued) entities belong to the same group ? */
410 static inline struct cfs_rq *
411 is_same_group(struct sched_entity *se, struct sched_entity *pse)
413 if (se->cfs_rq == pse->cfs_rq)
419 static inline struct sched_entity *parent_entity(struct sched_entity *se)
425 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
427 int se_depth, pse_depth;
430 * preemption test can be made between sibling entities who are in the
431 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
432 * both tasks until we find their ancestors who are siblings of common
436 /* First walk up until both entities are at same depth */
437 se_depth = (*se)->depth;
438 pse_depth = (*pse)->depth;
440 while (se_depth > pse_depth) {
442 *se = parent_entity(*se);
445 while (pse_depth > se_depth) {
447 *pse = parent_entity(*pse);
450 while (!is_same_group(*se, *pse)) {
451 *se = parent_entity(*se);
452 *pse = parent_entity(*pse);
456 #else /* !CONFIG_FAIR_GROUP_SCHED */
458 static inline struct task_struct *task_of(struct sched_entity *se)
460 return container_of(se, struct task_struct, se);
463 #define for_each_sched_entity(se) \
464 for (; se; se = NULL)
466 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
468 return &task_rq(p)->cfs;
471 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
473 struct task_struct *p = task_of(se);
474 struct rq *rq = task_rq(p);
479 /* runqueue "owned" by this group */
480 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
485 static inline void cfs_rq_tg_path(struct cfs_rq *cfs_rq, char *path, int len)
488 strlcpy(path, "(null)", len);
491 static inline bool list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
496 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
500 static inline void assert_list_leaf_cfs_rq(struct rq *rq)
504 #define for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) \
505 for (cfs_rq = &rq->cfs, pos = NULL; cfs_rq; cfs_rq = pos)
507 static inline struct sched_entity *parent_entity(struct sched_entity *se)
513 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
517 #endif /* CONFIG_FAIR_GROUP_SCHED */
519 static __always_inline
520 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec);
522 /**************************************************************
523 * Scheduling class tree data structure manipulation methods:
526 static inline u64 max_vruntime(u64 max_vruntime, u64 vruntime)
528 s64 delta = (s64)(vruntime - max_vruntime);
530 max_vruntime = vruntime;
535 static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime)
537 s64 delta = (s64)(vruntime - min_vruntime);
539 min_vruntime = vruntime;
544 static inline bool entity_before(struct sched_entity *a,
545 struct sched_entity *b)
547 return (s64)(a->vruntime - b->vruntime) < 0;
550 #define __node_2_se(node) \
551 rb_entry((node), struct sched_entity, run_node)
553 static void update_min_vruntime(struct cfs_rq *cfs_rq)
555 struct sched_entity *curr = cfs_rq->curr;
556 struct rb_node *leftmost = rb_first_cached(&cfs_rq->tasks_timeline);
558 u64 vruntime = cfs_rq->min_vruntime;
562 vruntime = curr->vruntime;
567 if (leftmost) { /* non-empty tree */
568 struct sched_entity *se = __node_2_se(leftmost);
571 vruntime = se->vruntime;
573 vruntime = min_vruntime(vruntime, se->vruntime);
576 /* ensure we never gain time by being placed backwards. */
577 cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime);
580 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
584 static inline bool __entity_less(struct rb_node *a, const struct rb_node *b)
586 return entity_before(__node_2_se(a), __node_2_se(b));
590 * Enqueue an entity into the rb-tree:
592 static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
594 rb_add_cached(&se->run_node, &cfs_rq->tasks_timeline, __entity_less);
597 static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
599 rb_erase_cached(&se->run_node, &cfs_rq->tasks_timeline);
602 struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq)
604 struct rb_node *left = rb_first_cached(&cfs_rq->tasks_timeline);
609 return __node_2_se(left);
612 static struct sched_entity *__pick_next_entity(struct sched_entity *se)
614 struct rb_node *next = rb_next(&se->run_node);
619 return __node_2_se(next);
622 #ifdef CONFIG_SCHED_DEBUG
623 struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
625 struct rb_node *last = rb_last(&cfs_rq->tasks_timeline.rb_root);
630 return __node_2_se(last);
633 /**************************************************************
634 * Scheduling class statistics methods:
637 int sched_update_scaling(void)
639 unsigned int factor = get_update_sysctl_factor();
641 sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency,
642 sysctl_sched_min_granularity);
644 #define WRT_SYSCTL(name) \
645 (normalized_sysctl_##name = sysctl_##name / (factor))
646 WRT_SYSCTL(sched_min_granularity);
647 WRT_SYSCTL(sched_latency);
648 WRT_SYSCTL(sched_wakeup_granularity);
658 static inline u64 calc_delta_fair(u64 delta, struct sched_entity *se)
660 if (unlikely(se->load.weight != NICE_0_LOAD))
661 delta = __calc_delta(delta, NICE_0_LOAD, &se->load);
667 * The idea is to set a period in which each task runs once.
669 * When there are too many tasks (sched_nr_latency) we have to stretch
670 * this period because otherwise the slices get too small.
672 * p = (nr <= nl) ? l : l*nr/nl
674 static u64 __sched_period(unsigned long nr_running)
676 if (unlikely(nr_running > sched_nr_latency))
677 return nr_running * sysctl_sched_min_granularity;
679 return sysctl_sched_latency;
683 * We calculate the wall-time slice from the period by taking a part
684 * proportional to the weight.
688 static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
690 unsigned int nr_running = cfs_rq->nr_running;
693 if (sched_feat(ALT_PERIOD))
694 nr_running = rq_of(cfs_rq)->cfs.h_nr_running;
696 slice = __sched_period(nr_running + !se->on_rq);
698 for_each_sched_entity(se) {
699 struct load_weight *load;
700 struct load_weight lw;
702 cfs_rq = cfs_rq_of(se);
703 load = &cfs_rq->load;
705 if (unlikely(!se->on_rq)) {
708 update_load_add(&lw, se->load.weight);
711 slice = __calc_delta(slice, se->load.weight, load);
714 if (sched_feat(BASE_SLICE))
715 slice = max(slice, (u64)sysctl_sched_min_granularity);
721 * We calculate the vruntime slice of a to-be-inserted task.
725 static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se)
727 return calc_delta_fair(sched_slice(cfs_rq, se), se);
733 static int select_idle_sibling(struct task_struct *p, int prev_cpu, int cpu);
734 static unsigned long task_h_load(struct task_struct *p);
735 static unsigned long capacity_of(int cpu);
737 /* Give new sched_entity start runnable values to heavy its load in infant time */
738 void init_entity_runnable_average(struct sched_entity *se)
740 struct sched_avg *sa = &se->avg;
742 memset(sa, 0, sizeof(*sa));
745 * Tasks are initialized with full load to be seen as heavy tasks until
746 * they get a chance to stabilize to their real load level.
747 * Group entities are initialized with zero load to reflect the fact that
748 * nothing has been attached to the task group yet.
750 if (entity_is_task(se))
751 sa->load_avg = scale_load_down(se->load.weight);
753 /* when this task enqueue'ed, it will contribute to its cfs_rq's load_avg */
756 static void attach_entity_cfs_rq(struct sched_entity *se);
759 * With new tasks being created, their initial util_avgs are extrapolated
760 * based on the cfs_rq's current util_avg:
762 * util_avg = cfs_rq->util_avg / (cfs_rq->load_avg + 1) * se.load.weight
764 * However, in many cases, the above util_avg does not give a desired
765 * value. Moreover, the sum of the util_avgs may be divergent, such
766 * as when the series is a harmonic series.
768 * To solve this problem, we also cap the util_avg of successive tasks to
769 * only 1/2 of the left utilization budget:
771 * util_avg_cap = (cpu_scale - cfs_rq->avg.util_avg) / 2^n
773 * where n denotes the nth task and cpu_scale the CPU capacity.
775 * For example, for a CPU with 1024 of capacity, a simplest series from
776 * the beginning would be like:
778 * task util_avg: 512, 256, 128, 64, 32, 16, 8, ...
779 * cfs_rq util_avg: 512, 768, 896, 960, 992, 1008, 1016, ...
781 * Finally, that extrapolated util_avg is clamped to the cap (util_avg_cap)
782 * if util_avg > util_avg_cap.
784 void post_init_entity_util_avg(struct task_struct *p)
786 struct sched_entity *se = &p->se;
787 struct cfs_rq *cfs_rq = cfs_rq_of(se);
788 struct sched_avg *sa = &se->avg;
789 long cpu_scale = arch_scale_cpu_capacity(cpu_of(rq_of(cfs_rq)));
790 long cap = (long)(cpu_scale - cfs_rq->avg.util_avg) / 2;
793 if (cfs_rq->avg.util_avg != 0) {
794 sa->util_avg = cfs_rq->avg.util_avg * se->load.weight;
795 sa->util_avg /= (cfs_rq->avg.load_avg + 1);
797 if (sa->util_avg > cap)
804 sa->runnable_avg = sa->util_avg;
806 if (p->sched_class != &fair_sched_class) {
808 * For !fair tasks do:
810 update_cfs_rq_load_avg(now, cfs_rq);
811 attach_entity_load_avg(cfs_rq, se);
812 switched_from_fair(rq, p);
814 * such that the next switched_to_fair() has the
817 se->avg.last_update_time = cfs_rq_clock_pelt(cfs_rq);
821 attach_entity_cfs_rq(se);
824 #else /* !CONFIG_SMP */
825 void init_entity_runnable_average(struct sched_entity *se)
828 void post_init_entity_util_avg(struct task_struct *p)
831 static void update_tg_load_avg(struct cfs_rq *cfs_rq)
834 #endif /* CONFIG_SMP */
837 * Update the current task's runtime statistics.
839 static void update_curr(struct cfs_rq *cfs_rq)
841 struct sched_entity *curr = cfs_rq->curr;
842 u64 now = rq_clock_task(rq_of(cfs_rq));
848 delta_exec = now - curr->exec_start;
849 if (unlikely((s64)delta_exec <= 0))
852 curr->exec_start = now;
854 schedstat_set(curr->statistics.exec_max,
855 max(delta_exec, curr->statistics.exec_max));
857 curr->sum_exec_runtime += delta_exec;
858 schedstat_add(cfs_rq->exec_clock, delta_exec);
860 curr->vruntime += calc_delta_fair(delta_exec, curr);
861 update_min_vruntime(cfs_rq);
863 if (entity_is_task(curr)) {
864 struct task_struct *curtask = task_of(curr);
866 trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime);
867 cgroup_account_cputime(curtask, delta_exec);
868 account_group_exec_runtime(curtask, delta_exec);
871 account_cfs_rq_runtime(cfs_rq, delta_exec);
874 static void update_curr_fair(struct rq *rq)
876 update_curr(cfs_rq_of(&rq->curr->se));
880 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
882 u64 wait_start, prev_wait_start;
884 if (!schedstat_enabled())
887 wait_start = rq_clock(rq_of(cfs_rq));
888 prev_wait_start = schedstat_val(se->statistics.wait_start);
890 if (entity_is_task(se) && task_on_rq_migrating(task_of(se)) &&
891 likely(wait_start > prev_wait_start))
892 wait_start -= prev_wait_start;
894 __schedstat_set(se->statistics.wait_start, wait_start);
898 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
900 struct task_struct *p;
903 if (!schedstat_enabled())
907 * When the sched_schedstat changes from 0 to 1, some sched se
908 * maybe already in the runqueue, the se->statistics.wait_start
909 * will be 0.So it will let the delta wrong. We need to avoid this
912 if (unlikely(!schedstat_val(se->statistics.wait_start)))
915 delta = rq_clock(rq_of(cfs_rq)) - schedstat_val(se->statistics.wait_start);
917 if (entity_is_task(se)) {
919 if (task_on_rq_migrating(p)) {
921 * Preserve migrating task's wait time so wait_start
922 * time stamp can be adjusted to accumulate wait time
923 * prior to migration.
925 __schedstat_set(se->statistics.wait_start, delta);
928 trace_sched_stat_wait(p, delta);
931 __schedstat_set(se->statistics.wait_max,
932 max(schedstat_val(se->statistics.wait_max), delta));
933 __schedstat_inc(se->statistics.wait_count);
934 __schedstat_add(se->statistics.wait_sum, delta);
935 __schedstat_set(se->statistics.wait_start, 0);
939 update_stats_enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
941 struct task_struct *tsk = NULL;
942 u64 sleep_start, block_start;
944 if (!schedstat_enabled())
947 sleep_start = schedstat_val(se->statistics.sleep_start);
948 block_start = schedstat_val(se->statistics.block_start);
950 if (entity_is_task(se))
954 u64 delta = rq_clock(rq_of(cfs_rq)) - sleep_start;
959 if (unlikely(delta > schedstat_val(se->statistics.sleep_max)))
960 __schedstat_set(se->statistics.sleep_max, delta);
962 __schedstat_set(se->statistics.sleep_start, 0);
963 __schedstat_add(se->statistics.sum_sleep_runtime, delta);
966 account_scheduler_latency(tsk, delta >> 10, 1);
967 trace_sched_stat_sleep(tsk, delta);
971 u64 delta = rq_clock(rq_of(cfs_rq)) - block_start;
976 if (unlikely(delta > schedstat_val(se->statistics.block_max)))
977 __schedstat_set(se->statistics.block_max, delta);
979 __schedstat_set(se->statistics.block_start, 0);
980 __schedstat_add(se->statistics.sum_sleep_runtime, delta);
983 if (tsk->in_iowait) {
984 __schedstat_add(se->statistics.iowait_sum, delta);
985 __schedstat_inc(se->statistics.iowait_count);
986 trace_sched_stat_iowait(tsk, delta);
989 trace_sched_stat_blocked(tsk, delta);
992 * Blocking time is in units of nanosecs, so shift by
993 * 20 to get a milliseconds-range estimation of the
994 * amount of time that the task spent sleeping:
996 if (unlikely(prof_on == SLEEP_PROFILING)) {
997 profile_hits(SLEEP_PROFILING,
998 (void *)get_wchan(tsk),
1001 account_scheduler_latency(tsk, delta >> 10, 0);
1007 * Task is being enqueued - update stats:
1010 update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
1012 if (!schedstat_enabled())
1016 * Are we enqueueing a waiting task? (for current tasks
1017 * a dequeue/enqueue event is a NOP)
1019 if (se != cfs_rq->curr)
1020 update_stats_wait_start(cfs_rq, se);
1022 if (flags & ENQUEUE_WAKEUP)
1023 update_stats_enqueue_sleeper(cfs_rq, se);
1027 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
1030 if (!schedstat_enabled())
1034 * Mark the end of the wait period if dequeueing a
1037 if (se != cfs_rq->curr)
1038 update_stats_wait_end(cfs_rq, se);
1040 if ((flags & DEQUEUE_SLEEP) && entity_is_task(se)) {
1041 struct task_struct *tsk = task_of(se);
1043 if (tsk->state & TASK_INTERRUPTIBLE)
1044 __schedstat_set(se->statistics.sleep_start,
1045 rq_clock(rq_of(cfs_rq)));
1046 if (tsk->state & TASK_UNINTERRUPTIBLE)
1047 __schedstat_set(se->statistics.block_start,
1048 rq_clock(rq_of(cfs_rq)));
1053 * We are picking a new current task - update its stats:
1056 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
1059 * We are starting a new run period:
1061 se->exec_start = rq_clock_task(rq_of(cfs_rq));
1064 /**************************************************
1065 * Scheduling class queueing methods:
1068 #ifdef CONFIG_NUMA_BALANCING
1070 * Approximate time to scan a full NUMA task in ms. The task scan period is
1071 * calculated based on the tasks virtual memory size and
1072 * numa_balancing_scan_size.
1074 unsigned int sysctl_numa_balancing_scan_period_min = 1000;
1075 unsigned int sysctl_numa_balancing_scan_period_max = 60000;
1077 /* Portion of address space to scan in MB */
1078 unsigned int sysctl_numa_balancing_scan_size = 256;
1080 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
1081 unsigned int sysctl_numa_balancing_scan_delay = 1000;
1084 refcount_t refcount;
1086 spinlock_t lock; /* nr_tasks, tasks */
1091 struct rcu_head rcu;
1092 unsigned long total_faults;
1093 unsigned long max_faults_cpu;
1095 * Faults_cpu is used to decide whether memory should move
1096 * towards the CPU. As a consequence, these stats are weighted
1097 * more by CPU use than by memory faults.
1099 unsigned long *faults_cpu;
1100 unsigned long faults[];
1104 * For functions that can be called in multiple contexts that permit reading
1105 * ->numa_group (see struct task_struct for locking rules).
1107 static struct numa_group *deref_task_numa_group(struct task_struct *p)
1109 return rcu_dereference_check(p->numa_group, p == current ||
1110 (lockdep_is_held(&task_rq(p)->lock) && !READ_ONCE(p->on_cpu)));
1113 static struct numa_group *deref_curr_numa_group(struct task_struct *p)
1115 return rcu_dereference_protected(p->numa_group, p == current);
1118 static inline unsigned long group_faults_priv(struct numa_group *ng);
1119 static inline unsigned long group_faults_shared(struct numa_group *ng);
1121 static unsigned int task_nr_scan_windows(struct task_struct *p)
1123 unsigned long rss = 0;
1124 unsigned long nr_scan_pages;
1127 * Calculations based on RSS as non-present and empty pages are skipped
1128 * by the PTE scanner and NUMA hinting faults should be trapped based
1131 nr_scan_pages = sysctl_numa_balancing_scan_size << (20 - PAGE_SHIFT);
1132 rss = get_mm_rss(p->mm);
1134 rss = nr_scan_pages;
1136 rss = round_up(rss, nr_scan_pages);
1137 return rss / nr_scan_pages;
1140 /* For sanity's sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */
1141 #define MAX_SCAN_WINDOW 2560
1143 static unsigned int task_scan_min(struct task_struct *p)
1145 unsigned int scan_size = READ_ONCE(sysctl_numa_balancing_scan_size);
1146 unsigned int scan, floor;
1147 unsigned int windows = 1;
1149 if (scan_size < MAX_SCAN_WINDOW)
1150 windows = MAX_SCAN_WINDOW / scan_size;
1151 floor = 1000 / windows;
1153 scan = sysctl_numa_balancing_scan_period_min / task_nr_scan_windows(p);
1154 return max_t(unsigned int, floor, scan);
1157 static unsigned int task_scan_start(struct task_struct *p)
1159 unsigned long smin = task_scan_min(p);
1160 unsigned long period = smin;
1161 struct numa_group *ng;
1163 /* Scale the maximum scan period with the amount of shared memory. */
1165 ng = rcu_dereference(p->numa_group);
1167 unsigned long shared = group_faults_shared(ng);
1168 unsigned long private = group_faults_priv(ng);
1170 period *= refcount_read(&ng->refcount);
1171 period *= shared + 1;
1172 period /= private + shared + 1;
1176 return max(smin, period);
1179 static unsigned int task_scan_max(struct task_struct *p)
1181 unsigned long smin = task_scan_min(p);
1183 struct numa_group *ng;
1185 /* Watch for min being lower than max due to floor calculations */
1186 smax = sysctl_numa_balancing_scan_period_max / task_nr_scan_windows(p);
1188 /* Scale the maximum scan period with the amount of shared memory. */
1189 ng = deref_curr_numa_group(p);
1191 unsigned long shared = group_faults_shared(ng);
1192 unsigned long private = group_faults_priv(ng);
1193 unsigned long period = smax;
1195 period *= refcount_read(&ng->refcount);
1196 period *= shared + 1;
1197 period /= private + shared + 1;
1199 smax = max(smax, period);
1202 return max(smin, smax);
1205 static void account_numa_enqueue(struct rq *rq, struct task_struct *p)
1207 rq->nr_numa_running += (p->numa_preferred_nid != NUMA_NO_NODE);
1208 rq->nr_preferred_running += (p->numa_preferred_nid == task_node(p));
1211 static void account_numa_dequeue(struct rq *rq, struct task_struct *p)
1213 rq->nr_numa_running -= (p->numa_preferred_nid != NUMA_NO_NODE);
1214 rq->nr_preferred_running -= (p->numa_preferred_nid == task_node(p));
1217 /* Shared or private faults. */
1218 #define NR_NUMA_HINT_FAULT_TYPES 2
1220 /* Memory and CPU locality */
1221 #define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2)
1223 /* Averaged statistics, and temporary buffers. */
1224 #define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2)
1226 pid_t task_numa_group_id(struct task_struct *p)
1228 struct numa_group *ng;
1232 ng = rcu_dereference(p->numa_group);
1241 * The averaged statistics, shared & private, memory & CPU,
1242 * occupy the first half of the array. The second half of the
1243 * array is for current counters, which are averaged into the
1244 * first set by task_numa_placement.
1246 static inline int task_faults_idx(enum numa_faults_stats s, int nid, int priv)
1248 return NR_NUMA_HINT_FAULT_TYPES * (s * nr_node_ids + nid) + priv;
1251 static inline unsigned long task_faults(struct task_struct *p, int nid)
1253 if (!p->numa_faults)
1256 return p->numa_faults[task_faults_idx(NUMA_MEM, nid, 0)] +
1257 p->numa_faults[task_faults_idx(NUMA_MEM, nid, 1)];
1260 static inline unsigned long group_faults(struct task_struct *p, int nid)
1262 struct numa_group *ng = deref_task_numa_group(p);
1267 return ng->faults[task_faults_idx(NUMA_MEM, nid, 0)] +
1268 ng->faults[task_faults_idx(NUMA_MEM, nid, 1)];
1271 static inline unsigned long group_faults_cpu(struct numa_group *group, int nid)
1273 return group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 0)] +
1274 group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 1)];
1277 static inline unsigned long group_faults_priv(struct numa_group *ng)
1279 unsigned long faults = 0;
1282 for_each_online_node(node) {
1283 faults += ng->faults[task_faults_idx(NUMA_MEM, node, 1)];
1289 static inline unsigned long group_faults_shared(struct numa_group *ng)
1291 unsigned long faults = 0;
1294 for_each_online_node(node) {
1295 faults += ng->faults[task_faults_idx(NUMA_MEM, node, 0)];
1302 * A node triggering more than 1/3 as many NUMA faults as the maximum is
1303 * considered part of a numa group's pseudo-interleaving set. Migrations
1304 * between these nodes are slowed down, to allow things to settle down.
1306 #define ACTIVE_NODE_FRACTION 3
1308 static bool numa_is_active_node(int nid, struct numa_group *ng)
1310 return group_faults_cpu(ng, nid) * ACTIVE_NODE_FRACTION > ng->max_faults_cpu;
1313 /* Handle placement on systems where not all nodes are directly connected. */
1314 static unsigned long score_nearby_nodes(struct task_struct *p, int nid,
1315 int maxdist, bool task)
1317 unsigned long score = 0;
1321 * All nodes are directly connected, and the same distance
1322 * from each other. No need for fancy placement algorithms.
1324 if (sched_numa_topology_type == NUMA_DIRECT)
1328 * This code is called for each node, introducing N^2 complexity,
1329 * which should be ok given the number of nodes rarely exceeds 8.
1331 for_each_online_node(node) {
1332 unsigned long faults;
1333 int dist = node_distance(nid, node);
1336 * The furthest away nodes in the system are not interesting
1337 * for placement; nid was already counted.
1339 if (dist == sched_max_numa_distance || node == nid)
1343 * On systems with a backplane NUMA topology, compare groups
1344 * of nodes, and move tasks towards the group with the most
1345 * memory accesses. When comparing two nodes at distance
1346 * "hoplimit", only nodes closer by than "hoplimit" are part
1347 * of each group. Skip other nodes.
1349 if (sched_numa_topology_type == NUMA_BACKPLANE &&
1353 /* Add up the faults from nearby nodes. */
1355 faults = task_faults(p, node);
1357 faults = group_faults(p, node);
1360 * On systems with a glueless mesh NUMA topology, there are
1361 * no fixed "groups of nodes". Instead, nodes that are not
1362 * directly connected bounce traffic through intermediate
1363 * nodes; a numa_group can occupy any set of nodes.
1364 * The further away a node is, the less the faults count.
1365 * This seems to result in good task placement.
1367 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
1368 faults *= (sched_max_numa_distance - dist);
1369 faults /= (sched_max_numa_distance - LOCAL_DISTANCE);
1379 * These return the fraction of accesses done by a particular task, or
1380 * task group, on a particular numa node. The group weight is given a
1381 * larger multiplier, in order to group tasks together that are almost
1382 * evenly spread out between numa nodes.
1384 static inline unsigned long task_weight(struct task_struct *p, int nid,
1387 unsigned long faults, total_faults;
1389 if (!p->numa_faults)
1392 total_faults = p->total_numa_faults;
1397 faults = task_faults(p, nid);
1398 faults += score_nearby_nodes(p, nid, dist, true);
1400 return 1000 * faults / total_faults;
1403 static inline unsigned long group_weight(struct task_struct *p, int nid,
1406 struct numa_group *ng = deref_task_numa_group(p);
1407 unsigned long faults, total_faults;
1412 total_faults = ng->total_faults;
1417 faults = group_faults(p, nid);
1418 faults += score_nearby_nodes(p, nid, dist, false);
1420 return 1000 * faults / total_faults;
1423 bool should_numa_migrate_memory(struct task_struct *p, struct page * page,
1424 int src_nid, int dst_cpu)
1426 struct numa_group *ng = deref_curr_numa_group(p);
1427 int dst_nid = cpu_to_node(dst_cpu);
1428 int last_cpupid, this_cpupid;
1430 this_cpupid = cpu_pid_to_cpupid(dst_cpu, current->pid);
1431 last_cpupid = page_cpupid_xchg_last(page, this_cpupid);
1434 * Allow first faults or private faults to migrate immediately early in
1435 * the lifetime of a task. The magic number 4 is based on waiting for
1436 * two full passes of the "multi-stage node selection" test that is
1439 if ((p->numa_preferred_nid == NUMA_NO_NODE || p->numa_scan_seq <= 4) &&
1440 (cpupid_pid_unset(last_cpupid) || cpupid_match_pid(p, last_cpupid)))
1444 * Multi-stage node selection is used in conjunction with a periodic
1445 * migration fault to build a temporal task<->page relation. By using
1446 * a two-stage filter we remove short/unlikely relations.
1448 * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate
1449 * a task's usage of a particular page (n_p) per total usage of this
1450 * page (n_t) (in a given time-span) to a probability.
1452 * Our periodic faults will sample this probability and getting the
1453 * same result twice in a row, given these samples are fully
1454 * independent, is then given by P(n)^2, provided our sample period
1455 * is sufficiently short compared to the usage pattern.
1457 * This quadric squishes small probabilities, making it less likely we
1458 * act on an unlikely task<->page relation.
1460 if (!cpupid_pid_unset(last_cpupid) &&
1461 cpupid_to_nid(last_cpupid) != dst_nid)
1464 /* Always allow migrate on private faults */
1465 if (cpupid_match_pid(p, last_cpupid))
1468 /* A shared fault, but p->numa_group has not been set up yet. */
1473 * Destination node is much more heavily used than the source
1474 * node? Allow migration.
1476 if (group_faults_cpu(ng, dst_nid) > group_faults_cpu(ng, src_nid) *
1477 ACTIVE_NODE_FRACTION)
1481 * Distribute memory according to CPU & memory use on each node,
1482 * with 3/4 hysteresis to avoid unnecessary memory migrations:
1484 * faults_cpu(dst) 3 faults_cpu(src)
1485 * --------------- * - > ---------------
1486 * faults_mem(dst) 4 faults_mem(src)
1488 return group_faults_cpu(ng, dst_nid) * group_faults(p, src_nid) * 3 >
1489 group_faults_cpu(ng, src_nid) * group_faults(p, dst_nid) * 4;
1493 * 'numa_type' describes the node at the moment of load balancing.
1496 /* The node has spare capacity that can be used to run more tasks. */
1499 * The node is fully used and the tasks don't compete for more CPU
1500 * cycles. Nevertheless, some tasks might wait before running.
1504 * The node is overloaded and can't provide expected CPU cycles to all
1510 /* Cached statistics for all CPUs within a node */
1513 unsigned long runnable;
1515 /* Total compute capacity of CPUs on a node */
1516 unsigned long compute_capacity;
1517 unsigned int nr_running;
1518 unsigned int weight;
1519 enum numa_type node_type;
1523 static inline bool is_core_idle(int cpu)
1525 #ifdef CONFIG_SCHED_SMT
1528 for_each_cpu(sibling, cpu_smt_mask(cpu)) {
1540 struct task_numa_env {
1541 struct task_struct *p;
1543 int src_cpu, src_nid;
1544 int dst_cpu, dst_nid;
1546 struct numa_stats src_stats, dst_stats;
1551 struct task_struct *best_task;
1556 static unsigned long cpu_load(struct rq *rq);
1557 static unsigned long cpu_runnable(struct rq *rq);
1558 static unsigned long cpu_util(int cpu);
1559 static inline long adjust_numa_imbalance(int imbalance,
1560 int dst_running, int dst_weight);
1563 numa_type numa_classify(unsigned int imbalance_pct,
1564 struct numa_stats *ns)
1566 if ((ns->nr_running > ns->weight) &&
1567 (((ns->compute_capacity * 100) < (ns->util * imbalance_pct)) ||
1568 ((ns->compute_capacity * imbalance_pct) < (ns->runnable * 100))))
1569 return node_overloaded;
1571 if ((ns->nr_running < ns->weight) ||
1572 (((ns->compute_capacity * 100) > (ns->util * imbalance_pct)) &&
1573 ((ns->compute_capacity * imbalance_pct) > (ns->runnable * 100))))
1574 return node_has_spare;
1576 return node_fully_busy;
1579 #ifdef CONFIG_SCHED_SMT
1580 /* Forward declarations of select_idle_sibling helpers */
1581 static inline bool test_idle_cores(int cpu, bool def);
1582 static inline int numa_idle_core(int idle_core, int cpu)
1584 if (!static_branch_likely(&sched_smt_present) ||
1585 idle_core >= 0 || !test_idle_cores(cpu, false))
1589 * Prefer cores instead of packing HT siblings
1590 * and triggering future load balancing.
1592 if (is_core_idle(cpu))
1598 static inline int numa_idle_core(int idle_core, int cpu)
1605 * Gather all necessary information to make NUMA balancing placement
1606 * decisions that are compatible with standard load balancer. This
1607 * borrows code and logic from update_sg_lb_stats but sharing a
1608 * common implementation is impractical.
1610 static void update_numa_stats(struct task_numa_env *env,
1611 struct numa_stats *ns, int nid,
1614 int cpu, idle_core = -1;
1616 memset(ns, 0, sizeof(*ns));
1620 for_each_cpu(cpu, cpumask_of_node(nid)) {
1621 struct rq *rq = cpu_rq(cpu);
1623 ns->load += cpu_load(rq);
1624 ns->runnable += cpu_runnable(rq);
1625 ns->util += cpu_util(cpu);
1626 ns->nr_running += rq->cfs.h_nr_running;
1627 ns->compute_capacity += capacity_of(cpu);
1629 if (find_idle && !rq->nr_running && idle_cpu(cpu)) {
1630 if (READ_ONCE(rq->numa_migrate_on) ||
1631 !cpumask_test_cpu(cpu, env->p->cpus_ptr))
1634 if (ns->idle_cpu == -1)
1637 idle_core = numa_idle_core(idle_core, cpu);
1642 ns->weight = cpumask_weight(cpumask_of_node(nid));
1644 ns->node_type = numa_classify(env->imbalance_pct, ns);
1647 ns->idle_cpu = idle_core;
1650 static void task_numa_assign(struct task_numa_env *env,
1651 struct task_struct *p, long imp)
1653 struct rq *rq = cpu_rq(env->dst_cpu);
1655 /* Check if run-queue part of active NUMA balance. */
1656 if (env->best_cpu != env->dst_cpu && xchg(&rq->numa_migrate_on, 1)) {
1658 int start = env->dst_cpu;
1660 /* Find alternative idle CPU. */
1661 for_each_cpu_wrap(cpu, cpumask_of_node(env->dst_nid), start) {
1662 if (cpu == env->best_cpu || !idle_cpu(cpu) ||
1663 !cpumask_test_cpu(cpu, env->p->cpus_ptr)) {
1668 rq = cpu_rq(env->dst_cpu);
1669 if (!xchg(&rq->numa_migrate_on, 1))
1673 /* Failed to find an alternative idle CPU */
1679 * Clear previous best_cpu/rq numa-migrate flag, since task now
1680 * found a better CPU to move/swap.
1682 if (env->best_cpu != -1 && env->best_cpu != env->dst_cpu) {
1683 rq = cpu_rq(env->best_cpu);
1684 WRITE_ONCE(rq->numa_migrate_on, 0);
1688 put_task_struct(env->best_task);
1693 env->best_imp = imp;
1694 env->best_cpu = env->dst_cpu;
1697 static bool load_too_imbalanced(long src_load, long dst_load,
1698 struct task_numa_env *env)
1701 long orig_src_load, orig_dst_load;
1702 long src_capacity, dst_capacity;
1705 * The load is corrected for the CPU capacity available on each node.
1708 * ------------ vs ---------
1709 * src_capacity dst_capacity
1711 src_capacity = env->src_stats.compute_capacity;
1712 dst_capacity = env->dst_stats.compute_capacity;
1714 imb = abs(dst_load * src_capacity - src_load * dst_capacity);
1716 orig_src_load = env->src_stats.load;
1717 orig_dst_load = env->dst_stats.load;
1719 old_imb = abs(orig_dst_load * src_capacity - orig_src_load * dst_capacity);
1721 /* Would this change make things worse? */
1722 return (imb > old_imb);
1726 * Maximum NUMA importance can be 1998 (2*999);
1727 * SMALLIMP @ 30 would be close to 1998/64.
1728 * Used to deter task migration.
1733 * This checks if the overall compute and NUMA accesses of the system would
1734 * be improved if the source tasks was migrated to the target dst_cpu taking
1735 * into account that it might be best if task running on the dst_cpu should
1736 * be exchanged with the source task
1738 static bool task_numa_compare(struct task_numa_env *env,
1739 long taskimp, long groupimp, bool maymove)
1741 struct numa_group *cur_ng, *p_ng = deref_curr_numa_group(env->p);
1742 struct rq *dst_rq = cpu_rq(env->dst_cpu);
1743 long imp = p_ng ? groupimp : taskimp;
1744 struct task_struct *cur;
1745 long src_load, dst_load;
1746 int dist = env->dist;
1749 bool stopsearch = false;
1751 if (READ_ONCE(dst_rq->numa_migrate_on))
1755 cur = rcu_dereference(dst_rq->curr);
1756 if (cur && ((cur->flags & PF_EXITING) || is_idle_task(cur)))
1760 * Because we have preemption enabled we can get migrated around and
1761 * end try selecting ourselves (current == env->p) as a swap candidate.
1763 if (cur == env->p) {
1769 if (maymove && moveimp >= env->best_imp)
1775 /* Skip this swap candidate if cannot move to the source cpu. */
1776 if (!cpumask_test_cpu(env->src_cpu, cur->cpus_ptr))
1780 * Skip this swap candidate if it is not moving to its preferred
1781 * node and the best task is.
1783 if (env->best_task &&
1784 env->best_task->numa_preferred_nid == env->src_nid &&
1785 cur->numa_preferred_nid != env->src_nid) {
1790 * "imp" is the fault differential for the source task between the
1791 * source and destination node. Calculate the total differential for
1792 * the source task and potential destination task. The more negative
1793 * the value is, the more remote accesses that would be expected to
1794 * be incurred if the tasks were swapped.
1796 * If dst and source tasks are in the same NUMA group, or not
1797 * in any group then look only at task weights.
1799 cur_ng = rcu_dereference(cur->numa_group);
1800 if (cur_ng == p_ng) {
1801 imp = taskimp + task_weight(cur, env->src_nid, dist) -
1802 task_weight(cur, env->dst_nid, dist);
1804 * Add some hysteresis to prevent swapping the
1805 * tasks within a group over tiny differences.
1811 * Compare the group weights. If a task is all by itself
1812 * (not part of a group), use the task weight instead.
1815 imp += group_weight(cur, env->src_nid, dist) -
1816 group_weight(cur, env->dst_nid, dist);
1818 imp += task_weight(cur, env->src_nid, dist) -
1819 task_weight(cur, env->dst_nid, dist);
1822 /* Discourage picking a task already on its preferred node */
1823 if (cur->numa_preferred_nid == env->dst_nid)
1827 * Encourage picking a task that moves to its preferred node.
1828 * This potentially makes imp larger than it's maximum of
1829 * 1998 (see SMALLIMP and task_weight for why) but in this
1830 * case, it does not matter.
1832 if (cur->numa_preferred_nid == env->src_nid)
1835 if (maymove && moveimp > imp && moveimp > env->best_imp) {
1842 * Prefer swapping with a task moving to its preferred node over a
1845 if (env->best_task && cur->numa_preferred_nid == env->src_nid &&
1846 env->best_task->numa_preferred_nid != env->src_nid) {
1851 * If the NUMA importance is less than SMALLIMP,
1852 * task migration might only result in ping pong
1853 * of tasks and also hurt performance due to cache
1856 if (imp < SMALLIMP || imp <= env->best_imp + SMALLIMP / 2)
1860 * In the overloaded case, try and keep the load balanced.
1862 load = task_h_load(env->p) - task_h_load(cur);
1866 dst_load = env->dst_stats.load + load;
1867 src_load = env->src_stats.load - load;
1869 if (load_too_imbalanced(src_load, dst_load, env))
1873 /* Evaluate an idle CPU for a task numa move. */
1875 int cpu = env->dst_stats.idle_cpu;
1877 /* Nothing cached so current CPU went idle since the search. */
1882 * If the CPU is no longer truly idle and the previous best CPU
1883 * is, keep using it.
1885 if (!idle_cpu(cpu) && env->best_cpu >= 0 &&
1886 idle_cpu(env->best_cpu)) {
1887 cpu = env->best_cpu;
1893 task_numa_assign(env, cur, imp);
1896 * If a move to idle is allowed because there is capacity or load
1897 * balance improves then stop the search. While a better swap
1898 * candidate may exist, a search is not free.
1900 if (maymove && !cur && env->best_cpu >= 0 && idle_cpu(env->best_cpu))
1904 * If a swap candidate must be identified and the current best task
1905 * moves its preferred node then stop the search.
1907 if (!maymove && env->best_task &&
1908 env->best_task->numa_preferred_nid == env->src_nid) {
1917 static void task_numa_find_cpu(struct task_numa_env *env,
1918 long taskimp, long groupimp)
1920 bool maymove = false;
1924 * If dst node has spare capacity, then check if there is an
1925 * imbalance that would be overruled by the load balancer.
1927 if (env->dst_stats.node_type == node_has_spare) {
1928 unsigned int imbalance;
1929 int src_running, dst_running;
1932 * Would movement cause an imbalance? Note that if src has
1933 * more running tasks that the imbalance is ignored as the
1934 * move improves the imbalance from the perspective of the
1935 * CPU load balancer.
1937 src_running = env->src_stats.nr_running - 1;
1938 dst_running = env->dst_stats.nr_running + 1;
1939 imbalance = max(0, dst_running - src_running);
1940 imbalance = adjust_numa_imbalance(imbalance, dst_running,
1941 env->dst_stats.weight);
1943 /* Use idle CPU if there is no imbalance */
1946 if (env->dst_stats.idle_cpu >= 0) {
1947 env->dst_cpu = env->dst_stats.idle_cpu;
1948 task_numa_assign(env, NULL, 0);
1953 long src_load, dst_load, load;
1955 * If the improvement from just moving env->p direction is better
1956 * than swapping tasks around, check if a move is possible.
1958 load = task_h_load(env->p);
1959 dst_load = env->dst_stats.load + load;
1960 src_load = env->src_stats.load - load;
1961 maymove = !load_too_imbalanced(src_load, dst_load, env);
1964 for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) {
1965 /* Skip this CPU if the source task cannot migrate */
1966 if (!cpumask_test_cpu(cpu, env->p->cpus_ptr))
1970 if (task_numa_compare(env, taskimp, groupimp, maymove))
1975 static int task_numa_migrate(struct task_struct *p)
1977 struct task_numa_env env = {
1980 .src_cpu = task_cpu(p),
1981 .src_nid = task_node(p),
1983 .imbalance_pct = 112,
1989 unsigned long taskweight, groupweight;
1990 struct sched_domain *sd;
1991 long taskimp, groupimp;
1992 struct numa_group *ng;
1997 * Pick the lowest SD_NUMA domain, as that would have the smallest
1998 * imbalance and would be the first to start moving tasks about.
2000 * And we want to avoid any moving of tasks about, as that would create
2001 * random movement of tasks -- counter the numa conditions we're trying
2005 sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu));
2007 env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2;
2011 * Cpusets can break the scheduler domain tree into smaller
2012 * balance domains, some of which do not cross NUMA boundaries.
2013 * Tasks that are "trapped" in such domains cannot be migrated
2014 * elsewhere, so there is no point in (re)trying.
2016 if (unlikely(!sd)) {
2017 sched_setnuma(p, task_node(p));
2021 env.dst_nid = p->numa_preferred_nid;
2022 dist = env.dist = node_distance(env.src_nid, env.dst_nid);
2023 taskweight = task_weight(p, env.src_nid, dist);
2024 groupweight = group_weight(p, env.src_nid, dist);
2025 update_numa_stats(&env, &env.src_stats, env.src_nid, false);
2026 taskimp = task_weight(p, env.dst_nid, dist) - taskweight;
2027 groupimp = group_weight(p, env.dst_nid, dist) - groupweight;
2028 update_numa_stats(&env, &env.dst_stats, env.dst_nid, true);
2030 /* Try to find a spot on the preferred nid. */
2031 task_numa_find_cpu(&env, taskimp, groupimp);
2034 * Look at other nodes in these cases:
2035 * - there is no space available on the preferred_nid
2036 * - the task is part of a numa_group that is interleaved across
2037 * multiple NUMA nodes; in order to better consolidate the group,
2038 * we need to check other locations.
2040 ng = deref_curr_numa_group(p);
2041 if (env.best_cpu == -1 || (ng && ng->active_nodes > 1)) {
2042 for_each_online_node(nid) {
2043 if (nid == env.src_nid || nid == p->numa_preferred_nid)
2046 dist = node_distance(env.src_nid, env.dst_nid);
2047 if (sched_numa_topology_type == NUMA_BACKPLANE &&
2049 taskweight = task_weight(p, env.src_nid, dist);
2050 groupweight = group_weight(p, env.src_nid, dist);
2053 /* Only consider nodes where both task and groups benefit */
2054 taskimp = task_weight(p, nid, dist) - taskweight;
2055 groupimp = group_weight(p, nid, dist) - groupweight;
2056 if (taskimp < 0 && groupimp < 0)
2061 update_numa_stats(&env, &env.dst_stats, env.dst_nid, true);
2062 task_numa_find_cpu(&env, taskimp, groupimp);
2067 * If the task is part of a workload that spans multiple NUMA nodes,
2068 * and is migrating into one of the workload's active nodes, remember
2069 * this node as the task's preferred numa node, so the workload can
2071 * A task that migrated to a second choice node will be better off
2072 * trying for a better one later. Do not set the preferred node here.
2075 if (env.best_cpu == -1)
2078 nid = cpu_to_node(env.best_cpu);
2080 if (nid != p->numa_preferred_nid)
2081 sched_setnuma(p, nid);
2084 /* No better CPU than the current one was found. */
2085 if (env.best_cpu == -1) {
2086 trace_sched_stick_numa(p, env.src_cpu, NULL, -1);
2090 best_rq = cpu_rq(env.best_cpu);
2091 if (env.best_task == NULL) {
2092 ret = migrate_task_to(p, env.best_cpu);
2093 WRITE_ONCE(best_rq->numa_migrate_on, 0);
2095 trace_sched_stick_numa(p, env.src_cpu, NULL, env.best_cpu);
2099 ret = migrate_swap(p, env.best_task, env.best_cpu, env.src_cpu);
2100 WRITE_ONCE(best_rq->numa_migrate_on, 0);
2103 trace_sched_stick_numa(p, env.src_cpu, env.best_task, env.best_cpu);
2104 put_task_struct(env.best_task);
2108 /* Attempt to migrate a task to a CPU on the preferred node. */
2109 static void numa_migrate_preferred(struct task_struct *p)
2111 unsigned long interval = HZ;
2113 /* This task has no NUMA fault statistics yet */
2114 if (unlikely(p->numa_preferred_nid == NUMA_NO_NODE || !p->numa_faults))
2117 /* Periodically retry migrating the task to the preferred node */
2118 interval = min(interval, msecs_to_jiffies(p->numa_scan_period) / 16);
2119 p->numa_migrate_retry = jiffies + interval;
2121 /* Success if task is already running on preferred CPU */
2122 if (task_node(p) == p->numa_preferred_nid)
2125 /* Otherwise, try migrate to a CPU on the preferred node */
2126 task_numa_migrate(p);
2130 * Find out how many nodes on the workload is actively running on. Do this by
2131 * tracking the nodes from which NUMA hinting faults are triggered. This can
2132 * be different from the set of nodes where the workload's memory is currently
2135 static void numa_group_count_active_nodes(struct numa_group *numa_group)
2137 unsigned long faults, max_faults = 0;
2138 int nid, active_nodes = 0;
2140 for_each_online_node(nid) {
2141 faults = group_faults_cpu(numa_group, nid);
2142 if (faults > max_faults)
2143 max_faults = faults;
2146 for_each_online_node(nid) {
2147 faults = group_faults_cpu(numa_group, nid);
2148 if (faults * ACTIVE_NODE_FRACTION > max_faults)
2152 numa_group->max_faults_cpu = max_faults;
2153 numa_group->active_nodes = active_nodes;
2157 * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
2158 * increments. The more local the fault statistics are, the higher the scan
2159 * period will be for the next scan window. If local/(local+remote) ratio is
2160 * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS)
2161 * the scan period will decrease. Aim for 70% local accesses.
2163 #define NUMA_PERIOD_SLOTS 10
2164 #define NUMA_PERIOD_THRESHOLD 7
2167 * Increase the scan period (slow down scanning) if the majority of
2168 * our memory is already on our local node, or if the majority of
2169 * the page accesses are shared with other processes.
2170 * Otherwise, decrease the scan period.
2172 static void update_task_scan_period(struct task_struct *p,
2173 unsigned long shared, unsigned long private)
2175 unsigned int period_slot;
2176 int lr_ratio, ps_ratio;
2179 unsigned long remote = p->numa_faults_locality[0];
2180 unsigned long local = p->numa_faults_locality[1];
2183 * If there were no record hinting faults then either the task is
2184 * completely idle or all activity is areas that are not of interest
2185 * to automatic numa balancing. Related to that, if there were failed
2186 * migration then it implies we are migrating too quickly or the local
2187 * node is overloaded. In either case, scan slower
2189 if (local + shared == 0 || p->numa_faults_locality[2]) {
2190 p->numa_scan_period = min(p->numa_scan_period_max,
2191 p->numa_scan_period << 1);
2193 p->mm->numa_next_scan = jiffies +
2194 msecs_to_jiffies(p->numa_scan_period);
2200 * Prepare to scale scan period relative to the current period.
2201 * == NUMA_PERIOD_THRESHOLD scan period stays the same
2202 * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
2203 * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
2205 period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS);
2206 lr_ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote);
2207 ps_ratio = (private * NUMA_PERIOD_SLOTS) / (private + shared);
2209 if (ps_ratio >= NUMA_PERIOD_THRESHOLD) {
2211 * Most memory accesses are local. There is no need to
2212 * do fast NUMA scanning, since memory is already local.
2214 int slot = ps_ratio - NUMA_PERIOD_THRESHOLD;
2217 diff = slot * period_slot;
2218 } else if (lr_ratio >= NUMA_PERIOD_THRESHOLD) {
2220 * Most memory accesses are shared with other tasks.
2221 * There is no point in continuing fast NUMA scanning,
2222 * since other tasks may just move the memory elsewhere.
2224 int slot = lr_ratio - NUMA_PERIOD_THRESHOLD;
2227 diff = slot * period_slot;
2230 * Private memory faults exceed (SLOTS-THRESHOLD)/SLOTS,
2231 * yet they are not on the local NUMA node. Speed up
2232 * NUMA scanning to get the memory moved over.
2234 int ratio = max(lr_ratio, ps_ratio);
2235 diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot;
2238 p->numa_scan_period = clamp(p->numa_scan_period + diff,
2239 task_scan_min(p), task_scan_max(p));
2240 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
2244 * Get the fraction of time the task has been running since the last
2245 * NUMA placement cycle. The scheduler keeps similar statistics, but
2246 * decays those on a 32ms period, which is orders of magnitude off
2247 * from the dozens-of-seconds NUMA balancing period. Use the scheduler
2248 * stats only if the task is so new there are no NUMA statistics yet.
2250 static u64 numa_get_avg_runtime(struct task_struct *p, u64 *period)
2252 u64 runtime, delta, now;
2253 /* Use the start of this time slice to avoid calculations. */
2254 now = p->se.exec_start;
2255 runtime = p->se.sum_exec_runtime;
2257 if (p->last_task_numa_placement) {
2258 delta = runtime - p->last_sum_exec_runtime;
2259 *period = now - p->last_task_numa_placement;
2261 /* Avoid time going backwards, prevent potential divide error: */
2262 if (unlikely((s64)*period < 0))
2265 delta = p->se.avg.load_sum;
2266 *period = LOAD_AVG_MAX;
2269 p->last_sum_exec_runtime = runtime;
2270 p->last_task_numa_placement = now;
2276 * Determine the preferred nid for a task in a numa_group. This needs to
2277 * be done in a way that produces consistent results with group_weight,
2278 * otherwise workloads might not converge.
2280 static int preferred_group_nid(struct task_struct *p, int nid)
2285 /* Direct connections between all NUMA nodes. */
2286 if (sched_numa_topology_type == NUMA_DIRECT)
2290 * On a system with glueless mesh NUMA topology, group_weight
2291 * scores nodes according to the number of NUMA hinting faults on
2292 * both the node itself, and on nearby nodes.
2294 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
2295 unsigned long score, max_score = 0;
2296 int node, max_node = nid;
2298 dist = sched_max_numa_distance;
2300 for_each_online_node(node) {
2301 score = group_weight(p, node, dist);
2302 if (score > max_score) {
2311 * Finding the preferred nid in a system with NUMA backplane
2312 * interconnect topology is more involved. The goal is to locate
2313 * tasks from numa_groups near each other in the system, and
2314 * untangle workloads from different sides of the system. This requires
2315 * searching down the hierarchy of node groups, recursively searching
2316 * inside the highest scoring group of nodes. The nodemask tricks
2317 * keep the complexity of the search down.
2319 nodes = node_online_map;
2320 for (dist = sched_max_numa_distance; dist > LOCAL_DISTANCE; dist--) {
2321 unsigned long max_faults = 0;
2322 nodemask_t max_group = NODE_MASK_NONE;
2325 /* Are there nodes at this distance from each other? */
2326 if (!find_numa_distance(dist))
2329 for_each_node_mask(a, nodes) {
2330 unsigned long faults = 0;
2331 nodemask_t this_group;
2332 nodes_clear(this_group);
2334 /* Sum group's NUMA faults; includes a==b case. */
2335 for_each_node_mask(b, nodes) {
2336 if (node_distance(a, b) < dist) {
2337 faults += group_faults(p, b);
2338 node_set(b, this_group);
2339 node_clear(b, nodes);
2343 /* Remember the top group. */
2344 if (faults > max_faults) {
2345 max_faults = faults;
2346 max_group = this_group;
2348 * subtle: at the smallest distance there is
2349 * just one node left in each "group", the
2350 * winner is the preferred nid.
2355 /* Next round, evaluate the nodes within max_group. */
2363 static void task_numa_placement(struct task_struct *p)
2365 int seq, nid, max_nid = NUMA_NO_NODE;
2366 unsigned long max_faults = 0;
2367 unsigned long fault_types[2] = { 0, 0 };
2368 unsigned long total_faults;
2369 u64 runtime, period;
2370 spinlock_t *group_lock = NULL;
2371 struct numa_group *ng;
2374 * The p->mm->numa_scan_seq field gets updated without
2375 * exclusive access. Use READ_ONCE() here to ensure
2376 * that the field is read in a single access:
2378 seq = READ_ONCE(p->mm->numa_scan_seq);
2379 if (p->numa_scan_seq == seq)
2381 p->numa_scan_seq = seq;
2382 p->numa_scan_period_max = task_scan_max(p);
2384 total_faults = p->numa_faults_locality[0] +
2385 p->numa_faults_locality[1];
2386 runtime = numa_get_avg_runtime(p, &period);
2388 /* If the task is part of a group prevent parallel updates to group stats */
2389 ng = deref_curr_numa_group(p);
2391 group_lock = &ng->lock;
2392 spin_lock_irq(group_lock);
2395 /* Find the node with the highest number of faults */
2396 for_each_online_node(nid) {
2397 /* Keep track of the offsets in numa_faults array */
2398 int mem_idx, membuf_idx, cpu_idx, cpubuf_idx;
2399 unsigned long faults = 0, group_faults = 0;
2402 for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) {
2403 long diff, f_diff, f_weight;
2405 mem_idx = task_faults_idx(NUMA_MEM, nid, priv);
2406 membuf_idx = task_faults_idx(NUMA_MEMBUF, nid, priv);
2407 cpu_idx = task_faults_idx(NUMA_CPU, nid, priv);
2408 cpubuf_idx = task_faults_idx(NUMA_CPUBUF, nid, priv);
2410 /* Decay existing window, copy faults since last scan */
2411 diff = p->numa_faults[membuf_idx] - p->numa_faults[mem_idx] / 2;
2412 fault_types[priv] += p->numa_faults[membuf_idx];
2413 p->numa_faults[membuf_idx] = 0;
2416 * Normalize the faults_from, so all tasks in a group
2417 * count according to CPU use, instead of by the raw
2418 * number of faults. Tasks with little runtime have
2419 * little over-all impact on throughput, and thus their
2420 * faults are less important.
2422 f_weight = div64_u64(runtime << 16, period + 1);
2423 f_weight = (f_weight * p->numa_faults[cpubuf_idx]) /
2425 f_diff = f_weight - p->numa_faults[cpu_idx] / 2;
2426 p->numa_faults[cpubuf_idx] = 0;
2428 p->numa_faults[mem_idx] += diff;
2429 p->numa_faults[cpu_idx] += f_diff;
2430 faults += p->numa_faults[mem_idx];
2431 p->total_numa_faults += diff;
2434 * safe because we can only change our own group
2436 * mem_idx represents the offset for a given
2437 * nid and priv in a specific region because it
2438 * is at the beginning of the numa_faults array.
2440 ng->faults[mem_idx] += diff;
2441 ng->faults_cpu[mem_idx] += f_diff;
2442 ng->total_faults += diff;
2443 group_faults += ng->faults[mem_idx];
2448 if (faults > max_faults) {
2449 max_faults = faults;
2452 } else if (group_faults > max_faults) {
2453 max_faults = group_faults;
2459 numa_group_count_active_nodes(ng);
2460 spin_unlock_irq(group_lock);
2461 max_nid = preferred_group_nid(p, max_nid);
2465 /* Set the new preferred node */
2466 if (max_nid != p->numa_preferred_nid)
2467 sched_setnuma(p, max_nid);
2470 update_task_scan_period(p, fault_types[0], fault_types[1]);
2473 static inline int get_numa_group(struct numa_group *grp)
2475 return refcount_inc_not_zero(&grp->refcount);
2478 static inline void put_numa_group(struct numa_group *grp)
2480 if (refcount_dec_and_test(&grp->refcount))
2481 kfree_rcu(grp, rcu);
2484 static void task_numa_group(struct task_struct *p, int cpupid, int flags,
2487 struct numa_group *grp, *my_grp;
2488 struct task_struct *tsk;
2490 int cpu = cpupid_to_cpu(cpupid);
2493 if (unlikely(!deref_curr_numa_group(p))) {
2494 unsigned int size = sizeof(struct numa_group) +
2495 4*nr_node_ids*sizeof(unsigned long);
2497 grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
2501 refcount_set(&grp->refcount, 1);
2502 grp->active_nodes = 1;
2503 grp->max_faults_cpu = 0;
2504 spin_lock_init(&grp->lock);
2506 /* Second half of the array tracks nids where faults happen */
2507 grp->faults_cpu = grp->faults + NR_NUMA_HINT_FAULT_TYPES *
2510 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2511 grp->faults[i] = p->numa_faults[i];
2513 grp->total_faults = p->total_numa_faults;
2516 rcu_assign_pointer(p->numa_group, grp);
2520 tsk = READ_ONCE(cpu_rq(cpu)->curr);
2522 if (!cpupid_match_pid(tsk, cpupid))
2525 grp = rcu_dereference(tsk->numa_group);
2529 my_grp = deref_curr_numa_group(p);
2534 * Only join the other group if its bigger; if we're the bigger group,
2535 * the other task will join us.
2537 if (my_grp->nr_tasks > grp->nr_tasks)
2541 * Tie-break on the grp address.
2543 if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp)
2546 /* Always join threads in the same process. */
2547 if (tsk->mm == current->mm)
2550 /* Simple filter to avoid false positives due to PID collisions */
2551 if (flags & TNF_SHARED)
2554 /* Update priv based on whether false sharing was detected */
2557 if (join && !get_numa_group(grp))
2565 BUG_ON(irqs_disabled());
2566 double_lock_irq(&my_grp->lock, &grp->lock);
2568 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) {
2569 my_grp->faults[i] -= p->numa_faults[i];
2570 grp->faults[i] += p->numa_faults[i];
2572 my_grp->total_faults -= p->total_numa_faults;
2573 grp->total_faults += p->total_numa_faults;
2578 spin_unlock(&my_grp->lock);
2579 spin_unlock_irq(&grp->lock);
2581 rcu_assign_pointer(p->numa_group, grp);
2583 put_numa_group(my_grp);
2592 * Get rid of NUMA statistics associated with a task (either current or dead).
2593 * If @final is set, the task is dead and has reached refcount zero, so we can
2594 * safely free all relevant data structures. Otherwise, there might be
2595 * concurrent reads from places like load balancing and procfs, and we should
2596 * reset the data back to default state without freeing ->numa_faults.
2598 void task_numa_free(struct task_struct *p, bool final)
2600 /* safe: p either is current or is being freed by current */
2601 struct numa_group *grp = rcu_dereference_raw(p->numa_group);
2602 unsigned long *numa_faults = p->numa_faults;
2603 unsigned long flags;
2610 spin_lock_irqsave(&grp->lock, flags);
2611 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2612 grp->faults[i] -= p->numa_faults[i];
2613 grp->total_faults -= p->total_numa_faults;
2616 spin_unlock_irqrestore(&grp->lock, flags);
2617 RCU_INIT_POINTER(p->numa_group, NULL);
2618 put_numa_group(grp);
2622 p->numa_faults = NULL;
2625 p->total_numa_faults = 0;
2626 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2632 * Got a PROT_NONE fault for a page on @node.
2634 void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags)
2636 struct task_struct *p = current;
2637 bool migrated = flags & TNF_MIGRATED;
2638 int cpu_node = task_node(current);
2639 int local = !!(flags & TNF_FAULT_LOCAL);
2640 struct numa_group *ng;
2643 if (!static_branch_likely(&sched_numa_balancing))
2646 /* for example, ksmd faulting in a user's mm */
2650 /* Allocate buffer to track faults on a per-node basis */
2651 if (unlikely(!p->numa_faults)) {
2652 int size = sizeof(*p->numa_faults) *
2653 NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids;
2655 p->numa_faults = kzalloc(size, GFP_KERNEL|__GFP_NOWARN);
2656 if (!p->numa_faults)
2659 p->total_numa_faults = 0;
2660 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
2664 * First accesses are treated as private, otherwise consider accesses
2665 * to be private if the accessing pid has not changed
2667 if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) {
2670 priv = cpupid_match_pid(p, last_cpupid);
2671 if (!priv && !(flags & TNF_NO_GROUP))
2672 task_numa_group(p, last_cpupid, flags, &priv);
2676 * If a workload spans multiple NUMA nodes, a shared fault that
2677 * occurs wholly within the set of nodes that the workload is
2678 * actively using should be counted as local. This allows the
2679 * scan rate to slow down when a workload has settled down.
2681 ng = deref_curr_numa_group(p);
2682 if (!priv && !local && ng && ng->active_nodes > 1 &&
2683 numa_is_active_node(cpu_node, ng) &&
2684 numa_is_active_node(mem_node, ng))
2688 * Retry to migrate task to preferred node periodically, in case it
2689 * previously failed, or the scheduler moved us.
2691 if (time_after(jiffies, p->numa_migrate_retry)) {
2692 task_numa_placement(p);
2693 numa_migrate_preferred(p);
2697 p->numa_pages_migrated += pages;
2698 if (flags & TNF_MIGRATE_FAIL)
2699 p->numa_faults_locality[2] += pages;
2701 p->numa_faults[task_faults_idx(NUMA_MEMBUF, mem_node, priv)] += pages;
2702 p->numa_faults[task_faults_idx(NUMA_CPUBUF, cpu_node, priv)] += pages;
2703 p->numa_faults_locality[local] += pages;
2706 static void reset_ptenuma_scan(struct task_struct *p)
2709 * We only did a read acquisition of the mmap sem, so
2710 * p->mm->numa_scan_seq is written to without exclusive access
2711 * and the update is not guaranteed to be atomic. That's not
2712 * much of an issue though, since this is just used for
2713 * statistical sampling. Use READ_ONCE/WRITE_ONCE, which are not
2714 * expensive, to avoid any form of compiler optimizations:
2716 WRITE_ONCE(p->mm->numa_scan_seq, READ_ONCE(p->mm->numa_scan_seq) + 1);
2717 p->mm->numa_scan_offset = 0;
2721 * The expensive part of numa migration is done from task_work context.
2722 * Triggered from task_tick_numa().
2724 static void task_numa_work(struct callback_head *work)
2726 unsigned long migrate, next_scan, now = jiffies;
2727 struct task_struct *p = current;
2728 struct mm_struct *mm = p->mm;
2729 u64 runtime = p->se.sum_exec_runtime;
2730 struct vm_area_struct *vma;
2731 unsigned long start, end;
2732 unsigned long nr_pte_updates = 0;
2733 long pages, virtpages;
2735 SCHED_WARN_ON(p != container_of(work, struct task_struct, numa_work));
2739 * Who cares about NUMA placement when they're dying.
2741 * NOTE: make sure not to dereference p->mm before this check,
2742 * exit_task_work() happens _after_ exit_mm() so we could be called
2743 * without p->mm even though we still had it when we enqueued this
2746 if (p->flags & PF_EXITING)
2749 if (!mm->numa_next_scan) {
2750 mm->numa_next_scan = now +
2751 msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2755 * Enforce maximal scan/migration frequency..
2757 migrate = mm->numa_next_scan;
2758 if (time_before(now, migrate))
2761 if (p->numa_scan_period == 0) {
2762 p->numa_scan_period_max = task_scan_max(p);
2763 p->numa_scan_period = task_scan_start(p);
2766 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
2767 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
2771 * Delay this task enough that another task of this mm will likely win
2772 * the next time around.
2774 p->node_stamp += 2 * TICK_NSEC;
2776 start = mm->numa_scan_offset;
2777 pages = sysctl_numa_balancing_scan_size;
2778 pages <<= 20 - PAGE_SHIFT; /* MB in pages */
2779 virtpages = pages * 8; /* Scan up to this much virtual space */
2784 if (!mmap_read_trylock(mm))
2786 vma = find_vma(mm, start);
2788 reset_ptenuma_scan(p);
2792 for (; vma; vma = vma->vm_next) {
2793 if (!vma_migratable(vma) || !vma_policy_mof(vma) ||
2794 is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_MIXEDMAP)) {
2799 * Shared library pages mapped by multiple processes are not
2800 * migrated as it is expected they are cache replicated. Avoid
2801 * hinting faults in read-only file-backed mappings or the vdso
2802 * as migrating the pages will be of marginal benefit.
2805 (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ)))
2809 * Skip inaccessible VMAs to avoid any confusion between
2810 * PROT_NONE and NUMA hinting ptes
2812 if (!vma_is_accessible(vma))
2816 start = max(start, vma->vm_start);
2817 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
2818 end = min(end, vma->vm_end);
2819 nr_pte_updates = change_prot_numa(vma, start, end);
2822 * Try to scan sysctl_numa_balancing_size worth of
2823 * hpages that have at least one present PTE that
2824 * is not already pte-numa. If the VMA contains
2825 * areas that are unused or already full of prot_numa
2826 * PTEs, scan up to virtpages, to skip through those
2830 pages -= (end - start) >> PAGE_SHIFT;
2831 virtpages -= (end - start) >> PAGE_SHIFT;
2834 if (pages <= 0 || virtpages <= 0)
2838 } while (end != vma->vm_end);
2843 * It is possible to reach the end of the VMA list but the last few
2844 * VMAs are not guaranteed to the vma_migratable. If they are not, we
2845 * would find the !migratable VMA on the next scan but not reset the
2846 * scanner to the start so check it now.
2849 mm->numa_scan_offset = start;
2851 reset_ptenuma_scan(p);
2852 mmap_read_unlock(mm);
2855 * Make sure tasks use at least 32x as much time to run other code
2856 * than they used here, to limit NUMA PTE scanning overhead to 3% max.
2857 * Usually update_task_scan_period slows down scanning enough; on an
2858 * overloaded system we need to limit overhead on a per task basis.
2860 if (unlikely(p->se.sum_exec_runtime != runtime)) {
2861 u64 diff = p->se.sum_exec_runtime - runtime;
2862 p->node_stamp += 32 * diff;
2866 void init_numa_balancing(unsigned long clone_flags, struct task_struct *p)
2869 struct mm_struct *mm = p->mm;
2872 mm_users = atomic_read(&mm->mm_users);
2873 if (mm_users == 1) {
2874 mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2875 mm->numa_scan_seq = 0;
2879 p->numa_scan_seq = mm ? mm->numa_scan_seq : 0;
2880 p->numa_scan_period = sysctl_numa_balancing_scan_delay;
2881 /* Protect against double add, see task_tick_numa and task_numa_work */
2882 p->numa_work.next = &p->numa_work;
2883 p->numa_faults = NULL;
2884 RCU_INIT_POINTER(p->numa_group, NULL);
2885 p->last_task_numa_placement = 0;
2886 p->last_sum_exec_runtime = 0;
2888 init_task_work(&p->numa_work, task_numa_work);
2890 /* New address space, reset the preferred nid */
2891 if (!(clone_flags & CLONE_VM)) {
2892 p->numa_preferred_nid = NUMA_NO_NODE;
2897 * New thread, keep existing numa_preferred_nid which should be copied
2898 * already by arch_dup_task_struct but stagger when scans start.
2903 delay = min_t(unsigned int, task_scan_max(current),
2904 current->numa_scan_period * mm_users * NSEC_PER_MSEC);
2905 delay += 2 * TICK_NSEC;
2906 p->node_stamp = delay;
2911 * Drive the periodic memory faults..
2913 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
2915 struct callback_head *work = &curr->numa_work;
2919 * We don't care about NUMA placement if we don't have memory.
2921 if ((curr->flags & (PF_EXITING | PF_KTHREAD)) || work->next != work)
2925 * Using runtime rather than walltime has the dual advantage that
2926 * we (mostly) drive the selection from busy threads and that the
2927 * task needs to have done some actual work before we bother with
2930 now = curr->se.sum_exec_runtime;
2931 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
2933 if (now > curr->node_stamp + period) {
2934 if (!curr->node_stamp)
2935 curr->numa_scan_period = task_scan_start(curr);
2936 curr->node_stamp += period;
2938 if (!time_before(jiffies, curr->mm->numa_next_scan))
2939 task_work_add(curr, work, TWA_RESUME);
2943 static void update_scan_period(struct task_struct *p, int new_cpu)
2945 int src_nid = cpu_to_node(task_cpu(p));
2946 int dst_nid = cpu_to_node(new_cpu);
2948 if (!static_branch_likely(&sched_numa_balancing))
2951 if (!p->mm || !p->numa_faults || (p->flags & PF_EXITING))
2954 if (src_nid == dst_nid)
2958 * Allow resets if faults have been trapped before one scan
2959 * has completed. This is most likely due to a new task that
2960 * is pulled cross-node due to wakeups or load balancing.
2962 if (p->numa_scan_seq) {
2964 * Avoid scan adjustments if moving to the preferred
2965 * node or if the task was not previously running on
2966 * the preferred node.
2968 if (dst_nid == p->numa_preferred_nid ||
2969 (p->numa_preferred_nid != NUMA_NO_NODE &&
2970 src_nid != p->numa_preferred_nid))
2974 p->numa_scan_period = task_scan_start(p);
2978 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
2982 static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p)
2986 static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p)
2990 static inline void update_scan_period(struct task_struct *p, int new_cpu)
2994 #endif /* CONFIG_NUMA_BALANCING */
2997 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2999 update_load_add(&cfs_rq->load, se->load.weight);
3001 if (entity_is_task(se)) {
3002 struct rq *rq = rq_of(cfs_rq);
3004 account_numa_enqueue(rq, task_of(se));
3005 list_add(&se->group_node, &rq->cfs_tasks);
3008 cfs_rq->nr_running++;
3012 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
3014 update_load_sub(&cfs_rq->load, se->load.weight);
3016 if (entity_is_task(se)) {
3017 account_numa_dequeue(rq_of(cfs_rq), task_of(se));
3018 list_del_init(&se->group_node);
3021 cfs_rq->nr_running--;
3025 * Signed add and clamp on underflow.
3027 * Explicitly do a load-store to ensure the intermediate value never hits
3028 * memory. This allows lockless observations without ever seeing the negative
3031 #define add_positive(_ptr, _val) do { \
3032 typeof(_ptr) ptr = (_ptr); \
3033 typeof(_val) val = (_val); \
3034 typeof(*ptr) res, var = READ_ONCE(*ptr); \
3038 if (val < 0 && res > var) \
3041 WRITE_ONCE(*ptr, res); \
3045 * Unsigned subtract and clamp on underflow.
3047 * Explicitly do a load-store to ensure the intermediate value never hits
3048 * memory. This allows lockless observations without ever seeing the negative
3051 #define sub_positive(_ptr, _val) do { \
3052 typeof(_ptr) ptr = (_ptr); \
3053 typeof(*ptr) val = (_val); \
3054 typeof(*ptr) res, var = READ_ONCE(*ptr); \
3058 WRITE_ONCE(*ptr, res); \
3062 * Remove and clamp on negative, from a local variable.
3064 * A variant of sub_positive(), which does not use explicit load-store
3065 * and is thus optimized for local variable updates.
3067 #define lsub_positive(_ptr, _val) do { \
3068 typeof(_ptr) ptr = (_ptr); \
3069 *ptr -= min_t(typeof(*ptr), *ptr, _val); \
3074 enqueue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3076 cfs_rq->avg.load_avg += se->avg.load_avg;
3077 cfs_rq->avg.load_sum += se_weight(se) * se->avg.load_sum;
3081 dequeue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3083 sub_positive(&cfs_rq->avg.load_avg, se->avg.load_avg);
3084 sub_positive(&cfs_rq->avg.load_sum, se_weight(se) * se->avg.load_sum);
3088 enqueue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
3090 dequeue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
3093 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
3094 unsigned long weight)
3097 /* commit outstanding execution time */
3098 if (cfs_rq->curr == se)
3099 update_curr(cfs_rq);
3100 update_load_sub(&cfs_rq->load, se->load.weight);
3102 dequeue_load_avg(cfs_rq, se);
3104 update_load_set(&se->load, weight);
3108 u32 divider = get_pelt_divider(&se->avg);
3110 se->avg.load_avg = div_u64(se_weight(se) * se->avg.load_sum, divider);
3114 enqueue_load_avg(cfs_rq, se);
3116 update_load_add(&cfs_rq->load, se->load.weight);
3120 void reweight_task(struct task_struct *p, int prio)
3122 struct sched_entity *se = &p->se;
3123 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3124 struct load_weight *load = &se->load;
3125 unsigned long weight = scale_load(sched_prio_to_weight[prio]);
3127 reweight_entity(cfs_rq, se, weight);
3128 load->inv_weight = sched_prio_to_wmult[prio];
3131 #ifdef CONFIG_FAIR_GROUP_SCHED
3134 * All this does is approximate the hierarchical proportion which includes that
3135 * global sum we all love to hate.
3137 * That is, the weight of a group entity, is the proportional share of the
3138 * group weight based on the group runqueue weights. That is:
3140 * tg->weight * grq->load.weight
3141 * ge->load.weight = ----------------------------- (1)
3142 * \Sum grq->load.weight
3144 * Now, because computing that sum is prohibitively expensive to compute (been
3145 * there, done that) we approximate it with this average stuff. The average
3146 * moves slower and therefore the approximation is cheaper and more stable.
3148 * So instead of the above, we substitute:
3150 * grq->load.weight -> grq->avg.load_avg (2)
3152 * which yields the following:
3154 * tg->weight * grq->avg.load_avg
3155 * ge->load.weight = ------------------------------ (3)
3158 * Where: tg->load_avg ~= \Sum grq->avg.load_avg
3160 * That is shares_avg, and it is right (given the approximation (2)).
3162 * The problem with it is that because the average is slow -- it was designed
3163 * to be exactly that of course -- this leads to transients in boundary
3164 * conditions. In specific, the case where the group was idle and we start the
3165 * one task. It takes time for our CPU's grq->avg.load_avg to build up,
3166 * yielding bad latency etc..
3168 * Now, in that special case (1) reduces to:
3170 * tg->weight * grq->load.weight
3171 * ge->load.weight = ----------------------------- = tg->weight (4)
3174 * That is, the sum collapses because all other CPUs are idle; the UP scenario.
3176 * So what we do is modify our approximation (3) to approach (4) in the (near)
3181 * tg->weight * grq->load.weight
3182 * --------------------------------------------------- (5)
3183 * tg->load_avg - grq->avg.load_avg + grq->load.weight
3185 * But because grq->load.weight can drop to 0, resulting in a divide by zero,
3186 * we need to use grq->avg.load_avg as its lower bound, which then gives:
3189 * tg->weight * grq->load.weight
3190 * ge->load.weight = ----------------------------- (6)
3195 * tg_load_avg' = tg->load_avg - grq->avg.load_avg +
3196 * max(grq->load.weight, grq->avg.load_avg)
3198 * And that is shares_weight and is icky. In the (near) UP case it approaches
3199 * (4) while in the normal case it approaches (3). It consistently
3200 * overestimates the ge->load.weight and therefore:
3202 * \Sum ge->load.weight >= tg->weight
3206 static long calc_group_shares(struct cfs_rq *cfs_rq)
3208 long tg_weight, tg_shares, load, shares;
3209 struct task_group *tg = cfs_rq->tg;
3211 tg_shares = READ_ONCE(tg->shares);
3213 load = max(scale_load_down(cfs_rq->load.weight), cfs_rq->avg.load_avg);
3215 tg_weight = atomic_long_read(&tg->load_avg);
3217 /* Ensure tg_weight >= load */
3218 tg_weight -= cfs_rq->tg_load_avg_contrib;
3221 shares = (tg_shares * load);
3223 shares /= tg_weight;
3226 * MIN_SHARES has to be unscaled here to support per-CPU partitioning
3227 * of a group with small tg->shares value. It is a floor value which is
3228 * assigned as a minimum load.weight to the sched_entity representing
3229 * the group on a CPU.
3231 * E.g. on 64-bit for a group with tg->shares of scale_load(15)=15*1024
3232 * on an 8-core system with 8 tasks each runnable on one CPU shares has
3233 * to be 15*1024*1/8=1920 instead of scale_load(MIN_SHARES)=2*1024. In
3234 * case no task is runnable on a CPU MIN_SHARES=2 should be returned
3237 return clamp_t(long, shares, MIN_SHARES, tg_shares);
3239 #endif /* CONFIG_SMP */
3241 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
3244 * Recomputes the group entity based on the current state of its group
3247 static void update_cfs_group(struct sched_entity *se)
3249 struct cfs_rq *gcfs_rq = group_cfs_rq(se);
3255 if (throttled_hierarchy(gcfs_rq))
3259 shares = READ_ONCE(gcfs_rq->tg->shares);
3261 if (likely(se->load.weight == shares))
3264 shares = calc_group_shares(gcfs_rq);
3267 reweight_entity(cfs_rq_of(se), se, shares);
3270 #else /* CONFIG_FAIR_GROUP_SCHED */
3271 static inline void update_cfs_group(struct sched_entity *se)
3274 #endif /* CONFIG_FAIR_GROUP_SCHED */
3276 static inline void cfs_rq_util_change(struct cfs_rq *cfs_rq, int flags)
3278 struct rq *rq = rq_of(cfs_rq);
3280 if (&rq->cfs == cfs_rq) {
3282 * There are a few boundary cases this might miss but it should
3283 * get called often enough that that should (hopefully) not be
3286 * It will not get called when we go idle, because the idle
3287 * thread is a different class (!fair), nor will the utilization
3288 * number include things like RT tasks.
3290 * As is, the util number is not freq-invariant (we'd have to
3291 * implement arch_scale_freq_capacity() for that).
3295 cpufreq_update_util(rq, flags);
3300 #ifdef CONFIG_FAIR_GROUP_SCHED
3302 * update_tg_load_avg - update the tg's load avg
3303 * @cfs_rq: the cfs_rq whose avg changed
3305 * This function 'ensures': tg->load_avg := \Sum tg->cfs_rq[]->avg.load.
3306 * However, because tg->load_avg is a global value there are performance
3309 * In order to avoid having to look at the other cfs_rq's, we use a
3310 * differential update where we store the last value we propagated. This in
3311 * turn allows skipping updates if the differential is 'small'.
3313 * Updating tg's load_avg is necessary before update_cfs_share().
3315 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq)
3317 long delta = cfs_rq->avg.load_avg - cfs_rq->tg_load_avg_contrib;
3320 * No need to update load_avg for root_task_group as it is not used.
3322 if (cfs_rq->tg == &root_task_group)
3325 if (abs(delta) > cfs_rq->tg_load_avg_contrib / 64) {
3326 atomic_long_add(delta, &cfs_rq->tg->load_avg);
3327 cfs_rq->tg_load_avg_contrib = cfs_rq->avg.load_avg;
3332 * Called within set_task_rq() right before setting a task's CPU. The
3333 * caller only guarantees p->pi_lock is held; no other assumptions,
3334 * including the state of rq->lock, should be made.
3336 void set_task_rq_fair(struct sched_entity *se,
3337 struct cfs_rq *prev, struct cfs_rq *next)
3339 u64 p_last_update_time;
3340 u64 n_last_update_time;
3342 if (!sched_feat(ATTACH_AGE_LOAD))
3346 * We are supposed to update the task to "current" time, then its up to
3347 * date and ready to go to new CPU/cfs_rq. But we have difficulty in
3348 * getting what current time is, so simply throw away the out-of-date
3349 * time. This will result in the wakee task is less decayed, but giving
3350 * the wakee more load sounds not bad.
3352 if (!(se->avg.last_update_time && prev))
3355 #ifndef CONFIG_64BIT
3357 u64 p_last_update_time_copy;
3358 u64 n_last_update_time_copy;
3361 p_last_update_time_copy = prev->load_last_update_time_copy;
3362 n_last_update_time_copy = next->load_last_update_time_copy;
3366 p_last_update_time = prev->avg.last_update_time;
3367 n_last_update_time = next->avg.last_update_time;
3369 } while (p_last_update_time != p_last_update_time_copy ||
3370 n_last_update_time != n_last_update_time_copy);
3373 p_last_update_time = prev->avg.last_update_time;
3374 n_last_update_time = next->avg.last_update_time;
3376 __update_load_avg_blocked_se(p_last_update_time, se);
3377 se->avg.last_update_time = n_last_update_time;
3382 * When on migration a sched_entity joins/leaves the PELT hierarchy, we need to
3383 * propagate its contribution. The key to this propagation is the invariant
3384 * that for each group:
3386 * ge->avg == grq->avg (1)
3388 * _IFF_ we look at the pure running and runnable sums. Because they
3389 * represent the very same entity, just at different points in the hierarchy.
3391 * Per the above update_tg_cfs_util() and update_tg_cfs_runnable() are trivial
3392 * and simply copies the running/runnable sum over (but still wrong, because
3393 * the group entity and group rq do not have their PELT windows aligned).
3395 * However, update_tg_cfs_load() is more complex. So we have:
3397 * ge->avg.load_avg = ge->load.weight * ge->avg.runnable_avg (2)
3399 * And since, like util, the runnable part should be directly transferable,
3400 * the following would _appear_ to be the straight forward approach:
3402 * grq->avg.load_avg = grq->load.weight * grq->avg.runnable_avg (3)
3404 * And per (1) we have:
3406 * ge->avg.runnable_avg == grq->avg.runnable_avg
3410 * ge->load.weight * grq->avg.load_avg
3411 * ge->avg.load_avg = ----------------------------------- (4)
3414 * Except that is wrong!
3416 * Because while for entities historical weight is not important and we
3417 * really only care about our future and therefore can consider a pure
3418 * runnable sum, runqueues can NOT do this.
3420 * We specifically want runqueues to have a load_avg that includes
3421 * historical weights. Those represent the blocked load, the load we expect
3422 * to (shortly) return to us. This only works by keeping the weights as
3423 * integral part of the sum. We therefore cannot decompose as per (3).
3425 * Another reason this doesn't work is that runnable isn't a 0-sum entity.
3426 * Imagine a rq with 2 tasks that each are runnable 2/3 of the time. Then the
3427 * rq itself is runnable anywhere between 2/3 and 1 depending on how the
3428 * runnable section of these tasks overlap (or not). If they were to perfectly
3429 * align the rq as a whole would be runnable 2/3 of the time. If however we
3430 * always have at least 1 runnable task, the rq as a whole is always runnable.
3432 * So we'll have to approximate.. :/
3434 * Given the constraint:
3436 * ge->avg.running_sum <= ge->avg.runnable_sum <= LOAD_AVG_MAX
3438 * We can construct a rule that adds runnable to a rq by assuming minimal
3441 * On removal, we'll assume each task is equally runnable; which yields:
3443 * grq->avg.runnable_sum = grq->avg.load_sum / grq->load.weight
3445 * XXX: only do this for the part of runnable > running ?
3450 update_tg_cfs_util(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq)
3452 long delta = gcfs_rq->avg.util_avg - se->avg.util_avg;
3455 /* Nothing to update */
3460 * cfs_rq->avg.period_contrib can be used for both cfs_rq and se.
3461 * See ___update_load_avg() for details.
3463 divider = get_pelt_divider(&cfs_rq->avg);
3465 /* Set new sched_entity's utilization */
3466 se->avg.util_avg = gcfs_rq->avg.util_avg;
3467 se->avg.util_sum = se->avg.util_avg * divider;
3469 /* Update parent cfs_rq utilization */
3470 add_positive(&cfs_rq->avg.util_avg, delta);
3471 cfs_rq->avg.util_sum = cfs_rq->avg.util_avg * divider;
3475 update_tg_cfs_runnable(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq)
3477 long delta = gcfs_rq->avg.runnable_avg - se->avg.runnable_avg;
3480 /* Nothing to update */
3485 * cfs_rq->avg.period_contrib can be used for both cfs_rq and se.
3486 * See ___update_load_avg() for details.
3488 divider = get_pelt_divider(&cfs_rq->avg);
3490 /* Set new sched_entity's runnable */
3491 se->avg.runnable_avg = gcfs_rq->avg.runnable_avg;
3492 se->avg.runnable_sum = se->avg.runnable_avg * divider;
3494 /* Update parent cfs_rq runnable */
3495 add_positive(&cfs_rq->avg.runnable_avg, delta);
3496 cfs_rq->avg.runnable_sum = cfs_rq->avg.runnable_avg * divider;
3500 update_tg_cfs_load(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq)
3502 long delta_avg, running_sum, runnable_sum = gcfs_rq->prop_runnable_sum;
3503 unsigned long load_avg;
3511 gcfs_rq->prop_runnable_sum = 0;
3514 * cfs_rq->avg.period_contrib can be used for both cfs_rq and se.
3515 * See ___update_load_avg() for details.
3517 divider = get_pelt_divider(&cfs_rq->avg);
3519 if (runnable_sum >= 0) {
3521 * Add runnable; clip at LOAD_AVG_MAX. Reflects that until
3522 * the CPU is saturated running == runnable.
3524 runnable_sum += se->avg.load_sum;
3525 runnable_sum = min_t(long, runnable_sum, divider);
3528 * Estimate the new unweighted runnable_sum of the gcfs_rq by
3529 * assuming all tasks are equally runnable.
3531 if (scale_load_down(gcfs_rq->load.weight)) {
3532 load_sum = div_s64(gcfs_rq->avg.load_sum,
3533 scale_load_down(gcfs_rq->load.weight));
3536 /* But make sure to not inflate se's runnable */
3537 runnable_sum = min(se->avg.load_sum, load_sum);
3541 * runnable_sum can't be lower than running_sum
3542 * Rescale running sum to be in the same range as runnable sum
3543 * running_sum is in [0 : LOAD_AVG_MAX << SCHED_CAPACITY_SHIFT]
3544 * runnable_sum is in [0 : LOAD_AVG_MAX]
3546 running_sum = se->avg.util_sum >> SCHED_CAPACITY_SHIFT;
3547 runnable_sum = max(runnable_sum, running_sum);
3549 load_sum = (s64)se_weight(se) * runnable_sum;
3550 load_avg = div_s64(load_sum, divider);
3552 delta_sum = load_sum - (s64)se_weight(se) * se->avg.load_sum;
3553 delta_avg = load_avg - se->avg.load_avg;
3555 se->avg.load_sum = runnable_sum;
3556 se->avg.load_avg = load_avg;
3557 add_positive(&cfs_rq->avg.load_avg, delta_avg);
3558 add_positive(&cfs_rq->avg.load_sum, delta_sum);
3561 static inline void add_tg_cfs_propagate(struct cfs_rq *cfs_rq, long runnable_sum)
3563 cfs_rq->propagate = 1;
3564 cfs_rq->prop_runnable_sum += runnable_sum;
3567 /* Update task and its cfs_rq load average */
3568 static inline int propagate_entity_load_avg(struct sched_entity *se)
3570 struct cfs_rq *cfs_rq, *gcfs_rq;
3572 if (entity_is_task(se))
3575 gcfs_rq = group_cfs_rq(se);
3576 if (!gcfs_rq->propagate)
3579 gcfs_rq->propagate = 0;
3581 cfs_rq = cfs_rq_of(se);
3583 add_tg_cfs_propagate(cfs_rq, gcfs_rq->prop_runnable_sum);
3585 update_tg_cfs_util(cfs_rq, se, gcfs_rq);
3586 update_tg_cfs_runnable(cfs_rq, se, gcfs_rq);
3587 update_tg_cfs_load(cfs_rq, se, gcfs_rq);
3589 trace_pelt_cfs_tp(cfs_rq);
3590 trace_pelt_se_tp(se);
3596 * Check if we need to update the load and the utilization of a blocked
3599 static inline bool skip_blocked_update(struct sched_entity *se)
3601 struct cfs_rq *gcfs_rq = group_cfs_rq(se);
3604 * If sched_entity still have not zero load or utilization, we have to
3607 if (se->avg.load_avg || se->avg.util_avg)
3611 * If there is a pending propagation, we have to update the load and
3612 * the utilization of the sched_entity:
3614 if (gcfs_rq->propagate)
3618 * Otherwise, the load and the utilization of the sched_entity is
3619 * already zero and there is no pending propagation, so it will be a
3620 * waste of time to try to decay it:
3625 #else /* CONFIG_FAIR_GROUP_SCHED */
3627 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq) {}
3629 static inline int propagate_entity_load_avg(struct sched_entity *se)
3634 static inline void add_tg_cfs_propagate(struct cfs_rq *cfs_rq, long runnable_sum) {}
3636 #endif /* CONFIG_FAIR_GROUP_SCHED */
3639 * update_cfs_rq_load_avg - update the cfs_rq's load/util averages
3640 * @now: current time, as per cfs_rq_clock_pelt()
3641 * @cfs_rq: cfs_rq to update
3643 * The cfs_rq avg is the direct sum of all its entities (blocked and runnable)
3644 * avg. The immediate corollary is that all (fair) tasks must be attached, see
3645 * post_init_entity_util_avg().
3647 * cfs_rq->avg is used for task_h_load() and update_cfs_share() for example.
3649 * Returns true if the load decayed or we removed load.
3651 * Since both these conditions indicate a changed cfs_rq->avg.load we should
3652 * call update_tg_load_avg() when this function returns true.
3655 update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq)
3657 unsigned long removed_load = 0, removed_util = 0, removed_runnable = 0;
3658 struct sched_avg *sa = &cfs_rq->avg;
3661 if (cfs_rq->removed.nr) {
3663 u32 divider = get_pelt_divider(&cfs_rq->avg);
3665 raw_spin_lock(&cfs_rq->removed.lock);
3666 swap(cfs_rq->removed.util_avg, removed_util);
3667 swap(cfs_rq->removed.load_avg, removed_load);
3668 swap(cfs_rq->removed.runnable_avg, removed_runnable);
3669 cfs_rq->removed.nr = 0;
3670 raw_spin_unlock(&cfs_rq->removed.lock);
3673 sub_positive(&sa->load_avg, r);
3674 sub_positive(&sa->load_sum, r * divider);
3677 sub_positive(&sa->util_avg, r);
3678 sub_positive(&sa->util_sum, r * divider);
3680 r = removed_runnable;
3681 sub_positive(&sa->runnable_avg, r);
3682 sub_positive(&sa->runnable_sum, r * divider);
3685 * removed_runnable is the unweighted version of removed_load so we
3686 * can use it to estimate removed_load_sum.
3688 add_tg_cfs_propagate(cfs_rq,
3689 -(long)(removed_runnable * divider) >> SCHED_CAPACITY_SHIFT);
3694 decayed |= __update_load_avg_cfs_rq(now, cfs_rq);
3696 #ifndef CONFIG_64BIT
3698 cfs_rq->load_last_update_time_copy = sa->last_update_time;
3705 * attach_entity_load_avg - attach this entity to its cfs_rq load avg
3706 * @cfs_rq: cfs_rq to attach to
3707 * @se: sched_entity to attach
3709 * Must call update_cfs_rq_load_avg() before this, since we rely on
3710 * cfs_rq->avg.last_update_time being current.
3712 static void attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3715 * cfs_rq->avg.period_contrib can be used for both cfs_rq and se.
3716 * See ___update_load_avg() for details.
3718 u32 divider = get_pelt_divider(&cfs_rq->avg);
3721 * When we attach the @se to the @cfs_rq, we must align the decay
3722 * window because without that, really weird and wonderful things can
3727 se->avg.last_update_time = cfs_rq->avg.last_update_time;
3728 se->avg.period_contrib = cfs_rq->avg.period_contrib;
3731 * Hell(o) Nasty stuff.. we need to recompute _sum based on the new
3732 * period_contrib. This isn't strictly correct, but since we're
3733 * entirely outside of the PELT hierarchy, nobody cares if we truncate
3736 se->avg.util_sum = se->avg.util_avg * divider;
3738 se->avg.runnable_sum = se->avg.runnable_avg * divider;
3740 se->avg.load_sum = divider;
3741 if (se_weight(se)) {
3743 div_u64(se->avg.load_avg * se->avg.load_sum, se_weight(se));
3746 enqueue_load_avg(cfs_rq, se);
3747 cfs_rq->avg.util_avg += se->avg.util_avg;
3748 cfs_rq->avg.util_sum += se->avg.util_sum;
3749 cfs_rq->avg.runnable_avg += se->avg.runnable_avg;
3750 cfs_rq->avg.runnable_sum += se->avg.runnable_sum;
3752 add_tg_cfs_propagate(cfs_rq, se->avg.load_sum);
3754 cfs_rq_util_change(cfs_rq, 0);
3756 trace_pelt_cfs_tp(cfs_rq);
3760 * detach_entity_load_avg - detach this entity from its cfs_rq load avg
3761 * @cfs_rq: cfs_rq to detach from
3762 * @se: sched_entity to detach
3764 * Must call update_cfs_rq_load_avg() before this, since we rely on
3765 * cfs_rq->avg.last_update_time being current.
3767 static void detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3769 dequeue_load_avg(cfs_rq, se);
3770 sub_positive(&cfs_rq->avg.util_avg, se->avg.util_avg);
3771 sub_positive(&cfs_rq->avg.util_sum, se->avg.util_sum);
3772 sub_positive(&cfs_rq->avg.runnable_avg, se->avg.runnable_avg);
3773 sub_positive(&cfs_rq->avg.runnable_sum, se->avg.runnable_sum);
3775 add_tg_cfs_propagate(cfs_rq, -se->avg.load_sum);
3777 cfs_rq_util_change(cfs_rq, 0);
3779 trace_pelt_cfs_tp(cfs_rq);
3783 * Optional action to be done while updating the load average
3785 #define UPDATE_TG 0x1
3786 #define SKIP_AGE_LOAD 0x2
3787 #define DO_ATTACH 0x4
3789 /* Update task and its cfs_rq load average */
3790 static inline void update_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3792 u64 now = cfs_rq_clock_pelt(cfs_rq);
3796 * Track task load average for carrying it to new CPU after migrated, and
3797 * track group sched_entity load average for task_h_load calc in migration
3799 if (se->avg.last_update_time && !(flags & SKIP_AGE_LOAD))
3800 __update_load_avg_se(now, cfs_rq, se);
3802 decayed = update_cfs_rq_load_avg(now, cfs_rq);
3803 decayed |= propagate_entity_load_avg(se);
3805 if (!se->avg.last_update_time && (flags & DO_ATTACH)) {
3808 * DO_ATTACH means we're here from enqueue_entity().
3809 * !last_update_time means we've passed through
3810 * migrate_task_rq_fair() indicating we migrated.
3812 * IOW we're enqueueing a task on a new CPU.
3814 attach_entity_load_avg(cfs_rq, se);
3815 update_tg_load_avg(cfs_rq);
3817 } else if (decayed) {
3818 cfs_rq_util_change(cfs_rq, 0);
3820 if (flags & UPDATE_TG)
3821 update_tg_load_avg(cfs_rq);
3825 #ifndef CONFIG_64BIT
3826 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
3828 u64 last_update_time_copy;
3829 u64 last_update_time;
3832 last_update_time_copy = cfs_rq->load_last_update_time_copy;
3834 last_update_time = cfs_rq->avg.last_update_time;
3835 } while (last_update_time != last_update_time_copy);
3837 return last_update_time;
3840 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
3842 return cfs_rq->avg.last_update_time;
3847 * Synchronize entity load avg of dequeued entity without locking
3850 static void sync_entity_load_avg(struct sched_entity *se)
3852 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3853 u64 last_update_time;
3855 last_update_time = cfs_rq_last_update_time(cfs_rq);
3856 __update_load_avg_blocked_se(last_update_time, se);
3860 * Task first catches up with cfs_rq, and then subtract
3861 * itself from the cfs_rq (task must be off the queue now).
3863 static void remove_entity_load_avg(struct sched_entity *se)
3865 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3866 unsigned long flags;
3869 * tasks cannot exit without having gone through wake_up_new_task() ->
3870 * post_init_entity_util_avg() which will have added things to the
3871 * cfs_rq, so we can remove unconditionally.
3874 sync_entity_load_avg(se);
3876 raw_spin_lock_irqsave(&cfs_rq->removed.lock, flags);
3877 ++cfs_rq->removed.nr;
3878 cfs_rq->removed.util_avg += se->avg.util_avg;
3879 cfs_rq->removed.load_avg += se->avg.load_avg;
3880 cfs_rq->removed.runnable_avg += se->avg.runnable_avg;
3881 raw_spin_unlock_irqrestore(&cfs_rq->removed.lock, flags);
3884 static inline unsigned long cfs_rq_runnable_avg(struct cfs_rq *cfs_rq)
3886 return cfs_rq->avg.runnable_avg;
3889 static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq)
3891 return cfs_rq->avg.load_avg;
3894 static int newidle_balance(struct rq *this_rq, struct rq_flags *rf);
3896 static inline unsigned long task_util(struct task_struct *p)
3898 return READ_ONCE(p->se.avg.util_avg);
3901 static inline unsigned long _task_util_est(struct task_struct *p)
3903 struct util_est ue = READ_ONCE(p->se.avg.util_est);
3905 return (max(ue.ewma, ue.enqueued) | UTIL_AVG_UNCHANGED);
3908 static inline unsigned long task_util_est(struct task_struct *p)
3910 return max(task_util(p), _task_util_est(p));
3913 #ifdef CONFIG_UCLAMP_TASK
3914 static inline unsigned long uclamp_task_util(struct task_struct *p)
3916 return clamp(task_util_est(p),
3917 uclamp_eff_value(p, UCLAMP_MIN),
3918 uclamp_eff_value(p, UCLAMP_MAX));
3921 static inline unsigned long uclamp_task_util(struct task_struct *p)
3923 return task_util_est(p);
3927 static inline void util_est_enqueue(struct cfs_rq *cfs_rq,
3928 struct task_struct *p)
3930 unsigned int enqueued;
3932 if (!sched_feat(UTIL_EST))
3935 /* Update root cfs_rq's estimated utilization */
3936 enqueued = cfs_rq->avg.util_est.enqueued;
3937 enqueued += _task_util_est(p);
3938 WRITE_ONCE(cfs_rq->avg.util_est.enqueued, enqueued);
3940 trace_sched_util_est_cfs_tp(cfs_rq);
3943 static inline void util_est_dequeue(struct cfs_rq *cfs_rq,
3944 struct task_struct *p)
3946 unsigned int enqueued;
3948 if (!sched_feat(UTIL_EST))
3951 /* Update root cfs_rq's estimated utilization */
3952 enqueued = cfs_rq->avg.util_est.enqueued;
3953 enqueued -= min_t(unsigned int, enqueued, _task_util_est(p));
3954 WRITE_ONCE(cfs_rq->avg.util_est.enqueued, enqueued);
3956 trace_sched_util_est_cfs_tp(cfs_rq);
3959 #define UTIL_EST_MARGIN (SCHED_CAPACITY_SCALE / 100)
3962 * Check if a (signed) value is within a specified (unsigned) margin,
3963 * based on the observation that:
3965 * abs(x) < y := (unsigned)(x + y - 1) < (2 * y - 1)
3967 * NOTE: this only works when value + margin < INT_MAX.
3969 static inline bool within_margin(int value, int margin)
3971 return ((unsigned int)(value + margin - 1) < (2 * margin - 1));
3974 static inline void util_est_update(struct cfs_rq *cfs_rq,
3975 struct task_struct *p,
3978 long last_ewma_diff, last_enqueued_diff;
3981 if (!sched_feat(UTIL_EST))
3985 * Skip update of task's estimated utilization when the task has not
3986 * yet completed an activation, e.g. being migrated.
3992 * If the PELT values haven't changed since enqueue time,
3993 * skip the util_est update.
3995 ue = p->se.avg.util_est;
3996 if (ue.enqueued & UTIL_AVG_UNCHANGED)
3999 last_enqueued_diff = ue.enqueued;
4002 * Reset EWMA on utilization increases, the moving average is used only
4003 * to smooth utilization decreases.
4005 ue.enqueued = (task_util(p) | UTIL_AVG_UNCHANGED);
4006 if (sched_feat(UTIL_EST_FASTUP)) {
4007 if (ue.ewma < ue.enqueued) {
4008 ue.ewma = ue.enqueued;
4014 * Skip update of task's estimated utilization when its members are
4015 * already ~1% close to its last activation value.
4017 last_ewma_diff = ue.enqueued - ue.ewma;
4018 last_enqueued_diff -= ue.enqueued;
4019 if (within_margin(last_ewma_diff, UTIL_EST_MARGIN)) {
4020 if (!within_margin(last_enqueued_diff, UTIL_EST_MARGIN))
4027 * To avoid overestimation of actual task utilization, skip updates if
4028 * we cannot grant there is idle time in this CPU.
4030 if (task_util(p) > capacity_orig_of(cpu_of(rq_of(cfs_rq))))
4034 * Update Task's estimated utilization
4036 * When *p completes an activation we can consolidate another sample
4037 * of the task size. This is done by storing the current PELT value
4038 * as ue.enqueued and by using this value to update the Exponential
4039 * Weighted Moving Average (EWMA):
4041 * ewma(t) = w * task_util(p) + (1-w) * ewma(t-1)
4042 * = w * task_util(p) + ewma(t-1) - w * ewma(t-1)
4043 * = w * (task_util(p) - ewma(t-1)) + ewma(t-1)
4044 * = w * ( last_ewma_diff ) + ewma(t-1)
4045 * = w * (last_ewma_diff + ewma(t-1) / w)
4047 * Where 'w' is the weight of new samples, which is configured to be
4048 * 0.25, thus making w=1/4 ( >>= UTIL_EST_WEIGHT_SHIFT)
4050 ue.ewma <<= UTIL_EST_WEIGHT_SHIFT;
4051 ue.ewma += last_ewma_diff;
4052 ue.ewma >>= UTIL_EST_WEIGHT_SHIFT;
4054 WRITE_ONCE(p->se.avg.util_est, ue);
4056 trace_sched_util_est_se_tp(&p->se);
4059 static inline int task_fits_capacity(struct task_struct *p, long capacity)
4061 return fits_capacity(uclamp_task_util(p), capacity);
4064 static inline void update_misfit_status(struct task_struct *p, struct rq *rq)
4066 if (!static_branch_unlikely(&sched_asym_cpucapacity))
4069 if (!p || p->nr_cpus_allowed == 1) {
4070 rq->misfit_task_load = 0;
4074 if (task_fits_capacity(p, capacity_of(cpu_of(rq)))) {
4075 rq->misfit_task_load = 0;
4080 * Make sure that misfit_task_load will not be null even if
4081 * task_h_load() returns 0.
4083 rq->misfit_task_load = max_t(unsigned long, task_h_load(p), 1);
4086 #else /* CONFIG_SMP */
4088 #define UPDATE_TG 0x0
4089 #define SKIP_AGE_LOAD 0x0
4090 #define DO_ATTACH 0x0
4092 static inline void update_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int not_used1)
4094 cfs_rq_util_change(cfs_rq, 0);
4097 static inline void remove_entity_load_avg(struct sched_entity *se) {}
4100 attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
4102 detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
4104 static inline int newidle_balance(struct rq *rq, struct rq_flags *rf)
4110 util_est_enqueue(struct cfs_rq *cfs_rq, struct task_struct *p) {}
4113 util_est_dequeue(struct cfs_rq *cfs_rq, struct task_struct *p) {}
4116 util_est_update(struct cfs_rq *cfs_rq, struct task_struct *p,
4118 static inline void update_misfit_status(struct task_struct *p, struct rq *rq) {}
4120 #endif /* CONFIG_SMP */
4122 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
4124 #ifdef CONFIG_SCHED_DEBUG
4125 s64 d = se->vruntime - cfs_rq->min_vruntime;
4130 if (d > 3*sysctl_sched_latency)
4131 schedstat_inc(cfs_rq->nr_spread_over);
4136 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
4138 u64 vruntime = cfs_rq->min_vruntime;
4141 * The 'current' period is already promised to the current tasks,
4142 * however the extra weight of the new task will slow them down a
4143 * little, place the new task so that it fits in the slot that
4144 * stays open at the end.
4146 if (initial && sched_feat(START_DEBIT))
4147 vruntime += sched_vslice(cfs_rq, se);
4149 /* sleeps up to a single latency don't count. */
4151 unsigned long thresh = sysctl_sched_latency;
4154 * Halve their sleep time's effect, to allow
4155 * for a gentler effect of sleepers:
4157 if (sched_feat(GENTLE_FAIR_SLEEPERS))
4163 /* ensure we never gain time by being placed backwards. */
4164 se->vruntime = max_vruntime(se->vruntime, vruntime);
4167 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
4169 static inline void check_schedstat_required(void)
4171 #ifdef CONFIG_SCHEDSTATS
4172 if (schedstat_enabled())
4175 /* Force schedstat enabled if a dependent tracepoint is active */
4176 if (trace_sched_stat_wait_enabled() ||
4177 trace_sched_stat_sleep_enabled() ||
4178 trace_sched_stat_iowait_enabled() ||
4179 trace_sched_stat_blocked_enabled() ||
4180 trace_sched_stat_runtime_enabled()) {
4181 printk_deferred_once("Scheduler tracepoints stat_sleep, stat_iowait, "
4182 "stat_blocked and stat_runtime require the "
4183 "kernel parameter schedstats=enable or "
4184 "kernel.sched_schedstats=1\n");
4189 static inline bool cfs_bandwidth_used(void);
4196 * update_min_vruntime()
4197 * vruntime -= min_vruntime
4201 * update_min_vruntime()
4202 * vruntime += min_vruntime
4204 * this way the vruntime transition between RQs is done when both
4205 * min_vruntime are up-to-date.
4209 * ->migrate_task_rq_fair() (p->state == TASK_WAKING)
4210 * vruntime -= min_vruntime
4214 * update_min_vruntime()
4215 * vruntime += min_vruntime
4217 * this way we don't have the most up-to-date min_vruntime on the originating
4218 * CPU and an up-to-date min_vruntime on the destination CPU.
4222 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
4224 bool renorm = !(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_MIGRATED);
4225 bool curr = cfs_rq->curr == se;
4228 * If we're the current task, we must renormalise before calling
4232 se->vruntime += cfs_rq->min_vruntime;
4234 update_curr(cfs_rq);
4237 * Otherwise, renormalise after, such that we're placed at the current
4238 * moment in time, instead of some random moment in the past. Being
4239 * placed in the past could significantly boost this task to the
4240 * fairness detriment of existing tasks.
4242 if (renorm && !curr)
4243 se->vruntime += cfs_rq->min_vruntime;
4246 * When enqueuing a sched_entity, we must:
4247 * - Update loads to have both entity and cfs_rq synced with now.
4248 * - Add its load to cfs_rq->runnable_avg
4249 * - For group_entity, update its weight to reflect the new share of
4251 * - Add its new weight to cfs_rq->load.weight
4253 update_load_avg(cfs_rq, se, UPDATE_TG | DO_ATTACH);
4254 se_update_runnable(se);
4255 update_cfs_group(se);
4256 account_entity_enqueue(cfs_rq, se);
4258 if (flags & ENQUEUE_WAKEUP)
4259 place_entity(cfs_rq, se, 0);
4261 check_schedstat_required();
4262 update_stats_enqueue(cfs_rq, se, flags);
4263 check_spread(cfs_rq, se);
4265 __enqueue_entity(cfs_rq, se);
4269 * When bandwidth control is enabled, cfs might have been removed
4270 * because of a parent been throttled but cfs->nr_running > 1. Try to
4271 * add it unconditionally.
4273 if (cfs_rq->nr_running == 1 || cfs_bandwidth_used())
4274 list_add_leaf_cfs_rq(cfs_rq);
4276 if (cfs_rq->nr_running == 1)
4277 check_enqueue_throttle(cfs_rq);
4280 static void __clear_buddies_last(struct sched_entity *se)
4282 for_each_sched_entity(se) {
4283 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4284 if (cfs_rq->last != se)
4287 cfs_rq->last = NULL;
4291 static void __clear_buddies_next(struct sched_entity *se)
4293 for_each_sched_entity(se) {
4294 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4295 if (cfs_rq->next != se)
4298 cfs_rq->next = NULL;
4302 static void __clear_buddies_skip(struct sched_entity *se)
4304 for_each_sched_entity(se) {
4305 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4306 if (cfs_rq->skip != se)
4309 cfs_rq->skip = NULL;
4313 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
4315 if (cfs_rq->last == se)
4316 __clear_buddies_last(se);
4318 if (cfs_rq->next == se)
4319 __clear_buddies_next(se);
4321 if (cfs_rq->skip == se)
4322 __clear_buddies_skip(se);
4325 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
4328 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
4331 * Update run-time statistics of the 'current'.
4333 update_curr(cfs_rq);
4336 * When dequeuing a sched_entity, we must:
4337 * - Update loads to have both entity and cfs_rq synced with now.
4338 * - Subtract its load from the cfs_rq->runnable_avg.
4339 * - Subtract its previous weight from cfs_rq->load.weight.
4340 * - For group entity, update its weight to reflect the new share
4341 * of its group cfs_rq.
4343 update_load_avg(cfs_rq, se, UPDATE_TG);
4344 se_update_runnable(se);
4346 update_stats_dequeue(cfs_rq, se, flags);
4348 clear_buddies(cfs_rq, se);
4350 if (se != cfs_rq->curr)
4351 __dequeue_entity(cfs_rq, se);
4353 account_entity_dequeue(cfs_rq, se);
4356 * Normalize after update_curr(); which will also have moved
4357 * min_vruntime if @se is the one holding it back. But before doing
4358 * update_min_vruntime() again, which will discount @se's position and
4359 * can move min_vruntime forward still more.
4361 if (!(flags & DEQUEUE_SLEEP))
4362 se->vruntime -= cfs_rq->min_vruntime;
4364 /* return excess runtime on last dequeue */
4365 return_cfs_rq_runtime(cfs_rq);
4367 update_cfs_group(se);
4370 * Now advance min_vruntime if @se was the entity holding it back,
4371 * except when: DEQUEUE_SAVE && !DEQUEUE_MOVE, in this case we'll be
4372 * put back on, and if we advance min_vruntime, we'll be placed back
4373 * further than we started -- ie. we'll be penalized.
4375 if ((flags & (DEQUEUE_SAVE | DEQUEUE_MOVE)) != DEQUEUE_SAVE)
4376 update_min_vruntime(cfs_rq);
4380 * Preempt the current task with a newly woken task if needed:
4383 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
4385 unsigned long ideal_runtime, delta_exec;
4386 struct sched_entity *se;
4389 ideal_runtime = sched_slice(cfs_rq, curr);
4390 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
4391 if (delta_exec > ideal_runtime) {
4392 resched_curr(rq_of(cfs_rq));
4394 * The current task ran long enough, ensure it doesn't get
4395 * re-elected due to buddy favours.
4397 clear_buddies(cfs_rq, curr);
4402 * Ensure that a task that missed wakeup preemption by a
4403 * narrow margin doesn't have to wait for a full slice.
4404 * This also mitigates buddy induced latencies under load.
4406 if (delta_exec < sysctl_sched_min_granularity)
4409 se = __pick_first_entity(cfs_rq);
4410 delta = curr->vruntime - se->vruntime;
4415 if (delta > ideal_runtime)
4416 resched_curr(rq_of(cfs_rq));
4420 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
4422 /* 'current' is not kept within the tree. */
4425 * Any task has to be enqueued before it get to execute on
4426 * a CPU. So account for the time it spent waiting on the
4429 update_stats_wait_end(cfs_rq, se);
4430 __dequeue_entity(cfs_rq, se);
4431 update_load_avg(cfs_rq, se, UPDATE_TG);
4434 update_stats_curr_start(cfs_rq, se);
4438 * Track our maximum slice length, if the CPU's load is at
4439 * least twice that of our own weight (i.e. dont track it
4440 * when there are only lesser-weight tasks around):
4442 if (schedstat_enabled() &&
4443 rq_of(cfs_rq)->cfs.load.weight >= 2*se->load.weight) {
4444 schedstat_set(se->statistics.slice_max,
4445 max((u64)schedstat_val(se->statistics.slice_max),
4446 se->sum_exec_runtime - se->prev_sum_exec_runtime));
4449 se->prev_sum_exec_runtime = se->sum_exec_runtime;
4453 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
4456 * Pick the next process, keeping these things in mind, in this order:
4457 * 1) keep things fair between processes/task groups
4458 * 2) pick the "next" process, since someone really wants that to run
4459 * 3) pick the "last" process, for cache locality
4460 * 4) do not run the "skip" process, if something else is available
4462 static struct sched_entity *
4463 pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr)
4465 struct sched_entity *left = __pick_first_entity(cfs_rq);
4466 struct sched_entity *se;
4469 * If curr is set we have to see if its left of the leftmost entity
4470 * still in the tree, provided there was anything in the tree at all.
4472 if (!left || (curr && entity_before(curr, left)))
4475 se = left; /* ideally we run the leftmost entity */
4478 * Avoid running the skip buddy, if running something else can
4479 * be done without getting too unfair.
4481 if (cfs_rq->skip == se) {
4482 struct sched_entity *second;
4485 second = __pick_first_entity(cfs_rq);
4487 second = __pick_next_entity(se);
4488 if (!second || (curr && entity_before(curr, second)))
4492 if (second && wakeup_preempt_entity(second, left) < 1)
4496 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1) {
4498 * Someone really wants this to run. If it's not unfair, run it.
4501 } else if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1) {
4503 * Prefer last buddy, try to return the CPU to a preempted task.
4508 clear_buddies(cfs_rq, se);
4513 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
4515 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
4518 * If still on the runqueue then deactivate_task()
4519 * was not called and update_curr() has to be done:
4522 update_curr(cfs_rq);
4524 /* throttle cfs_rqs exceeding runtime */
4525 check_cfs_rq_runtime(cfs_rq);
4527 check_spread(cfs_rq, prev);
4530 update_stats_wait_start(cfs_rq, prev);
4531 /* Put 'current' back into the tree. */
4532 __enqueue_entity(cfs_rq, prev);
4533 /* in !on_rq case, update occurred at dequeue */
4534 update_load_avg(cfs_rq, prev, 0);
4536 cfs_rq->curr = NULL;
4540 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
4543 * Update run-time statistics of the 'current'.
4545 update_curr(cfs_rq);
4548 * Ensure that runnable average is periodically updated.
4550 update_load_avg(cfs_rq, curr, UPDATE_TG);
4551 update_cfs_group(curr);
4553 #ifdef CONFIG_SCHED_HRTICK
4555 * queued ticks are scheduled to match the slice, so don't bother
4556 * validating it and just reschedule.
4559 resched_curr(rq_of(cfs_rq));
4563 * don't let the period tick interfere with the hrtick preemption
4565 if (!sched_feat(DOUBLE_TICK) &&
4566 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
4570 if (cfs_rq->nr_running > 1)
4571 check_preempt_tick(cfs_rq, curr);
4575 /**************************************************
4576 * CFS bandwidth control machinery
4579 #ifdef CONFIG_CFS_BANDWIDTH
4581 #ifdef CONFIG_JUMP_LABEL
4582 static struct static_key __cfs_bandwidth_used;
4584 static inline bool cfs_bandwidth_used(void)
4586 return static_key_false(&__cfs_bandwidth_used);
4589 void cfs_bandwidth_usage_inc(void)
4591 static_key_slow_inc_cpuslocked(&__cfs_bandwidth_used);
4594 void cfs_bandwidth_usage_dec(void)
4596 static_key_slow_dec_cpuslocked(&__cfs_bandwidth_used);
4598 #else /* CONFIG_JUMP_LABEL */
4599 static bool cfs_bandwidth_used(void)
4604 void cfs_bandwidth_usage_inc(void) {}
4605 void cfs_bandwidth_usage_dec(void) {}
4606 #endif /* CONFIG_JUMP_LABEL */
4609 * default period for cfs group bandwidth.
4610 * default: 0.1s, units: nanoseconds
4612 static inline u64 default_cfs_period(void)
4614 return 100000000ULL;
4617 static inline u64 sched_cfs_bandwidth_slice(void)
4619 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
4623 * Replenish runtime according to assigned quota. We use sched_clock_cpu
4624 * directly instead of rq->clock to avoid adding additional synchronization
4627 * requires cfs_b->lock
4629 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
4631 if (cfs_b->quota != RUNTIME_INF)
4632 cfs_b->runtime = cfs_b->quota;
4635 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
4637 return &tg->cfs_bandwidth;
4640 /* returns 0 on failure to allocate runtime */
4641 static int __assign_cfs_rq_runtime(struct cfs_bandwidth *cfs_b,
4642 struct cfs_rq *cfs_rq, u64 target_runtime)
4644 u64 min_amount, amount = 0;
4646 lockdep_assert_held(&cfs_b->lock);
4648 /* note: this is a positive sum as runtime_remaining <= 0 */
4649 min_amount = target_runtime - cfs_rq->runtime_remaining;
4651 if (cfs_b->quota == RUNTIME_INF)
4652 amount = min_amount;
4654 start_cfs_bandwidth(cfs_b);
4656 if (cfs_b->runtime > 0) {
4657 amount = min(cfs_b->runtime, min_amount);
4658 cfs_b->runtime -= amount;
4663 cfs_rq->runtime_remaining += amount;
4665 return cfs_rq->runtime_remaining > 0;
4668 /* returns 0 on failure to allocate runtime */
4669 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4671 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4674 raw_spin_lock(&cfs_b->lock);
4675 ret = __assign_cfs_rq_runtime(cfs_b, cfs_rq, sched_cfs_bandwidth_slice());
4676 raw_spin_unlock(&cfs_b->lock);
4681 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
4683 /* dock delta_exec before expiring quota (as it could span periods) */
4684 cfs_rq->runtime_remaining -= delta_exec;
4686 if (likely(cfs_rq->runtime_remaining > 0))
4689 if (cfs_rq->throttled)
4692 * if we're unable to extend our runtime we resched so that the active
4693 * hierarchy can be throttled
4695 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
4696 resched_curr(rq_of(cfs_rq));
4699 static __always_inline
4700 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
4702 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
4705 __account_cfs_rq_runtime(cfs_rq, delta_exec);
4708 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
4710 return cfs_bandwidth_used() && cfs_rq->throttled;
4713 /* check whether cfs_rq, or any parent, is throttled */
4714 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
4716 return cfs_bandwidth_used() && cfs_rq->throttle_count;
4720 * Ensure that neither of the group entities corresponding to src_cpu or
4721 * dest_cpu are members of a throttled hierarchy when performing group
4722 * load-balance operations.
4724 static inline int throttled_lb_pair(struct task_group *tg,
4725 int src_cpu, int dest_cpu)
4727 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
4729 src_cfs_rq = tg->cfs_rq[src_cpu];
4730 dest_cfs_rq = tg->cfs_rq[dest_cpu];
4732 return throttled_hierarchy(src_cfs_rq) ||
4733 throttled_hierarchy(dest_cfs_rq);
4736 static int tg_unthrottle_up(struct task_group *tg, void *data)
4738 struct rq *rq = data;
4739 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
4741 cfs_rq->throttle_count--;
4742 if (!cfs_rq->throttle_count) {
4743 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
4744 cfs_rq->throttled_clock_task;
4746 /* Add cfs_rq with already running entity in the list */
4747 if (cfs_rq->nr_running >= 1)
4748 list_add_leaf_cfs_rq(cfs_rq);
4754 static int tg_throttle_down(struct task_group *tg, void *data)
4756 struct rq *rq = data;
4757 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
4759 /* group is entering throttled state, stop time */
4760 if (!cfs_rq->throttle_count) {
4761 cfs_rq->throttled_clock_task = rq_clock_task(rq);
4762 list_del_leaf_cfs_rq(cfs_rq);
4764 cfs_rq->throttle_count++;
4769 static bool throttle_cfs_rq(struct cfs_rq *cfs_rq)
4771 struct rq *rq = rq_of(cfs_rq);
4772 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4773 struct sched_entity *se;
4774 long task_delta, idle_task_delta, dequeue = 1;
4776 raw_spin_lock(&cfs_b->lock);
4777 /* This will start the period timer if necessary */
4778 if (__assign_cfs_rq_runtime(cfs_b, cfs_rq, 1)) {
4780 * We have raced with bandwidth becoming available, and if we
4781 * actually throttled the timer might not unthrottle us for an
4782 * entire period. We additionally needed to make sure that any
4783 * subsequent check_cfs_rq_runtime calls agree not to throttle
4784 * us, as we may commit to do cfs put_prev+pick_next, so we ask
4785 * for 1ns of runtime rather than just check cfs_b.
4789 list_add_tail_rcu(&cfs_rq->throttled_list,
4790 &cfs_b->throttled_cfs_rq);
4792 raw_spin_unlock(&cfs_b->lock);
4795 return false; /* Throttle no longer required. */
4797 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
4799 /* freeze hierarchy runnable averages while throttled */
4801 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
4804 task_delta = cfs_rq->h_nr_running;
4805 idle_task_delta = cfs_rq->idle_h_nr_running;
4806 for_each_sched_entity(se) {
4807 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
4808 /* throttled entity or throttle-on-deactivate */
4812 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
4814 qcfs_rq->h_nr_running -= task_delta;
4815 qcfs_rq->idle_h_nr_running -= idle_task_delta;
4817 if (qcfs_rq->load.weight) {
4818 /* Avoid re-evaluating load for this entity: */
4819 se = parent_entity(se);
4824 for_each_sched_entity(se) {
4825 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
4826 /* throttled entity or throttle-on-deactivate */
4830 update_load_avg(qcfs_rq, se, 0);
4831 se_update_runnable(se);
4833 qcfs_rq->h_nr_running -= task_delta;
4834 qcfs_rq->idle_h_nr_running -= idle_task_delta;
4837 /* At this point se is NULL and we are at root level*/
4838 sub_nr_running(rq, task_delta);
4842 * Note: distribution will already see us throttled via the
4843 * throttled-list. rq->lock protects completion.
4845 cfs_rq->throttled = 1;
4846 cfs_rq->throttled_clock = rq_clock(rq);
4850 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
4852 struct rq *rq = rq_of(cfs_rq);
4853 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4854 struct sched_entity *se;
4855 long task_delta, idle_task_delta;
4857 se = cfs_rq->tg->se[cpu_of(rq)];
4859 cfs_rq->throttled = 0;
4861 update_rq_clock(rq);
4863 raw_spin_lock(&cfs_b->lock);
4864 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
4865 list_del_rcu(&cfs_rq->throttled_list);
4866 raw_spin_unlock(&cfs_b->lock);
4868 /* update hierarchical throttle state */
4869 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
4871 if (!cfs_rq->load.weight)
4874 task_delta = cfs_rq->h_nr_running;
4875 idle_task_delta = cfs_rq->idle_h_nr_running;
4876 for_each_sched_entity(se) {
4879 cfs_rq = cfs_rq_of(se);
4880 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
4882 cfs_rq->h_nr_running += task_delta;
4883 cfs_rq->idle_h_nr_running += idle_task_delta;
4885 /* end evaluation on encountering a throttled cfs_rq */
4886 if (cfs_rq_throttled(cfs_rq))
4887 goto unthrottle_throttle;
4890 for_each_sched_entity(se) {
4891 cfs_rq = cfs_rq_of(se);
4893 update_load_avg(cfs_rq, se, UPDATE_TG);
4894 se_update_runnable(se);
4896 cfs_rq->h_nr_running += task_delta;
4897 cfs_rq->idle_h_nr_running += idle_task_delta;
4900 /* end evaluation on encountering a throttled cfs_rq */
4901 if (cfs_rq_throttled(cfs_rq))
4902 goto unthrottle_throttle;
4905 * One parent has been throttled and cfs_rq removed from the
4906 * list. Add it back to not break the leaf list.
4908 if (throttled_hierarchy(cfs_rq))
4909 list_add_leaf_cfs_rq(cfs_rq);
4912 /* At this point se is NULL and we are at root level*/
4913 add_nr_running(rq, task_delta);
4915 unthrottle_throttle:
4917 * The cfs_rq_throttled() breaks in the above iteration can result in
4918 * incomplete leaf list maintenance, resulting in triggering the
4921 for_each_sched_entity(se) {
4922 cfs_rq = cfs_rq_of(se);
4924 if (list_add_leaf_cfs_rq(cfs_rq))
4928 assert_list_leaf_cfs_rq(rq);
4930 /* Determine whether we need to wake up potentially idle CPU: */
4931 if (rq->curr == rq->idle && rq->cfs.nr_running)
4935 static void distribute_cfs_runtime(struct cfs_bandwidth *cfs_b)
4937 struct cfs_rq *cfs_rq;
4938 u64 runtime, remaining = 1;
4941 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
4943 struct rq *rq = rq_of(cfs_rq);
4946 rq_lock_irqsave(rq, &rf);
4947 if (!cfs_rq_throttled(cfs_rq))
4950 /* By the above check, this should never be true */
4951 SCHED_WARN_ON(cfs_rq->runtime_remaining > 0);
4953 raw_spin_lock(&cfs_b->lock);
4954 runtime = -cfs_rq->runtime_remaining + 1;
4955 if (runtime > cfs_b->runtime)
4956 runtime = cfs_b->runtime;
4957 cfs_b->runtime -= runtime;
4958 remaining = cfs_b->runtime;
4959 raw_spin_unlock(&cfs_b->lock);
4961 cfs_rq->runtime_remaining += runtime;
4963 /* we check whether we're throttled above */
4964 if (cfs_rq->runtime_remaining > 0)
4965 unthrottle_cfs_rq(cfs_rq);
4968 rq_unlock_irqrestore(rq, &rf);
4977 * Responsible for refilling a task_group's bandwidth and unthrottling its
4978 * cfs_rqs as appropriate. If there has been no activity within the last
4979 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
4980 * used to track this state.
4982 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun, unsigned long flags)
4986 /* no need to continue the timer with no bandwidth constraint */
4987 if (cfs_b->quota == RUNTIME_INF)
4988 goto out_deactivate;
4990 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
4991 cfs_b->nr_periods += overrun;
4994 * idle depends on !throttled (for the case of a large deficit), and if
4995 * we're going inactive then everything else can be deferred
4997 if (cfs_b->idle && !throttled)
4998 goto out_deactivate;
5000 __refill_cfs_bandwidth_runtime(cfs_b);
5003 /* mark as potentially idle for the upcoming period */
5008 /* account preceding periods in which throttling occurred */
5009 cfs_b->nr_throttled += overrun;
5012 * This check is repeated as we release cfs_b->lock while we unthrottle.
5014 while (throttled && cfs_b->runtime > 0) {
5015 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
5016 /* we can't nest cfs_b->lock while distributing bandwidth */
5017 distribute_cfs_runtime(cfs_b);
5018 raw_spin_lock_irqsave(&cfs_b->lock, flags);
5020 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
5024 * While we are ensured activity in the period following an
5025 * unthrottle, this also covers the case in which the new bandwidth is
5026 * insufficient to cover the existing bandwidth deficit. (Forcing the
5027 * timer to remain active while there are any throttled entities.)
5037 /* a cfs_rq won't donate quota below this amount */
5038 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
5039 /* minimum remaining period time to redistribute slack quota */
5040 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
5041 /* how long we wait to gather additional slack before distributing */
5042 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
5045 * Are we near the end of the current quota period?
5047 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
5048 * hrtimer base being cleared by hrtimer_start. In the case of
5049 * migrate_hrtimers, base is never cleared, so we are fine.
5051 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
5053 struct hrtimer *refresh_timer = &cfs_b->period_timer;
5056 /* if the call-back is running a quota refresh is already occurring */
5057 if (hrtimer_callback_running(refresh_timer))
5060 /* is a quota refresh about to occur? */
5061 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
5062 if (remaining < min_expire)
5068 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
5070 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
5072 /* if there's a quota refresh soon don't bother with slack */
5073 if (runtime_refresh_within(cfs_b, min_left))
5076 /* don't push forwards an existing deferred unthrottle */
5077 if (cfs_b->slack_started)
5079 cfs_b->slack_started = true;
5081 hrtimer_start(&cfs_b->slack_timer,
5082 ns_to_ktime(cfs_bandwidth_slack_period),
5086 /* we know any runtime found here is valid as update_curr() precedes return */
5087 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
5089 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
5090 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
5092 if (slack_runtime <= 0)
5095 raw_spin_lock(&cfs_b->lock);
5096 if (cfs_b->quota != RUNTIME_INF) {
5097 cfs_b->runtime += slack_runtime;
5099 /* we are under rq->lock, defer unthrottling using a timer */
5100 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
5101 !list_empty(&cfs_b->throttled_cfs_rq))
5102 start_cfs_slack_bandwidth(cfs_b);
5104 raw_spin_unlock(&cfs_b->lock);
5106 /* even if it's not valid for return we don't want to try again */
5107 cfs_rq->runtime_remaining -= slack_runtime;
5110 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
5112 if (!cfs_bandwidth_used())
5115 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
5118 __return_cfs_rq_runtime(cfs_rq);
5122 * This is done with a timer (instead of inline with bandwidth return) since
5123 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
5125 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
5127 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
5128 unsigned long flags;
5130 /* confirm we're still not at a refresh boundary */
5131 raw_spin_lock_irqsave(&cfs_b->lock, flags);
5132 cfs_b->slack_started = false;
5134 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
5135 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
5139 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice)
5140 runtime = cfs_b->runtime;
5142 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
5147 distribute_cfs_runtime(cfs_b);
5151 * When a group wakes up we want to make sure that its quota is not already
5152 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
5153 * runtime as update_curr() throttling can not trigger until it's on-rq.
5155 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
5157 if (!cfs_bandwidth_used())
5160 /* an active group must be handled by the update_curr()->put() path */
5161 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
5164 /* ensure the group is not already throttled */
5165 if (cfs_rq_throttled(cfs_rq))
5168 /* update runtime allocation */
5169 account_cfs_rq_runtime(cfs_rq, 0);
5170 if (cfs_rq->runtime_remaining <= 0)
5171 throttle_cfs_rq(cfs_rq);
5174 static void sync_throttle(struct task_group *tg, int cpu)
5176 struct cfs_rq *pcfs_rq, *cfs_rq;
5178 if (!cfs_bandwidth_used())
5184 cfs_rq = tg->cfs_rq[cpu];
5185 pcfs_rq = tg->parent->cfs_rq[cpu];
5187 cfs_rq->throttle_count = pcfs_rq->throttle_count;
5188 cfs_rq->throttled_clock_task = rq_clock_task(cpu_rq(cpu));
5191 /* conditionally throttle active cfs_rq's from put_prev_entity() */
5192 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
5194 if (!cfs_bandwidth_used())
5197 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
5201 * it's possible for a throttled entity to be forced into a running
5202 * state (e.g. set_curr_task), in this case we're finished.
5204 if (cfs_rq_throttled(cfs_rq))
5207 return throttle_cfs_rq(cfs_rq);
5210 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
5212 struct cfs_bandwidth *cfs_b =
5213 container_of(timer, struct cfs_bandwidth, slack_timer);
5215 do_sched_cfs_slack_timer(cfs_b);
5217 return HRTIMER_NORESTART;
5220 extern const u64 max_cfs_quota_period;
5222 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
5224 struct cfs_bandwidth *cfs_b =
5225 container_of(timer, struct cfs_bandwidth, period_timer);
5226 unsigned long flags;
5231 raw_spin_lock_irqsave(&cfs_b->lock, flags);
5233 overrun = hrtimer_forward_now(timer, cfs_b->period);
5237 idle = do_sched_cfs_period_timer(cfs_b, overrun, flags);
5240 u64 new, old = ktime_to_ns(cfs_b->period);
5243 * Grow period by a factor of 2 to avoid losing precision.
5244 * Precision loss in the quota/period ratio can cause __cfs_schedulable
5248 if (new < max_cfs_quota_period) {
5249 cfs_b->period = ns_to_ktime(new);
5252 pr_warn_ratelimited(
5253 "cfs_period_timer[cpu%d]: period too short, scaling up (new cfs_period_us = %lld, cfs_quota_us = %lld)\n",
5255 div_u64(new, NSEC_PER_USEC),
5256 div_u64(cfs_b->quota, NSEC_PER_USEC));
5258 pr_warn_ratelimited(
5259 "cfs_period_timer[cpu%d]: period too short, but cannot scale up without losing precision (cfs_period_us = %lld, cfs_quota_us = %lld)\n",
5261 div_u64(old, NSEC_PER_USEC),
5262 div_u64(cfs_b->quota, NSEC_PER_USEC));
5265 /* reset count so we don't come right back in here */
5270 cfs_b->period_active = 0;
5271 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
5273 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
5276 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
5278 raw_spin_lock_init(&cfs_b->lock);
5280 cfs_b->quota = RUNTIME_INF;
5281 cfs_b->period = ns_to_ktime(default_cfs_period());
5283 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
5284 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED);
5285 cfs_b->period_timer.function = sched_cfs_period_timer;
5286 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
5287 cfs_b->slack_timer.function = sched_cfs_slack_timer;
5288 cfs_b->slack_started = false;
5291 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
5293 cfs_rq->runtime_enabled = 0;
5294 INIT_LIST_HEAD(&cfs_rq->throttled_list);
5297 void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
5299 lockdep_assert_held(&cfs_b->lock);
5301 if (cfs_b->period_active)
5304 cfs_b->period_active = 1;
5305 hrtimer_forward_now(&cfs_b->period_timer, cfs_b->period);
5306 hrtimer_start_expires(&cfs_b->period_timer, HRTIMER_MODE_ABS_PINNED);
5309 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
5311 /* init_cfs_bandwidth() was not called */
5312 if (!cfs_b->throttled_cfs_rq.next)
5315 hrtimer_cancel(&cfs_b->period_timer);
5316 hrtimer_cancel(&cfs_b->slack_timer);
5320 * Both these CPU hotplug callbacks race against unregister_fair_sched_group()
5322 * The race is harmless, since modifying bandwidth settings of unhooked group
5323 * bits doesn't do much.
5326 /* cpu online callback */
5327 static void __maybe_unused update_runtime_enabled(struct rq *rq)
5329 struct task_group *tg;
5331 lockdep_assert_held(&rq->lock);
5334 list_for_each_entry_rcu(tg, &task_groups, list) {
5335 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
5336 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
5338 raw_spin_lock(&cfs_b->lock);
5339 cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF;
5340 raw_spin_unlock(&cfs_b->lock);
5345 /* cpu offline callback */
5346 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
5348 struct task_group *tg;
5350 lockdep_assert_held(&rq->lock);
5353 list_for_each_entry_rcu(tg, &task_groups, list) {
5354 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
5356 if (!cfs_rq->runtime_enabled)
5360 * clock_task is not advancing so we just need to make sure
5361 * there's some valid quota amount
5363 cfs_rq->runtime_remaining = 1;
5365 * Offline rq is schedulable till CPU is completely disabled
5366 * in take_cpu_down(), so we prevent new cfs throttling here.
5368 cfs_rq->runtime_enabled = 0;
5370 if (cfs_rq_throttled(cfs_rq))
5371 unthrottle_cfs_rq(cfs_rq);
5376 #else /* CONFIG_CFS_BANDWIDTH */
5378 static inline bool cfs_bandwidth_used(void)
5383 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {}
5384 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; }
5385 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
5386 static inline void sync_throttle(struct task_group *tg, int cpu) {}
5387 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
5389 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
5394 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
5399 static inline int throttled_lb_pair(struct task_group *tg,
5400 int src_cpu, int dest_cpu)
5405 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
5407 #ifdef CONFIG_FAIR_GROUP_SCHED
5408 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
5411 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
5415 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
5416 static inline void update_runtime_enabled(struct rq *rq) {}
5417 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
5419 #endif /* CONFIG_CFS_BANDWIDTH */
5421 /**************************************************
5422 * CFS operations on tasks:
5425 #ifdef CONFIG_SCHED_HRTICK
5426 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
5428 struct sched_entity *se = &p->se;
5429 struct cfs_rq *cfs_rq = cfs_rq_of(se);
5431 SCHED_WARN_ON(task_rq(p) != rq);
5433 if (rq->cfs.h_nr_running > 1) {
5434 u64 slice = sched_slice(cfs_rq, se);
5435 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
5436 s64 delta = slice - ran;
5439 if (task_current(rq, p))
5443 hrtick_start(rq, delta);
5448 * called from enqueue/dequeue and updates the hrtick when the
5449 * current task is from our class and nr_running is low enough
5452 static void hrtick_update(struct rq *rq)
5454 struct task_struct *curr = rq->curr;
5456 if (!hrtick_enabled_fair(rq) || curr->sched_class != &fair_sched_class)
5459 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
5460 hrtick_start_fair(rq, curr);
5462 #else /* !CONFIG_SCHED_HRTICK */
5464 hrtick_start_fair(struct rq *rq, struct task_struct *p)
5468 static inline void hrtick_update(struct rq *rq)
5474 static inline unsigned long cpu_util(int cpu);
5476 static inline bool cpu_overutilized(int cpu)
5478 return !fits_capacity(cpu_util(cpu), capacity_of(cpu));
5481 static inline void update_overutilized_status(struct rq *rq)
5483 if (!READ_ONCE(rq->rd->overutilized) && cpu_overutilized(rq->cpu)) {
5484 WRITE_ONCE(rq->rd->overutilized, SG_OVERUTILIZED);
5485 trace_sched_overutilized_tp(rq->rd, SG_OVERUTILIZED);
5489 static inline void update_overutilized_status(struct rq *rq) { }
5492 /* Runqueue only has SCHED_IDLE tasks enqueued */
5493 static int sched_idle_rq(struct rq *rq)
5495 return unlikely(rq->nr_running == rq->cfs.idle_h_nr_running &&
5500 static int sched_idle_cpu(int cpu)
5502 return sched_idle_rq(cpu_rq(cpu));
5507 * The enqueue_task method is called before nr_running is
5508 * increased. Here we update the fair scheduling stats and
5509 * then put the task into the rbtree:
5512 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
5514 struct cfs_rq *cfs_rq;
5515 struct sched_entity *se = &p->se;
5516 int idle_h_nr_running = task_has_idle_policy(p);
5517 int task_new = !(flags & ENQUEUE_WAKEUP);
5520 * The code below (indirectly) updates schedutil which looks at
5521 * the cfs_rq utilization to select a frequency.
5522 * Let's add the task's estimated utilization to the cfs_rq's
5523 * estimated utilization, before we update schedutil.
5525 util_est_enqueue(&rq->cfs, p);
5528 * If in_iowait is set, the code below may not trigger any cpufreq
5529 * utilization updates, so do it here explicitly with the IOWAIT flag
5533 cpufreq_update_util(rq, SCHED_CPUFREQ_IOWAIT);
5535 for_each_sched_entity(se) {
5538 cfs_rq = cfs_rq_of(se);
5539 enqueue_entity(cfs_rq, se, flags);
5541 cfs_rq->h_nr_running++;
5542 cfs_rq->idle_h_nr_running += idle_h_nr_running;
5544 /* end evaluation on encountering a throttled cfs_rq */
5545 if (cfs_rq_throttled(cfs_rq))
5546 goto enqueue_throttle;
5548 flags = ENQUEUE_WAKEUP;
5551 for_each_sched_entity(se) {
5552 cfs_rq = cfs_rq_of(se);
5554 update_load_avg(cfs_rq, se, UPDATE_TG);
5555 se_update_runnable(se);
5556 update_cfs_group(se);
5558 cfs_rq->h_nr_running++;
5559 cfs_rq->idle_h_nr_running += idle_h_nr_running;
5561 /* end evaluation on encountering a throttled cfs_rq */
5562 if (cfs_rq_throttled(cfs_rq))
5563 goto enqueue_throttle;
5566 * One parent has been throttled and cfs_rq removed from the
5567 * list. Add it back to not break the leaf list.
5569 if (throttled_hierarchy(cfs_rq))
5570 list_add_leaf_cfs_rq(cfs_rq);
5573 /* At this point se is NULL and we are at root level*/
5574 add_nr_running(rq, 1);
5577 * Since new tasks are assigned an initial util_avg equal to
5578 * half of the spare capacity of their CPU, tiny tasks have the
5579 * ability to cross the overutilized threshold, which will
5580 * result in the load balancer ruining all the task placement
5581 * done by EAS. As a way to mitigate that effect, do not account
5582 * for the first enqueue operation of new tasks during the
5583 * overutilized flag detection.
5585 * A better way of solving this problem would be to wait for
5586 * the PELT signals of tasks to converge before taking them
5587 * into account, but that is not straightforward to implement,
5588 * and the following generally works well enough in practice.
5591 update_overutilized_status(rq);
5594 if (cfs_bandwidth_used()) {
5596 * When bandwidth control is enabled; the cfs_rq_throttled()
5597 * breaks in the above iteration can result in incomplete
5598 * leaf list maintenance, resulting in triggering the assertion
5601 for_each_sched_entity(se) {
5602 cfs_rq = cfs_rq_of(se);
5604 if (list_add_leaf_cfs_rq(cfs_rq))
5609 assert_list_leaf_cfs_rq(rq);
5614 static void set_next_buddy(struct sched_entity *se);
5617 * The dequeue_task method is called before nr_running is
5618 * decreased. We remove the task from the rbtree and
5619 * update the fair scheduling stats:
5621 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
5623 struct cfs_rq *cfs_rq;
5624 struct sched_entity *se = &p->se;
5625 int task_sleep = flags & DEQUEUE_SLEEP;
5626 int idle_h_nr_running = task_has_idle_policy(p);
5627 bool was_sched_idle = sched_idle_rq(rq);
5629 util_est_dequeue(&rq->cfs, p);
5631 for_each_sched_entity(se) {
5632 cfs_rq = cfs_rq_of(se);
5633 dequeue_entity(cfs_rq, se, flags);
5635 cfs_rq->h_nr_running--;
5636 cfs_rq->idle_h_nr_running -= idle_h_nr_running;
5638 /* end evaluation on encountering a throttled cfs_rq */
5639 if (cfs_rq_throttled(cfs_rq))
5640 goto dequeue_throttle;
5642 /* Don't dequeue parent if it has other entities besides us */
5643 if (cfs_rq->load.weight) {
5644 /* Avoid re-evaluating load for this entity: */
5645 se = parent_entity(se);
5647 * Bias pick_next to pick a task from this cfs_rq, as
5648 * p is sleeping when it is within its sched_slice.
5650 if (task_sleep && se && !throttled_hierarchy(cfs_rq))
5654 flags |= DEQUEUE_SLEEP;
5657 for_each_sched_entity(se) {
5658 cfs_rq = cfs_rq_of(se);
5660 update_load_avg(cfs_rq, se, UPDATE_TG);
5661 se_update_runnable(se);
5662 update_cfs_group(se);
5664 cfs_rq->h_nr_running--;
5665 cfs_rq->idle_h_nr_running -= idle_h_nr_running;
5667 /* end evaluation on encountering a throttled cfs_rq */
5668 if (cfs_rq_throttled(cfs_rq))
5669 goto dequeue_throttle;
5673 /* At this point se is NULL and we are at root level*/
5674 sub_nr_running(rq, 1);
5676 /* balance early to pull high priority tasks */
5677 if (unlikely(!was_sched_idle && sched_idle_rq(rq)))
5678 rq->next_balance = jiffies;
5681 util_est_update(&rq->cfs, p, task_sleep);
5687 /* Working cpumask for: load_balance, load_balance_newidle. */
5688 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
5689 DEFINE_PER_CPU(cpumask_var_t, select_idle_mask);
5691 #ifdef CONFIG_NO_HZ_COMMON
5694 cpumask_var_t idle_cpus_mask;
5696 int has_blocked; /* Idle CPUS has blocked load */
5697 unsigned long next_balance; /* in jiffy units */
5698 unsigned long next_blocked; /* Next update of blocked load in jiffies */
5699 } nohz ____cacheline_aligned;
5701 #endif /* CONFIG_NO_HZ_COMMON */
5703 static unsigned long cpu_load(struct rq *rq)
5705 return cfs_rq_load_avg(&rq->cfs);
5709 * cpu_load_without - compute CPU load without any contributions from *p
5710 * @cpu: the CPU which load is requested
5711 * @p: the task which load should be discounted
5713 * The load of a CPU is defined by the load of tasks currently enqueued on that
5714 * CPU as well as tasks which are currently sleeping after an execution on that
5717 * This method returns the load of the specified CPU by discounting the load of
5718 * the specified task, whenever the task is currently contributing to the CPU
5721 static unsigned long cpu_load_without(struct rq *rq, struct task_struct *p)
5723 struct cfs_rq *cfs_rq;
5726 /* Task has no contribution or is new */
5727 if (cpu_of(rq) != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time))
5728 return cpu_load(rq);
5731 load = READ_ONCE(cfs_rq->avg.load_avg);
5733 /* Discount task's util from CPU's util */
5734 lsub_positive(&load, task_h_load(p));
5739 static unsigned long cpu_runnable(struct rq *rq)
5741 return cfs_rq_runnable_avg(&rq->cfs);
5744 static unsigned long cpu_runnable_without(struct rq *rq, struct task_struct *p)
5746 struct cfs_rq *cfs_rq;
5747 unsigned int runnable;
5749 /* Task has no contribution or is new */
5750 if (cpu_of(rq) != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time))
5751 return cpu_runnable(rq);
5754 runnable = READ_ONCE(cfs_rq->avg.runnable_avg);
5756 /* Discount task's runnable from CPU's runnable */
5757 lsub_positive(&runnable, p->se.avg.runnable_avg);
5762 static unsigned long capacity_of(int cpu)
5764 return cpu_rq(cpu)->cpu_capacity;
5767 static void record_wakee(struct task_struct *p)
5770 * Only decay a single time; tasks that have less then 1 wakeup per
5771 * jiffy will not have built up many flips.
5773 if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) {
5774 current->wakee_flips >>= 1;
5775 current->wakee_flip_decay_ts = jiffies;
5778 if (current->last_wakee != p) {
5779 current->last_wakee = p;
5780 current->wakee_flips++;
5785 * Detect M:N waker/wakee relationships via a switching-frequency heuristic.
5787 * A waker of many should wake a different task than the one last awakened
5788 * at a frequency roughly N times higher than one of its wakees.
5790 * In order to determine whether we should let the load spread vs consolidating
5791 * to shared cache, we look for a minimum 'flip' frequency of llc_size in one
5792 * partner, and a factor of lls_size higher frequency in the other.
5794 * With both conditions met, we can be relatively sure that the relationship is
5795 * non-monogamous, with partner count exceeding socket size.
5797 * Waker/wakee being client/server, worker/dispatcher, interrupt source or
5798 * whatever is irrelevant, spread criteria is apparent partner count exceeds
5801 static int wake_wide(struct task_struct *p)
5803 unsigned int master = current->wakee_flips;
5804 unsigned int slave = p->wakee_flips;
5805 int factor = __this_cpu_read(sd_llc_size);
5808 swap(master, slave);
5809 if (slave < factor || master < slave * factor)
5815 * The purpose of wake_affine() is to quickly determine on which CPU we can run
5816 * soonest. For the purpose of speed we only consider the waking and previous
5819 * wake_affine_idle() - only considers 'now', it check if the waking CPU is
5820 * cache-affine and is (or will be) idle.
5822 * wake_affine_weight() - considers the weight to reflect the average
5823 * scheduling latency of the CPUs. This seems to work
5824 * for the overloaded case.
5827 wake_affine_idle(int this_cpu, int prev_cpu, int sync)
5830 * If this_cpu is idle, it implies the wakeup is from interrupt
5831 * context. Only allow the move if cache is shared. Otherwise an
5832 * interrupt intensive workload could force all tasks onto one
5833 * node depending on the IO topology or IRQ affinity settings.
5835 * If the prev_cpu is idle and cache affine then avoid a migration.
5836 * There is no guarantee that the cache hot data from an interrupt
5837 * is more important than cache hot data on the prev_cpu and from
5838 * a cpufreq perspective, it's better to have higher utilisation
5841 if (available_idle_cpu(this_cpu) && cpus_share_cache(this_cpu, prev_cpu))
5842 return available_idle_cpu(prev_cpu) ? prev_cpu : this_cpu;
5844 if (sync && cpu_rq(this_cpu)->nr_running == 1)
5847 if (available_idle_cpu(prev_cpu))
5850 return nr_cpumask_bits;
5854 wake_affine_weight(struct sched_domain *sd, struct task_struct *p,
5855 int this_cpu, int prev_cpu, int sync)
5857 s64 this_eff_load, prev_eff_load;
5858 unsigned long task_load;
5860 this_eff_load = cpu_load(cpu_rq(this_cpu));
5863 unsigned long current_load = task_h_load(current);
5865 if (current_load > this_eff_load)
5868 this_eff_load -= current_load;
5871 task_load = task_h_load(p);
5873 this_eff_load += task_load;
5874 if (sched_feat(WA_BIAS))
5875 this_eff_load *= 100;
5876 this_eff_load *= capacity_of(prev_cpu);
5878 prev_eff_load = cpu_load(cpu_rq(prev_cpu));
5879 prev_eff_load -= task_load;
5880 if (sched_feat(WA_BIAS))
5881 prev_eff_load *= 100 + (sd->imbalance_pct - 100) / 2;
5882 prev_eff_load *= capacity_of(this_cpu);
5885 * If sync, adjust the weight of prev_eff_load such that if
5886 * prev_eff == this_eff that select_idle_sibling() will consider
5887 * stacking the wakee on top of the waker if no other CPU is
5893 return this_eff_load < prev_eff_load ? this_cpu : nr_cpumask_bits;
5896 static int wake_affine(struct sched_domain *sd, struct task_struct *p,
5897 int this_cpu, int prev_cpu, int sync)
5899 int target = nr_cpumask_bits;
5901 if (sched_feat(WA_IDLE))
5902 target = wake_affine_idle(this_cpu, prev_cpu, sync);
5904 if (sched_feat(WA_WEIGHT) && target == nr_cpumask_bits)
5905 target = wake_affine_weight(sd, p, this_cpu, prev_cpu, sync);
5907 schedstat_inc(p->se.statistics.nr_wakeups_affine_attempts);
5908 if (target == nr_cpumask_bits)
5911 schedstat_inc(sd->ttwu_move_affine);
5912 schedstat_inc(p->se.statistics.nr_wakeups_affine);
5916 static struct sched_group *
5917 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu);
5920 * find_idlest_group_cpu - find the idlest CPU among the CPUs in the group.
5923 find_idlest_group_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
5925 unsigned long load, min_load = ULONG_MAX;
5926 unsigned int min_exit_latency = UINT_MAX;
5927 u64 latest_idle_timestamp = 0;
5928 int least_loaded_cpu = this_cpu;
5929 int shallowest_idle_cpu = -1;
5932 /* Check if we have any choice: */
5933 if (group->group_weight == 1)
5934 return cpumask_first(sched_group_span(group));
5936 /* Traverse only the allowed CPUs */
5937 for_each_cpu_and(i, sched_group_span(group), p->cpus_ptr) {
5938 if (sched_idle_cpu(i))
5941 if (available_idle_cpu(i)) {
5942 struct rq *rq = cpu_rq(i);
5943 struct cpuidle_state *idle = idle_get_state(rq);
5944 if (idle && idle->exit_latency < min_exit_latency) {
5946 * We give priority to a CPU whose idle state
5947 * has the smallest exit latency irrespective
5948 * of any idle timestamp.
5950 min_exit_latency = idle->exit_latency;
5951 latest_idle_timestamp = rq->idle_stamp;
5952 shallowest_idle_cpu = i;
5953 } else if ((!idle || idle->exit_latency == min_exit_latency) &&
5954 rq->idle_stamp > latest_idle_timestamp) {
5956 * If equal or no active idle state, then
5957 * the most recently idled CPU might have
5960 latest_idle_timestamp = rq->idle_stamp;
5961 shallowest_idle_cpu = i;
5963 } else if (shallowest_idle_cpu == -1) {
5964 load = cpu_load(cpu_rq(i));
5965 if (load < min_load) {
5967 least_loaded_cpu = i;
5972 return shallowest_idle_cpu != -1 ? shallowest_idle_cpu : least_loaded_cpu;
5975 static inline int find_idlest_cpu(struct sched_domain *sd, struct task_struct *p,
5976 int cpu, int prev_cpu, int sd_flag)
5980 if (!cpumask_intersects(sched_domain_span(sd), p->cpus_ptr))
5984 * We need task's util for cpu_util_without, sync it up to
5985 * prev_cpu's last_update_time.
5987 if (!(sd_flag & SD_BALANCE_FORK))
5988 sync_entity_load_avg(&p->se);
5991 struct sched_group *group;
5992 struct sched_domain *tmp;
5995 if (!(sd->flags & sd_flag)) {
6000 group = find_idlest_group(sd, p, cpu);
6006 new_cpu = find_idlest_group_cpu(group, p, cpu);
6007 if (new_cpu == cpu) {
6008 /* Now try balancing at a lower domain level of 'cpu': */
6013 /* Now try balancing at a lower domain level of 'new_cpu': */
6015 weight = sd->span_weight;
6017 for_each_domain(cpu, tmp) {
6018 if (weight <= tmp->span_weight)
6020 if (tmp->flags & sd_flag)
6028 static inline int __select_idle_cpu(int cpu)
6030 if (available_idle_cpu(cpu) || sched_idle_cpu(cpu))
6036 #ifdef CONFIG_SCHED_SMT
6037 DEFINE_STATIC_KEY_FALSE(sched_smt_present);
6038 EXPORT_SYMBOL_GPL(sched_smt_present);
6040 static inline void set_idle_cores(int cpu, int val)
6042 struct sched_domain_shared *sds;
6044 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
6046 WRITE_ONCE(sds->has_idle_cores, val);
6049 static inline bool test_idle_cores(int cpu, bool def)
6051 struct sched_domain_shared *sds;
6053 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
6055 return READ_ONCE(sds->has_idle_cores);
6061 * Scans the local SMT mask to see if the entire core is idle, and records this
6062 * information in sd_llc_shared->has_idle_cores.
6064 * Since SMT siblings share all cache levels, inspecting this limited remote
6065 * state should be fairly cheap.
6067 void __update_idle_core(struct rq *rq)
6069 int core = cpu_of(rq);
6073 if (test_idle_cores(core, true))
6076 for_each_cpu(cpu, cpu_smt_mask(core)) {
6080 if (!available_idle_cpu(cpu))
6084 set_idle_cores(core, 1);
6090 * Scan the entire LLC domain for idle cores; this dynamically switches off if
6091 * there are no idle cores left in the system; tracked through
6092 * sd_llc->shared->has_idle_cores and enabled through update_idle_core() above.
6094 static int select_idle_core(struct task_struct *p, int core, struct cpumask *cpus, int *idle_cpu)
6099 if (!static_branch_likely(&sched_smt_present))
6100 return __select_idle_cpu(core);
6102 for_each_cpu(cpu, cpu_smt_mask(core)) {
6103 if (!available_idle_cpu(cpu)) {
6105 if (*idle_cpu == -1) {
6106 if (sched_idle_cpu(cpu) && cpumask_test_cpu(cpu, p->cpus_ptr)) {
6114 if (*idle_cpu == -1 && cpumask_test_cpu(cpu, p->cpus_ptr))
6121 cpumask_andnot(cpus, cpus, cpu_smt_mask(core));
6126 * Scan the local SMT mask for idle CPUs.
6128 static int select_idle_smt(struct task_struct *p, struct sched_domain *sd, int target)
6132 for_each_cpu(cpu, cpu_smt_mask(target)) {
6133 if (!cpumask_test_cpu(cpu, p->cpus_ptr) ||
6134 !cpumask_test_cpu(cpu, sched_domain_span(sd)))
6136 if (available_idle_cpu(cpu) || sched_idle_cpu(cpu))
6143 #else /* CONFIG_SCHED_SMT */
6145 static inline void set_idle_cores(int cpu, int val)
6149 static inline bool test_idle_cores(int cpu, bool def)
6154 static inline int select_idle_core(struct task_struct *p, int core, struct cpumask *cpus, int *idle_cpu)
6156 return __select_idle_cpu(core);
6159 static inline int select_idle_smt(struct task_struct *p, struct sched_domain *sd, int target)
6164 #endif /* CONFIG_SCHED_SMT */
6167 * Scan the LLC domain for idle CPUs; this is dynamically regulated by
6168 * comparing the average scan cost (tracked in sd->avg_scan_cost) against the
6169 * average idle time for this rq (as found in rq->avg_idle).
6171 static int select_idle_cpu(struct task_struct *p, struct sched_domain *sd, bool has_idle_core, int target)
6173 struct cpumask *cpus = this_cpu_cpumask_var_ptr(select_idle_mask);
6174 int i, cpu, idle_cpu = -1, nr = INT_MAX;
6175 int this = smp_processor_id();
6176 struct sched_domain *this_sd;
6179 this_sd = rcu_dereference(*this_cpu_ptr(&sd_llc));
6183 cpumask_and(cpus, sched_domain_span(sd), p->cpus_ptr);
6185 if (sched_feat(SIS_PROP) && !has_idle_core) {
6186 u64 avg_cost, avg_idle, span_avg;
6189 * Due to large variance we need a large fuzz factor;
6190 * hackbench in particularly is sensitive here.
6192 avg_idle = this_rq()->avg_idle / 512;
6193 avg_cost = this_sd->avg_scan_cost + 1;
6195 span_avg = sd->span_weight * avg_idle;
6196 if (span_avg > 4*avg_cost)
6197 nr = div_u64(span_avg, avg_cost);
6201 time = cpu_clock(this);
6204 for_each_cpu_wrap(cpu, cpus, target) {
6205 if (has_idle_core) {
6206 i = select_idle_core(p, cpu, cpus, &idle_cpu);
6207 if ((unsigned int)i < nr_cpumask_bits)
6213 idle_cpu = __select_idle_cpu(cpu);
6214 if ((unsigned int)idle_cpu < nr_cpumask_bits)
6220 set_idle_cores(target, false);
6222 if (sched_feat(SIS_PROP) && !has_idle_core) {
6223 time = cpu_clock(this) - time;
6224 update_avg(&this_sd->avg_scan_cost, time);
6231 * Scan the asym_capacity domain for idle CPUs; pick the first idle one on which
6232 * the task fits. If no CPU is big enough, but there are idle ones, try to
6233 * maximize capacity.
6236 select_idle_capacity(struct task_struct *p, struct sched_domain *sd, int target)
6238 unsigned long task_util, best_cap = 0;
6239 int cpu, best_cpu = -1;
6240 struct cpumask *cpus;
6242 cpus = this_cpu_cpumask_var_ptr(select_idle_mask);
6243 cpumask_and(cpus, sched_domain_span(sd), p->cpus_ptr);
6245 task_util = uclamp_task_util(p);
6247 for_each_cpu_wrap(cpu, cpus, target) {
6248 unsigned long cpu_cap = capacity_of(cpu);
6250 if (!available_idle_cpu(cpu) && !sched_idle_cpu(cpu))
6252 if (fits_capacity(task_util, cpu_cap))
6255 if (cpu_cap > best_cap) {
6264 static inline bool asym_fits_capacity(int task_util, int cpu)
6266 if (static_branch_unlikely(&sched_asym_cpucapacity))
6267 return fits_capacity(task_util, capacity_of(cpu));
6273 * Try and locate an idle core/thread in the LLC cache domain.
6275 static int select_idle_sibling(struct task_struct *p, int prev, int target)
6277 bool has_idle_core = false;
6278 struct sched_domain *sd;
6279 unsigned long task_util;
6280 int i, recent_used_cpu;
6283 * On asymmetric system, update task utilization because we will check
6284 * that the task fits with cpu's capacity.
6286 if (static_branch_unlikely(&sched_asym_cpucapacity)) {
6287 sync_entity_load_avg(&p->se);
6288 task_util = uclamp_task_util(p);
6291 if ((available_idle_cpu(target) || sched_idle_cpu(target)) &&
6292 asym_fits_capacity(task_util, target))
6296 * If the previous CPU is cache affine and idle, don't be stupid:
6298 if (prev != target && cpus_share_cache(prev, target) &&
6299 (available_idle_cpu(prev) || sched_idle_cpu(prev)) &&
6300 asym_fits_capacity(task_util, prev))
6304 * Allow a per-cpu kthread to stack with the wakee if the
6305 * kworker thread and the tasks previous CPUs are the same.
6306 * The assumption is that the wakee queued work for the
6307 * per-cpu kthread that is now complete and the wakeup is
6308 * essentially a sync wakeup. An obvious example of this
6309 * pattern is IO completions.
6311 if (is_per_cpu_kthread(current) &&
6312 prev == smp_processor_id() &&
6313 this_rq()->nr_running <= 1) {
6317 /* Check a recently used CPU as a potential idle candidate: */
6318 recent_used_cpu = p->recent_used_cpu;
6319 if (recent_used_cpu != prev &&
6320 recent_used_cpu != target &&
6321 cpus_share_cache(recent_used_cpu, target) &&
6322 (available_idle_cpu(recent_used_cpu) || sched_idle_cpu(recent_used_cpu)) &&
6323 cpumask_test_cpu(p->recent_used_cpu, p->cpus_ptr) &&
6324 asym_fits_capacity(task_util, recent_used_cpu)) {
6326 * Replace recent_used_cpu with prev as it is a potential
6327 * candidate for the next wake:
6329 p->recent_used_cpu = prev;
6330 return recent_used_cpu;
6334 * For asymmetric CPU capacity systems, our domain of interest is
6335 * sd_asym_cpucapacity rather than sd_llc.
6337 if (static_branch_unlikely(&sched_asym_cpucapacity)) {
6338 sd = rcu_dereference(per_cpu(sd_asym_cpucapacity, target));
6340 * On an asymmetric CPU capacity system where an exclusive
6341 * cpuset defines a symmetric island (i.e. one unique
6342 * capacity_orig value through the cpuset), the key will be set
6343 * but the CPUs within that cpuset will not have a domain with
6344 * SD_ASYM_CPUCAPACITY. These should follow the usual symmetric
6348 i = select_idle_capacity(p, sd, target);
6349 return ((unsigned)i < nr_cpumask_bits) ? i : target;
6353 sd = rcu_dereference(per_cpu(sd_llc, target));
6357 if (sched_smt_active()) {
6358 has_idle_core = test_idle_cores(target, false);
6360 if (!has_idle_core && cpus_share_cache(prev, target)) {
6361 i = select_idle_smt(p, sd, prev);
6362 if ((unsigned int)i < nr_cpumask_bits)
6367 i = select_idle_cpu(p, sd, has_idle_core, target);
6368 if ((unsigned)i < nr_cpumask_bits)
6375 * cpu_util - Estimates the amount of capacity of a CPU used by CFS tasks.
6376 * @cpu: the CPU to get the utilization of
6378 * The unit of the return value must be the one of capacity so we can compare
6379 * the utilization with the capacity of the CPU that is available for CFS task
6380 * (ie cpu_capacity).
6382 * cfs_rq.avg.util_avg is the sum of running time of runnable tasks plus the
6383 * recent utilization of currently non-runnable tasks on a CPU. It represents
6384 * the amount of utilization of a CPU in the range [0..capacity_orig] where
6385 * capacity_orig is the cpu_capacity available at the highest frequency
6386 * (arch_scale_freq_capacity()).
6387 * The utilization of a CPU converges towards a sum equal to or less than the
6388 * current capacity (capacity_curr <= capacity_orig) of the CPU because it is
6389 * the running time on this CPU scaled by capacity_curr.
6391 * The estimated utilization of a CPU is defined to be the maximum between its
6392 * cfs_rq.avg.util_avg and the sum of the estimated utilization of the tasks
6393 * currently RUNNABLE on that CPU.
6394 * This allows to properly represent the expected utilization of a CPU which
6395 * has just got a big task running since a long sleep period. At the same time
6396 * however it preserves the benefits of the "blocked utilization" in
6397 * describing the potential for other tasks waking up on the same CPU.
6399 * Nevertheless, cfs_rq.avg.util_avg can be higher than capacity_curr or even
6400 * higher than capacity_orig because of unfortunate rounding in
6401 * cfs.avg.util_avg or just after migrating tasks and new task wakeups until
6402 * the average stabilizes with the new running time. We need to check that the
6403 * utilization stays within the range of [0..capacity_orig] and cap it if
6404 * necessary. Without utilization capping, a group could be seen as overloaded
6405 * (CPU0 utilization at 121% + CPU1 utilization at 80%) whereas CPU1 has 20% of
6406 * available capacity. We allow utilization to overshoot capacity_curr (but not
6407 * capacity_orig) as it useful for predicting the capacity required after task
6408 * migrations (scheduler-driven DVFS).
6410 * Return: the (estimated) utilization for the specified CPU
6412 static inline unsigned long cpu_util(int cpu)
6414 struct cfs_rq *cfs_rq;
6417 cfs_rq = &cpu_rq(cpu)->cfs;
6418 util = READ_ONCE(cfs_rq->avg.util_avg);
6420 if (sched_feat(UTIL_EST))
6421 util = max(util, READ_ONCE(cfs_rq->avg.util_est.enqueued));
6423 return min_t(unsigned long, util, capacity_orig_of(cpu));
6427 * cpu_util_without: compute cpu utilization without any contributions from *p
6428 * @cpu: the CPU which utilization is requested
6429 * @p: the task which utilization should be discounted
6431 * The utilization of a CPU is defined by the utilization of tasks currently
6432 * enqueued on that CPU as well as tasks which are currently sleeping after an
6433 * execution on that CPU.
6435 * This method returns the utilization of the specified CPU by discounting the
6436 * utilization of the specified task, whenever the task is currently
6437 * contributing to the CPU utilization.
6439 static unsigned long cpu_util_without(int cpu, struct task_struct *p)
6441 struct cfs_rq *cfs_rq;
6444 /* Task has no contribution or is new */
6445 if (cpu != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time))
6446 return cpu_util(cpu);
6448 cfs_rq = &cpu_rq(cpu)->cfs;
6449 util = READ_ONCE(cfs_rq->avg.util_avg);
6451 /* Discount task's util from CPU's util */
6452 lsub_positive(&util, task_util(p));
6457 * a) if *p is the only task sleeping on this CPU, then:
6458 * cpu_util (== task_util) > util_est (== 0)
6459 * and thus we return:
6460 * cpu_util_without = (cpu_util - task_util) = 0
6462 * b) if other tasks are SLEEPING on this CPU, which is now exiting
6464 * cpu_util >= task_util
6465 * cpu_util > util_est (== 0)
6466 * and thus we discount *p's blocked utilization to return:
6467 * cpu_util_without = (cpu_util - task_util) >= 0
6469 * c) if other tasks are RUNNABLE on that CPU and
6470 * util_est > cpu_util
6471 * then we use util_est since it returns a more restrictive
6472 * estimation of the spare capacity on that CPU, by just
6473 * considering the expected utilization of tasks already
6474 * runnable on that CPU.
6476 * Cases a) and b) are covered by the above code, while case c) is
6477 * covered by the following code when estimated utilization is
6480 if (sched_feat(UTIL_EST)) {
6481 unsigned int estimated =
6482 READ_ONCE(cfs_rq->avg.util_est.enqueued);
6485 * Despite the following checks we still have a small window
6486 * for a possible race, when an execl's select_task_rq_fair()
6487 * races with LB's detach_task():
6490 * p->on_rq = TASK_ON_RQ_MIGRATING;
6491 * ---------------------------------- A
6492 * deactivate_task() \
6493 * dequeue_task() + RaceTime
6494 * util_est_dequeue() /
6495 * ---------------------------------- B
6497 * The additional check on "current == p" it's required to
6498 * properly fix the execl regression and it helps in further
6499 * reducing the chances for the above race.
6501 if (unlikely(task_on_rq_queued(p) || current == p))
6502 lsub_positive(&estimated, _task_util_est(p));
6504 util = max(util, estimated);
6508 * Utilization (estimated) can exceed the CPU capacity, thus let's
6509 * clamp to the maximum CPU capacity to ensure consistency with
6510 * the cpu_util call.
6512 return min_t(unsigned long, util, capacity_orig_of(cpu));
6516 * Predicts what cpu_util(@cpu) would return if @p was migrated (and enqueued)
6519 static unsigned long cpu_util_next(int cpu, struct task_struct *p, int dst_cpu)
6521 struct cfs_rq *cfs_rq = &cpu_rq(cpu)->cfs;
6522 unsigned long util_est, util = READ_ONCE(cfs_rq->avg.util_avg);
6525 * If @p migrates from @cpu to another, remove its contribution. Or,
6526 * if @p migrates from another CPU to @cpu, add its contribution. In
6527 * the other cases, @cpu is not impacted by the migration, so the
6528 * util_avg should already be correct.
6530 if (task_cpu(p) == cpu && dst_cpu != cpu)
6531 lsub_positive(&util, task_util(p));
6532 else if (task_cpu(p) != cpu && dst_cpu == cpu)
6533 util += task_util(p);
6535 if (sched_feat(UTIL_EST)) {
6536 util_est = READ_ONCE(cfs_rq->avg.util_est.enqueued);
6539 * During wake-up, the task isn't enqueued yet and doesn't
6540 * appear in the cfs_rq->avg.util_est.enqueued of any rq,
6541 * so just add it (if needed) to "simulate" what will be
6542 * cpu_util() after the task has been enqueued.
6545 util_est += _task_util_est(p);
6547 util = max(util, util_est);
6550 return min(util, capacity_orig_of(cpu));
6554 * compute_energy(): Estimates the energy that @pd would consume if @p was
6555 * migrated to @dst_cpu. compute_energy() predicts what will be the utilization
6556 * landscape of @pd's CPUs after the task migration, and uses the Energy Model
6557 * to compute what would be the energy if we decided to actually migrate that
6561 compute_energy(struct task_struct *p, int dst_cpu, struct perf_domain *pd)
6563 struct cpumask *pd_mask = perf_domain_span(pd);
6564 unsigned long cpu_cap = arch_scale_cpu_capacity(cpumask_first(pd_mask));
6565 unsigned long max_util = 0, sum_util = 0;
6569 * The capacity state of CPUs of the current rd can be driven by CPUs
6570 * of another rd if they belong to the same pd. So, account for the
6571 * utilization of these CPUs too by masking pd with cpu_online_mask
6572 * instead of the rd span.
6574 * If an entire pd is outside of the current rd, it will not appear in
6575 * its pd list and will not be accounted by compute_energy().
6577 for_each_cpu_and(cpu, pd_mask, cpu_online_mask) {
6578 unsigned long util_freq = cpu_util_next(cpu, p, dst_cpu);
6579 unsigned long cpu_util, util_running = util_freq;
6580 struct task_struct *tsk = NULL;
6583 * When @p is placed on @cpu:
6585 * util_running = max(cpu_util, cpu_util_est) +
6586 * max(task_util, _task_util_est)
6588 * while cpu_util_next is: max(cpu_util + task_util,
6589 * cpu_util_est + _task_util_est)
6591 if (cpu == dst_cpu) {
6594 cpu_util_next(cpu, p, -1) + task_util_est(p);
6598 * Busy time computation: utilization clamping is not
6599 * required since the ratio (sum_util / cpu_capacity)
6600 * is already enough to scale the EM reported power
6601 * consumption at the (eventually clamped) cpu_capacity.
6603 sum_util += effective_cpu_util(cpu, util_running, cpu_cap,
6607 * Performance domain frequency: utilization clamping
6608 * must be considered since it affects the selection
6609 * of the performance domain frequency.
6610 * NOTE: in case RT tasks are running, by default the
6611 * FREQUENCY_UTIL's utilization can be max OPP.
6613 cpu_util = effective_cpu_util(cpu, util_freq, cpu_cap,
6614 FREQUENCY_UTIL, tsk);
6615 max_util = max(max_util, cpu_util);
6618 return em_cpu_energy(pd->em_pd, max_util, sum_util);
6622 * find_energy_efficient_cpu(): Find most energy-efficient target CPU for the
6623 * waking task. find_energy_efficient_cpu() looks for the CPU with maximum
6624 * spare capacity in each performance domain and uses it as a potential
6625 * candidate to execute the task. Then, it uses the Energy Model to figure
6626 * out which of the CPU candidates is the most energy-efficient.
6628 * The rationale for this heuristic is as follows. In a performance domain,
6629 * all the most energy efficient CPU candidates (according to the Energy
6630 * Model) are those for which we'll request a low frequency. When there are
6631 * several CPUs for which the frequency request will be the same, we don't
6632 * have enough data to break the tie between them, because the Energy Model
6633 * only includes active power costs. With this model, if we assume that
6634 * frequency requests follow utilization (e.g. using schedutil), the CPU with
6635 * the maximum spare capacity in a performance domain is guaranteed to be among
6636 * the best candidates of the performance domain.
6638 * In practice, it could be preferable from an energy standpoint to pack
6639 * small tasks on a CPU in order to let other CPUs go in deeper idle states,
6640 * but that could also hurt our chances to go cluster idle, and we have no
6641 * ways to tell with the current Energy Model if this is actually a good
6642 * idea or not. So, find_energy_efficient_cpu() basically favors
6643 * cluster-packing, and spreading inside a cluster. That should at least be
6644 * a good thing for latency, and this is consistent with the idea that most
6645 * of the energy savings of EAS come from the asymmetry of the system, and
6646 * not so much from breaking the tie between identical CPUs. That's also the
6647 * reason why EAS is enabled in the topology code only for systems where
6648 * SD_ASYM_CPUCAPACITY is set.
6650 * NOTE: Forkees are not accepted in the energy-aware wake-up path because
6651 * they don't have any useful utilization data yet and it's not possible to
6652 * forecast their impact on energy consumption. Consequently, they will be
6653 * placed by find_idlest_cpu() on the least loaded CPU, which might turn out
6654 * to be energy-inefficient in some use-cases. The alternative would be to
6655 * bias new tasks towards specific types of CPUs first, or to try to infer
6656 * their util_avg from the parent task, but those heuristics could hurt
6657 * other use-cases too. So, until someone finds a better way to solve this,
6658 * let's keep things simple by re-using the existing slow path.
6660 static int find_energy_efficient_cpu(struct task_struct *p, int prev_cpu)
6662 unsigned long prev_delta = ULONG_MAX, best_delta = ULONG_MAX;
6663 struct root_domain *rd = cpu_rq(smp_processor_id())->rd;
6664 unsigned long cpu_cap, util, base_energy = 0;
6665 int cpu, best_energy_cpu = prev_cpu;
6666 struct sched_domain *sd;
6667 struct perf_domain *pd;
6670 pd = rcu_dereference(rd->pd);
6671 if (!pd || READ_ONCE(rd->overutilized))
6675 * Energy-aware wake-up happens on the lowest sched_domain starting
6676 * from sd_asym_cpucapacity spanning over this_cpu and prev_cpu.
6678 sd = rcu_dereference(*this_cpu_ptr(&sd_asym_cpucapacity));
6679 while (sd && !cpumask_test_cpu(prev_cpu, sched_domain_span(sd)))
6684 sync_entity_load_avg(&p->se);
6685 if (!task_util_est(p))
6688 for (; pd; pd = pd->next) {
6689 unsigned long cur_delta, spare_cap, max_spare_cap = 0;
6690 bool compute_prev_delta = false;
6691 unsigned long base_energy_pd;
6692 int max_spare_cap_cpu = -1;
6694 for_each_cpu_and(cpu, perf_domain_span(pd), sched_domain_span(sd)) {
6695 if (!cpumask_test_cpu(cpu, p->cpus_ptr))
6698 util = cpu_util_next(cpu, p, cpu);
6699 cpu_cap = capacity_of(cpu);
6700 spare_cap = cpu_cap;
6701 lsub_positive(&spare_cap, util);
6704 * Skip CPUs that cannot satisfy the capacity request.
6705 * IOW, placing the task there would make the CPU
6706 * overutilized. Take uclamp into account to see how
6707 * much capacity we can get out of the CPU; this is
6708 * aligned with sched_cpu_util().
6710 util = uclamp_rq_util_with(cpu_rq(cpu), util, p);
6711 if (!fits_capacity(util, cpu_cap))
6714 if (cpu == prev_cpu) {
6715 /* Always use prev_cpu as a candidate. */
6716 compute_prev_delta = true;
6717 } else if (spare_cap > max_spare_cap) {
6719 * Find the CPU with the maximum spare capacity
6720 * in the performance domain.
6722 max_spare_cap = spare_cap;
6723 max_spare_cap_cpu = cpu;
6727 if (max_spare_cap_cpu < 0 && !compute_prev_delta)
6730 /* Compute the 'base' energy of the pd, without @p */
6731 base_energy_pd = compute_energy(p, -1, pd);
6732 base_energy += base_energy_pd;
6734 /* Evaluate the energy impact of using prev_cpu. */
6735 if (compute_prev_delta) {
6736 prev_delta = compute_energy(p, prev_cpu, pd);
6737 prev_delta -= base_energy_pd;
6738 best_delta = min(best_delta, prev_delta);
6741 /* Evaluate the energy impact of using max_spare_cap_cpu. */
6742 if (max_spare_cap_cpu >= 0) {
6743 cur_delta = compute_energy(p, max_spare_cap_cpu, pd);
6744 cur_delta -= base_energy_pd;
6745 if (cur_delta < best_delta) {
6746 best_delta = cur_delta;
6747 best_energy_cpu = max_spare_cap_cpu;
6755 * Pick the best CPU if prev_cpu cannot be used, or if it saves at
6756 * least 6% of the energy used by prev_cpu.
6758 if (prev_delta == ULONG_MAX)
6759 return best_energy_cpu;
6761 if ((prev_delta - best_delta) > ((prev_delta + base_energy) >> 4))
6762 return best_energy_cpu;
6773 * select_task_rq_fair: Select target runqueue for the waking task in domains
6774 * that have the relevant SD flag set. In practice, this is SD_BALANCE_WAKE,
6775 * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
6777 * Balances load by selecting the idlest CPU in the idlest group, or under
6778 * certain conditions an idle sibling CPU if the domain has SD_WAKE_AFFINE set.
6780 * Returns the target CPU number.
6782 * preempt must be disabled.
6785 select_task_rq_fair(struct task_struct *p, int prev_cpu, int wake_flags)
6787 int sync = (wake_flags & WF_SYNC) && !(current->flags & PF_EXITING);
6788 struct sched_domain *tmp, *sd = NULL;
6789 int cpu = smp_processor_id();
6790 int new_cpu = prev_cpu;
6791 int want_affine = 0;
6792 /* SD_flags and WF_flags share the first nibble */
6793 int sd_flag = wake_flags & 0xF;
6795 if (wake_flags & WF_TTWU) {
6798 if (sched_energy_enabled()) {
6799 new_cpu = find_energy_efficient_cpu(p, prev_cpu);
6805 want_affine = !wake_wide(p) && cpumask_test_cpu(cpu, p->cpus_ptr);
6809 for_each_domain(cpu, tmp) {
6811 * If both 'cpu' and 'prev_cpu' are part of this domain,
6812 * cpu is a valid SD_WAKE_AFFINE target.
6814 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
6815 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
6816 if (cpu != prev_cpu)
6817 new_cpu = wake_affine(tmp, p, cpu, prev_cpu, sync);
6819 sd = NULL; /* Prefer wake_affine over balance flags */
6823 if (tmp->flags & sd_flag)
6825 else if (!want_affine)
6831 new_cpu = find_idlest_cpu(sd, p, cpu, prev_cpu, sd_flag);
6832 } else if (wake_flags & WF_TTWU) { /* XXX always ? */
6834 new_cpu = select_idle_sibling(p, prev_cpu, new_cpu);
6837 current->recent_used_cpu = cpu;
6844 static void detach_entity_cfs_rq(struct sched_entity *se);
6847 * Called immediately before a task is migrated to a new CPU; task_cpu(p) and
6848 * cfs_rq_of(p) references at time of call are still valid and identify the
6849 * previous CPU. The caller guarantees p->pi_lock or task_rq(p)->lock is held.
6851 static void migrate_task_rq_fair(struct task_struct *p, int new_cpu)
6854 * As blocked tasks retain absolute vruntime the migration needs to
6855 * deal with this by subtracting the old and adding the new
6856 * min_vruntime -- the latter is done by enqueue_entity() when placing
6857 * the task on the new runqueue.
6859 if (p->state == TASK_WAKING) {
6860 struct sched_entity *se = &p->se;
6861 struct cfs_rq *cfs_rq = cfs_rq_of(se);
6864 #ifndef CONFIG_64BIT
6865 u64 min_vruntime_copy;
6868 min_vruntime_copy = cfs_rq->min_vruntime_copy;
6870 min_vruntime = cfs_rq->min_vruntime;
6871 } while (min_vruntime != min_vruntime_copy);
6873 min_vruntime = cfs_rq->min_vruntime;
6876 se->vruntime -= min_vruntime;
6879 if (p->on_rq == TASK_ON_RQ_MIGRATING) {
6881 * In case of TASK_ON_RQ_MIGRATING we in fact hold the 'old'
6882 * rq->lock and can modify state directly.
6884 lockdep_assert_held(&task_rq(p)->lock);
6885 detach_entity_cfs_rq(&p->se);
6889 * We are supposed to update the task to "current" time, then
6890 * its up to date and ready to go to new CPU/cfs_rq. But we
6891 * have difficulty in getting what current time is, so simply
6892 * throw away the out-of-date time. This will result in the
6893 * wakee task is less decayed, but giving the wakee more load
6896 remove_entity_load_avg(&p->se);
6899 /* Tell new CPU we are migrated */
6900 p->se.avg.last_update_time = 0;
6902 /* We have migrated, no longer consider this task hot */
6903 p->se.exec_start = 0;
6905 update_scan_period(p, new_cpu);
6908 static void task_dead_fair(struct task_struct *p)
6910 remove_entity_load_avg(&p->se);
6914 balance_fair(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
6919 return newidle_balance(rq, rf) != 0;
6921 #endif /* CONFIG_SMP */
6923 static unsigned long wakeup_gran(struct sched_entity *se)
6925 unsigned long gran = sysctl_sched_wakeup_granularity;
6928 * Since its curr running now, convert the gran from real-time
6929 * to virtual-time in his units.
6931 * By using 'se' instead of 'curr' we penalize light tasks, so
6932 * they get preempted easier. That is, if 'se' < 'curr' then
6933 * the resulting gran will be larger, therefore penalizing the
6934 * lighter, if otoh 'se' > 'curr' then the resulting gran will
6935 * be smaller, again penalizing the lighter task.
6937 * This is especially important for buddies when the leftmost
6938 * task is higher priority than the buddy.
6940 return calc_delta_fair(gran, se);
6944 * Should 'se' preempt 'curr'.
6958 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
6960 s64 gran, vdiff = curr->vruntime - se->vruntime;
6965 gran = wakeup_gran(se);
6972 static void set_last_buddy(struct sched_entity *se)
6974 if (entity_is_task(se) && unlikely(task_has_idle_policy(task_of(se))))
6977 for_each_sched_entity(se) {
6978 if (SCHED_WARN_ON(!se->on_rq))
6980 cfs_rq_of(se)->last = se;
6984 static void set_next_buddy(struct sched_entity *se)
6986 if (entity_is_task(se) && unlikely(task_has_idle_policy(task_of(se))))
6989 for_each_sched_entity(se) {
6990 if (SCHED_WARN_ON(!se->on_rq))
6992 cfs_rq_of(se)->next = se;
6996 static void set_skip_buddy(struct sched_entity *se)
6998 for_each_sched_entity(se)
6999 cfs_rq_of(se)->skip = se;
7003 * Preempt the current task with a newly woken task if needed:
7005 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
7007 struct task_struct *curr = rq->curr;
7008 struct sched_entity *se = &curr->se, *pse = &p->se;
7009 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
7010 int scale = cfs_rq->nr_running >= sched_nr_latency;
7011 int next_buddy_marked = 0;
7013 if (unlikely(se == pse))
7017 * This is possible from callers such as attach_tasks(), in which we
7018 * unconditionally check_preempt_curr() after an enqueue (which may have
7019 * lead to a throttle). This both saves work and prevents false
7020 * next-buddy nomination below.
7022 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
7025 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
7026 set_next_buddy(pse);
7027 next_buddy_marked = 1;
7031 * We can come here with TIF_NEED_RESCHED already set from new task
7034 * Note: this also catches the edge-case of curr being in a throttled
7035 * group (e.g. via set_curr_task), since update_curr() (in the
7036 * enqueue of curr) will have resulted in resched being set. This
7037 * prevents us from potentially nominating it as a false LAST_BUDDY
7040 if (test_tsk_need_resched(curr))
7043 /* Idle tasks are by definition preempted by non-idle tasks. */
7044 if (unlikely(task_has_idle_policy(curr)) &&
7045 likely(!task_has_idle_policy(p)))
7049 * Batch and idle tasks do not preempt non-idle tasks (their preemption
7050 * is driven by the tick):
7052 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
7055 find_matching_se(&se, &pse);
7056 update_curr(cfs_rq_of(se));
7058 if (wakeup_preempt_entity(se, pse) == 1) {
7060 * Bias pick_next to pick the sched entity that is
7061 * triggering this preemption.
7063 if (!next_buddy_marked)
7064 set_next_buddy(pse);
7073 * Only set the backward buddy when the current task is still
7074 * on the rq. This can happen when a wakeup gets interleaved
7075 * with schedule on the ->pre_schedule() or idle_balance()
7076 * point, either of which can * drop the rq lock.
7078 * Also, during early boot the idle thread is in the fair class,
7079 * for obvious reasons its a bad idea to schedule back to it.
7081 if (unlikely(!se->on_rq || curr == rq->idle))
7084 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
7088 struct task_struct *
7089 pick_next_task_fair(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
7091 struct cfs_rq *cfs_rq = &rq->cfs;
7092 struct sched_entity *se;
7093 struct task_struct *p;
7097 if (!sched_fair_runnable(rq))
7100 #ifdef CONFIG_FAIR_GROUP_SCHED
7101 if (!prev || prev->sched_class != &fair_sched_class)
7105 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
7106 * likely that a next task is from the same cgroup as the current.
7108 * Therefore attempt to avoid putting and setting the entire cgroup
7109 * hierarchy, only change the part that actually changes.
7113 struct sched_entity *curr = cfs_rq->curr;
7116 * Since we got here without doing put_prev_entity() we also
7117 * have to consider cfs_rq->curr. If it is still a runnable
7118 * entity, update_curr() will update its vruntime, otherwise
7119 * forget we've ever seen it.
7123 update_curr(cfs_rq);
7128 * This call to check_cfs_rq_runtime() will do the
7129 * throttle and dequeue its entity in the parent(s).
7130 * Therefore the nr_running test will indeed
7133 if (unlikely(check_cfs_rq_runtime(cfs_rq))) {
7136 if (!cfs_rq->nr_running)
7143 se = pick_next_entity(cfs_rq, curr);
7144 cfs_rq = group_cfs_rq(se);
7150 * Since we haven't yet done put_prev_entity and if the selected task
7151 * is a different task than we started out with, try and touch the
7152 * least amount of cfs_rqs.
7155 struct sched_entity *pse = &prev->se;
7157 while (!(cfs_rq = is_same_group(se, pse))) {
7158 int se_depth = se->depth;
7159 int pse_depth = pse->depth;
7161 if (se_depth <= pse_depth) {
7162 put_prev_entity(cfs_rq_of(pse), pse);
7163 pse = parent_entity(pse);
7165 if (se_depth >= pse_depth) {
7166 set_next_entity(cfs_rq_of(se), se);
7167 se = parent_entity(se);
7171 put_prev_entity(cfs_rq, pse);
7172 set_next_entity(cfs_rq, se);
7179 put_prev_task(rq, prev);
7182 se = pick_next_entity(cfs_rq, NULL);
7183 set_next_entity(cfs_rq, se);
7184 cfs_rq = group_cfs_rq(se);
7189 done: __maybe_unused;
7192 * Move the next running task to the front of
7193 * the list, so our cfs_tasks list becomes MRU
7196 list_move(&p->se.group_node, &rq->cfs_tasks);
7199 if (hrtick_enabled_fair(rq))
7200 hrtick_start_fair(rq, p);
7202 update_misfit_status(p, rq);
7210 new_tasks = newidle_balance(rq, rf);
7213 * Because newidle_balance() releases (and re-acquires) rq->lock, it is
7214 * possible for any higher priority task to appear. In that case we
7215 * must re-start the pick_next_entity() loop.
7224 * rq is about to be idle, check if we need to update the
7225 * lost_idle_time of clock_pelt
7227 update_idle_rq_clock_pelt(rq);
7232 static struct task_struct *__pick_next_task_fair(struct rq *rq)
7234 return pick_next_task_fair(rq, NULL, NULL);
7238 * Account for a descheduled task:
7240 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
7242 struct sched_entity *se = &prev->se;
7243 struct cfs_rq *cfs_rq;
7245 for_each_sched_entity(se) {
7246 cfs_rq = cfs_rq_of(se);
7247 put_prev_entity(cfs_rq, se);
7252 * sched_yield() is very simple
7254 * The magic of dealing with the ->skip buddy is in pick_next_entity.
7256 static void yield_task_fair(struct rq *rq)
7258 struct task_struct *curr = rq->curr;
7259 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
7260 struct sched_entity *se = &curr->se;
7263 * Are we the only task in the tree?
7265 if (unlikely(rq->nr_running == 1))
7268 clear_buddies(cfs_rq, se);
7270 if (curr->policy != SCHED_BATCH) {
7271 update_rq_clock(rq);
7273 * Update run-time statistics of the 'current'.
7275 update_curr(cfs_rq);
7277 * Tell update_rq_clock() that we've just updated,
7278 * so we don't do microscopic update in schedule()
7279 * and double the fastpath cost.
7281 rq_clock_skip_update(rq);
7287 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p)
7289 struct sched_entity *se = &p->se;
7291 /* throttled hierarchies are not runnable */
7292 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
7295 /* Tell the scheduler that we'd really like pse to run next. */
7298 yield_task_fair(rq);
7304 /**************************************************
7305 * Fair scheduling class load-balancing methods.
7309 * The purpose of load-balancing is to achieve the same basic fairness the
7310 * per-CPU scheduler provides, namely provide a proportional amount of compute
7311 * time to each task. This is expressed in the following equation:
7313 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
7315 * Where W_i,n is the n-th weight average for CPU i. The instantaneous weight
7316 * W_i,0 is defined as:
7318 * W_i,0 = \Sum_j w_i,j (2)
7320 * Where w_i,j is the weight of the j-th runnable task on CPU i. This weight
7321 * is derived from the nice value as per sched_prio_to_weight[].
7323 * The weight average is an exponential decay average of the instantaneous
7326 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
7328 * C_i is the compute capacity of CPU i, typically it is the
7329 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
7330 * can also include other factors [XXX].
7332 * To achieve this balance we define a measure of imbalance which follows
7333 * directly from (1):
7335 * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4)
7337 * We them move tasks around to minimize the imbalance. In the continuous
7338 * function space it is obvious this converges, in the discrete case we get
7339 * a few fun cases generally called infeasible weight scenarios.
7342 * - infeasible weights;
7343 * - local vs global optima in the discrete case. ]
7348 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
7349 * for all i,j solution, we create a tree of CPUs that follows the hardware
7350 * topology where each level pairs two lower groups (or better). This results
7351 * in O(log n) layers. Furthermore we reduce the number of CPUs going up the
7352 * tree to only the first of the previous level and we decrease the frequency
7353 * of load-balance at each level inv. proportional to the number of CPUs in
7359 * \Sum { --- * --- * 2^i } = O(n) (5)
7361 * `- size of each group
7362 * | | `- number of CPUs doing load-balance
7364 * `- sum over all levels
7366 * Coupled with a limit on how many tasks we can migrate every balance pass,
7367 * this makes (5) the runtime complexity of the balancer.
7369 * An important property here is that each CPU is still (indirectly) connected
7370 * to every other CPU in at most O(log n) steps:
7372 * The adjacency matrix of the resulting graph is given by:
7375 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
7378 * And you'll find that:
7380 * A^(log_2 n)_i,j != 0 for all i,j (7)
7382 * Showing there's indeed a path between every CPU in at most O(log n) steps.
7383 * The task movement gives a factor of O(m), giving a convergence complexity
7386 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
7391 * In order to avoid CPUs going idle while there's still work to do, new idle
7392 * balancing is more aggressive and has the newly idle CPU iterate up the domain
7393 * tree itself instead of relying on other CPUs to bring it work.
7395 * This adds some complexity to both (5) and (8) but it reduces the total idle
7403 * Cgroups make a horror show out of (2), instead of a simple sum we get:
7406 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
7411 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
7413 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on CPU i.
7415 * The big problem is S_k, its a global sum needed to compute a local (W_i)
7418 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
7419 * rewrite all of this once again.]
7422 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
7424 enum fbq_type { regular, remote, all };
7427 * 'group_type' describes the group of CPUs at the moment of load balancing.
7429 * The enum is ordered by pulling priority, with the group with lowest priority
7430 * first so the group_type can simply be compared when selecting the busiest
7431 * group. See update_sd_pick_busiest().
7434 /* The group has spare capacity that can be used to run more tasks. */
7435 group_has_spare = 0,
7437 * The group is fully used and the tasks don't compete for more CPU
7438 * cycles. Nevertheless, some tasks might wait before running.
7442 * SD_ASYM_CPUCAPACITY only: One task doesn't fit with CPU's capacity
7443 * and must be migrated to a more powerful CPU.
7447 * SD_ASYM_PACKING only: One local CPU with higher capacity is available,
7448 * and the task should be migrated to it instead of running on the
7453 * The tasks' affinity constraints previously prevented the scheduler
7454 * from balancing the load across the system.
7458 * The CPU is overloaded and can't provide expected CPU cycles to all
7464 enum migration_type {
7471 #define LBF_ALL_PINNED 0x01
7472 #define LBF_NEED_BREAK 0x02
7473 #define LBF_DST_PINNED 0x04
7474 #define LBF_SOME_PINNED 0x08
7475 #define LBF_ACTIVE_LB 0x10
7478 struct sched_domain *sd;
7486 struct cpumask *dst_grpmask;
7488 enum cpu_idle_type idle;
7490 /* The set of CPUs under consideration for load-balancing */
7491 struct cpumask *cpus;
7496 unsigned int loop_break;
7497 unsigned int loop_max;
7499 enum fbq_type fbq_type;
7500 enum migration_type migration_type;
7501 struct list_head tasks;
7505 * Is this task likely cache-hot:
7507 static int task_hot(struct task_struct *p, struct lb_env *env)
7511 lockdep_assert_held(&env->src_rq->lock);
7513 if (p->sched_class != &fair_sched_class)
7516 if (unlikely(task_has_idle_policy(p)))
7519 /* SMT siblings share cache */
7520 if (env->sd->flags & SD_SHARE_CPUCAPACITY)
7524 * Buddy candidates are cache hot:
7526 if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running &&
7527 (&p->se == cfs_rq_of(&p->se)->next ||
7528 &p->se == cfs_rq_of(&p->se)->last))
7531 if (sysctl_sched_migration_cost == -1)
7533 if (sysctl_sched_migration_cost == 0)
7536 delta = rq_clock_task(env->src_rq) - p->se.exec_start;
7538 return delta < (s64)sysctl_sched_migration_cost;
7541 #ifdef CONFIG_NUMA_BALANCING
7543 * Returns 1, if task migration degrades locality
7544 * Returns 0, if task migration improves locality i.e migration preferred.
7545 * Returns -1, if task migration is not affected by locality.
7547 static int migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
7549 struct numa_group *numa_group = rcu_dereference(p->numa_group);
7550 unsigned long src_weight, dst_weight;
7551 int src_nid, dst_nid, dist;
7553 if (!static_branch_likely(&sched_numa_balancing))
7556 if (!p->numa_faults || !(env->sd->flags & SD_NUMA))
7559 src_nid = cpu_to_node(env->src_cpu);
7560 dst_nid = cpu_to_node(env->dst_cpu);
7562 if (src_nid == dst_nid)
7565 /* Migrating away from the preferred node is always bad. */
7566 if (src_nid == p->numa_preferred_nid) {
7567 if (env->src_rq->nr_running > env->src_rq->nr_preferred_running)
7573 /* Encourage migration to the preferred node. */
7574 if (dst_nid == p->numa_preferred_nid)
7577 /* Leaving a core idle is often worse than degrading locality. */
7578 if (env->idle == CPU_IDLE)
7581 dist = node_distance(src_nid, dst_nid);
7583 src_weight = group_weight(p, src_nid, dist);
7584 dst_weight = group_weight(p, dst_nid, dist);
7586 src_weight = task_weight(p, src_nid, dist);
7587 dst_weight = task_weight(p, dst_nid, dist);
7590 return dst_weight < src_weight;
7594 static inline int migrate_degrades_locality(struct task_struct *p,
7602 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
7605 int can_migrate_task(struct task_struct *p, struct lb_env *env)
7609 lockdep_assert_held(&env->src_rq->lock);
7612 * We do not migrate tasks that are:
7613 * 1) throttled_lb_pair, or
7614 * 2) cannot be migrated to this CPU due to cpus_ptr, or
7615 * 3) running (obviously), or
7616 * 4) are cache-hot on their current CPU.
7618 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
7621 /* Disregard pcpu kthreads; they are where they need to be. */
7622 if (kthread_is_per_cpu(p))
7625 if (!cpumask_test_cpu(env->dst_cpu, p->cpus_ptr)) {
7628 schedstat_inc(p->se.statistics.nr_failed_migrations_affine);
7630 env->flags |= LBF_SOME_PINNED;
7633 * Remember if this task can be migrated to any other CPU in
7634 * our sched_group. We may want to revisit it if we couldn't
7635 * meet load balance goals by pulling other tasks on src_cpu.
7637 * Avoid computing new_dst_cpu
7639 * - if we have already computed one in current iteration
7640 * - if it's an active balance
7642 if (env->idle == CPU_NEWLY_IDLE ||
7643 env->flags & (LBF_DST_PINNED | LBF_ACTIVE_LB))
7646 /* Prevent to re-select dst_cpu via env's CPUs: */
7647 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
7648 if (cpumask_test_cpu(cpu, p->cpus_ptr)) {
7649 env->flags |= LBF_DST_PINNED;
7650 env->new_dst_cpu = cpu;
7658 /* Record that we found at least one task that could run on dst_cpu */
7659 env->flags &= ~LBF_ALL_PINNED;
7661 if (task_running(env->src_rq, p)) {
7662 schedstat_inc(p->se.statistics.nr_failed_migrations_running);
7667 * Aggressive migration if:
7669 * 2) destination numa is preferred
7670 * 3) task is cache cold, or
7671 * 4) too many balance attempts have failed.
7673 if (env->flags & LBF_ACTIVE_LB)
7676 tsk_cache_hot = migrate_degrades_locality(p, env);
7677 if (tsk_cache_hot == -1)
7678 tsk_cache_hot = task_hot(p, env);
7680 if (tsk_cache_hot <= 0 ||
7681 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
7682 if (tsk_cache_hot == 1) {
7683 schedstat_inc(env->sd->lb_hot_gained[env->idle]);
7684 schedstat_inc(p->se.statistics.nr_forced_migrations);
7689 schedstat_inc(p->se.statistics.nr_failed_migrations_hot);
7694 * detach_task() -- detach the task for the migration specified in env
7696 static void detach_task(struct task_struct *p, struct lb_env *env)
7698 lockdep_assert_held(&env->src_rq->lock);
7700 deactivate_task(env->src_rq, p, DEQUEUE_NOCLOCK);
7701 set_task_cpu(p, env->dst_cpu);
7705 * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
7706 * part of active balancing operations within "domain".
7708 * Returns a task if successful and NULL otherwise.
7710 static struct task_struct *detach_one_task(struct lb_env *env)
7712 struct task_struct *p;
7714 lockdep_assert_held(&env->src_rq->lock);
7716 list_for_each_entry_reverse(p,
7717 &env->src_rq->cfs_tasks, se.group_node) {
7718 if (!can_migrate_task(p, env))
7721 detach_task(p, env);
7724 * Right now, this is only the second place where
7725 * lb_gained[env->idle] is updated (other is detach_tasks)
7726 * so we can safely collect stats here rather than
7727 * inside detach_tasks().
7729 schedstat_inc(env->sd->lb_gained[env->idle]);
7735 static const unsigned int sched_nr_migrate_break = 32;
7738 * detach_tasks() -- tries to detach up to imbalance load/util/tasks from
7739 * busiest_rq, as part of a balancing operation within domain "sd".
7741 * Returns number of detached tasks if successful and 0 otherwise.
7743 static int detach_tasks(struct lb_env *env)
7745 struct list_head *tasks = &env->src_rq->cfs_tasks;
7746 unsigned long util, load;
7747 struct task_struct *p;
7750 lockdep_assert_held(&env->src_rq->lock);
7753 * Source run queue has been emptied by another CPU, clear
7754 * LBF_ALL_PINNED flag as we will not test any task.
7756 if (env->src_rq->nr_running <= 1) {
7757 env->flags &= ~LBF_ALL_PINNED;
7761 if (env->imbalance <= 0)
7764 while (!list_empty(tasks)) {
7766 * We don't want to steal all, otherwise we may be treated likewise,
7767 * which could at worst lead to a livelock crash.
7769 if (env->idle != CPU_NOT_IDLE && env->src_rq->nr_running <= 1)
7772 p = list_last_entry(tasks, struct task_struct, se.group_node);
7775 /* We've more or less seen every task there is, call it quits */
7776 if (env->loop > env->loop_max)
7779 /* take a breather every nr_migrate tasks */
7780 if (env->loop > env->loop_break) {
7781 env->loop_break += sched_nr_migrate_break;
7782 env->flags |= LBF_NEED_BREAK;
7786 if (!can_migrate_task(p, env))
7789 switch (env->migration_type) {
7792 * Depending of the number of CPUs and tasks and the
7793 * cgroup hierarchy, task_h_load() can return a null
7794 * value. Make sure that env->imbalance decreases
7795 * otherwise detach_tasks() will stop only after
7796 * detaching up to loop_max tasks.
7798 load = max_t(unsigned long, task_h_load(p), 1);
7800 if (sched_feat(LB_MIN) &&
7801 load < 16 && !env->sd->nr_balance_failed)
7805 * Make sure that we don't migrate too much load.
7806 * Nevertheless, let relax the constraint if
7807 * scheduler fails to find a good waiting task to
7810 if (shr_bound(load, env->sd->nr_balance_failed) > env->imbalance)
7813 env->imbalance -= load;
7817 util = task_util_est(p);
7819 if (util > env->imbalance)
7822 env->imbalance -= util;
7829 case migrate_misfit:
7830 /* This is not a misfit task */
7831 if (task_fits_capacity(p, capacity_of(env->src_cpu)))
7838 detach_task(p, env);
7839 list_add(&p->se.group_node, &env->tasks);
7843 #ifdef CONFIG_PREEMPTION
7845 * NEWIDLE balancing is a source of latency, so preemptible
7846 * kernels will stop after the first task is detached to minimize
7847 * the critical section.
7849 if (env->idle == CPU_NEWLY_IDLE)
7854 * We only want to steal up to the prescribed amount of
7857 if (env->imbalance <= 0)
7862 list_move(&p->se.group_node, tasks);
7866 * Right now, this is one of only two places we collect this stat
7867 * so we can safely collect detach_one_task() stats here rather
7868 * than inside detach_one_task().
7870 schedstat_add(env->sd->lb_gained[env->idle], detached);
7876 * attach_task() -- attach the task detached by detach_task() to its new rq.
7878 static void attach_task(struct rq *rq, struct task_struct *p)
7880 lockdep_assert_held(&rq->lock);
7882 BUG_ON(task_rq(p) != rq);
7883 activate_task(rq, p, ENQUEUE_NOCLOCK);
7884 check_preempt_curr(rq, p, 0);
7888 * attach_one_task() -- attaches the task returned from detach_one_task() to
7891 static void attach_one_task(struct rq *rq, struct task_struct *p)
7896 update_rq_clock(rq);
7902 * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
7905 static void attach_tasks(struct lb_env *env)
7907 struct list_head *tasks = &env->tasks;
7908 struct task_struct *p;
7911 rq_lock(env->dst_rq, &rf);
7912 update_rq_clock(env->dst_rq);
7914 while (!list_empty(tasks)) {
7915 p = list_first_entry(tasks, struct task_struct, se.group_node);
7916 list_del_init(&p->se.group_node);
7918 attach_task(env->dst_rq, p);
7921 rq_unlock(env->dst_rq, &rf);
7924 #ifdef CONFIG_NO_HZ_COMMON
7925 static inline bool cfs_rq_has_blocked(struct cfs_rq *cfs_rq)
7927 if (cfs_rq->avg.load_avg)
7930 if (cfs_rq->avg.util_avg)
7936 static inline bool others_have_blocked(struct rq *rq)
7938 if (READ_ONCE(rq->avg_rt.util_avg))
7941 if (READ_ONCE(rq->avg_dl.util_avg))
7944 if (thermal_load_avg(rq))
7947 #ifdef CONFIG_HAVE_SCHED_AVG_IRQ
7948 if (READ_ONCE(rq->avg_irq.util_avg))
7955 static inline void update_blocked_load_tick(struct rq *rq)
7957 WRITE_ONCE(rq->last_blocked_load_update_tick, jiffies);
7960 static inline void update_blocked_load_status(struct rq *rq, bool has_blocked)
7963 rq->has_blocked_load = 0;
7966 static inline bool cfs_rq_has_blocked(struct cfs_rq *cfs_rq) { return false; }
7967 static inline bool others_have_blocked(struct rq *rq) { return false; }
7968 static inline void update_blocked_load_tick(struct rq *rq) {}
7969 static inline void update_blocked_load_status(struct rq *rq, bool has_blocked) {}
7972 static bool __update_blocked_others(struct rq *rq, bool *done)
7974 const struct sched_class *curr_class;
7975 u64 now = rq_clock_pelt(rq);
7976 unsigned long thermal_pressure;
7980 * update_load_avg() can call cpufreq_update_util(). Make sure that RT,
7981 * DL and IRQ signals have been updated before updating CFS.
7983 curr_class = rq->curr->sched_class;
7985 thermal_pressure = arch_scale_thermal_pressure(cpu_of(rq));
7987 decayed = update_rt_rq_load_avg(now, rq, curr_class == &rt_sched_class) |
7988 update_dl_rq_load_avg(now, rq, curr_class == &dl_sched_class) |
7989 update_thermal_load_avg(rq_clock_thermal(rq), rq, thermal_pressure) |
7990 update_irq_load_avg(rq, 0);
7992 if (others_have_blocked(rq))
7998 #ifdef CONFIG_FAIR_GROUP_SCHED
8000 static inline bool cfs_rq_is_decayed(struct cfs_rq *cfs_rq)
8002 if (cfs_rq->load.weight)
8005 if (cfs_rq->avg.load_sum)
8008 if (cfs_rq->avg.util_sum)
8011 if (cfs_rq->avg.runnable_sum)
8017 static bool __update_blocked_fair(struct rq *rq, bool *done)
8019 struct cfs_rq *cfs_rq, *pos;
8020 bool decayed = false;
8021 int cpu = cpu_of(rq);
8024 * Iterates the task_group tree in a bottom up fashion, see
8025 * list_add_leaf_cfs_rq() for details.
8027 for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) {
8028 struct sched_entity *se;
8030 if (update_cfs_rq_load_avg(cfs_rq_clock_pelt(cfs_rq), cfs_rq)) {
8031 update_tg_load_avg(cfs_rq);
8033 if (cfs_rq == &rq->cfs)
8037 /* Propagate pending load changes to the parent, if any: */
8038 se = cfs_rq->tg->se[cpu];
8039 if (se && !skip_blocked_update(se))
8040 update_load_avg(cfs_rq_of(se), se, 0);
8043 * There can be a lot of idle CPU cgroups. Don't let fully
8044 * decayed cfs_rqs linger on the list.
8046 if (cfs_rq_is_decayed(cfs_rq))
8047 list_del_leaf_cfs_rq(cfs_rq);
8049 /* Don't need periodic decay once load/util_avg are null */
8050 if (cfs_rq_has_blocked(cfs_rq))
8058 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
8059 * This needs to be done in a top-down fashion because the load of a child
8060 * group is a fraction of its parents load.
8062 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
8064 struct rq *rq = rq_of(cfs_rq);
8065 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
8066 unsigned long now = jiffies;
8069 if (cfs_rq->last_h_load_update == now)
8072 WRITE_ONCE(cfs_rq->h_load_next, NULL);
8073 for_each_sched_entity(se) {
8074 cfs_rq = cfs_rq_of(se);
8075 WRITE_ONCE(cfs_rq->h_load_next, se);
8076 if (cfs_rq->last_h_load_update == now)
8081 cfs_rq->h_load = cfs_rq_load_avg(cfs_rq);
8082 cfs_rq->last_h_load_update = now;
8085 while ((se = READ_ONCE(cfs_rq->h_load_next)) != NULL) {
8086 load = cfs_rq->h_load;
8087 load = div64_ul(load * se->avg.load_avg,
8088 cfs_rq_load_avg(cfs_rq) + 1);
8089 cfs_rq = group_cfs_rq(se);
8090 cfs_rq->h_load = load;
8091 cfs_rq->last_h_load_update = now;
8095 static unsigned long task_h_load(struct task_struct *p)
8097 struct cfs_rq *cfs_rq = task_cfs_rq(p);
8099 update_cfs_rq_h_load(cfs_rq);
8100 return div64_ul(p->se.avg.load_avg * cfs_rq->h_load,
8101 cfs_rq_load_avg(cfs_rq) + 1);
8104 static bool __update_blocked_fair(struct rq *rq, bool *done)
8106 struct cfs_rq *cfs_rq = &rq->cfs;
8109 decayed = update_cfs_rq_load_avg(cfs_rq_clock_pelt(cfs_rq), cfs_rq);
8110 if (cfs_rq_has_blocked(cfs_rq))
8116 static unsigned long task_h_load(struct task_struct *p)
8118 return p->se.avg.load_avg;
8122 static void update_blocked_averages(int cpu)
8124 bool decayed = false, done = true;
8125 struct rq *rq = cpu_rq(cpu);
8128 rq_lock_irqsave(rq, &rf);
8129 update_blocked_load_tick(rq);
8130 update_rq_clock(rq);
8132 decayed |= __update_blocked_others(rq, &done);
8133 decayed |= __update_blocked_fair(rq, &done);
8135 update_blocked_load_status(rq, !done);
8137 cpufreq_update_util(rq, 0);
8138 rq_unlock_irqrestore(rq, &rf);
8141 /********** Helpers for find_busiest_group ************************/
8144 * sg_lb_stats - stats of a sched_group required for load_balancing
8146 struct sg_lb_stats {
8147 unsigned long avg_load; /*Avg load across the CPUs of the group */
8148 unsigned long group_load; /* Total load over the CPUs of the group */
8149 unsigned long group_capacity;
8150 unsigned long group_util; /* Total utilization over the CPUs of the group */
8151 unsigned long group_runnable; /* Total runnable time over the CPUs of the group */
8152 unsigned int sum_nr_running; /* Nr of tasks running in the group */
8153 unsigned int sum_h_nr_running; /* Nr of CFS tasks running in the group */
8154 unsigned int idle_cpus;
8155 unsigned int group_weight;
8156 enum group_type group_type;
8157 unsigned int group_asym_packing; /* Tasks should be moved to preferred CPU */
8158 unsigned long group_misfit_task_load; /* A CPU has a task too big for its capacity */
8159 #ifdef CONFIG_NUMA_BALANCING
8160 unsigned int nr_numa_running;
8161 unsigned int nr_preferred_running;
8166 * sd_lb_stats - Structure to store the statistics of a sched_domain
8167 * during load balancing.
8169 struct sd_lb_stats {
8170 struct sched_group *busiest; /* Busiest group in this sd */
8171 struct sched_group *local; /* Local group in this sd */
8172 unsigned long total_load; /* Total load of all groups in sd */
8173 unsigned long total_capacity; /* Total capacity of all groups in sd */
8174 unsigned long avg_load; /* Average load across all groups in sd */
8175 unsigned int prefer_sibling; /* tasks should go to sibling first */
8177 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
8178 struct sg_lb_stats local_stat; /* Statistics of the local group */
8181 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
8184 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
8185 * local_stat because update_sg_lb_stats() does a full clear/assignment.
8186 * We must however set busiest_stat::group_type and
8187 * busiest_stat::idle_cpus to the worst busiest group because
8188 * update_sd_pick_busiest() reads these before assignment.
8190 *sds = (struct sd_lb_stats){
8194 .total_capacity = 0UL,
8196 .idle_cpus = UINT_MAX,
8197 .group_type = group_has_spare,
8202 static unsigned long scale_rt_capacity(int cpu)
8204 struct rq *rq = cpu_rq(cpu);
8205 unsigned long max = arch_scale_cpu_capacity(cpu);
8206 unsigned long used, free;
8209 irq = cpu_util_irq(rq);
8211 if (unlikely(irq >= max))
8215 * avg_rt.util_avg and avg_dl.util_avg track binary signals
8216 * (running and not running) with weights 0 and 1024 respectively.
8217 * avg_thermal.load_avg tracks thermal pressure and the weighted
8218 * average uses the actual delta max capacity(load).
8220 used = READ_ONCE(rq->avg_rt.util_avg);
8221 used += READ_ONCE(rq->avg_dl.util_avg);
8222 used += thermal_load_avg(rq);
8224 if (unlikely(used >= max))
8229 return scale_irq_capacity(free, irq, max);
8232 static void update_cpu_capacity(struct sched_domain *sd, int cpu)
8234 unsigned long capacity = scale_rt_capacity(cpu);
8235 struct sched_group *sdg = sd->groups;
8237 cpu_rq(cpu)->cpu_capacity_orig = arch_scale_cpu_capacity(cpu);
8242 cpu_rq(cpu)->cpu_capacity = capacity;
8243 trace_sched_cpu_capacity_tp(cpu_rq(cpu));
8245 sdg->sgc->capacity = capacity;
8246 sdg->sgc->min_capacity = capacity;
8247 sdg->sgc->max_capacity = capacity;
8250 void update_group_capacity(struct sched_domain *sd, int cpu)
8252 struct sched_domain *child = sd->child;
8253 struct sched_group *group, *sdg = sd->groups;
8254 unsigned long capacity, min_capacity, max_capacity;
8255 unsigned long interval;
8257 interval = msecs_to_jiffies(sd->balance_interval);
8258 interval = clamp(interval, 1UL, max_load_balance_interval);
8259 sdg->sgc->next_update = jiffies + interval;
8262 update_cpu_capacity(sd, cpu);
8267 min_capacity = ULONG_MAX;
8270 if (child->flags & SD_OVERLAP) {
8272 * SD_OVERLAP domains cannot assume that child groups
8273 * span the current group.
8276 for_each_cpu(cpu, sched_group_span(sdg)) {
8277 unsigned long cpu_cap = capacity_of(cpu);
8279 capacity += cpu_cap;
8280 min_capacity = min(cpu_cap, min_capacity);
8281 max_capacity = max(cpu_cap, max_capacity);
8285 * !SD_OVERLAP domains can assume that child groups
8286 * span the current group.
8289 group = child->groups;
8291 struct sched_group_capacity *sgc = group->sgc;
8293 capacity += sgc->capacity;
8294 min_capacity = min(sgc->min_capacity, min_capacity);
8295 max_capacity = max(sgc->max_capacity, max_capacity);
8296 group = group->next;
8297 } while (group != child->groups);
8300 sdg->sgc->capacity = capacity;
8301 sdg->sgc->min_capacity = min_capacity;
8302 sdg->sgc->max_capacity = max_capacity;
8306 * Check whether the capacity of the rq has been noticeably reduced by side
8307 * activity. The imbalance_pct is used for the threshold.
8308 * Return true is the capacity is reduced
8311 check_cpu_capacity(struct rq *rq, struct sched_domain *sd)
8313 return ((rq->cpu_capacity * sd->imbalance_pct) <
8314 (rq->cpu_capacity_orig * 100));
8318 * Check whether a rq has a misfit task and if it looks like we can actually
8319 * help that task: we can migrate the task to a CPU of higher capacity, or
8320 * the task's current CPU is heavily pressured.
8322 static inline int check_misfit_status(struct rq *rq, struct sched_domain *sd)
8324 return rq->misfit_task_load &&
8325 (rq->cpu_capacity_orig < rq->rd->max_cpu_capacity ||
8326 check_cpu_capacity(rq, sd));
8330 * Group imbalance indicates (and tries to solve) the problem where balancing
8331 * groups is inadequate due to ->cpus_ptr constraints.
8333 * Imagine a situation of two groups of 4 CPUs each and 4 tasks each with a
8334 * cpumask covering 1 CPU of the first group and 3 CPUs of the second group.
8337 * { 0 1 2 3 } { 4 5 6 7 }
8340 * If we were to balance group-wise we'd place two tasks in the first group and
8341 * two tasks in the second group. Clearly this is undesired as it will overload
8342 * cpu 3 and leave one of the CPUs in the second group unused.
8344 * The current solution to this issue is detecting the skew in the first group
8345 * by noticing the lower domain failed to reach balance and had difficulty
8346 * moving tasks due to affinity constraints.
8348 * When this is so detected; this group becomes a candidate for busiest; see
8349 * update_sd_pick_busiest(). And calculate_imbalance() and
8350 * find_busiest_group() avoid some of the usual balance conditions to allow it
8351 * to create an effective group imbalance.
8353 * This is a somewhat tricky proposition since the next run might not find the
8354 * group imbalance and decide the groups need to be balanced again. A most
8355 * subtle and fragile situation.
8358 static inline int sg_imbalanced(struct sched_group *group)
8360 return group->sgc->imbalance;
8364 * group_has_capacity returns true if the group has spare capacity that could
8365 * be used by some tasks.
8366 * We consider that a group has spare capacity if the * number of task is
8367 * smaller than the number of CPUs or if the utilization is lower than the
8368 * available capacity for CFS tasks.
8369 * For the latter, we use a threshold to stabilize the state, to take into
8370 * account the variance of the tasks' load and to return true if the available
8371 * capacity in meaningful for the load balancer.
8372 * As an example, an available capacity of 1% can appear but it doesn't make
8373 * any benefit for the load balance.
8376 group_has_capacity(unsigned int imbalance_pct, struct sg_lb_stats *sgs)
8378 if (sgs->sum_nr_running < sgs->group_weight)
8381 if ((sgs->group_capacity * imbalance_pct) <
8382 (sgs->group_runnable * 100))
8385 if ((sgs->group_capacity * 100) >
8386 (sgs->group_util * imbalance_pct))
8393 * group_is_overloaded returns true if the group has more tasks than it can
8395 * group_is_overloaded is not equals to !group_has_capacity because a group
8396 * with the exact right number of tasks, has no more spare capacity but is not
8397 * overloaded so both group_has_capacity and group_is_overloaded return
8401 group_is_overloaded(unsigned int imbalance_pct, struct sg_lb_stats *sgs)
8403 if (sgs->sum_nr_running <= sgs->group_weight)
8406 if ((sgs->group_capacity * 100) <
8407 (sgs->group_util * imbalance_pct))
8410 if ((sgs->group_capacity * imbalance_pct) <
8411 (sgs->group_runnable * 100))
8418 group_type group_classify(unsigned int imbalance_pct,
8419 struct sched_group *group,
8420 struct sg_lb_stats *sgs)
8422 if (group_is_overloaded(imbalance_pct, sgs))
8423 return group_overloaded;
8425 if (sg_imbalanced(group))
8426 return group_imbalanced;
8428 if (sgs->group_asym_packing)
8429 return group_asym_packing;
8431 if (sgs->group_misfit_task_load)
8432 return group_misfit_task;
8434 if (!group_has_capacity(imbalance_pct, sgs))
8435 return group_fully_busy;
8437 return group_has_spare;
8441 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
8442 * @env: The load balancing environment.
8443 * @group: sched_group whose statistics are to be updated.
8444 * @sgs: variable to hold the statistics for this group.
8445 * @sg_status: Holds flag indicating the status of the sched_group
8447 static inline void update_sg_lb_stats(struct lb_env *env,
8448 struct sched_group *group,
8449 struct sg_lb_stats *sgs,
8452 int i, nr_running, local_group;
8454 memset(sgs, 0, sizeof(*sgs));
8456 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_span(group));
8458 for_each_cpu_and(i, sched_group_span(group), env->cpus) {
8459 struct rq *rq = cpu_rq(i);
8461 sgs->group_load += cpu_load(rq);
8462 sgs->group_util += cpu_util(i);
8463 sgs->group_runnable += cpu_runnable(rq);
8464 sgs->sum_h_nr_running += rq->cfs.h_nr_running;
8466 nr_running = rq->nr_running;
8467 sgs->sum_nr_running += nr_running;
8470 *sg_status |= SG_OVERLOAD;
8472 if (cpu_overutilized(i))
8473 *sg_status |= SG_OVERUTILIZED;
8475 #ifdef CONFIG_NUMA_BALANCING
8476 sgs->nr_numa_running += rq->nr_numa_running;
8477 sgs->nr_preferred_running += rq->nr_preferred_running;
8480 * No need to call idle_cpu() if nr_running is not 0
8482 if (!nr_running && idle_cpu(i)) {
8484 /* Idle cpu can't have misfit task */
8491 /* Check for a misfit task on the cpu */
8492 if (env->sd->flags & SD_ASYM_CPUCAPACITY &&
8493 sgs->group_misfit_task_load < rq->misfit_task_load) {
8494 sgs->group_misfit_task_load = rq->misfit_task_load;
8495 *sg_status |= SG_OVERLOAD;
8499 /* Check if dst CPU is idle and preferred to this group */
8500 if (env->sd->flags & SD_ASYM_PACKING &&
8501 env->idle != CPU_NOT_IDLE &&
8502 sgs->sum_h_nr_running &&
8503 sched_asym_prefer(env->dst_cpu, group->asym_prefer_cpu)) {
8504 sgs->group_asym_packing = 1;
8507 sgs->group_capacity = group->sgc->capacity;
8509 sgs->group_weight = group->group_weight;
8511 sgs->group_type = group_classify(env->sd->imbalance_pct, group, sgs);
8513 /* Computing avg_load makes sense only when group is overloaded */
8514 if (sgs->group_type == group_overloaded)
8515 sgs->avg_load = (sgs->group_load * SCHED_CAPACITY_SCALE) /
8516 sgs->group_capacity;
8520 * update_sd_pick_busiest - return 1 on busiest group
8521 * @env: The load balancing environment.
8522 * @sds: sched_domain statistics
8523 * @sg: sched_group candidate to be checked for being the busiest
8524 * @sgs: sched_group statistics
8526 * Determine if @sg is a busier group than the previously selected
8529 * Return: %true if @sg is a busier group than the previously selected
8530 * busiest group. %false otherwise.
8532 static bool update_sd_pick_busiest(struct lb_env *env,
8533 struct sd_lb_stats *sds,
8534 struct sched_group *sg,
8535 struct sg_lb_stats *sgs)
8537 struct sg_lb_stats *busiest = &sds->busiest_stat;
8539 /* Make sure that there is at least one task to pull */
8540 if (!sgs->sum_h_nr_running)
8544 * Don't try to pull misfit tasks we can't help.
8545 * We can use max_capacity here as reduction in capacity on some
8546 * CPUs in the group should either be possible to resolve
8547 * internally or be covered by avg_load imbalance (eventually).
8549 if (sgs->group_type == group_misfit_task &&
8550 (!capacity_greater(capacity_of(env->dst_cpu), sg->sgc->max_capacity) ||
8551 sds->local_stat.group_type != group_has_spare))
8554 if (sgs->group_type > busiest->group_type)
8557 if (sgs->group_type < busiest->group_type)
8561 * The candidate and the current busiest group are the same type of
8562 * group. Let check which one is the busiest according to the type.
8565 switch (sgs->group_type) {
8566 case group_overloaded:
8567 /* Select the overloaded group with highest avg_load. */
8568 if (sgs->avg_load <= busiest->avg_load)
8572 case group_imbalanced:
8574 * Select the 1st imbalanced group as we don't have any way to
8575 * choose one more than another.
8579 case group_asym_packing:
8580 /* Prefer to move from lowest priority CPU's work */
8581 if (sched_asym_prefer(sg->asym_prefer_cpu, sds->busiest->asym_prefer_cpu))
8585 case group_misfit_task:
8587 * If we have more than one misfit sg go with the biggest
8590 if (sgs->group_misfit_task_load < busiest->group_misfit_task_load)
8594 case group_fully_busy:
8596 * Select the fully busy group with highest avg_load. In
8597 * theory, there is no need to pull task from such kind of
8598 * group because tasks have all compute capacity that they need
8599 * but we can still improve the overall throughput by reducing
8600 * contention when accessing shared HW resources.
8602 * XXX for now avg_load is not computed and always 0 so we
8603 * select the 1st one.
8605 if (sgs->avg_load <= busiest->avg_load)
8609 case group_has_spare:
8611 * Select not overloaded group with lowest number of idle cpus
8612 * and highest number of running tasks. We could also compare
8613 * the spare capacity which is more stable but it can end up
8614 * that the group has less spare capacity but finally more idle
8615 * CPUs which means less opportunity to pull tasks.
8617 if (sgs->idle_cpus > busiest->idle_cpus)
8619 else if ((sgs->idle_cpus == busiest->idle_cpus) &&
8620 (sgs->sum_nr_running <= busiest->sum_nr_running))
8627 * Candidate sg has no more than one task per CPU and has higher
8628 * per-CPU capacity. Migrating tasks to less capable CPUs may harm
8629 * throughput. Maximize throughput, power/energy consequences are not
8632 if ((env->sd->flags & SD_ASYM_CPUCAPACITY) &&
8633 (sgs->group_type <= group_fully_busy) &&
8634 (capacity_greater(sg->sgc->min_capacity, capacity_of(env->dst_cpu))))
8640 #ifdef CONFIG_NUMA_BALANCING
8641 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
8643 if (sgs->sum_h_nr_running > sgs->nr_numa_running)
8645 if (sgs->sum_h_nr_running > sgs->nr_preferred_running)
8650 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
8652 if (rq->nr_running > rq->nr_numa_running)
8654 if (rq->nr_running > rq->nr_preferred_running)
8659 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
8664 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
8668 #endif /* CONFIG_NUMA_BALANCING */
8674 * task_running_on_cpu - return 1 if @p is running on @cpu.
8677 static unsigned int task_running_on_cpu(int cpu, struct task_struct *p)
8679 /* Task has no contribution or is new */
8680 if (cpu != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time))
8683 if (task_on_rq_queued(p))
8690 * idle_cpu_without - would a given CPU be idle without p ?
8691 * @cpu: the processor on which idleness is tested.
8692 * @p: task which should be ignored.
8694 * Return: 1 if the CPU would be idle. 0 otherwise.
8696 static int idle_cpu_without(int cpu, struct task_struct *p)
8698 struct rq *rq = cpu_rq(cpu);
8700 if (rq->curr != rq->idle && rq->curr != p)
8704 * rq->nr_running can't be used but an updated version without the
8705 * impact of p on cpu must be used instead. The updated nr_running
8706 * be computed and tested before calling idle_cpu_without().
8710 if (rq->ttwu_pending)
8718 * update_sg_wakeup_stats - Update sched_group's statistics for wakeup.
8719 * @sd: The sched_domain level to look for idlest group.
8720 * @group: sched_group whose statistics are to be updated.
8721 * @sgs: variable to hold the statistics for this group.
8722 * @p: The task for which we look for the idlest group/CPU.
8724 static inline void update_sg_wakeup_stats(struct sched_domain *sd,
8725 struct sched_group *group,
8726 struct sg_lb_stats *sgs,
8727 struct task_struct *p)
8731 memset(sgs, 0, sizeof(*sgs));
8733 for_each_cpu(i, sched_group_span(group)) {
8734 struct rq *rq = cpu_rq(i);
8737 sgs->group_load += cpu_load_without(rq, p);
8738 sgs->group_util += cpu_util_without(i, p);
8739 sgs->group_runnable += cpu_runnable_without(rq, p);
8740 local = task_running_on_cpu(i, p);
8741 sgs->sum_h_nr_running += rq->cfs.h_nr_running - local;
8743 nr_running = rq->nr_running - local;
8744 sgs->sum_nr_running += nr_running;
8747 * No need to call idle_cpu_without() if nr_running is not 0
8749 if (!nr_running && idle_cpu_without(i, p))
8754 /* Check if task fits in the group */
8755 if (sd->flags & SD_ASYM_CPUCAPACITY &&
8756 !task_fits_capacity(p, group->sgc->max_capacity)) {
8757 sgs->group_misfit_task_load = 1;
8760 sgs->group_capacity = group->sgc->capacity;
8762 sgs->group_weight = group->group_weight;
8764 sgs->group_type = group_classify(sd->imbalance_pct, group, sgs);
8767 * Computing avg_load makes sense only when group is fully busy or
8770 if (sgs->group_type == group_fully_busy ||
8771 sgs->group_type == group_overloaded)
8772 sgs->avg_load = (sgs->group_load * SCHED_CAPACITY_SCALE) /
8773 sgs->group_capacity;
8776 static bool update_pick_idlest(struct sched_group *idlest,
8777 struct sg_lb_stats *idlest_sgs,
8778 struct sched_group *group,
8779 struct sg_lb_stats *sgs)
8781 if (sgs->group_type < idlest_sgs->group_type)
8784 if (sgs->group_type > idlest_sgs->group_type)
8788 * The candidate and the current idlest group are the same type of
8789 * group. Let check which one is the idlest according to the type.
8792 switch (sgs->group_type) {
8793 case group_overloaded:
8794 case group_fully_busy:
8795 /* Select the group with lowest avg_load. */
8796 if (idlest_sgs->avg_load <= sgs->avg_load)
8800 case group_imbalanced:
8801 case group_asym_packing:
8802 /* Those types are not used in the slow wakeup path */
8805 case group_misfit_task:
8806 /* Select group with the highest max capacity */
8807 if (idlest->sgc->max_capacity >= group->sgc->max_capacity)
8811 case group_has_spare:
8812 /* Select group with most idle CPUs */
8813 if (idlest_sgs->idle_cpus > sgs->idle_cpus)
8816 /* Select group with lowest group_util */
8817 if (idlest_sgs->idle_cpus == sgs->idle_cpus &&
8818 idlest_sgs->group_util <= sgs->group_util)
8828 * Allow a NUMA imbalance if busy CPUs is less than 25% of the domain.
8829 * This is an approximation as the number of running tasks may not be
8830 * related to the number of busy CPUs due to sched_setaffinity.
8832 static inline bool allow_numa_imbalance(int dst_running, int dst_weight)
8834 return (dst_running < (dst_weight >> 2));
8838 * find_idlest_group() finds and returns the least busy CPU group within the
8841 * Assumes p is allowed on at least one CPU in sd.
8843 static struct sched_group *
8844 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
8846 struct sched_group *idlest = NULL, *local = NULL, *group = sd->groups;
8847 struct sg_lb_stats local_sgs, tmp_sgs;
8848 struct sg_lb_stats *sgs;
8849 unsigned long imbalance;
8850 struct sg_lb_stats idlest_sgs = {
8851 .avg_load = UINT_MAX,
8852 .group_type = group_overloaded,
8858 /* Skip over this group if it has no CPUs allowed */
8859 if (!cpumask_intersects(sched_group_span(group),
8863 local_group = cpumask_test_cpu(this_cpu,
8864 sched_group_span(group));
8873 update_sg_wakeup_stats(sd, group, sgs, p);
8875 if (!local_group && update_pick_idlest(idlest, &idlest_sgs, group, sgs)) {
8880 } while (group = group->next, group != sd->groups);
8883 /* There is no idlest group to push tasks to */
8887 /* The local group has been skipped because of CPU affinity */
8892 * If the local group is idler than the selected idlest group
8893 * don't try and push the task.
8895 if (local_sgs.group_type < idlest_sgs.group_type)
8899 * If the local group is busier than the selected idlest group
8900 * try and push the task.
8902 if (local_sgs.group_type > idlest_sgs.group_type)
8905 switch (local_sgs.group_type) {
8906 case group_overloaded:
8907 case group_fully_busy:
8909 /* Calculate allowed imbalance based on load */
8910 imbalance = scale_load_down(NICE_0_LOAD) *
8911 (sd->imbalance_pct-100) / 100;
8914 * When comparing groups across NUMA domains, it's possible for
8915 * the local domain to be very lightly loaded relative to the
8916 * remote domains but "imbalance" skews the comparison making
8917 * remote CPUs look much more favourable. When considering
8918 * cross-domain, add imbalance to the load on the remote node
8919 * and consider staying local.
8922 if ((sd->flags & SD_NUMA) &&
8923 ((idlest_sgs.avg_load + imbalance) >= local_sgs.avg_load))
8927 * If the local group is less loaded than the selected
8928 * idlest group don't try and push any tasks.
8930 if (idlest_sgs.avg_load >= (local_sgs.avg_load + imbalance))
8933 if (100 * local_sgs.avg_load <= sd->imbalance_pct * idlest_sgs.avg_load)
8937 case group_imbalanced:
8938 case group_asym_packing:
8939 /* Those type are not used in the slow wakeup path */
8942 case group_misfit_task:
8943 /* Select group with the highest max capacity */
8944 if (local->sgc->max_capacity >= idlest->sgc->max_capacity)
8948 case group_has_spare:
8949 if (sd->flags & SD_NUMA) {
8950 #ifdef CONFIG_NUMA_BALANCING
8953 * If there is spare capacity at NUMA, try to select
8954 * the preferred node
8956 if (cpu_to_node(this_cpu) == p->numa_preferred_nid)
8959 idlest_cpu = cpumask_first(sched_group_span(idlest));
8960 if (cpu_to_node(idlest_cpu) == p->numa_preferred_nid)
8964 * Otherwise, keep the task on this node to stay close
8965 * its wakeup source and improve locality. If there is
8966 * a real need of migration, periodic load balance will
8969 if (allow_numa_imbalance(local_sgs.sum_nr_running, sd->span_weight))
8974 * Select group with highest number of idle CPUs. We could also
8975 * compare the utilization which is more stable but it can end
8976 * up that the group has less spare capacity but finally more
8977 * idle CPUs which means more opportunity to run task.
8979 if (local_sgs.idle_cpus >= idlest_sgs.idle_cpus)
8988 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
8989 * @env: The load balancing environment.
8990 * @sds: variable to hold the statistics for this sched_domain.
8993 static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
8995 struct sched_domain *child = env->sd->child;
8996 struct sched_group *sg = env->sd->groups;
8997 struct sg_lb_stats *local = &sds->local_stat;
8998 struct sg_lb_stats tmp_sgs;
9002 struct sg_lb_stats *sgs = &tmp_sgs;
9005 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_span(sg));
9010 if (env->idle != CPU_NEWLY_IDLE ||
9011 time_after_eq(jiffies, sg->sgc->next_update))
9012 update_group_capacity(env->sd, env->dst_cpu);
9015 update_sg_lb_stats(env, sg, sgs, &sg_status);
9021 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
9023 sds->busiest_stat = *sgs;
9027 /* Now, start updating sd_lb_stats */
9028 sds->total_load += sgs->group_load;
9029 sds->total_capacity += sgs->group_capacity;
9032 } while (sg != env->sd->groups);
9034 /* Tag domain that child domain prefers tasks go to siblings first */
9035 sds->prefer_sibling = child && child->flags & SD_PREFER_SIBLING;
9038 if (env->sd->flags & SD_NUMA)
9039 env->fbq_type = fbq_classify_group(&sds->busiest_stat);
9041 if (!env->sd->parent) {
9042 struct root_domain *rd = env->dst_rq->rd;
9044 /* update overload indicator if we are at root domain */
9045 WRITE_ONCE(rd->overload, sg_status & SG_OVERLOAD);
9047 /* Update over-utilization (tipping point, U >= 0) indicator */
9048 WRITE_ONCE(rd->overutilized, sg_status & SG_OVERUTILIZED);
9049 trace_sched_overutilized_tp(rd, sg_status & SG_OVERUTILIZED);
9050 } else if (sg_status & SG_OVERUTILIZED) {
9051 struct root_domain *rd = env->dst_rq->rd;
9053 WRITE_ONCE(rd->overutilized, SG_OVERUTILIZED);
9054 trace_sched_overutilized_tp(rd, SG_OVERUTILIZED);
9058 #define NUMA_IMBALANCE_MIN 2
9060 static inline long adjust_numa_imbalance(int imbalance,
9061 int dst_running, int dst_weight)
9063 if (!allow_numa_imbalance(dst_running, dst_weight))
9067 * Allow a small imbalance based on a simple pair of communicating
9068 * tasks that remain local when the destination is lightly loaded.
9070 if (imbalance <= NUMA_IMBALANCE_MIN)
9077 * calculate_imbalance - Calculate the amount of imbalance present within the
9078 * groups of a given sched_domain during load balance.
9079 * @env: load balance environment
9080 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
9082 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
9084 struct sg_lb_stats *local, *busiest;
9086 local = &sds->local_stat;
9087 busiest = &sds->busiest_stat;
9089 if (busiest->group_type == group_misfit_task) {
9090 /* Set imbalance to allow misfit tasks to be balanced. */
9091 env->migration_type = migrate_misfit;
9096 if (busiest->group_type == group_asym_packing) {
9098 * In case of asym capacity, we will try to migrate all load to
9099 * the preferred CPU.
9101 env->migration_type = migrate_task;
9102 env->imbalance = busiest->sum_h_nr_running;
9106 if (busiest->group_type == group_imbalanced) {
9108 * In the group_imb case we cannot rely on group-wide averages
9109 * to ensure CPU-load equilibrium, try to move any task to fix
9110 * the imbalance. The next load balance will take care of
9111 * balancing back the system.
9113 env->migration_type = migrate_task;
9119 * Try to use spare capacity of local group without overloading it or
9122 if (local->group_type == group_has_spare) {
9123 if ((busiest->group_type > group_fully_busy) &&
9124 !(env->sd->flags & SD_SHARE_PKG_RESOURCES)) {
9126 * If busiest is overloaded, try to fill spare
9127 * capacity. This might end up creating spare capacity
9128 * in busiest or busiest still being overloaded but
9129 * there is no simple way to directly compute the
9130 * amount of load to migrate in order to balance the
9133 env->migration_type = migrate_util;
9134 env->imbalance = max(local->group_capacity, local->group_util) -
9138 * In some cases, the group's utilization is max or even
9139 * higher than capacity because of migrations but the
9140 * local CPU is (newly) idle. There is at least one
9141 * waiting task in this overloaded busiest group. Let's
9144 if (env->idle != CPU_NOT_IDLE && env->imbalance == 0) {
9145 env->migration_type = migrate_task;
9152 if (busiest->group_weight == 1 || sds->prefer_sibling) {
9153 unsigned int nr_diff = busiest->sum_nr_running;
9155 * When prefer sibling, evenly spread running tasks on
9158 env->migration_type = migrate_task;
9159 lsub_positive(&nr_diff, local->sum_nr_running);
9160 env->imbalance = nr_diff >> 1;
9164 * If there is no overload, we just want to even the number of
9167 env->migration_type = migrate_task;
9168 env->imbalance = max_t(long, 0, (local->idle_cpus -
9169 busiest->idle_cpus) >> 1);
9172 /* Consider allowing a small imbalance between NUMA groups */
9173 if (env->sd->flags & SD_NUMA) {
9174 env->imbalance = adjust_numa_imbalance(env->imbalance,
9175 busiest->sum_nr_running, busiest->group_weight);
9182 * Local is fully busy but has to take more load to relieve the
9185 if (local->group_type < group_overloaded) {
9187 * Local will become overloaded so the avg_load metrics are
9191 local->avg_load = (local->group_load * SCHED_CAPACITY_SCALE) /
9192 local->group_capacity;
9194 sds->avg_load = (sds->total_load * SCHED_CAPACITY_SCALE) /
9195 sds->total_capacity;
9197 * If the local group is more loaded than the selected
9198 * busiest group don't try to pull any tasks.
9200 if (local->avg_load >= busiest->avg_load) {
9207 * Both group are or will become overloaded and we're trying to get all
9208 * the CPUs to the average_load, so we don't want to push ourselves
9209 * above the average load, nor do we wish to reduce the max loaded CPU
9210 * below the average load. At the same time, we also don't want to
9211 * reduce the group load below the group capacity. Thus we look for
9212 * the minimum possible imbalance.
9214 env->migration_type = migrate_load;
9215 env->imbalance = min(
9216 (busiest->avg_load - sds->avg_load) * busiest->group_capacity,
9217 (sds->avg_load - local->avg_load) * local->group_capacity
9218 ) / SCHED_CAPACITY_SCALE;
9221 /******* find_busiest_group() helpers end here *********************/
9224 * Decision matrix according to the local and busiest group type:
9226 * busiest \ local has_spare fully_busy misfit asym imbalanced overloaded
9227 * has_spare nr_idle balanced N/A N/A balanced balanced
9228 * fully_busy nr_idle nr_idle N/A N/A balanced balanced
9229 * misfit_task force N/A N/A N/A force force
9230 * asym_packing force force N/A N/A force force
9231 * imbalanced force force N/A N/A force force
9232 * overloaded force force N/A N/A force avg_load
9234 * N/A : Not Applicable because already filtered while updating
9236 * balanced : The system is balanced for these 2 groups.
9237 * force : Calculate the imbalance as load migration is probably needed.
9238 * avg_load : Only if imbalance is significant enough.
9239 * nr_idle : dst_cpu is not busy and the number of idle CPUs is quite
9240 * different in groups.
9244 * find_busiest_group - Returns the busiest group within the sched_domain
9245 * if there is an imbalance.
9247 * Also calculates the amount of runnable load which should be moved
9248 * to restore balance.
9250 * @env: The load balancing environment.
9252 * Return: - The busiest group if imbalance exists.
9254 static struct sched_group *find_busiest_group(struct lb_env *env)
9256 struct sg_lb_stats *local, *busiest;
9257 struct sd_lb_stats sds;
9259 init_sd_lb_stats(&sds);
9262 * Compute the various statistics relevant for load balancing at
9265 update_sd_lb_stats(env, &sds);
9267 if (sched_energy_enabled()) {
9268 struct root_domain *rd = env->dst_rq->rd;
9270 if (rcu_dereference(rd->pd) && !READ_ONCE(rd->overutilized))
9274 local = &sds.local_stat;
9275 busiest = &sds.busiest_stat;
9277 /* There is no busy sibling group to pull tasks from */
9281 /* Misfit tasks should be dealt with regardless of the avg load */
9282 if (busiest->group_type == group_misfit_task)
9285 /* ASYM feature bypasses nice load balance check */
9286 if (busiest->group_type == group_asym_packing)
9290 * If the busiest group is imbalanced the below checks don't
9291 * work because they assume all things are equal, which typically
9292 * isn't true due to cpus_ptr constraints and the like.
9294 if (busiest->group_type == group_imbalanced)
9298 * If the local group is busier than the selected busiest group
9299 * don't try and pull any tasks.
9301 if (local->group_type > busiest->group_type)
9305 * When groups are overloaded, use the avg_load to ensure fairness
9308 if (local->group_type == group_overloaded) {
9310 * If the local group is more loaded than the selected
9311 * busiest group don't try to pull any tasks.
9313 if (local->avg_load >= busiest->avg_load)
9316 /* XXX broken for overlapping NUMA groups */
9317 sds.avg_load = (sds.total_load * SCHED_CAPACITY_SCALE) /
9321 * Don't pull any tasks if this group is already above the
9322 * domain average load.
9324 if (local->avg_load >= sds.avg_load)
9328 * If the busiest group is more loaded, use imbalance_pct to be
9331 if (100 * busiest->avg_load <=
9332 env->sd->imbalance_pct * local->avg_load)
9336 /* Try to move all excess tasks to child's sibling domain */
9337 if (sds.prefer_sibling && local->group_type == group_has_spare &&
9338 busiest->sum_nr_running > local->sum_nr_running + 1)
9341 if (busiest->group_type != group_overloaded) {
9342 if (env->idle == CPU_NOT_IDLE)
9344 * If the busiest group is not overloaded (and as a
9345 * result the local one too) but this CPU is already
9346 * busy, let another idle CPU try to pull task.
9350 if (busiest->group_weight > 1 &&
9351 local->idle_cpus <= (busiest->idle_cpus + 1))
9353 * If the busiest group is not overloaded
9354 * and there is no imbalance between this and busiest
9355 * group wrt idle CPUs, it is balanced. The imbalance
9356 * becomes significant if the diff is greater than 1
9357 * otherwise we might end up to just move the imbalance
9358 * on another group. Of course this applies only if
9359 * there is more than 1 CPU per group.
9363 if (busiest->sum_h_nr_running == 1)
9365 * busiest doesn't have any tasks waiting to run
9371 /* Looks like there is an imbalance. Compute it */
9372 calculate_imbalance(env, &sds);
9373 return env->imbalance ? sds.busiest : NULL;
9381 * find_busiest_queue - find the busiest runqueue among the CPUs in the group.
9383 static struct rq *find_busiest_queue(struct lb_env *env,
9384 struct sched_group *group)
9386 struct rq *busiest = NULL, *rq;
9387 unsigned long busiest_util = 0, busiest_load = 0, busiest_capacity = 1;
9388 unsigned int busiest_nr = 0;
9391 for_each_cpu_and(i, sched_group_span(group), env->cpus) {
9392 unsigned long capacity, load, util;
9393 unsigned int nr_running;
9397 rt = fbq_classify_rq(rq);
9400 * We classify groups/runqueues into three groups:
9401 * - regular: there are !numa tasks
9402 * - remote: there are numa tasks that run on the 'wrong' node
9403 * - all: there is no distinction
9405 * In order to avoid migrating ideally placed numa tasks,
9406 * ignore those when there's better options.
9408 * If we ignore the actual busiest queue to migrate another
9409 * task, the next balance pass can still reduce the busiest
9410 * queue by moving tasks around inside the node.
9412 * If we cannot move enough load due to this classification
9413 * the next pass will adjust the group classification and
9414 * allow migration of more tasks.
9416 * Both cases only affect the total convergence complexity.
9418 if (rt > env->fbq_type)
9421 nr_running = rq->cfs.h_nr_running;
9425 capacity = capacity_of(i);
9428 * For ASYM_CPUCAPACITY domains, don't pick a CPU that could
9429 * eventually lead to active_balancing high->low capacity.
9430 * Higher per-CPU capacity is considered better than balancing
9433 if (env->sd->flags & SD_ASYM_CPUCAPACITY &&
9434 !capacity_greater(capacity_of(env->dst_cpu), capacity) &&
9438 switch (env->migration_type) {
9441 * When comparing with load imbalance, use cpu_load()
9442 * which is not scaled with the CPU capacity.
9444 load = cpu_load(rq);
9446 if (nr_running == 1 && load > env->imbalance &&
9447 !check_cpu_capacity(rq, env->sd))
9451 * For the load comparisons with the other CPUs,
9452 * consider the cpu_load() scaled with the CPU
9453 * capacity, so that the load can be moved away
9454 * from the CPU that is potentially running at a
9457 * Thus we're looking for max(load_i / capacity_i),
9458 * crosswise multiplication to rid ourselves of the
9459 * division works out to:
9460 * load_i * capacity_j > load_j * capacity_i;
9461 * where j is our previous maximum.
9463 if (load * busiest_capacity > busiest_load * capacity) {
9464 busiest_load = load;
9465 busiest_capacity = capacity;
9471 util = cpu_util(cpu_of(rq));
9474 * Don't try to pull utilization from a CPU with one
9475 * running task. Whatever its utilization, we will fail
9478 if (nr_running <= 1)
9481 if (busiest_util < util) {
9482 busiest_util = util;
9488 if (busiest_nr < nr_running) {
9489 busiest_nr = nr_running;
9494 case migrate_misfit:
9496 * For ASYM_CPUCAPACITY domains with misfit tasks we
9497 * simply seek the "biggest" misfit task.
9499 if (rq->misfit_task_load > busiest_load) {
9500 busiest_load = rq->misfit_task_load;
9513 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
9514 * so long as it is large enough.
9516 #define MAX_PINNED_INTERVAL 512
9519 asym_active_balance(struct lb_env *env)
9522 * ASYM_PACKING needs to force migrate tasks from busy but
9523 * lower priority CPUs in order to pack all tasks in the
9524 * highest priority CPUs.
9526 return env->idle != CPU_NOT_IDLE && (env->sd->flags & SD_ASYM_PACKING) &&
9527 sched_asym_prefer(env->dst_cpu, env->src_cpu);
9531 imbalanced_active_balance(struct lb_env *env)
9533 struct sched_domain *sd = env->sd;
9536 * The imbalanced case includes the case of pinned tasks preventing a fair
9537 * distribution of the load on the system but also the even distribution of the
9538 * threads on a system with spare capacity
9540 if ((env->migration_type == migrate_task) &&
9541 (sd->nr_balance_failed > sd->cache_nice_tries+2))
9547 static int need_active_balance(struct lb_env *env)
9549 struct sched_domain *sd = env->sd;
9551 if (asym_active_balance(env))
9554 if (imbalanced_active_balance(env))
9558 * The dst_cpu is idle and the src_cpu CPU has only 1 CFS task.
9559 * It's worth migrating the task if the src_cpu's capacity is reduced
9560 * because of other sched_class or IRQs if more capacity stays
9561 * available on dst_cpu.
9563 if ((env->idle != CPU_NOT_IDLE) &&
9564 (env->src_rq->cfs.h_nr_running == 1)) {
9565 if ((check_cpu_capacity(env->src_rq, sd)) &&
9566 (capacity_of(env->src_cpu)*sd->imbalance_pct < capacity_of(env->dst_cpu)*100))
9570 if (env->migration_type == migrate_misfit)
9576 static int active_load_balance_cpu_stop(void *data);
9578 static int should_we_balance(struct lb_env *env)
9580 struct sched_group *sg = env->sd->groups;
9584 * Ensure the balancing environment is consistent; can happen
9585 * when the softirq triggers 'during' hotplug.
9587 if (!cpumask_test_cpu(env->dst_cpu, env->cpus))
9591 * In the newly idle case, we will allow all the CPUs
9592 * to do the newly idle load balance.
9594 if (env->idle == CPU_NEWLY_IDLE)
9597 /* Try to find first idle CPU */
9598 for_each_cpu_and(cpu, group_balance_mask(sg), env->cpus) {
9602 /* Are we the first idle CPU? */
9603 return cpu == env->dst_cpu;
9606 /* Are we the first CPU of this group ? */
9607 return group_balance_cpu(sg) == env->dst_cpu;
9611 * Check this_cpu to ensure it is balanced within domain. Attempt to move
9612 * tasks if there is an imbalance.
9614 static int load_balance(int this_cpu, struct rq *this_rq,
9615 struct sched_domain *sd, enum cpu_idle_type idle,
9616 int *continue_balancing)
9618 int ld_moved, cur_ld_moved, active_balance = 0;
9619 struct sched_domain *sd_parent = sd->parent;
9620 struct sched_group *group;
9623 struct cpumask *cpus = this_cpu_cpumask_var_ptr(load_balance_mask);
9625 struct lb_env env = {
9627 .dst_cpu = this_cpu,
9629 .dst_grpmask = sched_group_span(sd->groups),
9631 .loop_break = sched_nr_migrate_break,
9634 .tasks = LIST_HEAD_INIT(env.tasks),
9637 cpumask_and(cpus, sched_domain_span(sd), cpu_active_mask);
9639 schedstat_inc(sd->lb_count[idle]);
9642 if (!should_we_balance(&env)) {
9643 *continue_balancing = 0;
9647 group = find_busiest_group(&env);
9649 schedstat_inc(sd->lb_nobusyg[idle]);
9653 busiest = find_busiest_queue(&env, group);
9655 schedstat_inc(sd->lb_nobusyq[idle]);
9659 BUG_ON(busiest == env.dst_rq);
9661 schedstat_add(sd->lb_imbalance[idle], env.imbalance);
9663 env.src_cpu = busiest->cpu;
9664 env.src_rq = busiest;
9667 /* Clear this flag as soon as we find a pullable task */
9668 env.flags |= LBF_ALL_PINNED;
9669 if (busiest->nr_running > 1) {
9671 * Attempt to move tasks. If find_busiest_group has found
9672 * an imbalance but busiest->nr_running <= 1, the group is
9673 * still unbalanced. ld_moved simply stays zero, so it is
9674 * correctly treated as an imbalance.
9676 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
9679 rq_lock_irqsave(busiest, &rf);
9680 update_rq_clock(busiest);
9683 * cur_ld_moved - load moved in current iteration
9684 * ld_moved - cumulative load moved across iterations
9686 cur_ld_moved = detach_tasks(&env);
9689 * We've detached some tasks from busiest_rq. Every
9690 * task is masked "TASK_ON_RQ_MIGRATING", so we can safely
9691 * unlock busiest->lock, and we are able to be sure
9692 * that nobody can manipulate the tasks in parallel.
9693 * See task_rq_lock() family for the details.
9696 rq_unlock(busiest, &rf);
9700 ld_moved += cur_ld_moved;
9703 local_irq_restore(rf.flags);
9705 if (env.flags & LBF_NEED_BREAK) {
9706 env.flags &= ~LBF_NEED_BREAK;
9711 * Revisit (affine) tasks on src_cpu that couldn't be moved to
9712 * us and move them to an alternate dst_cpu in our sched_group
9713 * where they can run. The upper limit on how many times we
9714 * iterate on same src_cpu is dependent on number of CPUs in our
9717 * This changes load balance semantics a bit on who can move
9718 * load to a given_cpu. In addition to the given_cpu itself
9719 * (or a ilb_cpu acting on its behalf where given_cpu is
9720 * nohz-idle), we now have balance_cpu in a position to move
9721 * load to given_cpu. In rare situations, this may cause
9722 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
9723 * _independently_ and at _same_ time to move some load to
9724 * given_cpu) causing excess load to be moved to given_cpu.
9725 * This however should not happen so much in practice and
9726 * moreover subsequent load balance cycles should correct the
9727 * excess load moved.
9729 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
9731 /* Prevent to re-select dst_cpu via env's CPUs */
9732 __cpumask_clear_cpu(env.dst_cpu, env.cpus);
9734 env.dst_rq = cpu_rq(env.new_dst_cpu);
9735 env.dst_cpu = env.new_dst_cpu;
9736 env.flags &= ~LBF_DST_PINNED;
9738 env.loop_break = sched_nr_migrate_break;
9741 * Go back to "more_balance" rather than "redo" since we
9742 * need to continue with same src_cpu.
9748 * We failed to reach balance because of affinity.
9751 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
9753 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0)
9754 *group_imbalance = 1;
9757 /* All tasks on this runqueue were pinned by CPU affinity */
9758 if (unlikely(env.flags & LBF_ALL_PINNED)) {
9759 __cpumask_clear_cpu(cpu_of(busiest), cpus);
9761 * Attempting to continue load balancing at the current
9762 * sched_domain level only makes sense if there are
9763 * active CPUs remaining as possible busiest CPUs to
9764 * pull load from which are not contained within the
9765 * destination group that is receiving any migrated
9768 if (!cpumask_subset(cpus, env.dst_grpmask)) {
9770 env.loop_break = sched_nr_migrate_break;
9773 goto out_all_pinned;
9778 schedstat_inc(sd->lb_failed[idle]);
9780 * Increment the failure counter only on periodic balance.
9781 * We do not want newidle balance, which can be very
9782 * frequent, pollute the failure counter causing
9783 * excessive cache_hot migrations and active balances.
9785 if (idle != CPU_NEWLY_IDLE)
9786 sd->nr_balance_failed++;
9788 if (need_active_balance(&env)) {
9789 unsigned long flags;
9791 raw_spin_lock_irqsave(&busiest->lock, flags);
9794 * Don't kick the active_load_balance_cpu_stop,
9795 * if the curr task on busiest CPU can't be
9796 * moved to this_cpu:
9798 if (!cpumask_test_cpu(this_cpu, busiest->curr->cpus_ptr)) {
9799 raw_spin_unlock_irqrestore(&busiest->lock,
9801 goto out_one_pinned;
9804 /* Record that we found at least one task that could run on this_cpu */
9805 env.flags &= ~LBF_ALL_PINNED;
9808 * ->active_balance synchronizes accesses to
9809 * ->active_balance_work. Once set, it's cleared
9810 * only after active load balance is finished.
9812 if (!busiest->active_balance) {
9813 busiest->active_balance = 1;
9814 busiest->push_cpu = this_cpu;
9817 raw_spin_unlock_irqrestore(&busiest->lock, flags);
9819 if (active_balance) {
9820 stop_one_cpu_nowait(cpu_of(busiest),
9821 active_load_balance_cpu_stop, busiest,
9822 &busiest->active_balance_work);
9826 sd->nr_balance_failed = 0;
9829 if (likely(!active_balance) || need_active_balance(&env)) {
9830 /* We were unbalanced, so reset the balancing interval */
9831 sd->balance_interval = sd->min_interval;
9838 * We reach balance although we may have faced some affinity
9839 * constraints. Clear the imbalance flag only if other tasks got
9840 * a chance to move and fix the imbalance.
9842 if (sd_parent && !(env.flags & LBF_ALL_PINNED)) {
9843 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
9845 if (*group_imbalance)
9846 *group_imbalance = 0;
9851 * We reach balance because all tasks are pinned at this level so
9852 * we can't migrate them. Let the imbalance flag set so parent level
9853 * can try to migrate them.
9855 schedstat_inc(sd->lb_balanced[idle]);
9857 sd->nr_balance_failed = 0;
9863 * newidle_balance() disregards balance intervals, so we could
9864 * repeatedly reach this code, which would lead to balance_interval
9865 * skyrocketing in a short amount of time. Skip the balance_interval
9866 * increase logic to avoid that.
9868 if (env.idle == CPU_NEWLY_IDLE)
9871 /* tune up the balancing interval */
9872 if ((env.flags & LBF_ALL_PINNED &&
9873 sd->balance_interval < MAX_PINNED_INTERVAL) ||
9874 sd->balance_interval < sd->max_interval)
9875 sd->balance_interval *= 2;
9880 static inline unsigned long
9881 get_sd_balance_interval(struct sched_domain *sd, int cpu_busy)
9883 unsigned long interval = sd->balance_interval;
9886 interval *= sd->busy_factor;
9888 /* scale ms to jiffies */
9889 interval = msecs_to_jiffies(interval);
9892 * Reduce likelihood of busy balancing at higher domains racing with
9893 * balancing at lower domains by preventing their balancing periods
9894 * from being multiples of each other.
9899 interval = clamp(interval, 1UL, max_load_balance_interval);
9905 update_next_balance(struct sched_domain *sd, unsigned long *next_balance)
9907 unsigned long interval, next;
9909 /* used by idle balance, so cpu_busy = 0 */
9910 interval = get_sd_balance_interval(sd, 0);
9911 next = sd->last_balance + interval;
9913 if (time_after(*next_balance, next))
9914 *next_balance = next;
9918 * active_load_balance_cpu_stop is run by the CPU stopper. It pushes
9919 * running tasks off the busiest CPU onto idle CPUs. It requires at
9920 * least 1 task to be running on each physical CPU where possible, and
9921 * avoids physical / logical imbalances.
9923 static int active_load_balance_cpu_stop(void *data)
9925 struct rq *busiest_rq = data;
9926 int busiest_cpu = cpu_of(busiest_rq);
9927 int target_cpu = busiest_rq->push_cpu;
9928 struct rq *target_rq = cpu_rq(target_cpu);
9929 struct sched_domain *sd;
9930 struct task_struct *p = NULL;
9933 rq_lock_irq(busiest_rq, &rf);
9935 * Between queueing the stop-work and running it is a hole in which
9936 * CPUs can become inactive. We should not move tasks from or to
9939 if (!cpu_active(busiest_cpu) || !cpu_active(target_cpu))
9942 /* Make sure the requested CPU hasn't gone down in the meantime: */
9943 if (unlikely(busiest_cpu != smp_processor_id() ||
9944 !busiest_rq->active_balance))
9947 /* Is there any task to move? */
9948 if (busiest_rq->nr_running <= 1)
9952 * This condition is "impossible", if it occurs
9953 * we need to fix it. Originally reported by
9954 * Bjorn Helgaas on a 128-CPU setup.
9956 BUG_ON(busiest_rq == target_rq);
9958 /* Search for an sd spanning us and the target CPU. */
9960 for_each_domain(target_cpu, sd) {
9961 if (cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
9966 struct lb_env env = {
9968 .dst_cpu = target_cpu,
9969 .dst_rq = target_rq,
9970 .src_cpu = busiest_rq->cpu,
9971 .src_rq = busiest_rq,
9973 .flags = LBF_ACTIVE_LB,
9976 schedstat_inc(sd->alb_count);
9977 update_rq_clock(busiest_rq);
9979 p = detach_one_task(&env);
9981 schedstat_inc(sd->alb_pushed);
9982 /* Active balancing done, reset the failure counter. */
9983 sd->nr_balance_failed = 0;
9985 schedstat_inc(sd->alb_failed);
9990 busiest_rq->active_balance = 0;
9991 rq_unlock(busiest_rq, &rf);
9994 attach_one_task(target_rq, p);
10001 static DEFINE_SPINLOCK(balancing);
10004 * Scale the max load_balance interval with the number of CPUs in the system.
10005 * This trades load-balance latency on larger machines for less cross talk.
10007 void update_max_interval(void)
10009 max_load_balance_interval = HZ*num_online_cpus()/10;
10013 * It checks each scheduling domain to see if it is due to be balanced,
10014 * and initiates a balancing operation if so.
10016 * Balancing parameters are set up in init_sched_domains.
10018 static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle)
10020 int continue_balancing = 1;
10022 int busy = idle != CPU_IDLE && !sched_idle_cpu(cpu);
10023 unsigned long interval;
10024 struct sched_domain *sd;
10025 /* Earliest time when we have to do rebalance again */
10026 unsigned long next_balance = jiffies + 60*HZ;
10027 int update_next_balance = 0;
10028 int need_serialize, need_decay = 0;
10032 for_each_domain(cpu, sd) {
10034 * Decay the newidle max times here because this is a regular
10035 * visit to all the domains. Decay ~1% per second.
10037 if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
10038 sd->max_newidle_lb_cost =
10039 (sd->max_newidle_lb_cost * 253) / 256;
10040 sd->next_decay_max_lb_cost = jiffies + HZ;
10043 max_cost += sd->max_newidle_lb_cost;
10046 * Stop the load balance at this level. There is another
10047 * CPU in our sched group which is doing load balancing more
10050 if (!continue_balancing) {
10056 interval = get_sd_balance_interval(sd, busy);
10058 need_serialize = sd->flags & SD_SERIALIZE;
10059 if (need_serialize) {
10060 if (!spin_trylock(&balancing))
10064 if (time_after_eq(jiffies, sd->last_balance + interval)) {
10065 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
10067 * The LBF_DST_PINNED logic could have changed
10068 * env->dst_cpu, so we can't know our idle
10069 * state even if we migrated tasks. Update it.
10071 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
10072 busy = idle != CPU_IDLE && !sched_idle_cpu(cpu);
10074 sd->last_balance = jiffies;
10075 interval = get_sd_balance_interval(sd, busy);
10077 if (need_serialize)
10078 spin_unlock(&balancing);
10080 if (time_after(next_balance, sd->last_balance + interval)) {
10081 next_balance = sd->last_balance + interval;
10082 update_next_balance = 1;
10087 * Ensure the rq-wide value also decays but keep it at a
10088 * reasonable floor to avoid funnies with rq->avg_idle.
10090 rq->max_idle_balance_cost =
10091 max((u64)sysctl_sched_migration_cost, max_cost);
10096 * next_balance will be updated only when there is a need.
10097 * When the cpu is attached to null domain for ex, it will not be
10100 if (likely(update_next_balance))
10101 rq->next_balance = next_balance;
10105 static inline int on_null_domain(struct rq *rq)
10107 return unlikely(!rcu_dereference_sched(rq->sd));
10110 #ifdef CONFIG_NO_HZ_COMMON
10112 * idle load balancing details
10113 * - When one of the busy CPUs notice that there may be an idle rebalancing
10114 * needed, they will kick the idle load balancer, which then does idle
10115 * load balancing for all the idle CPUs.
10116 * - HK_FLAG_MISC CPUs are used for this task, because HK_FLAG_SCHED not set
10120 static inline int find_new_ilb(void)
10124 for_each_cpu_and(ilb, nohz.idle_cpus_mask,
10125 housekeeping_cpumask(HK_FLAG_MISC)) {
10127 if (ilb == smp_processor_id())
10138 * Kick a CPU to do the nohz balancing, if it is time for it. We pick any
10139 * idle CPU in the HK_FLAG_MISC housekeeping set (if there is one).
10141 static void kick_ilb(unsigned int flags)
10146 * Increase nohz.next_balance only when if full ilb is triggered but
10147 * not if we only update stats.
10149 if (flags & NOHZ_BALANCE_KICK)
10150 nohz.next_balance = jiffies+1;
10152 ilb_cpu = find_new_ilb();
10154 if (ilb_cpu >= nr_cpu_ids)
10158 * Access to rq::nohz_csd is serialized by NOHZ_KICK_MASK; he who sets
10159 * the first flag owns it; cleared by nohz_csd_func().
10161 flags = atomic_fetch_or(flags, nohz_flags(ilb_cpu));
10162 if (flags & NOHZ_KICK_MASK)
10166 * This way we generate an IPI on the target CPU which
10167 * is idle. And the softirq performing nohz idle load balance
10168 * will be run before returning from the IPI.
10170 smp_call_function_single_async(ilb_cpu, &cpu_rq(ilb_cpu)->nohz_csd);
10174 * Current decision point for kicking the idle load balancer in the presence
10175 * of idle CPUs in the system.
10177 static void nohz_balancer_kick(struct rq *rq)
10179 unsigned long now = jiffies;
10180 struct sched_domain_shared *sds;
10181 struct sched_domain *sd;
10182 int nr_busy, i, cpu = rq->cpu;
10183 unsigned int flags = 0;
10185 if (unlikely(rq->idle_balance))
10189 * We may be recently in ticked or tickless idle mode. At the first
10190 * busy tick after returning from idle, we will update the busy stats.
10192 nohz_balance_exit_idle(rq);
10195 * None are in tickless mode and hence no need for NOHZ idle load
10198 if (likely(!atomic_read(&nohz.nr_cpus)))
10201 if (READ_ONCE(nohz.has_blocked) &&
10202 time_after(now, READ_ONCE(nohz.next_blocked)))
10203 flags = NOHZ_STATS_KICK;
10205 if (time_before(now, nohz.next_balance))
10208 if (rq->nr_running >= 2) {
10209 flags = NOHZ_KICK_MASK;
10215 sd = rcu_dereference(rq->sd);
10218 * If there's a CFS task and the current CPU has reduced
10219 * capacity; kick the ILB to see if there's a better CPU to run
10222 if (rq->cfs.h_nr_running >= 1 && check_cpu_capacity(rq, sd)) {
10223 flags = NOHZ_KICK_MASK;
10228 sd = rcu_dereference(per_cpu(sd_asym_packing, cpu));
10231 * When ASYM_PACKING; see if there's a more preferred CPU
10232 * currently idle; in which case, kick the ILB to move tasks
10235 for_each_cpu_and(i, sched_domain_span(sd), nohz.idle_cpus_mask) {
10236 if (sched_asym_prefer(i, cpu)) {
10237 flags = NOHZ_KICK_MASK;
10243 sd = rcu_dereference(per_cpu(sd_asym_cpucapacity, cpu));
10246 * When ASYM_CPUCAPACITY; see if there's a higher capacity CPU
10247 * to run the misfit task on.
10249 if (check_misfit_status(rq, sd)) {
10250 flags = NOHZ_KICK_MASK;
10255 * For asymmetric systems, we do not want to nicely balance
10256 * cache use, instead we want to embrace asymmetry and only
10257 * ensure tasks have enough CPU capacity.
10259 * Skip the LLC logic because it's not relevant in that case.
10264 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
10267 * If there is an imbalance between LLC domains (IOW we could
10268 * increase the overall cache use), we need some less-loaded LLC
10269 * domain to pull some load. Likewise, we may need to spread
10270 * load within the current LLC domain (e.g. packed SMT cores but
10271 * other CPUs are idle). We can't really know from here how busy
10272 * the others are - so just get a nohz balance going if it looks
10273 * like this LLC domain has tasks we could move.
10275 nr_busy = atomic_read(&sds->nr_busy_cpus);
10277 flags = NOHZ_KICK_MASK;
10288 static void set_cpu_sd_state_busy(int cpu)
10290 struct sched_domain *sd;
10293 sd = rcu_dereference(per_cpu(sd_llc, cpu));
10295 if (!sd || !sd->nohz_idle)
10299 atomic_inc(&sd->shared->nr_busy_cpus);
10304 void nohz_balance_exit_idle(struct rq *rq)
10306 SCHED_WARN_ON(rq != this_rq());
10308 if (likely(!rq->nohz_tick_stopped))
10311 rq->nohz_tick_stopped = 0;
10312 cpumask_clear_cpu(rq->cpu, nohz.idle_cpus_mask);
10313 atomic_dec(&nohz.nr_cpus);
10315 set_cpu_sd_state_busy(rq->cpu);
10318 static void set_cpu_sd_state_idle(int cpu)
10320 struct sched_domain *sd;
10323 sd = rcu_dereference(per_cpu(sd_llc, cpu));
10325 if (!sd || sd->nohz_idle)
10329 atomic_dec(&sd->shared->nr_busy_cpus);
10335 * This routine will record that the CPU is going idle with tick stopped.
10336 * This info will be used in performing idle load balancing in the future.
10338 void nohz_balance_enter_idle(int cpu)
10340 struct rq *rq = cpu_rq(cpu);
10342 SCHED_WARN_ON(cpu != smp_processor_id());
10344 /* If this CPU is going down, then nothing needs to be done: */
10345 if (!cpu_active(cpu))
10348 /* Spare idle load balancing on CPUs that don't want to be disturbed: */
10349 if (!housekeeping_cpu(cpu, HK_FLAG_SCHED))
10353 * Can be set safely without rq->lock held
10354 * If a clear happens, it will have evaluated last additions because
10355 * rq->lock is held during the check and the clear
10357 rq->has_blocked_load = 1;
10360 * The tick is still stopped but load could have been added in the
10361 * meantime. We set the nohz.has_blocked flag to trig a check of the
10362 * *_avg. The CPU is already part of nohz.idle_cpus_mask so the clear
10363 * of nohz.has_blocked can only happen after checking the new load
10365 if (rq->nohz_tick_stopped)
10368 /* If we're a completely isolated CPU, we don't play: */
10369 if (on_null_domain(rq))
10372 rq->nohz_tick_stopped = 1;
10374 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
10375 atomic_inc(&nohz.nr_cpus);
10378 * Ensures that if nohz_idle_balance() fails to observe our
10379 * @idle_cpus_mask store, it must observe the @has_blocked
10382 smp_mb__after_atomic();
10384 set_cpu_sd_state_idle(cpu);
10388 * Each time a cpu enter idle, we assume that it has blocked load and
10389 * enable the periodic update of the load of idle cpus
10391 WRITE_ONCE(nohz.has_blocked, 1);
10394 static bool update_nohz_stats(struct rq *rq)
10396 unsigned int cpu = rq->cpu;
10398 if (!rq->has_blocked_load)
10401 if (!cpumask_test_cpu(cpu, nohz.idle_cpus_mask))
10404 if (!time_after(jiffies, READ_ONCE(rq->last_blocked_load_update_tick)))
10407 update_blocked_averages(cpu);
10409 return rq->has_blocked_load;
10413 * Internal function that runs load balance for all idle cpus. The load balance
10414 * can be a simple update of blocked load or a complete load balance with
10415 * tasks movement depending of flags.
10417 static void _nohz_idle_balance(struct rq *this_rq, unsigned int flags,
10418 enum cpu_idle_type idle)
10420 /* Earliest time when we have to do rebalance again */
10421 unsigned long now = jiffies;
10422 unsigned long next_balance = now + 60*HZ;
10423 bool has_blocked_load = false;
10424 int update_next_balance = 0;
10425 int this_cpu = this_rq->cpu;
10429 SCHED_WARN_ON((flags & NOHZ_KICK_MASK) == NOHZ_BALANCE_KICK);
10432 * We assume there will be no idle load after this update and clear
10433 * the has_blocked flag. If a cpu enters idle in the mean time, it will
10434 * set the has_blocked flag and trig another update of idle load.
10435 * Because a cpu that becomes idle, is added to idle_cpus_mask before
10436 * setting the flag, we are sure to not clear the state and not
10437 * check the load of an idle cpu.
10439 WRITE_ONCE(nohz.has_blocked, 0);
10442 * Ensures that if we miss the CPU, we must see the has_blocked
10443 * store from nohz_balance_enter_idle().
10448 * Start with the next CPU after this_cpu so we will end with this_cpu and let a
10449 * chance for other idle cpu to pull load.
10451 for_each_cpu_wrap(balance_cpu, nohz.idle_cpus_mask, this_cpu+1) {
10452 if (!idle_cpu(balance_cpu))
10456 * If this CPU gets work to do, stop the load balancing
10457 * work being done for other CPUs. Next load
10458 * balancing owner will pick it up.
10460 if (need_resched()) {
10461 has_blocked_load = true;
10465 rq = cpu_rq(balance_cpu);
10467 has_blocked_load |= update_nohz_stats(rq);
10470 * If time for next balance is due,
10473 if (time_after_eq(jiffies, rq->next_balance)) {
10474 struct rq_flags rf;
10476 rq_lock_irqsave(rq, &rf);
10477 update_rq_clock(rq);
10478 rq_unlock_irqrestore(rq, &rf);
10480 if (flags & NOHZ_BALANCE_KICK)
10481 rebalance_domains(rq, CPU_IDLE);
10484 if (time_after(next_balance, rq->next_balance)) {
10485 next_balance = rq->next_balance;
10486 update_next_balance = 1;
10491 * next_balance will be updated only when there is a need.
10492 * When the CPU is attached to null domain for ex, it will not be
10495 if (likely(update_next_balance))
10496 nohz.next_balance = next_balance;
10498 WRITE_ONCE(nohz.next_blocked,
10499 now + msecs_to_jiffies(LOAD_AVG_PERIOD));
10502 /* There is still blocked load, enable periodic update */
10503 if (has_blocked_load)
10504 WRITE_ONCE(nohz.has_blocked, 1);
10508 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
10509 * rebalancing for all the cpus for whom scheduler ticks are stopped.
10511 static bool nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
10513 unsigned int flags = this_rq->nohz_idle_balance;
10518 this_rq->nohz_idle_balance = 0;
10520 if (idle != CPU_IDLE)
10523 _nohz_idle_balance(this_rq, flags, idle);
10529 * Check if we need to run the ILB for updating blocked load before entering
10532 void nohz_run_idle_balance(int cpu)
10534 unsigned int flags;
10536 flags = atomic_fetch_andnot(NOHZ_NEWILB_KICK, nohz_flags(cpu));
10539 * Update the blocked load only if no SCHED_SOFTIRQ is about to happen
10540 * (ie NOHZ_STATS_KICK set) and will do the same.
10542 if ((flags == NOHZ_NEWILB_KICK) && !need_resched())
10543 _nohz_idle_balance(cpu_rq(cpu), NOHZ_STATS_KICK, CPU_IDLE);
10546 static void nohz_newidle_balance(struct rq *this_rq)
10548 int this_cpu = this_rq->cpu;
10551 * This CPU doesn't want to be disturbed by scheduler
10554 if (!housekeeping_cpu(this_cpu, HK_FLAG_SCHED))
10557 /* Will wake up very soon. No time for doing anything else*/
10558 if (this_rq->avg_idle < sysctl_sched_migration_cost)
10561 /* Don't need to update blocked load of idle CPUs*/
10562 if (!READ_ONCE(nohz.has_blocked) ||
10563 time_before(jiffies, READ_ONCE(nohz.next_blocked)))
10567 * Set the need to trigger ILB in order to update blocked load
10568 * before entering idle state.
10570 atomic_or(NOHZ_NEWILB_KICK, nohz_flags(this_cpu));
10573 #else /* !CONFIG_NO_HZ_COMMON */
10574 static inline void nohz_balancer_kick(struct rq *rq) { }
10576 static inline bool nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
10581 static inline void nohz_newidle_balance(struct rq *this_rq) { }
10582 #endif /* CONFIG_NO_HZ_COMMON */
10585 * newidle_balance is called by schedule() if this_cpu is about to become
10586 * idle. Attempts to pull tasks from other CPUs.
10589 * < 0 - we released the lock and there are !fair tasks present
10590 * 0 - failed, no new tasks
10591 * > 0 - success, new (fair) tasks present
10593 static int newidle_balance(struct rq *this_rq, struct rq_flags *rf)
10595 unsigned long next_balance = jiffies + HZ;
10596 int this_cpu = this_rq->cpu;
10597 struct sched_domain *sd;
10598 int pulled_task = 0;
10601 update_misfit_status(NULL, this_rq);
10604 * There is a task waiting to run. No need to search for one.
10605 * Return 0; the task will be enqueued when switching to idle.
10607 if (this_rq->ttwu_pending)
10611 * We must set idle_stamp _before_ calling idle_balance(), such that we
10612 * measure the duration of idle_balance() as idle time.
10614 this_rq->idle_stamp = rq_clock(this_rq);
10617 * Do not pull tasks towards !active CPUs...
10619 if (!cpu_active(this_cpu))
10623 * This is OK, because current is on_cpu, which avoids it being picked
10624 * for load-balance and preemption/IRQs are still disabled avoiding
10625 * further scheduler activity on it and we're being very careful to
10626 * re-start the picking loop.
10628 rq_unpin_lock(this_rq, rf);
10630 if (this_rq->avg_idle < sysctl_sched_migration_cost ||
10631 !READ_ONCE(this_rq->rd->overload)) {
10634 sd = rcu_dereference_check_sched_domain(this_rq->sd);
10636 update_next_balance(sd, &next_balance);
10642 raw_spin_unlock(&this_rq->lock);
10644 update_blocked_averages(this_cpu);
10646 for_each_domain(this_cpu, sd) {
10647 int continue_balancing = 1;
10648 u64 t0, domain_cost;
10650 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost) {
10651 update_next_balance(sd, &next_balance);
10655 if (sd->flags & SD_BALANCE_NEWIDLE) {
10656 t0 = sched_clock_cpu(this_cpu);
10658 pulled_task = load_balance(this_cpu, this_rq,
10659 sd, CPU_NEWLY_IDLE,
10660 &continue_balancing);
10662 domain_cost = sched_clock_cpu(this_cpu) - t0;
10663 if (domain_cost > sd->max_newidle_lb_cost)
10664 sd->max_newidle_lb_cost = domain_cost;
10666 curr_cost += domain_cost;
10669 update_next_balance(sd, &next_balance);
10672 * Stop searching for tasks to pull if there are
10673 * now runnable tasks on this rq.
10675 if (pulled_task || this_rq->nr_running > 0 ||
10676 this_rq->ttwu_pending)
10681 raw_spin_lock(&this_rq->lock);
10683 if (curr_cost > this_rq->max_idle_balance_cost)
10684 this_rq->max_idle_balance_cost = curr_cost;
10687 * While browsing the domains, we released the rq lock, a task could
10688 * have been enqueued in the meantime. Since we're not going idle,
10689 * pretend we pulled a task.
10691 if (this_rq->cfs.h_nr_running && !pulled_task)
10694 /* Is there a task of a high priority class? */
10695 if (this_rq->nr_running != this_rq->cfs.h_nr_running)
10699 /* Move the next balance forward */
10700 if (time_after(this_rq->next_balance, next_balance))
10701 this_rq->next_balance = next_balance;
10704 this_rq->idle_stamp = 0;
10706 nohz_newidle_balance(this_rq);
10708 rq_repin_lock(this_rq, rf);
10710 return pulled_task;
10714 * run_rebalance_domains is triggered when needed from the scheduler tick.
10715 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
10717 static __latent_entropy void run_rebalance_domains(struct softirq_action *h)
10719 struct rq *this_rq = this_rq();
10720 enum cpu_idle_type idle = this_rq->idle_balance ?
10721 CPU_IDLE : CPU_NOT_IDLE;
10724 * If this CPU has a pending nohz_balance_kick, then do the
10725 * balancing on behalf of the other idle CPUs whose ticks are
10726 * stopped. Do nohz_idle_balance *before* rebalance_domains to
10727 * give the idle CPUs a chance to load balance. Else we may
10728 * load balance only within the local sched_domain hierarchy
10729 * and abort nohz_idle_balance altogether if we pull some load.
10731 if (nohz_idle_balance(this_rq, idle))
10734 /* normal load balance */
10735 update_blocked_averages(this_rq->cpu);
10736 rebalance_domains(this_rq, idle);
10740 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
10742 void trigger_load_balance(struct rq *rq)
10745 * Don't need to rebalance while attached to NULL domain or
10746 * runqueue CPU is not active
10748 if (unlikely(on_null_domain(rq) || !cpu_active(cpu_of(rq))))
10751 if (time_after_eq(jiffies, rq->next_balance))
10752 raise_softirq(SCHED_SOFTIRQ);
10754 nohz_balancer_kick(rq);
10757 static void rq_online_fair(struct rq *rq)
10761 update_runtime_enabled(rq);
10764 static void rq_offline_fair(struct rq *rq)
10768 /* Ensure any throttled groups are reachable by pick_next_task */
10769 unthrottle_offline_cfs_rqs(rq);
10772 #endif /* CONFIG_SMP */
10775 * scheduler tick hitting a task of our scheduling class.
10777 * NOTE: This function can be called remotely by the tick offload that
10778 * goes along full dynticks. Therefore no local assumption can be made
10779 * and everything must be accessed through the @rq and @curr passed in
10782 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
10784 struct cfs_rq *cfs_rq;
10785 struct sched_entity *se = &curr->se;
10787 for_each_sched_entity(se) {
10788 cfs_rq = cfs_rq_of(se);
10789 entity_tick(cfs_rq, se, queued);
10792 if (static_branch_unlikely(&sched_numa_balancing))
10793 task_tick_numa(rq, curr);
10795 update_misfit_status(curr, rq);
10796 update_overutilized_status(task_rq(curr));
10800 * called on fork with the child task as argument from the parent's context
10801 * - child not yet on the tasklist
10802 * - preemption disabled
10804 static void task_fork_fair(struct task_struct *p)
10806 struct cfs_rq *cfs_rq;
10807 struct sched_entity *se = &p->se, *curr;
10808 struct rq *rq = this_rq();
10809 struct rq_flags rf;
10812 update_rq_clock(rq);
10814 cfs_rq = task_cfs_rq(current);
10815 curr = cfs_rq->curr;
10817 update_curr(cfs_rq);
10818 se->vruntime = curr->vruntime;
10820 place_entity(cfs_rq, se, 1);
10822 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
10824 * Upon rescheduling, sched_class::put_prev_task() will place
10825 * 'current' within the tree based on its new key value.
10827 swap(curr->vruntime, se->vruntime);
10831 se->vruntime -= cfs_rq->min_vruntime;
10832 rq_unlock(rq, &rf);
10836 * Priority of the task has changed. Check to see if we preempt
10837 * the current task.
10840 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
10842 if (!task_on_rq_queued(p))
10845 if (rq->cfs.nr_running == 1)
10849 * Reschedule if we are currently running on this runqueue and
10850 * our priority decreased, or if we are not currently running on
10851 * this runqueue and our priority is higher than the current's
10853 if (task_current(rq, p)) {
10854 if (p->prio > oldprio)
10857 check_preempt_curr(rq, p, 0);
10860 static inline bool vruntime_normalized(struct task_struct *p)
10862 struct sched_entity *se = &p->se;
10865 * In both the TASK_ON_RQ_QUEUED and TASK_ON_RQ_MIGRATING cases,
10866 * the dequeue_entity(.flags=0) will already have normalized the
10873 * When !on_rq, vruntime of the task has usually NOT been normalized.
10874 * But there are some cases where it has already been normalized:
10876 * - A forked child which is waiting for being woken up by
10877 * wake_up_new_task().
10878 * - A task which has been woken up by try_to_wake_up() and
10879 * waiting for actually being woken up by sched_ttwu_pending().
10881 if (!se->sum_exec_runtime ||
10882 (p->state == TASK_WAKING && p->sched_remote_wakeup))
10888 #ifdef CONFIG_FAIR_GROUP_SCHED
10890 * Propagate the changes of the sched_entity across the tg tree to make it
10891 * visible to the root
10893 static void propagate_entity_cfs_rq(struct sched_entity *se)
10895 struct cfs_rq *cfs_rq;
10897 list_add_leaf_cfs_rq(cfs_rq_of(se));
10899 /* Start to propagate at parent */
10902 for_each_sched_entity(se) {
10903 cfs_rq = cfs_rq_of(se);
10905 if (!cfs_rq_throttled(cfs_rq)){
10906 update_load_avg(cfs_rq, se, UPDATE_TG);
10907 list_add_leaf_cfs_rq(cfs_rq);
10911 if (list_add_leaf_cfs_rq(cfs_rq))
10916 static void propagate_entity_cfs_rq(struct sched_entity *se) { }
10919 static void detach_entity_cfs_rq(struct sched_entity *se)
10921 struct cfs_rq *cfs_rq = cfs_rq_of(se);
10923 /* Catch up with the cfs_rq and remove our load when we leave */
10924 update_load_avg(cfs_rq, se, 0);
10925 detach_entity_load_avg(cfs_rq, se);
10926 update_tg_load_avg(cfs_rq);
10927 propagate_entity_cfs_rq(se);
10930 static void attach_entity_cfs_rq(struct sched_entity *se)
10932 struct cfs_rq *cfs_rq = cfs_rq_of(se);
10934 #ifdef CONFIG_FAIR_GROUP_SCHED
10936 * Since the real-depth could have been changed (only FAIR
10937 * class maintain depth value), reset depth properly.
10939 se->depth = se->parent ? se->parent->depth + 1 : 0;
10942 /* Synchronize entity with its cfs_rq */
10943 update_load_avg(cfs_rq, se, sched_feat(ATTACH_AGE_LOAD) ? 0 : SKIP_AGE_LOAD);
10944 attach_entity_load_avg(cfs_rq, se);
10945 update_tg_load_avg(cfs_rq);
10946 propagate_entity_cfs_rq(se);
10949 static void detach_task_cfs_rq(struct task_struct *p)
10951 struct sched_entity *se = &p->se;
10952 struct cfs_rq *cfs_rq = cfs_rq_of(se);
10954 if (!vruntime_normalized(p)) {
10956 * Fix up our vruntime so that the current sleep doesn't
10957 * cause 'unlimited' sleep bonus.
10959 place_entity(cfs_rq, se, 0);
10960 se->vruntime -= cfs_rq->min_vruntime;
10963 detach_entity_cfs_rq(se);
10966 static void attach_task_cfs_rq(struct task_struct *p)
10968 struct sched_entity *se = &p->se;
10969 struct cfs_rq *cfs_rq = cfs_rq_of(se);
10971 attach_entity_cfs_rq(se);
10973 if (!vruntime_normalized(p))
10974 se->vruntime += cfs_rq->min_vruntime;
10977 static void switched_from_fair(struct rq *rq, struct task_struct *p)
10979 detach_task_cfs_rq(p);
10982 static void switched_to_fair(struct rq *rq, struct task_struct *p)
10984 attach_task_cfs_rq(p);
10986 if (task_on_rq_queued(p)) {
10988 * We were most likely switched from sched_rt, so
10989 * kick off the schedule if running, otherwise just see
10990 * if we can still preempt the current task.
10992 if (task_current(rq, p))
10995 check_preempt_curr(rq, p, 0);
10999 /* Account for a task changing its policy or group.
11001 * This routine is mostly called to set cfs_rq->curr field when a task
11002 * migrates between groups/classes.
11004 static void set_next_task_fair(struct rq *rq, struct task_struct *p, bool first)
11006 struct sched_entity *se = &p->se;
11009 if (task_on_rq_queued(p)) {
11011 * Move the next running task to the front of the list, so our
11012 * cfs_tasks list becomes MRU one.
11014 list_move(&se->group_node, &rq->cfs_tasks);
11018 for_each_sched_entity(se) {
11019 struct cfs_rq *cfs_rq = cfs_rq_of(se);
11021 set_next_entity(cfs_rq, se);
11022 /* ensure bandwidth has been allocated on our new cfs_rq */
11023 account_cfs_rq_runtime(cfs_rq, 0);
11027 void init_cfs_rq(struct cfs_rq *cfs_rq)
11029 cfs_rq->tasks_timeline = RB_ROOT_CACHED;
11030 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
11031 #ifndef CONFIG_64BIT
11032 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
11035 raw_spin_lock_init(&cfs_rq->removed.lock);
11039 #ifdef CONFIG_FAIR_GROUP_SCHED
11040 static void task_set_group_fair(struct task_struct *p)
11042 struct sched_entity *se = &p->se;
11044 set_task_rq(p, task_cpu(p));
11045 se->depth = se->parent ? se->parent->depth + 1 : 0;
11048 static void task_move_group_fair(struct task_struct *p)
11050 detach_task_cfs_rq(p);
11051 set_task_rq(p, task_cpu(p));
11054 /* Tell se's cfs_rq has been changed -- migrated */
11055 p->se.avg.last_update_time = 0;
11057 attach_task_cfs_rq(p);
11060 static void task_change_group_fair(struct task_struct *p, int type)
11063 case TASK_SET_GROUP:
11064 task_set_group_fair(p);
11067 case TASK_MOVE_GROUP:
11068 task_move_group_fair(p);
11073 void free_fair_sched_group(struct task_group *tg)
11077 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
11079 for_each_possible_cpu(i) {
11081 kfree(tg->cfs_rq[i]);
11090 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
11092 struct sched_entity *se;
11093 struct cfs_rq *cfs_rq;
11096 tg->cfs_rq = kcalloc(nr_cpu_ids, sizeof(cfs_rq), GFP_KERNEL);
11099 tg->se = kcalloc(nr_cpu_ids, sizeof(se), GFP_KERNEL);
11103 tg->shares = NICE_0_LOAD;
11105 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
11107 for_each_possible_cpu(i) {
11108 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
11109 GFP_KERNEL, cpu_to_node(i));
11113 se = kzalloc_node(sizeof(struct sched_entity),
11114 GFP_KERNEL, cpu_to_node(i));
11118 init_cfs_rq(cfs_rq);
11119 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
11120 init_entity_runnable_average(se);
11131 void online_fair_sched_group(struct task_group *tg)
11133 struct sched_entity *se;
11134 struct rq_flags rf;
11138 for_each_possible_cpu(i) {
11141 rq_lock_irq(rq, &rf);
11142 update_rq_clock(rq);
11143 attach_entity_cfs_rq(se);
11144 sync_throttle(tg, i);
11145 rq_unlock_irq(rq, &rf);
11149 void unregister_fair_sched_group(struct task_group *tg)
11151 unsigned long flags;
11155 for_each_possible_cpu(cpu) {
11157 remove_entity_load_avg(tg->se[cpu]);
11160 * Only empty task groups can be destroyed; so we can speculatively
11161 * check on_list without danger of it being re-added.
11163 if (!tg->cfs_rq[cpu]->on_list)
11168 raw_spin_lock_irqsave(&rq->lock, flags);
11169 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
11170 raw_spin_unlock_irqrestore(&rq->lock, flags);
11174 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
11175 struct sched_entity *se, int cpu,
11176 struct sched_entity *parent)
11178 struct rq *rq = cpu_rq(cpu);
11182 init_cfs_rq_runtime(cfs_rq);
11184 tg->cfs_rq[cpu] = cfs_rq;
11187 /* se could be NULL for root_task_group */
11192 se->cfs_rq = &rq->cfs;
11195 se->cfs_rq = parent->my_q;
11196 se->depth = parent->depth + 1;
11200 /* guarantee group entities always have weight */
11201 update_load_set(&se->load, NICE_0_LOAD);
11202 se->parent = parent;
11205 static DEFINE_MUTEX(shares_mutex);
11207 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
11212 * We can't change the weight of the root cgroup.
11217 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
11219 mutex_lock(&shares_mutex);
11220 if (tg->shares == shares)
11223 tg->shares = shares;
11224 for_each_possible_cpu(i) {
11225 struct rq *rq = cpu_rq(i);
11226 struct sched_entity *se = tg->se[i];
11227 struct rq_flags rf;
11229 /* Propagate contribution to hierarchy */
11230 rq_lock_irqsave(rq, &rf);
11231 update_rq_clock(rq);
11232 for_each_sched_entity(se) {
11233 update_load_avg(cfs_rq_of(se), se, UPDATE_TG);
11234 update_cfs_group(se);
11236 rq_unlock_irqrestore(rq, &rf);
11240 mutex_unlock(&shares_mutex);
11243 #else /* CONFIG_FAIR_GROUP_SCHED */
11245 void free_fair_sched_group(struct task_group *tg) { }
11247 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
11252 void online_fair_sched_group(struct task_group *tg) { }
11254 void unregister_fair_sched_group(struct task_group *tg) { }
11256 #endif /* CONFIG_FAIR_GROUP_SCHED */
11259 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
11261 struct sched_entity *se = &task->se;
11262 unsigned int rr_interval = 0;
11265 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
11268 if (rq->cfs.load.weight)
11269 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
11271 return rr_interval;
11275 * All the scheduling class methods:
11277 DEFINE_SCHED_CLASS(fair) = {
11279 .enqueue_task = enqueue_task_fair,
11280 .dequeue_task = dequeue_task_fair,
11281 .yield_task = yield_task_fair,
11282 .yield_to_task = yield_to_task_fair,
11284 .check_preempt_curr = check_preempt_wakeup,
11286 .pick_next_task = __pick_next_task_fair,
11287 .put_prev_task = put_prev_task_fair,
11288 .set_next_task = set_next_task_fair,
11291 .balance = balance_fair,
11292 .select_task_rq = select_task_rq_fair,
11293 .migrate_task_rq = migrate_task_rq_fair,
11295 .rq_online = rq_online_fair,
11296 .rq_offline = rq_offline_fair,
11298 .task_dead = task_dead_fair,
11299 .set_cpus_allowed = set_cpus_allowed_common,
11302 .task_tick = task_tick_fair,
11303 .task_fork = task_fork_fair,
11305 .prio_changed = prio_changed_fair,
11306 .switched_from = switched_from_fair,
11307 .switched_to = switched_to_fair,
11309 .get_rr_interval = get_rr_interval_fair,
11311 .update_curr = update_curr_fair,
11313 #ifdef CONFIG_FAIR_GROUP_SCHED
11314 .task_change_group = task_change_group_fair,
11317 #ifdef CONFIG_UCLAMP_TASK
11318 .uclamp_enabled = 1,
11322 #ifdef CONFIG_SCHED_DEBUG
11323 void print_cfs_stats(struct seq_file *m, int cpu)
11325 struct cfs_rq *cfs_rq, *pos;
11328 for_each_leaf_cfs_rq_safe(cpu_rq(cpu), cfs_rq, pos)
11329 print_cfs_rq(m, cpu, cfs_rq);
11333 #ifdef CONFIG_NUMA_BALANCING
11334 void show_numa_stats(struct task_struct *p, struct seq_file *m)
11337 unsigned long tsf = 0, tpf = 0, gsf = 0, gpf = 0;
11338 struct numa_group *ng;
11341 ng = rcu_dereference(p->numa_group);
11342 for_each_online_node(node) {
11343 if (p->numa_faults) {
11344 tsf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 0)];
11345 tpf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 1)];
11348 gsf = ng->faults[task_faults_idx(NUMA_MEM, node, 0)],
11349 gpf = ng->faults[task_faults_idx(NUMA_MEM, node, 1)];
11351 print_numa_stats(m, node, tsf, tpf, gsf, gpf);
11355 #endif /* CONFIG_NUMA_BALANCING */
11356 #endif /* CONFIG_SCHED_DEBUG */
11358 __init void init_sched_fair_class(void)
11361 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
11363 #ifdef CONFIG_NO_HZ_COMMON
11364 nohz.next_balance = jiffies;
11365 nohz.next_blocked = jiffies;
11366 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
11373 * Helper functions to facilitate extracting info from tracepoints.
11376 const struct sched_avg *sched_trace_cfs_rq_avg(struct cfs_rq *cfs_rq)
11379 return cfs_rq ? &cfs_rq->avg : NULL;
11384 EXPORT_SYMBOL_GPL(sched_trace_cfs_rq_avg);
11386 char *sched_trace_cfs_rq_path(struct cfs_rq *cfs_rq, char *str, int len)
11390 strlcpy(str, "(null)", len);
11395 cfs_rq_tg_path(cfs_rq, str, len);
11398 EXPORT_SYMBOL_GPL(sched_trace_cfs_rq_path);
11400 int sched_trace_cfs_rq_cpu(struct cfs_rq *cfs_rq)
11402 return cfs_rq ? cpu_of(rq_of(cfs_rq)) : -1;
11404 EXPORT_SYMBOL_GPL(sched_trace_cfs_rq_cpu);
11406 const struct sched_avg *sched_trace_rq_avg_rt(struct rq *rq)
11409 return rq ? &rq->avg_rt : NULL;
11414 EXPORT_SYMBOL_GPL(sched_trace_rq_avg_rt);
11416 const struct sched_avg *sched_trace_rq_avg_dl(struct rq *rq)
11419 return rq ? &rq->avg_dl : NULL;
11424 EXPORT_SYMBOL_GPL(sched_trace_rq_avg_dl);
11426 const struct sched_avg *sched_trace_rq_avg_irq(struct rq *rq)
11428 #if defined(CONFIG_SMP) && defined(CONFIG_HAVE_SCHED_AVG_IRQ)
11429 return rq ? &rq->avg_irq : NULL;
11434 EXPORT_SYMBOL_GPL(sched_trace_rq_avg_irq);
11436 int sched_trace_rq_cpu(struct rq *rq)
11438 return rq ? cpu_of(rq) : -1;
11440 EXPORT_SYMBOL_GPL(sched_trace_rq_cpu);
11442 int sched_trace_rq_cpu_capacity(struct rq *rq)
11448 SCHED_CAPACITY_SCALE
11452 EXPORT_SYMBOL_GPL(sched_trace_rq_cpu_capacity);
11454 const struct cpumask *sched_trace_rd_span(struct root_domain *rd)
11457 return rd ? rd->span : NULL;
11462 EXPORT_SYMBOL_GPL(sched_trace_rd_span);
11464 int sched_trace_rq_nr_running(struct rq *rq)
11466 return rq ? rq->nr_running : -1;
11468 EXPORT_SYMBOL_GPL(sched_trace_rq_nr_running);