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
23 #include <linux/energy_model.h>
24 #include <linux/mmap_lock.h>
25 #include <linux/hugetlb_inline.h>
26 #include <linux/jiffies.h>
27 #include <linux/mm_api.h>
28 #include <linux/highmem.h>
29 #include <linux/spinlock_api.h>
30 #include <linux/cpumask_api.h>
31 #include <linux/lockdep_api.h>
32 #include <linux/softirq.h>
33 #include <linux/refcount_api.h>
34 #include <linux/topology.h>
35 #include <linux/sched/clock.h>
36 #include <linux/sched/cond_resched.h>
37 #include <linux/sched/cputime.h>
38 #include <linux/sched/isolation.h>
39 #include <linux/sched/nohz.h>
41 #include <linux/cpuidle.h>
42 #include <linux/interrupt.h>
43 #include <linux/memory-tiers.h>
44 #include <linux/mempolicy.h>
45 #include <linux/mutex_api.h>
46 #include <linux/profile.h>
47 #include <linux/psi.h>
48 #include <linux/ratelimit.h>
49 #include <linux/task_work.h>
51 #include <asm/switch_to.h>
53 #include <linux/sched/cond_resched.h>
57 #include "autogroup.h"
60 * Targeted preemption latency for CPU-bound tasks:
62 * NOTE: this latency value is not the same as the concept of
63 * 'timeslice length' - timeslices in CFS are of variable length
64 * and have no persistent notion like in traditional, time-slice
65 * based scheduling concepts.
67 * (to see the precise effective timeslice length of your workload,
68 * run vmstat and monitor the context-switches (cs) field)
70 * (default: 6ms * (1 + ilog(ncpus)), units: nanoseconds)
72 unsigned int sysctl_sched_latency = 6000000ULL;
73 static unsigned int normalized_sysctl_sched_latency = 6000000ULL;
76 * The initial- and re-scaling of tunables is configurable
80 * SCHED_TUNABLESCALING_NONE - unscaled, always *1
81 * SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus)
82 * SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus
84 * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus))
86 unsigned int sysctl_sched_tunable_scaling = SCHED_TUNABLESCALING_LOG;
89 * Minimal preemption granularity for CPU-bound tasks:
91 * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds)
93 unsigned int sysctl_sched_min_granularity = 750000ULL;
94 static unsigned int normalized_sysctl_sched_min_granularity = 750000ULL;
97 * Minimal preemption granularity for CPU-bound SCHED_IDLE tasks.
98 * Applies only when SCHED_IDLE tasks compete with normal tasks.
100 * (default: 0.75 msec)
102 unsigned int sysctl_sched_idle_min_granularity = 750000ULL;
105 * This value is kept at sysctl_sched_latency/sysctl_sched_min_granularity
107 static unsigned int sched_nr_latency = 8;
110 * After fork, child runs first. If set to 0 (default) then
111 * parent will (try to) run first.
113 unsigned int sysctl_sched_child_runs_first __read_mostly;
116 * SCHED_OTHER wake-up granularity.
118 * This option delays the preemption effects of decoupled workloads
119 * and reduces their over-scheduling. Synchronous workloads will still
120 * have immediate wakeup/sleep latencies.
122 * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds)
124 unsigned int sysctl_sched_wakeup_granularity = 1000000UL;
125 static unsigned int normalized_sysctl_sched_wakeup_granularity = 1000000UL;
127 const_debug unsigned int sysctl_sched_migration_cost = 500000UL;
129 int sched_thermal_decay_shift;
130 static int __init setup_sched_thermal_decay_shift(char *str)
134 if (kstrtoint(str, 0, &_shift))
135 pr_warn("Unable to set scheduler thermal pressure decay shift parameter\n");
137 sched_thermal_decay_shift = clamp(_shift, 0, 10);
140 __setup("sched_thermal_decay_shift=", setup_sched_thermal_decay_shift);
144 * For asym packing, by default the lower numbered CPU has higher priority.
146 int __weak arch_asym_cpu_priority(int cpu)
152 * The margin used when comparing utilization with CPU capacity.
156 #define fits_capacity(cap, max) ((cap) * 1280 < (max) * 1024)
159 * The margin used when comparing CPU capacities.
160 * is 'cap1' noticeably greater than 'cap2'
164 #define capacity_greater(cap1, cap2) ((cap1) * 1024 > (cap2) * 1078)
167 #ifdef CONFIG_CFS_BANDWIDTH
169 * Amount of runtime to allocate from global (tg) to local (per-cfs_rq) pool
170 * each time a cfs_rq requests quota.
172 * Note: in the case that the slice exceeds the runtime remaining (either due
173 * to consumption or the quota being specified to be smaller than the slice)
174 * we will always only issue the remaining available time.
176 * (default: 5 msec, units: microseconds)
178 static unsigned int sysctl_sched_cfs_bandwidth_slice = 5000UL;
182 static struct ctl_table sched_fair_sysctls[] = {
184 .procname = "sched_child_runs_first",
185 .data = &sysctl_sched_child_runs_first,
186 .maxlen = sizeof(unsigned int),
188 .proc_handler = proc_dointvec,
190 #ifdef CONFIG_CFS_BANDWIDTH
192 .procname = "sched_cfs_bandwidth_slice_us",
193 .data = &sysctl_sched_cfs_bandwidth_slice,
194 .maxlen = sizeof(unsigned int),
196 .proc_handler = proc_dointvec_minmax,
197 .extra1 = SYSCTL_ONE,
203 static int __init sched_fair_sysctl_init(void)
205 register_sysctl_init("kernel", sched_fair_sysctls);
208 late_initcall(sched_fair_sysctl_init);
211 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
217 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
223 static inline void update_load_set(struct load_weight *lw, unsigned long w)
230 * Increase the granularity value when there are more CPUs,
231 * because with more CPUs the 'effective latency' as visible
232 * to users decreases. But the relationship is not linear,
233 * so pick a second-best guess by going with the log2 of the
236 * This idea comes from the SD scheduler of Con Kolivas:
238 static unsigned int get_update_sysctl_factor(void)
240 unsigned int cpus = min_t(unsigned int, num_online_cpus(), 8);
243 switch (sysctl_sched_tunable_scaling) {
244 case SCHED_TUNABLESCALING_NONE:
247 case SCHED_TUNABLESCALING_LINEAR:
250 case SCHED_TUNABLESCALING_LOG:
252 factor = 1 + ilog2(cpus);
259 static void update_sysctl(void)
261 unsigned int factor = get_update_sysctl_factor();
263 #define SET_SYSCTL(name) \
264 (sysctl_##name = (factor) * normalized_sysctl_##name)
265 SET_SYSCTL(sched_min_granularity);
266 SET_SYSCTL(sched_latency);
267 SET_SYSCTL(sched_wakeup_granularity);
271 void __init sched_init_granularity(void)
276 #define WMULT_CONST (~0U)
277 #define WMULT_SHIFT 32
279 static void __update_inv_weight(struct load_weight *lw)
283 if (likely(lw->inv_weight))
286 w = scale_load_down(lw->weight);
288 if (BITS_PER_LONG > 32 && unlikely(w >= WMULT_CONST))
290 else if (unlikely(!w))
291 lw->inv_weight = WMULT_CONST;
293 lw->inv_weight = WMULT_CONST / w;
297 * delta_exec * weight / lw.weight
299 * (delta_exec * (weight * lw->inv_weight)) >> WMULT_SHIFT
301 * Either weight := NICE_0_LOAD and lw \e sched_prio_to_wmult[], in which case
302 * we're guaranteed shift stays positive because inv_weight is guaranteed to
303 * fit 32 bits, and NICE_0_LOAD gives another 10 bits; therefore shift >= 22.
305 * Or, weight =< lw.weight (because lw.weight is the runqueue weight), thus
306 * weight/lw.weight <= 1, and therefore our shift will also be positive.
308 static u64 __calc_delta(u64 delta_exec, unsigned long weight, struct load_weight *lw)
310 u64 fact = scale_load_down(weight);
311 u32 fact_hi = (u32)(fact >> 32);
312 int shift = WMULT_SHIFT;
315 __update_inv_weight(lw);
317 if (unlikely(fact_hi)) {
323 fact = mul_u32_u32(fact, lw->inv_weight);
325 fact_hi = (u32)(fact >> 32);
332 return mul_u64_u32_shr(delta_exec, fact, shift);
336 const struct sched_class fair_sched_class;
338 /**************************************************************
339 * CFS operations on generic schedulable entities:
342 #ifdef CONFIG_FAIR_GROUP_SCHED
344 /* Walk up scheduling entities hierarchy */
345 #define for_each_sched_entity(se) \
346 for (; se; se = se->parent)
348 static inline bool list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
350 struct rq *rq = rq_of(cfs_rq);
351 int cpu = cpu_of(rq);
354 return rq->tmp_alone_branch == &rq->leaf_cfs_rq_list;
359 * Ensure we either appear before our parent (if already
360 * enqueued) or force our parent to appear after us when it is
361 * enqueued. The fact that we always enqueue bottom-up
362 * reduces this to two cases and a special case for the root
363 * cfs_rq. Furthermore, it also means that we will always reset
364 * tmp_alone_branch either when the branch is connected
365 * to a tree or when we reach the top of the tree
367 if (cfs_rq->tg->parent &&
368 cfs_rq->tg->parent->cfs_rq[cpu]->on_list) {
370 * If parent is already on the list, we add the child
371 * just before. Thanks to circular linked property of
372 * the list, this means to put the child at the tail
373 * of the list that starts by parent.
375 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
376 &(cfs_rq->tg->parent->cfs_rq[cpu]->leaf_cfs_rq_list));
378 * The branch is now connected to its tree so we can
379 * reset tmp_alone_branch to the beginning of the
382 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
386 if (!cfs_rq->tg->parent) {
388 * cfs rq without parent should be put
389 * at the tail of the list.
391 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
392 &rq->leaf_cfs_rq_list);
394 * We have reach the top of a tree so we can reset
395 * tmp_alone_branch to the beginning of the list.
397 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
402 * The parent has not already been added so we want to
403 * make sure that it will be put after us.
404 * tmp_alone_branch points to the begin of the branch
405 * where we will add parent.
407 list_add_rcu(&cfs_rq->leaf_cfs_rq_list, rq->tmp_alone_branch);
409 * update tmp_alone_branch to points to the new begin
412 rq->tmp_alone_branch = &cfs_rq->leaf_cfs_rq_list;
416 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
418 if (cfs_rq->on_list) {
419 struct rq *rq = rq_of(cfs_rq);
422 * With cfs_rq being unthrottled/throttled during an enqueue,
423 * it can happen the tmp_alone_branch points the a leaf that
424 * we finally want to del. In this case, tmp_alone_branch moves
425 * to the prev element but it will point to rq->leaf_cfs_rq_list
426 * at the end of the enqueue.
428 if (rq->tmp_alone_branch == &cfs_rq->leaf_cfs_rq_list)
429 rq->tmp_alone_branch = cfs_rq->leaf_cfs_rq_list.prev;
431 list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
436 static inline void assert_list_leaf_cfs_rq(struct rq *rq)
438 SCHED_WARN_ON(rq->tmp_alone_branch != &rq->leaf_cfs_rq_list);
441 /* Iterate thr' all leaf cfs_rq's on a runqueue */
442 #define for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) \
443 list_for_each_entry_safe(cfs_rq, pos, &rq->leaf_cfs_rq_list, \
446 /* Do the two (enqueued) entities belong to the same group ? */
447 static inline struct cfs_rq *
448 is_same_group(struct sched_entity *se, struct sched_entity *pse)
450 if (se->cfs_rq == pse->cfs_rq)
456 static inline struct sched_entity *parent_entity(struct sched_entity *se)
462 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
464 int se_depth, pse_depth;
467 * preemption test can be made between sibling entities who are in the
468 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
469 * both tasks until we find their ancestors who are siblings of common
473 /* First walk up until both entities are at same depth */
474 se_depth = (*se)->depth;
475 pse_depth = (*pse)->depth;
477 while (se_depth > pse_depth) {
479 *se = parent_entity(*se);
482 while (pse_depth > se_depth) {
484 *pse = parent_entity(*pse);
487 while (!is_same_group(*se, *pse)) {
488 *se = parent_entity(*se);
489 *pse = parent_entity(*pse);
493 static int tg_is_idle(struct task_group *tg)
498 static int cfs_rq_is_idle(struct cfs_rq *cfs_rq)
500 return cfs_rq->idle > 0;
503 static int se_is_idle(struct sched_entity *se)
505 if (entity_is_task(se))
506 return task_has_idle_policy(task_of(se));
507 return cfs_rq_is_idle(group_cfs_rq(se));
510 #else /* !CONFIG_FAIR_GROUP_SCHED */
512 #define for_each_sched_entity(se) \
513 for (; se; se = NULL)
515 static inline bool list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
520 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
524 static inline void assert_list_leaf_cfs_rq(struct rq *rq)
528 #define for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) \
529 for (cfs_rq = &rq->cfs, pos = NULL; cfs_rq; cfs_rq = pos)
531 static inline struct sched_entity *parent_entity(struct sched_entity *se)
537 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
541 static inline int tg_is_idle(struct task_group *tg)
546 static int cfs_rq_is_idle(struct cfs_rq *cfs_rq)
551 static int se_is_idle(struct sched_entity *se)
556 #endif /* CONFIG_FAIR_GROUP_SCHED */
558 static __always_inline
559 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec);
561 /**************************************************************
562 * Scheduling class tree data structure manipulation methods:
565 static inline u64 max_vruntime(u64 max_vruntime, u64 vruntime)
567 s64 delta = (s64)(vruntime - max_vruntime);
569 max_vruntime = vruntime;
574 static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime)
576 s64 delta = (s64)(vruntime - min_vruntime);
578 min_vruntime = vruntime;
583 static inline bool entity_before(struct sched_entity *a,
584 struct sched_entity *b)
586 return (s64)(a->vruntime - b->vruntime) < 0;
589 #define __node_2_se(node) \
590 rb_entry((node), struct sched_entity, run_node)
592 static void update_min_vruntime(struct cfs_rq *cfs_rq)
594 struct sched_entity *curr = cfs_rq->curr;
595 struct rb_node *leftmost = rb_first_cached(&cfs_rq->tasks_timeline);
597 u64 vruntime = cfs_rq->min_vruntime;
601 vruntime = curr->vruntime;
606 if (leftmost) { /* non-empty tree */
607 struct sched_entity *se = __node_2_se(leftmost);
610 vruntime = se->vruntime;
612 vruntime = min_vruntime(vruntime, se->vruntime);
615 /* ensure we never gain time by being placed backwards. */
616 u64_u32_store(cfs_rq->min_vruntime,
617 max_vruntime(cfs_rq->min_vruntime, vruntime));
620 static inline bool __entity_less(struct rb_node *a, const struct rb_node *b)
622 return entity_before(__node_2_se(a), __node_2_se(b));
626 * Enqueue an entity into the rb-tree:
628 static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
630 rb_add_cached(&se->run_node, &cfs_rq->tasks_timeline, __entity_less);
633 static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
635 rb_erase_cached(&se->run_node, &cfs_rq->tasks_timeline);
638 struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq)
640 struct rb_node *left = rb_first_cached(&cfs_rq->tasks_timeline);
645 return __node_2_se(left);
648 static struct sched_entity *__pick_next_entity(struct sched_entity *se)
650 struct rb_node *next = rb_next(&se->run_node);
655 return __node_2_se(next);
658 #ifdef CONFIG_SCHED_DEBUG
659 struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
661 struct rb_node *last = rb_last(&cfs_rq->tasks_timeline.rb_root);
666 return __node_2_se(last);
669 /**************************************************************
670 * Scheduling class statistics methods:
673 int sched_update_scaling(void)
675 unsigned int factor = get_update_sysctl_factor();
677 sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency,
678 sysctl_sched_min_granularity);
680 #define WRT_SYSCTL(name) \
681 (normalized_sysctl_##name = sysctl_##name / (factor))
682 WRT_SYSCTL(sched_min_granularity);
683 WRT_SYSCTL(sched_latency);
684 WRT_SYSCTL(sched_wakeup_granularity);
694 static inline u64 calc_delta_fair(u64 delta, struct sched_entity *se)
696 if (unlikely(se->load.weight != NICE_0_LOAD))
697 delta = __calc_delta(delta, NICE_0_LOAD, &se->load);
703 * The idea is to set a period in which each task runs once.
705 * When there are too many tasks (sched_nr_latency) we have to stretch
706 * this period because otherwise the slices get too small.
708 * p = (nr <= nl) ? l : l*nr/nl
710 static u64 __sched_period(unsigned long nr_running)
712 if (unlikely(nr_running > sched_nr_latency))
713 return nr_running * sysctl_sched_min_granularity;
715 return sysctl_sched_latency;
718 static bool sched_idle_cfs_rq(struct cfs_rq *cfs_rq);
721 * We calculate the wall-time slice from the period by taking a part
722 * proportional to the weight.
726 static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
728 unsigned int nr_running = cfs_rq->nr_running;
729 struct sched_entity *init_se = se;
730 unsigned int min_gran;
733 if (sched_feat(ALT_PERIOD))
734 nr_running = rq_of(cfs_rq)->cfs.h_nr_running;
736 slice = __sched_period(nr_running + !se->on_rq);
738 for_each_sched_entity(se) {
739 struct load_weight *load;
740 struct load_weight lw;
741 struct cfs_rq *qcfs_rq;
743 qcfs_rq = cfs_rq_of(se);
744 load = &qcfs_rq->load;
746 if (unlikely(!se->on_rq)) {
749 update_load_add(&lw, se->load.weight);
752 slice = __calc_delta(slice, se->load.weight, load);
755 if (sched_feat(BASE_SLICE)) {
756 if (se_is_idle(init_se) && !sched_idle_cfs_rq(cfs_rq))
757 min_gran = sysctl_sched_idle_min_granularity;
759 min_gran = sysctl_sched_min_granularity;
761 slice = max_t(u64, slice, min_gran);
768 * We calculate the vruntime slice of a to-be-inserted task.
772 static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se)
774 return calc_delta_fair(sched_slice(cfs_rq, se), se);
780 static int select_idle_sibling(struct task_struct *p, int prev_cpu, int cpu);
781 static unsigned long task_h_load(struct task_struct *p);
782 static unsigned long capacity_of(int cpu);
784 /* Give new sched_entity start runnable values to heavy its load in infant time */
785 void init_entity_runnable_average(struct sched_entity *se)
787 struct sched_avg *sa = &se->avg;
789 memset(sa, 0, sizeof(*sa));
792 * Tasks are initialized with full load to be seen as heavy tasks until
793 * they get a chance to stabilize to their real load level.
794 * Group entities are initialized with zero load to reflect the fact that
795 * nothing has been attached to the task group yet.
797 if (entity_is_task(se))
798 sa->load_avg = scale_load_down(se->load.weight);
800 /* when this task enqueue'ed, it will contribute to its cfs_rq's load_avg */
804 * With new tasks being created, their initial util_avgs are extrapolated
805 * based on the cfs_rq's current util_avg:
807 * util_avg = cfs_rq->util_avg / (cfs_rq->load_avg + 1) * se.load.weight
809 * However, in many cases, the above util_avg does not give a desired
810 * value. Moreover, the sum of the util_avgs may be divergent, such
811 * as when the series is a harmonic series.
813 * To solve this problem, we also cap the util_avg of successive tasks to
814 * only 1/2 of the left utilization budget:
816 * util_avg_cap = (cpu_scale - cfs_rq->avg.util_avg) / 2^n
818 * where n denotes the nth task and cpu_scale the CPU capacity.
820 * For example, for a CPU with 1024 of capacity, a simplest series from
821 * the beginning would be like:
823 * task util_avg: 512, 256, 128, 64, 32, 16, 8, ...
824 * cfs_rq util_avg: 512, 768, 896, 960, 992, 1008, 1016, ...
826 * Finally, that extrapolated util_avg is clamped to the cap (util_avg_cap)
827 * if util_avg > util_avg_cap.
829 void post_init_entity_util_avg(struct task_struct *p)
831 struct sched_entity *se = &p->se;
832 struct cfs_rq *cfs_rq = cfs_rq_of(se);
833 struct sched_avg *sa = &se->avg;
834 long cpu_scale = arch_scale_cpu_capacity(cpu_of(rq_of(cfs_rq)));
835 long cap = (long)(cpu_scale - cfs_rq->avg.util_avg) / 2;
837 if (p->sched_class != &fair_sched_class) {
839 * For !fair tasks do:
841 update_cfs_rq_load_avg(now, cfs_rq);
842 attach_entity_load_avg(cfs_rq, se);
843 switched_from_fair(rq, p);
845 * such that the next switched_to_fair() has the
848 se->avg.last_update_time = cfs_rq_clock_pelt(cfs_rq);
853 if (cfs_rq->avg.util_avg != 0) {
854 sa->util_avg = cfs_rq->avg.util_avg * se->load.weight;
855 sa->util_avg /= (cfs_rq->avg.load_avg + 1);
857 if (sa->util_avg > cap)
864 sa->runnable_avg = sa->util_avg;
867 #else /* !CONFIG_SMP */
868 void init_entity_runnable_average(struct sched_entity *se)
871 void post_init_entity_util_avg(struct task_struct *p)
874 static void update_tg_load_avg(struct cfs_rq *cfs_rq)
877 #endif /* CONFIG_SMP */
880 * Update the current task's runtime statistics.
882 static void update_curr(struct cfs_rq *cfs_rq)
884 struct sched_entity *curr = cfs_rq->curr;
885 u64 now = rq_clock_task(rq_of(cfs_rq));
891 delta_exec = now - curr->exec_start;
892 if (unlikely((s64)delta_exec <= 0))
895 curr->exec_start = now;
897 if (schedstat_enabled()) {
898 struct sched_statistics *stats;
900 stats = __schedstats_from_se(curr);
901 __schedstat_set(stats->exec_max,
902 max(delta_exec, stats->exec_max));
905 curr->sum_exec_runtime += delta_exec;
906 schedstat_add(cfs_rq->exec_clock, delta_exec);
908 curr->vruntime += calc_delta_fair(delta_exec, curr);
909 update_min_vruntime(cfs_rq);
911 if (entity_is_task(curr)) {
912 struct task_struct *curtask = task_of(curr);
914 trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime);
915 cgroup_account_cputime(curtask, delta_exec);
916 account_group_exec_runtime(curtask, delta_exec);
919 account_cfs_rq_runtime(cfs_rq, delta_exec);
922 static void update_curr_fair(struct rq *rq)
924 update_curr(cfs_rq_of(&rq->curr->se));
928 update_stats_wait_start_fair(struct cfs_rq *cfs_rq, struct sched_entity *se)
930 struct sched_statistics *stats;
931 struct task_struct *p = NULL;
933 if (!schedstat_enabled())
936 stats = __schedstats_from_se(se);
938 if (entity_is_task(se))
941 __update_stats_wait_start(rq_of(cfs_rq), p, stats);
945 update_stats_wait_end_fair(struct cfs_rq *cfs_rq, struct sched_entity *se)
947 struct sched_statistics *stats;
948 struct task_struct *p = NULL;
950 if (!schedstat_enabled())
953 stats = __schedstats_from_se(se);
956 * When the sched_schedstat changes from 0 to 1, some sched se
957 * maybe already in the runqueue, the se->statistics.wait_start
958 * will be 0.So it will let the delta wrong. We need to avoid this
961 if (unlikely(!schedstat_val(stats->wait_start)))
964 if (entity_is_task(se))
967 __update_stats_wait_end(rq_of(cfs_rq), p, stats);
971 update_stats_enqueue_sleeper_fair(struct cfs_rq *cfs_rq, struct sched_entity *se)
973 struct sched_statistics *stats;
974 struct task_struct *tsk = NULL;
976 if (!schedstat_enabled())
979 stats = __schedstats_from_se(se);
981 if (entity_is_task(se))
984 __update_stats_enqueue_sleeper(rq_of(cfs_rq), tsk, stats);
988 * Task is being enqueued - update stats:
991 update_stats_enqueue_fair(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
993 if (!schedstat_enabled())
997 * Are we enqueueing a waiting task? (for current tasks
998 * a dequeue/enqueue event is a NOP)
1000 if (se != cfs_rq->curr)
1001 update_stats_wait_start_fair(cfs_rq, se);
1003 if (flags & ENQUEUE_WAKEUP)
1004 update_stats_enqueue_sleeper_fair(cfs_rq, se);
1008 update_stats_dequeue_fair(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
1011 if (!schedstat_enabled())
1015 * Mark the end of the wait period if dequeueing a
1018 if (se != cfs_rq->curr)
1019 update_stats_wait_end_fair(cfs_rq, se);
1021 if ((flags & DEQUEUE_SLEEP) && entity_is_task(se)) {
1022 struct task_struct *tsk = task_of(se);
1025 /* XXX racy against TTWU */
1026 state = READ_ONCE(tsk->__state);
1027 if (state & TASK_INTERRUPTIBLE)
1028 __schedstat_set(tsk->stats.sleep_start,
1029 rq_clock(rq_of(cfs_rq)));
1030 if (state & TASK_UNINTERRUPTIBLE)
1031 __schedstat_set(tsk->stats.block_start,
1032 rq_clock(rq_of(cfs_rq)));
1037 * We are picking a new current task - update its stats:
1040 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
1043 * We are starting a new run period:
1045 se->exec_start = rq_clock_task(rq_of(cfs_rq));
1048 /**************************************************
1049 * Scheduling class queueing methods:
1053 #define NUMA_IMBALANCE_MIN 2
1056 adjust_numa_imbalance(int imbalance, int dst_running, int imb_numa_nr)
1059 * Allow a NUMA imbalance if busy CPUs is less than the maximum
1060 * threshold. Above this threshold, individual tasks may be contending
1061 * for both memory bandwidth and any shared HT resources. This is an
1062 * approximation as the number of running tasks may not be related to
1063 * the number of busy CPUs due to sched_setaffinity.
1065 if (dst_running > imb_numa_nr)
1069 * Allow a small imbalance based on a simple pair of communicating
1070 * tasks that remain local when the destination is lightly loaded.
1072 if (imbalance <= NUMA_IMBALANCE_MIN)
1077 #endif /* CONFIG_NUMA */
1079 #ifdef CONFIG_NUMA_BALANCING
1081 * Approximate time to scan a full NUMA task in ms. The task scan period is
1082 * calculated based on the tasks virtual memory size and
1083 * numa_balancing_scan_size.
1085 unsigned int sysctl_numa_balancing_scan_period_min = 1000;
1086 unsigned int sysctl_numa_balancing_scan_period_max = 60000;
1088 /* Portion of address space to scan in MB */
1089 unsigned int sysctl_numa_balancing_scan_size = 256;
1091 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
1092 unsigned int sysctl_numa_balancing_scan_delay = 1000;
1094 /* The page with hint page fault latency < threshold in ms is considered hot */
1095 unsigned int sysctl_numa_balancing_hot_threshold = MSEC_PER_SEC;
1097 /* Restrict the NUMA promotion throughput (MB/s) for each target node. */
1098 unsigned int sysctl_numa_balancing_promote_rate_limit = 65536;
1101 refcount_t refcount;
1103 spinlock_t lock; /* nr_tasks, tasks */
1108 struct rcu_head rcu;
1109 unsigned long total_faults;
1110 unsigned long max_faults_cpu;
1112 * faults[] array is split into two regions: faults_mem and faults_cpu.
1114 * Faults_cpu is used to decide whether memory should move
1115 * towards the CPU. As a consequence, these stats are weighted
1116 * more by CPU use than by memory faults.
1118 unsigned long faults[];
1122 * For functions that can be called in multiple contexts that permit reading
1123 * ->numa_group (see struct task_struct for locking rules).
1125 static struct numa_group *deref_task_numa_group(struct task_struct *p)
1127 return rcu_dereference_check(p->numa_group, p == current ||
1128 (lockdep_is_held(__rq_lockp(task_rq(p))) && !READ_ONCE(p->on_cpu)));
1131 static struct numa_group *deref_curr_numa_group(struct task_struct *p)
1133 return rcu_dereference_protected(p->numa_group, p == current);
1136 static inline unsigned long group_faults_priv(struct numa_group *ng);
1137 static inline unsigned long group_faults_shared(struct numa_group *ng);
1139 static unsigned int task_nr_scan_windows(struct task_struct *p)
1141 unsigned long rss = 0;
1142 unsigned long nr_scan_pages;
1145 * Calculations based on RSS as non-present and empty pages are skipped
1146 * by the PTE scanner and NUMA hinting faults should be trapped based
1149 nr_scan_pages = sysctl_numa_balancing_scan_size << (20 - PAGE_SHIFT);
1150 rss = get_mm_rss(p->mm);
1152 rss = nr_scan_pages;
1154 rss = round_up(rss, nr_scan_pages);
1155 return rss / nr_scan_pages;
1158 /* For sanity's sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */
1159 #define MAX_SCAN_WINDOW 2560
1161 static unsigned int task_scan_min(struct task_struct *p)
1163 unsigned int scan_size = READ_ONCE(sysctl_numa_balancing_scan_size);
1164 unsigned int scan, floor;
1165 unsigned int windows = 1;
1167 if (scan_size < MAX_SCAN_WINDOW)
1168 windows = MAX_SCAN_WINDOW / scan_size;
1169 floor = 1000 / windows;
1171 scan = sysctl_numa_balancing_scan_period_min / task_nr_scan_windows(p);
1172 return max_t(unsigned int, floor, scan);
1175 static unsigned int task_scan_start(struct task_struct *p)
1177 unsigned long smin = task_scan_min(p);
1178 unsigned long period = smin;
1179 struct numa_group *ng;
1181 /* Scale the maximum scan period with the amount of shared memory. */
1183 ng = rcu_dereference(p->numa_group);
1185 unsigned long shared = group_faults_shared(ng);
1186 unsigned long private = group_faults_priv(ng);
1188 period *= refcount_read(&ng->refcount);
1189 period *= shared + 1;
1190 period /= private + shared + 1;
1194 return max(smin, period);
1197 static unsigned int task_scan_max(struct task_struct *p)
1199 unsigned long smin = task_scan_min(p);
1201 struct numa_group *ng;
1203 /* Watch for min being lower than max due to floor calculations */
1204 smax = sysctl_numa_balancing_scan_period_max / task_nr_scan_windows(p);
1206 /* Scale the maximum scan period with the amount of shared memory. */
1207 ng = deref_curr_numa_group(p);
1209 unsigned long shared = group_faults_shared(ng);
1210 unsigned long private = group_faults_priv(ng);
1211 unsigned long period = smax;
1213 period *= refcount_read(&ng->refcount);
1214 period *= shared + 1;
1215 period /= private + shared + 1;
1217 smax = max(smax, period);
1220 return max(smin, smax);
1223 static void account_numa_enqueue(struct rq *rq, struct task_struct *p)
1225 rq->nr_numa_running += (p->numa_preferred_nid != NUMA_NO_NODE);
1226 rq->nr_preferred_running += (p->numa_preferred_nid == task_node(p));
1229 static void account_numa_dequeue(struct rq *rq, struct task_struct *p)
1231 rq->nr_numa_running -= (p->numa_preferred_nid != NUMA_NO_NODE);
1232 rq->nr_preferred_running -= (p->numa_preferred_nid == task_node(p));
1235 /* Shared or private faults. */
1236 #define NR_NUMA_HINT_FAULT_TYPES 2
1238 /* Memory and CPU locality */
1239 #define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2)
1241 /* Averaged statistics, and temporary buffers. */
1242 #define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2)
1244 pid_t task_numa_group_id(struct task_struct *p)
1246 struct numa_group *ng;
1250 ng = rcu_dereference(p->numa_group);
1259 * The averaged statistics, shared & private, memory & CPU,
1260 * occupy the first half of the array. The second half of the
1261 * array is for current counters, which are averaged into the
1262 * first set by task_numa_placement.
1264 static inline int task_faults_idx(enum numa_faults_stats s, int nid, int priv)
1266 return NR_NUMA_HINT_FAULT_TYPES * (s * nr_node_ids + nid) + priv;
1269 static inline unsigned long task_faults(struct task_struct *p, int nid)
1271 if (!p->numa_faults)
1274 return p->numa_faults[task_faults_idx(NUMA_MEM, nid, 0)] +
1275 p->numa_faults[task_faults_idx(NUMA_MEM, nid, 1)];
1278 static inline unsigned long group_faults(struct task_struct *p, int nid)
1280 struct numa_group *ng = deref_task_numa_group(p);
1285 return ng->faults[task_faults_idx(NUMA_MEM, nid, 0)] +
1286 ng->faults[task_faults_idx(NUMA_MEM, nid, 1)];
1289 static inline unsigned long group_faults_cpu(struct numa_group *group, int nid)
1291 return group->faults[task_faults_idx(NUMA_CPU, nid, 0)] +
1292 group->faults[task_faults_idx(NUMA_CPU, nid, 1)];
1295 static inline unsigned long group_faults_priv(struct numa_group *ng)
1297 unsigned long faults = 0;
1300 for_each_online_node(node) {
1301 faults += ng->faults[task_faults_idx(NUMA_MEM, node, 1)];
1307 static inline unsigned long group_faults_shared(struct numa_group *ng)
1309 unsigned long faults = 0;
1312 for_each_online_node(node) {
1313 faults += ng->faults[task_faults_idx(NUMA_MEM, node, 0)];
1320 * A node triggering more than 1/3 as many NUMA faults as the maximum is
1321 * considered part of a numa group's pseudo-interleaving set. Migrations
1322 * between these nodes are slowed down, to allow things to settle down.
1324 #define ACTIVE_NODE_FRACTION 3
1326 static bool numa_is_active_node(int nid, struct numa_group *ng)
1328 return group_faults_cpu(ng, nid) * ACTIVE_NODE_FRACTION > ng->max_faults_cpu;
1331 /* Handle placement on systems where not all nodes are directly connected. */
1332 static unsigned long score_nearby_nodes(struct task_struct *p, int nid,
1333 int lim_dist, bool task)
1335 unsigned long score = 0;
1339 * All nodes are directly connected, and the same distance
1340 * from each other. No need for fancy placement algorithms.
1342 if (sched_numa_topology_type == NUMA_DIRECT)
1345 /* sched_max_numa_distance may be changed in parallel. */
1346 max_dist = READ_ONCE(sched_max_numa_distance);
1348 * This code is called for each node, introducing N^2 complexity,
1349 * which should be ok given the number of nodes rarely exceeds 8.
1351 for_each_online_node(node) {
1352 unsigned long faults;
1353 int dist = node_distance(nid, node);
1356 * The furthest away nodes in the system are not interesting
1357 * for placement; nid was already counted.
1359 if (dist >= max_dist || node == nid)
1363 * On systems with a backplane NUMA topology, compare groups
1364 * of nodes, and move tasks towards the group with the most
1365 * memory accesses. When comparing two nodes at distance
1366 * "hoplimit", only nodes closer by than "hoplimit" are part
1367 * of each group. Skip other nodes.
1369 if (sched_numa_topology_type == NUMA_BACKPLANE && dist >= lim_dist)
1372 /* Add up the faults from nearby nodes. */
1374 faults = task_faults(p, node);
1376 faults = group_faults(p, node);
1379 * On systems with a glueless mesh NUMA topology, there are
1380 * no fixed "groups of nodes". Instead, nodes that are not
1381 * directly connected bounce traffic through intermediate
1382 * nodes; a numa_group can occupy any set of nodes.
1383 * The further away a node is, the less the faults count.
1384 * This seems to result in good task placement.
1386 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
1387 faults *= (max_dist - dist);
1388 faults /= (max_dist - LOCAL_DISTANCE);
1398 * These return the fraction of accesses done by a particular task, or
1399 * task group, on a particular numa node. The group weight is given a
1400 * larger multiplier, in order to group tasks together that are almost
1401 * evenly spread out between numa nodes.
1403 static inline unsigned long task_weight(struct task_struct *p, int nid,
1406 unsigned long faults, total_faults;
1408 if (!p->numa_faults)
1411 total_faults = p->total_numa_faults;
1416 faults = task_faults(p, nid);
1417 faults += score_nearby_nodes(p, nid, dist, true);
1419 return 1000 * faults / total_faults;
1422 static inline unsigned long group_weight(struct task_struct *p, int nid,
1425 struct numa_group *ng = deref_task_numa_group(p);
1426 unsigned long faults, total_faults;
1431 total_faults = ng->total_faults;
1436 faults = group_faults(p, nid);
1437 faults += score_nearby_nodes(p, nid, dist, false);
1439 return 1000 * faults / total_faults;
1443 * If memory tiering mode is enabled, cpupid of slow memory page is
1444 * used to record scan time instead of CPU and PID. When tiering mode
1445 * is disabled at run time, the scan time (in cpupid) will be
1446 * interpreted as CPU and PID. So CPU needs to be checked to avoid to
1447 * access out of array bound.
1449 static inline bool cpupid_valid(int cpupid)
1451 return cpupid_to_cpu(cpupid) < nr_cpu_ids;
1455 * For memory tiering mode, if there are enough free pages (more than
1456 * enough watermark defined here) in fast memory node, to take full
1457 * advantage of fast memory capacity, all recently accessed slow
1458 * memory pages will be migrated to fast memory node without
1459 * considering hot threshold.
1461 static bool pgdat_free_space_enough(struct pglist_data *pgdat)
1464 unsigned long enough_wmark;
1466 enough_wmark = max(1UL * 1024 * 1024 * 1024 >> PAGE_SHIFT,
1467 pgdat->node_present_pages >> 4);
1468 for (z = pgdat->nr_zones - 1; z >= 0; z--) {
1469 struct zone *zone = pgdat->node_zones + z;
1471 if (!populated_zone(zone))
1474 if (zone_watermark_ok(zone, 0,
1475 wmark_pages(zone, WMARK_PROMO) + enough_wmark,
1483 * For memory tiering mode, when page tables are scanned, the scan
1484 * time will be recorded in struct page in addition to make page
1485 * PROT_NONE for slow memory page. So when the page is accessed, in
1486 * hint page fault handler, the hint page fault latency is calculated
1489 * hint page fault latency = hint page fault time - scan time
1491 * The smaller the hint page fault latency, the higher the possibility
1492 * for the page to be hot.
1494 static int numa_hint_fault_latency(struct page *page)
1496 int last_time, time;
1498 time = jiffies_to_msecs(jiffies);
1499 last_time = xchg_page_access_time(page, time);
1501 return (time - last_time) & PAGE_ACCESS_TIME_MASK;
1505 * For memory tiering mode, too high promotion/demotion throughput may
1506 * hurt application latency. So we provide a mechanism to rate limit
1507 * the number of pages that are tried to be promoted.
1509 static bool numa_promotion_rate_limit(struct pglist_data *pgdat,
1510 unsigned long rate_limit, int nr)
1512 unsigned long nr_cand;
1513 unsigned int now, start;
1515 now = jiffies_to_msecs(jiffies);
1516 mod_node_page_state(pgdat, PGPROMOTE_CANDIDATE, nr);
1517 nr_cand = node_page_state(pgdat, PGPROMOTE_CANDIDATE);
1518 start = pgdat->nbp_rl_start;
1519 if (now - start > MSEC_PER_SEC &&
1520 cmpxchg(&pgdat->nbp_rl_start, start, now) == start)
1521 pgdat->nbp_rl_nr_cand = nr_cand;
1522 if (nr_cand - pgdat->nbp_rl_nr_cand >= rate_limit)
1527 #define NUMA_MIGRATION_ADJUST_STEPS 16
1529 static void numa_promotion_adjust_threshold(struct pglist_data *pgdat,
1530 unsigned long rate_limit,
1531 unsigned int ref_th)
1533 unsigned int now, start, th_period, unit_th, th;
1534 unsigned long nr_cand, ref_cand, diff_cand;
1536 now = jiffies_to_msecs(jiffies);
1537 th_period = sysctl_numa_balancing_scan_period_max;
1538 start = pgdat->nbp_th_start;
1539 if (now - start > th_period &&
1540 cmpxchg(&pgdat->nbp_th_start, start, now) == start) {
1541 ref_cand = rate_limit *
1542 sysctl_numa_balancing_scan_period_max / MSEC_PER_SEC;
1543 nr_cand = node_page_state(pgdat, PGPROMOTE_CANDIDATE);
1544 diff_cand = nr_cand - pgdat->nbp_th_nr_cand;
1545 unit_th = ref_th * 2 / NUMA_MIGRATION_ADJUST_STEPS;
1546 th = pgdat->nbp_threshold ? : ref_th;
1547 if (diff_cand > ref_cand * 11 / 10)
1548 th = max(th - unit_th, unit_th);
1549 else if (diff_cand < ref_cand * 9 / 10)
1550 th = min(th + unit_th, ref_th * 2);
1551 pgdat->nbp_th_nr_cand = nr_cand;
1552 pgdat->nbp_threshold = th;
1556 bool should_numa_migrate_memory(struct task_struct *p, struct page * page,
1557 int src_nid, int dst_cpu)
1559 struct numa_group *ng = deref_curr_numa_group(p);
1560 int dst_nid = cpu_to_node(dst_cpu);
1561 int last_cpupid, this_cpupid;
1564 * The pages in slow memory node should be migrated according
1565 * to hot/cold instead of private/shared.
1567 if (sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING &&
1568 !node_is_toptier(src_nid)) {
1569 struct pglist_data *pgdat;
1570 unsigned long rate_limit;
1571 unsigned int latency, th, def_th;
1573 pgdat = NODE_DATA(dst_nid);
1574 if (pgdat_free_space_enough(pgdat)) {
1575 /* workload changed, reset hot threshold */
1576 pgdat->nbp_threshold = 0;
1580 def_th = sysctl_numa_balancing_hot_threshold;
1581 rate_limit = sysctl_numa_balancing_promote_rate_limit << \
1583 numa_promotion_adjust_threshold(pgdat, rate_limit, def_th);
1585 th = pgdat->nbp_threshold ? : def_th;
1586 latency = numa_hint_fault_latency(page);
1590 return !numa_promotion_rate_limit(pgdat, rate_limit,
1591 thp_nr_pages(page));
1594 this_cpupid = cpu_pid_to_cpupid(dst_cpu, current->pid);
1595 last_cpupid = page_cpupid_xchg_last(page, this_cpupid);
1597 if (!(sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING) &&
1598 !node_is_toptier(src_nid) && !cpupid_valid(last_cpupid))
1602 * Allow first faults or private faults to migrate immediately early in
1603 * the lifetime of a task. The magic number 4 is based on waiting for
1604 * two full passes of the "multi-stage node selection" test that is
1607 if ((p->numa_preferred_nid == NUMA_NO_NODE || p->numa_scan_seq <= 4) &&
1608 (cpupid_pid_unset(last_cpupid) || cpupid_match_pid(p, last_cpupid)))
1612 * Multi-stage node selection is used in conjunction with a periodic
1613 * migration fault to build a temporal task<->page relation. By using
1614 * a two-stage filter we remove short/unlikely relations.
1616 * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate
1617 * a task's usage of a particular page (n_p) per total usage of this
1618 * page (n_t) (in a given time-span) to a probability.
1620 * Our periodic faults will sample this probability and getting the
1621 * same result twice in a row, given these samples are fully
1622 * independent, is then given by P(n)^2, provided our sample period
1623 * is sufficiently short compared to the usage pattern.
1625 * This quadric squishes small probabilities, making it less likely we
1626 * act on an unlikely task<->page relation.
1628 if (!cpupid_pid_unset(last_cpupid) &&
1629 cpupid_to_nid(last_cpupid) != dst_nid)
1632 /* Always allow migrate on private faults */
1633 if (cpupid_match_pid(p, last_cpupid))
1636 /* A shared fault, but p->numa_group has not been set up yet. */
1641 * Destination node is much more heavily used than the source
1642 * node? Allow migration.
1644 if (group_faults_cpu(ng, dst_nid) > group_faults_cpu(ng, src_nid) *
1645 ACTIVE_NODE_FRACTION)
1649 * Distribute memory according to CPU & memory use on each node,
1650 * with 3/4 hysteresis to avoid unnecessary memory migrations:
1652 * faults_cpu(dst) 3 faults_cpu(src)
1653 * --------------- * - > ---------------
1654 * faults_mem(dst) 4 faults_mem(src)
1656 return group_faults_cpu(ng, dst_nid) * group_faults(p, src_nid) * 3 >
1657 group_faults_cpu(ng, src_nid) * group_faults(p, dst_nid) * 4;
1661 * 'numa_type' describes the node at the moment of load balancing.
1664 /* The node has spare capacity that can be used to run more tasks. */
1667 * The node is fully used and the tasks don't compete for more CPU
1668 * cycles. Nevertheless, some tasks might wait before running.
1672 * The node is overloaded and can't provide expected CPU cycles to all
1678 /* Cached statistics for all CPUs within a node */
1681 unsigned long runnable;
1683 /* Total compute capacity of CPUs on a node */
1684 unsigned long compute_capacity;
1685 unsigned int nr_running;
1686 unsigned int weight;
1687 enum numa_type node_type;
1691 static inline bool is_core_idle(int cpu)
1693 #ifdef CONFIG_SCHED_SMT
1696 for_each_cpu(sibling, cpu_smt_mask(cpu)) {
1700 if (!idle_cpu(sibling))
1708 struct task_numa_env {
1709 struct task_struct *p;
1711 int src_cpu, src_nid;
1712 int dst_cpu, dst_nid;
1715 struct numa_stats src_stats, dst_stats;
1720 struct task_struct *best_task;
1725 static unsigned long cpu_load(struct rq *rq);
1726 static unsigned long cpu_runnable(struct rq *rq);
1729 numa_type numa_classify(unsigned int imbalance_pct,
1730 struct numa_stats *ns)
1732 if ((ns->nr_running > ns->weight) &&
1733 (((ns->compute_capacity * 100) < (ns->util * imbalance_pct)) ||
1734 ((ns->compute_capacity * imbalance_pct) < (ns->runnable * 100))))
1735 return node_overloaded;
1737 if ((ns->nr_running < ns->weight) ||
1738 (((ns->compute_capacity * 100) > (ns->util * imbalance_pct)) &&
1739 ((ns->compute_capacity * imbalance_pct) > (ns->runnable * 100))))
1740 return node_has_spare;
1742 return node_fully_busy;
1745 #ifdef CONFIG_SCHED_SMT
1746 /* Forward declarations of select_idle_sibling helpers */
1747 static inline bool test_idle_cores(int cpu);
1748 static inline int numa_idle_core(int idle_core, int cpu)
1750 if (!static_branch_likely(&sched_smt_present) ||
1751 idle_core >= 0 || !test_idle_cores(cpu))
1755 * Prefer cores instead of packing HT siblings
1756 * and triggering future load balancing.
1758 if (is_core_idle(cpu))
1764 static inline int numa_idle_core(int idle_core, int cpu)
1771 * Gather all necessary information to make NUMA balancing placement
1772 * decisions that are compatible with standard load balancer. This
1773 * borrows code and logic from update_sg_lb_stats but sharing a
1774 * common implementation is impractical.
1776 static void update_numa_stats(struct task_numa_env *env,
1777 struct numa_stats *ns, int nid,
1780 int cpu, idle_core = -1;
1782 memset(ns, 0, sizeof(*ns));
1786 for_each_cpu(cpu, cpumask_of_node(nid)) {
1787 struct rq *rq = cpu_rq(cpu);
1789 ns->load += cpu_load(rq);
1790 ns->runnable += cpu_runnable(rq);
1791 ns->util += cpu_util_cfs(cpu);
1792 ns->nr_running += rq->cfs.h_nr_running;
1793 ns->compute_capacity += capacity_of(cpu);
1795 if (find_idle && !rq->nr_running && idle_cpu(cpu)) {
1796 if (READ_ONCE(rq->numa_migrate_on) ||
1797 !cpumask_test_cpu(cpu, env->p->cpus_ptr))
1800 if (ns->idle_cpu == -1)
1803 idle_core = numa_idle_core(idle_core, cpu);
1808 ns->weight = cpumask_weight(cpumask_of_node(nid));
1810 ns->node_type = numa_classify(env->imbalance_pct, ns);
1813 ns->idle_cpu = idle_core;
1816 static void task_numa_assign(struct task_numa_env *env,
1817 struct task_struct *p, long imp)
1819 struct rq *rq = cpu_rq(env->dst_cpu);
1821 /* Check if run-queue part of active NUMA balance. */
1822 if (env->best_cpu != env->dst_cpu && xchg(&rq->numa_migrate_on, 1)) {
1824 int start = env->dst_cpu;
1826 /* Find alternative idle CPU. */
1827 for_each_cpu_wrap(cpu, cpumask_of_node(env->dst_nid), start) {
1828 if (cpu == env->best_cpu || !idle_cpu(cpu) ||
1829 !cpumask_test_cpu(cpu, env->p->cpus_ptr)) {
1834 rq = cpu_rq(env->dst_cpu);
1835 if (!xchg(&rq->numa_migrate_on, 1))
1839 /* Failed to find an alternative idle CPU */
1845 * Clear previous best_cpu/rq numa-migrate flag, since task now
1846 * found a better CPU to move/swap.
1848 if (env->best_cpu != -1 && env->best_cpu != env->dst_cpu) {
1849 rq = cpu_rq(env->best_cpu);
1850 WRITE_ONCE(rq->numa_migrate_on, 0);
1854 put_task_struct(env->best_task);
1859 env->best_imp = imp;
1860 env->best_cpu = env->dst_cpu;
1863 static bool load_too_imbalanced(long src_load, long dst_load,
1864 struct task_numa_env *env)
1867 long orig_src_load, orig_dst_load;
1868 long src_capacity, dst_capacity;
1871 * The load is corrected for the CPU capacity available on each node.
1874 * ------------ vs ---------
1875 * src_capacity dst_capacity
1877 src_capacity = env->src_stats.compute_capacity;
1878 dst_capacity = env->dst_stats.compute_capacity;
1880 imb = abs(dst_load * src_capacity - src_load * dst_capacity);
1882 orig_src_load = env->src_stats.load;
1883 orig_dst_load = env->dst_stats.load;
1885 old_imb = abs(orig_dst_load * src_capacity - orig_src_load * dst_capacity);
1887 /* Would this change make things worse? */
1888 return (imb > old_imb);
1892 * Maximum NUMA importance can be 1998 (2*999);
1893 * SMALLIMP @ 30 would be close to 1998/64.
1894 * Used to deter task migration.
1899 * This checks if the overall compute and NUMA accesses of the system would
1900 * be improved if the source tasks was migrated to the target dst_cpu taking
1901 * into account that it might be best if task running on the dst_cpu should
1902 * be exchanged with the source task
1904 static bool task_numa_compare(struct task_numa_env *env,
1905 long taskimp, long groupimp, bool maymove)
1907 struct numa_group *cur_ng, *p_ng = deref_curr_numa_group(env->p);
1908 struct rq *dst_rq = cpu_rq(env->dst_cpu);
1909 long imp = p_ng ? groupimp : taskimp;
1910 struct task_struct *cur;
1911 long src_load, dst_load;
1912 int dist = env->dist;
1915 bool stopsearch = false;
1917 if (READ_ONCE(dst_rq->numa_migrate_on))
1921 cur = rcu_dereference(dst_rq->curr);
1922 if (cur && ((cur->flags & PF_EXITING) || is_idle_task(cur)))
1926 * Because we have preemption enabled we can get migrated around and
1927 * end try selecting ourselves (current == env->p) as a swap candidate.
1929 if (cur == env->p) {
1935 if (maymove && moveimp >= env->best_imp)
1941 /* Skip this swap candidate if cannot move to the source cpu. */
1942 if (!cpumask_test_cpu(env->src_cpu, cur->cpus_ptr))
1946 * Skip this swap candidate if it is not moving to its preferred
1947 * node and the best task is.
1949 if (env->best_task &&
1950 env->best_task->numa_preferred_nid == env->src_nid &&
1951 cur->numa_preferred_nid != env->src_nid) {
1956 * "imp" is the fault differential for the source task between the
1957 * source and destination node. Calculate the total differential for
1958 * the source task and potential destination task. The more negative
1959 * the value is, the more remote accesses that would be expected to
1960 * be incurred if the tasks were swapped.
1962 * If dst and source tasks are in the same NUMA group, or not
1963 * in any group then look only at task weights.
1965 cur_ng = rcu_dereference(cur->numa_group);
1966 if (cur_ng == p_ng) {
1968 * Do not swap within a group or between tasks that have
1969 * no group if there is spare capacity. Swapping does
1970 * not address the load imbalance and helps one task at
1971 * the cost of punishing another.
1973 if (env->dst_stats.node_type == node_has_spare)
1976 imp = taskimp + task_weight(cur, env->src_nid, dist) -
1977 task_weight(cur, env->dst_nid, dist);
1979 * Add some hysteresis to prevent swapping the
1980 * tasks within a group over tiny differences.
1986 * Compare the group weights. If a task is all by itself
1987 * (not part of a group), use the task weight instead.
1990 imp += group_weight(cur, env->src_nid, dist) -
1991 group_weight(cur, env->dst_nid, dist);
1993 imp += task_weight(cur, env->src_nid, dist) -
1994 task_weight(cur, env->dst_nid, dist);
1997 /* Discourage picking a task already on its preferred node */
1998 if (cur->numa_preferred_nid == env->dst_nid)
2002 * Encourage picking a task that moves to its preferred node.
2003 * This potentially makes imp larger than it's maximum of
2004 * 1998 (see SMALLIMP and task_weight for why) but in this
2005 * case, it does not matter.
2007 if (cur->numa_preferred_nid == env->src_nid)
2010 if (maymove && moveimp > imp && moveimp > env->best_imp) {
2017 * Prefer swapping with a task moving to its preferred node over a
2020 if (env->best_task && cur->numa_preferred_nid == env->src_nid &&
2021 env->best_task->numa_preferred_nid != env->src_nid) {
2026 * If the NUMA importance is less than SMALLIMP,
2027 * task migration might only result in ping pong
2028 * of tasks and also hurt performance due to cache
2031 if (imp < SMALLIMP || imp <= env->best_imp + SMALLIMP / 2)
2035 * In the overloaded case, try and keep the load balanced.
2037 load = task_h_load(env->p) - task_h_load(cur);
2041 dst_load = env->dst_stats.load + load;
2042 src_load = env->src_stats.load - load;
2044 if (load_too_imbalanced(src_load, dst_load, env))
2048 /* Evaluate an idle CPU for a task numa move. */
2050 int cpu = env->dst_stats.idle_cpu;
2052 /* Nothing cached so current CPU went idle since the search. */
2057 * If the CPU is no longer truly idle and the previous best CPU
2058 * is, keep using it.
2060 if (!idle_cpu(cpu) && env->best_cpu >= 0 &&
2061 idle_cpu(env->best_cpu)) {
2062 cpu = env->best_cpu;
2068 task_numa_assign(env, cur, imp);
2071 * If a move to idle is allowed because there is capacity or load
2072 * balance improves then stop the search. While a better swap
2073 * candidate may exist, a search is not free.
2075 if (maymove && !cur && env->best_cpu >= 0 && idle_cpu(env->best_cpu))
2079 * If a swap candidate must be identified and the current best task
2080 * moves its preferred node then stop the search.
2082 if (!maymove && env->best_task &&
2083 env->best_task->numa_preferred_nid == env->src_nid) {
2092 static void task_numa_find_cpu(struct task_numa_env *env,
2093 long taskimp, long groupimp)
2095 bool maymove = false;
2099 * If dst node has spare capacity, then check if there is an
2100 * imbalance that would be overruled by the load balancer.
2102 if (env->dst_stats.node_type == node_has_spare) {
2103 unsigned int imbalance;
2104 int src_running, dst_running;
2107 * Would movement cause an imbalance? Note that if src has
2108 * more running tasks that the imbalance is ignored as the
2109 * move improves the imbalance from the perspective of the
2110 * CPU load balancer.
2112 src_running = env->src_stats.nr_running - 1;
2113 dst_running = env->dst_stats.nr_running + 1;
2114 imbalance = max(0, dst_running - src_running);
2115 imbalance = adjust_numa_imbalance(imbalance, dst_running,
2118 /* Use idle CPU if there is no imbalance */
2121 if (env->dst_stats.idle_cpu >= 0) {
2122 env->dst_cpu = env->dst_stats.idle_cpu;
2123 task_numa_assign(env, NULL, 0);
2128 long src_load, dst_load, load;
2130 * If the improvement from just moving env->p direction is better
2131 * than swapping tasks around, check if a move is possible.
2133 load = task_h_load(env->p);
2134 dst_load = env->dst_stats.load + load;
2135 src_load = env->src_stats.load - load;
2136 maymove = !load_too_imbalanced(src_load, dst_load, env);
2139 for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) {
2140 /* Skip this CPU if the source task cannot migrate */
2141 if (!cpumask_test_cpu(cpu, env->p->cpus_ptr))
2145 if (task_numa_compare(env, taskimp, groupimp, maymove))
2150 static int task_numa_migrate(struct task_struct *p)
2152 struct task_numa_env env = {
2155 .src_cpu = task_cpu(p),
2156 .src_nid = task_node(p),
2158 .imbalance_pct = 112,
2164 unsigned long taskweight, groupweight;
2165 struct sched_domain *sd;
2166 long taskimp, groupimp;
2167 struct numa_group *ng;
2172 * Pick the lowest SD_NUMA domain, as that would have the smallest
2173 * imbalance and would be the first to start moving tasks about.
2175 * And we want to avoid any moving of tasks about, as that would create
2176 * random movement of tasks -- counter the numa conditions we're trying
2180 sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu));
2182 env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2;
2183 env.imb_numa_nr = sd->imb_numa_nr;
2188 * Cpusets can break the scheduler domain tree into smaller
2189 * balance domains, some of which do not cross NUMA boundaries.
2190 * Tasks that are "trapped" in such domains cannot be migrated
2191 * elsewhere, so there is no point in (re)trying.
2193 if (unlikely(!sd)) {
2194 sched_setnuma(p, task_node(p));
2198 env.dst_nid = p->numa_preferred_nid;
2199 dist = env.dist = node_distance(env.src_nid, env.dst_nid);
2200 taskweight = task_weight(p, env.src_nid, dist);
2201 groupweight = group_weight(p, env.src_nid, dist);
2202 update_numa_stats(&env, &env.src_stats, env.src_nid, false);
2203 taskimp = task_weight(p, env.dst_nid, dist) - taskweight;
2204 groupimp = group_weight(p, env.dst_nid, dist) - groupweight;
2205 update_numa_stats(&env, &env.dst_stats, env.dst_nid, true);
2207 /* Try to find a spot on the preferred nid. */
2208 task_numa_find_cpu(&env, taskimp, groupimp);
2211 * Look at other nodes in these cases:
2212 * - there is no space available on the preferred_nid
2213 * - the task is part of a numa_group that is interleaved across
2214 * multiple NUMA nodes; in order to better consolidate the group,
2215 * we need to check other locations.
2217 ng = deref_curr_numa_group(p);
2218 if (env.best_cpu == -1 || (ng && ng->active_nodes > 1)) {
2219 for_each_node_state(nid, N_CPU) {
2220 if (nid == env.src_nid || nid == p->numa_preferred_nid)
2223 dist = node_distance(env.src_nid, env.dst_nid);
2224 if (sched_numa_topology_type == NUMA_BACKPLANE &&
2226 taskweight = task_weight(p, env.src_nid, dist);
2227 groupweight = group_weight(p, env.src_nid, dist);
2230 /* Only consider nodes where both task and groups benefit */
2231 taskimp = task_weight(p, nid, dist) - taskweight;
2232 groupimp = group_weight(p, nid, dist) - groupweight;
2233 if (taskimp < 0 && groupimp < 0)
2238 update_numa_stats(&env, &env.dst_stats, env.dst_nid, true);
2239 task_numa_find_cpu(&env, taskimp, groupimp);
2244 * If the task is part of a workload that spans multiple NUMA nodes,
2245 * and is migrating into one of the workload's active nodes, remember
2246 * this node as the task's preferred numa node, so the workload can
2248 * A task that migrated to a second choice node will be better off
2249 * trying for a better one later. Do not set the preferred node here.
2252 if (env.best_cpu == -1)
2255 nid = cpu_to_node(env.best_cpu);
2257 if (nid != p->numa_preferred_nid)
2258 sched_setnuma(p, nid);
2261 /* No better CPU than the current one was found. */
2262 if (env.best_cpu == -1) {
2263 trace_sched_stick_numa(p, env.src_cpu, NULL, -1);
2267 best_rq = cpu_rq(env.best_cpu);
2268 if (env.best_task == NULL) {
2269 ret = migrate_task_to(p, env.best_cpu);
2270 WRITE_ONCE(best_rq->numa_migrate_on, 0);
2272 trace_sched_stick_numa(p, env.src_cpu, NULL, env.best_cpu);
2276 ret = migrate_swap(p, env.best_task, env.best_cpu, env.src_cpu);
2277 WRITE_ONCE(best_rq->numa_migrate_on, 0);
2280 trace_sched_stick_numa(p, env.src_cpu, env.best_task, env.best_cpu);
2281 put_task_struct(env.best_task);
2285 /* Attempt to migrate a task to a CPU on the preferred node. */
2286 static void numa_migrate_preferred(struct task_struct *p)
2288 unsigned long interval = HZ;
2290 /* This task has no NUMA fault statistics yet */
2291 if (unlikely(p->numa_preferred_nid == NUMA_NO_NODE || !p->numa_faults))
2294 /* Periodically retry migrating the task to the preferred node */
2295 interval = min(interval, msecs_to_jiffies(p->numa_scan_period) / 16);
2296 p->numa_migrate_retry = jiffies + interval;
2298 /* Success if task is already running on preferred CPU */
2299 if (task_node(p) == p->numa_preferred_nid)
2302 /* Otherwise, try migrate to a CPU on the preferred node */
2303 task_numa_migrate(p);
2307 * Find out how many nodes the workload is actively running on. Do this by
2308 * tracking the nodes from which NUMA hinting faults are triggered. This can
2309 * be different from the set of nodes where the workload's memory is currently
2312 static void numa_group_count_active_nodes(struct numa_group *numa_group)
2314 unsigned long faults, max_faults = 0;
2315 int nid, active_nodes = 0;
2317 for_each_node_state(nid, N_CPU) {
2318 faults = group_faults_cpu(numa_group, nid);
2319 if (faults > max_faults)
2320 max_faults = faults;
2323 for_each_node_state(nid, N_CPU) {
2324 faults = group_faults_cpu(numa_group, nid);
2325 if (faults * ACTIVE_NODE_FRACTION > max_faults)
2329 numa_group->max_faults_cpu = max_faults;
2330 numa_group->active_nodes = active_nodes;
2334 * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
2335 * increments. The more local the fault statistics are, the higher the scan
2336 * period will be for the next scan window. If local/(local+remote) ratio is
2337 * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS)
2338 * the scan period will decrease. Aim for 70% local accesses.
2340 #define NUMA_PERIOD_SLOTS 10
2341 #define NUMA_PERIOD_THRESHOLD 7
2344 * Increase the scan period (slow down scanning) if the majority of
2345 * our memory is already on our local node, or if the majority of
2346 * the page accesses are shared with other processes.
2347 * Otherwise, decrease the scan period.
2349 static void update_task_scan_period(struct task_struct *p,
2350 unsigned long shared, unsigned long private)
2352 unsigned int period_slot;
2353 int lr_ratio, ps_ratio;
2356 unsigned long remote = p->numa_faults_locality[0];
2357 unsigned long local = p->numa_faults_locality[1];
2360 * If there were no record hinting faults then either the task is
2361 * completely idle or all activity is in areas that are not of interest
2362 * to automatic numa balancing. Related to that, if there were failed
2363 * migration then it implies we are migrating too quickly or the local
2364 * node is overloaded. In either case, scan slower
2366 if (local + shared == 0 || p->numa_faults_locality[2]) {
2367 p->numa_scan_period = min(p->numa_scan_period_max,
2368 p->numa_scan_period << 1);
2370 p->mm->numa_next_scan = jiffies +
2371 msecs_to_jiffies(p->numa_scan_period);
2377 * Prepare to scale scan period relative to the current period.
2378 * == NUMA_PERIOD_THRESHOLD scan period stays the same
2379 * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
2380 * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
2382 period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS);
2383 lr_ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote);
2384 ps_ratio = (private * NUMA_PERIOD_SLOTS) / (private + shared);
2386 if (ps_ratio >= NUMA_PERIOD_THRESHOLD) {
2388 * Most memory accesses are local. There is no need to
2389 * do fast NUMA scanning, since memory is already local.
2391 int slot = ps_ratio - NUMA_PERIOD_THRESHOLD;
2394 diff = slot * period_slot;
2395 } else if (lr_ratio >= NUMA_PERIOD_THRESHOLD) {
2397 * Most memory accesses are shared with other tasks.
2398 * There is no point in continuing fast NUMA scanning,
2399 * since other tasks may just move the memory elsewhere.
2401 int slot = lr_ratio - NUMA_PERIOD_THRESHOLD;
2404 diff = slot * period_slot;
2407 * Private memory faults exceed (SLOTS-THRESHOLD)/SLOTS,
2408 * yet they are not on the local NUMA node. Speed up
2409 * NUMA scanning to get the memory moved over.
2411 int ratio = max(lr_ratio, ps_ratio);
2412 diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot;
2415 p->numa_scan_period = clamp(p->numa_scan_period + diff,
2416 task_scan_min(p), task_scan_max(p));
2417 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
2421 * Get the fraction of time the task has been running since the last
2422 * NUMA placement cycle. The scheduler keeps similar statistics, but
2423 * decays those on a 32ms period, which is orders of magnitude off
2424 * from the dozens-of-seconds NUMA balancing period. Use the scheduler
2425 * stats only if the task is so new there are no NUMA statistics yet.
2427 static u64 numa_get_avg_runtime(struct task_struct *p, u64 *period)
2429 u64 runtime, delta, now;
2430 /* Use the start of this time slice to avoid calculations. */
2431 now = p->se.exec_start;
2432 runtime = p->se.sum_exec_runtime;
2434 if (p->last_task_numa_placement) {
2435 delta = runtime - p->last_sum_exec_runtime;
2436 *period = now - p->last_task_numa_placement;
2438 /* Avoid time going backwards, prevent potential divide error: */
2439 if (unlikely((s64)*period < 0))
2442 delta = p->se.avg.load_sum;
2443 *period = LOAD_AVG_MAX;
2446 p->last_sum_exec_runtime = runtime;
2447 p->last_task_numa_placement = now;
2453 * Determine the preferred nid for a task in a numa_group. This needs to
2454 * be done in a way that produces consistent results with group_weight,
2455 * otherwise workloads might not converge.
2457 static int preferred_group_nid(struct task_struct *p, int nid)
2462 /* Direct connections between all NUMA nodes. */
2463 if (sched_numa_topology_type == NUMA_DIRECT)
2467 * On a system with glueless mesh NUMA topology, group_weight
2468 * scores nodes according to the number of NUMA hinting faults on
2469 * both the node itself, and on nearby nodes.
2471 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
2472 unsigned long score, max_score = 0;
2473 int node, max_node = nid;
2475 dist = sched_max_numa_distance;
2477 for_each_node_state(node, N_CPU) {
2478 score = group_weight(p, node, dist);
2479 if (score > max_score) {
2488 * Finding the preferred nid in a system with NUMA backplane
2489 * interconnect topology is more involved. The goal is to locate
2490 * tasks from numa_groups near each other in the system, and
2491 * untangle workloads from different sides of the system. This requires
2492 * searching down the hierarchy of node groups, recursively searching
2493 * inside the highest scoring group of nodes. The nodemask tricks
2494 * keep the complexity of the search down.
2496 nodes = node_states[N_CPU];
2497 for (dist = sched_max_numa_distance; dist > LOCAL_DISTANCE; dist--) {
2498 unsigned long max_faults = 0;
2499 nodemask_t max_group = NODE_MASK_NONE;
2502 /* Are there nodes at this distance from each other? */
2503 if (!find_numa_distance(dist))
2506 for_each_node_mask(a, nodes) {
2507 unsigned long faults = 0;
2508 nodemask_t this_group;
2509 nodes_clear(this_group);
2511 /* Sum group's NUMA faults; includes a==b case. */
2512 for_each_node_mask(b, nodes) {
2513 if (node_distance(a, b) < dist) {
2514 faults += group_faults(p, b);
2515 node_set(b, this_group);
2516 node_clear(b, nodes);
2520 /* Remember the top group. */
2521 if (faults > max_faults) {
2522 max_faults = faults;
2523 max_group = this_group;
2525 * subtle: at the smallest distance there is
2526 * just one node left in each "group", the
2527 * winner is the preferred nid.
2532 /* Next round, evaluate the nodes within max_group. */
2540 static void task_numa_placement(struct task_struct *p)
2542 int seq, nid, max_nid = NUMA_NO_NODE;
2543 unsigned long max_faults = 0;
2544 unsigned long fault_types[2] = { 0, 0 };
2545 unsigned long total_faults;
2546 u64 runtime, period;
2547 spinlock_t *group_lock = NULL;
2548 struct numa_group *ng;
2551 * The p->mm->numa_scan_seq field gets updated without
2552 * exclusive access. Use READ_ONCE() here to ensure
2553 * that the field is read in a single access:
2555 seq = READ_ONCE(p->mm->numa_scan_seq);
2556 if (p->numa_scan_seq == seq)
2558 p->numa_scan_seq = seq;
2559 p->numa_scan_period_max = task_scan_max(p);
2561 total_faults = p->numa_faults_locality[0] +
2562 p->numa_faults_locality[1];
2563 runtime = numa_get_avg_runtime(p, &period);
2565 /* If the task is part of a group prevent parallel updates to group stats */
2566 ng = deref_curr_numa_group(p);
2568 group_lock = &ng->lock;
2569 spin_lock_irq(group_lock);
2572 /* Find the node with the highest number of faults */
2573 for_each_online_node(nid) {
2574 /* Keep track of the offsets in numa_faults array */
2575 int mem_idx, membuf_idx, cpu_idx, cpubuf_idx;
2576 unsigned long faults = 0, group_faults = 0;
2579 for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) {
2580 long diff, f_diff, f_weight;
2582 mem_idx = task_faults_idx(NUMA_MEM, nid, priv);
2583 membuf_idx = task_faults_idx(NUMA_MEMBUF, nid, priv);
2584 cpu_idx = task_faults_idx(NUMA_CPU, nid, priv);
2585 cpubuf_idx = task_faults_idx(NUMA_CPUBUF, nid, priv);
2587 /* Decay existing window, copy faults since last scan */
2588 diff = p->numa_faults[membuf_idx] - p->numa_faults[mem_idx] / 2;
2589 fault_types[priv] += p->numa_faults[membuf_idx];
2590 p->numa_faults[membuf_idx] = 0;
2593 * Normalize the faults_from, so all tasks in a group
2594 * count according to CPU use, instead of by the raw
2595 * number of faults. Tasks with little runtime have
2596 * little over-all impact on throughput, and thus their
2597 * faults are less important.
2599 f_weight = div64_u64(runtime << 16, period + 1);
2600 f_weight = (f_weight * p->numa_faults[cpubuf_idx]) /
2602 f_diff = f_weight - p->numa_faults[cpu_idx] / 2;
2603 p->numa_faults[cpubuf_idx] = 0;
2605 p->numa_faults[mem_idx] += diff;
2606 p->numa_faults[cpu_idx] += f_diff;
2607 faults += p->numa_faults[mem_idx];
2608 p->total_numa_faults += diff;
2611 * safe because we can only change our own group
2613 * mem_idx represents the offset for a given
2614 * nid and priv in a specific region because it
2615 * is at the beginning of the numa_faults array.
2617 ng->faults[mem_idx] += diff;
2618 ng->faults[cpu_idx] += f_diff;
2619 ng->total_faults += diff;
2620 group_faults += ng->faults[mem_idx];
2625 if (faults > max_faults) {
2626 max_faults = faults;
2629 } else if (group_faults > max_faults) {
2630 max_faults = group_faults;
2635 /* Cannot migrate task to CPU-less node */
2636 if (max_nid != NUMA_NO_NODE && !node_state(max_nid, N_CPU)) {
2637 int near_nid = max_nid;
2638 int distance, near_distance = INT_MAX;
2640 for_each_node_state(nid, N_CPU) {
2641 distance = node_distance(max_nid, nid);
2642 if (distance < near_distance) {
2644 near_distance = distance;
2651 numa_group_count_active_nodes(ng);
2652 spin_unlock_irq(group_lock);
2653 max_nid = preferred_group_nid(p, max_nid);
2657 /* Set the new preferred node */
2658 if (max_nid != p->numa_preferred_nid)
2659 sched_setnuma(p, max_nid);
2662 update_task_scan_period(p, fault_types[0], fault_types[1]);
2665 static inline int get_numa_group(struct numa_group *grp)
2667 return refcount_inc_not_zero(&grp->refcount);
2670 static inline void put_numa_group(struct numa_group *grp)
2672 if (refcount_dec_and_test(&grp->refcount))
2673 kfree_rcu(grp, rcu);
2676 static void task_numa_group(struct task_struct *p, int cpupid, int flags,
2679 struct numa_group *grp, *my_grp;
2680 struct task_struct *tsk;
2682 int cpu = cpupid_to_cpu(cpupid);
2685 if (unlikely(!deref_curr_numa_group(p))) {
2686 unsigned int size = sizeof(struct numa_group) +
2687 NR_NUMA_HINT_FAULT_STATS *
2688 nr_node_ids * sizeof(unsigned long);
2690 grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
2694 refcount_set(&grp->refcount, 1);
2695 grp->active_nodes = 1;
2696 grp->max_faults_cpu = 0;
2697 spin_lock_init(&grp->lock);
2700 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2701 grp->faults[i] = p->numa_faults[i];
2703 grp->total_faults = p->total_numa_faults;
2706 rcu_assign_pointer(p->numa_group, grp);
2710 tsk = READ_ONCE(cpu_rq(cpu)->curr);
2712 if (!cpupid_match_pid(tsk, cpupid))
2715 grp = rcu_dereference(tsk->numa_group);
2719 my_grp = deref_curr_numa_group(p);
2724 * Only join the other group if its bigger; if we're the bigger group,
2725 * the other task will join us.
2727 if (my_grp->nr_tasks > grp->nr_tasks)
2731 * Tie-break on the grp address.
2733 if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp)
2736 /* Always join threads in the same process. */
2737 if (tsk->mm == current->mm)
2740 /* Simple filter to avoid false positives due to PID collisions */
2741 if (flags & TNF_SHARED)
2744 /* Update priv based on whether false sharing was detected */
2747 if (join && !get_numa_group(grp))
2755 WARN_ON_ONCE(irqs_disabled());
2756 double_lock_irq(&my_grp->lock, &grp->lock);
2758 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) {
2759 my_grp->faults[i] -= p->numa_faults[i];
2760 grp->faults[i] += p->numa_faults[i];
2762 my_grp->total_faults -= p->total_numa_faults;
2763 grp->total_faults += p->total_numa_faults;
2768 spin_unlock(&my_grp->lock);
2769 spin_unlock_irq(&grp->lock);
2771 rcu_assign_pointer(p->numa_group, grp);
2773 put_numa_group(my_grp);
2782 * Get rid of NUMA statistics associated with a task (either current or dead).
2783 * If @final is set, the task is dead and has reached refcount zero, so we can
2784 * safely free all relevant data structures. Otherwise, there might be
2785 * concurrent reads from places like load balancing and procfs, and we should
2786 * reset the data back to default state without freeing ->numa_faults.
2788 void task_numa_free(struct task_struct *p, bool final)
2790 /* safe: p either is current or is being freed by current */
2791 struct numa_group *grp = rcu_dereference_raw(p->numa_group);
2792 unsigned long *numa_faults = p->numa_faults;
2793 unsigned long flags;
2800 spin_lock_irqsave(&grp->lock, flags);
2801 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2802 grp->faults[i] -= p->numa_faults[i];
2803 grp->total_faults -= p->total_numa_faults;
2806 spin_unlock_irqrestore(&grp->lock, flags);
2807 RCU_INIT_POINTER(p->numa_group, NULL);
2808 put_numa_group(grp);
2812 p->numa_faults = NULL;
2815 p->total_numa_faults = 0;
2816 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2822 * Got a PROT_NONE fault for a page on @node.
2824 void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags)
2826 struct task_struct *p = current;
2827 bool migrated = flags & TNF_MIGRATED;
2828 int cpu_node = task_node(current);
2829 int local = !!(flags & TNF_FAULT_LOCAL);
2830 struct numa_group *ng;
2833 if (!static_branch_likely(&sched_numa_balancing))
2836 /* for example, ksmd faulting in a user's mm */
2841 * NUMA faults statistics are unnecessary for the slow memory
2842 * node for memory tiering mode.
2844 if (!node_is_toptier(mem_node) &&
2845 (sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING ||
2846 !cpupid_valid(last_cpupid)))
2849 /* Allocate buffer to track faults on a per-node basis */
2850 if (unlikely(!p->numa_faults)) {
2851 int size = sizeof(*p->numa_faults) *
2852 NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids;
2854 p->numa_faults = kzalloc(size, GFP_KERNEL|__GFP_NOWARN);
2855 if (!p->numa_faults)
2858 p->total_numa_faults = 0;
2859 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
2863 * First accesses are treated as private, otherwise consider accesses
2864 * to be private if the accessing pid has not changed
2866 if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) {
2869 priv = cpupid_match_pid(p, last_cpupid);
2870 if (!priv && !(flags & TNF_NO_GROUP))
2871 task_numa_group(p, last_cpupid, flags, &priv);
2875 * If a workload spans multiple NUMA nodes, a shared fault that
2876 * occurs wholly within the set of nodes that the workload is
2877 * actively using should be counted as local. This allows the
2878 * scan rate to slow down when a workload has settled down.
2880 ng = deref_curr_numa_group(p);
2881 if (!priv && !local && ng && ng->active_nodes > 1 &&
2882 numa_is_active_node(cpu_node, ng) &&
2883 numa_is_active_node(mem_node, ng))
2887 * Retry to migrate task to preferred node periodically, in case it
2888 * previously failed, or the scheduler moved us.
2890 if (time_after(jiffies, p->numa_migrate_retry)) {
2891 task_numa_placement(p);
2892 numa_migrate_preferred(p);
2896 p->numa_pages_migrated += pages;
2897 if (flags & TNF_MIGRATE_FAIL)
2898 p->numa_faults_locality[2] += pages;
2900 p->numa_faults[task_faults_idx(NUMA_MEMBUF, mem_node, priv)] += pages;
2901 p->numa_faults[task_faults_idx(NUMA_CPUBUF, cpu_node, priv)] += pages;
2902 p->numa_faults_locality[local] += pages;
2905 static void reset_ptenuma_scan(struct task_struct *p)
2908 * We only did a read acquisition of the mmap sem, so
2909 * p->mm->numa_scan_seq is written to without exclusive access
2910 * and the update is not guaranteed to be atomic. That's not
2911 * much of an issue though, since this is just used for
2912 * statistical sampling. Use READ_ONCE/WRITE_ONCE, which are not
2913 * expensive, to avoid any form of compiler optimizations:
2915 WRITE_ONCE(p->mm->numa_scan_seq, READ_ONCE(p->mm->numa_scan_seq) + 1);
2916 p->mm->numa_scan_offset = 0;
2920 * The expensive part of numa migration is done from task_work context.
2921 * Triggered from task_tick_numa().
2923 static void task_numa_work(struct callback_head *work)
2925 unsigned long migrate, next_scan, now = jiffies;
2926 struct task_struct *p = current;
2927 struct mm_struct *mm = p->mm;
2928 u64 runtime = p->se.sum_exec_runtime;
2929 MA_STATE(mas, &mm->mm_mt, 0, 0);
2930 struct vm_area_struct *vma;
2931 unsigned long start, end;
2932 unsigned long nr_pte_updates = 0;
2933 long pages, virtpages;
2935 SCHED_WARN_ON(p != container_of(work, struct task_struct, numa_work));
2939 * Who cares about NUMA placement when they're dying.
2941 * NOTE: make sure not to dereference p->mm before this check,
2942 * exit_task_work() happens _after_ exit_mm() so we could be called
2943 * without p->mm even though we still had it when we enqueued this
2946 if (p->flags & PF_EXITING)
2949 if (!mm->numa_next_scan) {
2950 mm->numa_next_scan = now +
2951 msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2955 * Enforce maximal scan/migration frequency..
2957 migrate = mm->numa_next_scan;
2958 if (time_before(now, migrate))
2961 if (p->numa_scan_period == 0) {
2962 p->numa_scan_period_max = task_scan_max(p);
2963 p->numa_scan_period = task_scan_start(p);
2966 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
2967 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
2971 * Delay this task enough that another task of this mm will likely win
2972 * the next time around.
2974 p->node_stamp += 2 * TICK_NSEC;
2976 start = mm->numa_scan_offset;
2977 pages = sysctl_numa_balancing_scan_size;
2978 pages <<= 20 - PAGE_SHIFT; /* MB in pages */
2979 virtpages = pages * 8; /* Scan up to this much virtual space */
2984 if (!mmap_read_trylock(mm))
2986 mas_set(&mas, start);
2987 vma = mas_find(&mas, ULONG_MAX);
2989 reset_ptenuma_scan(p);
2991 mas_set(&mas, start);
2992 vma = mas_find(&mas, ULONG_MAX);
2995 for (; vma; vma = mas_find(&mas, ULONG_MAX)) {
2996 if (!vma_migratable(vma) || !vma_policy_mof(vma) ||
2997 is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_MIXEDMAP)) {
3002 * Shared library pages mapped by multiple processes are not
3003 * migrated as it is expected they are cache replicated. Avoid
3004 * hinting faults in read-only file-backed mappings or the vdso
3005 * as migrating the pages will be of marginal benefit.
3008 (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ)))
3012 * Skip inaccessible VMAs to avoid any confusion between
3013 * PROT_NONE and NUMA hinting ptes
3015 if (!vma_is_accessible(vma))
3019 start = max(start, vma->vm_start);
3020 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
3021 end = min(end, vma->vm_end);
3022 nr_pte_updates = change_prot_numa(vma, start, end);
3025 * Try to scan sysctl_numa_balancing_size worth of
3026 * hpages that have at least one present PTE that
3027 * is not already pte-numa. If the VMA contains
3028 * areas that are unused or already full of prot_numa
3029 * PTEs, scan up to virtpages, to skip through those
3033 pages -= (end - start) >> PAGE_SHIFT;
3034 virtpages -= (end - start) >> PAGE_SHIFT;
3037 if (pages <= 0 || virtpages <= 0)
3041 } while (end != vma->vm_end);
3046 * It is possible to reach the end of the VMA list but the last few
3047 * VMAs are not guaranteed to the vma_migratable. If they are not, we
3048 * would find the !migratable VMA on the next scan but not reset the
3049 * scanner to the start so check it now.
3052 mm->numa_scan_offset = start;
3054 reset_ptenuma_scan(p);
3055 mmap_read_unlock(mm);
3058 * Make sure tasks use at least 32x as much time to run other code
3059 * than they used here, to limit NUMA PTE scanning overhead to 3% max.
3060 * Usually update_task_scan_period slows down scanning enough; on an
3061 * overloaded system we need to limit overhead on a per task basis.
3063 if (unlikely(p->se.sum_exec_runtime != runtime)) {
3064 u64 diff = p->se.sum_exec_runtime - runtime;
3065 p->node_stamp += 32 * diff;
3069 void init_numa_balancing(unsigned long clone_flags, struct task_struct *p)
3072 struct mm_struct *mm = p->mm;
3075 mm_users = atomic_read(&mm->mm_users);
3076 if (mm_users == 1) {
3077 mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
3078 mm->numa_scan_seq = 0;
3082 p->numa_scan_seq = mm ? mm->numa_scan_seq : 0;
3083 p->numa_scan_period = sysctl_numa_balancing_scan_delay;
3084 p->numa_migrate_retry = 0;
3085 /* Protect against double add, see task_tick_numa and task_numa_work */
3086 p->numa_work.next = &p->numa_work;
3087 p->numa_faults = NULL;
3088 p->numa_pages_migrated = 0;
3089 p->total_numa_faults = 0;
3090 RCU_INIT_POINTER(p->numa_group, NULL);
3091 p->last_task_numa_placement = 0;
3092 p->last_sum_exec_runtime = 0;
3094 init_task_work(&p->numa_work, task_numa_work);
3096 /* New address space, reset the preferred nid */
3097 if (!(clone_flags & CLONE_VM)) {
3098 p->numa_preferred_nid = NUMA_NO_NODE;
3103 * New thread, keep existing numa_preferred_nid which should be copied
3104 * already by arch_dup_task_struct but stagger when scans start.
3109 delay = min_t(unsigned int, task_scan_max(current),
3110 current->numa_scan_period * mm_users * NSEC_PER_MSEC);
3111 delay += 2 * TICK_NSEC;
3112 p->node_stamp = delay;
3117 * Drive the periodic memory faults..
3119 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
3121 struct callback_head *work = &curr->numa_work;
3125 * We don't care about NUMA placement if we don't have memory.
3127 if (!curr->mm || (curr->flags & (PF_EXITING | PF_KTHREAD)) || work->next != work)
3131 * Using runtime rather than walltime has the dual advantage that
3132 * we (mostly) drive the selection from busy threads and that the
3133 * task needs to have done some actual work before we bother with
3136 now = curr->se.sum_exec_runtime;
3137 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
3139 if (now > curr->node_stamp + period) {
3140 if (!curr->node_stamp)
3141 curr->numa_scan_period = task_scan_start(curr);
3142 curr->node_stamp += period;
3144 if (!time_before(jiffies, curr->mm->numa_next_scan))
3145 task_work_add(curr, work, TWA_RESUME);
3149 static void update_scan_period(struct task_struct *p, int new_cpu)
3151 int src_nid = cpu_to_node(task_cpu(p));
3152 int dst_nid = cpu_to_node(new_cpu);
3154 if (!static_branch_likely(&sched_numa_balancing))
3157 if (!p->mm || !p->numa_faults || (p->flags & PF_EXITING))
3160 if (src_nid == dst_nid)
3164 * Allow resets if faults have been trapped before one scan
3165 * has completed. This is most likely due to a new task that
3166 * is pulled cross-node due to wakeups or load balancing.
3168 if (p->numa_scan_seq) {
3170 * Avoid scan adjustments if moving to the preferred
3171 * node or if the task was not previously running on
3172 * the preferred node.
3174 if (dst_nid == p->numa_preferred_nid ||
3175 (p->numa_preferred_nid != NUMA_NO_NODE &&
3176 src_nid != p->numa_preferred_nid))
3180 p->numa_scan_period = task_scan_start(p);
3184 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
3188 static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p)
3192 static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p)
3196 static inline void update_scan_period(struct task_struct *p, int new_cpu)
3200 #endif /* CONFIG_NUMA_BALANCING */
3203 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
3205 update_load_add(&cfs_rq->load, se->load.weight);
3207 if (entity_is_task(se)) {
3208 struct rq *rq = rq_of(cfs_rq);
3210 account_numa_enqueue(rq, task_of(se));
3211 list_add(&se->group_node, &rq->cfs_tasks);
3214 cfs_rq->nr_running++;
3216 cfs_rq->idle_nr_running++;
3220 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
3222 update_load_sub(&cfs_rq->load, se->load.weight);
3224 if (entity_is_task(se)) {
3225 account_numa_dequeue(rq_of(cfs_rq), task_of(se));
3226 list_del_init(&se->group_node);
3229 cfs_rq->nr_running--;
3231 cfs_rq->idle_nr_running--;
3235 * Signed add and clamp on underflow.
3237 * Explicitly do a load-store to ensure the intermediate value never hits
3238 * memory. This allows lockless observations without ever seeing the negative
3241 #define add_positive(_ptr, _val) do { \
3242 typeof(_ptr) ptr = (_ptr); \
3243 typeof(_val) val = (_val); \
3244 typeof(*ptr) res, var = READ_ONCE(*ptr); \
3248 if (val < 0 && res > var) \
3251 WRITE_ONCE(*ptr, res); \
3255 * Unsigned subtract and clamp on underflow.
3257 * Explicitly do a load-store to ensure the intermediate value never hits
3258 * memory. This allows lockless observations without ever seeing the negative
3261 #define sub_positive(_ptr, _val) do { \
3262 typeof(_ptr) ptr = (_ptr); \
3263 typeof(*ptr) val = (_val); \
3264 typeof(*ptr) res, var = READ_ONCE(*ptr); \
3268 WRITE_ONCE(*ptr, res); \
3272 * Remove and clamp on negative, from a local variable.
3274 * A variant of sub_positive(), which does not use explicit load-store
3275 * and is thus optimized for local variable updates.
3277 #define lsub_positive(_ptr, _val) do { \
3278 typeof(_ptr) ptr = (_ptr); \
3279 *ptr -= min_t(typeof(*ptr), *ptr, _val); \
3284 enqueue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3286 cfs_rq->avg.load_avg += se->avg.load_avg;
3287 cfs_rq->avg.load_sum += se_weight(se) * se->avg.load_sum;
3291 dequeue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3293 sub_positive(&cfs_rq->avg.load_avg, se->avg.load_avg);
3294 sub_positive(&cfs_rq->avg.load_sum, se_weight(se) * se->avg.load_sum);
3295 /* See update_cfs_rq_load_avg() */
3296 cfs_rq->avg.load_sum = max_t(u32, cfs_rq->avg.load_sum,
3297 cfs_rq->avg.load_avg * PELT_MIN_DIVIDER);
3301 enqueue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
3303 dequeue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
3306 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
3307 unsigned long weight)
3310 /* commit outstanding execution time */
3311 if (cfs_rq->curr == se)
3312 update_curr(cfs_rq);
3313 update_load_sub(&cfs_rq->load, se->load.weight);
3315 dequeue_load_avg(cfs_rq, se);
3317 update_load_set(&se->load, weight);
3321 u32 divider = get_pelt_divider(&se->avg);
3323 se->avg.load_avg = div_u64(se_weight(se) * se->avg.load_sum, divider);
3327 enqueue_load_avg(cfs_rq, se);
3329 update_load_add(&cfs_rq->load, se->load.weight);
3333 void reweight_task(struct task_struct *p, int prio)
3335 struct sched_entity *se = &p->se;
3336 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3337 struct load_weight *load = &se->load;
3338 unsigned long weight = scale_load(sched_prio_to_weight[prio]);
3340 reweight_entity(cfs_rq, se, weight);
3341 load->inv_weight = sched_prio_to_wmult[prio];
3344 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
3346 #ifdef CONFIG_FAIR_GROUP_SCHED
3349 * All this does is approximate the hierarchical proportion which includes that
3350 * global sum we all love to hate.
3352 * That is, the weight of a group entity, is the proportional share of the
3353 * group weight based on the group runqueue weights. That is:
3355 * tg->weight * grq->load.weight
3356 * ge->load.weight = ----------------------------- (1)
3357 * \Sum grq->load.weight
3359 * Now, because computing that sum is prohibitively expensive to compute (been
3360 * there, done that) we approximate it with this average stuff. The average
3361 * moves slower and therefore the approximation is cheaper and more stable.
3363 * So instead of the above, we substitute:
3365 * grq->load.weight -> grq->avg.load_avg (2)
3367 * which yields the following:
3369 * tg->weight * grq->avg.load_avg
3370 * ge->load.weight = ------------------------------ (3)
3373 * Where: tg->load_avg ~= \Sum grq->avg.load_avg
3375 * That is shares_avg, and it is right (given the approximation (2)).
3377 * The problem with it is that because the average is slow -- it was designed
3378 * to be exactly that of course -- this leads to transients in boundary
3379 * conditions. In specific, the case where the group was idle and we start the
3380 * one task. It takes time for our CPU's grq->avg.load_avg to build up,
3381 * yielding bad latency etc..
3383 * Now, in that special case (1) reduces to:
3385 * tg->weight * grq->load.weight
3386 * ge->load.weight = ----------------------------- = tg->weight (4)
3389 * That is, the sum collapses because all other CPUs are idle; the UP scenario.
3391 * So what we do is modify our approximation (3) to approach (4) in the (near)
3396 * tg->weight * grq->load.weight
3397 * --------------------------------------------------- (5)
3398 * tg->load_avg - grq->avg.load_avg + grq->load.weight
3400 * But because grq->load.weight can drop to 0, resulting in a divide by zero,
3401 * we need to use grq->avg.load_avg as its lower bound, which then gives:
3404 * tg->weight * grq->load.weight
3405 * ge->load.weight = ----------------------------- (6)
3410 * tg_load_avg' = tg->load_avg - grq->avg.load_avg +
3411 * max(grq->load.weight, grq->avg.load_avg)
3413 * And that is shares_weight and is icky. In the (near) UP case it approaches
3414 * (4) while in the normal case it approaches (3). It consistently
3415 * overestimates the ge->load.weight and therefore:
3417 * \Sum ge->load.weight >= tg->weight
3421 static long calc_group_shares(struct cfs_rq *cfs_rq)
3423 long tg_weight, tg_shares, load, shares;
3424 struct task_group *tg = cfs_rq->tg;
3426 tg_shares = READ_ONCE(tg->shares);
3428 load = max(scale_load_down(cfs_rq->load.weight), cfs_rq->avg.load_avg);
3430 tg_weight = atomic_long_read(&tg->load_avg);
3432 /* Ensure tg_weight >= load */
3433 tg_weight -= cfs_rq->tg_load_avg_contrib;
3436 shares = (tg_shares * load);
3438 shares /= tg_weight;
3441 * MIN_SHARES has to be unscaled here to support per-CPU partitioning
3442 * of a group with small tg->shares value. It is a floor value which is
3443 * assigned as a minimum load.weight to the sched_entity representing
3444 * the group on a CPU.
3446 * E.g. on 64-bit for a group with tg->shares of scale_load(15)=15*1024
3447 * on an 8-core system with 8 tasks each runnable on one CPU shares has
3448 * to be 15*1024*1/8=1920 instead of scale_load(MIN_SHARES)=2*1024. In
3449 * case no task is runnable on a CPU MIN_SHARES=2 should be returned
3452 return clamp_t(long, shares, MIN_SHARES, tg_shares);
3454 #endif /* CONFIG_SMP */
3457 * Recomputes the group entity based on the current state of its group
3460 static void update_cfs_group(struct sched_entity *se)
3462 struct cfs_rq *gcfs_rq = group_cfs_rq(se);
3468 if (throttled_hierarchy(gcfs_rq))
3472 shares = READ_ONCE(gcfs_rq->tg->shares);
3474 if (likely(se->load.weight == shares))
3477 shares = calc_group_shares(gcfs_rq);
3480 reweight_entity(cfs_rq_of(se), se, shares);
3483 #else /* CONFIG_FAIR_GROUP_SCHED */
3484 static inline void update_cfs_group(struct sched_entity *se)
3487 #endif /* CONFIG_FAIR_GROUP_SCHED */
3489 static inline void cfs_rq_util_change(struct cfs_rq *cfs_rq, int flags)
3491 struct rq *rq = rq_of(cfs_rq);
3493 if (&rq->cfs == cfs_rq) {
3495 * There are a few boundary cases this might miss but it should
3496 * get called often enough that that should (hopefully) not be
3499 * It will not get called when we go idle, because the idle
3500 * thread is a different class (!fair), nor will the utilization
3501 * number include things like RT tasks.
3503 * As is, the util number is not freq-invariant (we'd have to
3504 * implement arch_scale_freq_capacity() for that).
3506 * See cpu_util_cfs().
3508 cpufreq_update_util(rq, flags);
3513 static inline bool load_avg_is_decayed(struct sched_avg *sa)
3521 if (sa->runnable_sum)
3525 * _avg must be null when _sum are null because _avg = _sum / divider
3526 * Make sure that rounding and/or propagation of PELT values never
3529 SCHED_WARN_ON(sa->load_avg ||
3536 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
3538 return u64_u32_load_copy(cfs_rq->avg.last_update_time,
3539 cfs_rq->last_update_time_copy);
3541 #ifdef CONFIG_FAIR_GROUP_SCHED
3543 * Because list_add_leaf_cfs_rq always places a child cfs_rq on the list
3544 * immediately before a parent cfs_rq, and cfs_rqs are removed from the list
3545 * bottom-up, we only have to test whether the cfs_rq before us on the list
3547 * If cfs_rq is not on the list, test whether a child needs its to be added to
3548 * connect a branch to the tree * (see list_add_leaf_cfs_rq() for details).
3550 static inline bool child_cfs_rq_on_list(struct cfs_rq *cfs_rq)
3552 struct cfs_rq *prev_cfs_rq;
3553 struct list_head *prev;
3555 if (cfs_rq->on_list) {
3556 prev = cfs_rq->leaf_cfs_rq_list.prev;
3558 struct rq *rq = rq_of(cfs_rq);
3560 prev = rq->tmp_alone_branch;
3563 prev_cfs_rq = container_of(prev, struct cfs_rq, leaf_cfs_rq_list);
3565 return (prev_cfs_rq->tg->parent == cfs_rq->tg);
3568 static inline bool cfs_rq_is_decayed(struct cfs_rq *cfs_rq)
3570 if (cfs_rq->load.weight)
3573 if (!load_avg_is_decayed(&cfs_rq->avg))
3576 if (child_cfs_rq_on_list(cfs_rq))
3583 * update_tg_load_avg - update the tg's load avg
3584 * @cfs_rq: the cfs_rq whose avg changed
3586 * This function 'ensures': tg->load_avg := \Sum tg->cfs_rq[]->avg.load.
3587 * However, because tg->load_avg is a global value there are performance
3590 * In order to avoid having to look at the other cfs_rq's, we use a
3591 * differential update where we store the last value we propagated. This in
3592 * turn allows skipping updates if the differential is 'small'.
3594 * Updating tg's load_avg is necessary before update_cfs_share().
3596 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq)
3598 long delta = cfs_rq->avg.load_avg - cfs_rq->tg_load_avg_contrib;
3601 * No need to update load_avg for root_task_group as it is not used.
3603 if (cfs_rq->tg == &root_task_group)
3606 if (abs(delta) > cfs_rq->tg_load_avg_contrib / 64) {
3607 atomic_long_add(delta, &cfs_rq->tg->load_avg);
3608 cfs_rq->tg_load_avg_contrib = cfs_rq->avg.load_avg;
3613 * Called within set_task_rq() right before setting a task's CPU. The
3614 * caller only guarantees p->pi_lock is held; no other assumptions,
3615 * including the state of rq->lock, should be made.
3617 void set_task_rq_fair(struct sched_entity *se,
3618 struct cfs_rq *prev, struct cfs_rq *next)
3620 u64 p_last_update_time;
3621 u64 n_last_update_time;
3623 if (!sched_feat(ATTACH_AGE_LOAD))
3627 * We are supposed to update the task to "current" time, then its up to
3628 * date and ready to go to new CPU/cfs_rq. But we have difficulty in
3629 * getting what current time is, so simply throw away the out-of-date
3630 * time. This will result in the wakee task is less decayed, but giving
3631 * the wakee more load sounds not bad.
3633 if (!(se->avg.last_update_time && prev))
3636 p_last_update_time = cfs_rq_last_update_time(prev);
3637 n_last_update_time = cfs_rq_last_update_time(next);
3639 __update_load_avg_blocked_se(p_last_update_time, se);
3640 se->avg.last_update_time = n_last_update_time;
3644 * When on migration a sched_entity joins/leaves the PELT hierarchy, we need to
3645 * propagate its contribution. The key to this propagation is the invariant
3646 * that for each group:
3648 * ge->avg == grq->avg (1)
3650 * _IFF_ we look at the pure running and runnable sums. Because they
3651 * represent the very same entity, just at different points in the hierarchy.
3653 * Per the above update_tg_cfs_util() and update_tg_cfs_runnable() are trivial
3654 * and simply copies the running/runnable sum over (but still wrong, because
3655 * the group entity and group rq do not have their PELT windows aligned).
3657 * However, update_tg_cfs_load() is more complex. So we have:
3659 * ge->avg.load_avg = ge->load.weight * ge->avg.runnable_avg (2)
3661 * And since, like util, the runnable part should be directly transferable,
3662 * the following would _appear_ to be the straight forward approach:
3664 * grq->avg.load_avg = grq->load.weight * grq->avg.runnable_avg (3)
3666 * And per (1) we have:
3668 * ge->avg.runnable_avg == grq->avg.runnable_avg
3672 * ge->load.weight * grq->avg.load_avg
3673 * ge->avg.load_avg = ----------------------------------- (4)
3676 * Except that is wrong!
3678 * Because while for entities historical weight is not important and we
3679 * really only care about our future and therefore can consider a pure
3680 * runnable sum, runqueues can NOT do this.
3682 * We specifically want runqueues to have a load_avg that includes
3683 * historical weights. Those represent the blocked load, the load we expect
3684 * to (shortly) return to us. This only works by keeping the weights as
3685 * integral part of the sum. We therefore cannot decompose as per (3).
3687 * Another reason this doesn't work is that runnable isn't a 0-sum entity.
3688 * Imagine a rq with 2 tasks that each are runnable 2/3 of the time. Then the
3689 * rq itself is runnable anywhere between 2/3 and 1 depending on how the
3690 * runnable section of these tasks overlap (or not). If they were to perfectly
3691 * align the rq as a whole would be runnable 2/3 of the time. If however we
3692 * always have at least 1 runnable task, the rq as a whole is always runnable.
3694 * So we'll have to approximate.. :/
3696 * Given the constraint:
3698 * ge->avg.running_sum <= ge->avg.runnable_sum <= LOAD_AVG_MAX
3700 * We can construct a rule that adds runnable to a rq by assuming minimal
3703 * On removal, we'll assume each task is equally runnable; which yields:
3705 * grq->avg.runnable_sum = grq->avg.load_sum / grq->load.weight
3707 * XXX: only do this for the part of runnable > running ?
3711 update_tg_cfs_util(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq)
3713 long delta_sum, delta_avg = gcfs_rq->avg.util_avg - se->avg.util_avg;
3714 u32 new_sum, divider;
3716 /* Nothing to update */
3721 * cfs_rq->avg.period_contrib can be used for both cfs_rq and se.
3722 * See ___update_load_avg() for details.
3724 divider = get_pelt_divider(&cfs_rq->avg);
3727 /* Set new sched_entity's utilization */
3728 se->avg.util_avg = gcfs_rq->avg.util_avg;
3729 new_sum = se->avg.util_avg * divider;
3730 delta_sum = (long)new_sum - (long)se->avg.util_sum;
3731 se->avg.util_sum = new_sum;
3733 /* Update parent cfs_rq utilization */
3734 add_positive(&cfs_rq->avg.util_avg, delta_avg);
3735 add_positive(&cfs_rq->avg.util_sum, delta_sum);
3737 /* See update_cfs_rq_load_avg() */
3738 cfs_rq->avg.util_sum = max_t(u32, cfs_rq->avg.util_sum,
3739 cfs_rq->avg.util_avg * PELT_MIN_DIVIDER);
3743 update_tg_cfs_runnable(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq)
3745 long delta_sum, delta_avg = gcfs_rq->avg.runnable_avg - se->avg.runnable_avg;
3746 u32 new_sum, divider;
3748 /* Nothing to update */
3753 * cfs_rq->avg.period_contrib can be used for both cfs_rq and se.
3754 * See ___update_load_avg() for details.
3756 divider = get_pelt_divider(&cfs_rq->avg);
3758 /* Set new sched_entity's runnable */
3759 se->avg.runnable_avg = gcfs_rq->avg.runnable_avg;
3760 new_sum = se->avg.runnable_avg * divider;
3761 delta_sum = (long)new_sum - (long)se->avg.runnable_sum;
3762 se->avg.runnable_sum = new_sum;
3764 /* Update parent cfs_rq runnable */
3765 add_positive(&cfs_rq->avg.runnable_avg, delta_avg);
3766 add_positive(&cfs_rq->avg.runnable_sum, delta_sum);
3767 /* See update_cfs_rq_load_avg() */
3768 cfs_rq->avg.runnable_sum = max_t(u32, cfs_rq->avg.runnable_sum,
3769 cfs_rq->avg.runnable_avg * PELT_MIN_DIVIDER);
3773 update_tg_cfs_load(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq)
3775 long delta_avg, running_sum, runnable_sum = gcfs_rq->prop_runnable_sum;
3776 unsigned long load_avg;
3784 gcfs_rq->prop_runnable_sum = 0;
3787 * cfs_rq->avg.period_contrib can be used for both cfs_rq and se.
3788 * See ___update_load_avg() for details.
3790 divider = get_pelt_divider(&cfs_rq->avg);
3792 if (runnable_sum >= 0) {
3794 * Add runnable; clip at LOAD_AVG_MAX. Reflects that until
3795 * the CPU is saturated running == runnable.
3797 runnable_sum += se->avg.load_sum;
3798 runnable_sum = min_t(long, runnable_sum, divider);
3801 * Estimate the new unweighted runnable_sum of the gcfs_rq by
3802 * assuming all tasks are equally runnable.
3804 if (scale_load_down(gcfs_rq->load.weight)) {
3805 load_sum = div_u64(gcfs_rq->avg.load_sum,
3806 scale_load_down(gcfs_rq->load.weight));
3809 /* But make sure to not inflate se's runnable */
3810 runnable_sum = min(se->avg.load_sum, load_sum);
3814 * runnable_sum can't be lower than running_sum
3815 * Rescale running sum to be in the same range as runnable sum
3816 * running_sum is in [0 : LOAD_AVG_MAX << SCHED_CAPACITY_SHIFT]
3817 * runnable_sum is in [0 : LOAD_AVG_MAX]
3819 running_sum = se->avg.util_sum >> SCHED_CAPACITY_SHIFT;
3820 runnable_sum = max(runnable_sum, running_sum);
3822 load_sum = se_weight(se) * runnable_sum;
3823 load_avg = div_u64(load_sum, divider);
3825 delta_avg = load_avg - se->avg.load_avg;
3829 delta_sum = load_sum - (s64)se_weight(se) * se->avg.load_sum;
3831 se->avg.load_sum = runnable_sum;
3832 se->avg.load_avg = load_avg;
3833 add_positive(&cfs_rq->avg.load_avg, delta_avg);
3834 add_positive(&cfs_rq->avg.load_sum, delta_sum);
3835 /* See update_cfs_rq_load_avg() */
3836 cfs_rq->avg.load_sum = max_t(u32, cfs_rq->avg.load_sum,
3837 cfs_rq->avg.load_avg * PELT_MIN_DIVIDER);
3840 static inline void add_tg_cfs_propagate(struct cfs_rq *cfs_rq, long runnable_sum)
3842 cfs_rq->propagate = 1;
3843 cfs_rq->prop_runnable_sum += runnable_sum;
3846 /* Update task and its cfs_rq load average */
3847 static inline int propagate_entity_load_avg(struct sched_entity *se)
3849 struct cfs_rq *cfs_rq, *gcfs_rq;
3851 if (entity_is_task(se))
3854 gcfs_rq = group_cfs_rq(se);
3855 if (!gcfs_rq->propagate)
3858 gcfs_rq->propagate = 0;
3860 cfs_rq = cfs_rq_of(se);
3862 add_tg_cfs_propagate(cfs_rq, gcfs_rq->prop_runnable_sum);
3864 update_tg_cfs_util(cfs_rq, se, gcfs_rq);
3865 update_tg_cfs_runnable(cfs_rq, se, gcfs_rq);
3866 update_tg_cfs_load(cfs_rq, se, gcfs_rq);
3868 trace_pelt_cfs_tp(cfs_rq);
3869 trace_pelt_se_tp(se);
3875 * Check if we need to update the load and the utilization of a blocked
3878 static inline bool skip_blocked_update(struct sched_entity *se)
3880 struct cfs_rq *gcfs_rq = group_cfs_rq(se);
3883 * If sched_entity still have not zero load or utilization, we have to
3886 if (se->avg.load_avg || se->avg.util_avg)
3890 * If there is a pending propagation, we have to update the load and
3891 * the utilization of the sched_entity:
3893 if (gcfs_rq->propagate)
3897 * Otherwise, the load and the utilization of the sched_entity is
3898 * already zero and there is no pending propagation, so it will be a
3899 * waste of time to try to decay it:
3904 #else /* CONFIG_FAIR_GROUP_SCHED */
3906 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq) {}
3908 static inline int propagate_entity_load_avg(struct sched_entity *se)
3913 static inline void add_tg_cfs_propagate(struct cfs_rq *cfs_rq, long runnable_sum) {}
3915 #endif /* CONFIG_FAIR_GROUP_SCHED */
3917 #ifdef CONFIG_NO_HZ_COMMON
3918 static inline void migrate_se_pelt_lag(struct sched_entity *se)
3920 u64 throttled = 0, now, lut;
3921 struct cfs_rq *cfs_rq;
3925 if (load_avg_is_decayed(&se->avg))
3928 cfs_rq = cfs_rq_of(se);
3932 is_idle = is_idle_task(rcu_dereference(rq->curr));
3936 * The lag estimation comes with a cost we don't want to pay all the
3937 * time. Hence, limiting to the case where the source CPU is idle and
3938 * we know we are at the greatest risk to have an outdated clock.
3944 * Estimated "now" is: last_update_time + cfs_idle_lag + rq_idle_lag, where:
3946 * last_update_time (the cfs_rq's last_update_time)
3947 * = cfs_rq_clock_pelt()@cfs_rq_idle
3948 * = rq_clock_pelt()@cfs_rq_idle
3949 * - cfs->throttled_clock_pelt_time@cfs_rq_idle
3951 * cfs_idle_lag (delta between rq's update and cfs_rq's update)
3952 * = rq_clock_pelt()@rq_idle - rq_clock_pelt()@cfs_rq_idle
3954 * rq_idle_lag (delta between now and rq's update)
3955 * = sched_clock_cpu() - rq_clock()@rq_idle
3957 * We can then write:
3959 * now = rq_clock_pelt()@rq_idle - cfs->throttled_clock_pelt_time +
3960 * sched_clock_cpu() - rq_clock()@rq_idle
3962 * rq_clock_pelt()@rq_idle is rq->clock_pelt_idle
3963 * rq_clock()@rq_idle is rq->clock_idle
3964 * cfs->throttled_clock_pelt_time@cfs_rq_idle
3965 * is cfs_rq->throttled_pelt_idle
3968 #ifdef CONFIG_CFS_BANDWIDTH
3969 throttled = u64_u32_load(cfs_rq->throttled_pelt_idle);
3970 /* The clock has been stopped for throttling */
3971 if (throttled == U64_MAX)
3974 now = u64_u32_load(rq->clock_pelt_idle);
3976 * Paired with _update_idle_rq_clock_pelt(). It ensures at the worst case
3977 * is observed the old clock_pelt_idle value and the new clock_idle,
3978 * which lead to an underestimation. The opposite would lead to an
3982 lut = cfs_rq_last_update_time(cfs_rq);
3987 * cfs_rq->avg.last_update_time is more recent than our
3988 * estimation, let's use it.
3992 now += sched_clock_cpu(cpu_of(rq)) - u64_u32_load(rq->clock_idle);
3994 __update_load_avg_blocked_se(now, se);
3997 static void migrate_se_pelt_lag(struct sched_entity *se) {}
4001 * update_cfs_rq_load_avg - update the cfs_rq's load/util averages
4002 * @now: current time, as per cfs_rq_clock_pelt()
4003 * @cfs_rq: cfs_rq to update
4005 * The cfs_rq avg is the direct sum of all its entities (blocked and runnable)
4006 * avg. The immediate corollary is that all (fair) tasks must be attached.
4008 * cfs_rq->avg is used for task_h_load() and update_cfs_share() for example.
4010 * Return: true if the load decayed or we removed load.
4012 * Since both these conditions indicate a changed cfs_rq->avg.load we should
4013 * call update_tg_load_avg() when this function returns true.
4016 update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq)
4018 unsigned long removed_load = 0, removed_util = 0, removed_runnable = 0;
4019 struct sched_avg *sa = &cfs_rq->avg;
4022 if (cfs_rq->removed.nr) {
4024 u32 divider = get_pelt_divider(&cfs_rq->avg);
4026 raw_spin_lock(&cfs_rq->removed.lock);
4027 swap(cfs_rq->removed.util_avg, removed_util);
4028 swap(cfs_rq->removed.load_avg, removed_load);
4029 swap(cfs_rq->removed.runnable_avg, removed_runnable);
4030 cfs_rq->removed.nr = 0;
4031 raw_spin_unlock(&cfs_rq->removed.lock);
4034 sub_positive(&sa->load_avg, r);
4035 sub_positive(&sa->load_sum, r * divider);
4036 /* See sa->util_sum below */
4037 sa->load_sum = max_t(u32, sa->load_sum, sa->load_avg * PELT_MIN_DIVIDER);
4040 sub_positive(&sa->util_avg, r);
4041 sub_positive(&sa->util_sum, r * divider);
4043 * Because of rounding, se->util_sum might ends up being +1 more than
4044 * cfs->util_sum. Although this is not a problem by itself, detaching
4045 * a lot of tasks with the rounding problem between 2 updates of
4046 * util_avg (~1ms) can make cfs->util_sum becoming null whereas
4047 * cfs_util_avg is not.
4048 * Check that util_sum is still above its lower bound for the new
4049 * util_avg. Given that period_contrib might have moved since the last
4050 * sync, we are only sure that util_sum must be above or equal to
4051 * util_avg * minimum possible divider
4053 sa->util_sum = max_t(u32, sa->util_sum, sa->util_avg * PELT_MIN_DIVIDER);
4055 r = removed_runnable;
4056 sub_positive(&sa->runnable_avg, r);
4057 sub_positive(&sa->runnable_sum, r * divider);
4058 /* See sa->util_sum above */
4059 sa->runnable_sum = max_t(u32, sa->runnable_sum,
4060 sa->runnable_avg * PELT_MIN_DIVIDER);
4063 * removed_runnable is the unweighted version of removed_load so we
4064 * can use it to estimate removed_load_sum.
4066 add_tg_cfs_propagate(cfs_rq,
4067 -(long)(removed_runnable * divider) >> SCHED_CAPACITY_SHIFT);
4072 decayed |= __update_load_avg_cfs_rq(now, cfs_rq);
4073 u64_u32_store_copy(sa->last_update_time,
4074 cfs_rq->last_update_time_copy,
4075 sa->last_update_time);
4080 * attach_entity_load_avg - attach this entity to its cfs_rq load avg
4081 * @cfs_rq: cfs_rq to attach to
4082 * @se: sched_entity to attach
4084 * Must call update_cfs_rq_load_avg() before this, since we rely on
4085 * cfs_rq->avg.last_update_time being current.
4087 static void attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
4090 * cfs_rq->avg.period_contrib can be used for both cfs_rq and se.
4091 * See ___update_load_avg() for details.
4093 u32 divider = get_pelt_divider(&cfs_rq->avg);
4096 * When we attach the @se to the @cfs_rq, we must align the decay
4097 * window because without that, really weird and wonderful things can
4102 se->avg.last_update_time = cfs_rq->avg.last_update_time;
4103 se->avg.period_contrib = cfs_rq->avg.period_contrib;
4106 * Hell(o) Nasty stuff.. we need to recompute _sum based on the new
4107 * period_contrib. This isn't strictly correct, but since we're
4108 * entirely outside of the PELT hierarchy, nobody cares if we truncate
4111 se->avg.util_sum = se->avg.util_avg * divider;
4113 se->avg.runnable_sum = se->avg.runnable_avg * divider;
4115 se->avg.load_sum = se->avg.load_avg * divider;
4116 if (se_weight(se) < se->avg.load_sum)
4117 se->avg.load_sum = div_u64(se->avg.load_sum, se_weight(se));
4119 se->avg.load_sum = 1;
4121 enqueue_load_avg(cfs_rq, se);
4122 cfs_rq->avg.util_avg += se->avg.util_avg;
4123 cfs_rq->avg.util_sum += se->avg.util_sum;
4124 cfs_rq->avg.runnable_avg += se->avg.runnable_avg;
4125 cfs_rq->avg.runnable_sum += se->avg.runnable_sum;
4127 add_tg_cfs_propagate(cfs_rq, se->avg.load_sum);
4129 cfs_rq_util_change(cfs_rq, 0);
4131 trace_pelt_cfs_tp(cfs_rq);
4135 * detach_entity_load_avg - detach this entity from its cfs_rq load avg
4136 * @cfs_rq: cfs_rq to detach from
4137 * @se: sched_entity to detach
4139 * Must call update_cfs_rq_load_avg() before this, since we rely on
4140 * cfs_rq->avg.last_update_time being current.
4142 static void detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
4144 dequeue_load_avg(cfs_rq, se);
4145 sub_positive(&cfs_rq->avg.util_avg, se->avg.util_avg);
4146 sub_positive(&cfs_rq->avg.util_sum, se->avg.util_sum);
4147 /* See update_cfs_rq_load_avg() */
4148 cfs_rq->avg.util_sum = max_t(u32, cfs_rq->avg.util_sum,
4149 cfs_rq->avg.util_avg * PELT_MIN_DIVIDER);
4151 sub_positive(&cfs_rq->avg.runnable_avg, se->avg.runnable_avg);
4152 sub_positive(&cfs_rq->avg.runnable_sum, se->avg.runnable_sum);
4153 /* See update_cfs_rq_load_avg() */
4154 cfs_rq->avg.runnable_sum = max_t(u32, cfs_rq->avg.runnable_sum,
4155 cfs_rq->avg.runnable_avg * PELT_MIN_DIVIDER);
4157 add_tg_cfs_propagate(cfs_rq, -se->avg.load_sum);
4159 cfs_rq_util_change(cfs_rq, 0);
4161 trace_pelt_cfs_tp(cfs_rq);
4165 * Optional action to be done while updating the load average
4167 #define UPDATE_TG 0x1
4168 #define SKIP_AGE_LOAD 0x2
4169 #define DO_ATTACH 0x4
4170 #define DO_DETACH 0x8
4172 /* Update task and its cfs_rq load average */
4173 static inline void update_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
4175 u64 now = cfs_rq_clock_pelt(cfs_rq);
4179 * Track task load average for carrying it to new CPU after migrated, and
4180 * track group sched_entity load average for task_h_load calc in migration
4182 if (se->avg.last_update_time && !(flags & SKIP_AGE_LOAD))
4183 __update_load_avg_se(now, cfs_rq, se);
4185 decayed = update_cfs_rq_load_avg(now, cfs_rq);
4186 decayed |= propagate_entity_load_avg(se);
4188 if (!se->avg.last_update_time && (flags & DO_ATTACH)) {
4191 * DO_ATTACH means we're here from enqueue_entity().
4192 * !last_update_time means we've passed through
4193 * migrate_task_rq_fair() indicating we migrated.
4195 * IOW we're enqueueing a task on a new CPU.
4197 attach_entity_load_avg(cfs_rq, se);
4198 update_tg_load_avg(cfs_rq);
4200 } else if (flags & DO_DETACH) {
4202 * DO_DETACH means we're here from dequeue_entity()
4203 * and we are migrating task out of the CPU.
4205 detach_entity_load_avg(cfs_rq, se);
4206 update_tg_load_avg(cfs_rq);
4207 } else if (decayed) {
4208 cfs_rq_util_change(cfs_rq, 0);
4210 if (flags & UPDATE_TG)
4211 update_tg_load_avg(cfs_rq);
4216 * Synchronize entity load avg of dequeued entity without locking
4219 static void sync_entity_load_avg(struct sched_entity *se)
4221 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4222 u64 last_update_time;
4224 last_update_time = cfs_rq_last_update_time(cfs_rq);
4225 __update_load_avg_blocked_se(last_update_time, se);
4229 * Task first catches up with cfs_rq, and then subtract
4230 * itself from the cfs_rq (task must be off the queue now).
4232 static void remove_entity_load_avg(struct sched_entity *se)
4234 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4235 unsigned long flags;
4238 * tasks cannot exit without having gone through wake_up_new_task() ->
4239 * enqueue_task_fair() which will have added things to the cfs_rq,
4240 * so we can remove unconditionally.
4243 sync_entity_load_avg(se);
4245 raw_spin_lock_irqsave(&cfs_rq->removed.lock, flags);
4246 ++cfs_rq->removed.nr;
4247 cfs_rq->removed.util_avg += se->avg.util_avg;
4248 cfs_rq->removed.load_avg += se->avg.load_avg;
4249 cfs_rq->removed.runnable_avg += se->avg.runnable_avg;
4250 raw_spin_unlock_irqrestore(&cfs_rq->removed.lock, flags);
4253 static inline unsigned long cfs_rq_runnable_avg(struct cfs_rq *cfs_rq)
4255 return cfs_rq->avg.runnable_avg;
4258 static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq)
4260 return cfs_rq->avg.load_avg;
4263 static int newidle_balance(struct rq *this_rq, struct rq_flags *rf);
4265 static inline unsigned long task_util(struct task_struct *p)
4267 return READ_ONCE(p->se.avg.util_avg);
4270 static inline unsigned long _task_util_est(struct task_struct *p)
4272 struct util_est ue = READ_ONCE(p->se.avg.util_est);
4274 return max(ue.ewma, (ue.enqueued & ~UTIL_AVG_UNCHANGED));
4277 static inline unsigned long task_util_est(struct task_struct *p)
4279 return max(task_util(p), _task_util_est(p));
4282 #ifdef CONFIG_UCLAMP_TASK
4283 static inline unsigned long uclamp_task_util(struct task_struct *p)
4285 return clamp(task_util_est(p),
4286 uclamp_eff_value(p, UCLAMP_MIN),
4287 uclamp_eff_value(p, UCLAMP_MAX));
4290 static inline unsigned long uclamp_task_util(struct task_struct *p)
4292 return task_util_est(p);
4296 static inline void util_est_enqueue(struct cfs_rq *cfs_rq,
4297 struct task_struct *p)
4299 unsigned int enqueued;
4301 if (!sched_feat(UTIL_EST))
4304 /* Update root cfs_rq's estimated utilization */
4305 enqueued = cfs_rq->avg.util_est.enqueued;
4306 enqueued += _task_util_est(p);
4307 WRITE_ONCE(cfs_rq->avg.util_est.enqueued, enqueued);
4309 trace_sched_util_est_cfs_tp(cfs_rq);
4312 static inline void util_est_dequeue(struct cfs_rq *cfs_rq,
4313 struct task_struct *p)
4315 unsigned int enqueued;
4317 if (!sched_feat(UTIL_EST))
4320 /* Update root cfs_rq's estimated utilization */
4321 enqueued = cfs_rq->avg.util_est.enqueued;
4322 enqueued -= min_t(unsigned int, enqueued, _task_util_est(p));
4323 WRITE_ONCE(cfs_rq->avg.util_est.enqueued, enqueued);
4325 trace_sched_util_est_cfs_tp(cfs_rq);
4328 #define UTIL_EST_MARGIN (SCHED_CAPACITY_SCALE / 100)
4331 * Check if a (signed) value is within a specified (unsigned) margin,
4332 * based on the observation that:
4334 * abs(x) < y := (unsigned)(x + y - 1) < (2 * y - 1)
4336 * NOTE: this only works when value + margin < INT_MAX.
4338 static inline bool within_margin(int value, int margin)
4340 return ((unsigned int)(value + margin - 1) < (2 * margin - 1));
4343 static inline void util_est_update(struct cfs_rq *cfs_rq,
4344 struct task_struct *p,
4347 long last_ewma_diff, last_enqueued_diff;
4350 if (!sched_feat(UTIL_EST))
4354 * Skip update of task's estimated utilization when the task has not
4355 * yet completed an activation, e.g. being migrated.
4361 * If the PELT values haven't changed since enqueue time,
4362 * skip the util_est update.
4364 ue = p->se.avg.util_est;
4365 if (ue.enqueued & UTIL_AVG_UNCHANGED)
4368 last_enqueued_diff = ue.enqueued;
4371 * Reset EWMA on utilization increases, the moving average is used only
4372 * to smooth utilization decreases.
4374 ue.enqueued = task_util(p);
4375 if (sched_feat(UTIL_EST_FASTUP)) {
4376 if (ue.ewma < ue.enqueued) {
4377 ue.ewma = ue.enqueued;
4383 * Skip update of task's estimated utilization when its members are
4384 * already ~1% close to its last activation value.
4386 last_ewma_diff = ue.enqueued - ue.ewma;
4387 last_enqueued_diff -= ue.enqueued;
4388 if (within_margin(last_ewma_diff, UTIL_EST_MARGIN)) {
4389 if (!within_margin(last_enqueued_diff, UTIL_EST_MARGIN))
4396 * To avoid overestimation of actual task utilization, skip updates if
4397 * we cannot grant there is idle time in this CPU.
4399 if (task_util(p) > capacity_orig_of(cpu_of(rq_of(cfs_rq))))
4403 * Update Task's estimated utilization
4405 * When *p completes an activation we can consolidate another sample
4406 * of the task size. This is done by storing the current PELT value
4407 * as ue.enqueued and by using this value to update the Exponential
4408 * Weighted Moving Average (EWMA):
4410 * ewma(t) = w * task_util(p) + (1-w) * ewma(t-1)
4411 * = w * task_util(p) + ewma(t-1) - w * ewma(t-1)
4412 * = w * (task_util(p) - ewma(t-1)) + ewma(t-1)
4413 * = w * ( last_ewma_diff ) + ewma(t-1)
4414 * = w * (last_ewma_diff + ewma(t-1) / w)
4416 * Where 'w' is the weight of new samples, which is configured to be
4417 * 0.25, thus making w=1/4 ( >>= UTIL_EST_WEIGHT_SHIFT)
4419 ue.ewma <<= UTIL_EST_WEIGHT_SHIFT;
4420 ue.ewma += last_ewma_diff;
4421 ue.ewma >>= UTIL_EST_WEIGHT_SHIFT;
4423 ue.enqueued |= UTIL_AVG_UNCHANGED;
4424 WRITE_ONCE(p->se.avg.util_est, ue);
4426 trace_sched_util_est_se_tp(&p->se);
4429 static inline int task_fits_capacity(struct task_struct *p,
4430 unsigned long capacity)
4432 return fits_capacity(uclamp_task_util(p), capacity);
4435 static inline void update_misfit_status(struct task_struct *p, struct rq *rq)
4437 if (!sched_asym_cpucap_active())
4440 if (!p || p->nr_cpus_allowed == 1) {
4441 rq->misfit_task_load = 0;
4445 if (task_fits_capacity(p, capacity_of(cpu_of(rq)))) {
4446 rq->misfit_task_load = 0;
4451 * Make sure that misfit_task_load will not be null even if
4452 * task_h_load() returns 0.
4454 rq->misfit_task_load = max_t(unsigned long, task_h_load(p), 1);
4457 #else /* CONFIG_SMP */
4459 static inline bool cfs_rq_is_decayed(struct cfs_rq *cfs_rq)
4464 #define UPDATE_TG 0x0
4465 #define SKIP_AGE_LOAD 0x0
4466 #define DO_ATTACH 0x0
4467 #define DO_DETACH 0x0
4469 static inline void update_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int not_used1)
4471 cfs_rq_util_change(cfs_rq, 0);
4474 static inline void remove_entity_load_avg(struct sched_entity *se) {}
4477 attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
4479 detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
4481 static inline int newidle_balance(struct rq *rq, struct rq_flags *rf)
4487 util_est_enqueue(struct cfs_rq *cfs_rq, struct task_struct *p) {}
4490 util_est_dequeue(struct cfs_rq *cfs_rq, struct task_struct *p) {}
4493 util_est_update(struct cfs_rq *cfs_rq, struct task_struct *p,
4495 static inline void update_misfit_status(struct task_struct *p, struct rq *rq) {}
4497 #endif /* CONFIG_SMP */
4499 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
4501 #ifdef CONFIG_SCHED_DEBUG
4502 s64 d = se->vruntime - cfs_rq->min_vruntime;
4507 if (d > 3*sysctl_sched_latency)
4508 schedstat_inc(cfs_rq->nr_spread_over);
4513 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
4515 u64 vruntime = cfs_rq->min_vruntime;
4518 * The 'current' period is already promised to the current tasks,
4519 * however the extra weight of the new task will slow them down a
4520 * little, place the new task so that it fits in the slot that
4521 * stays open at the end.
4523 if (initial && sched_feat(START_DEBIT))
4524 vruntime += sched_vslice(cfs_rq, se);
4526 /* sleeps up to a single latency don't count. */
4528 unsigned long thresh;
4531 thresh = sysctl_sched_min_granularity;
4533 thresh = sysctl_sched_latency;
4536 * Halve their sleep time's effect, to allow
4537 * for a gentler effect of sleepers:
4539 if (sched_feat(GENTLE_FAIR_SLEEPERS))
4545 /* ensure we never gain time by being placed backwards. */
4546 se->vruntime = max_vruntime(se->vruntime, vruntime);
4549 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
4551 static inline bool cfs_bandwidth_used(void);
4558 * update_min_vruntime()
4559 * vruntime -= min_vruntime
4563 * update_min_vruntime()
4564 * vruntime += min_vruntime
4566 * this way the vruntime transition between RQs is done when both
4567 * min_vruntime are up-to-date.
4571 * ->migrate_task_rq_fair() (p->state == TASK_WAKING)
4572 * vruntime -= min_vruntime
4576 * update_min_vruntime()
4577 * vruntime += min_vruntime
4579 * this way we don't have the most up-to-date min_vruntime on the originating
4580 * CPU and an up-to-date min_vruntime on the destination CPU.
4584 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
4586 bool renorm = !(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_MIGRATED);
4587 bool curr = cfs_rq->curr == se;
4590 * If we're the current task, we must renormalise before calling
4594 se->vruntime += cfs_rq->min_vruntime;
4596 update_curr(cfs_rq);
4599 * Otherwise, renormalise after, such that we're placed at the current
4600 * moment in time, instead of some random moment in the past. Being
4601 * placed in the past could significantly boost this task to the
4602 * fairness detriment of existing tasks.
4604 if (renorm && !curr)
4605 se->vruntime += cfs_rq->min_vruntime;
4608 * When enqueuing a sched_entity, we must:
4609 * - Update loads to have both entity and cfs_rq synced with now.
4610 * - For group_entity, update its runnable_weight to reflect the new
4611 * h_nr_running of its group cfs_rq.
4612 * - For group_entity, update its weight to reflect the new share of
4614 * - Add its new weight to cfs_rq->load.weight
4616 update_load_avg(cfs_rq, se, UPDATE_TG | DO_ATTACH);
4617 se_update_runnable(se);
4618 update_cfs_group(se);
4619 account_entity_enqueue(cfs_rq, se);
4621 if (flags & ENQUEUE_WAKEUP)
4622 place_entity(cfs_rq, se, 0);
4624 check_schedstat_required();
4625 update_stats_enqueue_fair(cfs_rq, se, flags);
4626 check_spread(cfs_rq, se);
4628 __enqueue_entity(cfs_rq, se);
4631 if (cfs_rq->nr_running == 1) {
4632 check_enqueue_throttle(cfs_rq);
4633 if (!throttled_hierarchy(cfs_rq))
4634 list_add_leaf_cfs_rq(cfs_rq);
4638 static void __clear_buddies_last(struct sched_entity *se)
4640 for_each_sched_entity(se) {
4641 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4642 if (cfs_rq->last != se)
4645 cfs_rq->last = NULL;
4649 static void __clear_buddies_next(struct sched_entity *se)
4651 for_each_sched_entity(se) {
4652 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4653 if (cfs_rq->next != se)
4656 cfs_rq->next = NULL;
4660 static void __clear_buddies_skip(struct sched_entity *se)
4662 for_each_sched_entity(se) {
4663 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4664 if (cfs_rq->skip != se)
4667 cfs_rq->skip = NULL;
4671 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
4673 if (cfs_rq->last == se)
4674 __clear_buddies_last(se);
4676 if (cfs_rq->next == se)
4677 __clear_buddies_next(se);
4679 if (cfs_rq->skip == se)
4680 __clear_buddies_skip(se);
4683 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
4686 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
4688 int action = UPDATE_TG;
4690 if (entity_is_task(se) && task_on_rq_migrating(task_of(se)))
4691 action |= DO_DETACH;
4694 * Update run-time statistics of the 'current'.
4696 update_curr(cfs_rq);
4699 * When dequeuing a sched_entity, we must:
4700 * - Update loads to have both entity and cfs_rq synced with now.
4701 * - For group_entity, update its runnable_weight to reflect the new
4702 * h_nr_running of its group cfs_rq.
4703 * - Subtract its previous weight from cfs_rq->load.weight.
4704 * - For group entity, update its weight to reflect the new share
4705 * of its group cfs_rq.
4707 update_load_avg(cfs_rq, se, action);
4708 se_update_runnable(se);
4710 update_stats_dequeue_fair(cfs_rq, se, flags);
4712 clear_buddies(cfs_rq, se);
4714 if (se != cfs_rq->curr)
4715 __dequeue_entity(cfs_rq, se);
4717 account_entity_dequeue(cfs_rq, se);
4720 * Normalize after update_curr(); which will also have moved
4721 * min_vruntime if @se is the one holding it back. But before doing
4722 * update_min_vruntime() again, which will discount @se's position and
4723 * can move min_vruntime forward still more.
4725 if (!(flags & DEQUEUE_SLEEP))
4726 se->vruntime -= cfs_rq->min_vruntime;
4728 /* return excess runtime on last dequeue */
4729 return_cfs_rq_runtime(cfs_rq);
4731 update_cfs_group(se);
4734 * Now advance min_vruntime if @se was the entity holding it back,
4735 * except when: DEQUEUE_SAVE && !DEQUEUE_MOVE, in this case we'll be
4736 * put back on, and if we advance min_vruntime, we'll be placed back
4737 * further than we started -- ie. we'll be penalized.
4739 if ((flags & (DEQUEUE_SAVE | DEQUEUE_MOVE)) != DEQUEUE_SAVE)
4740 update_min_vruntime(cfs_rq);
4742 if (cfs_rq->nr_running == 0)
4743 update_idle_cfs_rq_clock_pelt(cfs_rq);
4747 * Preempt the current task with a newly woken task if needed:
4750 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
4752 unsigned long ideal_runtime, delta_exec;
4753 struct sched_entity *se;
4756 ideal_runtime = sched_slice(cfs_rq, curr);
4757 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
4758 if (delta_exec > ideal_runtime) {
4759 resched_curr(rq_of(cfs_rq));
4761 * The current task ran long enough, ensure it doesn't get
4762 * re-elected due to buddy favours.
4764 clear_buddies(cfs_rq, curr);
4769 * Ensure that a task that missed wakeup preemption by a
4770 * narrow margin doesn't have to wait for a full slice.
4771 * This also mitigates buddy induced latencies under load.
4773 if (delta_exec < sysctl_sched_min_granularity)
4776 se = __pick_first_entity(cfs_rq);
4777 delta = curr->vruntime - se->vruntime;
4782 if (delta > ideal_runtime)
4783 resched_curr(rq_of(cfs_rq));
4787 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
4789 clear_buddies(cfs_rq, se);
4791 /* 'current' is not kept within the tree. */
4794 * Any task has to be enqueued before it get to execute on
4795 * a CPU. So account for the time it spent waiting on the
4798 update_stats_wait_end_fair(cfs_rq, se);
4799 __dequeue_entity(cfs_rq, se);
4800 update_load_avg(cfs_rq, se, UPDATE_TG);
4803 update_stats_curr_start(cfs_rq, se);
4807 * Track our maximum slice length, if the CPU's load is at
4808 * least twice that of our own weight (i.e. dont track it
4809 * when there are only lesser-weight tasks around):
4811 if (schedstat_enabled() &&
4812 rq_of(cfs_rq)->cfs.load.weight >= 2*se->load.weight) {
4813 struct sched_statistics *stats;
4815 stats = __schedstats_from_se(se);
4816 __schedstat_set(stats->slice_max,
4817 max((u64)stats->slice_max,
4818 se->sum_exec_runtime - se->prev_sum_exec_runtime));
4821 se->prev_sum_exec_runtime = se->sum_exec_runtime;
4825 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
4828 * Pick the next process, keeping these things in mind, in this order:
4829 * 1) keep things fair between processes/task groups
4830 * 2) pick the "next" process, since someone really wants that to run
4831 * 3) pick the "last" process, for cache locality
4832 * 4) do not run the "skip" process, if something else is available
4834 static struct sched_entity *
4835 pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr)
4837 struct sched_entity *left = __pick_first_entity(cfs_rq);
4838 struct sched_entity *se;
4841 * If curr is set we have to see if its left of the leftmost entity
4842 * still in the tree, provided there was anything in the tree at all.
4844 if (!left || (curr && entity_before(curr, left)))
4847 se = left; /* ideally we run the leftmost entity */
4850 * Avoid running the skip buddy, if running something else can
4851 * be done without getting too unfair.
4853 if (cfs_rq->skip && cfs_rq->skip == se) {
4854 struct sched_entity *second;
4857 second = __pick_first_entity(cfs_rq);
4859 second = __pick_next_entity(se);
4860 if (!second || (curr && entity_before(curr, second)))
4864 if (second && wakeup_preempt_entity(second, left) < 1)
4868 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1) {
4870 * Someone really wants this to run. If it's not unfair, run it.
4873 } else if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1) {
4875 * Prefer last buddy, try to return the CPU to a preempted task.
4883 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
4885 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
4888 * If still on the runqueue then deactivate_task()
4889 * was not called and update_curr() has to be done:
4892 update_curr(cfs_rq);
4894 /* throttle cfs_rqs exceeding runtime */
4895 check_cfs_rq_runtime(cfs_rq);
4897 check_spread(cfs_rq, prev);
4900 update_stats_wait_start_fair(cfs_rq, prev);
4901 /* Put 'current' back into the tree. */
4902 __enqueue_entity(cfs_rq, prev);
4903 /* in !on_rq case, update occurred at dequeue */
4904 update_load_avg(cfs_rq, prev, 0);
4906 cfs_rq->curr = NULL;
4910 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
4913 * Update run-time statistics of the 'current'.
4915 update_curr(cfs_rq);
4918 * Ensure that runnable average is periodically updated.
4920 update_load_avg(cfs_rq, curr, UPDATE_TG);
4921 update_cfs_group(curr);
4923 #ifdef CONFIG_SCHED_HRTICK
4925 * queued ticks are scheduled to match the slice, so don't bother
4926 * validating it and just reschedule.
4929 resched_curr(rq_of(cfs_rq));
4933 * don't let the period tick interfere with the hrtick preemption
4935 if (!sched_feat(DOUBLE_TICK) &&
4936 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
4940 if (cfs_rq->nr_running > 1)
4941 check_preempt_tick(cfs_rq, curr);
4945 /**************************************************
4946 * CFS bandwidth control machinery
4949 #ifdef CONFIG_CFS_BANDWIDTH
4951 #ifdef CONFIG_JUMP_LABEL
4952 static struct static_key __cfs_bandwidth_used;
4954 static inline bool cfs_bandwidth_used(void)
4956 return static_key_false(&__cfs_bandwidth_used);
4959 void cfs_bandwidth_usage_inc(void)
4961 static_key_slow_inc_cpuslocked(&__cfs_bandwidth_used);
4964 void cfs_bandwidth_usage_dec(void)
4966 static_key_slow_dec_cpuslocked(&__cfs_bandwidth_used);
4968 #else /* CONFIG_JUMP_LABEL */
4969 static bool cfs_bandwidth_used(void)
4974 void cfs_bandwidth_usage_inc(void) {}
4975 void cfs_bandwidth_usage_dec(void) {}
4976 #endif /* CONFIG_JUMP_LABEL */
4979 * default period for cfs group bandwidth.
4980 * default: 0.1s, units: nanoseconds
4982 static inline u64 default_cfs_period(void)
4984 return 100000000ULL;
4987 static inline u64 sched_cfs_bandwidth_slice(void)
4989 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
4993 * Replenish runtime according to assigned quota. We use sched_clock_cpu
4994 * directly instead of rq->clock to avoid adding additional synchronization
4997 * requires cfs_b->lock
4999 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
5003 if (unlikely(cfs_b->quota == RUNTIME_INF))
5006 cfs_b->runtime += cfs_b->quota;
5007 runtime = cfs_b->runtime_snap - cfs_b->runtime;
5009 cfs_b->burst_time += runtime;
5013 cfs_b->runtime = min(cfs_b->runtime, cfs_b->quota + cfs_b->burst);
5014 cfs_b->runtime_snap = cfs_b->runtime;
5017 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
5019 return &tg->cfs_bandwidth;
5022 /* returns 0 on failure to allocate runtime */
5023 static int __assign_cfs_rq_runtime(struct cfs_bandwidth *cfs_b,
5024 struct cfs_rq *cfs_rq, u64 target_runtime)
5026 u64 min_amount, amount = 0;
5028 lockdep_assert_held(&cfs_b->lock);
5030 /* note: this is a positive sum as runtime_remaining <= 0 */
5031 min_amount = target_runtime - cfs_rq->runtime_remaining;
5033 if (cfs_b->quota == RUNTIME_INF)
5034 amount = min_amount;
5036 start_cfs_bandwidth(cfs_b);
5038 if (cfs_b->runtime > 0) {
5039 amount = min(cfs_b->runtime, min_amount);
5040 cfs_b->runtime -= amount;
5045 cfs_rq->runtime_remaining += amount;
5047 return cfs_rq->runtime_remaining > 0;
5050 /* returns 0 on failure to allocate runtime */
5051 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
5053 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
5056 raw_spin_lock(&cfs_b->lock);
5057 ret = __assign_cfs_rq_runtime(cfs_b, cfs_rq, sched_cfs_bandwidth_slice());
5058 raw_spin_unlock(&cfs_b->lock);
5063 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
5065 /* dock delta_exec before expiring quota (as it could span periods) */
5066 cfs_rq->runtime_remaining -= delta_exec;
5068 if (likely(cfs_rq->runtime_remaining > 0))
5071 if (cfs_rq->throttled)
5074 * if we're unable to extend our runtime we resched so that the active
5075 * hierarchy can be throttled
5077 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
5078 resched_curr(rq_of(cfs_rq));
5081 static __always_inline
5082 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
5084 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
5087 __account_cfs_rq_runtime(cfs_rq, delta_exec);
5090 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
5092 return cfs_bandwidth_used() && cfs_rq->throttled;
5095 /* check whether cfs_rq, or any parent, is throttled */
5096 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
5098 return cfs_bandwidth_used() && cfs_rq->throttle_count;
5102 * Ensure that neither of the group entities corresponding to src_cpu or
5103 * dest_cpu are members of a throttled hierarchy when performing group
5104 * load-balance operations.
5106 static inline int throttled_lb_pair(struct task_group *tg,
5107 int src_cpu, int dest_cpu)
5109 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
5111 src_cfs_rq = tg->cfs_rq[src_cpu];
5112 dest_cfs_rq = tg->cfs_rq[dest_cpu];
5114 return throttled_hierarchy(src_cfs_rq) ||
5115 throttled_hierarchy(dest_cfs_rq);
5118 static int tg_unthrottle_up(struct task_group *tg, void *data)
5120 struct rq *rq = data;
5121 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
5123 cfs_rq->throttle_count--;
5124 if (!cfs_rq->throttle_count) {
5125 cfs_rq->throttled_clock_pelt_time += rq_clock_pelt(rq) -
5126 cfs_rq->throttled_clock_pelt;
5128 /* Add cfs_rq with load or one or more already running entities to the list */
5129 if (!cfs_rq_is_decayed(cfs_rq))
5130 list_add_leaf_cfs_rq(cfs_rq);
5136 static int tg_throttle_down(struct task_group *tg, void *data)
5138 struct rq *rq = data;
5139 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
5141 /* group is entering throttled state, stop time */
5142 if (!cfs_rq->throttle_count) {
5143 cfs_rq->throttled_clock_pelt = rq_clock_pelt(rq);
5144 list_del_leaf_cfs_rq(cfs_rq);
5146 cfs_rq->throttle_count++;
5151 static bool throttle_cfs_rq(struct cfs_rq *cfs_rq)
5153 struct rq *rq = rq_of(cfs_rq);
5154 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
5155 struct sched_entity *se;
5156 long task_delta, idle_task_delta, dequeue = 1;
5158 raw_spin_lock(&cfs_b->lock);
5159 /* This will start the period timer if necessary */
5160 if (__assign_cfs_rq_runtime(cfs_b, cfs_rq, 1)) {
5162 * We have raced with bandwidth becoming available, and if we
5163 * actually throttled the timer might not unthrottle us for an
5164 * entire period. We additionally needed to make sure that any
5165 * subsequent check_cfs_rq_runtime calls agree not to throttle
5166 * us, as we may commit to do cfs put_prev+pick_next, so we ask
5167 * for 1ns of runtime rather than just check cfs_b.
5171 list_add_tail_rcu(&cfs_rq->throttled_list,
5172 &cfs_b->throttled_cfs_rq);
5174 raw_spin_unlock(&cfs_b->lock);
5177 return false; /* Throttle no longer required. */
5179 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
5181 /* freeze hierarchy runnable averages while throttled */
5183 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
5186 task_delta = cfs_rq->h_nr_running;
5187 idle_task_delta = cfs_rq->idle_h_nr_running;
5188 for_each_sched_entity(se) {
5189 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
5190 /* throttled entity or throttle-on-deactivate */
5194 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
5196 if (cfs_rq_is_idle(group_cfs_rq(se)))
5197 idle_task_delta = cfs_rq->h_nr_running;
5199 qcfs_rq->h_nr_running -= task_delta;
5200 qcfs_rq->idle_h_nr_running -= idle_task_delta;
5202 if (qcfs_rq->load.weight) {
5203 /* Avoid re-evaluating load for this entity: */
5204 se = parent_entity(se);
5209 for_each_sched_entity(se) {
5210 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
5211 /* throttled entity or throttle-on-deactivate */
5215 update_load_avg(qcfs_rq, se, 0);
5216 se_update_runnable(se);
5218 if (cfs_rq_is_idle(group_cfs_rq(se)))
5219 idle_task_delta = cfs_rq->h_nr_running;
5221 qcfs_rq->h_nr_running -= task_delta;
5222 qcfs_rq->idle_h_nr_running -= idle_task_delta;
5225 /* At this point se is NULL and we are at root level*/
5226 sub_nr_running(rq, task_delta);
5230 * Note: distribution will already see us throttled via the
5231 * throttled-list. rq->lock protects completion.
5233 cfs_rq->throttled = 1;
5234 cfs_rq->throttled_clock = rq_clock(rq);
5238 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
5240 struct rq *rq = rq_of(cfs_rq);
5241 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
5242 struct sched_entity *se;
5243 long task_delta, idle_task_delta;
5245 se = cfs_rq->tg->se[cpu_of(rq)];
5247 cfs_rq->throttled = 0;
5249 update_rq_clock(rq);
5251 raw_spin_lock(&cfs_b->lock);
5252 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
5253 list_del_rcu(&cfs_rq->throttled_list);
5254 raw_spin_unlock(&cfs_b->lock);
5256 /* update hierarchical throttle state */
5257 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
5259 if (!cfs_rq->load.weight) {
5260 if (!cfs_rq->on_list)
5263 * Nothing to run but something to decay (on_list)?
5264 * Complete the branch.
5266 for_each_sched_entity(se) {
5267 if (list_add_leaf_cfs_rq(cfs_rq_of(se)))
5270 goto unthrottle_throttle;
5273 task_delta = cfs_rq->h_nr_running;
5274 idle_task_delta = cfs_rq->idle_h_nr_running;
5275 for_each_sched_entity(se) {
5276 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
5280 enqueue_entity(qcfs_rq, se, ENQUEUE_WAKEUP);
5282 if (cfs_rq_is_idle(group_cfs_rq(se)))
5283 idle_task_delta = cfs_rq->h_nr_running;
5285 qcfs_rq->h_nr_running += task_delta;
5286 qcfs_rq->idle_h_nr_running += idle_task_delta;
5288 /* end evaluation on encountering a throttled cfs_rq */
5289 if (cfs_rq_throttled(qcfs_rq))
5290 goto unthrottle_throttle;
5293 for_each_sched_entity(se) {
5294 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
5296 update_load_avg(qcfs_rq, se, UPDATE_TG);
5297 se_update_runnable(se);
5299 if (cfs_rq_is_idle(group_cfs_rq(se)))
5300 idle_task_delta = cfs_rq->h_nr_running;
5302 qcfs_rq->h_nr_running += task_delta;
5303 qcfs_rq->idle_h_nr_running += idle_task_delta;
5305 /* end evaluation on encountering a throttled cfs_rq */
5306 if (cfs_rq_throttled(qcfs_rq))
5307 goto unthrottle_throttle;
5310 /* At this point se is NULL and we are at root level*/
5311 add_nr_running(rq, task_delta);
5313 unthrottle_throttle:
5314 assert_list_leaf_cfs_rq(rq);
5316 /* Determine whether we need to wake up potentially idle CPU: */
5317 if (rq->curr == rq->idle && rq->cfs.nr_running)
5321 static void distribute_cfs_runtime(struct cfs_bandwidth *cfs_b)
5323 struct cfs_rq *cfs_rq;
5324 u64 runtime, remaining = 1;
5327 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
5329 struct rq *rq = rq_of(cfs_rq);
5332 rq_lock_irqsave(rq, &rf);
5333 if (!cfs_rq_throttled(cfs_rq))
5336 /* By the above check, this should never be true */
5337 SCHED_WARN_ON(cfs_rq->runtime_remaining > 0);
5339 raw_spin_lock(&cfs_b->lock);
5340 runtime = -cfs_rq->runtime_remaining + 1;
5341 if (runtime > cfs_b->runtime)
5342 runtime = cfs_b->runtime;
5343 cfs_b->runtime -= runtime;
5344 remaining = cfs_b->runtime;
5345 raw_spin_unlock(&cfs_b->lock);
5347 cfs_rq->runtime_remaining += runtime;
5349 /* we check whether we're throttled above */
5350 if (cfs_rq->runtime_remaining > 0)
5351 unthrottle_cfs_rq(cfs_rq);
5354 rq_unlock_irqrestore(rq, &rf);
5363 * Responsible for refilling a task_group's bandwidth and unthrottling its
5364 * cfs_rqs as appropriate. If there has been no activity within the last
5365 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
5366 * used to track this state.
5368 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun, unsigned long flags)
5372 /* no need to continue the timer with no bandwidth constraint */
5373 if (cfs_b->quota == RUNTIME_INF)
5374 goto out_deactivate;
5376 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
5377 cfs_b->nr_periods += overrun;
5379 /* Refill extra burst quota even if cfs_b->idle */
5380 __refill_cfs_bandwidth_runtime(cfs_b);
5383 * idle depends on !throttled (for the case of a large deficit), and if
5384 * we're going inactive then everything else can be deferred
5386 if (cfs_b->idle && !throttled)
5387 goto out_deactivate;
5390 /* mark as potentially idle for the upcoming period */
5395 /* account preceding periods in which throttling occurred */
5396 cfs_b->nr_throttled += overrun;
5399 * This check is repeated as we release cfs_b->lock while we unthrottle.
5401 while (throttled && cfs_b->runtime > 0) {
5402 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
5403 /* we can't nest cfs_b->lock while distributing bandwidth */
5404 distribute_cfs_runtime(cfs_b);
5405 raw_spin_lock_irqsave(&cfs_b->lock, flags);
5407 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
5411 * While we are ensured activity in the period following an
5412 * unthrottle, this also covers the case in which the new bandwidth is
5413 * insufficient to cover the existing bandwidth deficit. (Forcing the
5414 * timer to remain active while there are any throttled entities.)
5424 /* a cfs_rq won't donate quota below this amount */
5425 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
5426 /* minimum remaining period time to redistribute slack quota */
5427 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
5428 /* how long we wait to gather additional slack before distributing */
5429 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
5432 * Are we near the end of the current quota period?
5434 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
5435 * hrtimer base being cleared by hrtimer_start. In the case of
5436 * migrate_hrtimers, base is never cleared, so we are fine.
5438 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
5440 struct hrtimer *refresh_timer = &cfs_b->period_timer;
5443 /* if the call-back is running a quota refresh is already occurring */
5444 if (hrtimer_callback_running(refresh_timer))
5447 /* is a quota refresh about to occur? */
5448 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
5449 if (remaining < (s64)min_expire)
5455 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
5457 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
5459 /* if there's a quota refresh soon don't bother with slack */
5460 if (runtime_refresh_within(cfs_b, min_left))
5463 /* don't push forwards an existing deferred unthrottle */
5464 if (cfs_b->slack_started)
5466 cfs_b->slack_started = true;
5468 hrtimer_start(&cfs_b->slack_timer,
5469 ns_to_ktime(cfs_bandwidth_slack_period),
5473 /* we know any runtime found here is valid as update_curr() precedes return */
5474 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
5476 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
5477 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
5479 if (slack_runtime <= 0)
5482 raw_spin_lock(&cfs_b->lock);
5483 if (cfs_b->quota != RUNTIME_INF) {
5484 cfs_b->runtime += slack_runtime;
5486 /* we are under rq->lock, defer unthrottling using a timer */
5487 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
5488 !list_empty(&cfs_b->throttled_cfs_rq))
5489 start_cfs_slack_bandwidth(cfs_b);
5491 raw_spin_unlock(&cfs_b->lock);
5493 /* even if it's not valid for return we don't want to try again */
5494 cfs_rq->runtime_remaining -= slack_runtime;
5497 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
5499 if (!cfs_bandwidth_used())
5502 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
5505 __return_cfs_rq_runtime(cfs_rq);
5509 * This is done with a timer (instead of inline with bandwidth return) since
5510 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
5512 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
5514 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
5515 unsigned long flags;
5517 /* confirm we're still not at a refresh boundary */
5518 raw_spin_lock_irqsave(&cfs_b->lock, flags);
5519 cfs_b->slack_started = false;
5521 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
5522 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
5526 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice)
5527 runtime = cfs_b->runtime;
5529 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
5534 distribute_cfs_runtime(cfs_b);
5538 * When a group wakes up we want to make sure that its quota is not already
5539 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
5540 * runtime as update_curr() throttling can not trigger until it's on-rq.
5542 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
5544 if (!cfs_bandwidth_used())
5547 /* an active group must be handled by the update_curr()->put() path */
5548 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
5551 /* ensure the group is not already throttled */
5552 if (cfs_rq_throttled(cfs_rq))
5555 /* update runtime allocation */
5556 account_cfs_rq_runtime(cfs_rq, 0);
5557 if (cfs_rq->runtime_remaining <= 0)
5558 throttle_cfs_rq(cfs_rq);
5561 static void sync_throttle(struct task_group *tg, int cpu)
5563 struct cfs_rq *pcfs_rq, *cfs_rq;
5565 if (!cfs_bandwidth_used())
5571 cfs_rq = tg->cfs_rq[cpu];
5572 pcfs_rq = tg->parent->cfs_rq[cpu];
5574 cfs_rq->throttle_count = pcfs_rq->throttle_count;
5575 cfs_rq->throttled_clock_pelt = rq_clock_pelt(cpu_rq(cpu));
5578 /* conditionally throttle active cfs_rq's from put_prev_entity() */
5579 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
5581 if (!cfs_bandwidth_used())
5584 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
5588 * it's possible for a throttled entity to be forced into a running
5589 * state (e.g. set_curr_task), in this case we're finished.
5591 if (cfs_rq_throttled(cfs_rq))
5594 return throttle_cfs_rq(cfs_rq);
5597 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
5599 struct cfs_bandwidth *cfs_b =
5600 container_of(timer, struct cfs_bandwidth, slack_timer);
5602 do_sched_cfs_slack_timer(cfs_b);
5604 return HRTIMER_NORESTART;
5607 extern const u64 max_cfs_quota_period;
5609 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
5611 struct cfs_bandwidth *cfs_b =
5612 container_of(timer, struct cfs_bandwidth, period_timer);
5613 unsigned long flags;
5618 raw_spin_lock_irqsave(&cfs_b->lock, flags);
5620 overrun = hrtimer_forward_now(timer, cfs_b->period);
5624 idle = do_sched_cfs_period_timer(cfs_b, overrun, flags);
5627 u64 new, old = ktime_to_ns(cfs_b->period);
5630 * Grow period by a factor of 2 to avoid losing precision.
5631 * Precision loss in the quota/period ratio can cause __cfs_schedulable
5635 if (new < max_cfs_quota_period) {
5636 cfs_b->period = ns_to_ktime(new);
5640 pr_warn_ratelimited(
5641 "cfs_period_timer[cpu%d]: period too short, scaling up (new cfs_period_us = %lld, cfs_quota_us = %lld)\n",
5643 div_u64(new, NSEC_PER_USEC),
5644 div_u64(cfs_b->quota, NSEC_PER_USEC));
5646 pr_warn_ratelimited(
5647 "cfs_period_timer[cpu%d]: period too short, but cannot scale up without losing precision (cfs_period_us = %lld, cfs_quota_us = %lld)\n",
5649 div_u64(old, NSEC_PER_USEC),
5650 div_u64(cfs_b->quota, NSEC_PER_USEC));
5653 /* reset count so we don't come right back in here */
5658 cfs_b->period_active = 0;
5659 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
5661 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
5664 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
5666 raw_spin_lock_init(&cfs_b->lock);
5668 cfs_b->quota = RUNTIME_INF;
5669 cfs_b->period = ns_to_ktime(default_cfs_period());
5672 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
5673 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED);
5674 cfs_b->period_timer.function = sched_cfs_period_timer;
5675 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
5676 cfs_b->slack_timer.function = sched_cfs_slack_timer;
5677 cfs_b->slack_started = false;
5680 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
5682 cfs_rq->runtime_enabled = 0;
5683 INIT_LIST_HEAD(&cfs_rq->throttled_list);
5686 void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
5688 lockdep_assert_held(&cfs_b->lock);
5690 if (cfs_b->period_active)
5693 cfs_b->period_active = 1;
5694 hrtimer_forward_now(&cfs_b->period_timer, cfs_b->period);
5695 hrtimer_start_expires(&cfs_b->period_timer, HRTIMER_MODE_ABS_PINNED);
5698 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
5700 /* init_cfs_bandwidth() was not called */
5701 if (!cfs_b->throttled_cfs_rq.next)
5704 hrtimer_cancel(&cfs_b->period_timer);
5705 hrtimer_cancel(&cfs_b->slack_timer);
5709 * Both these CPU hotplug callbacks race against unregister_fair_sched_group()
5711 * The race is harmless, since modifying bandwidth settings of unhooked group
5712 * bits doesn't do much.
5715 /* cpu online callback */
5716 static void __maybe_unused update_runtime_enabled(struct rq *rq)
5718 struct task_group *tg;
5720 lockdep_assert_rq_held(rq);
5723 list_for_each_entry_rcu(tg, &task_groups, list) {
5724 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
5725 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
5727 raw_spin_lock(&cfs_b->lock);
5728 cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF;
5729 raw_spin_unlock(&cfs_b->lock);
5734 /* cpu offline callback */
5735 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
5737 struct task_group *tg;
5739 lockdep_assert_rq_held(rq);
5742 list_for_each_entry_rcu(tg, &task_groups, list) {
5743 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
5745 if (!cfs_rq->runtime_enabled)
5749 * clock_task is not advancing so we just need to make sure
5750 * there's some valid quota amount
5752 cfs_rq->runtime_remaining = 1;
5754 * Offline rq is schedulable till CPU is completely disabled
5755 * in take_cpu_down(), so we prevent new cfs throttling here.
5757 cfs_rq->runtime_enabled = 0;
5759 if (cfs_rq_throttled(cfs_rq))
5760 unthrottle_cfs_rq(cfs_rq);
5765 #else /* CONFIG_CFS_BANDWIDTH */
5767 static inline bool cfs_bandwidth_used(void)
5772 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {}
5773 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; }
5774 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
5775 static inline void sync_throttle(struct task_group *tg, int cpu) {}
5776 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
5778 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
5783 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
5788 static inline int throttled_lb_pair(struct task_group *tg,
5789 int src_cpu, int dest_cpu)
5794 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
5796 #ifdef CONFIG_FAIR_GROUP_SCHED
5797 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
5800 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
5804 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
5805 static inline void update_runtime_enabled(struct rq *rq) {}
5806 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
5808 #endif /* CONFIG_CFS_BANDWIDTH */
5810 /**************************************************
5811 * CFS operations on tasks:
5814 #ifdef CONFIG_SCHED_HRTICK
5815 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
5817 struct sched_entity *se = &p->se;
5818 struct cfs_rq *cfs_rq = cfs_rq_of(se);
5820 SCHED_WARN_ON(task_rq(p) != rq);
5822 if (rq->cfs.h_nr_running > 1) {
5823 u64 slice = sched_slice(cfs_rq, se);
5824 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
5825 s64 delta = slice - ran;
5828 if (task_current(rq, p))
5832 hrtick_start(rq, delta);
5837 * called from enqueue/dequeue and updates the hrtick when the
5838 * current task is from our class and nr_running is low enough
5841 static void hrtick_update(struct rq *rq)
5843 struct task_struct *curr = rq->curr;
5845 if (!hrtick_enabled_fair(rq) || curr->sched_class != &fair_sched_class)
5848 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
5849 hrtick_start_fair(rq, curr);
5851 #else /* !CONFIG_SCHED_HRTICK */
5853 hrtick_start_fair(struct rq *rq, struct task_struct *p)
5857 static inline void hrtick_update(struct rq *rq)
5863 static inline bool cpu_overutilized(int cpu)
5865 return !fits_capacity(cpu_util_cfs(cpu), capacity_of(cpu));
5868 static inline void update_overutilized_status(struct rq *rq)
5870 if (!READ_ONCE(rq->rd->overutilized) && cpu_overutilized(rq->cpu)) {
5871 WRITE_ONCE(rq->rd->overutilized, SG_OVERUTILIZED);
5872 trace_sched_overutilized_tp(rq->rd, SG_OVERUTILIZED);
5876 static inline void update_overutilized_status(struct rq *rq) { }
5879 /* Runqueue only has SCHED_IDLE tasks enqueued */
5880 static int sched_idle_rq(struct rq *rq)
5882 return unlikely(rq->nr_running == rq->cfs.idle_h_nr_running &&
5887 * Returns true if cfs_rq only has SCHED_IDLE entities enqueued. Note the use
5888 * of idle_nr_running, which does not consider idle descendants of normal
5891 static bool sched_idle_cfs_rq(struct cfs_rq *cfs_rq)
5893 return cfs_rq->nr_running &&
5894 cfs_rq->nr_running == cfs_rq->idle_nr_running;
5898 static int sched_idle_cpu(int cpu)
5900 return sched_idle_rq(cpu_rq(cpu));
5905 * The enqueue_task method is called before nr_running is
5906 * increased. Here we update the fair scheduling stats and
5907 * then put the task into the rbtree:
5910 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
5912 struct cfs_rq *cfs_rq;
5913 struct sched_entity *se = &p->se;
5914 int idle_h_nr_running = task_has_idle_policy(p);
5915 int task_new = !(flags & ENQUEUE_WAKEUP);
5918 * The code below (indirectly) updates schedutil which looks at
5919 * the cfs_rq utilization to select a frequency.
5920 * Let's add the task's estimated utilization to the cfs_rq's
5921 * estimated utilization, before we update schedutil.
5923 util_est_enqueue(&rq->cfs, p);
5926 * If in_iowait is set, the code below may not trigger any cpufreq
5927 * utilization updates, so do it here explicitly with the IOWAIT flag
5931 cpufreq_update_util(rq, SCHED_CPUFREQ_IOWAIT);
5933 for_each_sched_entity(se) {
5936 cfs_rq = cfs_rq_of(se);
5937 enqueue_entity(cfs_rq, se, flags);
5939 cfs_rq->h_nr_running++;
5940 cfs_rq->idle_h_nr_running += idle_h_nr_running;
5942 if (cfs_rq_is_idle(cfs_rq))
5943 idle_h_nr_running = 1;
5945 /* end evaluation on encountering a throttled cfs_rq */
5946 if (cfs_rq_throttled(cfs_rq))
5947 goto enqueue_throttle;
5949 flags = ENQUEUE_WAKEUP;
5952 for_each_sched_entity(se) {
5953 cfs_rq = cfs_rq_of(se);
5955 update_load_avg(cfs_rq, se, UPDATE_TG);
5956 se_update_runnable(se);
5957 update_cfs_group(se);
5959 cfs_rq->h_nr_running++;
5960 cfs_rq->idle_h_nr_running += idle_h_nr_running;
5962 if (cfs_rq_is_idle(cfs_rq))
5963 idle_h_nr_running = 1;
5965 /* end evaluation on encountering a throttled cfs_rq */
5966 if (cfs_rq_throttled(cfs_rq))
5967 goto enqueue_throttle;
5970 /* At this point se is NULL and we are at root level*/
5971 add_nr_running(rq, 1);
5974 * Since new tasks are assigned an initial util_avg equal to
5975 * half of the spare capacity of their CPU, tiny tasks have the
5976 * ability to cross the overutilized threshold, which will
5977 * result in the load balancer ruining all the task placement
5978 * done by EAS. As a way to mitigate that effect, do not account
5979 * for the first enqueue operation of new tasks during the
5980 * overutilized flag detection.
5982 * A better way of solving this problem would be to wait for
5983 * the PELT signals of tasks to converge before taking them
5984 * into account, but that is not straightforward to implement,
5985 * and the following generally works well enough in practice.
5988 update_overutilized_status(rq);
5991 assert_list_leaf_cfs_rq(rq);
5996 static void set_next_buddy(struct sched_entity *se);
5999 * The dequeue_task method is called before nr_running is
6000 * decreased. We remove the task from the rbtree and
6001 * update the fair scheduling stats:
6003 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
6005 struct cfs_rq *cfs_rq;
6006 struct sched_entity *se = &p->se;
6007 int task_sleep = flags & DEQUEUE_SLEEP;
6008 int idle_h_nr_running = task_has_idle_policy(p);
6009 bool was_sched_idle = sched_idle_rq(rq);
6011 util_est_dequeue(&rq->cfs, p);
6013 for_each_sched_entity(se) {
6014 cfs_rq = cfs_rq_of(se);
6015 dequeue_entity(cfs_rq, se, flags);
6017 cfs_rq->h_nr_running--;
6018 cfs_rq->idle_h_nr_running -= idle_h_nr_running;
6020 if (cfs_rq_is_idle(cfs_rq))
6021 idle_h_nr_running = 1;
6023 /* end evaluation on encountering a throttled cfs_rq */
6024 if (cfs_rq_throttled(cfs_rq))
6025 goto dequeue_throttle;
6027 /* Don't dequeue parent if it has other entities besides us */
6028 if (cfs_rq->load.weight) {
6029 /* Avoid re-evaluating load for this entity: */
6030 se = parent_entity(se);
6032 * Bias pick_next to pick a task from this cfs_rq, as
6033 * p is sleeping when it is within its sched_slice.
6035 if (task_sleep && se && !throttled_hierarchy(cfs_rq))
6039 flags |= DEQUEUE_SLEEP;
6042 for_each_sched_entity(se) {
6043 cfs_rq = cfs_rq_of(se);
6045 update_load_avg(cfs_rq, se, UPDATE_TG);
6046 se_update_runnable(se);
6047 update_cfs_group(se);
6049 cfs_rq->h_nr_running--;
6050 cfs_rq->idle_h_nr_running -= idle_h_nr_running;
6052 if (cfs_rq_is_idle(cfs_rq))
6053 idle_h_nr_running = 1;
6055 /* end evaluation on encountering a throttled cfs_rq */
6056 if (cfs_rq_throttled(cfs_rq))
6057 goto dequeue_throttle;
6061 /* At this point se is NULL and we are at root level*/
6062 sub_nr_running(rq, 1);
6064 /* balance early to pull high priority tasks */
6065 if (unlikely(!was_sched_idle && sched_idle_rq(rq)))
6066 rq->next_balance = jiffies;
6069 util_est_update(&rq->cfs, p, task_sleep);
6075 /* Working cpumask for: load_balance, load_balance_newidle. */
6076 static DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
6077 static DEFINE_PER_CPU(cpumask_var_t, select_rq_mask);
6079 #ifdef CONFIG_NO_HZ_COMMON
6082 cpumask_var_t idle_cpus_mask;
6084 int has_blocked; /* Idle CPUS has blocked load */
6085 int needs_update; /* Newly idle CPUs need their next_balance collated */
6086 unsigned long next_balance; /* in jiffy units */
6087 unsigned long next_blocked; /* Next update of blocked load in jiffies */
6088 } nohz ____cacheline_aligned;
6090 #endif /* CONFIG_NO_HZ_COMMON */
6092 static unsigned long cpu_load(struct rq *rq)
6094 return cfs_rq_load_avg(&rq->cfs);
6098 * cpu_load_without - compute CPU load without any contributions from *p
6099 * @cpu: the CPU which load is requested
6100 * @p: the task which load should be discounted
6102 * The load of a CPU is defined by the load of tasks currently enqueued on that
6103 * CPU as well as tasks which are currently sleeping after an execution on that
6106 * This method returns the load of the specified CPU by discounting the load of
6107 * the specified task, whenever the task is currently contributing to the CPU
6110 static unsigned long cpu_load_without(struct rq *rq, struct task_struct *p)
6112 struct cfs_rq *cfs_rq;
6115 /* Task has no contribution or is new */
6116 if (cpu_of(rq) != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time))
6117 return cpu_load(rq);
6120 load = READ_ONCE(cfs_rq->avg.load_avg);
6122 /* Discount task's util from CPU's util */
6123 lsub_positive(&load, task_h_load(p));
6128 static unsigned long cpu_runnable(struct rq *rq)
6130 return cfs_rq_runnable_avg(&rq->cfs);
6133 static unsigned long cpu_runnable_without(struct rq *rq, struct task_struct *p)
6135 struct cfs_rq *cfs_rq;
6136 unsigned int runnable;
6138 /* Task has no contribution or is new */
6139 if (cpu_of(rq) != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time))
6140 return cpu_runnable(rq);
6143 runnable = READ_ONCE(cfs_rq->avg.runnable_avg);
6145 /* Discount task's runnable from CPU's runnable */
6146 lsub_positive(&runnable, p->se.avg.runnable_avg);
6151 static unsigned long capacity_of(int cpu)
6153 return cpu_rq(cpu)->cpu_capacity;
6156 static void record_wakee(struct task_struct *p)
6159 * Only decay a single time; tasks that have less then 1 wakeup per
6160 * jiffy will not have built up many flips.
6162 if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) {
6163 current->wakee_flips >>= 1;
6164 current->wakee_flip_decay_ts = jiffies;
6167 if (current->last_wakee != p) {
6168 current->last_wakee = p;
6169 current->wakee_flips++;
6174 * Detect M:N waker/wakee relationships via a switching-frequency heuristic.
6176 * A waker of many should wake a different task than the one last awakened
6177 * at a frequency roughly N times higher than one of its wakees.
6179 * In order to determine whether we should let the load spread vs consolidating
6180 * to shared cache, we look for a minimum 'flip' frequency of llc_size in one
6181 * partner, and a factor of lls_size higher frequency in the other.
6183 * With both conditions met, we can be relatively sure that the relationship is
6184 * non-monogamous, with partner count exceeding socket size.
6186 * Waker/wakee being client/server, worker/dispatcher, interrupt source or
6187 * whatever is irrelevant, spread criteria is apparent partner count exceeds
6190 static int wake_wide(struct task_struct *p)
6192 unsigned int master = current->wakee_flips;
6193 unsigned int slave = p->wakee_flips;
6194 int factor = __this_cpu_read(sd_llc_size);
6197 swap(master, slave);
6198 if (slave < factor || master < slave * factor)
6204 * The purpose of wake_affine() is to quickly determine on which CPU we can run
6205 * soonest. For the purpose of speed we only consider the waking and previous
6208 * wake_affine_idle() - only considers 'now', it check if the waking CPU is
6209 * cache-affine and is (or will be) idle.
6211 * wake_affine_weight() - considers the weight to reflect the average
6212 * scheduling latency of the CPUs. This seems to work
6213 * for the overloaded case.
6216 wake_affine_idle(int this_cpu, int prev_cpu, int sync)
6219 * If this_cpu is idle, it implies the wakeup is from interrupt
6220 * context. Only allow the move if cache is shared. Otherwise an
6221 * interrupt intensive workload could force all tasks onto one
6222 * node depending on the IO topology or IRQ affinity settings.
6224 * If the prev_cpu is idle and cache affine then avoid a migration.
6225 * There is no guarantee that the cache hot data from an interrupt
6226 * is more important than cache hot data on the prev_cpu and from
6227 * a cpufreq perspective, it's better to have higher utilisation
6230 if (available_idle_cpu(this_cpu) && cpus_share_cache(this_cpu, prev_cpu))
6231 return available_idle_cpu(prev_cpu) ? prev_cpu : this_cpu;
6233 if (sync && cpu_rq(this_cpu)->nr_running == 1)
6236 if (available_idle_cpu(prev_cpu))
6239 return nr_cpumask_bits;
6243 wake_affine_weight(struct sched_domain *sd, struct task_struct *p,
6244 int this_cpu, int prev_cpu, int sync)
6246 s64 this_eff_load, prev_eff_load;
6247 unsigned long task_load;
6249 this_eff_load = cpu_load(cpu_rq(this_cpu));
6252 unsigned long current_load = task_h_load(current);
6254 if (current_load > this_eff_load)
6257 this_eff_load -= current_load;
6260 task_load = task_h_load(p);
6262 this_eff_load += task_load;
6263 if (sched_feat(WA_BIAS))
6264 this_eff_load *= 100;
6265 this_eff_load *= capacity_of(prev_cpu);
6267 prev_eff_load = cpu_load(cpu_rq(prev_cpu));
6268 prev_eff_load -= task_load;
6269 if (sched_feat(WA_BIAS))
6270 prev_eff_load *= 100 + (sd->imbalance_pct - 100) / 2;
6271 prev_eff_load *= capacity_of(this_cpu);
6274 * If sync, adjust the weight of prev_eff_load such that if
6275 * prev_eff == this_eff that select_idle_sibling() will consider
6276 * stacking the wakee on top of the waker if no other CPU is
6282 return this_eff_load < prev_eff_load ? this_cpu : nr_cpumask_bits;
6285 static int wake_affine(struct sched_domain *sd, struct task_struct *p,
6286 int this_cpu, int prev_cpu, int sync)
6288 int target = nr_cpumask_bits;
6290 if (sched_feat(WA_IDLE))
6291 target = wake_affine_idle(this_cpu, prev_cpu, sync);
6293 if (sched_feat(WA_WEIGHT) && target == nr_cpumask_bits)
6294 target = wake_affine_weight(sd, p, this_cpu, prev_cpu, sync);
6296 schedstat_inc(p->stats.nr_wakeups_affine_attempts);
6297 if (target == nr_cpumask_bits)
6300 schedstat_inc(sd->ttwu_move_affine);
6301 schedstat_inc(p->stats.nr_wakeups_affine);
6305 static struct sched_group *
6306 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu);
6309 * find_idlest_group_cpu - find the idlest CPU among the CPUs in the group.
6312 find_idlest_group_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
6314 unsigned long load, min_load = ULONG_MAX;
6315 unsigned int min_exit_latency = UINT_MAX;
6316 u64 latest_idle_timestamp = 0;
6317 int least_loaded_cpu = this_cpu;
6318 int shallowest_idle_cpu = -1;
6321 /* Check if we have any choice: */
6322 if (group->group_weight == 1)
6323 return cpumask_first(sched_group_span(group));
6325 /* Traverse only the allowed CPUs */
6326 for_each_cpu_and(i, sched_group_span(group), p->cpus_ptr) {
6327 struct rq *rq = cpu_rq(i);
6329 if (!sched_core_cookie_match(rq, p))
6332 if (sched_idle_cpu(i))
6335 if (available_idle_cpu(i)) {
6336 struct cpuidle_state *idle = idle_get_state(rq);
6337 if (idle && idle->exit_latency < min_exit_latency) {
6339 * We give priority to a CPU whose idle state
6340 * has the smallest exit latency irrespective
6341 * of any idle timestamp.
6343 min_exit_latency = idle->exit_latency;
6344 latest_idle_timestamp = rq->idle_stamp;
6345 shallowest_idle_cpu = i;
6346 } else if ((!idle || idle->exit_latency == min_exit_latency) &&
6347 rq->idle_stamp > latest_idle_timestamp) {
6349 * If equal or no active idle state, then
6350 * the most recently idled CPU might have
6353 latest_idle_timestamp = rq->idle_stamp;
6354 shallowest_idle_cpu = i;
6356 } else if (shallowest_idle_cpu == -1) {
6357 load = cpu_load(cpu_rq(i));
6358 if (load < min_load) {
6360 least_loaded_cpu = i;
6365 return shallowest_idle_cpu != -1 ? shallowest_idle_cpu : least_loaded_cpu;
6368 static inline int find_idlest_cpu(struct sched_domain *sd, struct task_struct *p,
6369 int cpu, int prev_cpu, int sd_flag)
6373 if (!cpumask_intersects(sched_domain_span(sd), p->cpus_ptr))
6377 * We need task's util for cpu_util_without, sync it up to
6378 * prev_cpu's last_update_time.
6380 if (!(sd_flag & SD_BALANCE_FORK))
6381 sync_entity_load_avg(&p->se);
6384 struct sched_group *group;
6385 struct sched_domain *tmp;
6388 if (!(sd->flags & sd_flag)) {
6393 group = find_idlest_group(sd, p, cpu);
6399 new_cpu = find_idlest_group_cpu(group, p, cpu);
6400 if (new_cpu == cpu) {
6401 /* Now try balancing at a lower domain level of 'cpu': */
6406 /* Now try balancing at a lower domain level of 'new_cpu': */
6408 weight = sd->span_weight;
6410 for_each_domain(cpu, tmp) {
6411 if (weight <= tmp->span_weight)
6413 if (tmp->flags & sd_flag)
6421 static inline int __select_idle_cpu(int cpu, struct task_struct *p)
6423 if ((available_idle_cpu(cpu) || sched_idle_cpu(cpu)) &&
6424 sched_cpu_cookie_match(cpu_rq(cpu), p))
6430 #ifdef CONFIG_SCHED_SMT
6431 DEFINE_STATIC_KEY_FALSE(sched_smt_present);
6432 EXPORT_SYMBOL_GPL(sched_smt_present);
6434 static inline void set_idle_cores(int cpu, int val)
6436 struct sched_domain_shared *sds;
6438 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
6440 WRITE_ONCE(sds->has_idle_cores, val);
6443 static inline bool test_idle_cores(int cpu)
6445 struct sched_domain_shared *sds;
6447 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
6449 return READ_ONCE(sds->has_idle_cores);
6455 * Scans the local SMT mask to see if the entire core is idle, and records this
6456 * information in sd_llc_shared->has_idle_cores.
6458 * Since SMT siblings share all cache levels, inspecting this limited remote
6459 * state should be fairly cheap.
6461 void __update_idle_core(struct rq *rq)
6463 int core = cpu_of(rq);
6467 if (test_idle_cores(core))
6470 for_each_cpu(cpu, cpu_smt_mask(core)) {
6474 if (!available_idle_cpu(cpu))
6478 set_idle_cores(core, 1);
6484 * Scan the entire LLC domain for idle cores; this dynamically switches off if
6485 * there are no idle cores left in the system; tracked through
6486 * sd_llc->shared->has_idle_cores and enabled through update_idle_core() above.
6488 static int select_idle_core(struct task_struct *p, int core, struct cpumask *cpus, int *idle_cpu)
6493 for_each_cpu(cpu, cpu_smt_mask(core)) {
6494 if (!available_idle_cpu(cpu)) {
6496 if (*idle_cpu == -1) {
6497 if (sched_idle_cpu(cpu) && cpumask_test_cpu(cpu, p->cpus_ptr)) {
6505 if (*idle_cpu == -1 && cpumask_test_cpu(cpu, p->cpus_ptr))
6512 cpumask_andnot(cpus, cpus, cpu_smt_mask(core));
6517 * Scan the local SMT mask for idle CPUs.
6519 static int select_idle_smt(struct task_struct *p, int target)
6523 for_each_cpu_and(cpu, cpu_smt_mask(target), p->cpus_ptr) {
6526 if (available_idle_cpu(cpu) || sched_idle_cpu(cpu))
6533 #else /* CONFIG_SCHED_SMT */
6535 static inline void set_idle_cores(int cpu, int val)
6539 static inline bool test_idle_cores(int cpu)
6544 static inline int select_idle_core(struct task_struct *p, int core, struct cpumask *cpus, int *idle_cpu)
6546 return __select_idle_cpu(core, p);
6549 static inline int select_idle_smt(struct task_struct *p, int target)
6554 #endif /* CONFIG_SCHED_SMT */
6557 * Scan the LLC domain for idle CPUs; this is dynamically regulated by
6558 * comparing the average scan cost (tracked in sd->avg_scan_cost) against the
6559 * average idle time for this rq (as found in rq->avg_idle).
6561 static int select_idle_cpu(struct task_struct *p, struct sched_domain *sd, bool has_idle_core, int target)
6563 struct cpumask *cpus = this_cpu_cpumask_var_ptr(select_rq_mask);
6564 int i, cpu, idle_cpu = -1, nr = INT_MAX;
6565 struct sched_domain_shared *sd_share;
6566 struct rq *this_rq = this_rq();
6567 int this = smp_processor_id();
6568 struct sched_domain *this_sd = NULL;
6571 cpumask_and(cpus, sched_domain_span(sd), p->cpus_ptr);
6573 if (sched_feat(SIS_PROP) && !has_idle_core) {
6574 u64 avg_cost, avg_idle, span_avg;
6575 unsigned long now = jiffies;
6577 this_sd = rcu_dereference(*this_cpu_ptr(&sd_llc));
6582 * If we're busy, the assumption that the last idle period
6583 * predicts the future is flawed; age away the remaining
6584 * predicted idle time.
6586 if (unlikely(this_rq->wake_stamp < now)) {
6587 while (this_rq->wake_stamp < now && this_rq->wake_avg_idle) {
6588 this_rq->wake_stamp++;
6589 this_rq->wake_avg_idle >>= 1;
6593 avg_idle = this_rq->wake_avg_idle;
6594 avg_cost = this_sd->avg_scan_cost + 1;
6596 span_avg = sd->span_weight * avg_idle;
6597 if (span_avg > 4*avg_cost)
6598 nr = div_u64(span_avg, avg_cost);
6602 time = cpu_clock(this);
6605 if (sched_feat(SIS_UTIL)) {
6606 sd_share = rcu_dereference(per_cpu(sd_llc_shared, target));
6608 /* because !--nr is the condition to stop scan */
6609 nr = READ_ONCE(sd_share->nr_idle_scan) + 1;
6610 /* overloaded LLC is unlikely to have idle cpu/core */
6616 for_each_cpu_wrap(cpu, cpus, target + 1) {
6617 if (has_idle_core) {
6618 i = select_idle_core(p, cpu, cpus, &idle_cpu);
6619 if ((unsigned int)i < nr_cpumask_bits)
6625 idle_cpu = __select_idle_cpu(cpu, p);
6626 if ((unsigned int)idle_cpu < nr_cpumask_bits)
6632 set_idle_cores(target, false);
6634 if (sched_feat(SIS_PROP) && this_sd && !has_idle_core) {
6635 time = cpu_clock(this) - time;
6638 * Account for the scan cost of wakeups against the average
6641 this_rq->wake_avg_idle -= min(this_rq->wake_avg_idle, time);
6643 update_avg(&this_sd->avg_scan_cost, time);
6650 * Scan the asym_capacity domain for idle CPUs; pick the first idle one on which
6651 * the task fits. If no CPU is big enough, but there are idle ones, try to
6652 * maximize capacity.
6655 select_idle_capacity(struct task_struct *p, struct sched_domain *sd, int target)
6657 unsigned long task_util, best_cap = 0;
6658 int cpu, best_cpu = -1;
6659 struct cpumask *cpus;
6661 cpus = this_cpu_cpumask_var_ptr(select_rq_mask);
6662 cpumask_and(cpus, sched_domain_span(sd), p->cpus_ptr);
6664 task_util = uclamp_task_util(p);
6666 for_each_cpu_wrap(cpu, cpus, target) {
6667 unsigned long cpu_cap = capacity_of(cpu);
6669 if (!available_idle_cpu(cpu) && !sched_idle_cpu(cpu))
6671 if (fits_capacity(task_util, cpu_cap))
6674 if (cpu_cap > best_cap) {
6683 static inline bool asym_fits_capacity(unsigned long task_util, int cpu)
6685 if (sched_asym_cpucap_active())
6686 return fits_capacity(task_util, capacity_of(cpu));
6692 * Try and locate an idle core/thread in the LLC cache domain.
6694 static int select_idle_sibling(struct task_struct *p, int prev, int target)
6696 bool has_idle_core = false;
6697 struct sched_domain *sd;
6698 unsigned long task_util;
6699 int i, recent_used_cpu;
6702 * On asymmetric system, update task utilization because we will check
6703 * that the task fits with cpu's capacity.
6705 if (sched_asym_cpucap_active()) {
6706 sync_entity_load_avg(&p->se);
6707 task_util = uclamp_task_util(p);
6711 * per-cpu select_rq_mask usage
6713 lockdep_assert_irqs_disabled();
6715 if ((available_idle_cpu(target) || sched_idle_cpu(target)) &&
6716 asym_fits_capacity(task_util, target))
6720 * If the previous CPU is cache affine and idle, don't be stupid:
6722 if (prev != target && cpus_share_cache(prev, target) &&
6723 (available_idle_cpu(prev) || sched_idle_cpu(prev)) &&
6724 asym_fits_capacity(task_util, prev))
6728 * Allow a per-cpu kthread to stack with the wakee if the
6729 * kworker thread and the tasks previous CPUs are the same.
6730 * The assumption is that the wakee queued work for the
6731 * per-cpu kthread that is now complete and the wakeup is
6732 * essentially a sync wakeup. An obvious example of this
6733 * pattern is IO completions.
6735 if (is_per_cpu_kthread(current) &&
6737 prev == smp_processor_id() &&
6738 this_rq()->nr_running <= 1 &&
6739 asym_fits_capacity(task_util, prev)) {
6743 /* Check a recently used CPU as a potential idle candidate: */
6744 recent_used_cpu = p->recent_used_cpu;
6745 p->recent_used_cpu = prev;
6746 if (recent_used_cpu != prev &&
6747 recent_used_cpu != target &&
6748 cpus_share_cache(recent_used_cpu, target) &&
6749 (available_idle_cpu(recent_used_cpu) || sched_idle_cpu(recent_used_cpu)) &&
6750 cpumask_test_cpu(p->recent_used_cpu, p->cpus_ptr) &&
6751 asym_fits_capacity(task_util, recent_used_cpu)) {
6752 return recent_used_cpu;
6756 * For asymmetric CPU capacity systems, our domain of interest is
6757 * sd_asym_cpucapacity rather than sd_llc.
6759 if (sched_asym_cpucap_active()) {
6760 sd = rcu_dereference(per_cpu(sd_asym_cpucapacity, target));
6762 * On an asymmetric CPU capacity system where an exclusive
6763 * cpuset defines a symmetric island (i.e. one unique
6764 * capacity_orig value through the cpuset), the key will be set
6765 * but the CPUs within that cpuset will not have a domain with
6766 * SD_ASYM_CPUCAPACITY. These should follow the usual symmetric
6770 i = select_idle_capacity(p, sd, target);
6771 return ((unsigned)i < nr_cpumask_bits) ? i : target;
6775 sd = rcu_dereference(per_cpu(sd_llc, target));
6779 if (sched_smt_active()) {
6780 has_idle_core = test_idle_cores(target);
6782 if (!has_idle_core && cpus_share_cache(prev, target)) {
6783 i = select_idle_smt(p, prev);
6784 if ((unsigned int)i < nr_cpumask_bits)
6789 i = select_idle_cpu(p, sd, has_idle_core, target);
6790 if ((unsigned)i < nr_cpumask_bits)
6797 * Predicts what cpu_util(@cpu) would return if @p was removed from @cpu
6798 * (@dst_cpu = -1) or migrated to @dst_cpu.
6800 static unsigned long cpu_util_next(int cpu, struct task_struct *p, int dst_cpu)
6802 struct cfs_rq *cfs_rq = &cpu_rq(cpu)->cfs;
6803 unsigned long util = READ_ONCE(cfs_rq->avg.util_avg);
6806 * If @dst_cpu is -1 or @p migrates from @cpu to @dst_cpu remove its
6807 * contribution. If @p migrates from another CPU to @cpu add its
6808 * contribution. In all the other cases @cpu is not impacted by the
6809 * migration so its util_avg is already correct.
6811 if (task_cpu(p) == cpu && dst_cpu != cpu)
6812 lsub_positive(&util, task_util(p));
6813 else if (task_cpu(p) != cpu && dst_cpu == cpu)
6814 util += task_util(p);
6816 if (sched_feat(UTIL_EST)) {
6817 unsigned long util_est;
6819 util_est = READ_ONCE(cfs_rq->avg.util_est.enqueued);
6822 * During wake-up @p isn't enqueued yet and doesn't contribute
6823 * to any cpu_rq(cpu)->cfs.avg.util_est.enqueued.
6824 * If @dst_cpu == @cpu add it to "simulate" cpu_util after @p
6825 * has been enqueued.
6827 * During exec (@dst_cpu = -1) @p is enqueued and does
6828 * contribute to cpu_rq(cpu)->cfs.util_est.enqueued.
6829 * Remove it to "simulate" cpu_util without @p's contribution.
6831 * Despite the task_on_rq_queued(@p) check there is still a
6832 * small window for a possible race when an exec
6833 * select_task_rq_fair() races with LB's detach_task().
6837 * p->on_rq = TASK_ON_RQ_MIGRATING;
6838 * -------------------------------- A
6840 * dequeue_task_fair() + Race Time
6841 * util_est_dequeue() /
6842 * -------------------------------- B
6844 * The additional check "current == p" is required to further
6845 * reduce the race window.
6848 util_est += _task_util_est(p);
6849 else if (unlikely(task_on_rq_queued(p) || current == p))
6850 lsub_positive(&util_est, _task_util_est(p));
6852 util = max(util, util_est);
6855 return min(util, capacity_orig_of(cpu));
6859 * cpu_util_without: compute cpu utilization without any contributions from *p
6860 * @cpu: the CPU which utilization is requested
6861 * @p: the task which utilization should be discounted
6863 * The utilization of a CPU is defined by the utilization of tasks currently
6864 * enqueued on that CPU as well as tasks which are currently sleeping after an
6865 * execution on that CPU.
6867 * This method returns the utilization of the specified CPU by discounting the
6868 * utilization of the specified task, whenever the task is currently
6869 * contributing to the CPU utilization.
6871 static unsigned long cpu_util_without(int cpu, struct task_struct *p)
6873 /* Task has no contribution or is new */
6874 if (cpu != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time))
6875 return cpu_util_cfs(cpu);
6877 return cpu_util_next(cpu, p, -1);
6881 * energy_env - Utilization landscape for energy estimation.
6882 * @task_busy_time: Utilization contribution by the task for which we test the
6883 * placement. Given by eenv_task_busy_time().
6884 * @pd_busy_time: Utilization of the whole perf domain without the task
6885 * contribution. Given by eenv_pd_busy_time().
6886 * @cpu_cap: Maximum CPU capacity for the perf domain.
6887 * @pd_cap: Entire perf domain capacity. (pd->nr_cpus * cpu_cap).
6890 unsigned long task_busy_time;
6891 unsigned long pd_busy_time;
6892 unsigned long cpu_cap;
6893 unsigned long pd_cap;
6897 * Compute the task busy time for compute_energy(). This time cannot be
6898 * injected directly into effective_cpu_util() because of the IRQ scaling.
6899 * The latter only makes sense with the most recent CPUs where the task has
6902 static inline void eenv_task_busy_time(struct energy_env *eenv,
6903 struct task_struct *p, int prev_cpu)
6905 unsigned long busy_time, max_cap = arch_scale_cpu_capacity(prev_cpu);
6906 unsigned long irq = cpu_util_irq(cpu_rq(prev_cpu));
6908 if (unlikely(irq >= max_cap))
6909 busy_time = max_cap;
6911 busy_time = scale_irq_capacity(task_util_est(p), irq, max_cap);
6913 eenv->task_busy_time = busy_time;
6917 * Compute the perf_domain (PD) busy time for compute_energy(). Based on the
6918 * utilization for each @pd_cpus, it however doesn't take into account
6919 * clamping since the ratio (utilization / cpu_capacity) is already enough to
6920 * scale the EM reported power consumption at the (eventually clamped)
6923 * The contribution of the task @p for which we want to estimate the
6924 * energy cost is removed (by cpu_util_next()) and must be calculated
6925 * separately (see eenv_task_busy_time). This ensures:
6927 * - A stable PD utilization, no matter which CPU of that PD we want to place
6930 * - A fair comparison between CPUs as the task contribution (task_util())
6931 * will always be the same no matter which CPU utilization we rely on
6932 * (util_avg or util_est).
6934 * Set @eenv busy time for the PD that spans @pd_cpus. This busy time can't
6935 * exceed @eenv->pd_cap.
6937 static inline void eenv_pd_busy_time(struct energy_env *eenv,
6938 struct cpumask *pd_cpus,
6939 struct task_struct *p)
6941 unsigned long busy_time = 0;
6944 for_each_cpu(cpu, pd_cpus) {
6945 unsigned long util = cpu_util_next(cpu, p, -1);
6947 busy_time += effective_cpu_util(cpu, util, ENERGY_UTIL, NULL);
6950 eenv->pd_busy_time = min(eenv->pd_cap, busy_time);
6954 * Compute the maximum utilization for compute_energy() when the task @p
6955 * is placed on the cpu @dst_cpu.
6957 * Returns the maximum utilization among @eenv->cpus. This utilization can't
6958 * exceed @eenv->cpu_cap.
6960 static inline unsigned long
6961 eenv_pd_max_util(struct energy_env *eenv, struct cpumask *pd_cpus,
6962 struct task_struct *p, int dst_cpu)
6964 unsigned long max_util = 0;
6967 for_each_cpu(cpu, pd_cpus) {
6968 struct task_struct *tsk = (cpu == dst_cpu) ? p : NULL;
6969 unsigned long util = cpu_util_next(cpu, p, dst_cpu);
6970 unsigned long cpu_util;
6973 * Performance domain frequency: utilization clamping
6974 * must be considered since it affects the selection
6975 * of the performance domain frequency.
6976 * NOTE: in case RT tasks are running, by default the
6977 * FREQUENCY_UTIL's utilization can be max OPP.
6979 cpu_util = effective_cpu_util(cpu, util, FREQUENCY_UTIL, tsk);
6980 max_util = max(max_util, cpu_util);
6983 return min(max_util, eenv->cpu_cap);
6987 * compute_energy(): Use the Energy Model to estimate the energy that @pd would
6988 * consume for a given utilization landscape @eenv. When @dst_cpu < 0, the task
6989 * contribution is ignored.
6991 static inline unsigned long
6992 compute_energy(struct energy_env *eenv, struct perf_domain *pd,
6993 struct cpumask *pd_cpus, struct task_struct *p, int dst_cpu)
6995 unsigned long max_util = eenv_pd_max_util(eenv, pd_cpus, p, dst_cpu);
6996 unsigned long busy_time = eenv->pd_busy_time;
6999 busy_time = min(eenv->pd_cap, busy_time + eenv->task_busy_time);
7001 return em_cpu_energy(pd->em_pd, max_util, busy_time, eenv->cpu_cap);
7005 * find_energy_efficient_cpu(): Find most energy-efficient target CPU for the
7006 * waking task. find_energy_efficient_cpu() looks for the CPU with maximum
7007 * spare capacity in each performance domain and uses it as a potential
7008 * candidate to execute the task. Then, it uses the Energy Model to figure
7009 * out which of the CPU candidates is the most energy-efficient.
7011 * The rationale for this heuristic is as follows. In a performance domain,
7012 * all the most energy efficient CPU candidates (according to the Energy
7013 * Model) are those for which we'll request a low frequency. When there are
7014 * several CPUs for which the frequency request will be the same, we don't
7015 * have enough data to break the tie between them, because the Energy Model
7016 * only includes active power costs. With this model, if we assume that
7017 * frequency requests follow utilization (e.g. using schedutil), the CPU with
7018 * the maximum spare capacity in a performance domain is guaranteed to be among
7019 * the best candidates of the performance domain.
7021 * In practice, it could be preferable from an energy standpoint to pack
7022 * small tasks on a CPU in order to let other CPUs go in deeper idle states,
7023 * but that could also hurt our chances to go cluster idle, and we have no
7024 * ways to tell with the current Energy Model if this is actually a good
7025 * idea or not. So, find_energy_efficient_cpu() basically favors
7026 * cluster-packing, and spreading inside a cluster. That should at least be
7027 * a good thing for latency, and this is consistent with the idea that most
7028 * of the energy savings of EAS come from the asymmetry of the system, and
7029 * not so much from breaking the tie between identical CPUs. That's also the
7030 * reason why EAS is enabled in the topology code only for systems where
7031 * SD_ASYM_CPUCAPACITY is set.
7033 * NOTE: Forkees are not accepted in the energy-aware wake-up path because
7034 * they don't have any useful utilization data yet and it's not possible to
7035 * forecast their impact on energy consumption. Consequently, they will be
7036 * placed by find_idlest_cpu() on the least loaded CPU, which might turn out
7037 * to be energy-inefficient in some use-cases. The alternative would be to
7038 * bias new tasks towards specific types of CPUs first, or to try to infer
7039 * their util_avg from the parent task, but those heuristics could hurt
7040 * other use-cases too. So, until someone finds a better way to solve this,
7041 * let's keep things simple by re-using the existing slow path.
7043 static int find_energy_efficient_cpu(struct task_struct *p, int prev_cpu)
7045 struct cpumask *cpus = this_cpu_cpumask_var_ptr(select_rq_mask);
7046 unsigned long prev_delta = ULONG_MAX, best_delta = ULONG_MAX;
7047 struct root_domain *rd = this_rq()->rd;
7048 int cpu, best_energy_cpu, target = -1;
7049 struct sched_domain *sd;
7050 struct perf_domain *pd;
7051 struct energy_env eenv;
7054 pd = rcu_dereference(rd->pd);
7055 if (!pd || READ_ONCE(rd->overutilized))
7059 * Energy-aware wake-up happens on the lowest sched_domain starting
7060 * from sd_asym_cpucapacity spanning over this_cpu and prev_cpu.
7062 sd = rcu_dereference(*this_cpu_ptr(&sd_asym_cpucapacity));
7063 while (sd && !cpumask_test_cpu(prev_cpu, sched_domain_span(sd)))
7070 sync_entity_load_avg(&p->se);
7071 if (!task_util_est(p))
7074 eenv_task_busy_time(&eenv, p, prev_cpu);
7076 for (; pd; pd = pd->next) {
7077 unsigned long cpu_cap, cpu_thermal_cap, util;
7078 unsigned long cur_delta, max_spare_cap = 0;
7079 bool compute_prev_delta = false;
7080 int max_spare_cap_cpu = -1;
7081 unsigned long base_energy;
7083 cpumask_and(cpus, perf_domain_span(pd), cpu_online_mask);
7085 if (cpumask_empty(cpus))
7088 /* Account thermal pressure for the energy estimation */
7089 cpu = cpumask_first(cpus);
7090 cpu_thermal_cap = arch_scale_cpu_capacity(cpu);
7091 cpu_thermal_cap -= arch_scale_thermal_pressure(cpu);
7093 eenv.cpu_cap = cpu_thermal_cap;
7096 for_each_cpu(cpu, cpus) {
7097 eenv.pd_cap += cpu_thermal_cap;
7099 if (!cpumask_test_cpu(cpu, sched_domain_span(sd)))
7102 if (!cpumask_test_cpu(cpu, p->cpus_ptr))
7105 util = cpu_util_next(cpu, p, cpu);
7106 cpu_cap = capacity_of(cpu);
7109 * Skip CPUs that cannot satisfy the capacity request.
7110 * IOW, placing the task there would make the CPU
7111 * overutilized. Take uclamp into account to see how
7112 * much capacity we can get out of the CPU; this is
7113 * aligned with sched_cpu_util().
7115 util = uclamp_rq_util_with(cpu_rq(cpu), util, p);
7116 if (!fits_capacity(util, cpu_cap))
7119 lsub_positive(&cpu_cap, util);
7121 if (cpu == prev_cpu) {
7122 /* Always use prev_cpu as a candidate. */
7123 compute_prev_delta = true;
7124 } else if (cpu_cap > max_spare_cap) {
7126 * Find the CPU with the maximum spare capacity
7127 * in the performance domain.
7129 max_spare_cap = cpu_cap;
7130 max_spare_cap_cpu = cpu;
7134 if (max_spare_cap_cpu < 0 && !compute_prev_delta)
7137 eenv_pd_busy_time(&eenv, cpus, p);
7138 /* Compute the 'base' energy of the pd, without @p */
7139 base_energy = compute_energy(&eenv, pd, cpus, p, -1);
7141 /* Evaluate the energy impact of using prev_cpu. */
7142 if (compute_prev_delta) {
7143 prev_delta = compute_energy(&eenv, pd, cpus, p,
7145 /* CPU utilization has changed */
7146 if (prev_delta < base_energy)
7148 prev_delta -= base_energy;
7149 best_delta = min(best_delta, prev_delta);
7152 /* Evaluate the energy impact of using max_spare_cap_cpu. */
7153 if (max_spare_cap_cpu >= 0) {
7154 cur_delta = compute_energy(&eenv, pd, cpus, p,
7156 /* CPU utilization has changed */
7157 if (cur_delta < base_energy)
7159 cur_delta -= base_energy;
7160 if (cur_delta < best_delta) {
7161 best_delta = cur_delta;
7162 best_energy_cpu = max_spare_cap_cpu;
7168 if (best_delta < prev_delta)
7169 target = best_energy_cpu;
7180 * select_task_rq_fair: Select target runqueue for the waking task in domains
7181 * that have the relevant SD flag set. In practice, this is SD_BALANCE_WAKE,
7182 * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
7184 * Balances load by selecting the idlest CPU in the idlest group, or under
7185 * certain conditions an idle sibling CPU if the domain has SD_WAKE_AFFINE set.
7187 * Returns the target CPU number.
7190 select_task_rq_fair(struct task_struct *p, int prev_cpu, int wake_flags)
7192 int sync = (wake_flags & WF_SYNC) && !(current->flags & PF_EXITING);
7193 struct sched_domain *tmp, *sd = NULL;
7194 int cpu = smp_processor_id();
7195 int new_cpu = prev_cpu;
7196 int want_affine = 0;
7197 /* SD_flags and WF_flags share the first nibble */
7198 int sd_flag = wake_flags & 0xF;
7201 * required for stable ->cpus_allowed
7203 lockdep_assert_held(&p->pi_lock);
7204 if (wake_flags & WF_TTWU) {
7207 if (sched_energy_enabled()) {
7208 new_cpu = find_energy_efficient_cpu(p, prev_cpu);
7214 want_affine = !wake_wide(p) && cpumask_test_cpu(cpu, p->cpus_ptr);
7218 for_each_domain(cpu, tmp) {
7220 * If both 'cpu' and 'prev_cpu' are part of this domain,
7221 * cpu is a valid SD_WAKE_AFFINE target.
7223 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
7224 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
7225 if (cpu != prev_cpu)
7226 new_cpu = wake_affine(tmp, p, cpu, prev_cpu, sync);
7228 sd = NULL; /* Prefer wake_affine over balance flags */
7233 * Usually only true for WF_EXEC and WF_FORK, as sched_domains
7234 * usually do not have SD_BALANCE_WAKE set. That means wakeup
7235 * will usually go to the fast path.
7237 if (tmp->flags & sd_flag)
7239 else if (!want_affine)
7245 new_cpu = find_idlest_cpu(sd, p, cpu, prev_cpu, sd_flag);
7246 } else if (wake_flags & WF_TTWU) { /* XXX always ? */
7248 new_cpu = select_idle_sibling(p, prev_cpu, new_cpu);
7256 * Called immediately before a task is migrated to a new CPU; task_cpu(p) and
7257 * cfs_rq_of(p) references at time of call are still valid and identify the
7258 * previous CPU. The caller guarantees p->pi_lock or task_rq(p)->lock is held.
7260 static void migrate_task_rq_fair(struct task_struct *p, int new_cpu)
7262 struct sched_entity *se = &p->se;
7265 * As blocked tasks retain absolute vruntime the migration needs to
7266 * deal with this by subtracting the old and adding the new
7267 * min_vruntime -- the latter is done by enqueue_entity() when placing
7268 * the task on the new runqueue.
7270 if (READ_ONCE(p->__state) == TASK_WAKING) {
7271 struct cfs_rq *cfs_rq = cfs_rq_of(se);
7273 se->vruntime -= u64_u32_load(cfs_rq->min_vruntime);
7276 if (!task_on_rq_migrating(p)) {
7277 remove_entity_load_avg(se);
7280 * Here, the task's PELT values have been updated according to
7281 * the current rq's clock. But if that clock hasn't been
7282 * updated in a while, a substantial idle time will be missed,
7283 * leading to an inflation after wake-up on the new rq.
7285 * Estimate the missing time from the cfs_rq last_update_time
7286 * and update sched_avg to improve the PELT continuity after
7289 migrate_se_pelt_lag(se);
7292 /* Tell new CPU we are migrated */
7293 se->avg.last_update_time = 0;
7295 /* We have migrated, no longer consider this task hot */
7298 update_scan_period(p, new_cpu);
7301 static void task_dead_fair(struct task_struct *p)
7303 remove_entity_load_avg(&p->se);
7307 balance_fair(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
7312 return newidle_balance(rq, rf) != 0;
7314 #endif /* CONFIG_SMP */
7316 static unsigned long wakeup_gran(struct sched_entity *se)
7318 unsigned long gran = sysctl_sched_wakeup_granularity;
7321 * Since its curr running now, convert the gran from real-time
7322 * to virtual-time in his units.
7324 * By using 'se' instead of 'curr' we penalize light tasks, so
7325 * they get preempted easier. That is, if 'se' < 'curr' then
7326 * the resulting gran will be larger, therefore penalizing the
7327 * lighter, if otoh 'se' > 'curr' then the resulting gran will
7328 * be smaller, again penalizing the lighter task.
7330 * This is especially important for buddies when the leftmost
7331 * task is higher priority than the buddy.
7333 return calc_delta_fair(gran, se);
7337 * Should 'se' preempt 'curr'.
7351 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
7353 s64 gran, vdiff = curr->vruntime - se->vruntime;
7358 gran = wakeup_gran(se);
7365 static void set_last_buddy(struct sched_entity *se)
7367 for_each_sched_entity(se) {
7368 if (SCHED_WARN_ON(!se->on_rq))
7372 cfs_rq_of(se)->last = se;
7376 static void set_next_buddy(struct sched_entity *se)
7378 for_each_sched_entity(se) {
7379 if (SCHED_WARN_ON(!se->on_rq))
7383 cfs_rq_of(se)->next = se;
7387 static void set_skip_buddy(struct sched_entity *se)
7389 for_each_sched_entity(se)
7390 cfs_rq_of(se)->skip = se;
7394 * Preempt the current task with a newly woken task if needed:
7396 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
7398 struct task_struct *curr = rq->curr;
7399 struct sched_entity *se = &curr->se, *pse = &p->se;
7400 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
7401 int scale = cfs_rq->nr_running >= sched_nr_latency;
7402 int next_buddy_marked = 0;
7403 int cse_is_idle, pse_is_idle;
7405 if (unlikely(se == pse))
7409 * This is possible from callers such as attach_tasks(), in which we
7410 * unconditionally check_preempt_curr() after an enqueue (which may have
7411 * lead to a throttle). This both saves work and prevents false
7412 * next-buddy nomination below.
7414 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
7417 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
7418 set_next_buddy(pse);
7419 next_buddy_marked = 1;
7423 * We can come here with TIF_NEED_RESCHED already set from new task
7426 * Note: this also catches the edge-case of curr being in a throttled
7427 * group (e.g. via set_curr_task), since update_curr() (in the
7428 * enqueue of curr) will have resulted in resched being set. This
7429 * prevents us from potentially nominating it as a false LAST_BUDDY
7432 if (test_tsk_need_resched(curr))
7435 /* Idle tasks are by definition preempted by non-idle tasks. */
7436 if (unlikely(task_has_idle_policy(curr)) &&
7437 likely(!task_has_idle_policy(p)))
7441 * Batch and idle tasks do not preempt non-idle tasks (their preemption
7442 * is driven by the tick):
7444 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
7447 find_matching_se(&se, &pse);
7450 cse_is_idle = se_is_idle(se);
7451 pse_is_idle = se_is_idle(pse);
7454 * Preempt an idle group in favor of a non-idle group (and don't preempt
7455 * in the inverse case).
7457 if (cse_is_idle && !pse_is_idle)
7459 if (cse_is_idle != pse_is_idle)
7462 update_curr(cfs_rq_of(se));
7463 if (wakeup_preempt_entity(se, pse) == 1) {
7465 * Bias pick_next to pick the sched entity that is
7466 * triggering this preemption.
7468 if (!next_buddy_marked)
7469 set_next_buddy(pse);
7478 * Only set the backward buddy when the current task is still
7479 * on the rq. This can happen when a wakeup gets interleaved
7480 * with schedule on the ->pre_schedule() or idle_balance()
7481 * point, either of which can * drop the rq lock.
7483 * Also, during early boot the idle thread is in the fair class,
7484 * for obvious reasons its a bad idea to schedule back to it.
7486 if (unlikely(!se->on_rq || curr == rq->idle))
7489 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
7494 static struct task_struct *pick_task_fair(struct rq *rq)
7496 struct sched_entity *se;
7497 struct cfs_rq *cfs_rq;
7501 if (!cfs_rq->nr_running)
7505 struct sched_entity *curr = cfs_rq->curr;
7507 /* When we pick for a remote RQ, we'll not have done put_prev_entity() */
7510 update_curr(cfs_rq);
7514 if (unlikely(check_cfs_rq_runtime(cfs_rq)))
7518 se = pick_next_entity(cfs_rq, curr);
7519 cfs_rq = group_cfs_rq(se);
7526 struct task_struct *
7527 pick_next_task_fair(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
7529 struct cfs_rq *cfs_rq = &rq->cfs;
7530 struct sched_entity *se;
7531 struct task_struct *p;
7535 if (!sched_fair_runnable(rq))
7538 #ifdef CONFIG_FAIR_GROUP_SCHED
7539 if (!prev || prev->sched_class != &fair_sched_class)
7543 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
7544 * likely that a next task is from the same cgroup as the current.
7546 * Therefore attempt to avoid putting and setting the entire cgroup
7547 * hierarchy, only change the part that actually changes.
7551 struct sched_entity *curr = cfs_rq->curr;
7554 * Since we got here without doing put_prev_entity() we also
7555 * have to consider cfs_rq->curr. If it is still a runnable
7556 * entity, update_curr() will update its vruntime, otherwise
7557 * forget we've ever seen it.
7561 update_curr(cfs_rq);
7566 * This call to check_cfs_rq_runtime() will do the
7567 * throttle and dequeue its entity in the parent(s).
7568 * Therefore the nr_running test will indeed
7571 if (unlikely(check_cfs_rq_runtime(cfs_rq))) {
7574 if (!cfs_rq->nr_running)
7581 se = pick_next_entity(cfs_rq, curr);
7582 cfs_rq = group_cfs_rq(se);
7588 * Since we haven't yet done put_prev_entity and if the selected task
7589 * is a different task than we started out with, try and touch the
7590 * least amount of cfs_rqs.
7593 struct sched_entity *pse = &prev->se;
7595 while (!(cfs_rq = is_same_group(se, pse))) {
7596 int se_depth = se->depth;
7597 int pse_depth = pse->depth;
7599 if (se_depth <= pse_depth) {
7600 put_prev_entity(cfs_rq_of(pse), pse);
7601 pse = parent_entity(pse);
7603 if (se_depth >= pse_depth) {
7604 set_next_entity(cfs_rq_of(se), se);
7605 se = parent_entity(se);
7609 put_prev_entity(cfs_rq, pse);
7610 set_next_entity(cfs_rq, se);
7617 put_prev_task(rq, prev);
7620 se = pick_next_entity(cfs_rq, NULL);
7621 set_next_entity(cfs_rq, se);
7622 cfs_rq = group_cfs_rq(se);
7627 done: __maybe_unused;
7630 * Move the next running task to the front of
7631 * the list, so our cfs_tasks list becomes MRU
7634 list_move(&p->se.group_node, &rq->cfs_tasks);
7637 if (hrtick_enabled_fair(rq))
7638 hrtick_start_fair(rq, p);
7640 update_misfit_status(p, rq);
7648 new_tasks = newidle_balance(rq, rf);
7651 * Because newidle_balance() releases (and re-acquires) rq->lock, it is
7652 * possible for any higher priority task to appear. In that case we
7653 * must re-start the pick_next_entity() loop.
7662 * rq is about to be idle, check if we need to update the
7663 * lost_idle_time of clock_pelt
7665 update_idle_rq_clock_pelt(rq);
7670 static struct task_struct *__pick_next_task_fair(struct rq *rq)
7672 return pick_next_task_fair(rq, NULL, NULL);
7676 * Account for a descheduled task:
7678 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
7680 struct sched_entity *se = &prev->se;
7681 struct cfs_rq *cfs_rq;
7683 for_each_sched_entity(se) {
7684 cfs_rq = cfs_rq_of(se);
7685 put_prev_entity(cfs_rq, se);
7690 * sched_yield() is very simple
7692 * The magic of dealing with the ->skip buddy is in pick_next_entity.
7694 static void yield_task_fair(struct rq *rq)
7696 struct task_struct *curr = rq->curr;
7697 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
7698 struct sched_entity *se = &curr->se;
7701 * Are we the only task in the tree?
7703 if (unlikely(rq->nr_running == 1))
7706 clear_buddies(cfs_rq, se);
7708 if (curr->policy != SCHED_BATCH) {
7709 update_rq_clock(rq);
7711 * Update run-time statistics of the 'current'.
7713 update_curr(cfs_rq);
7715 * Tell update_rq_clock() that we've just updated,
7716 * so we don't do microscopic update in schedule()
7717 * and double the fastpath cost.
7719 rq_clock_skip_update(rq);
7725 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p)
7727 struct sched_entity *se = &p->se;
7729 /* throttled hierarchies are not runnable */
7730 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
7733 /* Tell the scheduler that we'd really like pse to run next. */
7736 yield_task_fair(rq);
7742 /**************************************************
7743 * Fair scheduling class load-balancing methods.
7747 * The purpose of load-balancing is to achieve the same basic fairness the
7748 * per-CPU scheduler provides, namely provide a proportional amount of compute
7749 * time to each task. This is expressed in the following equation:
7751 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
7753 * Where W_i,n is the n-th weight average for CPU i. The instantaneous weight
7754 * W_i,0 is defined as:
7756 * W_i,0 = \Sum_j w_i,j (2)
7758 * Where w_i,j is the weight of the j-th runnable task on CPU i. This weight
7759 * is derived from the nice value as per sched_prio_to_weight[].
7761 * The weight average is an exponential decay average of the instantaneous
7764 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
7766 * C_i is the compute capacity of CPU i, typically it is the
7767 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
7768 * can also include other factors [XXX].
7770 * To achieve this balance we define a measure of imbalance which follows
7771 * directly from (1):
7773 * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4)
7775 * We them move tasks around to minimize the imbalance. In the continuous
7776 * function space it is obvious this converges, in the discrete case we get
7777 * a few fun cases generally called infeasible weight scenarios.
7780 * - infeasible weights;
7781 * - local vs global optima in the discrete case. ]
7786 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
7787 * for all i,j solution, we create a tree of CPUs that follows the hardware
7788 * topology where each level pairs two lower groups (or better). This results
7789 * in O(log n) layers. Furthermore we reduce the number of CPUs going up the
7790 * tree to only the first of the previous level and we decrease the frequency
7791 * of load-balance at each level inv. proportional to the number of CPUs in
7797 * \Sum { --- * --- * 2^i } = O(n) (5)
7799 * `- size of each group
7800 * | | `- number of CPUs doing load-balance
7802 * `- sum over all levels
7804 * Coupled with a limit on how many tasks we can migrate every balance pass,
7805 * this makes (5) the runtime complexity of the balancer.
7807 * An important property here is that each CPU is still (indirectly) connected
7808 * to every other CPU in at most O(log n) steps:
7810 * The adjacency matrix of the resulting graph is given by:
7813 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
7816 * And you'll find that:
7818 * A^(log_2 n)_i,j != 0 for all i,j (7)
7820 * Showing there's indeed a path between every CPU in at most O(log n) steps.
7821 * The task movement gives a factor of O(m), giving a convergence complexity
7824 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
7829 * In order to avoid CPUs going idle while there's still work to do, new idle
7830 * balancing is more aggressive and has the newly idle CPU iterate up the domain
7831 * tree itself instead of relying on other CPUs to bring it work.
7833 * This adds some complexity to both (5) and (8) but it reduces the total idle
7841 * Cgroups make a horror show out of (2), instead of a simple sum we get:
7844 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
7849 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
7851 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on CPU i.
7853 * The big problem is S_k, its a global sum needed to compute a local (W_i)
7856 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
7857 * rewrite all of this once again.]
7860 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
7862 enum fbq_type { regular, remote, all };
7865 * 'group_type' describes the group of CPUs at the moment of load balancing.
7867 * The enum is ordered by pulling priority, with the group with lowest priority
7868 * first so the group_type can simply be compared when selecting the busiest
7869 * group. See update_sd_pick_busiest().
7872 /* The group has spare capacity that can be used to run more tasks. */
7873 group_has_spare = 0,
7875 * The group is fully used and the tasks don't compete for more CPU
7876 * cycles. Nevertheless, some tasks might wait before running.
7880 * One task doesn't fit with CPU's capacity and must be migrated to a
7881 * more powerful CPU.
7885 * SD_ASYM_PACKING only: One local CPU with higher capacity is available,
7886 * and the task should be migrated to it instead of running on the
7891 * The tasks' affinity constraints previously prevented the scheduler
7892 * from balancing the load across the system.
7896 * The CPU is overloaded and can't provide expected CPU cycles to all
7902 enum migration_type {
7909 #define LBF_ALL_PINNED 0x01
7910 #define LBF_NEED_BREAK 0x02
7911 #define LBF_DST_PINNED 0x04
7912 #define LBF_SOME_PINNED 0x08
7913 #define LBF_ACTIVE_LB 0x10
7916 struct sched_domain *sd;
7924 struct cpumask *dst_grpmask;
7926 enum cpu_idle_type idle;
7928 /* The set of CPUs under consideration for load-balancing */
7929 struct cpumask *cpus;
7934 unsigned int loop_break;
7935 unsigned int loop_max;
7937 enum fbq_type fbq_type;
7938 enum migration_type migration_type;
7939 struct list_head tasks;
7943 * Is this task likely cache-hot:
7945 static int task_hot(struct task_struct *p, struct lb_env *env)
7949 lockdep_assert_rq_held(env->src_rq);
7951 if (p->sched_class != &fair_sched_class)
7954 if (unlikely(task_has_idle_policy(p)))
7957 /* SMT siblings share cache */
7958 if (env->sd->flags & SD_SHARE_CPUCAPACITY)
7962 * Buddy candidates are cache hot:
7964 if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running &&
7965 (&p->se == cfs_rq_of(&p->se)->next ||
7966 &p->se == cfs_rq_of(&p->se)->last))
7969 if (sysctl_sched_migration_cost == -1)
7973 * Don't migrate task if the task's cookie does not match
7974 * with the destination CPU's core cookie.
7976 if (!sched_core_cookie_match(cpu_rq(env->dst_cpu), p))
7979 if (sysctl_sched_migration_cost == 0)
7982 delta = rq_clock_task(env->src_rq) - p->se.exec_start;
7984 return delta < (s64)sysctl_sched_migration_cost;
7987 #ifdef CONFIG_NUMA_BALANCING
7989 * Returns 1, if task migration degrades locality
7990 * Returns 0, if task migration improves locality i.e migration preferred.
7991 * Returns -1, if task migration is not affected by locality.
7993 static int migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
7995 struct numa_group *numa_group = rcu_dereference(p->numa_group);
7996 unsigned long src_weight, dst_weight;
7997 int src_nid, dst_nid, dist;
7999 if (!static_branch_likely(&sched_numa_balancing))
8002 if (!p->numa_faults || !(env->sd->flags & SD_NUMA))
8005 src_nid = cpu_to_node(env->src_cpu);
8006 dst_nid = cpu_to_node(env->dst_cpu);
8008 if (src_nid == dst_nid)
8011 /* Migrating away from the preferred node is always bad. */
8012 if (src_nid == p->numa_preferred_nid) {
8013 if (env->src_rq->nr_running > env->src_rq->nr_preferred_running)
8019 /* Encourage migration to the preferred node. */
8020 if (dst_nid == p->numa_preferred_nid)
8023 /* Leaving a core idle is often worse than degrading locality. */
8024 if (env->idle == CPU_IDLE)
8027 dist = node_distance(src_nid, dst_nid);
8029 src_weight = group_weight(p, src_nid, dist);
8030 dst_weight = group_weight(p, dst_nid, dist);
8032 src_weight = task_weight(p, src_nid, dist);
8033 dst_weight = task_weight(p, dst_nid, dist);
8036 return dst_weight < src_weight;
8040 static inline int migrate_degrades_locality(struct task_struct *p,
8048 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
8051 int can_migrate_task(struct task_struct *p, struct lb_env *env)
8055 lockdep_assert_rq_held(env->src_rq);
8058 * We do not migrate tasks that are:
8059 * 1) throttled_lb_pair, or
8060 * 2) cannot be migrated to this CPU due to cpus_ptr, or
8061 * 3) running (obviously), or
8062 * 4) are cache-hot on their current CPU.
8064 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
8067 /* Disregard pcpu kthreads; they are where they need to be. */
8068 if (kthread_is_per_cpu(p))
8071 if (!cpumask_test_cpu(env->dst_cpu, p->cpus_ptr)) {
8074 schedstat_inc(p->stats.nr_failed_migrations_affine);
8076 env->flags |= LBF_SOME_PINNED;
8079 * Remember if this task can be migrated to any other CPU in
8080 * our sched_group. We may want to revisit it if we couldn't
8081 * meet load balance goals by pulling other tasks on src_cpu.
8083 * Avoid computing new_dst_cpu
8085 * - if we have already computed one in current iteration
8086 * - if it's an active balance
8088 if (env->idle == CPU_NEWLY_IDLE ||
8089 env->flags & (LBF_DST_PINNED | LBF_ACTIVE_LB))
8092 /* Prevent to re-select dst_cpu via env's CPUs: */
8093 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
8094 if (cpumask_test_cpu(cpu, p->cpus_ptr)) {
8095 env->flags |= LBF_DST_PINNED;
8096 env->new_dst_cpu = cpu;
8104 /* Record that we found at least one task that could run on dst_cpu */
8105 env->flags &= ~LBF_ALL_PINNED;
8107 if (task_on_cpu(env->src_rq, p)) {
8108 schedstat_inc(p->stats.nr_failed_migrations_running);
8113 * Aggressive migration if:
8115 * 2) destination numa is preferred
8116 * 3) task is cache cold, or
8117 * 4) too many balance attempts have failed.
8119 if (env->flags & LBF_ACTIVE_LB)
8122 tsk_cache_hot = migrate_degrades_locality(p, env);
8123 if (tsk_cache_hot == -1)
8124 tsk_cache_hot = task_hot(p, env);
8126 if (tsk_cache_hot <= 0 ||
8127 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
8128 if (tsk_cache_hot == 1) {
8129 schedstat_inc(env->sd->lb_hot_gained[env->idle]);
8130 schedstat_inc(p->stats.nr_forced_migrations);
8135 schedstat_inc(p->stats.nr_failed_migrations_hot);
8140 * detach_task() -- detach the task for the migration specified in env
8142 static void detach_task(struct task_struct *p, struct lb_env *env)
8144 lockdep_assert_rq_held(env->src_rq);
8146 deactivate_task(env->src_rq, p, DEQUEUE_NOCLOCK);
8147 set_task_cpu(p, env->dst_cpu);
8151 * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
8152 * part of active balancing operations within "domain".
8154 * Returns a task if successful and NULL otherwise.
8156 static struct task_struct *detach_one_task(struct lb_env *env)
8158 struct task_struct *p;
8160 lockdep_assert_rq_held(env->src_rq);
8162 list_for_each_entry_reverse(p,
8163 &env->src_rq->cfs_tasks, se.group_node) {
8164 if (!can_migrate_task(p, env))
8167 detach_task(p, env);
8170 * Right now, this is only the second place where
8171 * lb_gained[env->idle] is updated (other is detach_tasks)
8172 * so we can safely collect stats here rather than
8173 * inside detach_tasks().
8175 schedstat_inc(env->sd->lb_gained[env->idle]);
8182 * detach_tasks() -- tries to detach up to imbalance load/util/tasks from
8183 * busiest_rq, as part of a balancing operation within domain "sd".
8185 * Returns number of detached tasks if successful and 0 otherwise.
8187 static int detach_tasks(struct lb_env *env)
8189 struct list_head *tasks = &env->src_rq->cfs_tasks;
8190 unsigned long util, load;
8191 struct task_struct *p;
8194 lockdep_assert_rq_held(env->src_rq);
8197 * Source run queue has been emptied by another CPU, clear
8198 * LBF_ALL_PINNED flag as we will not test any task.
8200 if (env->src_rq->nr_running <= 1) {
8201 env->flags &= ~LBF_ALL_PINNED;
8205 if (env->imbalance <= 0)
8208 while (!list_empty(tasks)) {
8210 * We don't want to steal all, otherwise we may be treated likewise,
8211 * which could at worst lead to a livelock crash.
8213 if (env->idle != CPU_NOT_IDLE && env->src_rq->nr_running <= 1)
8218 * We've more or less seen every task there is, call it quits
8219 * unless we haven't found any movable task yet.
8221 if (env->loop > env->loop_max &&
8222 !(env->flags & LBF_ALL_PINNED))
8225 /* take a breather every nr_migrate tasks */
8226 if (env->loop > env->loop_break) {
8227 env->loop_break += SCHED_NR_MIGRATE_BREAK;
8228 env->flags |= LBF_NEED_BREAK;
8232 p = list_last_entry(tasks, struct task_struct, se.group_node);
8234 if (!can_migrate_task(p, env))
8237 switch (env->migration_type) {
8240 * Depending of the number of CPUs and tasks and the
8241 * cgroup hierarchy, task_h_load() can return a null
8242 * value. Make sure that env->imbalance decreases
8243 * otherwise detach_tasks() will stop only after
8244 * detaching up to loop_max tasks.
8246 load = max_t(unsigned long, task_h_load(p), 1);
8248 if (sched_feat(LB_MIN) &&
8249 load < 16 && !env->sd->nr_balance_failed)
8253 * Make sure that we don't migrate too much load.
8254 * Nevertheless, let relax the constraint if
8255 * scheduler fails to find a good waiting task to
8258 if (shr_bound(load, env->sd->nr_balance_failed) > env->imbalance)
8261 env->imbalance -= load;
8265 util = task_util_est(p);
8267 if (util > env->imbalance)
8270 env->imbalance -= util;
8277 case migrate_misfit:
8278 /* This is not a misfit task */
8279 if (task_fits_capacity(p, capacity_of(env->src_cpu)))
8286 detach_task(p, env);
8287 list_add(&p->se.group_node, &env->tasks);
8291 #ifdef CONFIG_PREEMPTION
8293 * NEWIDLE balancing is a source of latency, so preemptible
8294 * kernels will stop after the first task is detached to minimize
8295 * the critical section.
8297 if (env->idle == CPU_NEWLY_IDLE)
8302 * We only want to steal up to the prescribed amount of
8305 if (env->imbalance <= 0)
8310 list_move(&p->se.group_node, tasks);
8314 * Right now, this is one of only two places we collect this stat
8315 * so we can safely collect detach_one_task() stats here rather
8316 * than inside detach_one_task().
8318 schedstat_add(env->sd->lb_gained[env->idle], detached);
8324 * attach_task() -- attach the task detached by detach_task() to its new rq.
8326 static void attach_task(struct rq *rq, struct task_struct *p)
8328 lockdep_assert_rq_held(rq);
8330 WARN_ON_ONCE(task_rq(p) != rq);
8331 activate_task(rq, p, ENQUEUE_NOCLOCK);
8332 check_preempt_curr(rq, p, 0);
8336 * attach_one_task() -- attaches the task returned from detach_one_task() to
8339 static void attach_one_task(struct rq *rq, struct task_struct *p)
8344 update_rq_clock(rq);
8350 * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
8353 static void attach_tasks(struct lb_env *env)
8355 struct list_head *tasks = &env->tasks;
8356 struct task_struct *p;
8359 rq_lock(env->dst_rq, &rf);
8360 update_rq_clock(env->dst_rq);
8362 while (!list_empty(tasks)) {
8363 p = list_first_entry(tasks, struct task_struct, se.group_node);
8364 list_del_init(&p->se.group_node);
8366 attach_task(env->dst_rq, p);
8369 rq_unlock(env->dst_rq, &rf);
8372 #ifdef CONFIG_NO_HZ_COMMON
8373 static inline bool cfs_rq_has_blocked(struct cfs_rq *cfs_rq)
8375 if (cfs_rq->avg.load_avg)
8378 if (cfs_rq->avg.util_avg)
8384 static inline bool others_have_blocked(struct rq *rq)
8386 if (READ_ONCE(rq->avg_rt.util_avg))
8389 if (READ_ONCE(rq->avg_dl.util_avg))
8392 if (thermal_load_avg(rq))
8395 #ifdef CONFIG_HAVE_SCHED_AVG_IRQ
8396 if (READ_ONCE(rq->avg_irq.util_avg))
8403 static inline void update_blocked_load_tick(struct rq *rq)
8405 WRITE_ONCE(rq->last_blocked_load_update_tick, jiffies);
8408 static inline void update_blocked_load_status(struct rq *rq, bool has_blocked)
8411 rq->has_blocked_load = 0;
8414 static inline bool cfs_rq_has_blocked(struct cfs_rq *cfs_rq) { return false; }
8415 static inline bool others_have_blocked(struct rq *rq) { return false; }
8416 static inline void update_blocked_load_tick(struct rq *rq) {}
8417 static inline void update_blocked_load_status(struct rq *rq, bool has_blocked) {}
8420 static bool __update_blocked_others(struct rq *rq, bool *done)
8422 const struct sched_class *curr_class;
8423 u64 now = rq_clock_pelt(rq);
8424 unsigned long thermal_pressure;
8428 * update_load_avg() can call cpufreq_update_util(). Make sure that RT,
8429 * DL and IRQ signals have been updated before updating CFS.
8431 curr_class = rq->curr->sched_class;
8433 thermal_pressure = arch_scale_thermal_pressure(cpu_of(rq));
8435 decayed = update_rt_rq_load_avg(now, rq, curr_class == &rt_sched_class) |
8436 update_dl_rq_load_avg(now, rq, curr_class == &dl_sched_class) |
8437 update_thermal_load_avg(rq_clock_thermal(rq), rq, thermal_pressure) |
8438 update_irq_load_avg(rq, 0);
8440 if (others_have_blocked(rq))
8446 #ifdef CONFIG_FAIR_GROUP_SCHED
8448 static bool __update_blocked_fair(struct rq *rq, bool *done)
8450 struct cfs_rq *cfs_rq, *pos;
8451 bool decayed = false;
8452 int cpu = cpu_of(rq);
8455 * Iterates the task_group tree in a bottom up fashion, see
8456 * list_add_leaf_cfs_rq() for details.
8458 for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) {
8459 struct sched_entity *se;
8461 if (update_cfs_rq_load_avg(cfs_rq_clock_pelt(cfs_rq), cfs_rq)) {
8462 update_tg_load_avg(cfs_rq);
8464 if (cfs_rq->nr_running == 0)
8465 update_idle_cfs_rq_clock_pelt(cfs_rq);
8467 if (cfs_rq == &rq->cfs)
8471 /* Propagate pending load changes to the parent, if any: */
8472 se = cfs_rq->tg->se[cpu];
8473 if (se && !skip_blocked_update(se))
8474 update_load_avg(cfs_rq_of(se), se, UPDATE_TG);
8477 * There can be a lot of idle CPU cgroups. Don't let fully
8478 * decayed cfs_rqs linger on the list.
8480 if (cfs_rq_is_decayed(cfs_rq))
8481 list_del_leaf_cfs_rq(cfs_rq);
8483 /* Don't need periodic decay once load/util_avg are null */
8484 if (cfs_rq_has_blocked(cfs_rq))
8492 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
8493 * This needs to be done in a top-down fashion because the load of a child
8494 * group is a fraction of its parents load.
8496 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
8498 struct rq *rq = rq_of(cfs_rq);
8499 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
8500 unsigned long now = jiffies;
8503 if (cfs_rq->last_h_load_update == now)
8506 WRITE_ONCE(cfs_rq->h_load_next, NULL);
8507 for_each_sched_entity(se) {
8508 cfs_rq = cfs_rq_of(se);
8509 WRITE_ONCE(cfs_rq->h_load_next, se);
8510 if (cfs_rq->last_h_load_update == now)
8515 cfs_rq->h_load = cfs_rq_load_avg(cfs_rq);
8516 cfs_rq->last_h_load_update = now;
8519 while ((se = READ_ONCE(cfs_rq->h_load_next)) != NULL) {
8520 load = cfs_rq->h_load;
8521 load = div64_ul(load * se->avg.load_avg,
8522 cfs_rq_load_avg(cfs_rq) + 1);
8523 cfs_rq = group_cfs_rq(se);
8524 cfs_rq->h_load = load;
8525 cfs_rq->last_h_load_update = now;
8529 static unsigned long task_h_load(struct task_struct *p)
8531 struct cfs_rq *cfs_rq = task_cfs_rq(p);
8533 update_cfs_rq_h_load(cfs_rq);
8534 return div64_ul(p->se.avg.load_avg * cfs_rq->h_load,
8535 cfs_rq_load_avg(cfs_rq) + 1);
8538 static bool __update_blocked_fair(struct rq *rq, bool *done)
8540 struct cfs_rq *cfs_rq = &rq->cfs;
8543 decayed = update_cfs_rq_load_avg(cfs_rq_clock_pelt(cfs_rq), cfs_rq);
8544 if (cfs_rq_has_blocked(cfs_rq))
8550 static unsigned long task_h_load(struct task_struct *p)
8552 return p->se.avg.load_avg;
8556 static void update_blocked_averages(int cpu)
8558 bool decayed = false, done = true;
8559 struct rq *rq = cpu_rq(cpu);
8562 rq_lock_irqsave(rq, &rf);
8563 update_blocked_load_tick(rq);
8564 update_rq_clock(rq);
8566 decayed |= __update_blocked_others(rq, &done);
8567 decayed |= __update_blocked_fair(rq, &done);
8569 update_blocked_load_status(rq, !done);
8571 cpufreq_update_util(rq, 0);
8572 rq_unlock_irqrestore(rq, &rf);
8575 /********** Helpers for find_busiest_group ************************/
8578 * sg_lb_stats - stats of a sched_group required for load_balancing
8580 struct sg_lb_stats {
8581 unsigned long avg_load; /*Avg load across the CPUs of the group */
8582 unsigned long group_load; /* Total load over the CPUs of the group */
8583 unsigned long group_capacity;
8584 unsigned long group_util; /* Total utilization over the CPUs of the group */
8585 unsigned long group_runnable; /* Total runnable time over the CPUs of the group */
8586 unsigned int sum_nr_running; /* Nr of tasks running in the group */
8587 unsigned int sum_h_nr_running; /* Nr of CFS tasks running in the group */
8588 unsigned int idle_cpus;
8589 unsigned int group_weight;
8590 enum group_type group_type;
8591 unsigned int group_asym_packing; /* Tasks should be moved to preferred CPU */
8592 unsigned long group_misfit_task_load; /* A CPU has a task too big for its capacity */
8593 #ifdef CONFIG_NUMA_BALANCING
8594 unsigned int nr_numa_running;
8595 unsigned int nr_preferred_running;
8600 * sd_lb_stats - Structure to store the statistics of a sched_domain
8601 * during load balancing.
8603 struct sd_lb_stats {
8604 struct sched_group *busiest; /* Busiest group in this sd */
8605 struct sched_group *local; /* Local group in this sd */
8606 unsigned long total_load; /* Total load of all groups in sd */
8607 unsigned long total_capacity; /* Total capacity of all groups in sd */
8608 unsigned long avg_load; /* Average load across all groups in sd */
8609 unsigned int prefer_sibling; /* tasks should go to sibling first */
8611 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
8612 struct sg_lb_stats local_stat; /* Statistics of the local group */
8615 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
8618 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
8619 * local_stat because update_sg_lb_stats() does a full clear/assignment.
8620 * We must however set busiest_stat::group_type and
8621 * busiest_stat::idle_cpus to the worst busiest group because
8622 * update_sd_pick_busiest() reads these before assignment.
8624 *sds = (struct sd_lb_stats){
8628 .total_capacity = 0UL,
8630 .idle_cpus = UINT_MAX,
8631 .group_type = group_has_spare,
8636 static unsigned long scale_rt_capacity(int cpu)
8638 struct rq *rq = cpu_rq(cpu);
8639 unsigned long max = arch_scale_cpu_capacity(cpu);
8640 unsigned long used, free;
8643 irq = cpu_util_irq(rq);
8645 if (unlikely(irq >= max))
8649 * avg_rt.util_avg and avg_dl.util_avg track binary signals
8650 * (running and not running) with weights 0 and 1024 respectively.
8651 * avg_thermal.load_avg tracks thermal pressure and the weighted
8652 * average uses the actual delta max capacity(load).
8654 used = READ_ONCE(rq->avg_rt.util_avg);
8655 used += READ_ONCE(rq->avg_dl.util_avg);
8656 used += thermal_load_avg(rq);
8658 if (unlikely(used >= max))
8663 return scale_irq_capacity(free, irq, max);
8666 static void update_cpu_capacity(struct sched_domain *sd, int cpu)
8668 unsigned long capacity = scale_rt_capacity(cpu);
8669 struct sched_group *sdg = sd->groups;
8671 cpu_rq(cpu)->cpu_capacity_orig = arch_scale_cpu_capacity(cpu);
8676 cpu_rq(cpu)->cpu_capacity = capacity;
8677 trace_sched_cpu_capacity_tp(cpu_rq(cpu));
8679 sdg->sgc->capacity = capacity;
8680 sdg->sgc->min_capacity = capacity;
8681 sdg->sgc->max_capacity = capacity;
8684 void update_group_capacity(struct sched_domain *sd, int cpu)
8686 struct sched_domain *child = sd->child;
8687 struct sched_group *group, *sdg = sd->groups;
8688 unsigned long capacity, min_capacity, max_capacity;
8689 unsigned long interval;
8691 interval = msecs_to_jiffies(sd->balance_interval);
8692 interval = clamp(interval, 1UL, max_load_balance_interval);
8693 sdg->sgc->next_update = jiffies + interval;
8696 update_cpu_capacity(sd, cpu);
8701 min_capacity = ULONG_MAX;
8704 if (child->flags & SD_OVERLAP) {
8706 * SD_OVERLAP domains cannot assume that child groups
8707 * span the current group.
8710 for_each_cpu(cpu, sched_group_span(sdg)) {
8711 unsigned long cpu_cap = capacity_of(cpu);
8713 capacity += cpu_cap;
8714 min_capacity = min(cpu_cap, min_capacity);
8715 max_capacity = max(cpu_cap, max_capacity);
8719 * !SD_OVERLAP domains can assume that child groups
8720 * span the current group.
8723 group = child->groups;
8725 struct sched_group_capacity *sgc = group->sgc;
8727 capacity += sgc->capacity;
8728 min_capacity = min(sgc->min_capacity, min_capacity);
8729 max_capacity = max(sgc->max_capacity, max_capacity);
8730 group = group->next;
8731 } while (group != child->groups);
8734 sdg->sgc->capacity = capacity;
8735 sdg->sgc->min_capacity = min_capacity;
8736 sdg->sgc->max_capacity = max_capacity;
8740 * Check whether the capacity of the rq has been noticeably reduced by side
8741 * activity. The imbalance_pct is used for the threshold.
8742 * Return true is the capacity is reduced
8745 check_cpu_capacity(struct rq *rq, struct sched_domain *sd)
8747 return ((rq->cpu_capacity * sd->imbalance_pct) <
8748 (rq->cpu_capacity_orig * 100));
8752 * Check whether a rq has a misfit task and if it looks like we can actually
8753 * help that task: we can migrate the task to a CPU of higher capacity, or
8754 * the task's current CPU is heavily pressured.
8756 static inline int check_misfit_status(struct rq *rq, struct sched_domain *sd)
8758 return rq->misfit_task_load &&
8759 (rq->cpu_capacity_orig < rq->rd->max_cpu_capacity ||
8760 check_cpu_capacity(rq, sd));
8764 * Group imbalance indicates (and tries to solve) the problem where balancing
8765 * groups is inadequate due to ->cpus_ptr constraints.
8767 * Imagine a situation of two groups of 4 CPUs each and 4 tasks each with a
8768 * cpumask covering 1 CPU of the first group and 3 CPUs of the second group.
8771 * { 0 1 2 3 } { 4 5 6 7 }
8774 * If we were to balance group-wise we'd place two tasks in the first group and
8775 * two tasks in the second group. Clearly this is undesired as it will overload
8776 * cpu 3 and leave one of the CPUs in the second group unused.
8778 * The current solution to this issue is detecting the skew in the first group
8779 * by noticing the lower domain failed to reach balance and had difficulty
8780 * moving tasks due to affinity constraints.
8782 * When this is so detected; this group becomes a candidate for busiest; see
8783 * update_sd_pick_busiest(). And calculate_imbalance() and
8784 * find_busiest_group() avoid some of the usual balance conditions to allow it
8785 * to create an effective group imbalance.
8787 * This is a somewhat tricky proposition since the next run might not find the
8788 * group imbalance and decide the groups need to be balanced again. A most
8789 * subtle and fragile situation.
8792 static inline int sg_imbalanced(struct sched_group *group)
8794 return group->sgc->imbalance;
8798 * group_has_capacity returns true if the group has spare capacity that could
8799 * be used by some tasks.
8800 * We consider that a group has spare capacity if the number of task is
8801 * smaller than the number of CPUs or if the utilization is lower than the
8802 * available capacity for CFS tasks.
8803 * For the latter, we use a threshold to stabilize the state, to take into
8804 * account the variance of the tasks' load and to return true if the available
8805 * capacity in meaningful for the load balancer.
8806 * As an example, an available capacity of 1% can appear but it doesn't make
8807 * any benefit for the load balance.
8810 group_has_capacity(unsigned int imbalance_pct, struct sg_lb_stats *sgs)
8812 if (sgs->sum_nr_running < sgs->group_weight)
8815 if ((sgs->group_capacity * imbalance_pct) <
8816 (sgs->group_runnable * 100))
8819 if ((sgs->group_capacity * 100) >
8820 (sgs->group_util * imbalance_pct))
8827 * group_is_overloaded returns true if the group has more tasks than it can
8829 * group_is_overloaded is not equals to !group_has_capacity because a group
8830 * with the exact right number of tasks, has no more spare capacity but is not
8831 * overloaded so both group_has_capacity and group_is_overloaded return
8835 group_is_overloaded(unsigned int imbalance_pct, struct sg_lb_stats *sgs)
8837 if (sgs->sum_nr_running <= sgs->group_weight)
8840 if ((sgs->group_capacity * 100) <
8841 (sgs->group_util * imbalance_pct))
8844 if ((sgs->group_capacity * imbalance_pct) <
8845 (sgs->group_runnable * 100))
8852 group_type group_classify(unsigned int imbalance_pct,
8853 struct sched_group *group,
8854 struct sg_lb_stats *sgs)
8856 if (group_is_overloaded(imbalance_pct, sgs))
8857 return group_overloaded;
8859 if (sg_imbalanced(group))
8860 return group_imbalanced;
8862 if (sgs->group_asym_packing)
8863 return group_asym_packing;
8865 if (sgs->group_misfit_task_load)
8866 return group_misfit_task;
8868 if (!group_has_capacity(imbalance_pct, sgs))
8869 return group_fully_busy;
8871 return group_has_spare;
8875 * asym_smt_can_pull_tasks - Check whether the load balancing CPU can pull tasks
8876 * @dst_cpu: Destination CPU of the load balancing
8877 * @sds: Load-balancing data with statistics of the local group
8878 * @sgs: Load-balancing statistics of the candidate busiest group
8879 * @sg: The candidate busiest group
8881 * Check the state of the SMT siblings of both @sds::local and @sg and decide
8882 * if @dst_cpu can pull tasks.
8884 * If @dst_cpu does not have SMT siblings, it can pull tasks if two or more of
8885 * the SMT siblings of @sg are busy. If only one CPU in @sg is busy, pull tasks
8886 * only if @dst_cpu has higher priority.
8888 * If both @dst_cpu and @sg have SMT siblings, and @sg has exactly one more
8889 * busy CPU than @sds::local, let @dst_cpu pull tasks if it has higher priority.
8890 * Bigger imbalances in the number of busy CPUs will be dealt with in
8891 * update_sd_pick_busiest().
8893 * If @sg does not have SMT siblings, only pull tasks if all of the SMT siblings
8894 * of @dst_cpu are idle and @sg has lower priority.
8896 * Return: true if @dst_cpu can pull tasks, false otherwise.
8898 static bool asym_smt_can_pull_tasks(int dst_cpu, struct sd_lb_stats *sds,
8899 struct sg_lb_stats *sgs,
8900 struct sched_group *sg)
8902 #ifdef CONFIG_SCHED_SMT
8903 bool local_is_smt, sg_is_smt;
8906 local_is_smt = sds->local->flags & SD_SHARE_CPUCAPACITY;
8907 sg_is_smt = sg->flags & SD_SHARE_CPUCAPACITY;
8909 sg_busy_cpus = sgs->group_weight - sgs->idle_cpus;
8911 if (!local_is_smt) {
8913 * If we are here, @dst_cpu is idle and does not have SMT
8914 * siblings. Pull tasks if candidate group has two or more
8917 if (sg_busy_cpus >= 2) /* implies sg_is_smt */
8921 * @dst_cpu does not have SMT siblings. @sg may have SMT
8922 * siblings and only one is busy. In such case, @dst_cpu
8923 * can help if it has higher priority and is idle (i.e.,
8924 * it has no running tasks).
8926 return sched_asym_prefer(dst_cpu, sg->asym_prefer_cpu);
8929 /* @dst_cpu has SMT siblings. */
8932 int local_busy_cpus = sds->local->group_weight -
8933 sds->local_stat.idle_cpus;
8934 int busy_cpus_delta = sg_busy_cpus - local_busy_cpus;
8936 if (busy_cpus_delta == 1)
8937 return sched_asym_prefer(dst_cpu, sg->asym_prefer_cpu);
8943 * @sg does not have SMT siblings. Ensure that @sds::local does not end
8944 * up with more than one busy SMT sibling and only pull tasks if there
8945 * are not busy CPUs (i.e., no CPU has running tasks).
8947 if (!sds->local_stat.sum_nr_running)
8948 return sched_asym_prefer(dst_cpu, sg->asym_prefer_cpu);
8952 /* Always return false so that callers deal with non-SMT cases. */
8958 sched_asym(struct lb_env *env, struct sd_lb_stats *sds, struct sg_lb_stats *sgs,
8959 struct sched_group *group)
8961 /* Only do SMT checks if either local or candidate have SMT siblings */
8962 if ((sds->local->flags & SD_SHARE_CPUCAPACITY) ||
8963 (group->flags & SD_SHARE_CPUCAPACITY))
8964 return asym_smt_can_pull_tasks(env->dst_cpu, sds, sgs, group);
8966 return sched_asym_prefer(env->dst_cpu, group->asym_prefer_cpu);
8970 sched_reduced_capacity(struct rq *rq, struct sched_domain *sd)
8973 * When there is more than 1 task, the group_overloaded case already
8974 * takes care of cpu with reduced capacity
8976 if (rq->cfs.h_nr_running != 1)
8979 return check_cpu_capacity(rq, sd);
8983 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
8984 * @env: The load balancing environment.
8985 * @sds: Load-balancing data with statistics of the local group.
8986 * @group: sched_group whose statistics are to be updated.
8987 * @sgs: variable to hold the statistics for this group.
8988 * @sg_status: Holds flag indicating the status of the sched_group
8990 static inline void update_sg_lb_stats(struct lb_env *env,
8991 struct sd_lb_stats *sds,
8992 struct sched_group *group,
8993 struct sg_lb_stats *sgs,
8996 int i, nr_running, local_group;
8998 memset(sgs, 0, sizeof(*sgs));
9000 local_group = group == sds->local;
9002 for_each_cpu_and(i, sched_group_span(group), env->cpus) {
9003 struct rq *rq = cpu_rq(i);
9004 unsigned long load = cpu_load(rq);
9006 sgs->group_load += load;
9007 sgs->group_util += cpu_util_cfs(i);
9008 sgs->group_runnable += cpu_runnable(rq);
9009 sgs->sum_h_nr_running += rq->cfs.h_nr_running;
9011 nr_running = rq->nr_running;
9012 sgs->sum_nr_running += nr_running;
9015 *sg_status |= SG_OVERLOAD;
9017 if (cpu_overutilized(i))
9018 *sg_status |= SG_OVERUTILIZED;
9020 #ifdef CONFIG_NUMA_BALANCING
9021 sgs->nr_numa_running += rq->nr_numa_running;
9022 sgs->nr_preferred_running += rq->nr_preferred_running;
9025 * No need to call idle_cpu() if nr_running is not 0
9027 if (!nr_running && idle_cpu(i)) {
9029 /* Idle cpu can't have misfit task */
9036 if (env->sd->flags & SD_ASYM_CPUCAPACITY) {
9037 /* Check for a misfit task on the cpu */
9038 if (sgs->group_misfit_task_load < rq->misfit_task_load) {
9039 sgs->group_misfit_task_load = rq->misfit_task_load;
9040 *sg_status |= SG_OVERLOAD;
9042 } else if ((env->idle != CPU_NOT_IDLE) &&
9043 sched_reduced_capacity(rq, env->sd)) {
9044 /* Check for a task running on a CPU with reduced capacity */
9045 if (sgs->group_misfit_task_load < load)
9046 sgs->group_misfit_task_load = load;
9050 sgs->group_capacity = group->sgc->capacity;
9052 sgs->group_weight = group->group_weight;
9054 /* Check if dst CPU is idle and preferred to this group */
9055 if (!local_group && env->sd->flags & SD_ASYM_PACKING &&
9056 env->idle != CPU_NOT_IDLE && sgs->sum_h_nr_running &&
9057 sched_asym(env, sds, sgs, group)) {
9058 sgs->group_asym_packing = 1;
9061 sgs->group_type = group_classify(env->sd->imbalance_pct, group, sgs);
9063 /* Computing avg_load makes sense only when group is overloaded */
9064 if (sgs->group_type == group_overloaded)
9065 sgs->avg_load = (sgs->group_load * SCHED_CAPACITY_SCALE) /
9066 sgs->group_capacity;
9070 * update_sd_pick_busiest - return 1 on busiest group
9071 * @env: The load balancing environment.
9072 * @sds: sched_domain statistics
9073 * @sg: sched_group candidate to be checked for being the busiest
9074 * @sgs: sched_group statistics
9076 * Determine if @sg is a busier group than the previously selected
9079 * Return: %true if @sg is a busier group than the previously selected
9080 * busiest group. %false otherwise.
9082 static bool update_sd_pick_busiest(struct lb_env *env,
9083 struct sd_lb_stats *sds,
9084 struct sched_group *sg,
9085 struct sg_lb_stats *sgs)
9087 struct sg_lb_stats *busiest = &sds->busiest_stat;
9089 /* Make sure that there is at least one task to pull */
9090 if (!sgs->sum_h_nr_running)
9094 * Don't try to pull misfit tasks we can't help.
9095 * We can use max_capacity here as reduction in capacity on some
9096 * CPUs in the group should either be possible to resolve
9097 * internally or be covered by avg_load imbalance (eventually).
9099 if ((env->sd->flags & SD_ASYM_CPUCAPACITY) &&
9100 (sgs->group_type == group_misfit_task) &&
9101 (!capacity_greater(capacity_of(env->dst_cpu), sg->sgc->max_capacity) ||
9102 sds->local_stat.group_type != group_has_spare))
9105 if (sgs->group_type > busiest->group_type)
9108 if (sgs->group_type < busiest->group_type)
9112 * The candidate and the current busiest group are the same type of
9113 * group. Let check which one is the busiest according to the type.
9116 switch (sgs->group_type) {
9117 case group_overloaded:
9118 /* Select the overloaded group with highest avg_load. */
9119 if (sgs->avg_load <= busiest->avg_load)
9123 case group_imbalanced:
9125 * Select the 1st imbalanced group as we don't have any way to
9126 * choose one more than another.
9130 case group_asym_packing:
9131 /* Prefer to move from lowest priority CPU's work */
9132 if (sched_asym_prefer(sg->asym_prefer_cpu, sds->busiest->asym_prefer_cpu))
9136 case group_misfit_task:
9138 * If we have more than one misfit sg go with the biggest
9141 if (sgs->group_misfit_task_load < busiest->group_misfit_task_load)
9145 case group_fully_busy:
9147 * Select the fully busy group with highest avg_load. In
9148 * theory, there is no need to pull task from such kind of
9149 * group because tasks have all compute capacity that they need
9150 * but we can still improve the overall throughput by reducing
9151 * contention when accessing shared HW resources.
9153 * XXX for now avg_load is not computed and always 0 so we
9154 * select the 1st one.
9156 if (sgs->avg_load <= busiest->avg_load)
9160 case group_has_spare:
9162 * Select not overloaded group with lowest number of idle cpus
9163 * and highest number of running tasks. We could also compare
9164 * the spare capacity which is more stable but it can end up
9165 * that the group has less spare capacity but finally more idle
9166 * CPUs which means less opportunity to pull tasks.
9168 if (sgs->idle_cpus > busiest->idle_cpus)
9170 else if ((sgs->idle_cpus == busiest->idle_cpus) &&
9171 (sgs->sum_nr_running <= busiest->sum_nr_running))
9178 * Candidate sg has no more than one task per CPU and has higher
9179 * per-CPU capacity. Migrating tasks to less capable CPUs may harm
9180 * throughput. Maximize throughput, power/energy consequences are not
9183 if ((env->sd->flags & SD_ASYM_CPUCAPACITY) &&
9184 (sgs->group_type <= group_fully_busy) &&
9185 (capacity_greater(sg->sgc->min_capacity, capacity_of(env->dst_cpu))))
9191 #ifdef CONFIG_NUMA_BALANCING
9192 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
9194 if (sgs->sum_h_nr_running > sgs->nr_numa_running)
9196 if (sgs->sum_h_nr_running > sgs->nr_preferred_running)
9201 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
9203 if (rq->nr_running > rq->nr_numa_running)
9205 if (rq->nr_running > rq->nr_preferred_running)
9210 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
9215 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
9219 #endif /* CONFIG_NUMA_BALANCING */
9225 * task_running_on_cpu - return 1 if @p is running on @cpu.
9228 static unsigned int task_running_on_cpu(int cpu, struct task_struct *p)
9230 /* Task has no contribution or is new */
9231 if (cpu != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time))
9234 if (task_on_rq_queued(p))
9241 * idle_cpu_without - would a given CPU be idle without p ?
9242 * @cpu: the processor on which idleness is tested.
9243 * @p: task which should be ignored.
9245 * Return: 1 if the CPU would be idle. 0 otherwise.
9247 static int idle_cpu_without(int cpu, struct task_struct *p)
9249 struct rq *rq = cpu_rq(cpu);
9251 if (rq->curr != rq->idle && rq->curr != p)
9255 * rq->nr_running can't be used but an updated version without the
9256 * impact of p on cpu must be used instead. The updated nr_running
9257 * be computed and tested before calling idle_cpu_without().
9261 if (rq->ttwu_pending)
9269 * update_sg_wakeup_stats - Update sched_group's statistics for wakeup.
9270 * @sd: The sched_domain level to look for idlest group.
9271 * @group: sched_group whose statistics are to be updated.
9272 * @sgs: variable to hold the statistics for this group.
9273 * @p: The task for which we look for the idlest group/CPU.
9275 static inline void update_sg_wakeup_stats(struct sched_domain *sd,
9276 struct sched_group *group,
9277 struct sg_lb_stats *sgs,
9278 struct task_struct *p)
9282 memset(sgs, 0, sizeof(*sgs));
9284 for_each_cpu(i, sched_group_span(group)) {
9285 struct rq *rq = cpu_rq(i);
9288 sgs->group_load += cpu_load_without(rq, p);
9289 sgs->group_util += cpu_util_without(i, p);
9290 sgs->group_runnable += cpu_runnable_without(rq, p);
9291 local = task_running_on_cpu(i, p);
9292 sgs->sum_h_nr_running += rq->cfs.h_nr_running - local;
9294 nr_running = rq->nr_running - local;
9295 sgs->sum_nr_running += nr_running;
9298 * No need to call idle_cpu_without() if nr_running is not 0
9300 if (!nr_running && idle_cpu_without(i, p))
9305 /* Check if task fits in the group */
9306 if (sd->flags & SD_ASYM_CPUCAPACITY &&
9307 !task_fits_capacity(p, group->sgc->max_capacity)) {
9308 sgs->group_misfit_task_load = 1;
9311 sgs->group_capacity = group->sgc->capacity;
9313 sgs->group_weight = group->group_weight;
9315 sgs->group_type = group_classify(sd->imbalance_pct, group, sgs);
9318 * Computing avg_load makes sense only when group is fully busy or
9321 if (sgs->group_type == group_fully_busy ||
9322 sgs->group_type == group_overloaded)
9323 sgs->avg_load = (sgs->group_load * SCHED_CAPACITY_SCALE) /
9324 sgs->group_capacity;
9327 static bool update_pick_idlest(struct sched_group *idlest,
9328 struct sg_lb_stats *idlest_sgs,
9329 struct sched_group *group,
9330 struct sg_lb_stats *sgs)
9332 if (sgs->group_type < idlest_sgs->group_type)
9335 if (sgs->group_type > idlest_sgs->group_type)
9339 * The candidate and the current idlest group are the same type of
9340 * group. Let check which one is the idlest according to the type.
9343 switch (sgs->group_type) {
9344 case group_overloaded:
9345 case group_fully_busy:
9346 /* Select the group with lowest avg_load. */
9347 if (idlest_sgs->avg_load <= sgs->avg_load)
9351 case group_imbalanced:
9352 case group_asym_packing:
9353 /* Those types are not used in the slow wakeup path */
9356 case group_misfit_task:
9357 /* Select group with the highest max capacity */
9358 if (idlest->sgc->max_capacity >= group->sgc->max_capacity)
9362 case group_has_spare:
9363 /* Select group with most idle CPUs */
9364 if (idlest_sgs->idle_cpus > sgs->idle_cpus)
9367 /* Select group with lowest group_util */
9368 if (idlest_sgs->idle_cpus == sgs->idle_cpus &&
9369 idlest_sgs->group_util <= sgs->group_util)
9379 * find_idlest_group() finds and returns the least busy CPU group within the
9382 * Assumes p is allowed on at least one CPU in sd.
9384 static struct sched_group *
9385 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
9387 struct sched_group *idlest = NULL, *local = NULL, *group = sd->groups;
9388 struct sg_lb_stats local_sgs, tmp_sgs;
9389 struct sg_lb_stats *sgs;
9390 unsigned long imbalance;
9391 struct sg_lb_stats idlest_sgs = {
9392 .avg_load = UINT_MAX,
9393 .group_type = group_overloaded,
9399 /* Skip over this group if it has no CPUs allowed */
9400 if (!cpumask_intersects(sched_group_span(group),
9404 /* Skip over this group if no cookie matched */
9405 if (!sched_group_cookie_match(cpu_rq(this_cpu), p, group))
9408 local_group = cpumask_test_cpu(this_cpu,
9409 sched_group_span(group));
9418 update_sg_wakeup_stats(sd, group, sgs, p);
9420 if (!local_group && update_pick_idlest(idlest, &idlest_sgs, group, sgs)) {
9425 } while (group = group->next, group != sd->groups);
9428 /* There is no idlest group to push tasks to */
9432 /* The local group has been skipped because of CPU affinity */
9437 * If the local group is idler than the selected idlest group
9438 * don't try and push the task.
9440 if (local_sgs.group_type < idlest_sgs.group_type)
9444 * If the local group is busier than the selected idlest group
9445 * try and push the task.
9447 if (local_sgs.group_type > idlest_sgs.group_type)
9450 switch (local_sgs.group_type) {
9451 case group_overloaded:
9452 case group_fully_busy:
9454 /* Calculate allowed imbalance based on load */
9455 imbalance = scale_load_down(NICE_0_LOAD) *
9456 (sd->imbalance_pct-100) / 100;
9459 * When comparing groups across NUMA domains, it's possible for
9460 * the local domain to be very lightly loaded relative to the
9461 * remote domains but "imbalance" skews the comparison making
9462 * remote CPUs look much more favourable. When considering
9463 * cross-domain, add imbalance to the load on the remote node
9464 * and consider staying local.
9467 if ((sd->flags & SD_NUMA) &&
9468 ((idlest_sgs.avg_load + imbalance) >= local_sgs.avg_load))
9472 * If the local group is less loaded than the selected
9473 * idlest group don't try and push any tasks.
9475 if (idlest_sgs.avg_load >= (local_sgs.avg_load + imbalance))
9478 if (100 * local_sgs.avg_load <= sd->imbalance_pct * idlest_sgs.avg_load)
9482 case group_imbalanced:
9483 case group_asym_packing:
9484 /* Those type are not used in the slow wakeup path */
9487 case group_misfit_task:
9488 /* Select group with the highest max capacity */
9489 if (local->sgc->max_capacity >= idlest->sgc->max_capacity)
9493 case group_has_spare:
9495 if (sd->flags & SD_NUMA) {
9496 int imb_numa_nr = sd->imb_numa_nr;
9497 #ifdef CONFIG_NUMA_BALANCING
9500 * If there is spare capacity at NUMA, try to select
9501 * the preferred node
9503 if (cpu_to_node(this_cpu) == p->numa_preferred_nid)
9506 idlest_cpu = cpumask_first(sched_group_span(idlest));
9507 if (cpu_to_node(idlest_cpu) == p->numa_preferred_nid)
9509 #endif /* CONFIG_NUMA_BALANCING */
9511 * Otherwise, keep the task close to the wakeup source
9512 * and improve locality if the number of running tasks
9513 * would remain below threshold where an imbalance is
9514 * allowed while accounting for the possibility the
9515 * task is pinned to a subset of CPUs. If there is a
9516 * real need of migration, periodic load balance will
9519 if (p->nr_cpus_allowed != NR_CPUS) {
9520 struct cpumask *cpus = this_cpu_cpumask_var_ptr(select_rq_mask);
9522 cpumask_and(cpus, sched_group_span(local), p->cpus_ptr);
9523 imb_numa_nr = min(cpumask_weight(cpus), sd->imb_numa_nr);
9526 imbalance = abs(local_sgs.idle_cpus - idlest_sgs.idle_cpus);
9527 if (!adjust_numa_imbalance(imbalance,
9528 local_sgs.sum_nr_running + 1,
9533 #endif /* CONFIG_NUMA */
9536 * Select group with highest number of idle CPUs. We could also
9537 * compare the utilization which is more stable but it can end
9538 * up that the group has less spare capacity but finally more
9539 * idle CPUs which means more opportunity to run task.
9541 if (local_sgs.idle_cpus >= idlest_sgs.idle_cpus)
9549 static void update_idle_cpu_scan(struct lb_env *env,
9550 unsigned long sum_util)
9552 struct sched_domain_shared *sd_share;
9553 int llc_weight, pct;
9556 * Update the number of CPUs to scan in LLC domain, which could
9557 * be used as a hint in select_idle_cpu(). The update of sd_share
9558 * could be expensive because it is within a shared cache line.
9559 * So the write of this hint only occurs during periodic load
9560 * balancing, rather than CPU_NEWLY_IDLE, because the latter
9561 * can fire way more frequently than the former.
9563 if (!sched_feat(SIS_UTIL) || env->idle == CPU_NEWLY_IDLE)
9566 llc_weight = per_cpu(sd_llc_size, env->dst_cpu);
9567 if (env->sd->span_weight != llc_weight)
9570 sd_share = rcu_dereference(per_cpu(sd_llc_shared, env->dst_cpu));
9575 * The number of CPUs to search drops as sum_util increases, when
9576 * sum_util hits 85% or above, the scan stops.
9577 * The reason to choose 85% as the threshold is because this is the
9578 * imbalance_pct(117) when a LLC sched group is overloaded.
9580 * let y = SCHED_CAPACITY_SCALE - p * x^2 [1]
9581 * and y'= y / SCHED_CAPACITY_SCALE
9583 * x is the ratio of sum_util compared to the CPU capacity:
9584 * x = sum_util / (llc_weight * SCHED_CAPACITY_SCALE)
9585 * y' is the ratio of CPUs to be scanned in the LLC domain,
9586 * and the number of CPUs to scan is calculated by:
9588 * nr_scan = llc_weight * y' [2]
9590 * When x hits the threshold of overloaded, AKA, when
9591 * x = 100 / pct, y drops to 0. According to [1],
9592 * p should be SCHED_CAPACITY_SCALE * pct^2 / 10000
9594 * Scale x by SCHED_CAPACITY_SCALE:
9595 * x' = sum_util / llc_weight; [3]
9597 * and finally [1] becomes:
9598 * y = SCHED_CAPACITY_SCALE -
9599 * x'^2 * pct^2 / (10000 * SCHED_CAPACITY_SCALE) [4]
9604 do_div(x, llc_weight);
9607 pct = env->sd->imbalance_pct;
9608 tmp = x * x * pct * pct;
9609 do_div(tmp, 10000 * SCHED_CAPACITY_SCALE);
9610 tmp = min_t(long, tmp, SCHED_CAPACITY_SCALE);
9611 y = SCHED_CAPACITY_SCALE - tmp;
9615 do_div(y, SCHED_CAPACITY_SCALE);
9616 if ((int)y != sd_share->nr_idle_scan)
9617 WRITE_ONCE(sd_share->nr_idle_scan, (int)y);
9621 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
9622 * @env: The load balancing environment.
9623 * @sds: variable to hold the statistics for this sched_domain.
9626 static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
9628 struct sched_domain *child = env->sd->child;
9629 struct sched_group *sg = env->sd->groups;
9630 struct sg_lb_stats *local = &sds->local_stat;
9631 struct sg_lb_stats tmp_sgs;
9632 unsigned long sum_util = 0;
9636 struct sg_lb_stats *sgs = &tmp_sgs;
9639 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_span(sg));
9644 if (env->idle != CPU_NEWLY_IDLE ||
9645 time_after_eq(jiffies, sg->sgc->next_update))
9646 update_group_capacity(env->sd, env->dst_cpu);
9649 update_sg_lb_stats(env, sds, sg, sgs, &sg_status);
9655 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
9657 sds->busiest_stat = *sgs;
9661 /* Now, start updating sd_lb_stats */
9662 sds->total_load += sgs->group_load;
9663 sds->total_capacity += sgs->group_capacity;
9665 sum_util += sgs->group_util;
9667 } while (sg != env->sd->groups);
9669 /* Tag domain that child domain prefers tasks go to siblings first */
9670 sds->prefer_sibling = child && child->flags & SD_PREFER_SIBLING;
9673 if (env->sd->flags & SD_NUMA)
9674 env->fbq_type = fbq_classify_group(&sds->busiest_stat);
9676 if (!env->sd->parent) {
9677 struct root_domain *rd = env->dst_rq->rd;
9679 /* update overload indicator if we are at root domain */
9680 WRITE_ONCE(rd->overload, sg_status & SG_OVERLOAD);
9682 /* Update over-utilization (tipping point, U >= 0) indicator */
9683 WRITE_ONCE(rd->overutilized, sg_status & SG_OVERUTILIZED);
9684 trace_sched_overutilized_tp(rd, sg_status & SG_OVERUTILIZED);
9685 } else if (sg_status & SG_OVERUTILIZED) {
9686 struct root_domain *rd = env->dst_rq->rd;
9688 WRITE_ONCE(rd->overutilized, SG_OVERUTILIZED);
9689 trace_sched_overutilized_tp(rd, SG_OVERUTILIZED);
9692 update_idle_cpu_scan(env, sum_util);
9696 * calculate_imbalance - Calculate the amount of imbalance present within the
9697 * groups of a given sched_domain during load balance.
9698 * @env: load balance environment
9699 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
9701 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
9703 struct sg_lb_stats *local, *busiest;
9705 local = &sds->local_stat;
9706 busiest = &sds->busiest_stat;
9708 if (busiest->group_type == group_misfit_task) {
9709 if (env->sd->flags & SD_ASYM_CPUCAPACITY) {
9710 /* Set imbalance to allow misfit tasks to be balanced. */
9711 env->migration_type = migrate_misfit;
9715 * Set load imbalance to allow moving task from cpu
9716 * with reduced capacity.
9718 env->migration_type = migrate_load;
9719 env->imbalance = busiest->group_misfit_task_load;
9724 if (busiest->group_type == group_asym_packing) {
9726 * In case of asym capacity, we will try to migrate all load to
9727 * the preferred CPU.
9729 env->migration_type = migrate_task;
9730 env->imbalance = busiest->sum_h_nr_running;
9734 if (busiest->group_type == group_imbalanced) {
9736 * In the group_imb case we cannot rely on group-wide averages
9737 * to ensure CPU-load equilibrium, try to move any task to fix
9738 * the imbalance. The next load balance will take care of
9739 * balancing back the system.
9741 env->migration_type = migrate_task;
9747 * Try to use spare capacity of local group without overloading it or
9750 if (local->group_type == group_has_spare) {
9751 if ((busiest->group_type > group_fully_busy) &&
9752 !(env->sd->flags & SD_SHARE_PKG_RESOURCES)) {
9754 * If busiest is overloaded, try to fill spare
9755 * capacity. This might end up creating spare capacity
9756 * in busiest or busiest still being overloaded but
9757 * there is no simple way to directly compute the
9758 * amount of load to migrate in order to balance the
9761 env->migration_type = migrate_util;
9762 env->imbalance = max(local->group_capacity, local->group_util) -
9766 * In some cases, the group's utilization is max or even
9767 * higher than capacity because of migrations but the
9768 * local CPU is (newly) idle. There is at least one
9769 * waiting task in this overloaded busiest group. Let's
9772 if (env->idle != CPU_NOT_IDLE && env->imbalance == 0) {
9773 env->migration_type = migrate_task;
9780 if (busiest->group_weight == 1 || sds->prefer_sibling) {
9781 unsigned int nr_diff = busiest->sum_nr_running;
9783 * When prefer sibling, evenly spread running tasks on
9786 env->migration_type = migrate_task;
9787 lsub_positive(&nr_diff, local->sum_nr_running);
9788 env->imbalance = nr_diff;
9792 * If there is no overload, we just want to even the number of
9795 env->migration_type = migrate_task;
9796 env->imbalance = max_t(long, 0,
9797 (local->idle_cpus - busiest->idle_cpus));
9801 /* Consider allowing a small imbalance between NUMA groups */
9802 if (env->sd->flags & SD_NUMA) {
9803 env->imbalance = adjust_numa_imbalance(env->imbalance,
9804 local->sum_nr_running + 1,
9805 env->sd->imb_numa_nr);
9809 /* Number of tasks to move to restore balance */
9810 env->imbalance >>= 1;
9816 * Local is fully busy but has to take more load to relieve the
9819 if (local->group_type < group_overloaded) {
9821 * Local will become overloaded so the avg_load metrics are
9825 local->avg_load = (local->group_load * SCHED_CAPACITY_SCALE) /
9826 local->group_capacity;
9829 * If the local group is more loaded than the selected
9830 * busiest group don't try to pull any tasks.
9832 if (local->avg_load >= busiest->avg_load) {
9837 sds->avg_load = (sds->total_load * SCHED_CAPACITY_SCALE) /
9838 sds->total_capacity;
9842 * Both group are or will become overloaded and we're trying to get all
9843 * the CPUs to the average_load, so we don't want to push ourselves
9844 * above the average load, nor do we wish to reduce the max loaded CPU
9845 * below the average load. At the same time, we also don't want to
9846 * reduce the group load below the group capacity. Thus we look for
9847 * the minimum possible imbalance.
9849 env->migration_type = migrate_load;
9850 env->imbalance = min(
9851 (busiest->avg_load - sds->avg_load) * busiest->group_capacity,
9852 (sds->avg_load - local->avg_load) * local->group_capacity
9853 ) / SCHED_CAPACITY_SCALE;
9856 /******* find_busiest_group() helpers end here *********************/
9859 * Decision matrix according to the local and busiest group type:
9861 * busiest \ local has_spare fully_busy misfit asym imbalanced overloaded
9862 * has_spare nr_idle balanced N/A N/A balanced balanced
9863 * fully_busy nr_idle nr_idle N/A N/A balanced balanced
9864 * misfit_task force N/A N/A N/A N/A N/A
9865 * asym_packing force force N/A N/A force force
9866 * imbalanced force force N/A N/A force force
9867 * overloaded force force N/A N/A force avg_load
9869 * N/A : Not Applicable because already filtered while updating
9871 * balanced : The system is balanced for these 2 groups.
9872 * force : Calculate the imbalance as load migration is probably needed.
9873 * avg_load : Only if imbalance is significant enough.
9874 * nr_idle : dst_cpu is not busy and the number of idle CPUs is quite
9875 * different in groups.
9879 * find_busiest_group - Returns the busiest group within the sched_domain
9880 * if there is an imbalance.
9881 * @env: The load balancing environment.
9883 * Also calculates the amount of runnable load which should be moved
9884 * to restore balance.
9886 * Return: - The busiest group if imbalance exists.
9888 static struct sched_group *find_busiest_group(struct lb_env *env)
9890 struct sg_lb_stats *local, *busiest;
9891 struct sd_lb_stats sds;
9893 init_sd_lb_stats(&sds);
9896 * Compute the various statistics relevant for load balancing at
9899 update_sd_lb_stats(env, &sds);
9901 if (sched_energy_enabled()) {
9902 struct root_domain *rd = env->dst_rq->rd;
9904 if (rcu_dereference(rd->pd) && !READ_ONCE(rd->overutilized))
9908 local = &sds.local_stat;
9909 busiest = &sds.busiest_stat;
9911 /* There is no busy sibling group to pull tasks from */
9915 /* Misfit tasks should be dealt with regardless of the avg load */
9916 if (busiest->group_type == group_misfit_task)
9919 /* ASYM feature bypasses nice load balance check */
9920 if (busiest->group_type == group_asym_packing)
9924 * If the busiest group is imbalanced the below checks don't
9925 * work because they assume all things are equal, which typically
9926 * isn't true due to cpus_ptr constraints and the like.
9928 if (busiest->group_type == group_imbalanced)
9932 * If the local group is busier than the selected busiest group
9933 * don't try and pull any tasks.
9935 if (local->group_type > busiest->group_type)
9939 * When groups are overloaded, use the avg_load to ensure fairness
9942 if (local->group_type == group_overloaded) {
9944 * If the local group is more loaded than the selected
9945 * busiest group don't try to pull any tasks.
9947 if (local->avg_load >= busiest->avg_load)
9950 /* XXX broken for overlapping NUMA groups */
9951 sds.avg_load = (sds.total_load * SCHED_CAPACITY_SCALE) /
9955 * Don't pull any tasks if this group is already above the
9956 * domain average load.
9958 if (local->avg_load >= sds.avg_load)
9962 * If the busiest group is more loaded, use imbalance_pct to be
9965 if (100 * busiest->avg_load <=
9966 env->sd->imbalance_pct * local->avg_load)
9970 /* Try to move all excess tasks to child's sibling domain */
9971 if (sds.prefer_sibling && local->group_type == group_has_spare &&
9972 busiest->sum_nr_running > local->sum_nr_running + 1)
9975 if (busiest->group_type != group_overloaded) {
9976 if (env->idle == CPU_NOT_IDLE)
9978 * If the busiest group is not overloaded (and as a
9979 * result the local one too) but this CPU is already
9980 * busy, let another idle CPU try to pull task.
9984 if (busiest->group_weight > 1 &&
9985 local->idle_cpus <= (busiest->idle_cpus + 1))
9987 * If the busiest group is not overloaded
9988 * and there is no imbalance between this and busiest
9989 * group wrt idle CPUs, it is balanced. The imbalance
9990 * becomes significant if the diff is greater than 1
9991 * otherwise we might end up to just move the imbalance
9992 * on another group. Of course this applies only if
9993 * there is more than 1 CPU per group.
9997 if (busiest->sum_h_nr_running == 1)
9999 * busiest doesn't have any tasks waiting to run
10005 /* Looks like there is an imbalance. Compute it */
10006 calculate_imbalance(env, &sds);
10007 return env->imbalance ? sds.busiest : NULL;
10010 env->imbalance = 0;
10015 * find_busiest_queue - find the busiest runqueue among the CPUs in the group.
10017 static struct rq *find_busiest_queue(struct lb_env *env,
10018 struct sched_group *group)
10020 struct rq *busiest = NULL, *rq;
10021 unsigned long busiest_util = 0, busiest_load = 0, busiest_capacity = 1;
10022 unsigned int busiest_nr = 0;
10025 for_each_cpu_and(i, sched_group_span(group), env->cpus) {
10026 unsigned long capacity, load, util;
10027 unsigned int nr_running;
10031 rt = fbq_classify_rq(rq);
10034 * We classify groups/runqueues into three groups:
10035 * - regular: there are !numa tasks
10036 * - remote: there are numa tasks that run on the 'wrong' node
10037 * - all: there is no distinction
10039 * In order to avoid migrating ideally placed numa tasks,
10040 * ignore those when there's better options.
10042 * If we ignore the actual busiest queue to migrate another
10043 * task, the next balance pass can still reduce the busiest
10044 * queue by moving tasks around inside the node.
10046 * If we cannot move enough load due to this classification
10047 * the next pass will adjust the group classification and
10048 * allow migration of more tasks.
10050 * Both cases only affect the total convergence complexity.
10052 if (rt > env->fbq_type)
10055 nr_running = rq->cfs.h_nr_running;
10059 capacity = capacity_of(i);
10062 * For ASYM_CPUCAPACITY domains, don't pick a CPU that could
10063 * eventually lead to active_balancing high->low capacity.
10064 * Higher per-CPU capacity is considered better than balancing
10067 if (env->sd->flags & SD_ASYM_CPUCAPACITY &&
10068 !capacity_greater(capacity_of(env->dst_cpu), capacity) &&
10072 /* Make sure we only pull tasks from a CPU of lower priority */
10073 if ((env->sd->flags & SD_ASYM_PACKING) &&
10074 sched_asym_prefer(i, env->dst_cpu) &&
10078 switch (env->migration_type) {
10081 * When comparing with load imbalance, use cpu_load()
10082 * which is not scaled with the CPU capacity.
10084 load = cpu_load(rq);
10086 if (nr_running == 1 && load > env->imbalance &&
10087 !check_cpu_capacity(rq, env->sd))
10091 * For the load comparisons with the other CPUs,
10092 * consider the cpu_load() scaled with the CPU
10093 * capacity, so that the load can be moved away
10094 * from the CPU that is potentially running at a
10097 * Thus we're looking for max(load_i / capacity_i),
10098 * crosswise multiplication to rid ourselves of the
10099 * division works out to:
10100 * load_i * capacity_j > load_j * capacity_i;
10101 * where j is our previous maximum.
10103 if (load * busiest_capacity > busiest_load * capacity) {
10104 busiest_load = load;
10105 busiest_capacity = capacity;
10111 util = cpu_util_cfs(i);
10114 * Don't try to pull utilization from a CPU with one
10115 * running task. Whatever its utilization, we will fail
10118 if (nr_running <= 1)
10121 if (busiest_util < util) {
10122 busiest_util = util;
10128 if (busiest_nr < nr_running) {
10129 busiest_nr = nr_running;
10134 case migrate_misfit:
10136 * For ASYM_CPUCAPACITY domains with misfit tasks we
10137 * simply seek the "biggest" misfit task.
10139 if (rq->misfit_task_load > busiest_load) {
10140 busiest_load = rq->misfit_task_load;
10153 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
10154 * so long as it is large enough.
10156 #define MAX_PINNED_INTERVAL 512
10159 asym_active_balance(struct lb_env *env)
10162 * ASYM_PACKING needs to force migrate tasks from busy but
10163 * lower priority CPUs in order to pack all tasks in the
10164 * highest priority CPUs.
10166 return env->idle != CPU_NOT_IDLE && (env->sd->flags & SD_ASYM_PACKING) &&
10167 sched_asym_prefer(env->dst_cpu, env->src_cpu);
10171 imbalanced_active_balance(struct lb_env *env)
10173 struct sched_domain *sd = env->sd;
10176 * The imbalanced case includes the case of pinned tasks preventing a fair
10177 * distribution of the load on the system but also the even distribution of the
10178 * threads on a system with spare capacity
10180 if ((env->migration_type == migrate_task) &&
10181 (sd->nr_balance_failed > sd->cache_nice_tries+2))
10187 static int need_active_balance(struct lb_env *env)
10189 struct sched_domain *sd = env->sd;
10191 if (asym_active_balance(env))
10194 if (imbalanced_active_balance(env))
10198 * The dst_cpu is idle and the src_cpu CPU has only 1 CFS task.
10199 * It's worth migrating the task if the src_cpu's capacity is reduced
10200 * because of other sched_class or IRQs if more capacity stays
10201 * available on dst_cpu.
10203 if ((env->idle != CPU_NOT_IDLE) &&
10204 (env->src_rq->cfs.h_nr_running == 1)) {
10205 if ((check_cpu_capacity(env->src_rq, sd)) &&
10206 (capacity_of(env->src_cpu)*sd->imbalance_pct < capacity_of(env->dst_cpu)*100))
10210 if (env->migration_type == migrate_misfit)
10216 static int active_load_balance_cpu_stop(void *data);
10218 static int should_we_balance(struct lb_env *env)
10220 struct sched_group *sg = env->sd->groups;
10224 * Ensure the balancing environment is consistent; can happen
10225 * when the softirq triggers 'during' hotplug.
10227 if (!cpumask_test_cpu(env->dst_cpu, env->cpus))
10231 * In the newly idle case, we will allow all the CPUs
10232 * to do the newly idle load balance.
10234 * However, we bail out if we already have tasks or a wakeup pending,
10235 * to optimize wakeup latency.
10237 if (env->idle == CPU_NEWLY_IDLE) {
10238 if (env->dst_rq->nr_running > 0 || env->dst_rq->ttwu_pending)
10243 /* Try to find first idle CPU */
10244 for_each_cpu_and(cpu, group_balance_mask(sg), env->cpus) {
10245 if (!idle_cpu(cpu))
10248 /* Are we the first idle CPU? */
10249 return cpu == env->dst_cpu;
10252 /* Are we the first CPU of this group ? */
10253 return group_balance_cpu(sg) == env->dst_cpu;
10257 * Check this_cpu to ensure it is balanced within domain. Attempt to move
10258 * tasks if there is an imbalance.
10260 static int load_balance(int this_cpu, struct rq *this_rq,
10261 struct sched_domain *sd, enum cpu_idle_type idle,
10262 int *continue_balancing)
10264 int ld_moved, cur_ld_moved, active_balance = 0;
10265 struct sched_domain *sd_parent = sd->parent;
10266 struct sched_group *group;
10267 struct rq *busiest;
10268 struct rq_flags rf;
10269 struct cpumask *cpus = this_cpu_cpumask_var_ptr(load_balance_mask);
10270 struct lb_env env = {
10272 .dst_cpu = this_cpu,
10274 .dst_grpmask = sched_group_span(sd->groups),
10276 .loop_break = SCHED_NR_MIGRATE_BREAK,
10279 .tasks = LIST_HEAD_INIT(env.tasks),
10282 cpumask_and(cpus, sched_domain_span(sd), cpu_active_mask);
10284 schedstat_inc(sd->lb_count[idle]);
10287 if (!should_we_balance(&env)) {
10288 *continue_balancing = 0;
10292 group = find_busiest_group(&env);
10294 schedstat_inc(sd->lb_nobusyg[idle]);
10298 busiest = find_busiest_queue(&env, group);
10300 schedstat_inc(sd->lb_nobusyq[idle]);
10304 WARN_ON_ONCE(busiest == env.dst_rq);
10306 schedstat_add(sd->lb_imbalance[idle], env.imbalance);
10308 env.src_cpu = busiest->cpu;
10309 env.src_rq = busiest;
10312 /* Clear this flag as soon as we find a pullable task */
10313 env.flags |= LBF_ALL_PINNED;
10314 if (busiest->nr_running > 1) {
10316 * Attempt to move tasks. If find_busiest_group has found
10317 * an imbalance but busiest->nr_running <= 1, the group is
10318 * still unbalanced. ld_moved simply stays zero, so it is
10319 * correctly treated as an imbalance.
10321 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
10324 rq_lock_irqsave(busiest, &rf);
10325 update_rq_clock(busiest);
10328 * cur_ld_moved - load moved in current iteration
10329 * ld_moved - cumulative load moved across iterations
10331 cur_ld_moved = detach_tasks(&env);
10334 * We've detached some tasks from busiest_rq. Every
10335 * task is masked "TASK_ON_RQ_MIGRATING", so we can safely
10336 * unlock busiest->lock, and we are able to be sure
10337 * that nobody can manipulate the tasks in parallel.
10338 * See task_rq_lock() family for the details.
10341 rq_unlock(busiest, &rf);
10343 if (cur_ld_moved) {
10344 attach_tasks(&env);
10345 ld_moved += cur_ld_moved;
10348 local_irq_restore(rf.flags);
10350 if (env.flags & LBF_NEED_BREAK) {
10351 env.flags &= ~LBF_NEED_BREAK;
10352 /* Stop if we tried all running tasks */
10353 if (env.loop < busiest->nr_running)
10358 * Revisit (affine) tasks on src_cpu that couldn't be moved to
10359 * us and move them to an alternate dst_cpu in our sched_group
10360 * where they can run. The upper limit on how many times we
10361 * iterate on same src_cpu is dependent on number of CPUs in our
10364 * This changes load balance semantics a bit on who can move
10365 * load to a given_cpu. In addition to the given_cpu itself
10366 * (or a ilb_cpu acting on its behalf where given_cpu is
10367 * nohz-idle), we now have balance_cpu in a position to move
10368 * load to given_cpu. In rare situations, this may cause
10369 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
10370 * _independently_ and at _same_ time to move some load to
10371 * given_cpu) causing excess load to be moved to given_cpu.
10372 * This however should not happen so much in practice and
10373 * moreover subsequent load balance cycles should correct the
10374 * excess load moved.
10376 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
10378 /* Prevent to re-select dst_cpu via env's CPUs */
10379 __cpumask_clear_cpu(env.dst_cpu, env.cpus);
10381 env.dst_rq = cpu_rq(env.new_dst_cpu);
10382 env.dst_cpu = env.new_dst_cpu;
10383 env.flags &= ~LBF_DST_PINNED;
10385 env.loop_break = SCHED_NR_MIGRATE_BREAK;
10388 * Go back to "more_balance" rather than "redo" since we
10389 * need to continue with same src_cpu.
10395 * We failed to reach balance because of affinity.
10398 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
10400 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0)
10401 *group_imbalance = 1;
10404 /* All tasks on this runqueue were pinned by CPU affinity */
10405 if (unlikely(env.flags & LBF_ALL_PINNED)) {
10406 __cpumask_clear_cpu(cpu_of(busiest), cpus);
10408 * Attempting to continue load balancing at the current
10409 * sched_domain level only makes sense if there are
10410 * active CPUs remaining as possible busiest CPUs to
10411 * pull load from which are not contained within the
10412 * destination group that is receiving any migrated
10415 if (!cpumask_subset(cpus, env.dst_grpmask)) {
10417 env.loop_break = SCHED_NR_MIGRATE_BREAK;
10420 goto out_all_pinned;
10425 schedstat_inc(sd->lb_failed[idle]);
10427 * Increment the failure counter only on periodic balance.
10428 * We do not want newidle balance, which can be very
10429 * frequent, pollute the failure counter causing
10430 * excessive cache_hot migrations and active balances.
10432 if (idle != CPU_NEWLY_IDLE)
10433 sd->nr_balance_failed++;
10435 if (need_active_balance(&env)) {
10436 unsigned long flags;
10438 raw_spin_rq_lock_irqsave(busiest, flags);
10441 * Don't kick the active_load_balance_cpu_stop,
10442 * if the curr task on busiest CPU can't be
10443 * moved to this_cpu:
10445 if (!cpumask_test_cpu(this_cpu, busiest->curr->cpus_ptr)) {
10446 raw_spin_rq_unlock_irqrestore(busiest, flags);
10447 goto out_one_pinned;
10450 /* Record that we found at least one task that could run on this_cpu */
10451 env.flags &= ~LBF_ALL_PINNED;
10454 * ->active_balance synchronizes accesses to
10455 * ->active_balance_work. Once set, it's cleared
10456 * only after active load balance is finished.
10458 if (!busiest->active_balance) {
10459 busiest->active_balance = 1;
10460 busiest->push_cpu = this_cpu;
10461 active_balance = 1;
10463 raw_spin_rq_unlock_irqrestore(busiest, flags);
10465 if (active_balance) {
10466 stop_one_cpu_nowait(cpu_of(busiest),
10467 active_load_balance_cpu_stop, busiest,
10468 &busiest->active_balance_work);
10472 sd->nr_balance_failed = 0;
10475 if (likely(!active_balance) || need_active_balance(&env)) {
10476 /* We were unbalanced, so reset the balancing interval */
10477 sd->balance_interval = sd->min_interval;
10484 * We reach balance although we may have faced some affinity
10485 * constraints. Clear the imbalance flag only if other tasks got
10486 * a chance to move and fix the imbalance.
10488 if (sd_parent && !(env.flags & LBF_ALL_PINNED)) {
10489 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
10491 if (*group_imbalance)
10492 *group_imbalance = 0;
10497 * We reach balance because all tasks are pinned at this level so
10498 * we can't migrate them. Let the imbalance flag set so parent level
10499 * can try to migrate them.
10501 schedstat_inc(sd->lb_balanced[idle]);
10503 sd->nr_balance_failed = 0;
10509 * newidle_balance() disregards balance intervals, so we could
10510 * repeatedly reach this code, which would lead to balance_interval
10511 * skyrocketing in a short amount of time. Skip the balance_interval
10512 * increase logic to avoid that.
10514 if (env.idle == CPU_NEWLY_IDLE)
10517 /* tune up the balancing interval */
10518 if ((env.flags & LBF_ALL_PINNED &&
10519 sd->balance_interval < MAX_PINNED_INTERVAL) ||
10520 sd->balance_interval < sd->max_interval)
10521 sd->balance_interval *= 2;
10526 static inline unsigned long
10527 get_sd_balance_interval(struct sched_domain *sd, int cpu_busy)
10529 unsigned long interval = sd->balance_interval;
10532 interval *= sd->busy_factor;
10534 /* scale ms to jiffies */
10535 interval = msecs_to_jiffies(interval);
10538 * Reduce likelihood of busy balancing at higher domains racing with
10539 * balancing at lower domains by preventing their balancing periods
10540 * from being multiples of each other.
10545 interval = clamp(interval, 1UL, max_load_balance_interval);
10551 update_next_balance(struct sched_domain *sd, unsigned long *next_balance)
10553 unsigned long interval, next;
10555 /* used by idle balance, so cpu_busy = 0 */
10556 interval = get_sd_balance_interval(sd, 0);
10557 next = sd->last_balance + interval;
10559 if (time_after(*next_balance, next))
10560 *next_balance = next;
10564 * active_load_balance_cpu_stop is run by the CPU stopper. It pushes
10565 * running tasks off the busiest CPU onto idle CPUs. It requires at
10566 * least 1 task to be running on each physical CPU where possible, and
10567 * avoids physical / logical imbalances.
10569 static int active_load_balance_cpu_stop(void *data)
10571 struct rq *busiest_rq = data;
10572 int busiest_cpu = cpu_of(busiest_rq);
10573 int target_cpu = busiest_rq->push_cpu;
10574 struct rq *target_rq = cpu_rq(target_cpu);
10575 struct sched_domain *sd;
10576 struct task_struct *p = NULL;
10577 struct rq_flags rf;
10579 rq_lock_irq(busiest_rq, &rf);
10581 * Between queueing the stop-work and running it is a hole in which
10582 * CPUs can become inactive. We should not move tasks from or to
10585 if (!cpu_active(busiest_cpu) || !cpu_active(target_cpu))
10588 /* Make sure the requested CPU hasn't gone down in the meantime: */
10589 if (unlikely(busiest_cpu != smp_processor_id() ||
10590 !busiest_rq->active_balance))
10593 /* Is there any task to move? */
10594 if (busiest_rq->nr_running <= 1)
10598 * This condition is "impossible", if it occurs
10599 * we need to fix it. Originally reported by
10600 * Bjorn Helgaas on a 128-CPU setup.
10602 WARN_ON_ONCE(busiest_rq == target_rq);
10604 /* Search for an sd spanning us and the target CPU. */
10606 for_each_domain(target_cpu, sd) {
10607 if (cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
10612 struct lb_env env = {
10614 .dst_cpu = target_cpu,
10615 .dst_rq = target_rq,
10616 .src_cpu = busiest_rq->cpu,
10617 .src_rq = busiest_rq,
10619 .flags = LBF_ACTIVE_LB,
10622 schedstat_inc(sd->alb_count);
10623 update_rq_clock(busiest_rq);
10625 p = detach_one_task(&env);
10627 schedstat_inc(sd->alb_pushed);
10628 /* Active balancing done, reset the failure counter. */
10629 sd->nr_balance_failed = 0;
10631 schedstat_inc(sd->alb_failed);
10636 busiest_rq->active_balance = 0;
10637 rq_unlock(busiest_rq, &rf);
10640 attach_one_task(target_rq, p);
10642 local_irq_enable();
10647 static DEFINE_SPINLOCK(balancing);
10650 * Scale the max load_balance interval with the number of CPUs in the system.
10651 * This trades load-balance latency on larger machines for less cross talk.
10653 void update_max_interval(void)
10655 max_load_balance_interval = HZ*num_online_cpus()/10;
10658 static inline bool update_newidle_cost(struct sched_domain *sd, u64 cost)
10660 if (cost > sd->max_newidle_lb_cost) {
10662 * Track max cost of a domain to make sure to not delay the
10663 * next wakeup on the CPU.
10665 sd->max_newidle_lb_cost = cost;
10666 sd->last_decay_max_lb_cost = jiffies;
10667 } else if (time_after(jiffies, sd->last_decay_max_lb_cost + HZ)) {
10669 * Decay the newidle max times by ~1% per second to ensure that
10670 * it is not outdated and the current max cost is actually
10673 sd->max_newidle_lb_cost = (sd->max_newidle_lb_cost * 253) / 256;
10674 sd->last_decay_max_lb_cost = jiffies;
10683 * It checks each scheduling domain to see if it is due to be balanced,
10684 * and initiates a balancing operation if so.
10686 * Balancing parameters are set up in init_sched_domains.
10688 static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle)
10690 int continue_balancing = 1;
10692 int busy = idle != CPU_IDLE && !sched_idle_cpu(cpu);
10693 unsigned long interval;
10694 struct sched_domain *sd;
10695 /* Earliest time when we have to do rebalance again */
10696 unsigned long next_balance = jiffies + 60*HZ;
10697 int update_next_balance = 0;
10698 int need_serialize, need_decay = 0;
10702 for_each_domain(cpu, sd) {
10704 * Decay the newidle max times here because this is a regular
10705 * visit to all the domains.
10707 need_decay = update_newidle_cost(sd, 0);
10708 max_cost += sd->max_newidle_lb_cost;
10711 * Stop the load balance at this level. There is another
10712 * CPU in our sched group which is doing load balancing more
10715 if (!continue_balancing) {
10721 interval = get_sd_balance_interval(sd, busy);
10723 need_serialize = sd->flags & SD_SERIALIZE;
10724 if (need_serialize) {
10725 if (!spin_trylock(&balancing))
10729 if (time_after_eq(jiffies, sd->last_balance + interval)) {
10730 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
10732 * The LBF_DST_PINNED logic could have changed
10733 * env->dst_cpu, so we can't know our idle
10734 * state even if we migrated tasks. Update it.
10736 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
10737 busy = idle != CPU_IDLE && !sched_idle_cpu(cpu);
10739 sd->last_balance = jiffies;
10740 interval = get_sd_balance_interval(sd, busy);
10742 if (need_serialize)
10743 spin_unlock(&balancing);
10745 if (time_after(next_balance, sd->last_balance + interval)) {
10746 next_balance = sd->last_balance + interval;
10747 update_next_balance = 1;
10752 * Ensure the rq-wide value also decays but keep it at a
10753 * reasonable floor to avoid funnies with rq->avg_idle.
10755 rq->max_idle_balance_cost =
10756 max((u64)sysctl_sched_migration_cost, max_cost);
10761 * next_balance will be updated only when there is a need.
10762 * When the cpu is attached to null domain for ex, it will not be
10765 if (likely(update_next_balance))
10766 rq->next_balance = next_balance;
10770 static inline int on_null_domain(struct rq *rq)
10772 return unlikely(!rcu_dereference_sched(rq->sd));
10775 #ifdef CONFIG_NO_HZ_COMMON
10777 * idle load balancing details
10778 * - When one of the busy CPUs notice that there may be an idle rebalancing
10779 * needed, they will kick the idle load balancer, which then does idle
10780 * load balancing for all the idle CPUs.
10781 * - HK_TYPE_MISC CPUs are used for this task, because HK_TYPE_SCHED not set
10785 static inline int find_new_ilb(void)
10788 const struct cpumask *hk_mask;
10790 hk_mask = housekeeping_cpumask(HK_TYPE_MISC);
10792 for_each_cpu_and(ilb, nohz.idle_cpus_mask, hk_mask) {
10794 if (ilb == smp_processor_id())
10805 * Kick a CPU to do the nohz balancing, if it is time for it. We pick any
10806 * idle CPU in the HK_TYPE_MISC housekeeping set (if there is one).
10808 static void kick_ilb(unsigned int flags)
10813 * Increase nohz.next_balance only when if full ilb is triggered but
10814 * not if we only update stats.
10816 if (flags & NOHZ_BALANCE_KICK)
10817 nohz.next_balance = jiffies+1;
10819 ilb_cpu = find_new_ilb();
10821 if (ilb_cpu >= nr_cpu_ids)
10825 * Access to rq::nohz_csd is serialized by NOHZ_KICK_MASK; he who sets
10826 * the first flag owns it; cleared by nohz_csd_func().
10828 flags = atomic_fetch_or(flags, nohz_flags(ilb_cpu));
10829 if (flags & NOHZ_KICK_MASK)
10833 * This way we generate an IPI on the target CPU which
10834 * is idle. And the softirq performing nohz idle load balance
10835 * will be run before returning from the IPI.
10837 smp_call_function_single_async(ilb_cpu, &cpu_rq(ilb_cpu)->nohz_csd);
10841 * Current decision point for kicking the idle load balancer in the presence
10842 * of idle CPUs in the system.
10844 static void nohz_balancer_kick(struct rq *rq)
10846 unsigned long now = jiffies;
10847 struct sched_domain_shared *sds;
10848 struct sched_domain *sd;
10849 int nr_busy, i, cpu = rq->cpu;
10850 unsigned int flags = 0;
10852 if (unlikely(rq->idle_balance))
10856 * We may be recently in ticked or tickless idle mode. At the first
10857 * busy tick after returning from idle, we will update the busy stats.
10859 nohz_balance_exit_idle(rq);
10862 * None are in tickless mode and hence no need for NOHZ idle load
10865 if (likely(!atomic_read(&nohz.nr_cpus)))
10868 if (READ_ONCE(nohz.has_blocked) &&
10869 time_after(now, READ_ONCE(nohz.next_blocked)))
10870 flags = NOHZ_STATS_KICK;
10872 if (time_before(now, nohz.next_balance))
10875 if (rq->nr_running >= 2) {
10876 flags = NOHZ_STATS_KICK | NOHZ_BALANCE_KICK;
10882 sd = rcu_dereference(rq->sd);
10885 * If there's a CFS task and the current CPU has reduced
10886 * capacity; kick the ILB to see if there's a better CPU to run
10889 if (rq->cfs.h_nr_running >= 1 && check_cpu_capacity(rq, sd)) {
10890 flags = NOHZ_STATS_KICK | NOHZ_BALANCE_KICK;
10895 sd = rcu_dereference(per_cpu(sd_asym_packing, cpu));
10898 * When ASYM_PACKING; see if there's a more preferred CPU
10899 * currently idle; in which case, kick the ILB to move tasks
10902 for_each_cpu_and(i, sched_domain_span(sd), nohz.idle_cpus_mask) {
10903 if (sched_asym_prefer(i, cpu)) {
10904 flags = NOHZ_STATS_KICK | NOHZ_BALANCE_KICK;
10910 sd = rcu_dereference(per_cpu(sd_asym_cpucapacity, cpu));
10913 * When ASYM_CPUCAPACITY; see if there's a higher capacity CPU
10914 * to run the misfit task on.
10916 if (check_misfit_status(rq, sd)) {
10917 flags = NOHZ_STATS_KICK | NOHZ_BALANCE_KICK;
10922 * For asymmetric systems, we do not want to nicely balance
10923 * cache use, instead we want to embrace asymmetry and only
10924 * ensure tasks have enough CPU capacity.
10926 * Skip the LLC logic because it's not relevant in that case.
10931 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
10934 * If there is an imbalance between LLC domains (IOW we could
10935 * increase the overall cache use), we need some less-loaded LLC
10936 * domain to pull some load. Likewise, we may need to spread
10937 * load within the current LLC domain (e.g. packed SMT cores but
10938 * other CPUs are idle). We can't really know from here how busy
10939 * the others are - so just get a nohz balance going if it looks
10940 * like this LLC domain has tasks we could move.
10942 nr_busy = atomic_read(&sds->nr_busy_cpus);
10944 flags = NOHZ_STATS_KICK | NOHZ_BALANCE_KICK;
10951 if (READ_ONCE(nohz.needs_update))
10952 flags |= NOHZ_NEXT_KICK;
10958 static void set_cpu_sd_state_busy(int cpu)
10960 struct sched_domain *sd;
10963 sd = rcu_dereference(per_cpu(sd_llc, cpu));
10965 if (!sd || !sd->nohz_idle)
10969 atomic_inc(&sd->shared->nr_busy_cpus);
10974 void nohz_balance_exit_idle(struct rq *rq)
10976 SCHED_WARN_ON(rq != this_rq());
10978 if (likely(!rq->nohz_tick_stopped))
10981 rq->nohz_tick_stopped = 0;
10982 cpumask_clear_cpu(rq->cpu, nohz.idle_cpus_mask);
10983 atomic_dec(&nohz.nr_cpus);
10985 set_cpu_sd_state_busy(rq->cpu);
10988 static void set_cpu_sd_state_idle(int cpu)
10990 struct sched_domain *sd;
10993 sd = rcu_dereference(per_cpu(sd_llc, cpu));
10995 if (!sd || sd->nohz_idle)
10999 atomic_dec(&sd->shared->nr_busy_cpus);
11005 * This routine will record that the CPU is going idle with tick stopped.
11006 * This info will be used in performing idle load balancing in the future.
11008 void nohz_balance_enter_idle(int cpu)
11010 struct rq *rq = cpu_rq(cpu);
11012 SCHED_WARN_ON(cpu != smp_processor_id());
11014 /* If this CPU is going down, then nothing needs to be done: */
11015 if (!cpu_active(cpu))
11018 /* Spare idle load balancing on CPUs that don't want to be disturbed: */
11019 if (!housekeeping_cpu(cpu, HK_TYPE_SCHED))
11023 * Can be set safely without rq->lock held
11024 * If a clear happens, it will have evaluated last additions because
11025 * rq->lock is held during the check and the clear
11027 rq->has_blocked_load = 1;
11030 * The tick is still stopped but load could have been added in the
11031 * meantime. We set the nohz.has_blocked flag to trig a check of the
11032 * *_avg. The CPU is already part of nohz.idle_cpus_mask so the clear
11033 * of nohz.has_blocked can only happen after checking the new load
11035 if (rq->nohz_tick_stopped)
11038 /* If we're a completely isolated CPU, we don't play: */
11039 if (on_null_domain(rq))
11042 rq->nohz_tick_stopped = 1;
11044 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
11045 atomic_inc(&nohz.nr_cpus);
11048 * Ensures that if nohz_idle_balance() fails to observe our
11049 * @idle_cpus_mask store, it must observe the @has_blocked
11050 * and @needs_update stores.
11052 smp_mb__after_atomic();
11054 set_cpu_sd_state_idle(cpu);
11056 WRITE_ONCE(nohz.needs_update, 1);
11059 * Each time a cpu enter idle, we assume that it has blocked load and
11060 * enable the periodic update of the load of idle cpus
11062 WRITE_ONCE(nohz.has_blocked, 1);
11065 static bool update_nohz_stats(struct rq *rq)
11067 unsigned int cpu = rq->cpu;
11069 if (!rq->has_blocked_load)
11072 if (!cpumask_test_cpu(cpu, nohz.idle_cpus_mask))
11075 if (!time_after(jiffies, READ_ONCE(rq->last_blocked_load_update_tick)))
11078 update_blocked_averages(cpu);
11080 return rq->has_blocked_load;
11084 * Internal function that runs load balance for all idle cpus. The load balance
11085 * can be a simple update of blocked load or a complete load balance with
11086 * tasks movement depending of flags.
11088 static void _nohz_idle_balance(struct rq *this_rq, unsigned int flags)
11090 /* Earliest time when we have to do rebalance again */
11091 unsigned long now = jiffies;
11092 unsigned long next_balance = now + 60*HZ;
11093 bool has_blocked_load = false;
11094 int update_next_balance = 0;
11095 int this_cpu = this_rq->cpu;
11099 SCHED_WARN_ON((flags & NOHZ_KICK_MASK) == NOHZ_BALANCE_KICK);
11102 * We assume there will be no idle load after this update and clear
11103 * the has_blocked flag. If a cpu enters idle in the mean time, it will
11104 * set the has_blocked flag and trigger another update of idle load.
11105 * Because a cpu that becomes idle, is added to idle_cpus_mask before
11106 * setting the flag, we are sure to not clear the state and not
11107 * check the load of an idle cpu.
11109 * Same applies to idle_cpus_mask vs needs_update.
11111 if (flags & NOHZ_STATS_KICK)
11112 WRITE_ONCE(nohz.has_blocked, 0);
11113 if (flags & NOHZ_NEXT_KICK)
11114 WRITE_ONCE(nohz.needs_update, 0);
11117 * Ensures that if we miss the CPU, we must see the has_blocked
11118 * store from nohz_balance_enter_idle().
11123 * Start with the next CPU after this_cpu so we will end with this_cpu and let a
11124 * chance for other idle cpu to pull load.
11126 for_each_cpu_wrap(balance_cpu, nohz.idle_cpus_mask, this_cpu+1) {
11127 if (!idle_cpu(balance_cpu))
11131 * If this CPU gets work to do, stop the load balancing
11132 * work being done for other CPUs. Next load
11133 * balancing owner will pick it up.
11135 if (need_resched()) {
11136 if (flags & NOHZ_STATS_KICK)
11137 has_blocked_load = true;
11138 if (flags & NOHZ_NEXT_KICK)
11139 WRITE_ONCE(nohz.needs_update, 1);
11143 rq = cpu_rq(balance_cpu);
11145 if (flags & NOHZ_STATS_KICK)
11146 has_blocked_load |= update_nohz_stats(rq);
11149 * If time for next balance is due,
11152 if (time_after_eq(jiffies, rq->next_balance)) {
11153 struct rq_flags rf;
11155 rq_lock_irqsave(rq, &rf);
11156 update_rq_clock(rq);
11157 rq_unlock_irqrestore(rq, &rf);
11159 if (flags & NOHZ_BALANCE_KICK)
11160 rebalance_domains(rq, CPU_IDLE);
11163 if (time_after(next_balance, rq->next_balance)) {
11164 next_balance = rq->next_balance;
11165 update_next_balance = 1;
11170 * next_balance will be updated only when there is a need.
11171 * When the CPU is attached to null domain for ex, it will not be
11174 if (likely(update_next_balance))
11175 nohz.next_balance = next_balance;
11177 if (flags & NOHZ_STATS_KICK)
11178 WRITE_ONCE(nohz.next_blocked,
11179 now + msecs_to_jiffies(LOAD_AVG_PERIOD));
11182 /* There is still blocked load, enable periodic update */
11183 if (has_blocked_load)
11184 WRITE_ONCE(nohz.has_blocked, 1);
11188 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
11189 * rebalancing for all the cpus for whom scheduler ticks are stopped.
11191 static bool nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
11193 unsigned int flags = this_rq->nohz_idle_balance;
11198 this_rq->nohz_idle_balance = 0;
11200 if (idle != CPU_IDLE)
11203 _nohz_idle_balance(this_rq, flags);
11209 * Check if we need to run the ILB for updating blocked load before entering
11212 void nohz_run_idle_balance(int cpu)
11214 unsigned int flags;
11216 flags = atomic_fetch_andnot(NOHZ_NEWILB_KICK, nohz_flags(cpu));
11219 * Update the blocked load only if no SCHED_SOFTIRQ is about to happen
11220 * (ie NOHZ_STATS_KICK set) and will do the same.
11222 if ((flags == NOHZ_NEWILB_KICK) && !need_resched())
11223 _nohz_idle_balance(cpu_rq(cpu), NOHZ_STATS_KICK);
11226 static void nohz_newidle_balance(struct rq *this_rq)
11228 int this_cpu = this_rq->cpu;
11231 * This CPU doesn't want to be disturbed by scheduler
11234 if (!housekeeping_cpu(this_cpu, HK_TYPE_SCHED))
11237 /* Will wake up very soon. No time for doing anything else*/
11238 if (this_rq->avg_idle < sysctl_sched_migration_cost)
11241 /* Don't need to update blocked load of idle CPUs*/
11242 if (!READ_ONCE(nohz.has_blocked) ||
11243 time_before(jiffies, READ_ONCE(nohz.next_blocked)))
11247 * Set the need to trigger ILB in order to update blocked load
11248 * before entering idle state.
11250 atomic_or(NOHZ_NEWILB_KICK, nohz_flags(this_cpu));
11253 #else /* !CONFIG_NO_HZ_COMMON */
11254 static inline void nohz_balancer_kick(struct rq *rq) { }
11256 static inline bool nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
11261 static inline void nohz_newidle_balance(struct rq *this_rq) { }
11262 #endif /* CONFIG_NO_HZ_COMMON */
11265 * newidle_balance is called by schedule() if this_cpu is about to become
11266 * idle. Attempts to pull tasks from other CPUs.
11269 * < 0 - we released the lock and there are !fair tasks present
11270 * 0 - failed, no new tasks
11271 * > 0 - success, new (fair) tasks present
11273 static int newidle_balance(struct rq *this_rq, struct rq_flags *rf)
11275 unsigned long next_balance = jiffies + HZ;
11276 int this_cpu = this_rq->cpu;
11277 u64 t0, t1, curr_cost = 0;
11278 struct sched_domain *sd;
11279 int pulled_task = 0;
11281 update_misfit_status(NULL, this_rq);
11284 * There is a task waiting to run. No need to search for one.
11285 * Return 0; the task will be enqueued when switching to idle.
11287 if (this_rq->ttwu_pending)
11291 * We must set idle_stamp _before_ calling idle_balance(), such that we
11292 * measure the duration of idle_balance() as idle time.
11294 this_rq->idle_stamp = rq_clock(this_rq);
11297 * Do not pull tasks towards !active CPUs...
11299 if (!cpu_active(this_cpu))
11303 * This is OK, because current is on_cpu, which avoids it being picked
11304 * for load-balance and preemption/IRQs are still disabled avoiding
11305 * further scheduler activity on it and we're being very careful to
11306 * re-start the picking loop.
11308 rq_unpin_lock(this_rq, rf);
11311 sd = rcu_dereference_check_sched_domain(this_rq->sd);
11313 if (!READ_ONCE(this_rq->rd->overload) ||
11314 (sd && this_rq->avg_idle < sd->max_newidle_lb_cost)) {
11317 update_next_balance(sd, &next_balance);
11324 raw_spin_rq_unlock(this_rq);
11326 t0 = sched_clock_cpu(this_cpu);
11327 update_blocked_averages(this_cpu);
11330 for_each_domain(this_cpu, sd) {
11331 int continue_balancing = 1;
11334 update_next_balance(sd, &next_balance);
11336 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost)
11339 if (sd->flags & SD_BALANCE_NEWIDLE) {
11341 pulled_task = load_balance(this_cpu, this_rq,
11342 sd, CPU_NEWLY_IDLE,
11343 &continue_balancing);
11345 t1 = sched_clock_cpu(this_cpu);
11346 domain_cost = t1 - t0;
11347 update_newidle_cost(sd, domain_cost);
11349 curr_cost += domain_cost;
11354 * Stop searching for tasks to pull if there are
11355 * now runnable tasks on this rq.
11357 if (pulled_task || this_rq->nr_running > 0 ||
11358 this_rq->ttwu_pending)
11363 raw_spin_rq_lock(this_rq);
11365 if (curr_cost > this_rq->max_idle_balance_cost)
11366 this_rq->max_idle_balance_cost = curr_cost;
11369 * While browsing the domains, we released the rq lock, a task could
11370 * have been enqueued in the meantime. Since we're not going idle,
11371 * pretend we pulled a task.
11373 if (this_rq->cfs.h_nr_running && !pulled_task)
11376 /* Is there a task of a high priority class? */
11377 if (this_rq->nr_running != this_rq->cfs.h_nr_running)
11381 /* Move the next balance forward */
11382 if (time_after(this_rq->next_balance, next_balance))
11383 this_rq->next_balance = next_balance;
11386 this_rq->idle_stamp = 0;
11388 nohz_newidle_balance(this_rq);
11390 rq_repin_lock(this_rq, rf);
11392 return pulled_task;
11396 * run_rebalance_domains is triggered when needed from the scheduler tick.
11397 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
11399 static __latent_entropy void run_rebalance_domains(struct softirq_action *h)
11401 struct rq *this_rq = this_rq();
11402 enum cpu_idle_type idle = this_rq->idle_balance ?
11403 CPU_IDLE : CPU_NOT_IDLE;
11406 * If this CPU has a pending nohz_balance_kick, then do the
11407 * balancing on behalf of the other idle CPUs whose ticks are
11408 * stopped. Do nohz_idle_balance *before* rebalance_domains to
11409 * give the idle CPUs a chance to load balance. Else we may
11410 * load balance only within the local sched_domain hierarchy
11411 * and abort nohz_idle_balance altogether if we pull some load.
11413 if (nohz_idle_balance(this_rq, idle))
11416 /* normal load balance */
11417 update_blocked_averages(this_rq->cpu);
11418 rebalance_domains(this_rq, idle);
11422 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
11424 void trigger_load_balance(struct rq *rq)
11427 * Don't need to rebalance while attached to NULL domain or
11428 * runqueue CPU is not active
11430 if (unlikely(on_null_domain(rq) || !cpu_active(cpu_of(rq))))
11433 if (time_after_eq(jiffies, rq->next_balance))
11434 raise_softirq(SCHED_SOFTIRQ);
11436 nohz_balancer_kick(rq);
11439 static void rq_online_fair(struct rq *rq)
11443 update_runtime_enabled(rq);
11446 static void rq_offline_fair(struct rq *rq)
11450 /* Ensure any throttled groups are reachable by pick_next_task */
11451 unthrottle_offline_cfs_rqs(rq);
11454 #endif /* CONFIG_SMP */
11456 #ifdef CONFIG_SCHED_CORE
11458 __entity_slice_used(struct sched_entity *se, int min_nr_tasks)
11460 u64 slice = sched_slice(cfs_rq_of(se), se);
11461 u64 rtime = se->sum_exec_runtime - se->prev_sum_exec_runtime;
11463 return (rtime * min_nr_tasks > slice);
11466 #define MIN_NR_TASKS_DURING_FORCEIDLE 2
11467 static inline void task_tick_core(struct rq *rq, struct task_struct *curr)
11469 if (!sched_core_enabled(rq))
11473 * If runqueue has only one task which used up its slice and
11474 * if the sibling is forced idle, then trigger schedule to
11475 * give forced idle task a chance.
11477 * sched_slice() considers only this active rq and it gets the
11478 * whole slice. But during force idle, we have siblings acting
11479 * like a single runqueue and hence we need to consider runnable
11480 * tasks on this CPU and the forced idle CPU. Ideally, we should
11481 * go through the forced idle rq, but that would be a perf hit.
11482 * We can assume that the forced idle CPU has at least
11483 * MIN_NR_TASKS_DURING_FORCEIDLE - 1 tasks and use that to check
11484 * if we need to give up the CPU.
11486 if (rq->core->core_forceidle_count && rq->cfs.nr_running == 1 &&
11487 __entity_slice_used(&curr->se, MIN_NR_TASKS_DURING_FORCEIDLE))
11492 * se_fi_update - Update the cfs_rq->min_vruntime_fi in a CFS hierarchy if needed.
11494 static void se_fi_update(struct sched_entity *se, unsigned int fi_seq, bool forceidle)
11496 for_each_sched_entity(se) {
11497 struct cfs_rq *cfs_rq = cfs_rq_of(se);
11500 if (cfs_rq->forceidle_seq == fi_seq)
11502 cfs_rq->forceidle_seq = fi_seq;
11505 cfs_rq->min_vruntime_fi = cfs_rq->min_vruntime;
11509 void task_vruntime_update(struct rq *rq, struct task_struct *p, bool in_fi)
11511 struct sched_entity *se = &p->se;
11513 if (p->sched_class != &fair_sched_class)
11516 se_fi_update(se, rq->core->core_forceidle_seq, in_fi);
11519 bool cfs_prio_less(struct task_struct *a, struct task_struct *b, bool in_fi)
11521 struct rq *rq = task_rq(a);
11522 struct sched_entity *sea = &a->se;
11523 struct sched_entity *seb = &b->se;
11524 struct cfs_rq *cfs_rqa;
11525 struct cfs_rq *cfs_rqb;
11528 SCHED_WARN_ON(task_rq(b)->core != rq->core);
11530 #ifdef CONFIG_FAIR_GROUP_SCHED
11532 * Find an se in the hierarchy for tasks a and b, such that the se's
11533 * are immediate siblings.
11535 while (sea->cfs_rq->tg != seb->cfs_rq->tg) {
11536 int sea_depth = sea->depth;
11537 int seb_depth = seb->depth;
11539 if (sea_depth >= seb_depth)
11540 sea = parent_entity(sea);
11541 if (sea_depth <= seb_depth)
11542 seb = parent_entity(seb);
11545 se_fi_update(sea, rq->core->core_forceidle_seq, in_fi);
11546 se_fi_update(seb, rq->core->core_forceidle_seq, in_fi);
11548 cfs_rqa = sea->cfs_rq;
11549 cfs_rqb = seb->cfs_rq;
11551 cfs_rqa = &task_rq(a)->cfs;
11552 cfs_rqb = &task_rq(b)->cfs;
11556 * Find delta after normalizing se's vruntime with its cfs_rq's
11557 * min_vruntime_fi, which would have been updated in prior calls
11558 * to se_fi_update().
11560 delta = (s64)(sea->vruntime - seb->vruntime) +
11561 (s64)(cfs_rqb->min_vruntime_fi - cfs_rqa->min_vruntime_fi);
11566 static inline void task_tick_core(struct rq *rq, struct task_struct *curr) {}
11570 * scheduler tick hitting a task of our scheduling class.
11572 * NOTE: This function can be called remotely by the tick offload that
11573 * goes along full dynticks. Therefore no local assumption can be made
11574 * and everything must be accessed through the @rq and @curr passed in
11577 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
11579 struct cfs_rq *cfs_rq;
11580 struct sched_entity *se = &curr->se;
11582 for_each_sched_entity(se) {
11583 cfs_rq = cfs_rq_of(se);
11584 entity_tick(cfs_rq, se, queued);
11587 if (static_branch_unlikely(&sched_numa_balancing))
11588 task_tick_numa(rq, curr);
11590 update_misfit_status(curr, rq);
11591 update_overutilized_status(task_rq(curr));
11593 task_tick_core(rq, curr);
11597 * called on fork with the child task as argument from the parent's context
11598 * - child not yet on the tasklist
11599 * - preemption disabled
11601 static void task_fork_fair(struct task_struct *p)
11603 struct cfs_rq *cfs_rq;
11604 struct sched_entity *se = &p->se, *curr;
11605 struct rq *rq = this_rq();
11606 struct rq_flags rf;
11609 update_rq_clock(rq);
11611 cfs_rq = task_cfs_rq(current);
11612 curr = cfs_rq->curr;
11614 update_curr(cfs_rq);
11615 se->vruntime = curr->vruntime;
11617 place_entity(cfs_rq, se, 1);
11619 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
11621 * Upon rescheduling, sched_class::put_prev_task() will place
11622 * 'current' within the tree based on its new key value.
11624 swap(curr->vruntime, se->vruntime);
11628 se->vruntime -= cfs_rq->min_vruntime;
11629 rq_unlock(rq, &rf);
11633 * Priority of the task has changed. Check to see if we preempt
11634 * the current task.
11637 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
11639 if (!task_on_rq_queued(p))
11642 if (rq->cfs.nr_running == 1)
11646 * Reschedule if we are currently running on this runqueue and
11647 * our priority decreased, or if we are not currently running on
11648 * this runqueue and our priority is higher than the current's
11650 if (task_current(rq, p)) {
11651 if (p->prio > oldprio)
11654 check_preempt_curr(rq, p, 0);
11657 static inline bool vruntime_normalized(struct task_struct *p)
11659 struct sched_entity *se = &p->se;
11662 * In both the TASK_ON_RQ_QUEUED and TASK_ON_RQ_MIGRATING cases,
11663 * the dequeue_entity(.flags=0) will already have normalized the
11670 * When !on_rq, vruntime of the task has usually NOT been normalized.
11671 * But there are some cases where it has already been normalized:
11673 * - A forked child which is waiting for being woken up by
11674 * wake_up_new_task().
11675 * - A task which has been woken up by try_to_wake_up() and
11676 * waiting for actually being woken up by sched_ttwu_pending().
11678 if (!se->sum_exec_runtime ||
11679 (READ_ONCE(p->__state) == TASK_WAKING && p->sched_remote_wakeup))
11685 #ifdef CONFIG_FAIR_GROUP_SCHED
11687 * Propagate the changes of the sched_entity across the tg tree to make it
11688 * visible to the root
11690 static void propagate_entity_cfs_rq(struct sched_entity *se)
11692 struct cfs_rq *cfs_rq = cfs_rq_of(se);
11694 if (cfs_rq_throttled(cfs_rq))
11697 if (!throttled_hierarchy(cfs_rq))
11698 list_add_leaf_cfs_rq(cfs_rq);
11700 /* Start to propagate at parent */
11703 for_each_sched_entity(se) {
11704 cfs_rq = cfs_rq_of(se);
11706 update_load_avg(cfs_rq, se, UPDATE_TG);
11708 if (cfs_rq_throttled(cfs_rq))
11711 if (!throttled_hierarchy(cfs_rq))
11712 list_add_leaf_cfs_rq(cfs_rq);
11716 static void propagate_entity_cfs_rq(struct sched_entity *se) { }
11719 static void detach_entity_cfs_rq(struct sched_entity *se)
11721 struct cfs_rq *cfs_rq = cfs_rq_of(se);
11725 * In case the task sched_avg hasn't been attached:
11726 * - A forked task which hasn't been woken up by wake_up_new_task().
11727 * - A task which has been woken up by try_to_wake_up() but is
11728 * waiting for actually being woken up by sched_ttwu_pending().
11730 if (!se->avg.last_update_time)
11734 /* Catch up with the cfs_rq and remove our load when we leave */
11735 update_load_avg(cfs_rq, se, 0);
11736 detach_entity_load_avg(cfs_rq, se);
11737 update_tg_load_avg(cfs_rq);
11738 propagate_entity_cfs_rq(se);
11741 static void attach_entity_cfs_rq(struct sched_entity *se)
11743 struct cfs_rq *cfs_rq = cfs_rq_of(se);
11745 /* Synchronize entity with its cfs_rq */
11746 update_load_avg(cfs_rq, se, sched_feat(ATTACH_AGE_LOAD) ? 0 : SKIP_AGE_LOAD);
11747 attach_entity_load_avg(cfs_rq, se);
11748 update_tg_load_avg(cfs_rq);
11749 propagate_entity_cfs_rq(se);
11752 static void detach_task_cfs_rq(struct task_struct *p)
11754 struct sched_entity *se = &p->se;
11755 struct cfs_rq *cfs_rq = cfs_rq_of(se);
11757 if (!vruntime_normalized(p)) {
11759 * Fix up our vruntime so that the current sleep doesn't
11760 * cause 'unlimited' sleep bonus.
11762 place_entity(cfs_rq, se, 0);
11763 se->vruntime -= cfs_rq->min_vruntime;
11766 detach_entity_cfs_rq(se);
11769 static void attach_task_cfs_rq(struct task_struct *p)
11771 struct sched_entity *se = &p->se;
11772 struct cfs_rq *cfs_rq = cfs_rq_of(se);
11774 attach_entity_cfs_rq(se);
11776 if (!vruntime_normalized(p))
11777 se->vruntime += cfs_rq->min_vruntime;
11780 static void switched_from_fair(struct rq *rq, struct task_struct *p)
11782 detach_task_cfs_rq(p);
11785 static void switched_to_fair(struct rq *rq, struct task_struct *p)
11787 attach_task_cfs_rq(p);
11789 if (task_on_rq_queued(p)) {
11791 * We were most likely switched from sched_rt, so
11792 * kick off the schedule if running, otherwise just see
11793 * if we can still preempt the current task.
11795 if (task_current(rq, p))
11798 check_preempt_curr(rq, p, 0);
11802 /* Account for a task changing its policy or group.
11804 * This routine is mostly called to set cfs_rq->curr field when a task
11805 * migrates between groups/classes.
11807 static void set_next_task_fair(struct rq *rq, struct task_struct *p, bool first)
11809 struct sched_entity *se = &p->se;
11812 if (task_on_rq_queued(p)) {
11814 * Move the next running task to the front of the list, so our
11815 * cfs_tasks list becomes MRU one.
11817 list_move(&se->group_node, &rq->cfs_tasks);
11821 for_each_sched_entity(se) {
11822 struct cfs_rq *cfs_rq = cfs_rq_of(se);
11824 set_next_entity(cfs_rq, se);
11825 /* ensure bandwidth has been allocated on our new cfs_rq */
11826 account_cfs_rq_runtime(cfs_rq, 0);
11830 void init_cfs_rq(struct cfs_rq *cfs_rq)
11832 cfs_rq->tasks_timeline = RB_ROOT_CACHED;
11833 u64_u32_store(cfs_rq->min_vruntime, (u64)(-(1LL << 20)));
11835 raw_spin_lock_init(&cfs_rq->removed.lock);
11839 #ifdef CONFIG_FAIR_GROUP_SCHED
11840 static void task_change_group_fair(struct task_struct *p)
11843 * We couldn't detach or attach a forked task which
11844 * hasn't been woken up by wake_up_new_task().
11846 if (READ_ONCE(p->__state) == TASK_NEW)
11849 detach_task_cfs_rq(p);
11852 /* Tell se's cfs_rq has been changed -- migrated */
11853 p->se.avg.last_update_time = 0;
11855 set_task_rq(p, task_cpu(p));
11856 attach_task_cfs_rq(p);
11859 void free_fair_sched_group(struct task_group *tg)
11863 for_each_possible_cpu(i) {
11865 kfree(tg->cfs_rq[i]);
11874 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
11876 struct sched_entity *se;
11877 struct cfs_rq *cfs_rq;
11880 tg->cfs_rq = kcalloc(nr_cpu_ids, sizeof(cfs_rq), GFP_KERNEL);
11883 tg->se = kcalloc(nr_cpu_ids, sizeof(se), GFP_KERNEL);
11887 tg->shares = NICE_0_LOAD;
11889 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
11891 for_each_possible_cpu(i) {
11892 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
11893 GFP_KERNEL, cpu_to_node(i));
11897 se = kzalloc_node(sizeof(struct sched_entity_stats),
11898 GFP_KERNEL, cpu_to_node(i));
11902 init_cfs_rq(cfs_rq);
11903 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
11904 init_entity_runnable_average(se);
11915 void online_fair_sched_group(struct task_group *tg)
11917 struct sched_entity *se;
11918 struct rq_flags rf;
11922 for_each_possible_cpu(i) {
11925 rq_lock_irq(rq, &rf);
11926 update_rq_clock(rq);
11927 attach_entity_cfs_rq(se);
11928 sync_throttle(tg, i);
11929 rq_unlock_irq(rq, &rf);
11933 void unregister_fair_sched_group(struct task_group *tg)
11935 unsigned long flags;
11939 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
11941 for_each_possible_cpu(cpu) {
11943 remove_entity_load_avg(tg->se[cpu]);
11946 * Only empty task groups can be destroyed; so we can speculatively
11947 * check on_list without danger of it being re-added.
11949 if (!tg->cfs_rq[cpu]->on_list)
11954 raw_spin_rq_lock_irqsave(rq, flags);
11955 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
11956 raw_spin_rq_unlock_irqrestore(rq, flags);
11960 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
11961 struct sched_entity *se, int cpu,
11962 struct sched_entity *parent)
11964 struct rq *rq = cpu_rq(cpu);
11968 init_cfs_rq_runtime(cfs_rq);
11970 tg->cfs_rq[cpu] = cfs_rq;
11973 /* se could be NULL for root_task_group */
11978 se->cfs_rq = &rq->cfs;
11981 se->cfs_rq = parent->my_q;
11982 se->depth = parent->depth + 1;
11986 /* guarantee group entities always have weight */
11987 update_load_set(&se->load, NICE_0_LOAD);
11988 se->parent = parent;
11991 static DEFINE_MUTEX(shares_mutex);
11993 static int __sched_group_set_shares(struct task_group *tg, unsigned long shares)
11997 lockdep_assert_held(&shares_mutex);
12000 * We can't change the weight of the root cgroup.
12005 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
12007 if (tg->shares == shares)
12010 tg->shares = shares;
12011 for_each_possible_cpu(i) {
12012 struct rq *rq = cpu_rq(i);
12013 struct sched_entity *se = tg->se[i];
12014 struct rq_flags rf;
12016 /* Propagate contribution to hierarchy */
12017 rq_lock_irqsave(rq, &rf);
12018 update_rq_clock(rq);
12019 for_each_sched_entity(se) {
12020 update_load_avg(cfs_rq_of(se), se, UPDATE_TG);
12021 update_cfs_group(se);
12023 rq_unlock_irqrestore(rq, &rf);
12029 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
12033 mutex_lock(&shares_mutex);
12034 if (tg_is_idle(tg))
12037 ret = __sched_group_set_shares(tg, shares);
12038 mutex_unlock(&shares_mutex);
12043 int sched_group_set_idle(struct task_group *tg, long idle)
12047 if (tg == &root_task_group)
12050 if (idle < 0 || idle > 1)
12053 mutex_lock(&shares_mutex);
12055 if (tg->idle == idle) {
12056 mutex_unlock(&shares_mutex);
12062 for_each_possible_cpu(i) {
12063 struct rq *rq = cpu_rq(i);
12064 struct sched_entity *se = tg->se[i];
12065 struct cfs_rq *parent_cfs_rq, *grp_cfs_rq = tg->cfs_rq[i];
12066 bool was_idle = cfs_rq_is_idle(grp_cfs_rq);
12067 long idle_task_delta;
12068 struct rq_flags rf;
12070 rq_lock_irqsave(rq, &rf);
12072 grp_cfs_rq->idle = idle;
12073 if (WARN_ON_ONCE(was_idle == cfs_rq_is_idle(grp_cfs_rq)))
12077 parent_cfs_rq = cfs_rq_of(se);
12078 if (cfs_rq_is_idle(grp_cfs_rq))
12079 parent_cfs_rq->idle_nr_running++;
12081 parent_cfs_rq->idle_nr_running--;
12084 idle_task_delta = grp_cfs_rq->h_nr_running -
12085 grp_cfs_rq->idle_h_nr_running;
12086 if (!cfs_rq_is_idle(grp_cfs_rq))
12087 idle_task_delta *= -1;
12089 for_each_sched_entity(se) {
12090 struct cfs_rq *cfs_rq = cfs_rq_of(se);
12095 cfs_rq->idle_h_nr_running += idle_task_delta;
12097 /* Already accounted at parent level and above. */
12098 if (cfs_rq_is_idle(cfs_rq))
12103 rq_unlock_irqrestore(rq, &rf);
12106 /* Idle groups have minimum weight. */
12107 if (tg_is_idle(tg))
12108 __sched_group_set_shares(tg, scale_load(WEIGHT_IDLEPRIO));
12110 __sched_group_set_shares(tg, NICE_0_LOAD);
12112 mutex_unlock(&shares_mutex);
12116 #else /* CONFIG_FAIR_GROUP_SCHED */
12118 void free_fair_sched_group(struct task_group *tg) { }
12120 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
12125 void online_fair_sched_group(struct task_group *tg) { }
12127 void unregister_fair_sched_group(struct task_group *tg) { }
12129 #endif /* CONFIG_FAIR_GROUP_SCHED */
12132 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
12134 struct sched_entity *se = &task->se;
12135 unsigned int rr_interval = 0;
12138 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
12141 if (rq->cfs.load.weight)
12142 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
12144 return rr_interval;
12148 * All the scheduling class methods:
12150 DEFINE_SCHED_CLASS(fair) = {
12152 .enqueue_task = enqueue_task_fair,
12153 .dequeue_task = dequeue_task_fair,
12154 .yield_task = yield_task_fair,
12155 .yield_to_task = yield_to_task_fair,
12157 .check_preempt_curr = check_preempt_wakeup,
12159 .pick_next_task = __pick_next_task_fair,
12160 .put_prev_task = put_prev_task_fair,
12161 .set_next_task = set_next_task_fair,
12164 .balance = balance_fair,
12165 .pick_task = pick_task_fair,
12166 .select_task_rq = select_task_rq_fair,
12167 .migrate_task_rq = migrate_task_rq_fair,
12169 .rq_online = rq_online_fair,
12170 .rq_offline = rq_offline_fair,
12172 .task_dead = task_dead_fair,
12173 .set_cpus_allowed = set_cpus_allowed_common,
12176 .task_tick = task_tick_fair,
12177 .task_fork = task_fork_fair,
12179 .prio_changed = prio_changed_fair,
12180 .switched_from = switched_from_fair,
12181 .switched_to = switched_to_fair,
12183 .get_rr_interval = get_rr_interval_fair,
12185 .update_curr = update_curr_fair,
12187 #ifdef CONFIG_FAIR_GROUP_SCHED
12188 .task_change_group = task_change_group_fair,
12191 #ifdef CONFIG_UCLAMP_TASK
12192 .uclamp_enabled = 1,
12196 #ifdef CONFIG_SCHED_DEBUG
12197 void print_cfs_stats(struct seq_file *m, int cpu)
12199 struct cfs_rq *cfs_rq, *pos;
12202 for_each_leaf_cfs_rq_safe(cpu_rq(cpu), cfs_rq, pos)
12203 print_cfs_rq(m, cpu, cfs_rq);
12207 #ifdef CONFIG_NUMA_BALANCING
12208 void show_numa_stats(struct task_struct *p, struct seq_file *m)
12211 unsigned long tsf = 0, tpf = 0, gsf = 0, gpf = 0;
12212 struct numa_group *ng;
12215 ng = rcu_dereference(p->numa_group);
12216 for_each_online_node(node) {
12217 if (p->numa_faults) {
12218 tsf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 0)];
12219 tpf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 1)];
12222 gsf = ng->faults[task_faults_idx(NUMA_MEM, node, 0)],
12223 gpf = ng->faults[task_faults_idx(NUMA_MEM, node, 1)];
12225 print_numa_stats(m, node, tsf, tpf, gsf, gpf);
12229 #endif /* CONFIG_NUMA_BALANCING */
12230 #endif /* CONFIG_SCHED_DEBUG */
12232 __init void init_sched_fair_class(void)
12237 for_each_possible_cpu(i) {
12238 zalloc_cpumask_var_node(&per_cpu(load_balance_mask, i), GFP_KERNEL, cpu_to_node(i));
12239 zalloc_cpumask_var_node(&per_cpu(select_rq_mask, i), GFP_KERNEL, cpu_to_node(i));
12242 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
12244 #ifdef CONFIG_NO_HZ_COMMON
12245 nohz.next_balance = jiffies;
12246 nohz.next_blocked = jiffies;
12247 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);