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>
50 #include <linux/rbtree_augmented.h>
52 #include <asm/switch_to.h>
56 #include "autogroup.h"
59 * The initial- and re-scaling of tunables is configurable
63 * SCHED_TUNABLESCALING_NONE - unscaled, always *1
64 * SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus)
65 * SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus
67 * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus))
69 unsigned int sysctl_sched_tunable_scaling = SCHED_TUNABLESCALING_LOG;
72 * Minimal preemption granularity for CPU-bound tasks:
74 * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds)
76 unsigned int sysctl_sched_base_slice = 750000ULL;
77 static unsigned int normalized_sysctl_sched_base_slice = 750000ULL;
79 const_debug unsigned int sysctl_sched_migration_cost = 500000UL;
81 static int __init setup_sched_thermal_decay_shift(char *str)
83 pr_warn("Ignoring the deprecated sched_thermal_decay_shift= option\n");
86 __setup("sched_thermal_decay_shift=", setup_sched_thermal_decay_shift);
90 * For asym packing, by default the lower numbered CPU has higher priority.
92 int __weak arch_asym_cpu_priority(int cpu)
98 * The margin used when comparing utilization with CPU capacity.
102 #define fits_capacity(cap, max) ((cap) * 1280 < (max) * 1024)
105 * The margin used when comparing CPU capacities.
106 * is 'cap1' noticeably greater than 'cap2'
110 #define capacity_greater(cap1, cap2) ((cap1) * 1024 > (cap2) * 1078)
113 #ifdef CONFIG_CFS_BANDWIDTH
115 * Amount of runtime to allocate from global (tg) to local (per-cfs_rq) pool
116 * each time a cfs_rq requests quota.
118 * Note: in the case that the slice exceeds the runtime remaining (either due
119 * to consumption or the quota being specified to be smaller than the slice)
120 * we will always only issue the remaining available time.
122 * (default: 5 msec, units: microseconds)
124 static unsigned int sysctl_sched_cfs_bandwidth_slice = 5000UL;
127 #ifdef CONFIG_NUMA_BALANCING
128 /* Restrict the NUMA promotion throughput (MB/s) for each target node. */
129 static unsigned int sysctl_numa_balancing_promote_rate_limit = 65536;
133 static struct ctl_table sched_fair_sysctls[] = {
134 #ifdef CONFIG_CFS_BANDWIDTH
136 .procname = "sched_cfs_bandwidth_slice_us",
137 .data = &sysctl_sched_cfs_bandwidth_slice,
138 .maxlen = sizeof(unsigned int),
140 .proc_handler = proc_dointvec_minmax,
141 .extra1 = SYSCTL_ONE,
144 #ifdef CONFIG_NUMA_BALANCING
146 .procname = "numa_balancing_promote_rate_limit_MBps",
147 .data = &sysctl_numa_balancing_promote_rate_limit,
148 .maxlen = sizeof(unsigned int),
150 .proc_handler = proc_dointvec_minmax,
151 .extra1 = SYSCTL_ZERO,
153 #endif /* CONFIG_NUMA_BALANCING */
156 static int __init sched_fair_sysctl_init(void)
158 register_sysctl_init("kernel", sched_fair_sysctls);
161 late_initcall(sched_fair_sysctl_init);
164 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
170 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
176 static inline void update_load_set(struct load_weight *lw, unsigned long w)
183 * Increase the granularity value when there are more CPUs,
184 * because with more CPUs the 'effective latency' as visible
185 * to users decreases. But the relationship is not linear,
186 * so pick a second-best guess by going with the log2 of the
189 * This idea comes from the SD scheduler of Con Kolivas:
191 static unsigned int get_update_sysctl_factor(void)
193 unsigned int cpus = min_t(unsigned int, num_online_cpus(), 8);
196 switch (sysctl_sched_tunable_scaling) {
197 case SCHED_TUNABLESCALING_NONE:
200 case SCHED_TUNABLESCALING_LINEAR:
203 case SCHED_TUNABLESCALING_LOG:
205 factor = 1 + ilog2(cpus);
212 static void update_sysctl(void)
214 unsigned int factor = get_update_sysctl_factor();
216 #define SET_SYSCTL(name) \
217 (sysctl_##name = (factor) * normalized_sysctl_##name)
218 SET_SYSCTL(sched_base_slice);
222 void __init sched_init_granularity(void)
227 #define WMULT_CONST (~0U)
228 #define WMULT_SHIFT 32
230 static void __update_inv_weight(struct load_weight *lw)
234 if (likely(lw->inv_weight))
237 w = scale_load_down(lw->weight);
239 if (BITS_PER_LONG > 32 && unlikely(w >= WMULT_CONST))
241 else if (unlikely(!w))
242 lw->inv_weight = WMULT_CONST;
244 lw->inv_weight = WMULT_CONST / w;
248 * delta_exec * weight / lw.weight
250 * (delta_exec * (weight * lw->inv_weight)) >> WMULT_SHIFT
252 * Either weight := NICE_0_LOAD and lw \e sched_prio_to_wmult[], in which case
253 * we're guaranteed shift stays positive because inv_weight is guaranteed to
254 * fit 32 bits, and NICE_0_LOAD gives another 10 bits; therefore shift >= 22.
256 * Or, weight =< lw.weight (because lw.weight is the runqueue weight), thus
257 * weight/lw.weight <= 1, and therefore our shift will also be positive.
259 static u64 __calc_delta(u64 delta_exec, unsigned long weight, struct load_weight *lw)
261 u64 fact = scale_load_down(weight);
262 u32 fact_hi = (u32)(fact >> 32);
263 int shift = WMULT_SHIFT;
266 __update_inv_weight(lw);
268 if (unlikely(fact_hi)) {
274 fact = mul_u32_u32(fact, lw->inv_weight);
276 fact_hi = (u32)(fact >> 32);
283 return mul_u64_u32_shr(delta_exec, fact, shift);
289 static inline u64 calc_delta_fair(u64 delta, struct sched_entity *se)
291 if (unlikely(se->load.weight != NICE_0_LOAD))
292 delta = __calc_delta(delta, NICE_0_LOAD, &se->load);
297 const struct sched_class fair_sched_class;
299 /**************************************************************
300 * CFS operations on generic schedulable entities:
303 #ifdef CONFIG_FAIR_GROUP_SCHED
305 /* Walk up scheduling entities hierarchy */
306 #define for_each_sched_entity(se) \
307 for (; se; se = se->parent)
309 static inline bool list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
311 struct rq *rq = rq_of(cfs_rq);
312 int cpu = cpu_of(rq);
315 return rq->tmp_alone_branch == &rq->leaf_cfs_rq_list;
320 * Ensure we either appear before our parent (if already
321 * enqueued) or force our parent to appear after us when it is
322 * enqueued. The fact that we always enqueue bottom-up
323 * reduces this to two cases and a special case for the root
324 * cfs_rq. Furthermore, it also means that we will always reset
325 * tmp_alone_branch either when the branch is connected
326 * to a tree or when we reach the top of the tree
328 if (cfs_rq->tg->parent &&
329 cfs_rq->tg->parent->cfs_rq[cpu]->on_list) {
331 * If parent is already on the list, we add the child
332 * just before. Thanks to circular linked property of
333 * the list, this means to put the child at the tail
334 * of the list that starts by parent.
336 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
337 &(cfs_rq->tg->parent->cfs_rq[cpu]->leaf_cfs_rq_list));
339 * The branch is now connected to its tree so we can
340 * reset tmp_alone_branch to the beginning of the
343 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
347 if (!cfs_rq->tg->parent) {
349 * cfs rq without parent should be put
350 * at the tail of the list.
352 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
353 &rq->leaf_cfs_rq_list);
355 * We have reach the top of a tree so we can reset
356 * tmp_alone_branch to the beginning of the list.
358 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
363 * The parent has not already been added so we want to
364 * make sure that it will be put after us.
365 * tmp_alone_branch points to the begin of the branch
366 * where we will add parent.
368 list_add_rcu(&cfs_rq->leaf_cfs_rq_list, rq->tmp_alone_branch);
370 * update tmp_alone_branch to points to the new begin
373 rq->tmp_alone_branch = &cfs_rq->leaf_cfs_rq_list;
377 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
379 if (cfs_rq->on_list) {
380 struct rq *rq = rq_of(cfs_rq);
383 * With cfs_rq being unthrottled/throttled during an enqueue,
384 * it can happen the tmp_alone_branch points to the leaf that
385 * we finally want to delete. In this case, tmp_alone_branch moves
386 * to the prev element but it will point to rq->leaf_cfs_rq_list
387 * at the end of the enqueue.
389 if (rq->tmp_alone_branch == &cfs_rq->leaf_cfs_rq_list)
390 rq->tmp_alone_branch = cfs_rq->leaf_cfs_rq_list.prev;
392 list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
397 static inline void assert_list_leaf_cfs_rq(struct rq *rq)
399 SCHED_WARN_ON(rq->tmp_alone_branch != &rq->leaf_cfs_rq_list);
402 /* Iterate through all leaf cfs_rq's on a runqueue */
403 #define for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) \
404 list_for_each_entry_safe(cfs_rq, pos, &rq->leaf_cfs_rq_list, \
407 /* Do the two (enqueued) entities belong to the same group ? */
408 static inline struct cfs_rq *
409 is_same_group(struct sched_entity *se, struct sched_entity *pse)
411 if (se->cfs_rq == pse->cfs_rq)
417 static inline struct sched_entity *parent_entity(const struct sched_entity *se)
423 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
425 int se_depth, pse_depth;
428 * preemption test can be made between sibling entities who are in the
429 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
430 * both tasks until we find their ancestors who are siblings of common
434 /* First walk up until both entities are at same depth */
435 se_depth = (*se)->depth;
436 pse_depth = (*pse)->depth;
438 while (se_depth > pse_depth) {
440 *se = parent_entity(*se);
443 while (pse_depth > se_depth) {
445 *pse = parent_entity(*pse);
448 while (!is_same_group(*se, *pse)) {
449 *se = parent_entity(*se);
450 *pse = parent_entity(*pse);
454 static int tg_is_idle(struct task_group *tg)
459 static int cfs_rq_is_idle(struct cfs_rq *cfs_rq)
461 return cfs_rq->idle > 0;
464 static int se_is_idle(struct sched_entity *se)
466 if (entity_is_task(se))
467 return task_has_idle_policy(task_of(se));
468 return cfs_rq_is_idle(group_cfs_rq(se));
471 #else /* !CONFIG_FAIR_GROUP_SCHED */
473 #define for_each_sched_entity(se) \
474 for (; se; se = NULL)
476 static inline bool list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
481 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
485 static inline void assert_list_leaf_cfs_rq(struct rq *rq)
489 #define for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) \
490 for (cfs_rq = &rq->cfs, pos = NULL; cfs_rq; cfs_rq = pos)
492 static inline struct sched_entity *parent_entity(struct sched_entity *se)
498 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
502 static inline int tg_is_idle(struct task_group *tg)
507 static int cfs_rq_is_idle(struct cfs_rq *cfs_rq)
512 static int se_is_idle(struct sched_entity *se)
517 #endif /* CONFIG_FAIR_GROUP_SCHED */
519 static __always_inline
520 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec);
522 /**************************************************************
523 * Scheduling class tree data structure manipulation methods:
526 static inline u64 max_vruntime(u64 max_vruntime, u64 vruntime)
528 s64 delta = (s64)(vruntime - max_vruntime);
530 max_vruntime = vruntime;
535 static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime)
537 s64 delta = (s64)(vruntime - min_vruntime);
539 min_vruntime = vruntime;
544 static inline bool entity_before(const struct sched_entity *a,
545 const struct sched_entity *b)
548 * Tiebreak on vruntime seems unnecessary since it can
551 return (s64)(a->deadline - b->deadline) < 0;
554 static inline s64 entity_key(struct cfs_rq *cfs_rq, struct sched_entity *se)
556 return (s64)(se->vruntime - cfs_rq->min_vruntime);
559 #define __node_2_se(node) \
560 rb_entry((node), struct sched_entity, run_node)
563 * Compute virtual time from the per-task service numbers:
565 * Fair schedulers conserve lag:
569 * Where lag_i is given by:
571 * lag_i = S - s_i = w_i * (V - v_i)
573 * Where S is the ideal service time and V is it's virtual time counterpart.
577 * \Sum w_i * (V - v_i) = 0
578 * \Sum w_i * V - w_i * v_i = 0
580 * From which we can solve an expression for V in v_i (which we have in
583 * \Sum v_i * w_i \Sum v_i * w_i
584 * V = -------------- = --------------
587 * Specifically, this is the weighted average of all entity virtual runtimes.
589 * [[ NOTE: this is only equal to the ideal scheduler under the condition
590 * that join/leave operations happen at lag_i = 0, otherwise the
591 * virtual time has non-contiguous motion equivalent to:
595 * Also see the comment in place_entity() that deals with this. ]]
597 * However, since v_i is u64, and the multiplication could easily overflow
598 * transform it into a relative form that uses smaller quantities:
600 * Substitute: v_i == (v_i - v0) + v0
602 * \Sum ((v_i - v0) + v0) * w_i \Sum (v_i - v0) * w_i
603 * V = ---------------------------- = --------------------- + v0
606 * Which we track using:
608 * v0 := cfs_rq->min_vruntime
609 * \Sum (v_i - v0) * w_i := cfs_rq->avg_vruntime
610 * \Sum w_i := cfs_rq->avg_load
612 * Since min_vruntime is a monotonic increasing variable that closely tracks
613 * the per-task service, these deltas: (v_i - v), will be in the order of the
614 * maximal (virtual) lag induced in the system due to quantisation.
616 * Also, we use scale_load_down() to reduce the size.
618 * As measured, the max (key * weight) value was ~44 bits for a kernel build.
621 avg_vruntime_add(struct cfs_rq *cfs_rq, struct sched_entity *se)
623 unsigned long weight = scale_load_down(se->load.weight);
624 s64 key = entity_key(cfs_rq, se);
626 cfs_rq->avg_vruntime += key * weight;
627 cfs_rq->avg_load += weight;
631 avg_vruntime_sub(struct cfs_rq *cfs_rq, struct sched_entity *se)
633 unsigned long weight = scale_load_down(se->load.weight);
634 s64 key = entity_key(cfs_rq, se);
636 cfs_rq->avg_vruntime -= key * weight;
637 cfs_rq->avg_load -= weight;
641 void avg_vruntime_update(struct cfs_rq *cfs_rq, s64 delta)
644 * v' = v + d ==> avg_vruntime' = avg_runtime - d*avg_load
646 cfs_rq->avg_vruntime -= cfs_rq->avg_load * delta;
650 * Specifically: avg_runtime() + 0 must result in entity_eligible() := true
651 * For this to be so, the result of this function must have a left bias.
653 u64 avg_vruntime(struct cfs_rq *cfs_rq)
655 struct sched_entity *curr = cfs_rq->curr;
656 s64 avg = cfs_rq->avg_vruntime;
657 long load = cfs_rq->avg_load;
659 if (curr && curr->on_rq) {
660 unsigned long weight = scale_load_down(curr->load.weight);
662 avg += entity_key(cfs_rq, curr) * weight;
667 /* sign flips effective floor / ceiling */
670 avg = div_s64(avg, load);
673 return cfs_rq->min_vruntime + avg;
677 * lag_i = S - s_i = w_i * (V - v_i)
679 * However, since V is approximated by the weighted average of all entities it
680 * is possible -- by addition/removal/reweight to the tree -- to move V around
681 * and end up with a larger lag than we started with.
683 * Limit this to either double the slice length with a minimum of TICK_NSEC
684 * since that is the timing granularity.
686 * EEVDF gives the following limit for a steady state system:
688 * -r_max < lag < max(r_max, q)
690 * XXX could add max_slice to the augmented data to track this.
692 static s64 entity_lag(u64 avruntime, struct sched_entity *se)
696 vlag = avruntime - se->vruntime;
697 limit = calc_delta_fair(max_t(u64, 2*se->slice, TICK_NSEC), se);
699 return clamp(vlag, -limit, limit);
702 static void update_entity_lag(struct cfs_rq *cfs_rq, struct sched_entity *se)
704 SCHED_WARN_ON(!se->on_rq);
706 se->vlag = entity_lag(avg_vruntime(cfs_rq), se);
710 * Entity is eligible once it received less service than it ought to have,
713 * lag_i = S - s_i = w_i*(V - v_i)
715 * lag_i >= 0 -> V >= v_i
718 * V = ------------------ + v
721 * lag_i >= 0 -> \Sum (v_i - v)*w_i >= (v_i - v)*(\Sum w_i)
723 * Note: using 'avg_vruntime() > se->vruntime' is inaccurate due
724 * to the loss in precision caused by the division.
726 static int vruntime_eligible(struct cfs_rq *cfs_rq, u64 vruntime)
728 struct sched_entity *curr = cfs_rq->curr;
729 s64 avg = cfs_rq->avg_vruntime;
730 long load = cfs_rq->avg_load;
732 if (curr && curr->on_rq) {
733 unsigned long weight = scale_load_down(curr->load.weight);
735 avg += entity_key(cfs_rq, curr) * weight;
739 return avg >= (s64)(vruntime - cfs_rq->min_vruntime) * load;
742 int entity_eligible(struct cfs_rq *cfs_rq, struct sched_entity *se)
744 return vruntime_eligible(cfs_rq, se->vruntime);
747 static u64 __update_min_vruntime(struct cfs_rq *cfs_rq, u64 vruntime)
749 u64 min_vruntime = cfs_rq->min_vruntime;
751 * open coded max_vruntime() to allow updating avg_vruntime
753 s64 delta = (s64)(vruntime - min_vruntime);
755 avg_vruntime_update(cfs_rq, delta);
756 min_vruntime = vruntime;
761 static void update_min_vruntime(struct cfs_rq *cfs_rq)
763 struct sched_entity *se = __pick_root_entity(cfs_rq);
764 struct sched_entity *curr = cfs_rq->curr;
765 u64 vruntime = cfs_rq->min_vruntime;
769 vruntime = curr->vruntime;
776 vruntime = se->min_vruntime;
778 vruntime = min_vruntime(vruntime, se->min_vruntime);
781 /* ensure we never gain time by being placed backwards. */
782 u64_u32_store(cfs_rq->min_vruntime,
783 __update_min_vruntime(cfs_rq, vruntime));
786 static inline bool __entity_less(struct rb_node *a, const struct rb_node *b)
788 return entity_before(__node_2_se(a), __node_2_se(b));
791 #define vruntime_gt(field, lse, rse) ({ (s64)((lse)->field - (rse)->field) > 0; })
793 static inline void __min_vruntime_update(struct sched_entity *se, struct rb_node *node)
796 struct sched_entity *rse = __node_2_se(node);
797 if (vruntime_gt(min_vruntime, se, rse))
798 se->min_vruntime = rse->min_vruntime;
803 * se->min_vruntime = min(se->vruntime, {left,right}->min_vruntime)
805 static inline bool min_vruntime_update(struct sched_entity *se, bool exit)
807 u64 old_min_vruntime = se->min_vruntime;
808 struct rb_node *node = &se->run_node;
810 se->min_vruntime = se->vruntime;
811 __min_vruntime_update(se, node->rb_right);
812 __min_vruntime_update(se, node->rb_left);
814 return se->min_vruntime == old_min_vruntime;
817 RB_DECLARE_CALLBACKS(static, min_vruntime_cb, struct sched_entity,
818 run_node, min_vruntime, min_vruntime_update);
821 * Enqueue an entity into the rb-tree:
823 static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
825 avg_vruntime_add(cfs_rq, se);
826 se->min_vruntime = se->vruntime;
827 rb_add_augmented_cached(&se->run_node, &cfs_rq->tasks_timeline,
828 __entity_less, &min_vruntime_cb);
831 static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
833 rb_erase_augmented_cached(&se->run_node, &cfs_rq->tasks_timeline,
835 avg_vruntime_sub(cfs_rq, se);
838 struct sched_entity *__pick_root_entity(struct cfs_rq *cfs_rq)
840 struct rb_node *root = cfs_rq->tasks_timeline.rb_root.rb_node;
845 return __node_2_se(root);
848 struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq)
850 struct rb_node *left = rb_first_cached(&cfs_rq->tasks_timeline);
855 return __node_2_se(left);
859 * Earliest Eligible Virtual Deadline First
861 * In order to provide latency guarantees for different request sizes
862 * EEVDF selects the best runnable task from two criteria:
864 * 1) the task must be eligible (must be owed service)
866 * 2) from those tasks that meet 1), we select the one
867 * with the earliest virtual deadline.
869 * We can do this in O(log n) time due to an augmented RB-tree. The
870 * tree keeps the entries sorted on deadline, but also functions as a
871 * heap based on the vruntime by keeping:
873 * se->min_vruntime = min(se->vruntime, se->{left,right}->min_vruntime)
875 * Which allows tree pruning through eligibility.
877 static struct sched_entity *pick_eevdf(struct cfs_rq *cfs_rq)
879 struct rb_node *node = cfs_rq->tasks_timeline.rb_root.rb_node;
880 struct sched_entity *se = __pick_first_entity(cfs_rq);
881 struct sched_entity *curr = cfs_rq->curr;
882 struct sched_entity *best = NULL;
885 * We can safely skip eligibility check if there is only one entity
886 * in this cfs_rq, saving some cycles.
888 if (cfs_rq->nr_running == 1)
889 return curr && curr->on_rq ? curr : se;
891 if (curr && (!curr->on_rq || !entity_eligible(cfs_rq, curr)))
895 * Once selected, run a task until it either becomes non-eligible or
896 * until it gets a new slice. See the HACK in set_next_entity().
898 if (sched_feat(RUN_TO_PARITY) && curr && curr->vlag == curr->deadline)
901 /* Pick the leftmost entity if it's eligible */
902 if (se && entity_eligible(cfs_rq, se)) {
907 /* Heap search for the EEVD entity */
909 struct rb_node *left = node->rb_left;
912 * Eligible entities in left subtree are always better
913 * choices, since they have earlier deadlines.
915 if (left && vruntime_eligible(cfs_rq,
916 __node_2_se(left)->min_vruntime)) {
921 se = __node_2_se(node);
924 * The left subtree either is empty or has no eligible
925 * entity, so check the current node since it is the one
926 * with earliest deadline that might be eligible.
928 if (entity_eligible(cfs_rq, se)) {
933 node = node->rb_right;
936 if (!best || (curr && entity_before(curr, best)))
942 #ifdef CONFIG_SCHED_DEBUG
943 struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
945 struct rb_node *last = rb_last(&cfs_rq->tasks_timeline.rb_root);
950 return __node_2_se(last);
953 /**************************************************************
954 * Scheduling class statistics methods:
957 int sched_update_scaling(void)
959 unsigned int factor = get_update_sysctl_factor();
961 #define WRT_SYSCTL(name) \
962 (normalized_sysctl_##name = sysctl_##name / (factor))
963 WRT_SYSCTL(sched_base_slice);
971 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se);
974 * XXX: strictly: vd_i += N*r_i/w_i such that: vd_i > ve_i
975 * this is probably good enough.
977 static void update_deadline(struct cfs_rq *cfs_rq, struct sched_entity *se)
979 if ((s64)(se->vruntime - se->deadline) < 0)
983 * For EEVDF the virtual time slope is determined by w_i (iow.
984 * nice) while the request time r_i is determined by
985 * sysctl_sched_base_slice.
987 se->slice = sysctl_sched_base_slice;
990 * EEVDF: vd_i = ve_i + r_i / w_i
992 se->deadline = se->vruntime + calc_delta_fair(se->slice, se);
995 * The task has consumed its request, reschedule.
997 if (cfs_rq->nr_running > 1) {
998 resched_curr(rq_of(cfs_rq));
999 clear_buddies(cfs_rq, se);
1006 static int select_idle_sibling(struct task_struct *p, int prev_cpu, int cpu);
1007 static unsigned long task_h_load(struct task_struct *p);
1008 static unsigned long capacity_of(int cpu);
1010 /* Give new sched_entity start runnable values to heavy its load in infant time */
1011 void init_entity_runnable_average(struct sched_entity *se)
1013 struct sched_avg *sa = &se->avg;
1015 memset(sa, 0, sizeof(*sa));
1018 * Tasks are initialized with full load to be seen as heavy tasks until
1019 * they get a chance to stabilize to their real load level.
1020 * Group entities are initialized with zero load to reflect the fact that
1021 * nothing has been attached to the task group yet.
1023 if (entity_is_task(se))
1024 sa->load_avg = scale_load_down(se->load.weight);
1026 /* when this task is enqueued, it will contribute to its cfs_rq's load_avg */
1030 * With new tasks being created, their initial util_avgs are extrapolated
1031 * based on the cfs_rq's current util_avg:
1033 * util_avg = cfs_rq->avg.util_avg / (cfs_rq->avg.load_avg + 1)
1036 * However, in many cases, the above util_avg does not give a desired
1037 * value. Moreover, the sum of the util_avgs may be divergent, such
1038 * as when the series is a harmonic series.
1040 * To solve this problem, we also cap the util_avg of successive tasks to
1041 * only 1/2 of the left utilization budget:
1043 * util_avg_cap = (cpu_scale - cfs_rq->avg.util_avg) / 2^n
1045 * where n denotes the nth task and cpu_scale the CPU capacity.
1047 * For example, for a CPU with 1024 of capacity, a simplest series from
1048 * the beginning would be like:
1050 * task util_avg: 512, 256, 128, 64, 32, 16, 8, ...
1051 * cfs_rq util_avg: 512, 768, 896, 960, 992, 1008, 1016, ...
1053 * Finally, that extrapolated util_avg is clamped to the cap (util_avg_cap)
1054 * if util_avg > util_avg_cap.
1056 void post_init_entity_util_avg(struct task_struct *p)
1058 struct sched_entity *se = &p->se;
1059 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1060 struct sched_avg *sa = &se->avg;
1061 long cpu_scale = arch_scale_cpu_capacity(cpu_of(rq_of(cfs_rq)));
1062 long cap = (long)(cpu_scale - cfs_rq->avg.util_avg) / 2;
1064 if (p->sched_class != &fair_sched_class) {
1066 * For !fair tasks do:
1068 update_cfs_rq_load_avg(now, cfs_rq);
1069 attach_entity_load_avg(cfs_rq, se);
1070 switched_from_fair(rq, p);
1072 * such that the next switched_to_fair() has the
1075 se->avg.last_update_time = cfs_rq_clock_pelt(cfs_rq);
1080 if (cfs_rq->avg.util_avg != 0) {
1081 sa->util_avg = cfs_rq->avg.util_avg * se_weight(se);
1082 sa->util_avg /= (cfs_rq->avg.load_avg + 1);
1084 if (sa->util_avg > cap)
1091 sa->runnable_avg = sa->util_avg;
1094 #else /* !CONFIG_SMP */
1095 void init_entity_runnable_average(struct sched_entity *se)
1098 void post_init_entity_util_avg(struct task_struct *p)
1101 static void update_tg_load_avg(struct cfs_rq *cfs_rq)
1104 #endif /* CONFIG_SMP */
1106 static s64 update_curr_se(struct rq *rq, struct sched_entity *curr)
1108 u64 now = rq_clock_task(rq);
1111 delta_exec = now - curr->exec_start;
1112 if (unlikely(delta_exec <= 0))
1115 curr->exec_start = now;
1116 curr->sum_exec_runtime += delta_exec;
1118 if (schedstat_enabled()) {
1119 struct sched_statistics *stats;
1121 stats = __schedstats_from_se(curr);
1122 __schedstat_set(stats->exec_max,
1123 max(delta_exec, stats->exec_max));
1129 static inline void update_curr_task(struct task_struct *p, s64 delta_exec)
1131 trace_sched_stat_runtime(p, delta_exec);
1132 account_group_exec_runtime(p, delta_exec);
1133 cgroup_account_cputime(p, delta_exec);
1135 dl_server_update(p->dl_server, delta_exec);
1139 * Used by other classes to account runtime.
1141 s64 update_curr_common(struct rq *rq)
1143 struct task_struct *curr = rq->curr;
1146 delta_exec = update_curr_se(rq, &curr->se);
1147 if (likely(delta_exec > 0))
1148 update_curr_task(curr, delta_exec);
1154 * Update the current task's runtime statistics.
1156 static void update_curr(struct cfs_rq *cfs_rq)
1158 struct sched_entity *curr = cfs_rq->curr;
1161 if (unlikely(!curr))
1164 delta_exec = update_curr_se(rq_of(cfs_rq), curr);
1165 if (unlikely(delta_exec <= 0))
1168 curr->vruntime += calc_delta_fair(delta_exec, curr);
1169 update_deadline(cfs_rq, curr);
1170 update_min_vruntime(cfs_rq);
1172 if (entity_is_task(curr))
1173 update_curr_task(task_of(curr), delta_exec);
1175 account_cfs_rq_runtime(cfs_rq, delta_exec);
1178 static void update_curr_fair(struct rq *rq)
1180 update_curr(cfs_rq_of(&rq->curr->se));
1184 update_stats_wait_start_fair(struct cfs_rq *cfs_rq, struct sched_entity *se)
1186 struct sched_statistics *stats;
1187 struct task_struct *p = NULL;
1189 if (!schedstat_enabled())
1192 stats = __schedstats_from_se(se);
1194 if (entity_is_task(se))
1197 __update_stats_wait_start(rq_of(cfs_rq), p, stats);
1201 update_stats_wait_end_fair(struct cfs_rq *cfs_rq, struct sched_entity *se)
1203 struct sched_statistics *stats;
1204 struct task_struct *p = NULL;
1206 if (!schedstat_enabled())
1209 stats = __schedstats_from_se(se);
1212 * When the sched_schedstat changes from 0 to 1, some sched se
1213 * maybe already in the runqueue, the se->statistics.wait_start
1214 * will be 0.So it will let the delta wrong. We need to avoid this
1217 if (unlikely(!schedstat_val(stats->wait_start)))
1220 if (entity_is_task(se))
1223 __update_stats_wait_end(rq_of(cfs_rq), p, stats);
1227 update_stats_enqueue_sleeper_fair(struct cfs_rq *cfs_rq, struct sched_entity *se)
1229 struct sched_statistics *stats;
1230 struct task_struct *tsk = NULL;
1232 if (!schedstat_enabled())
1235 stats = __schedstats_from_se(se);
1237 if (entity_is_task(se))
1240 __update_stats_enqueue_sleeper(rq_of(cfs_rq), tsk, stats);
1244 * Task is being enqueued - update stats:
1247 update_stats_enqueue_fair(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
1249 if (!schedstat_enabled())
1253 * Are we enqueueing a waiting task? (for current tasks
1254 * a dequeue/enqueue event is a NOP)
1256 if (se != cfs_rq->curr)
1257 update_stats_wait_start_fair(cfs_rq, se);
1259 if (flags & ENQUEUE_WAKEUP)
1260 update_stats_enqueue_sleeper_fair(cfs_rq, se);
1264 update_stats_dequeue_fair(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
1267 if (!schedstat_enabled())
1271 * Mark the end of the wait period if dequeueing a
1274 if (se != cfs_rq->curr)
1275 update_stats_wait_end_fair(cfs_rq, se);
1277 if ((flags & DEQUEUE_SLEEP) && entity_is_task(se)) {
1278 struct task_struct *tsk = task_of(se);
1281 /* XXX racy against TTWU */
1282 state = READ_ONCE(tsk->__state);
1283 if (state & TASK_INTERRUPTIBLE)
1284 __schedstat_set(tsk->stats.sleep_start,
1285 rq_clock(rq_of(cfs_rq)));
1286 if (state & TASK_UNINTERRUPTIBLE)
1287 __schedstat_set(tsk->stats.block_start,
1288 rq_clock(rq_of(cfs_rq)));
1293 * We are picking a new current task - update its stats:
1296 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
1299 * We are starting a new run period:
1301 se->exec_start = rq_clock_task(rq_of(cfs_rq));
1304 /**************************************************
1305 * Scheduling class queueing methods:
1308 static inline bool is_core_idle(int cpu)
1310 #ifdef CONFIG_SCHED_SMT
1313 for_each_cpu(sibling, cpu_smt_mask(cpu)) {
1317 if (!idle_cpu(sibling))
1326 #define NUMA_IMBALANCE_MIN 2
1329 adjust_numa_imbalance(int imbalance, int dst_running, int imb_numa_nr)
1332 * Allow a NUMA imbalance if busy CPUs is less than the maximum
1333 * threshold. Above this threshold, individual tasks may be contending
1334 * for both memory bandwidth and any shared HT resources. This is an
1335 * approximation as the number of running tasks may not be related to
1336 * the number of busy CPUs due to sched_setaffinity.
1338 if (dst_running > imb_numa_nr)
1342 * Allow a small imbalance based on a simple pair of communicating
1343 * tasks that remain local when the destination is lightly loaded.
1345 if (imbalance <= NUMA_IMBALANCE_MIN)
1350 #endif /* CONFIG_NUMA */
1352 #ifdef CONFIG_NUMA_BALANCING
1354 * Approximate time to scan a full NUMA task in ms. The task scan period is
1355 * calculated based on the tasks virtual memory size and
1356 * numa_balancing_scan_size.
1358 unsigned int sysctl_numa_balancing_scan_period_min = 1000;
1359 unsigned int sysctl_numa_balancing_scan_period_max = 60000;
1361 /* Portion of address space to scan in MB */
1362 unsigned int sysctl_numa_balancing_scan_size = 256;
1364 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
1365 unsigned int sysctl_numa_balancing_scan_delay = 1000;
1367 /* The page with hint page fault latency < threshold in ms is considered hot */
1368 unsigned int sysctl_numa_balancing_hot_threshold = MSEC_PER_SEC;
1371 refcount_t refcount;
1373 spinlock_t lock; /* nr_tasks, tasks */
1378 struct rcu_head rcu;
1379 unsigned long total_faults;
1380 unsigned long max_faults_cpu;
1382 * faults[] array is split into two regions: faults_mem and faults_cpu.
1384 * Faults_cpu is used to decide whether memory should move
1385 * towards the CPU. As a consequence, these stats are weighted
1386 * more by CPU use than by memory faults.
1388 unsigned long faults[];
1392 * For functions that can be called in multiple contexts that permit reading
1393 * ->numa_group (see struct task_struct for locking rules).
1395 static struct numa_group *deref_task_numa_group(struct task_struct *p)
1397 return rcu_dereference_check(p->numa_group, p == current ||
1398 (lockdep_is_held(__rq_lockp(task_rq(p))) && !READ_ONCE(p->on_cpu)));
1401 static struct numa_group *deref_curr_numa_group(struct task_struct *p)
1403 return rcu_dereference_protected(p->numa_group, p == current);
1406 static inline unsigned long group_faults_priv(struct numa_group *ng);
1407 static inline unsigned long group_faults_shared(struct numa_group *ng);
1409 static unsigned int task_nr_scan_windows(struct task_struct *p)
1411 unsigned long rss = 0;
1412 unsigned long nr_scan_pages;
1415 * Calculations based on RSS as non-present and empty pages are skipped
1416 * by the PTE scanner and NUMA hinting faults should be trapped based
1419 nr_scan_pages = sysctl_numa_balancing_scan_size << (20 - PAGE_SHIFT);
1420 rss = get_mm_rss(p->mm);
1422 rss = nr_scan_pages;
1424 rss = round_up(rss, nr_scan_pages);
1425 return rss / nr_scan_pages;
1428 /* For sanity's sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */
1429 #define MAX_SCAN_WINDOW 2560
1431 static unsigned int task_scan_min(struct task_struct *p)
1433 unsigned int scan_size = READ_ONCE(sysctl_numa_balancing_scan_size);
1434 unsigned int scan, floor;
1435 unsigned int windows = 1;
1437 if (scan_size < MAX_SCAN_WINDOW)
1438 windows = MAX_SCAN_WINDOW / scan_size;
1439 floor = 1000 / windows;
1441 scan = sysctl_numa_balancing_scan_period_min / task_nr_scan_windows(p);
1442 return max_t(unsigned int, floor, scan);
1445 static unsigned int task_scan_start(struct task_struct *p)
1447 unsigned long smin = task_scan_min(p);
1448 unsigned long period = smin;
1449 struct numa_group *ng;
1451 /* Scale the maximum scan period with the amount of shared memory. */
1453 ng = rcu_dereference(p->numa_group);
1455 unsigned long shared = group_faults_shared(ng);
1456 unsigned long private = group_faults_priv(ng);
1458 period *= refcount_read(&ng->refcount);
1459 period *= shared + 1;
1460 period /= private + shared + 1;
1464 return max(smin, period);
1467 static unsigned int task_scan_max(struct task_struct *p)
1469 unsigned long smin = task_scan_min(p);
1471 struct numa_group *ng;
1473 /* Watch for min being lower than max due to floor calculations */
1474 smax = sysctl_numa_balancing_scan_period_max / task_nr_scan_windows(p);
1476 /* Scale the maximum scan period with the amount of shared memory. */
1477 ng = deref_curr_numa_group(p);
1479 unsigned long shared = group_faults_shared(ng);
1480 unsigned long private = group_faults_priv(ng);
1481 unsigned long period = smax;
1483 period *= refcount_read(&ng->refcount);
1484 period *= shared + 1;
1485 period /= private + shared + 1;
1487 smax = max(smax, period);
1490 return max(smin, smax);
1493 static void account_numa_enqueue(struct rq *rq, struct task_struct *p)
1495 rq->nr_numa_running += (p->numa_preferred_nid != NUMA_NO_NODE);
1496 rq->nr_preferred_running += (p->numa_preferred_nid == task_node(p));
1499 static void account_numa_dequeue(struct rq *rq, struct task_struct *p)
1501 rq->nr_numa_running -= (p->numa_preferred_nid != NUMA_NO_NODE);
1502 rq->nr_preferred_running -= (p->numa_preferred_nid == task_node(p));
1505 /* Shared or private faults. */
1506 #define NR_NUMA_HINT_FAULT_TYPES 2
1508 /* Memory and CPU locality */
1509 #define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2)
1511 /* Averaged statistics, and temporary buffers. */
1512 #define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2)
1514 pid_t task_numa_group_id(struct task_struct *p)
1516 struct numa_group *ng;
1520 ng = rcu_dereference(p->numa_group);
1529 * The averaged statistics, shared & private, memory & CPU,
1530 * occupy the first half of the array. The second half of the
1531 * array is for current counters, which are averaged into the
1532 * first set by task_numa_placement.
1534 static inline int task_faults_idx(enum numa_faults_stats s, int nid, int priv)
1536 return NR_NUMA_HINT_FAULT_TYPES * (s * nr_node_ids + nid) + priv;
1539 static inline unsigned long task_faults(struct task_struct *p, int nid)
1541 if (!p->numa_faults)
1544 return p->numa_faults[task_faults_idx(NUMA_MEM, nid, 0)] +
1545 p->numa_faults[task_faults_idx(NUMA_MEM, nid, 1)];
1548 static inline unsigned long group_faults(struct task_struct *p, int nid)
1550 struct numa_group *ng = deref_task_numa_group(p);
1555 return ng->faults[task_faults_idx(NUMA_MEM, nid, 0)] +
1556 ng->faults[task_faults_idx(NUMA_MEM, nid, 1)];
1559 static inline unsigned long group_faults_cpu(struct numa_group *group, int nid)
1561 return group->faults[task_faults_idx(NUMA_CPU, nid, 0)] +
1562 group->faults[task_faults_idx(NUMA_CPU, nid, 1)];
1565 static inline unsigned long group_faults_priv(struct numa_group *ng)
1567 unsigned long faults = 0;
1570 for_each_online_node(node) {
1571 faults += ng->faults[task_faults_idx(NUMA_MEM, node, 1)];
1577 static inline unsigned long group_faults_shared(struct numa_group *ng)
1579 unsigned long faults = 0;
1582 for_each_online_node(node) {
1583 faults += ng->faults[task_faults_idx(NUMA_MEM, node, 0)];
1590 * A node triggering more than 1/3 as many NUMA faults as the maximum is
1591 * considered part of a numa group's pseudo-interleaving set. Migrations
1592 * between these nodes are slowed down, to allow things to settle down.
1594 #define ACTIVE_NODE_FRACTION 3
1596 static bool numa_is_active_node(int nid, struct numa_group *ng)
1598 return group_faults_cpu(ng, nid) * ACTIVE_NODE_FRACTION > ng->max_faults_cpu;
1601 /* Handle placement on systems where not all nodes are directly connected. */
1602 static unsigned long score_nearby_nodes(struct task_struct *p, int nid,
1603 int lim_dist, bool task)
1605 unsigned long score = 0;
1609 * All nodes are directly connected, and the same distance
1610 * from each other. No need for fancy placement algorithms.
1612 if (sched_numa_topology_type == NUMA_DIRECT)
1615 /* sched_max_numa_distance may be changed in parallel. */
1616 max_dist = READ_ONCE(sched_max_numa_distance);
1618 * This code is called for each node, introducing N^2 complexity,
1619 * which should be OK given the number of nodes rarely exceeds 8.
1621 for_each_online_node(node) {
1622 unsigned long faults;
1623 int dist = node_distance(nid, node);
1626 * The furthest away nodes in the system are not interesting
1627 * for placement; nid was already counted.
1629 if (dist >= max_dist || node == nid)
1633 * On systems with a backplane NUMA topology, compare groups
1634 * of nodes, and move tasks towards the group with the most
1635 * memory accesses. When comparing two nodes at distance
1636 * "hoplimit", only nodes closer by than "hoplimit" are part
1637 * of each group. Skip other nodes.
1639 if (sched_numa_topology_type == NUMA_BACKPLANE && dist >= lim_dist)
1642 /* Add up the faults from nearby nodes. */
1644 faults = task_faults(p, node);
1646 faults = group_faults(p, node);
1649 * On systems with a glueless mesh NUMA topology, there are
1650 * no fixed "groups of nodes". Instead, nodes that are not
1651 * directly connected bounce traffic through intermediate
1652 * nodes; a numa_group can occupy any set of nodes.
1653 * The further away a node is, the less the faults count.
1654 * This seems to result in good task placement.
1656 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
1657 faults *= (max_dist - dist);
1658 faults /= (max_dist - LOCAL_DISTANCE);
1668 * These return the fraction of accesses done by a particular task, or
1669 * task group, on a particular numa node. The group weight is given a
1670 * larger multiplier, in order to group tasks together that are almost
1671 * evenly spread out between numa nodes.
1673 static inline unsigned long task_weight(struct task_struct *p, int nid,
1676 unsigned long faults, total_faults;
1678 if (!p->numa_faults)
1681 total_faults = p->total_numa_faults;
1686 faults = task_faults(p, nid);
1687 faults += score_nearby_nodes(p, nid, dist, true);
1689 return 1000 * faults / total_faults;
1692 static inline unsigned long group_weight(struct task_struct *p, int nid,
1695 struct numa_group *ng = deref_task_numa_group(p);
1696 unsigned long faults, total_faults;
1701 total_faults = ng->total_faults;
1706 faults = group_faults(p, nid);
1707 faults += score_nearby_nodes(p, nid, dist, false);
1709 return 1000 * faults / total_faults;
1713 * If memory tiering mode is enabled, cpupid of slow memory page is
1714 * used to record scan time instead of CPU and PID. When tiering mode
1715 * is disabled at run time, the scan time (in cpupid) will be
1716 * interpreted as CPU and PID. So CPU needs to be checked to avoid to
1717 * access out of array bound.
1719 static inline bool cpupid_valid(int cpupid)
1721 return cpupid_to_cpu(cpupid) < nr_cpu_ids;
1725 * For memory tiering mode, if there are enough free pages (more than
1726 * enough watermark defined here) in fast memory node, to take full
1727 * advantage of fast memory capacity, all recently accessed slow
1728 * memory pages will be migrated to fast memory node without
1729 * considering hot threshold.
1731 static bool pgdat_free_space_enough(struct pglist_data *pgdat)
1734 unsigned long enough_wmark;
1736 enough_wmark = max(1UL * 1024 * 1024 * 1024 >> PAGE_SHIFT,
1737 pgdat->node_present_pages >> 4);
1738 for (z = pgdat->nr_zones - 1; z >= 0; z--) {
1739 struct zone *zone = pgdat->node_zones + z;
1741 if (!populated_zone(zone))
1744 if (zone_watermark_ok(zone, 0,
1745 wmark_pages(zone, WMARK_PROMO) + enough_wmark,
1753 * For memory tiering mode, when page tables are scanned, the scan
1754 * time will be recorded in struct page in addition to make page
1755 * PROT_NONE for slow memory page. So when the page is accessed, in
1756 * hint page fault handler, the hint page fault latency is calculated
1759 * hint page fault latency = hint page fault time - scan time
1761 * The smaller the hint page fault latency, the higher the possibility
1762 * for the page to be hot.
1764 static int numa_hint_fault_latency(struct folio *folio)
1766 int last_time, time;
1768 time = jiffies_to_msecs(jiffies);
1769 last_time = folio_xchg_access_time(folio, time);
1771 return (time - last_time) & PAGE_ACCESS_TIME_MASK;
1775 * For memory tiering mode, too high promotion/demotion throughput may
1776 * hurt application latency. So we provide a mechanism to rate limit
1777 * the number of pages that are tried to be promoted.
1779 static bool numa_promotion_rate_limit(struct pglist_data *pgdat,
1780 unsigned long rate_limit, int nr)
1782 unsigned long nr_cand;
1783 unsigned int now, start;
1785 now = jiffies_to_msecs(jiffies);
1786 mod_node_page_state(pgdat, PGPROMOTE_CANDIDATE, nr);
1787 nr_cand = node_page_state(pgdat, PGPROMOTE_CANDIDATE);
1788 start = pgdat->nbp_rl_start;
1789 if (now - start > MSEC_PER_SEC &&
1790 cmpxchg(&pgdat->nbp_rl_start, start, now) == start)
1791 pgdat->nbp_rl_nr_cand = nr_cand;
1792 if (nr_cand - pgdat->nbp_rl_nr_cand >= rate_limit)
1797 #define NUMA_MIGRATION_ADJUST_STEPS 16
1799 static void numa_promotion_adjust_threshold(struct pglist_data *pgdat,
1800 unsigned long rate_limit,
1801 unsigned int ref_th)
1803 unsigned int now, start, th_period, unit_th, th;
1804 unsigned long nr_cand, ref_cand, diff_cand;
1806 now = jiffies_to_msecs(jiffies);
1807 th_period = sysctl_numa_balancing_scan_period_max;
1808 start = pgdat->nbp_th_start;
1809 if (now - start > th_period &&
1810 cmpxchg(&pgdat->nbp_th_start, start, now) == start) {
1811 ref_cand = rate_limit *
1812 sysctl_numa_balancing_scan_period_max / MSEC_PER_SEC;
1813 nr_cand = node_page_state(pgdat, PGPROMOTE_CANDIDATE);
1814 diff_cand = nr_cand - pgdat->nbp_th_nr_cand;
1815 unit_th = ref_th * 2 / NUMA_MIGRATION_ADJUST_STEPS;
1816 th = pgdat->nbp_threshold ? : ref_th;
1817 if (diff_cand > ref_cand * 11 / 10)
1818 th = max(th - unit_th, unit_th);
1819 else if (diff_cand < ref_cand * 9 / 10)
1820 th = min(th + unit_th, ref_th * 2);
1821 pgdat->nbp_th_nr_cand = nr_cand;
1822 pgdat->nbp_threshold = th;
1826 bool should_numa_migrate_memory(struct task_struct *p, struct folio *folio,
1827 int src_nid, int dst_cpu)
1829 struct numa_group *ng = deref_curr_numa_group(p);
1830 int dst_nid = cpu_to_node(dst_cpu);
1831 int last_cpupid, this_cpupid;
1834 * Cannot migrate to memoryless nodes.
1836 if (!node_state(dst_nid, N_MEMORY))
1840 * The pages in slow memory node should be migrated according
1841 * to hot/cold instead of private/shared.
1843 if (sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING &&
1844 !node_is_toptier(src_nid)) {
1845 struct pglist_data *pgdat;
1846 unsigned long rate_limit;
1847 unsigned int latency, th, def_th;
1849 pgdat = NODE_DATA(dst_nid);
1850 if (pgdat_free_space_enough(pgdat)) {
1851 /* workload changed, reset hot threshold */
1852 pgdat->nbp_threshold = 0;
1856 def_th = sysctl_numa_balancing_hot_threshold;
1857 rate_limit = sysctl_numa_balancing_promote_rate_limit << \
1859 numa_promotion_adjust_threshold(pgdat, rate_limit, def_th);
1861 th = pgdat->nbp_threshold ? : def_th;
1862 latency = numa_hint_fault_latency(folio);
1866 return !numa_promotion_rate_limit(pgdat, rate_limit,
1867 folio_nr_pages(folio));
1870 this_cpupid = cpu_pid_to_cpupid(dst_cpu, current->pid);
1871 last_cpupid = folio_xchg_last_cpupid(folio, this_cpupid);
1873 if (!(sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING) &&
1874 !node_is_toptier(src_nid) && !cpupid_valid(last_cpupid))
1878 * Allow first faults or private faults to migrate immediately early in
1879 * the lifetime of a task. The magic number 4 is based on waiting for
1880 * two full passes of the "multi-stage node selection" test that is
1883 if ((p->numa_preferred_nid == NUMA_NO_NODE || p->numa_scan_seq <= 4) &&
1884 (cpupid_pid_unset(last_cpupid) || cpupid_match_pid(p, last_cpupid)))
1888 * Multi-stage node selection is used in conjunction with a periodic
1889 * migration fault to build a temporal task<->page relation. By using
1890 * a two-stage filter we remove short/unlikely relations.
1892 * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate
1893 * a task's usage of a particular page (n_p) per total usage of this
1894 * page (n_t) (in a given time-span) to a probability.
1896 * Our periodic faults will sample this probability and getting the
1897 * same result twice in a row, given these samples are fully
1898 * independent, is then given by P(n)^2, provided our sample period
1899 * is sufficiently short compared to the usage pattern.
1901 * This quadric squishes small probabilities, making it less likely we
1902 * act on an unlikely task<->page relation.
1904 if (!cpupid_pid_unset(last_cpupid) &&
1905 cpupid_to_nid(last_cpupid) != dst_nid)
1908 /* Always allow migrate on private faults */
1909 if (cpupid_match_pid(p, last_cpupid))
1912 /* A shared fault, but p->numa_group has not been set up yet. */
1917 * Destination node is much more heavily used than the source
1918 * node? Allow migration.
1920 if (group_faults_cpu(ng, dst_nid) > group_faults_cpu(ng, src_nid) *
1921 ACTIVE_NODE_FRACTION)
1925 * Distribute memory according to CPU & memory use on each node,
1926 * with 3/4 hysteresis to avoid unnecessary memory migrations:
1928 * faults_cpu(dst) 3 faults_cpu(src)
1929 * --------------- * - > ---------------
1930 * faults_mem(dst) 4 faults_mem(src)
1932 return group_faults_cpu(ng, dst_nid) * group_faults(p, src_nid) * 3 >
1933 group_faults_cpu(ng, src_nid) * group_faults(p, dst_nid) * 4;
1937 * 'numa_type' describes the node at the moment of load balancing.
1940 /* The node has spare capacity that can be used to run more tasks. */
1943 * The node is fully used and the tasks don't compete for more CPU
1944 * cycles. Nevertheless, some tasks might wait before running.
1948 * The node is overloaded and can't provide expected CPU cycles to all
1954 /* Cached statistics for all CPUs within a node */
1957 unsigned long runnable;
1959 /* Total compute capacity of CPUs on a node */
1960 unsigned long compute_capacity;
1961 unsigned int nr_running;
1962 unsigned int weight;
1963 enum numa_type node_type;
1967 struct task_numa_env {
1968 struct task_struct *p;
1970 int src_cpu, src_nid;
1971 int dst_cpu, dst_nid;
1974 struct numa_stats src_stats, dst_stats;
1979 struct task_struct *best_task;
1984 static unsigned long cpu_load(struct rq *rq);
1985 static unsigned long cpu_runnable(struct rq *rq);
1988 numa_type numa_classify(unsigned int imbalance_pct,
1989 struct numa_stats *ns)
1991 if ((ns->nr_running > ns->weight) &&
1992 (((ns->compute_capacity * 100) < (ns->util * imbalance_pct)) ||
1993 ((ns->compute_capacity * imbalance_pct) < (ns->runnable * 100))))
1994 return node_overloaded;
1996 if ((ns->nr_running < ns->weight) ||
1997 (((ns->compute_capacity * 100) > (ns->util * imbalance_pct)) &&
1998 ((ns->compute_capacity * imbalance_pct) > (ns->runnable * 100))))
1999 return node_has_spare;
2001 return node_fully_busy;
2004 #ifdef CONFIG_SCHED_SMT
2005 /* Forward declarations of select_idle_sibling helpers */
2006 static inline bool test_idle_cores(int cpu);
2007 static inline int numa_idle_core(int idle_core, int cpu)
2009 if (!static_branch_likely(&sched_smt_present) ||
2010 idle_core >= 0 || !test_idle_cores(cpu))
2014 * Prefer cores instead of packing HT siblings
2015 * and triggering future load balancing.
2017 if (is_core_idle(cpu))
2023 static inline int numa_idle_core(int idle_core, int cpu)
2030 * Gather all necessary information to make NUMA balancing placement
2031 * decisions that are compatible with standard load balancer. This
2032 * borrows code and logic from update_sg_lb_stats but sharing a
2033 * common implementation is impractical.
2035 static void update_numa_stats(struct task_numa_env *env,
2036 struct numa_stats *ns, int nid,
2039 int cpu, idle_core = -1;
2041 memset(ns, 0, sizeof(*ns));
2045 for_each_cpu(cpu, cpumask_of_node(nid)) {
2046 struct rq *rq = cpu_rq(cpu);
2048 ns->load += cpu_load(rq);
2049 ns->runnable += cpu_runnable(rq);
2050 ns->util += cpu_util_cfs(cpu);
2051 ns->nr_running += rq->cfs.h_nr_running;
2052 ns->compute_capacity += capacity_of(cpu);
2054 if (find_idle && idle_core < 0 && !rq->nr_running && idle_cpu(cpu)) {
2055 if (READ_ONCE(rq->numa_migrate_on) ||
2056 !cpumask_test_cpu(cpu, env->p->cpus_ptr))
2059 if (ns->idle_cpu == -1)
2062 idle_core = numa_idle_core(idle_core, cpu);
2067 ns->weight = cpumask_weight(cpumask_of_node(nid));
2069 ns->node_type = numa_classify(env->imbalance_pct, ns);
2072 ns->idle_cpu = idle_core;
2075 static void task_numa_assign(struct task_numa_env *env,
2076 struct task_struct *p, long imp)
2078 struct rq *rq = cpu_rq(env->dst_cpu);
2080 /* Check if run-queue part of active NUMA balance. */
2081 if (env->best_cpu != env->dst_cpu && xchg(&rq->numa_migrate_on, 1)) {
2083 int start = env->dst_cpu;
2085 /* Find alternative idle CPU. */
2086 for_each_cpu_wrap(cpu, cpumask_of_node(env->dst_nid), start + 1) {
2087 if (cpu == env->best_cpu || !idle_cpu(cpu) ||
2088 !cpumask_test_cpu(cpu, env->p->cpus_ptr)) {
2093 rq = cpu_rq(env->dst_cpu);
2094 if (!xchg(&rq->numa_migrate_on, 1))
2098 /* Failed to find an alternative idle CPU */
2104 * Clear previous best_cpu/rq numa-migrate flag, since task now
2105 * found a better CPU to move/swap.
2107 if (env->best_cpu != -1 && env->best_cpu != env->dst_cpu) {
2108 rq = cpu_rq(env->best_cpu);
2109 WRITE_ONCE(rq->numa_migrate_on, 0);
2113 put_task_struct(env->best_task);
2118 env->best_imp = imp;
2119 env->best_cpu = env->dst_cpu;
2122 static bool load_too_imbalanced(long src_load, long dst_load,
2123 struct task_numa_env *env)
2126 long orig_src_load, orig_dst_load;
2127 long src_capacity, dst_capacity;
2130 * The load is corrected for the CPU capacity available on each node.
2133 * ------------ vs ---------
2134 * src_capacity dst_capacity
2136 src_capacity = env->src_stats.compute_capacity;
2137 dst_capacity = env->dst_stats.compute_capacity;
2139 imb = abs(dst_load * src_capacity - src_load * dst_capacity);
2141 orig_src_load = env->src_stats.load;
2142 orig_dst_load = env->dst_stats.load;
2144 old_imb = abs(orig_dst_load * src_capacity - orig_src_load * dst_capacity);
2146 /* Would this change make things worse? */
2147 return (imb > old_imb);
2151 * Maximum NUMA importance can be 1998 (2*999);
2152 * SMALLIMP @ 30 would be close to 1998/64.
2153 * Used to deter task migration.
2158 * This checks if the overall compute and NUMA accesses of the system would
2159 * be improved if the source tasks was migrated to the target dst_cpu taking
2160 * into account that it might be best if task running on the dst_cpu should
2161 * be exchanged with the source task
2163 static bool task_numa_compare(struct task_numa_env *env,
2164 long taskimp, long groupimp, bool maymove)
2166 struct numa_group *cur_ng, *p_ng = deref_curr_numa_group(env->p);
2167 struct rq *dst_rq = cpu_rq(env->dst_cpu);
2168 long imp = p_ng ? groupimp : taskimp;
2169 struct task_struct *cur;
2170 long src_load, dst_load;
2171 int dist = env->dist;
2174 bool stopsearch = false;
2176 if (READ_ONCE(dst_rq->numa_migrate_on))
2180 cur = rcu_dereference(dst_rq->curr);
2181 if (cur && ((cur->flags & PF_EXITING) || is_idle_task(cur)))
2185 * Because we have preemption enabled we can get migrated around and
2186 * end try selecting ourselves (current == env->p) as a swap candidate.
2188 if (cur == env->p) {
2194 if (maymove && moveimp >= env->best_imp)
2200 /* Skip this swap candidate if cannot move to the source cpu. */
2201 if (!cpumask_test_cpu(env->src_cpu, cur->cpus_ptr))
2205 * Skip this swap candidate if it is not moving to its preferred
2206 * node and the best task is.
2208 if (env->best_task &&
2209 env->best_task->numa_preferred_nid == env->src_nid &&
2210 cur->numa_preferred_nid != env->src_nid) {
2215 * "imp" is the fault differential for the source task between the
2216 * source and destination node. Calculate the total differential for
2217 * the source task and potential destination task. The more negative
2218 * the value is, the more remote accesses that would be expected to
2219 * be incurred if the tasks were swapped.
2221 * If dst and source tasks are in the same NUMA group, or not
2222 * in any group then look only at task weights.
2224 cur_ng = rcu_dereference(cur->numa_group);
2225 if (cur_ng == p_ng) {
2227 * Do not swap within a group or between tasks that have
2228 * no group if there is spare capacity. Swapping does
2229 * not address the load imbalance and helps one task at
2230 * the cost of punishing another.
2232 if (env->dst_stats.node_type == node_has_spare)
2235 imp = taskimp + task_weight(cur, env->src_nid, dist) -
2236 task_weight(cur, env->dst_nid, dist);
2238 * Add some hysteresis to prevent swapping the
2239 * tasks within a group over tiny differences.
2245 * Compare the group weights. If a task is all by itself
2246 * (not part of a group), use the task weight instead.
2249 imp += group_weight(cur, env->src_nid, dist) -
2250 group_weight(cur, env->dst_nid, dist);
2252 imp += task_weight(cur, env->src_nid, dist) -
2253 task_weight(cur, env->dst_nid, dist);
2256 /* Discourage picking a task already on its preferred node */
2257 if (cur->numa_preferred_nid == env->dst_nid)
2261 * Encourage picking a task that moves to its preferred node.
2262 * This potentially makes imp larger than it's maximum of
2263 * 1998 (see SMALLIMP and task_weight for why) but in this
2264 * case, it does not matter.
2266 if (cur->numa_preferred_nid == env->src_nid)
2269 if (maymove && moveimp > imp && moveimp > env->best_imp) {
2276 * Prefer swapping with a task moving to its preferred node over a
2279 if (env->best_task && cur->numa_preferred_nid == env->src_nid &&
2280 env->best_task->numa_preferred_nid != env->src_nid) {
2285 * If the NUMA importance is less than SMALLIMP,
2286 * task migration might only result in ping pong
2287 * of tasks and also hurt performance due to cache
2290 if (imp < SMALLIMP || imp <= env->best_imp + SMALLIMP / 2)
2294 * In the overloaded case, try and keep the load balanced.
2296 load = task_h_load(env->p) - task_h_load(cur);
2300 dst_load = env->dst_stats.load + load;
2301 src_load = env->src_stats.load - load;
2303 if (load_too_imbalanced(src_load, dst_load, env))
2307 /* Evaluate an idle CPU for a task numa move. */
2309 int cpu = env->dst_stats.idle_cpu;
2311 /* Nothing cached so current CPU went idle since the search. */
2316 * If the CPU is no longer truly idle and the previous best CPU
2317 * is, keep using it.
2319 if (!idle_cpu(cpu) && env->best_cpu >= 0 &&
2320 idle_cpu(env->best_cpu)) {
2321 cpu = env->best_cpu;
2327 task_numa_assign(env, cur, imp);
2330 * If a move to idle is allowed because there is capacity or load
2331 * balance improves then stop the search. While a better swap
2332 * candidate may exist, a search is not free.
2334 if (maymove && !cur && env->best_cpu >= 0 && idle_cpu(env->best_cpu))
2338 * If a swap candidate must be identified and the current best task
2339 * moves its preferred node then stop the search.
2341 if (!maymove && env->best_task &&
2342 env->best_task->numa_preferred_nid == env->src_nid) {
2351 static void task_numa_find_cpu(struct task_numa_env *env,
2352 long taskimp, long groupimp)
2354 bool maymove = false;
2358 * If dst node has spare capacity, then check if there is an
2359 * imbalance that would be overruled by the load balancer.
2361 if (env->dst_stats.node_type == node_has_spare) {
2362 unsigned int imbalance;
2363 int src_running, dst_running;
2366 * Would movement cause an imbalance? Note that if src has
2367 * more running tasks that the imbalance is ignored as the
2368 * move improves the imbalance from the perspective of the
2369 * CPU load balancer.
2371 src_running = env->src_stats.nr_running - 1;
2372 dst_running = env->dst_stats.nr_running + 1;
2373 imbalance = max(0, dst_running - src_running);
2374 imbalance = adjust_numa_imbalance(imbalance, dst_running,
2377 /* Use idle CPU if there is no imbalance */
2380 if (env->dst_stats.idle_cpu >= 0) {
2381 env->dst_cpu = env->dst_stats.idle_cpu;
2382 task_numa_assign(env, NULL, 0);
2387 long src_load, dst_load, load;
2389 * If the improvement from just moving env->p direction is better
2390 * than swapping tasks around, check if a move is possible.
2392 load = task_h_load(env->p);
2393 dst_load = env->dst_stats.load + load;
2394 src_load = env->src_stats.load - load;
2395 maymove = !load_too_imbalanced(src_load, dst_load, env);
2398 for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) {
2399 /* Skip this CPU if the source task cannot migrate */
2400 if (!cpumask_test_cpu(cpu, env->p->cpus_ptr))
2404 if (task_numa_compare(env, taskimp, groupimp, maymove))
2409 static int task_numa_migrate(struct task_struct *p)
2411 struct task_numa_env env = {
2414 .src_cpu = task_cpu(p),
2415 .src_nid = task_node(p),
2417 .imbalance_pct = 112,
2423 unsigned long taskweight, groupweight;
2424 struct sched_domain *sd;
2425 long taskimp, groupimp;
2426 struct numa_group *ng;
2431 * Pick the lowest SD_NUMA domain, as that would have the smallest
2432 * imbalance and would be the first to start moving tasks about.
2434 * And we want to avoid any moving of tasks about, as that would create
2435 * random movement of tasks -- counter the numa conditions we're trying
2439 sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu));
2441 env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2;
2442 env.imb_numa_nr = sd->imb_numa_nr;
2447 * Cpusets can break the scheduler domain tree into smaller
2448 * balance domains, some of which do not cross NUMA boundaries.
2449 * Tasks that are "trapped" in such domains cannot be migrated
2450 * elsewhere, so there is no point in (re)trying.
2452 if (unlikely(!sd)) {
2453 sched_setnuma(p, task_node(p));
2457 env.dst_nid = p->numa_preferred_nid;
2458 dist = env.dist = node_distance(env.src_nid, env.dst_nid);
2459 taskweight = task_weight(p, env.src_nid, dist);
2460 groupweight = group_weight(p, env.src_nid, dist);
2461 update_numa_stats(&env, &env.src_stats, env.src_nid, false);
2462 taskimp = task_weight(p, env.dst_nid, dist) - taskweight;
2463 groupimp = group_weight(p, env.dst_nid, dist) - groupweight;
2464 update_numa_stats(&env, &env.dst_stats, env.dst_nid, true);
2466 /* Try to find a spot on the preferred nid. */
2467 task_numa_find_cpu(&env, taskimp, groupimp);
2470 * Look at other nodes in these cases:
2471 * - there is no space available on the preferred_nid
2472 * - the task is part of a numa_group that is interleaved across
2473 * multiple NUMA nodes; in order to better consolidate the group,
2474 * we need to check other locations.
2476 ng = deref_curr_numa_group(p);
2477 if (env.best_cpu == -1 || (ng && ng->active_nodes > 1)) {
2478 for_each_node_state(nid, N_CPU) {
2479 if (nid == env.src_nid || nid == p->numa_preferred_nid)
2482 dist = node_distance(env.src_nid, env.dst_nid);
2483 if (sched_numa_topology_type == NUMA_BACKPLANE &&
2485 taskweight = task_weight(p, env.src_nid, dist);
2486 groupweight = group_weight(p, env.src_nid, dist);
2489 /* Only consider nodes where both task and groups benefit */
2490 taskimp = task_weight(p, nid, dist) - taskweight;
2491 groupimp = group_weight(p, nid, dist) - groupweight;
2492 if (taskimp < 0 && groupimp < 0)
2497 update_numa_stats(&env, &env.dst_stats, env.dst_nid, true);
2498 task_numa_find_cpu(&env, taskimp, groupimp);
2503 * If the task is part of a workload that spans multiple NUMA nodes,
2504 * and is migrating into one of the workload's active nodes, remember
2505 * this node as the task's preferred numa node, so the workload can
2507 * A task that migrated to a second choice node will be better off
2508 * trying for a better one later. Do not set the preferred node here.
2511 if (env.best_cpu == -1)
2514 nid = cpu_to_node(env.best_cpu);
2516 if (nid != p->numa_preferred_nid)
2517 sched_setnuma(p, nid);
2520 /* No better CPU than the current one was found. */
2521 if (env.best_cpu == -1) {
2522 trace_sched_stick_numa(p, env.src_cpu, NULL, -1);
2526 best_rq = cpu_rq(env.best_cpu);
2527 if (env.best_task == NULL) {
2528 ret = migrate_task_to(p, env.best_cpu);
2529 WRITE_ONCE(best_rq->numa_migrate_on, 0);
2531 trace_sched_stick_numa(p, env.src_cpu, NULL, env.best_cpu);
2535 ret = migrate_swap(p, env.best_task, env.best_cpu, env.src_cpu);
2536 WRITE_ONCE(best_rq->numa_migrate_on, 0);
2539 trace_sched_stick_numa(p, env.src_cpu, env.best_task, env.best_cpu);
2540 put_task_struct(env.best_task);
2544 /* Attempt to migrate a task to a CPU on the preferred node. */
2545 static void numa_migrate_preferred(struct task_struct *p)
2547 unsigned long interval = HZ;
2549 /* This task has no NUMA fault statistics yet */
2550 if (unlikely(p->numa_preferred_nid == NUMA_NO_NODE || !p->numa_faults))
2553 /* Periodically retry migrating the task to the preferred node */
2554 interval = min(interval, msecs_to_jiffies(p->numa_scan_period) / 16);
2555 p->numa_migrate_retry = jiffies + interval;
2557 /* Success if task is already running on preferred CPU */
2558 if (task_node(p) == p->numa_preferred_nid)
2561 /* Otherwise, try migrate to a CPU on the preferred node */
2562 task_numa_migrate(p);
2566 * Find out how many nodes the workload is actively running on. Do this by
2567 * tracking the nodes from which NUMA hinting faults are triggered. This can
2568 * be different from the set of nodes where the workload's memory is currently
2571 static void numa_group_count_active_nodes(struct numa_group *numa_group)
2573 unsigned long faults, max_faults = 0;
2574 int nid, active_nodes = 0;
2576 for_each_node_state(nid, N_CPU) {
2577 faults = group_faults_cpu(numa_group, nid);
2578 if (faults > max_faults)
2579 max_faults = faults;
2582 for_each_node_state(nid, N_CPU) {
2583 faults = group_faults_cpu(numa_group, nid);
2584 if (faults * ACTIVE_NODE_FRACTION > max_faults)
2588 numa_group->max_faults_cpu = max_faults;
2589 numa_group->active_nodes = active_nodes;
2593 * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
2594 * increments. The more local the fault statistics are, the higher the scan
2595 * period will be for the next scan window. If local/(local+remote) ratio is
2596 * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS)
2597 * the scan period will decrease. Aim for 70% local accesses.
2599 #define NUMA_PERIOD_SLOTS 10
2600 #define NUMA_PERIOD_THRESHOLD 7
2603 * Increase the scan period (slow down scanning) if the majority of
2604 * our memory is already on our local node, or if the majority of
2605 * the page accesses are shared with other processes.
2606 * Otherwise, decrease the scan period.
2608 static void update_task_scan_period(struct task_struct *p,
2609 unsigned long shared, unsigned long private)
2611 unsigned int period_slot;
2612 int lr_ratio, ps_ratio;
2615 unsigned long remote = p->numa_faults_locality[0];
2616 unsigned long local = p->numa_faults_locality[1];
2619 * If there were no record hinting faults then either the task is
2620 * completely idle or all activity is in areas that are not of interest
2621 * to automatic numa balancing. Related to that, if there were failed
2622 * migration then it implies we are migrating too quickly or the local
2623 * node is overloaded. In either case, scan slower
2625 if (local + shared == 0 || p->numa_faults_locality[2]) {
2626 p->numa_scan_period = min(p->numa_scan_period_max,
2627 p->numa_scan_period << 1);
2629 p->mm->numa_next_scan = jiffies +
2630 msecs_to_jiffies(p->numa_scan_period);
2636 * Prepare to scale scan period relative to the current period.
2637 * == NUMA_PERIOD_THRESHOLD scan period stays the same
2638 * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
2639 * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
2641 period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS);
2642 lr_ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote);
2643 ps_ratio = (private * NUMA_PERIOD_SLOTS) / (private + shared);
2645 if (ps_ratio >= NUMA_PERIOD_THRESHOLD) {
2647 * Most memory accesses are local. There is no need to
2648 * do fast NUMA scanning, since memory is already local.
2650 int slot = ps_ratio - NUMA_PERIOD_THRESHOLD;
2653 diff = slot * period_slot;
2654 } else if (lr_ratio >= NUMA_PERIOD_THRESHOLD) {
2656 * Most memory accesses are shared with other tasks.
2657 * There is no point in continuing fast NUMA scanning,
2658 * since other tasks may just move the memory elsewhere.
2660 int slot = lr_ratio - NUMA_PERIOD_THRESHOLD;
2663 diff = slot * period_slot;
2666 * Private memory faults exceed (SLOTS-THRESHOLD)/SLOTS,
2667 * yet they are not on the local NUMA node. Speed up
2668 * NUMA scanning to get the memory moved over.
2670 int ratio = max(lr_ratio, ps_ratio);
2671 diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot;
2674 p->numa_scan_period = clamp(p->numa_scan_period + diff,
2675 task_scan_min(p), task_scan_max(p));
2676 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
2680 * Get the fraction of time the task has been running since the last
2681 * NUMA placement cycle. The scheduler keeps similar statistics, but
2682 * decays those on a 32ms period, which is orders of magnitude off
2683 * from the dozens-of-seconds NUMA balancing period. Use the scheduler
2684 * stats only if the task is so new there are no NUMA statistics yet.
2686 static u64 numa_get_avg_runtime(struct task_struct *p, u64 *period)
2688 u64 runtime, delta, now;
2689 /* Use the start of this time slice to avoid calculations. */
2690 now = p->se.exec_start;
2691 runtime = p->se.sum_exec_runtime;
2693 if (p->last_task_numa_placement) {
2694 delta = runtime - p->last_sum_exec_runtime;
2695 *period = now - p->last_task_numa_placement;
2697 /* Avoid time going backwards, prevent potential divide error: */
2698 if (unlikely((s64)*period < 0))
2701 delta = p->se.avg.load_sum;
2702 *period = LOAD_AVG_MAX;
2705 p->last_sum_exec_runtime = runtime;
2706 p->last_task_numa_placement = now;
2712 * Determine the preferred nid for a task in a numa_group. This needs to
2713 * be done in a way that produces consistent results with group_weight,
2714 * otherwise workloads might not converge.
2716 static int preferred_group_nid(struct task_struct *p, int nid)
2721 /* Direct connections between all NUMA nodes. */
2722 if (sched_numa_topology_type == NUMA_DIRECT)
2726 * On a system with glueless mesh NUMA topology, group_weight
2727 * scores nodes according to the number of NUMA hinting faults on
2728 * both the node itself, and on nearby nodes.
2730 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
2731 unsigned long score, max_score = 0;
2732 int node, max_node = nid;
2734 dist = sched_max_numa_distance;
2736 for_each_node_state(node, N_CPU) {
2737 score = group_weight(p, node, dist);
2738 if (score > max_score) {
2747 * Finding the preferred nid in a system with NUMA backplane
2748 * interconnect topology is more involved. The goal is to locate
2749 * tasks from numa_groups near each other in the system, and
2750 * untangle workloads from different sides of the system. This requires
2751 * searching down the hierarchy of node groups, recursively searching
2752 * inside the highest scoring group of nodes. The nodemask tricks
2753 * keep the complexity of the search down.
2755 nodes = node_states[N_CPU];
2756 for (dist = sched_max_numa_distance; dist > LOCAL_DISTANCE; dist--) {
2757 unsigned long max_faults = 0;
2758 nodemask_t max_group = NODE_MASK_NONE;
2761 /* Are there nodes at this distance from each other? */
2762 if (!find_numa_distance(dist))
2765 for_each_node_mask(a, nodes) {
2766 unsigned long faults = 0;
2767 nodemask_t this_group;
2768 nodes_clear(this_group);
2770 /* Sum group's NUMA faults; includes a==b case. */
2771 for_each_node_mask(b, nodes) {
2772 if (node_distance(a, b) < dist) {
2773 faults += group_faults(p, b);
2774 node_set(b, this_group);
2775 node_clear(b, nodes);
2779 /* Remember the top group. */
2780 if (faults > max_faults) {
2781 max_faults = faults;
2782 max_group = this_group;
2784 * subtle: at the smallest distance there is
2785 * just one node left in each "group", the
2786 * winner is the preferred nid.
2791 /* Next round, evaluate the nodes within max_group. */
2799 static void task_numa_placement(struct task_struct *p)
2801 int seq, nid, max_nid = NUMA_NO_NODE;
2802 unsigned long max_faults = 0;
2803 unsigned long fault_types[2] = { 0, 0 };
2804 unsigned long total_faults;
2805 u64 runtime, period;
2806 spinlock_t *group_lock = NULL;
2807 struct numa_group *ng;
2810 * The p->mm->numa_scan_seq field gets updated without
2811 * exclusive access. Use READ_ONCE() here to ensure
2812 * that the field is read in a single access:
2814 seq = READ_ONCE(p->mm->numa_scan_seq);
2815 if (p->numa_scan_seq == seq)
2817 p->numa_scan_seq = seq;
2818 p->numa_scan_period_max = task_scan_max(p);
2820 total_faults = p->numa_faults_locality[0] +
2821 p->numa_faults_locality[1];
2822 runtime = numa_get_avg_runtime(p, &period);
2824 /* If the task is part of a group prevent parallel updates to group stats */
2825 ng = deref_curr_numa_group(p);
2827 group_lock = &ng->lock;
2828 spin_lock_irq(group_lock);
2831 /* Find the node with the highest number of faults */
2832 for_each_online_node(nid) {
2833 /* Keep track of the offsets in numa_faults array */
2834 int mem_idx, membuf_idx, cpu_idx, cpubuf_idx;
2835 unsigned long faults = 0, group_faults = 0;
2838 for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) {
2839 long diff, f_diff, f_weight;
2841 mem_idx = task_faults_idx(NUMA_MEM, nid, priv);
2842 membuf_idx = task_faults_idx(NUMA_MEMBUF, nid, priv);
2843 cpu_idx = task_faults_idx(NUMA_CPU, nid, priv);
2844 cpubuf_idx = task_faults_idx(NUMA_CPUBUF, nid, priv);
2846 /* Decay existing window, copy faults since last scan */
2847 diff = p->numa_faults[membuf_idx] - p->numa_faults[mem_idx] / 2;
2848 fault_types[priv] += p->numa_faults[membuf_idx];
2849 p->numa_faults[membuf_idx] = 0;
2852 * Normalize the faults_from, so all tasks in a group
2853 * count according to CPU use, instead of by the raw
2854 * number of faults. Tasks with little runtime have
2855 * little over-all impact on throughput, and thus their
2856 * faults are less important.
2858 f_weight = div64_u64(runtime << 16, period + 1);
2859 f_weight = (f_weight * p->numa_faults[cpubuf_idx]) /
2861 f_diff = f_weight - p->numa_faults[cpu_idx] / 2;
2862 p->numa_faults[cpubuf_idx] = 0;
2864 p->numa_faults[mem_idx] += diff;
2865 p->numa_faults[cpu_idx] += f_diff;
2866 faults += p->numa_faults[mem_idx];
2867 p->total_numa_faults += diff;
2870 * safe because we can only change our own group
2872 * mem_idx represents the offset for a given
2873 * nid and priv in a specific region because it
2874 * is at the beginning of the numa_faults array.
2876 ng->faults[mem_idx] += diff;
2877 ng->faults[cpu_idx] += f_diff;
2878 ng->total_faults += diff;
2879 group_faults += ng->faults[mem_idx];
2884 if (faults > max_faults) {
2885 max_faults = faults;
2888 } else if (group_faults > max_faults) {
2889 max_faults = group_faults;
2894 /* Cannot migrate task to CPU-less node */
2895 max_nid = numa_nearest_node(max_nid, N_CPU);
2898 numa_group_count_active_nodes(ng);
2899 spin_unlock_irq(group_lock);
2900 max_nid = preferred_group_nid(p, max_nid);
2904 /* Set the new preferred node */
2905 if (max_nid != p->numa_preferred_nid)
2906 sched_setnuma(p, max_nid);
2909 update_task_scan_period(p, fault_types[0], fault_types[1]);
2912 static inline int get_numa_group(struct numa_group *grp)
2914 return refcount_inc_not_zero(&grp->refcount);
2917 static inline void put_numa_group(struct numa_group *grp)
2919 if (refcount_dec_and_test(&grp->refcount))
2920 kfree_rcu(grp, rcu);
2923 static void task_numa_group(struct task_struct *p, int cpupid, int flags,
2926 struct numa_group *grp, *my_grp;
2927 struct task_struct *tsk;
2929 int cpu = cpupid_to_cpu(cpupid);
2932 if (unlikely(!deref_curr_numa_group(p))) {
2933 unsigned int size = sizeof(struct numa_group) +
2934 NR_NUMA_HINT_FAULT_STATS *
2935 nr_node_ids * sizeof(unsigned long);
2937 grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
2941 refcount_set(&grp->refcount, 1);
2942 grp->active_nodes = 1;
2943 grp->max_faults_cpu = 0;
2944 spin_lock_init(&grp->lock);
2947 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2948 grp->faults[i] = p->numa_faults[i];
2950 grp->total_faults = p->total_numa_faults;
2953 rcu_assign_pointer(p->numa_group, grp);
2957 tsk = READ_ONCE(cpu_rq(cpu)->curr);
2959 if (!cpupid_match_pid(tsk, cpupid))
2962 grp = rcu_dereference(tsk->numa_group);
2966 my_grp = deref_curr_numa_group(p);
2971 * Only join the other group if its bigger; if we're the bigger group,
2972 * the other task will join us.
2974 if (my_grp->nr_tasks > grp->nr_tasks)
2978 * Tie-break on the grp address.
2980 if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp)
2983 /* Always join threads in the same process. */
2984 if (tsk->mm == current->mm)
2987 /* Simple filter to avoid false positives due to PID collisions */
2988 if (flags & TNF_SHARED)
2991 /* Update priv based on whether false sharing was detected */
2994 if (join && !get_numa_group(grp))
3002 WARN_ON_ONCE(irqs_disabled());
3003 double_lock_irq(&my_grp->lock, &grp->lock);
3005 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) {
3006 my_grp->faults[i] -= p->numa_faults[i];
3007 grp->faults[i] += p->numa_faults[i];
3009 my_grp->total_faults -= p->total_numa_faults;
3010 grp->total_faults += p->total_numa_faults;
3015 spin_unlock(&my_grp->lock);
3016 spin_unlock_irq(&grp->lock);
3018 rcu_assign_pointer(p->numa_group, grp);
3020 put_numa_group(my_grp);
3029 * Get rid of NUMA statistics associated with a task (either current or dead).
3030 * If @final is set, the task is dead and has reached refcount zero, so we can
3031 * safely free all relevant data structures. Otherwise, there might be
3032 * concurrent reads from places like load balancing and procfs, and we should
3033 * reset the data back to default state without freeing ->numa_faults.
3035 void task_numa_free(struct task_struct *p, bool final)
3037 /* safe: p either is current or is being freed by current */
3038 struct numa_group *grp = rcu_dereference_raw(p->numa_group);
3039 unsigned long *numa_faults = p->numa_faults;
3040 unsigned long flags;
3047 spin_lock_irqsave(&grp->lock, flags);
3048 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
3049 grp->faults[i] -= p->numa_faults[i];
3050 grp->total_faults -= p->total_numa_faults;
3053 spin_unlock_irqrestore(&grp->lock, flags);
3054 RCU_INIT_POINTER(p->numa_group, NULL);
3055 put_numa_group(grp);
3059 p->numa_faults = NULL;
3062 p->total_numa_faults = 0;
3063 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
3069 * Got a PROT_NONE fault for a page on @node.
3071 void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags)
3073 struct task_struct *p = current;
3074 bool migrated = flags & TNF_MIGRATED;
3075 int cpu_node = task_node(current);
3076 int local = !!(flags & TNF_FAULT_LOCAL);
3077 struct numa_group *ng;
3080 if (!static_branch_likely(&sched_numa_balancing))
3083 /* for example, ksmd faulting in a user's mm */
3088 * NUMA faults statistics are unnecessary for the slow memory
3089 * node for memory tiering mode.
3091 if (!node_is_toptier(mem_node) &&
3092 (sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING ||
3093 !cpupid_valid(last_cpupid)))
3096 /* Allocate buffer to track faults on a per-node basis */
3097 if (unlikely(!p->numa_faults)) {
3098 int size = sizeof(*p->numa_faults) *
3099 NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids;
3101 p->numa_faults = kzalloc(size, GFP_KERNEL|__GFP_NOWARN);
3102 if (!p->numa_faults)
3105 p->total_numa_faults = 0;
3106 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
3110 * First accesses are treated as private, otherwise consider accesses
3111 * to be private if the accessing pid has not changed
3113 if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) {
3116 priv = cpupid_match_pid(p, last_cpupid);
3117 if (!priv && !(flags & TNF_NO_GROUP))
3118 task_numa_group(p, last_cpupid, flags, &priv);
3122 * If a workload spans multiple NUMA nodes, a shared fault that
3123 * occurs wholly within the set of nodes that the workload is
3124 * actively using should be counted as local. This allows the
3125 * scan rate to slow down when a workload has settled down.
3127 ng = deref_curr_numa_group(p);
3128 if (!priv && !local && ng && ng->active_nodes > 1 &&
3129 numa_is_active_node(cpu_node, ng) &&
3130 numa_is_active_node(mem_node, ng))
3134 * Retry to migrate task to preferred node periodically, in case it
3135 * previously failed, or the scheduler moved us.
3137 if (time_after(jiffies, p->numa_migrate_retry)) {
3138 task_numa_placement(p);
3139 numa_migrate_preferred(p);
3143 p->numa_pages_migrated += pages;
3144 if (flags & TNF_MIGRATE_FAIL)
3145 p->numa_faults_locality[2] += pages;
3147 p->numa_faults[task_faults_idx(NUMA_MEMBUF, mem_node, priv)] += pages;
3148 p->numa_faults[task_faults_idx(NUMA_CPUBUF, cpu_node, priv)] += pages;
3149 p->numa_faults_locality[local] += pages;
3152 static void reset_ptenuma_scan(struct task_struct *p)
3155 * We only did a read acquisition of the mmap sem, so
3156 * p->mm->numa_scan_seq is written to without exclusive access
3157 * and the update is not guaranteed to be atomic. That's not
3158 * much of an issue though, since this is just used for
3159 * statistical sampling. Use READ_ONCE/WRITE_ONCE, which are not
3160 * expensive, to avoid any form of compiler optimizations:
3162 WRITE_ONCE(p->mm->numa_scan_seq, READ_ONCE(p->mm->numa_scan_seq) + 1);
3163 p->mm->numa_scan_offset = 0;
3166 static bool vma_is_accessed(struct mm_struct *mm, struct vm_area_struct *vma)
3170 * Allow unconditional access first two times, so that all the (pages)
3171 * of VMAs get prot_none fault introduced irrespective of accesses.
3172 * This is also done to avoid any side effect of task scanning
3173 * amplifying the unfairness of disjoint set of VMAs' access.
3175 if ((READ_ONCE(current->mm->numa_scan_seq) - vma->numab_state->start_scan_seq) < 2)
3178 pids = vma->numab_state->pids_active[0] | vma->numab_state->pids_active[1];
3179 if (test_bit(hash_32(current->pid, ilog2(BITS_PER_LONG)), &pids))
3183 * Complete a scan that has already started regardless of PID access, or
3184 * some VMAs may never be scanned in multi-threaded applications:
3186 if (mm->numa_scan_offset > vma->vm_start) {
3187 trace_sched_skip_vma_numa(mm, vma, NUMAB_SKIP_IGNORE_PID);
3194 #define VMA_PID_RESET_PERIOD (4 * sysctl_numa_balancing_scan_delay)
3197 * The expensive part of numa migration is done from task_work context.
3198 * Triggered from task_tick_numa().
3200 static void task_numa_work(struct callback_head *work)
3202 unsigned long migrate, next_scan, now = jiffies;
3203 struct task_struct *p = current;
3204 struct mm_struct *mm = p->mm;
3205 u64 runtime = p->se.sum_exec_runtime;
3206 struct vm_area_struct *vma;
3207 unsigned long start, end;
3208 unsigned long nr_pte_updates = 0;
3209 long pages, virtpages;
3210 struct vma_iterator vmi;
3211 bool vma_pids_skipped;
3212 bool vma_pids_forced = false;
3214 SCHED_WARN_ON(p != container_of(work, struct task_struct, numa_work));
3218 * Who cares about NUMA placement when they're dying.
3220 * NOTE: make sure not to dereference p->mm before this check,
3221 * exit_task_work() happens _after_ exit_mm() so we could be called
3222 * without p->mm even though we still had it when we enqueued this
3225 if (p->flags & PF_EXITING)
3228 if (!mm->numa_next_scan) {
3229 mm->numa_next_scan = now +
3230 msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
3234 * Enforce maximal scan/migration frequency..
3236 migrate = mm->numa_next_scan;
3237 if (time_before(now, migrate))
3240 if (p->numa_scan_period == 0) {
3241 p->numa_scan_period_max = task_scan_max(p);
3242 p->numa_scan_period = task_scan_start(p);
3245 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
3246 if (!try_cmpxchg(&mm->numa_next_scan, &migrate, next_scan))
3250 * Delay this task enough that another task of this mm will likely win
3251 * the next time around.
3253 p->node_stamp += 2 * TICK_NSEC;
3255 pages = sysctl_numa_balancing_scan_size;
3256 pages <<= 20 - PAGE_SHIFT; /* MB in pages */
3257 virtpages = pages * 8; /* Scan up to this much virtual space */
3262 if (!mmap_read_trylock(mm))
3266 * VMAs are skipped if the current PID has not trapped a fault within
3267 * the VMA recently. Allow scanning to be forced if there is no
3268 * suitable VMA remaining.
3270 vma_pids_skipped = false;
3273 start = mm->numa_scan_offset;
3274 vma_iter_init(&vmi, mm, start);
3275 vma = vma_next(&vmi);
3277 reset_ptenuma_scan(p);
3279 vma_iter_set(&vmi, start);
3280 vma = vma_next(&vmi);
3284 if (!vma_migratable(vma) || !vma_policy_mof(vma) ||
3285 is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_MIXEDMAP)) {
3286 trace_sched_skip_vma_numa(mm, vma, NUMAB_SKIP_UNSUITABLE);
3291 * Shared library pages mapped by multiple processes are not
3292 * migrated as it is expected they are cache replicated. Avoid
3293 * hinting faults in read-only file-backed mappings or the vDSO
3294 * as migrating the pages will be of marginal benefit.
3297 (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ))) {
3298 trace_sched_skip_vma_numa(mm, vma, NUMAB_SKIP_SHARED_RO);
3303 * Skip inaccessible VMAs to avoid any confusion between
3304 * PROT_NONE and NUMA hinting PTEs
3306 if (!vma_is_accessible(vma)) {
3307 trace_sched_skip_vma_numa(mm, vma, NUMAB_SKIP_INACCESSIBLE);
3311 /* Initialise new per-VMA NUMAB state. */
3312 if (!vma->numab_state) {
3313 vma->numab_state = kzalloc(sizeof(struct vma_numab_state),
3315 if (!vma->numab_state)
3318 vma->numab_state->start_scan_seq = mm->numa_scan_seq;
3320 vma->numab_state->next_scan = now +
3321 msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
3323 /* Reset happens after 4 times scan delay of scan start */
3324 vma->numab_state->pids_active_reset = vma->numab_state->next_scan +
3325 msecs_to_jiffies(VMA_PID_RESET_PERIOD);
3328 * Ensure prev_scan_seq does not match numa_scan_seq,
3329 * to prevent VMAs being skipped prematurely on the
3332 vma->numab_state->prev_scan_seq = mm->numa_scan_seq - 1;
3336 * Scanning the VMAs of short lived tasks add more overhead. So
3337 * delay the scan for new VMAs.
3339 if (mm->numa_scan_seq && time_before(jiffies,
3340 vma->numab_state->next_scan)) {
3341 trace_sched_skip_vma_numa(mm, vma, NUMAB_SKIP_SCAN_DELAY);
3345 /* RESET access PIDs regularly for old VMAs. */
3346 if (mm->numa_scan_seq &&
3347 time_after(jiffies, vma->numab_state->pids_active_reset)) {
3348 vma->numab_state->pids_active_reset = vma->numab_state->pids_active_reset +
3349 msecs_to_jiffies(VMA_PID_RESET_PERIOD);
3350 vma->numab_state->pids_active[0] = READ_ONCE(vma->numab_state->pids_active[1]);
3351 vma->numab_state->pids_active[1] = 0;
3354 /* Do not rescan VMAs twice within the same sequence. */
3355 if (vma->numab_state->prev_scan_seq == mm->numa_scan_seq) {
3356 mm->numa_scan_offset = vma->vm_end;
3357 trace_sched_skip_vma_numa(mm, vma, NUMAB_SKIP_SEQ_COMPLETED);
3362 * Do not scan the VMA if task has not accessed it, unless no other
3363 * VMA candidate exists.
3365 if (!vma_pids_forced && !vma_is_accessed(mm, vma)) {
3366 vma_pids_skipped = true;
3367 trace_sched_skip_vma_numa(mm, vma, NUMAB_SKIP_PID_INACTIVE);
3372 start = max(start, vma->vm_start);
3373 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
3374 end = min(end, vma->vm_end);
3375 nr_pte_updates = change_prot_numa(vma, start, end);
3378 * Try to scan sysctl_numa_balancing_size worth of
3379 * hpages that have at least one present PTE that
3380 * is not already PTE-numa. If the VMA contains
3381 * areas that are unused or already full of prot_numa
3382 * PTEs, scan up to virtpages, to skip through those
3386 pages -= (end - start) >> PAGE_SHIFT;
3387 virtpages -= (end - start) >> PAGE_SHIFT;
3390 if (pages <= 0 || virtpages <= 0)
3394 } while (end != vma->vm_end);
3396 /* VMA scan is complete, do not scan until next sequence. */
3397 vma->numab_state->prev_scan_seq = mm->numa_scan_seq;
3400 * Only force scan within one VMA at a time, to limit the
3401 * cost of scanning a potentially uninteresting VMA.
3403 if (vma_pids_forced)
3405 } for_each_vma(vmi, vma);
3408 * If no VMAs are remaining and VMAs were skipped due to the PID
3409 * not accessing the VMA previously, then force a scan to ensure
3412 if (!vma && !vma_pids_forced && vma_pids_skipped) {
3413 vma_pids_forced = true;
3419 * It is possible to reach the end of the VMA list but the last few
3420 * VMAs are not guaranteed to the vma_migratable. If they are not, we
3421 * would find the !migratable VMA on the next scan but not reset the
3422 * scanner to the start so check it now.
3425 mm->numa_scan_offset = start;
3427 reset_ptenuma_scan(p);
3428 mmap_read_unlock(mm);
3431 * Make sure tasks use at least 32x as much time to run other code
3432 * than they used here, to limit NUMA PTE scanning overhead to 3% max.
3433 * Usually update_task_scan_period slows down scanning enough; on an
3434 * overloaded system we need to limit overhead on a per task basis.
3436 if (unlikely(p->se.sum_exec_runtime != runtime)) {
3437 u64 diff = p->se.sum_exec_runtime - runtime;
3438 p->node_stamp += 32 * diff;
3442 void init_numa_balancing(unsigned long clone_flags, struct task_struct *p)
3445 struct mm_struct *mm = p->mm;
3448 mm_users = atomic_read(&mm->mm_users);
3449 if (mm_users == 1) {
3450 mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
3451 mm->numa_scan_seq = 0;
3455 p->numa_scan_seq = mm ? mm->numa_scan_seq : 0;
3456 p->numa_scan_period = sysctl_numa_balancing_scan_delay;
3457 p->numa_migrate_retry = 0;
3458 /* Protect against double add, see task_tick_numa and task_numa_work */
3459 p->numa_work.next = &p->numa_work;
3460 p->numa_faults = NULL;
3461 p->numa_pages_migrated = 0;
3462 p->total_numa_faults = 0;
3463 RCU_INIT_POINTER(p->numa_group, NULL);
3464 p->last_task_numa_placement = 0;
3465 p->last_sum_exec_runtime = 0;
3467 init_task_work(&p->numa_work, task_numa_work);
3469 /* New address space, reset the preferred nid */
3470 if (!(clone_flags & CLONE_VM)) {
3471 p->numa_preferred_nid = NUMA_NO_NODE;
3476 * New thread, keep existing numa_preferred_nid which should be copied
3477 * already by arch_dup_task_struct but stagger when scans start.
3482 delay = min_t(unsigned int, task_scan_max(current),
3483 current->numa_scan_period * mm_users * NSEC_PER_MSEC);
3484 delay += 2 * TICK_NSEC;
3485 p->node_stamp = delay;
3490 * Drive the periodic memory faults..
3492 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
3494 struct callback_head *work = &curr->numa_work;
3498 * We don't care about NUMA placement if we don't have memory.
3500 if (!curr->mm || (curr->flags & (PF_EXITING | PF_KTHREAD)) || work->next != work)
3504 * Using runtime rather than walltime has the dual advantage that
3505 * we (mostly) drive the selection from busy threads and that the
3506 * task needs to have done some actual work before we bother with
3509 now = curr->se.sum_exec_runtime;
3510 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
3512 if (now > curr->node_stamp + period) {
3513 if (!curr->node_stamp)
3514 curr->numa_scan_period = task_scan_start(curr);
3515 curr->node_stamp += period;
3517 if (!time_before(jiffies, curr->mm->numa_next_scan))
3518 task_work_add(curr, work, TWA_RESUME);
3522 static void update_scan_period(struct task_struct *p, int new_cpu)
3524 int src_nid = cpu_to_node(task_cpu(p));
3525 int dst_nid = cpu_to_node(new_cpu);
3527 if (!static_branch_likely(&sched_numa_balancing))
3530 if (!p->mm || !p->numa_faults || (p->flags & PF_EXITING))
3533 if (src_nid == dst_nid)
3537 * Allow resets if faults have been trapped before one scan
3538 * has completed. This is most likely due to a new task that
3539 * is pulled cross-node due to wakeups or load balancing.
3541 if (p->numa_scan_seq) {
3543 * Avoid scan adjustments if moving to the preferred
3544 * node or if the task was not previously running on
3545 * the preferred node.
3547 if (dst_nid == p->numa_preferred_nid ||
3548 (p->numa_preferred_nid != NUMA_NO_NODE &&
3549 src_nid != p->numa_preferred_nid))
3553 p->numa_scan_period = task_scan_start(p);
3557 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
3561 static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p)
3565 static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p)
3569 static inline void update_scan_period(struct task_struct *p, int new_cpu)
3573 #endif /* CONFIG_NUMA_BALANCING */
3576 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
3578 update_load_add(&cfs_rq->load, se->load.weight);
3580 if (entity_is_task(se)) {
3581 struct rq *rq = rq_of(cfs_rq);
3583 account_numa_enqueue(rq, task_of(se));
3584 list_add(&se->group_node, &rq->cfs_tasks);
3587 cfs_rq->nr_running++;
3589 cfs_rq->idle_nr_running++;
3593 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
3595 update_load_sub(&cfs_rq->load, se->load.weight);
3597 if (entity_is_task(se)) {
3598 account_numa_dequeue(rq_of(cfs_rq), task_of(se));
3599 list_del_init(&se->group_node);
3602 cfs_rq->nr_running--;
3604 cfs_rq->idle_nr_running--;
3608 * Signed add and clamp on underflow.
3610 * Explicitly do a load-store to ensure the intermediate value never hits
3611 * memory. This allows lockless observations without ever seeing the negative
3614 #define add_positive(_ptr, _val) do { \
3615 typeof(_ptr) ptr = (_ptr); \
3616 typeof(_val) val = (_val); \
3617 typeof(*ptr) res, var = READ_ONCE(*ptr); \
3621 if (val < 0 && res > var) \
3624 WRITE_ONCE(*ptr, res); \
3628 * Unsigned subtract and clamp on underflow.
3630 * Explicitly do a load-store to ensure the intermediate value never hits
3631 * memory. This allows lockless observations without ever seeing the negative
3634 #define sub_positive(_ptr, _val) do { \
3635 typeof(_ptr) ptr = (_ptr); \
3636 typeof(*ptr) val = (_val); \
3637 typeof(*ptr) res, var = READ_ONCE(*ptr); \
3641 WRITE_ONCE(*ptr, res); \
3645 * Remove and clamp on negative, from a local variable.
3647 * A variant of sub_positive(), which does not use explicit load-store
3648 * and is thus optimized for local variable updates.
3650 #define lsub_positive(_ptr, _val) do { \
3651 typeof(_ptr) ptr = (_ptr); \
3652 *ptr -= min_t(typeof(*ptr), *ptr, _val); \
3657 enqueue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3659 cfs_rq->avg.load_avg += se->avg.load_avg;
3660 cfs_rq->avg.load_sum += se_weight(se) * se->avg.load_sum;
3664 dequeue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3666 sub_positive(&cfs_rq->avg.load_avg, se->avg.load_avg);
3667 sub_positive(&cfs_rq->avg.load_sum, se_weight(se) * se->avg.load_sum);
3668 /* See update_cfs_rq_load_avg() */
3669 cfs_rq->avg.load_sum = max_t(u32, cfs_rq->avg.load_sum,
3670 cfs_rq->avg.load_avg * PELT_MIN_DIVIDER);
3674 enqueue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
3676 dequeue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
3679 static void reweight_eevdf(struct sched_entity *se, u64 avruntime,
3680 unsigned long weight)
3682 unsigned long old_weight = se->load.weight;
3689 * COROLLARY #1: The virtual runtime of the entity needs to be
3690 * adjusted if re-weight at !0-lag point.
3692 * Proof: For contradiction assume this is not true, so we can
3693 * re-weight without changing vruntime at !0-lag point.
3695 * Weight VRuntime Avg-VRuntime
3699 * Since lag needs to be preserved through re-weight:
3701 * lag = (V - v)*w = (V'- v')*w', where v = v'
3702 * ==> V' = (V - v)*w/w' + v (1)
3704 * Let W be the total weight of the entities before reweight,
3705 * since V' is the new weighted average of entities:
3707 * V' = (WV + w'v - wv) / (W + w' - w) (2)
3709 * by using (1) & (2) we obtain:
3711 * (WV + w'v - wv) / (W + w' - w) = (V - v)*w/w' + v
3712 * ==> (WV-Wv+Wv+w'v-wv)/(W+w'-w) = (V - v)*w/w' + v
3713 * ==> (WV - Wv)/(W + w' - w) + v = (V - v)*w/w' + v
3714 * ==> (V - v)*W/(W + w' - w) = (V - v)*w/w' (3)
3716 * Since we are doing at !0-lag point which means V != v, we
3719 * ==> W / (W + w' - w) = w / w'
3720 * ==> Ww' = Ww + ww' - ww
3721 * ==> W * (w' - w) = w * (w' - w)
3722 * ==> W = w (re-weight indicates w' != w)
3724 * So the cfs_rq contains only one entity, hence vruntime of
3725 * the entity @v should always equal to the cfs_rq's weighted
3726 * average vruntime @V, which means we will always re-weight
3727 * at 0-lag point, thus breach assumption. Proof completed.
3730 * COROLLARY #2: Re-weight does NOT affect weighted average
3731 * vruntime of all the entities.
3733 * Proof: According to corollary #1, Eq. (1) should be:
3735 * (V - v)*w = (V' - v')*w'
3736 * ==> v' = V' - (V - v)*w/w' (4)
3738 * According to the weighted average formula, we have:
3740 * V' = (WV - wv + w'v') / (W - w + w')
3741 * = (WV - wv + w'(V' - (V - v)w/w')) / (W - w + w')
3742 * = (WV - wv + w'V' - Vw + wv) / (W - w + w')
3743 * = (WV + w'V' - Vw) / (W - w + w')
3745 * ==> V'*(W - w + w') = WV + w'V' - Vw
3746 * ==> V' * (W - w) = (W - w) * V (5)
3748 * If the entity is the only one in the cfs_rq, then reweight
3749 * always occurs at 0-lag point, so V won't change. Or else
3750 * there are other entities, hence W != w, then Eq. (5) turns
3751 * into V' = V. So V won't change in either case, proof done.
3754 * So according to corollary #1 & #2, the effect of re-weight
3755 * on vruntime should be:
3757 * v' = V' - (V - v) * w / w' (4)
3758 * = V - (V - v) * w / w'
3762 if (avruntime != se->vruntime) {
3763 vlag = entity_lag(avruntime, se);
3764 vlag = div_s64(vlag * old_weight, weight);
3765 se->vruntime = avruntime - vlag;
3772 * When the weight changes, the virtual time slope changes and
3773 * we should adjust the relative virtual deadline accordingly.
3775 * d' = v' + (d - v)*w/w'
3776 * = V' - (V - v)*w/w' + (d - v)*w/w'
3777 * = V - (V - v)*w/w' + (d - v)*w/w'
3778 * = V + (d - V)*w/w'
3780 vslice = (s64)(se->deadline - avruntime);
3781 vslice = div_s64(vslice * old_weight, weight);
3782 se->deadline = avruntime + vslice;
3785 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
3786 unsigned long weight)
3788 bool curr = cfs_rq->curr == se;
3792 /* commit outstanding execution time */
3793 update_curr(cfs_rq);
3794 avruntime = avg_vruntime(cfs_rq);
3796 __dequeue_entity(cfs_rq, se);
3797 update_load_sub(&cfs_rq->load, se->load.weight);
3799 dequeue_load_avg(cfs_rq, se);
3802 reweight_eevdf(se, avruntime, weight);
3805 * Because we keep se->vlag = V - v_i, while: lag_i = w_i*(V - v_i),
3806 * we need to scale se->vlag when w_i changes.
3808 se->vlag = div_s64(se->vlag * se->load.weight, weight);
3811 update_load_set(&se->load, weight);
3815 u32 divider = get_pelt_divider(&se->avg);
3817 se->avg.load_avg = div_u64(se_weight(se) * se->avg.load_sum, divider);
3821 enqueue_load_avg(cfs_rq, se);
3823 update_load_add(&cfs_rq->load, se->load.weight);
3825 __enqueue_entity(cfs_rq, se);
3828 * The entity's vruntime has been adjusted, so let's check
3829 * whether the rq-wide min_vruntime needs updated too. Since
3830 * the calculations above require stable min_vruntime rather
3831 * than up-to-date one, we do the update at the end of the
3834 update_min_vruntime(cfs_rq);
3838 void reweight_task(struct task_struct *p, int prio)
3840 struct sched_entity *se = &p->se;
3841 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3842 struct load_weight *load = &se->load;
3843 unsigned long weight = scale_load(sched_prio_to_weight[prio]);
3845 reweight_entity(cfs_rq, se, weight);
3846 load->inv_weight = sched_prio_to_wmult[prio];
3849 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
3851 #ifdef CONFIG_FAIR_GROUP_SCHED
3854 * All this does is approximate the hierarchical proportion which includes that
3855 * global sum we all love to hate.
3857 * That is, the weight of a group entity, is the proportional share of the
3858 * group weight based on the group runqueue weights. That is:
3860 * tg->weight * grq->load.weight
3861 * ge->load.weight = ----------------------------- (1)
3862 * \Sum grq->load.weight
3864 * Now, because computing that sum is prohibitively expensive to compute (been
3865 * there, done that) we approximate it with this average stuff. The average
3866 * moves slower and therefore the approximation is cheaper and more stable.
3868 * So instead of the above, we substitute:
3870 * grq->load.weight -> grq->avg.load_avg (2)
3872 * which yields the following:
3874 * tg->weight * grq->avg.load_avg
3875 * ge->load.weight = ------------------------------ (3)
3878 * Where: tg->load_avg ~= \Sum grq->avg.load_avg
3880 * That is shares_avg, and it is right (given the approximation (2)).
3882 * The problem with it is that because the average is slow -- it was designed
3883 * to be exactly that of course -- this leads to transients in boundary
3884 * conditions. In specific, the case where the group was idle and we start the
3885 * one task. It takes time for our CPU's grq->avg.load_avg to build up,
3886 * yielding bad latency etc..
3888 * Now, in that special case (1) reduces to:
3890 * tg->weight * grq->load.weight
3891 * ge->load.weight = ----------------------------- = tg->weight (4)
3894 * That is, the sum collapses because all other CPUs are idle; the UP scenario.
3896 * So what we do is modify our approximation (3) to approach (4) in the (near)
3901 * tg->weight * grq->load.weight
3902 * --------------------------------------------------- (5)
3903 * tg->load_avg - grq->avg.load_avg + grq->load.weight
3905 * But because grq->load.weight can drop to 0, resulting in a divide by zero,
3906 * we need to use grq->avg.load_avg as its lower bound, which then gives:
3909 * tg->weight * grq->load.weight
3910 * ge->load.weight = ----------------------------- (6)
3915 * tg_load_avg' = tg->load_avg - grq->avg.load_avg +
3916 * max(grq->load.weight, grq->avg.load_avg)
3918 * And that is shares_weight and is icky. In the (near) UP case it approaches
3919 * (4) while in the normal case it approaches (3). It consistently
3920 * overestimates the ge->load.weight and therefore:
3922 * \Sum ge->load.weight >= tg->weight
3926 static long calc_group_shares(struct cfs_rq *cfs_rq)
3928 long tg_weight, tg_shares, load, shares;
3929 struct task_group *tg = cfs_rq->tg;
3931 tg_shares = READ_ONCE(tg->shares);
3933 load = max(scale_load_down(cfs_rq->load.weight), cfs_rq->avg.load_avg);
3935 tg_weight = atomic_long_read(&tg->load_avg);
3937 /* Ensure tg_weight >= load */
3938 tg_weight -= cfs_rq->tg_load_avg_contrib;
3941 shares = (tg_shares * load);
3943 shares /= tg_weight;
3946 * MIN_SHARES has to be unscaled here to support per-CPU partitioning
3947 * of a group with small tg->shares value. It is a floor value which is
3948 * assigned as a minimum load.weight to the sched_entity representing
3949 * the group on a CPU.
3951 * E.g. on 64-bit for a group with tg->shares of scale_load(15)=15*1024
3952 * on an 8-core system with 8 tasks each runnable on one CPU shares has
3953 * to be 15*1024*1/8=1920 instead of scale_load(MIN_SHARES)=2*1024. In
3954 * case no task is runnable on a CPU MIN_SHARES=2 should be returned
3957 return clamp_t(long, shares, MIN_SHARES, tg_shares);
3959 #endif /* CONFIG_SMP */
3962 * Recomputes the group entity based on the current state of its group
3965 static void update_cfs_group(struct sched_entity *se)
3967 struct cfs_rq *gcfs_rq = group_cfs_rq(se);
3973 if (throttled_hierarchy(gcfs_rq))
3977 shares = READ_ONCE(gcfs_rq->tg->shares);
3979 shares = calc_group_shares(gcfs_rq);
3981 if (unlikely(se->load.weight != shares))
3982 reweight_entity(cfs_rq_of(se), se, shares);
3985 #else /* CONFIG_FAIR_GROUP_SCHED */
3986 static inline void update_cfs_group(struct sched_entity *se)
3989 #endif /* CONFIG_FAIR_GROUP_SCHED */
3991 static inline void cfs_rq_util_change(struct cfs_rq *cfs_rq, int flags)
3993 struct rq *rq = rq_of(cfs_rq);
3995 if (&rq->cfs == cfs_rq) {
3997 * There are a few boundary cases this might miss but it should
3998 * get called often enough that that should (hopefully) not be
4001 * It will not get called when we go idle, because the idle
4002 * thread is a different class (!fair), nor will the utilization
4003 * number include things like RT tasks.
4005 * As is, the util number is not freq-invariant (we'd have to
4006 * implement arch_scale_freq_capacity() for that).
4008 * See cpu_util_cfs().
4010 cpufreq_update_util(rq, flags);
4015 static inline bool load_avg_is_decayed(struct sched_avg *sa)
4023 if (sa->runnable_sum)
4027 * _avg must be null when _sum are null because _avg = _sum / divider
4028 * Make sure that rounding and/or propagation of PELT values never
4031 SCHED_WARN_ON(sa->load_avg ||
4038 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
4040 return u64_u32_load_copy(cfs_rq->avg.last_update_time,
4041 cfs_rq->last_update_time_copy);
4043 #ifdef CONFIG_FAIR_GROUP_SCHED
4045 * Because list_add_leaf_cfs_rq always places a child cfs_rq on the list
4046 * immediately before a parent cfs_rq, and cfs_rqs are removed from the list
4047 * bottom-up, we only have to test whether the cfs_rq before us on the list
4049 * If cfs_rq is not on the list, test whether a child needs its to be added to
4050 * connect a branch to the tree * (see list_add_leaf_cfs_rq() for details).
4052 static inline bool child_cfs_rq_on_list(struct cfs_rq *cfs_rq)
4054 struct cfs_rq *prev_cfs_rq;
4055 struct list_head *prev;
4057 if (cfs_rq->on_list) {
4058 prev = cfs_rq->leaf_cfs_rq_list.prev;
4060 struct rq *rq = rq_of(cfs_rq);
4062 prev = rq->tmp_alone_branch;
4065 prev_cfs_rq = container_of(prev, struct cfs_rq, leaf_cfs_rq_list);
4067 return (prev_cfs_rq->tg->parent == cfs_rq->tg);
4070 static inline bool cfs_rq_is_decayed(struct cfs_rq *cfs_rq)
4072 if (cfs_rq->load.weight)
4075 if (!load_avg_is_decayed(&cfs_rq->avg))
4078 if (child_cfs_rq_on_list(cfs_rq))
4085 * update_tg_load_avg - update the tg's load avg
4086 * @cfs_rq: the cfs_rq whose avg changed
4088 * This function 'ensures': tg->load_avg := \Sum tg->cfs_rq[]->avg.load.
4089 * However, because tg->load_avg is a global value there are performance
4092 * In order to avoid having to look at the other cfs_rq's, we use a
4093 * differential update where we store the last value we propagated. This in
4094 * turn allows skipping updates if the differential is 'small'.
4096 * Updating tg's load_avg is necessary before update_cfs_share().
4098 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq)
4104 * No need to update load_avg for root_task_group as it is not used.
4106 if (cfs_rq->tg == &root_task_group)
4109 /* rq has been offline and doesn't contribute to the share anymore: */
4110 if (!cpu_active(cpu_of(rq_of(cfs_rq))))
4114 * For migration heavy workloads, access to tg->load_avg can be
4115 * unbound. Limit the update rate to at most once per ms.
4117 now = sched_clock_cpu(cpu_of(rq_of(cfs_rq)));
4118 if (now - cfs_rq->last_update_tg_load_avg < NSEC_PER_MSEC)
4121 delta = cfs_rq->avg.load_avg - cfs_rq->tg_load_avg_contrib;
4122 if (abs(delta) > cfs_rq->tg_load_avg_contrib / 64) {
4123 atomic_long_add(delta, &cfs_rq->tg->load_avg);
4124 cfs_rq->tg_load_avg_contrib = cfs_rq->avg.load_avg;
4125 cfs_rq->last_update_tg_load_avg = now;
4129 static inline void clear_tg_load_avg(struct cfs_rq *cfs_rq)
4135 * No need to update load_avg for root_task_group, as it is not used.
4137 if (cfs_rq->tg == &root_task_group)
4140 now = sched_clock_cpu(cpu_of(rq_of(cfs_rq)));
4141 delta = 0 - cfs_rq->tg_load_avg_contrib;
4142 atomic_long_add(delta, &cfs_rq->tg->load_avg);
4143 cfs_rq->tg_load_avg_contrib = 0;
4144 cfs_rq->last_update_tg_load_avg = now;
4147 /* CPU offline callback: */
4148 static void __maybe_unused clear_tg_offline_cfs_rqs(struct rq *rq)
4150 struct task_group *tg;
4152 lockdep_assert_rq_held(rq);
4155 * The rq clock has already been updated in
4156 * set_rq_offline(), so we should skip updating
4157 * the rq clock again in unthrottle_cfs_rq().
4159 rq_clock_start_loop_update(rq);
4162 list_for_each_entry_rcu(tg, &task_groups, list) {
4163 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
4165 clear_tg_load_avg(cfs_rq);
4169 rq_clock_stop_loop_update(rq);
4173 * Called within set_task_rq() right before setting a task's CPU. The
4174 * caller only guarantees p->pi_lock is held; no other assumptions,
4175 * including the state of rq->lock, should be made.
4177 void set_task_rq_fair(struct sched_entity *se,
4178 struct cfs_rq *prev, struct cfs_rq *next)
4180 u64 p_last_update_time;
4181 u64 n_last_update_time;
4183 if (!sched_feat(ATTACH_AGE_LOAD))
4187 * We are supposed to update the task to "current" time, then its up to
4188 * date and ready to go to new CPU/cfs_rq. But we have difficulty in
4189 * getting what current time is, so simply throw away the out-of-date
4190 * time. This will result in the wakee task is less decayed, but giving
4191 * the wakee more load sounds not bad.
4193 if (!(se->avg.last_update_time && prev))
4196 p_last_update_time = cfs_rq_last_update_time(prev);
4197 n_last_update_time = cfs_rq_last_update_time(next);
4199 __update_load_avg_blocked_se(p_last_update_time, se);
4200 se->avg.last_update_time = n_last_update_time;
4204 * When on migration a sched_entity joins/leaves the PELT hierarchy, we need to
4205 * propagate its contribution. The key to this propagation is the invariant
4206 * that for each group:
4208 * ge->avg == grq->avg (1)
4210 * _IFF_ we look at the pure running and runnable sums. Because they
4211 * represent the very same entity, just at different points in the hierarchy.
4213 * Per the above update_tg_cfs_util() and update_tg_cfs_runnable() are trivial
4214 * and simply copies the running/runnable sum over (but still wrong, because
4215 * the group entity and group rq do not have their PELT windows aligned).
4217 * However, update_tg_cfs_load() is more complex. So we have:
4219 * ge->avg.load_avg = ge->load.weight * ge->avg.runnable_avg (2)
4221 * And since, like util, the runnable part should be directly transferable,
4222 * the following would _appear_ to be the straight forward approach:
4224 * grq->avg.load_avg = grq->load.weight * grq->avg.runnable_avg (3)
4226 * And per (1) we have:
4228 * ge->avg.runnable_avg == grq->avg.runnable_avg
4232 * ge->load.weight * grq->avg.load_avg
4233 * ge->avg.load_avg = ----------------------------------- (4)
4236 * Except that is wrong!
4238 * Because while for entities historical weight is not important and we
4239 * really only care about our future and therefore can consider a pure
4240 * runnable sum, runqueues can NOT do this.
4242 * We specifically want runqueues to have a load_avg that includes
4243 * historical weights. Those represent the blocked load, the load we expect
4244 * to (shortly) return to us. This only works by keeping the weights as
4245 * integral part of the sum. We therefore cannot decompose as per (3).
4247 * Another reason this doesn't work is that runnable isn't a 0-sum entity.
4248 * Imagine a rq with 2 tasks that each are runnable 2/3 of the time. Then the
4249 * rq itself is runnable anywhere between 2/3 and 1 depending on how the
4250 * runnable section of these tasks overlap (or not). If they were to perfectly
4251 * align the rq as a whole would be runnable 2/3 of the time. If however we
4252 * always have at least 1 runnable task, the rq as a whole is always runnable.
4254 * So we'll have to approximate.. :/
4256 * Given the constraint:
4258 * ge->avg.running_sum <= ge->avg.runnable_sum <= LOAD_AVG_MAX
4260 * We can construct a rule that adds runnable to a rq by assuming minimal
4263 * On removal, we'll assume each task is equally runnable; which yields:
4265 * grq->avg.runnable_sum = grq->avg.load_sum / grq->load.weight
4267 * XXX: only do this for the part of runnable > running ?
4271 update_tg_cfs_util(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq)
4273 long delta_sum, delta_avg = gcfs_rq->avg.util_avg - se->avg.util_avg;
4274 u32 new_sum, divider;
4276 /* Nothing to update */
4281 * cfs_rq->avg.period_contrib can be used for both cfs_rq and se.
4282 * See ___update_load_avg() for details.
4284 divider = get_pelt_divider(&cfs_rq->avg);
4287 /* Set new sched_entity's utilization */
4288 se->avg.util_avg = gcfs_rq->avg.util_avg;
4289 new_sum = se->avg.util_avg * divider;
4290 delta_sum = (long)new_sum - (long)se->avg.util_sum;
4291 se->avg.util_sum = new_sum;
4293 /* Update parent cfs_rq utilization */
4294 add_positive(&cfs_rq->avg.util_avg, delta_avg);
4295 add_positive(&cfs_rq->avg.util_sum, delta_sum);
4297 /* See update_cfs_rq_load_avg() */
4298 cfs_rq->avg.util_sum = max_t(u32, cfs_rq->avg.util_sum,
4299 cfs_rq->avg.util_avg * PELT_MIN_DIVIDER);
4303 update_tg_cfs_runnable(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq)
4305 long delta_sum, delta_avg = gcfs_rq->avg.runnable_avg - se->avg.runnable_avg;
4306 u32 new_sum, divider;
4308 /* Nothing to update */
4313 * cfs_rq->avg.period_contrib can be used for both cfs_rq and se.
4314 * See ___update_load_avg() for details.
4316 divider = get_pelt_divider(&cfs_rq->avg);
4318 /* Set new sched_entity's runnable */
4319 se->avg.runnable_avg = gcfs_rq->avg.runnable_avg;
4320 new_sum = se->avg.runnable_avg * divider;
4321 delta_sum = (long)new_sum - (long)se->avg.runnable_sum;
4322 se->avg.runnable_sum = new_sum;
4324 /* Update parent cfs_rq runnable */
4325 add_positive(&cfs_rq->avg.runnable_avg, delta_avg);
4326 add_positive(&cfs_rq->avg.runnable_sum, delta_sum);
4327 /* See update_cfs_rq_load_avg() */
4328 cfs_rq->avg.runnable_sum = max_t(u32, cfs_rq->avg.runnable_sum,
4329 cfs_rq->avg.runnable_avg * PELT_MIN_DIVIDER);
4333 update_tg_cfs_load(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq)
4335 long delta_avg, running_sum, runnable_sum = gcfs_rq->prop_runnable_sum;
4336 unsigned long load_avg;
4344 gcfs_rq->prop_runnable_sum = 0;
4347 * cfs_rq->avg.period_contrib can be used for both cfs_rq and se.
4348 * See ___update_load_avg() for details.
4350 divider = get_pelt_divider(&cfs_rq->avg);
4352 if (runnable_sum >= 0) {
4354 * Add runnable; clip at LOAD_AVG_MAX. Reflects that until
4355 * the CPU is saturated running == runnable.
4357 runnable_sum += se->avg.load_sum;
4358 runnable_sum = min_t(long, runnable_sum, divider);
4361 * Estimate the new unweighted runnable_sum of the gcfs_rq by
4362 * assuming all tasks are equally runnable.
4364 if (scale_load_down(gcfs_rq->load.weight)) {
4365 load_sum = div_u64(gcfs_rq->avg.load_sum,
4366 scale_load_down(gcfs_rq->load.weight));
4369 /* But make sure to not inflate se's runnable */
4370 runnable_sum = min(se->avg.load_sum, load_sum);
4374 * runnable_sum can't be lower than running_sum
4375 * Rescale running sum to be in the same range as runnable sum
4376 * running_sum is in [0 : LOAD_AVG_MAX << SCHED_CAPACITY_SHIFT]
4377 * runnable_sum is in [0 : LOAD_AVG_MAX]
4379 running_sum = se->avg.util_sum >> SCHED_CAPACITY_SHIFT;
4380 runnable_sum = max(runnable_sum, running_sum);
4382 load_sum = se_weight(se) * runnable_sum;
4383 load_avg = div_u64(load_sum, divider);
4385 delta_avg = load_avg - se->avg.load_avg;
4389 delta_sum = load_sum - (s64)se_weight(se) * se->avg.load_sum;
4391 se->avg.load_sum = runnable_sum;
4392 se->avg.load_avg = load_avg;
4393 add_positive(&cfs_rq->avg.load_avg, delta_avg);
4394 add_positive(&cfs_rq->avg.load_sum, delta_sum);
4395 /* See update_cfs_rq_load_avg() */
4396 cfs_rq->avg.load_sum = max_t(u32, cfs_rq->avg.load_sum,
4397 cfs_rq->avg.load_avg * PELT_MIN_DIVIDER);
4400 static inline void add_tg_cfs_propagate(struct cfs_rq *cfs_rq, long runnable_sum)
4402 cfs_rq->propagate = 1;
4403 cfs_rq->prop_runnable_sum += runnable_sum;
4406 /* Update task and its cfs_rq load average */
4407 static inline int propagate_entity_load_avg(struct sched_entity *se)
4409 struct cfs_rq *cfs_rq, *gcfs_rq;
4411 if (entity_is_task(se))
4414 gcfs_rq = group_cfs_rq(se);
4415 if (!gcfs_rq->propagate)
4418 gcfs_rq->propagate = 0;
4420 cfs_rq = cfs_rq_of(se);
4422 add_tg_cfs_propagate(cfs_rq, gcfs_rq->prop_runnable_sum);
4424 update_tg_cfs_util(cfs_rq, se, gcfs_rq);
4425 update_tg_cfs_runnable(cfs_rq, se, gcfs_rq);
4426 update_tg_cfs_load(cfs_rq, se, gcfs_rq);
4428 trace_pelt_cfs_tp(cfs_rq);
4429 trace_pelt_se_tp(se);
4435 * Check if we need to update the load and the utilization of a blocked
4438 static inline bool skip_blocked_update(struct sched_entity *se)
4440 struct cfs_rq *gcfs_rq = group_cfs_rq(se);
4443 * If sched_entity still have not zero load or utilization, we have to
4446 if (se->avg.load_avg || se->avg.util_avg)
4450 * If there is a pending propagation, we have to update the load and
4451 * the utilization of the sched_entity:
4453 if (gcfs_rq->propagate)
4457 * Otherwise, the load and the utilization of the sched_entity is
4458 * already zero and there is no pending propagation, so it will be a
4459 * waste of time to try to decay it:
4464 #else /* CONFIG_FAIR_GROUP_SCHED */
4466 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq) {}
4468 static inline void clear_tg_offline_cfs_rqs(struct rq *rq) {}
4470 static inline int propagate_entity_load_avg(struct sched_entity *se)
4475 static inline void add_tg_cfs_propagate(struct cfs_rq *cfs_rq, long runnable_sum) {}
4477 #endif /* CONFIG_FAIR_GROUP_SCHED */
4479 #ifdef CONFIG_NO_HZ_COMMON
4480 static inline void migrate_se_pelt_lag(struct sched_entity *se)
4482 u64 throttled = 0, now, lut;
4483 struct cfs_rq *cfs_rq;
4487 if (load_avg_is_decayed(&se->avg))
4490 cfs_rq = cfs_rq_of(se);
4494 is_idle = is_idle_task(rcu_dereference(rq->curr));
4498 * The lag estimation comes with a cost we don't want to pay all the
4499 * time. Hence, limiting to the case where the source CPU is idle and
4500 * we know we are at the greatest risk to have an outdated clock.
4506 * Estimated "now" is: last_update_time + cfs_idle_lag + rq_idle_lag, where:
4508 * last_update_time (the cfs_rq's last_update_time)
4509 * = cfs_rq_clock_pelt()@cfs_rq_idle
4510 * = rq_clock_pelt()@cfs_rq_idle
4511 * - cfs->throttled_clock_pelt_time@cfs_rq_idle
4513 * cfs_idle_lag (delta between rq's update and cfs_rq's update)
4514 * = rq_clock_pelt()@rq_idle - rq_clock_pelt()@cfs_rq_idle
4516 * rq_idle_lag (delta between now and rq's update)
4517 * = sched_clock_cpu() - rq_clock()@rq_idle
4519 * We can then write:
4521 * now = rq_clock_pelt()@rq_idle - cfs->throttled_clock_pelt_time +
4522 * sched_clock_cpu() - rq_clock()@rq_idle
4524 * rq_clock_pelt()@rq_idle is rq->clock_pelt_idle
4525 * rq_clock()@rq_idle is rq->clock_idle
4526 * cfs->throttled_clock_pelt_time@cfs_rq_idle
4527 * is cfs_rq->throttled_pelt_idle
4530 #ifdef CONFIG_CFS_BANDWIDTH
4531 throttled = u64_u32_load(cfs_rq->throttled_pelt_idle);
4532 /* The clock has been stopped for throttling */
4533 if (throttled == U64_MAX)
4536 now = u64_u32_load(rq->clock_pelt_idle);
4538 * Paired with _update_idle_rq_clock_pelt(). It ensures at the worst case
4539 * is observed the old clock_pelt_idle value and the new clock_idle,
4540 * which lead to an underestimation. The opposite would lead to an
4544 lut = cfs_rq_last_update_time(cfs_rq);
4549 * cfs_rq->avg.last_update_time is more recent than our
4550 * estimation, let's use it.
4554 now += sched_clock_cpu(cpu_of(rq)) - u64_u32_load(rq->clock_idle);
4556 __update_load_avg_blocked_se(now, se);
4559 static void migrate_se_pelt_lag(struct sched_entity *se) {}
4563 * update_cfs_rq_load_avg - update the cfs_rq's load/util averages
4564 * @now: current time, as per cfs_rq_clock_pelt()
4565 * @cfs_rq: cfs_rq to update
4567 * The cfs_rq avg is the direct sum of all its entities (blocked and runnable)
4568 * avg. The immediate corollary is that all (fair) tasks must be attached.
4570 * cfs_rq->avg is used for task_h_load() and update_cfs_share() for example.
4572 * Return: true if the load decayed or we removed load.
4574 * Since both these conditions indicate a changed cfs_rq->avg.load we should
4575 * call update_tg_load_avg() when this function returns true.
4578 update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq)
4580 unsigned long removed_load = 0, removed_util = 0, removed_runnable = 0;
4581 struct sched_avg *sa = &cfs_rq->avg;
4584 if (cfs_rq->removed.nr) {
4586 u32 divider = get_pelt_divider(&cfs_rq->avg);
4588 raw_spin_lock(&cfs_rq->removed.lock);
4589 swap(cfs_rq->removed.util_avg, removed_util);
4590 swap(cfs_rq->removed.load_avg, removed_load);
4591 swap(cfs_rq->removed.runnable_avg, removed_runnable);
4592 cfs_rq->removed.nr = 0;
4593 raw_spin_unlock(&cfs_rq->removed.lock);
4596 sub_positive(&sa->load_avg, r);
4597 sub_positive(&sa->load_sum, r * divider);
4598 /* See sa->util_sum below */
4599 sa->load_sum = max_t(u32, sa->load_sum, sa->load_avg * PELT_MIN_DIVIDER);
4602 sub_positive(&sa->util_avg, r);
4603 sub_positive(&sa->util_sum, r * divider);
4605 * Because of rounding, se->util_sum might ends up being +1 more than
4606 * cfs->util_sum. Although this is not a problem by itself, detaching
4607 * a lot of tasks with the rounding problem between 2 updates of
4608 * util_avg (~1ms) can make cfs->util_sum becoming null whereas
4609 * cfs_util_avg is not.
4610 * Check that util_sum is still above its lower bound for the new
4611 * util_avg. Given that period_contrib might have moved since the last
4612 * sync, we are only sure that util_sum must be above or equal to
4613 * util_avg * minimum possible divider
4615 sa->util_sum = max_t(u32, sa->util_sum, sa->util_avg * PELT_MIN_DIVIDER);
4617 r = removed_runnable;
4618 sub_positive(&sa->runnable_avg, r);
4619 sub_positive(&sa->runnable_sum, r * divider);
4620 /* See sa->util_sum above */
4621 sa->runnable_sum = max_t(u32, sa->runnable_sum,
4622 sa->runnable_avg * PELT_MIN_DIVIDER);
4625 * removed_runnable is the unweighted version of removed_load so we
4626 * can use it to estimate removed_load_sum.
4628 add_tg_cfs_propagate(cfs_rq,
4629 -(long)(removed_runnable * divider) >> SCHED_CAPACITY_SHIFT);
4634 decayed |= __update_load_avg_cfs_rq(now, cfs_rq);
4635 u64_u32_store_copy(sa->last_update_time,
4636 cfs_rq->last_update_time_copy,
4637 sa->last_update_time);
4642 * attach_entity_load_avg - attach this entity to its cfs_rq load avg
4643 * @cfs_rq: cfs_rq to attach to
4644 * @se: sched_entity to attach
4646 * Must call update_cfs_rq_load_avg() before this, since we rely on
4647 * cfs_rq->avg.last_update_time being current.
4649 static void attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
4652 * cfs_rq->avg.period_contrib can be used for both cfs_rq and se.
4653 * See ___update_load_avg() for details.
4655 u32 divider = get_pelt_divider(&cfs_rq->avg);
4658 * When we attach the @se to the @cfs_rq, we must align the decay
4659 * window because without that, really weird and wonderful things can
4664 se->avg.last_update_time = cfs_rq->avg.last_update_time;
4665 se->avg.period_contrib = cfs_rq->avg.period_contrib;
4668 * Hell(o) Nasty stuff.. we need to recompute _sum based on the new
4669 * period_contrib. This isn't strictly correct, but since we're
4670 * entirely outside of the PELT hierarchy, nobody cares if we truncate
4673 se->avg.util_sum = se->avg.util_avg * divider;
4675 se->avg.runnable_sum = se->avg.runnable_avg * divider;
4677 se->avg.load_sum = se->avg.load_avg * divider;
4678 if (se_weight(se) < se->avg.load_sum)
4679 se->avg.load_sum = div_u64(se->avg.load_sum, se_weight(se));
4681 se->avg.load_sum = 1;
4683 enqueue_load_avg(cfs_rq, se);
4684 cfs_rq->avg.util_avg += se->avg.util_avg;
4685 cfs_rq->avg.util_sum += se->avg.util_sum;
4686 cfs_rq->avg.runnable_avg += se->avg.runnable_avg;
4687 cfs_rq->avg.runnable_sum += se->avg.runnable_sum;
4689 add_tg_cfs_propagate(cfs_rq, se->avg.load_sum);
4691 cfs_rq_util_change(cfs_rq, 0);
4693 trace_pelt_cfs_tp(cfs_rq);
4697 * detach_entity_load_avg - detach this entity from its cfs_rq load avg
4698 * @cfs_rq: cfs_rq to detach from
4699 * @se: sched_entity to detach
4701 * Must call update_cfs_rq_load_avg() before this, since we rely on
4702 * cfs_rq->avg.last_update_time being current.
4704 static void detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
4706 dequeue_load_avg(cfs_rq, se);
4707 sub_positive(&cfs_rq->avg.util_avg, se->avg.util_avg);
4708 sub_positive(&cfs_rq->avg.util_sum, se->avg.util_sum);
4709 /* See update_cfs_rq_load_avg() */
4710 cfs_rq->avg.util_sum = max_t(u32, cfs_rq->avg.util_sum,
4711 cfs_rq->avg.util_avg * PELT_MIN_DIVIDER);
4713 sub_positive(&cfs_rq->avg.runnable_avg, se->avg.runnable_avg);
4714 sub_positive(&cfs_rq->avg.runnable_sum, se->avg.runnable_sum);
4715 /* See update_cfs_rq_load_avg() */
4716 cfs_rq->avg.runnable_sum = max_t(u32, cfs_rq->avg.runnable_sum,
4717 cfs_rq->avg.runnable_avg * PELT_MIN_DIVIDER);
4719 add_tg_cfs_propagate(cfs_rq, -se->avg.load_sum);
4721 cfs_rq_util_change(cfs_rq, 0);
4723 trace_pelt_cfs_tp(cfs_rq);
4727 * Optional action to be done while updating the load average
4729 #define UPDATE_TG 0x1
4730 #define SKIP_AGE_LOAD 0x2
4731 #define DO_ATTACH 0x4
4732 #define DO_DETACH 0x8
4734 /* Update task and its cfs_rq load average */
4735 static inline void update_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
4737 u64 now = cfs_rq_clock_pelt(cfs_rq);
4741 * Track task load average for carrying it to new CPU after migrated, and
4742 * track group sched_entity load average for task_h_load calculation in migration
4744 if (se->avg.last_update_time && !(flags & SKIP_AGE_LOAD))
4745 __update_load_avg_se(now, cfs_rq, se);
4747 decayed = update_cfs_rq_load_avg(now, cfs_rq);
4748 decayed |= propagate_entity_load_avg(se);
4750 if (!se->avg.last_update_time && (flags & DO_ATTACH)) {
4753 * DO_ATTACH means we're here from enqueue_entity().
4754 * !last_update_time means we've passed through
4755 * migrate_task_rq_fair() indicating we migrated.
4757 * IOW we're enqueueing a task on a new CPU.
4759 attach_entity_load_avg(cfs_rq, se);
4760 update_tg_load_avg(cfs_rq);
4762 } else if (flags & DO_DETACH) {
4764 * DO_DETACH means we're here from dequeue_entity()
4765 * and we are migrating task out of the CPU.
4767 detach_entity_load_avg(cfs_rq, se);
4768 update_tg_load_avg(cfs_rq);
4769 } else if (decayed) {
4770 cfs_rq_util_change(cfs_rq, 0);
4772 if (flags & UPDATE_TG)
4773 update_tg_load_avg(cfs_rq);
4778 * Synchronize entity load avg of dequeued entity without locking
4781 static void sync_entity_load_avg(struct sched_entity *se)
4783 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4784 u64 last_update_time;
4786 last_update_time = cfs_rq_last_update_time(cfs_rq);
4787 __update_load_avg_blocked_se(last_update_time, se);
4791 * Task first catches up with cfs_rq, and then subtract
4792 * itself from the cfs_rq (task must be off the queue now).
4794 static void remove_entity_load_avg(struct sched_entity *se)
4796 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4797 unsigned long flags;
4800 * tasks cannot exit without having gone through wake_up_new_task() ->
4801 * enqueue_task_fair() which will have added things to the cfs_rq,
4802 * so we can remove unconditionally.
4805 sync_entity_load_avg(se);
4807 raw_spin_lock_irqsave(&cfs_rq->removed.lock, flags);
4808 ++cfs_rq->removed.nr;
4809 cfs_rq->removed.util_avg += se->avg.util_avg;
4810 cfs_rq->removed.load_avg += se->avg.load_avg;
4811 cfs_rq->removed.runnable_avg += se->avg.runnable_avg;
4812 raw_spin_unlock_irqrestore(&cfs_rq->removed.lock, flags);
4815 static inline unsigned long cfs_rq_runnable_avg(struct cfs_rq *cfs_rq)
4817 return cfs_rq->avg.runnable_avg;
4820 static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq)
4822 return cfs_rq->avg.load_avg;
4825 static int sched_balance_newidle(struct rq *this_rq, struct rq_flags *rf);
4827 static inline unsigned long task_util(struct task_struct *p)
4829 return READ_ONCE(p->se.avg.util_avg);
4832 static inline unsigned long task_runnable(struct task_struct *p)
4834 return READ_ONCE(p->se.avg.runnable_avg);
4837 static inline unsigned long _task_util_est(struct task_struct *p)
4839 return READ_ONCE(p->se.avg.util_est) & ~UTIL_AVG_UNCHANGED;
4842 static inline unsigned long task_util_est(struct task_struct *p)
4844 return max(task_util(p), _task_util_est(p));
4847 static inline void util_est_enqueue(struct cfs_rq *cfs_rq,
4848 struct task_struct *p)
4850 unsigned int enqueued;
4852 if (!sched_feat(UTIL_EST))
4855 /* Update root cfs_rq's estimated utilization */
4856 enqueued = cfs_rq->avg.util_est;
4857 enqueued += _task_util_est(p);
4858 WRITE_ONCE(cfs_rq->avg.util_est, enqueued);
4860 trace_sched_util_est_cfs_tp(cfs_rq);
4863 static inline void util_est_dequeue(struct cfs_rq *cfs_rq,
4864 struct task_struct *p)
4866 unsigned int enqueued;
4868 if (!sched_feat(UTIL_EST))
4871 /* Update root cfs_rq's estimated utilization */
4872 enqueued = cfs_rq->avg.util_est;
4873 enqueued -= min_t(unsigned int, enqueued, _task_util_est(p));
4874 WRITE_ONCE(cfs_rq->avg.util_est, enqueued);
4876 trace_sched_util_est_cfs_tp(cfs_rq);
4879 #define UTIL_EST_MARGIN (SCHED_CAPACITY_SCALE / 100)
4881 static inline void util_est_update(struct cfs_rq *cfs_rq,
4882 struct task_struct *p,
4885 unsigned int ewma, dequeued, last_ewma_diff;
4887 if (!sched_feat(UTIL_EST))
4891 * Skip update of task's estimated utilization when the task has not
4892 * yet completed an activation, e.g. being migrated.
4897 /* Get current estimate of utilization */
4898 ewma = READ_ONCE(p->se.avg.util_est);
4901 * If the PELT values haven't changed since enqueue time,
4902 * skip the util_est update.
4904 if (ewma & UTIL_AVG_UNCHANGED)
4907 /* Get utilization at dequeue */
4908 dequeued = task_util(p);
4911 * Reset EWMA on utilization increases, the moving average is used only
4912 * to smooth utilization decreases.
4914 if (ewma <= dequeued) {
4920 * Skip update of task's estimated utilization when its members are
4921 * already ~1% close to its last activation value.
4923 last_ewma_diff = ewma - dequeued;
4924 if (last_ewma_diff < UTIL_EST_MARGIN)
4928 * To avoid overestimation of actual task utilization, skip updates if
4929 * we cannot grant there is idle time in this CPU.
4931 if (dequeued > arch_scale_cpu_capacity(cpu_of(rq_of(cfs_rq))))
4935 * To avoid underestimate of task utilization, skip updates of EWMA if
4936 * we cannot grant that thread got all CPU time it wanted.
4938 if ((dequeued + UTIL_EST_MARGIN) < task_runnable(p))
4943 * Update Task's estimated utilization
4945 * When *p completes an activation we can consolidate another sample
4946 * of the task size. This is done by using this value to update the
4947 * Exponential Weighted Moving Average (EWMA):
4949 * ewma(t) = w * task_util(p) + (1-w) * ewma(t-1)
4950 * = w * task_util(p) + ewma(t-1) - w * ewma(t-1)
4951 * = w * (task_util(p) - ewma(t-1)) + ewma(t-1)
4952 * = w * ( -last_ewma_diff ) + ewma(t-1)
4953 * = w * (-last_ewma_diff + ewma(t-1) / w)
4955 * Where 'w' is the weight of new samples, which is configured to be
4956 * 0.25, thus making w=1/4 ( >>= UTIL_EST_WEIGHT_SHIFT)
4958 ewma <<= UTIL_EST_WEIGHT_SHIFT;
4959 ewma -= last_ewma_diff;
4960 ewma >>= UTIL_EST_WEIGHT_SHIFT;
4962 ewma |= UTIL_AVG_UNCHANGED;
4963 WRITE_ONCE(p->se.avg.util_est, ewma);
4965 trace_sched_util_est_se_tp(&p->se);
4968 static inline unsigned long get_actual_cpu_capacity(int cpu)
4970 unsigned long capacity = arch_scale_cpu_capacity(cpu);
4972 capacity -= max(hw_load_avg(cpu_rq(cpu)), cpufreq_get_pressure(cpu));
4977 static inline int util_fits_cpu(unsigned long util,
4978 unsigned long uclamp_min,
4979 unsigned long uclamp_max,
4982 unsigned long capacity = capacity_of(cpu);
4983 unsigned long capacity_orig;
4984 bool fits, uclamp_max_fits;
4987 * Check if the real util fits without any uclamp boost/cap applied.
4989 fits = fits_capacity(util, capacity);
4991 if (!uclamp_is_used())
4995 * We must use arch_scale_cpu_capacity() for comparing against uclamp_min and
4996 * uclamp_max. We only care about capacity pressure (by using
4997 * capacity_of()) for comparing against the real util.
4999 * If a task is boosted to 1024 for example, we don't want a tiny
5000 * pressure to skew the check whether it fits a CPU or not.
5002 * Similarly if a task is capped to arch_scale_cpu_capacity(little_cpu), it
5003 * should fit a little cpu even if there's some pressure.
5005 * Only exception is for HW or cpufreq pressure since it has a direct impact
5006 * on available OPP of the system.
5008 * We honour it for uclamp_min only as a drop in performance level
5009 * could result in not getting the requested minimum performance level.
5011 * For uclamp_max, we can tolerate a drop in performance level as the
5012 * goal is to cap the task. So it's okay if it's getting less.
5014 capacity_orig = arch_scale_cpu_capacity(cpu);
5017 * We want to force a task to fit a cpu as implied by uclamp_max.
5018 * But we do have some corner cases to cater for..
5024 * |_ _ _ _ _ _ _ _ _ ___ _ _ _ | _ | _ _ _ _ _ uclamp_max
5027 * | | | | | | | (util somewhere in this region)
5030 * +----------------------------------------
5033 * In the above example if a task is capped to a specific performance
5034 * point, y, then when:
5036 * * util = 80% of x then it does not fit on CPU0 and should migrate
5038 * * util = 80% of y then it is forced to fit on CPU1 to honour
5039 * uclamp_max request.
5041 * which is what we're enforcing here. A task always fits if
5042 * uclamp_max <= capacity_orig. But when uclamp_max > capacity_orig,
5043 * the normal upmigration rules should withhold still.
5045 * Only exception is when we are on max capacity, then we need to be
5046 * careful not to block overutilized state. This is so because:
5048 * 1. There's no concept of capping at max_capacity! We can't go
5049 * beyond this performance level anyway.
5050 * 2. The system is being saturated when we're operating near
5051 * max capacity, it doesn't make sense to block overutilized.
5053 uclamp_max_fits = (capacity_orig == SCHED_CAPACITY_SCALE) && (uclamp_max == SCHED_CAPACITY_SCALE);
5054 uclamp_max_fits = !uclamp_max_fits && (uclamp_max <= capacity_orig);
5055 fits = fits || uclamp_max_fits;
5060 * | ___ (region a, capped, util >= uclamp_max)
5062 * |_ _ _ _ _ _ _ _ _ ___ _ _ _ | _ | _ _ _ _ _ uclamp_max
5064 * | ___ | | | | (region b, uclamp_min <= util <= uclamp_max)
5065 * |_ _ _|_ _|_ _ _ _| _ | _ _ _| _ | _ _ _ _ _ uclamp_min
5067 * | | | | | | | (region c, boosted, util < uclamp_min)
5068 * +----------------------------------------
5071 * a) If util > uclamp_max, then we're capped, we don't care about
5072 * actual fitness value here. We only care if uclamp_max fits
5073 * capacity without taking margin/pressure into account.
5074 * See comment above.
5076 * b) If uclamp_min <= util <= uclamp_max, then the normal
5077 * fits_capacity() rules apply. Except we need to ensure that we
5078 * enforce we remain within uclamp_max, see comment above.
5080 * c) If util < uclamp_min, then we are boosted. Same as (b) but we
5081 * need to take into account the boosted value fits the CPU without
5082 * taking margin/pressure into account.
5084 * Cases (a) and (b) are handled in the 'fits' variable already. We
5085 * just need to consider an extra check for case (c) after ensuring we
5086 * handle the case uclamp_min > uclamp_max.
5088 uclamp_min = min(uclamp_min, uclamp_max);
5089 if (fits && (util < uclamp_min) &&
5090 (uclamp_min > get_actual_cpu_capacity(cpu)))
5096 static inline int task_fits_cpu(struct task_struct *p, int cpu)
5098 unsigned long uclamp_min = uclamp_eff_value(p, UCLAMP_MIN);
5099 unsigned long uclamp_max = uclamp_eff_value(p, UCLAMP_MAX);
5100 unsigned long util = task_util_est(p);
5102 * Return true only if the cpu fully fits the task requirements, which
5103 * include the utilization but also the performance hints.
5105 return (util_fits_cpu(util, uclamp_min, uclamp_max, cpu) > 0);
5108 static inline void update_misfit_status(struct task_struct *p, struct rq *rq)
5110 int cpu = cpu_of(rq);
5112 if (!sched_asym_cpucap_active())
5116 * Affinity allows us to go somewhere higher? Or are we on biggest
5117 * available CPU already? Or do we fit into this CPU ?
5119 if (!p || (p->nr_cpus_allowed == 1) ||
5120 (arch_scale_cpu_capacity(cpu) == p->max_allowed_capacity) ||
5121 task_fits_cpu(p, cpu)) {
5123 rq->misfit_task_load = 0;
5128 * Make sure that misfit_task_load will not be null even if
5129 * task_h_load() returns 0.
5131 rq->misfit_task_load = max_t(unsigned long, task_h_load(p), 1);
5134 #else /* CONFIG_SMP */
5136 static inline bool cfs_rq_is_decayed(struct cfs_rq *cfs_rq)
5138 return !cfs_rq->nr_running;
5141 #define UPDATE_TG 0x0
5142 #define SKIP_AGE_LOAD 0x0
5143 #define DO_ATTACH 0x0
5144 #define DO_DETACH 0x0
5146 static inline void update_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int not_used1)
5148 cfs_rq_util_change(cfs_rq, 0);
5151 static inline void remove_entity_load_avg(struct sched_entity *se) {}
5154 attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
5156 detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
5158 static inline int sched_balance_newidle(struct rq *rq, struct rq_flags *rf)
5164 util_est_enqueue(struct cfs_rq *cfs_rq, struct task_struct *p) {}
5167 util_est_dequeue(struct cfs_rq *cfs_rq, struct task_struct *p) {}
5170 util_est_update(struct cfs_rq *cfs_rq, struct task_struct *p,
5172 static inline void update_misfit_status(struct task_struct *p, struct rq *rq) {}
5174 #endif /* CONFIG_SMP */
5177 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
5179 u64 vslice, vruntime = avg_vruntime(cfs_rq);
5182 se->slice = sysctl_sched_base_slice;
5183 vslice = calc_delta_fair(se->slice, se);
5186 * Due to how V is constructed as the weighted average of entities,
5187 * adding tasks with positive lag, or removing tasks with negative lag
5188 * will move 'time' backwards, this can screw around with the lag of
5191 * EEVDF: placement strategy #1 / #2
5193 if (sched_feat(PLACE_LAG) && cfs_rq->nr_running) {
5194 struct sched_entity *curr = cfs_rq->curr;
5200 * If we want to place a task and preserve lag, we have to
5201 * consider the effect of the new entity on the weighted
5202 * average and compensate for this, otherwise lag can quickly
5205 * Lag is defined as:
5207 * lag_i = S - s_i = w_i * (V - v_i)
5209 * To avoid the 'w_i' term all over the place, we only track
5212 * vl_i = V - v_i <=> v_i = V - vl_i
5214 * And we take V to be the weighted average of all v:
5216 * V = (\Sum w_j*v_j) / W
5218 * Where W is: \Sum w_j
5220 * Then, the weighted average after adding an entity with lag
5223 * V' = (\Sum w_j*v_j + w_i*v_i) / (W + w_i)
5224 * = (W*V + w_i*(V - vl_i)) / (W + w_i)
5225 * = (W*V + w_i*V - w_i*vl_i) / (W + w_i)
5226 * = (V*(W + w_i) - w_i*l) / (W + w_i)
5227 * = V - w_i*vl_i / (W + w_i)
5229 * And the actual lag after adding an entity with vl_i is:
5232 * = V - w_i*vl_i / (W + w_i) - (V - vl_i)
5233 * = vl_i - w_i*vl_i / (W + w_i)
5235 * Which is strictly less than vl_i. So in order to preserve lag
5236 * we should inflate the lag before placement such that the
5237 * effective lag after placement comes out right.
5239 * As such, invert the above relation for vl'_i to get the vl_i
5240 * we need to use such that the lag after placement is the lag
5241 * we computed before dequeue.
5243 * vl'_i = vl_i - w_i*vl_i / (W + w_i)
5244 * = ((W + w_i)*vl_i - w_i*vl_i) / (W + w_i)
5246 * (W + w_i)*vl'_i = (W + w_i)*vl_i - w_i*vl_i
5249 * vl_i = (W + w_i)*vl'_i / W
5251 load = cfs_rq->avg_load;
5252 if (curr && curr->on_rq)
5253 load += scale_load_down(curr->load.weight);
5255 lag *= load + scale_load_down(se->load.weight);
5256 if (WARN_ON_ONCE(!load))
5258 lag = div_s64(lag, load);
5261 se->vruntime = vruntime - lag;
5264 * When joining the competition; the existing tasks will be,
5265 * on average, halfway through their slice, as such start tasks
5266 * off with half a slice to ease into the competition.
5268 if (sched_feat(PLACE_DEADLINE_INITIAL) && (flags & ENQUEUE_INITIAL))
5272 * EEVDF: vd_i = ve_i + r_i/w_i
5274 se->deadline = se->vruntime + vslice;
5277 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
5278 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq);
5280 static inline bool cfs_bandwidth_used(void);
5283 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
5285 bool curr = cfs_rq->curr == se;
5288 * If we're the current task, we must renormalise before calling
5292 place_entity(cfs_rq, se, flags);
5294 update_curr(cfs_rq);
5297 * When enqueuing a sched_entity, we must:
5298 * - Update loads to have both entity and cfs_rq synced with now.
5299 * - For group_entity, update its runnable_weight to reflect the new
5300 * h_nr_running of its group cfs_rq.
5301 * - For group_entity, update its weight to reflect the new share of
5303 * - Add its new weight to cfs_rq->load.weight
5305 update_load_avg(cfs_rq, se, UPDATE_TG | DO_ATTACH);
5306 se_update_runnable(se);
5308 * XXX update_load_avg() above will have attached us to the pelt sum;
5309 * but update_cfs_group() here will re-adjust the weight and have to
5310 * undo/redo all that. Seems wasteful.
5312 update_cfs_group(se);
5315 * XXX now that the entity has been re-weighted, and it's lag adjusted,
5316 * we can place the entity.
5319 place_entity(cfs_rq, se, flags);
5321 account_entity_enqueue(cfs_rq, se);
5323 /* Entity has migrated, no longer consider this task hot */
5324 if (flags & ENQUEUE_MIGRATED)
5327 check_schedstat_required();
5328 update_stats_enqueue_fair(cfs_rq, se, flags);
5330 __enqueue_entity(cfs_rq, se);
5333 if (cfs_rq->nr_running == 1) {
5334 check_enqueue_throttle(cfs_rq);
5335 if (!throttled_hierarchy(cfs_rq)) {
5336 list_add_leaf_cfs_rq(cfs_rq);
5338 #ifdef CONFIG_CFS_BANDWIDTH
5339 struct rq *rq = rq_of(cfs_rq);
5341 if (cfs_rq_throttled(cfs_rq) && !cfs_rq->throttled_clock)
5342 cfs_rq->throttled_clock = rq_clock(rq);
5343 if (!cfs_rq->throttled_clock_self)
5344 cfs_rq->throttled_clock_self = rq_clock(rq);
5350 static void __clear_buddies_next(struct sched_entity *se)
5352 for_each_sched_entity(se) {
5353 struct cfs_rq *cfs_rq = cfs_rq_of(se);
5354 if (cfs_rq->next != se)
5357 cfs_rq->next = NULL;
5361 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
5363 if (cfs_rq->next == se)
5364 __clear_buddies_next(se);
5367 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
5370 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
5372 int action = UPDATE_TG;
5374 if (entity_is_task(se) && task_on_rq_migrating(task_of(se)))
5375 action |= DO_DETACH;
5378 * Update run-time statistics of the 'current'.
5380 update_curr(cfs_rq);
5383 * When dequeuing a sched_entity, we must:
5384 * - Update loads to have both entity and cfs_rq synced with now.
5385 * - For group_entity, update its runnable_weight to reflect the new
5386 * h_nr_running of its group cfs_rq.
5387 * - Subtract its previous weight from cfs_rq->load.weight.
5388 * - For group entity, update its weight to reflect the new share
5389 * of its group cfs_rq.
5391 update_load_avg(cfs_rq, se, action);
5392 se_update_runnable(se);
5394 update_stats_dequeue_fair(cfs_rq, se, flags);
5396 clear_buddies(cfs_rq, se);
5398 update_entity_lag(cfs_rq, se);
5399 if (se != cfs_rq->curr)
5400 __dequeue_entity(cfs_rq, se);
5402 account_entity_dequeue(cfs_rq, se);
5404 /* return excess runtime on last dequeue */
5405 return_cfs_rq_runtime(cfs_rq);
5407 update_cfs_group(se);
5410 * Now advance min_vruntime if @se was the entity holding it back,
5411 * except when: DEQUEUE_SAVE && !DEQUEUE_MOVE, in this case we'll be
5412 * put back on, and if we advance min_vruntime, we'll be placed back
5413 * further than we started -- i.e. we'll be penalized.
5415 if ((flags & (DEQUEUE_SAVE | DEQUEUE_MOVE)) != DEQUEUE_SAVE)
5416 update_min_vruntime(cfs_rq);
5418 if (cfs_rq->nr_running == 0)
5419 update_idle_cfs_rq_clock_pelt(cfs_rq);
5423 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
5425 clear_buddies(cfs_rq, se);
5427 /* 'current' is not kept within the tree. */
5430 * Any task has to be enqueued before it get to execute on
5431 * a CPU. So account for the time it spent waiting on the
5434 update_stats_wait_end_fair(cfs_rq, se);
5435 __dequeue_entity(cfs_rq, se);
5436 update_load_avg(cfs_rq, se, UPDATE_TG);
5438 * HACK, stash a copy of deadline at the point of pick in vlag,
5439 * which isn't used until dequeue.
5441 se->vlag = se->deadline;
5444 update_stats_curr_start(cfs_rq, se);
5448 * Track our maximum slice length, if the CPU's load is at
5449 * least twice that of our own weight (i.e. don't track it
5450 * when there are only lesser-weight tasks around):
5452 if (schedstat_enabled() &&
5453 rq_of(cfs_rq)->cfs.load.weight >= 2*se->load.weight) {
5454 struct sched_statistics *stats;
5456 stats = __schedstats_from_se(se);
5457 __schedstat_set(stats->slice_max,
5458 max((u64)stats->slice_max,
5459 se->sum_exec_runtime - se->prev_sum_exec_runtime));
5462 se->prev_sum_exec_runtime = se->sum_exec_runtime;
5466 * Pick the next process, keeping these things in mind, in this order:
5467 * 1) keep things fair between processes/task groups
5468 * 2) pick the "next" process, since someone really wants that to run
5469 * 3) pick the "last" process, for cache locality
5470 * 4) do not run the "skip" process, if something else is available
5472 static struct sched_entity *
5473 pick_next_entity(struct cfs_rq *cfs_rq)
5476 * Enabling NEXT_BUDDY will affect latency but not fairness.
5478 if (sched_feat(NEXT_BUDDY) &&
5479 cfs_rq->next && entity_eligible(cfs_rq, cfs_rq->next))
5480 return cfs_rq->next;
5482 return pick_eevdf(cfs_rq);
5485 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
5487 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
5490 * If still on the runqueue then deactivate_task()
5491 * was not called and update_curr() has to be done:
5494 update_curr(cfs_rq);
5496 /* throttle cfs_rqs exceeding runtime */
5497 check_cfs_rq_runtime(cfs_rq);
5500 update_stats_wait_start_fair(cfs_rq, prev);
5501 /* Put 'current' back into the tree. */
5502 __enqueue_entity(cfs_rq, prev);
5503 /* in !on_rq case, update occurred at dequeue */
5504 update_load_avg(cfs_rq, prev, 0);
5506 cfs_rq->curr = NULL;
5510 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
5513 * Update run-time statistics of the 'current'.
5515 update_curr(cfs_rq);
5518 * Ensure that runnable average is periodically updated.
5520 update_load_avg(cfs_rq, curr, UPDATE_TG);
5521 update_cfs_group(curr);
5523 #ifdef CONFIG_SCHED_HRTICK
5525 * queued ticks are scheduled to match the slice, so don't bother
5526 * validating it and just reschedule.
5529 resched_curr(rq_of(cfs_rq));
5533 * don't let the period tick interfere with the hrtick preemption
5535 if (!sched_feat(DOUBLE_TICK) &&
5536 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
5542 /**************************************************
5543 * CFS bandwidth control machinery
5546 #ifdef CONFIG_CFS_BANDWIDTH
5548 #ifdef CONFIG_JUMP_LABEL
5549 static struct static_key __cfs_bandwidth_used;
5551 static inline bool cfs_bandwidth_used(void)
5553 return static_key_false(&__cfs_bandwidth_used);
5556 void cfs_bandwidth_usage_inc(void)
5558 static_key_slow_inc_cpuslocked(&__cfs_bandwidth_used);
5561 void cfs_bandwidth_usage_dec(void)
5563 static_key_slow_dec_cpuslocked(&__cfs_bandwidth_used);
5565 #else /* CONFIG_JUMP_LABEL */
5566 static bool cfs_bandwidth_used(void)
5571 void cfs_bandwidth_usage_inc(void) {}
5572 void cfs_bandwidth_usage_dec(void) {}
5573 #endif /* CONFIG_JUMP_LABEL */
5576 * default period for cfs group bandwidth.
5577 * default: 0.1s, units: nanoseconds
5579 static inline u64 default_cfs_period(void)
5581 return 100000000ULL;
5584 static inline u64 sched_cfs_bandwidth_slice(void)
5586 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
5590 * Replenish runtime according to assigned quota. We use sched_clock_cpu
5591 * directly instead of rq->clock to avoid adding additional synchronization
5594 * requires cfs_b->lock
5596 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
5600 if (unlikely(cfs_b->quota == RUNTIME_INF))
5603 cfs_b->runtime += cfs_b->quota;
5604 runtime = cfs_b->runtime_snap - cfs_b->runtime;
5606 cfs_b->burst_time += runtime;
5610 cfs_b->runtime = min(cfs_b->runtime, cfs_b->quota + cfs_b->burst);
5611 cfs_b->runtime_snap = cfs_b->runtime;
5614 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
5616 return &tg->cfs_bandwidth;
5619 /* returns 0 on failure to allocate runtime */
5620 static int __assign_cfs_rq_runtime(struct cfs_bandwidth *cfs_b,
5621 struct cfs_rq *cfs_rq, u64 target_runtime)
5623 u64 min_amount, amount = 0;
5625 lockdep_assert_held(&cfs_b->lock);
5627 /* note: this is a positive sum as runtime_remaining <= 0 */
5628 min_amount = target_runtime - cfs_rq->runtime_remaining;
5630 if (cfs_b->quota == RUNTIME_INF)
5631 amount = min_amount;
5633 start_cfs_bandwidth(cfs_b);
5635 if (cfs_b->runtime > 0) {
5636 amount = min(cfs_b->runtime, min_amount);
5637 cfs_b->runtime -= amount;
5642 cfs_rq->runtime_remaining += amount;
5644 return cfs_rq->runtime_remaining > 0;
5647 /* returns 0 on failure to allocate runtime */
5648 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
5650 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
5653 raw_spin_lock(&cfs_b->lock);
5654 ret = __assign_cfs_rq_runtime(cfs_b, cfs_rq, sched_cfs_bandwidth_slice());
5655 raw_spin_unlock(&cfs_b->lock);
5660 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
5662 /* dock delta_exec before expiring quota (as it could span periods) */
5663 cfs_rq->runtime_remaining -= delta_exec;
5665 if (likely(cfs_rq->runtime_remaining > 0))
5668 if (cfs_rq->throttled)
5671 * if we're unable to extend our runtime we resched so that the active
5672 * hierarchy can be throttled
5674 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
5675 resched_curr(rq_of(cfs_rq));
5678 static __always_inline
5679 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
5681 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
5684 __account_cfs_rq_runtime(cfs_rq, delta_exec);
5687 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
5689 return cfs_bandwidth_used() && cfs_rq->throttled;
5692 /* check whether cfs_rq, or any parent, is throttled */
5693 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
5695 return cfs_bandwidth_used() && cfs_rq->throttle_count;
5699 * Ensure that neither of the group entities corresponding to src_cpu or
5700 * dest_cpu are members of a throttled hierarchy when performing group
5701 * load-balance operations.
5703 static inline int throttled_lb_pair(struct task_group *tg,
5704 int src_cpu, int dest_cpu)
5706 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
5708 src_cfs_rq = tg->cfs_rq[src_cpu];
5709 dest_cfs_rq = tg->cfs_rq[dest_cpu];
5711 return throttled_hierarchy(src_cfs_rq) ||
5712 throttled_hierarchy(dest_cfs_rq);
5715 static int tg_unthrottle_up(struct task_group *tg, void *data)
5717 struct rq *rq = data;
5718 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
5720 cfs_rq->throttle_count--;
5721 if (!cfs_rq->throttle_count) {
5722 cfs_rq->throttled_clock_pelt_time += rq_clock_pelt(rq) -
5723 cfs_rq->throttled_clock_pelt;
5725 /* Add cfs_rq with load or one or more already running entities to the list */
5726 if (!cfs_rq_is_decayed(cfs_rq))
5727 list_add_leaf_cfs_rq(cfs_rq);
5729 if (cfs_rq->throttled_clock_self) {
5730 u64 delta = rq_clock(rq) - cfs_rq->throttled_clock_self;
5732 cfs_rq->throttled_clock_self = 0;
5734 if (SCHED_WARN_ON((s64)delta < 0))
5737 cfs_rq->throttled_clock_self_time += delta;
5744 static int tg_throttle_down(struct task_group *tg, void *data)
5746 struct rq *rq = data;
5747 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
5749 /* group is entering throttled state, stop time */
5750 if (!cfs_rq->throttle_count) {
5751 cfs_rq->throttled_clock_pelt = rq_clock_pelt(rq);
5752 list_del_leaf_cfs_rq(cfs_rq);
5754 SCHED_WARN_ON(cfs_rq->throttled_clock_self);
5755 if (cfs_rq->nr_running)
5756 cfs_rq->throttled_clock_self = rq_clock(rq);
5758 cfs_rq->throttle_count++;
5763 static bool throttle_cfs_rq(struct cfs_rq *cfs_rq)
5765 struct rq *rq = rq_of(cfs_rq);
5766 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
5767 struct sched_entity *se;
5768 long task_delta, idle_task_delta, dequeue = 1;
5770 raw_spin_lock(&cfs_b->lock);
5771 /* This will start the period timer if necessary */
5772 if (__assign_cfs_rq_runtime(cfs_b, cfs_rq, 1)) {
5774 * We have raced with bandwidth becoming available, and if we
5775 * actually throttled the timer might not unthrottle us for an
5776 * entire period. We additionally needed to make sure that any
5777 * subsequent check_cfs_rq_runtime calls agree not to throttle
5778 * us, as we may commit to do cfs put_prev+pick_next, so we ask
5779 * for 1ns of runtime rather than just check cfs_b.
5783 list_add_tail_rcu(&cfs_rq->throttled_list,
5784 &cfs_b->throttled_cfs_rq);
5786 raw_spin_unlock(&cfs_b->lock);
5789 return false; /* Throttle no longer required. */
5791 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
5793 /* freeze hierarchy runnable averages while throttled */
5795 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
5798 task_delta = cfs_rq->h_nr_running;
5799 idle_task_delta = cfs_rq->idle_h_nr_running;
5800 for_each_sched_entity(se) {
5801 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
5802 /* throttled entity or throttle-on-deactivate */
5806 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
5808 if (cfs_rq_is_idle(group_cfs_rq(se)))
5809 idle_task_delta = cfs_rq->h_nr_running;
5811 qcfs_rq->h_nr_running -= task_delta;
5812 qcfs_rq->idle_h_nr_running -= idle_task_delta;
5814 if (qcfs_rq->load.weight) {
5815 /* Avoid re-evaluating load for this entity: */
5816 se = parent_entity(se);
5821 for_each_sched_entity(se) {
5822 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
5823 /* throttled entity or throttle-on-deactivate */
5827 update_load_avg(qcfs_rq, se, 0);
5828 se_update_runnable(se);
5830 if (cfs_rq_is_idle(group_cfs_rq(se)))
5831 idle_task_delta = cfs_rq->h_nr_running;
5833 qcfs_rq->h_nr_running -= task_delta;
5834 qcfs_rq->idle_h_nr_running -= idle_task_delta;
5837 /* At this point se is NULL and we are at root level*/
5838 sub_nr_running(rq, task_delta);
5842 * Note: distribution will already see us throttled via the
5843 * throttled-list. rq->lock protects completion.
5845 cfs_rq->throttled = 1;
5846 SCHED_WARN_ON(cfs_rq->throttled_clock);
5847 if (cfs_rq->nr_running)
5848 cfs_rq->throttled_clock = rq_clock(rq);
5852 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
5854 struct rq *rq = rq_of(cfs_rq);
5855 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
5856 struct sched_entity *se;
5857 long task_delta, idle_task_delta;
5859 se = cfs_rq->tg->se[cpu_of(rq)];
5861 cfs_rq->throttled = 0;
5863 update_rq_clock(rq);
5865 raw_spin_lock(&cfs_b->lock);
5866 if (cfs_rq->throttled_clock) {
5867 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
5868 cfs_rq->throttled_clock = 0;
5870 list_del_rcu(&cfs_rq->throttled_list);
5871 raw_spin_unlock(&cfs_b->lock);
5873 /* update hierarchical throttle state */
5874 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
5876 if (!cfs_rq->load.weight) {
5877 if (!cfs_rq->on_list)
5880 * Nothing to run but something to decay (on_list)?
5881 * Complete the branch.
5883 for_each_sched_entity(se) {
5884 if (list_add_leaf_cfs_rq(cfs_rq_of(se)))
5887 goto unthrottle_throttle;
5890 task_delta = cfs_rq->h_nr_running;
5891 idle_task_delta = cfs_rq->idle_h_nr_running;
5892 for_each_sched_entity(se) {
5893 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
5897 enqueue_entity(qcfs_rq, se, ENQUEUE_WAKEUP);
5899 if (cfs_rq_is_idle(group_cfs_rq(se)))
5900 idle_task_delta = cfs_rq->h_nr_running;
5902 qcfs_rq->h_nr_running += task_delta;
5903 qcfs_rq->idle_h_nr_running += idle_task_delta;
5905 /* end evaluation on encountering a throttled cfs_rq */
5906 if (cfs_rq_throttled(qcfs_rq))
5907 goto unthrottle_throttle;
5910 for_each_sched_entity(se) {
5911 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
5913 update_load_avg(qcfs_rq, se, UPDATE_TG);
5914 se_update_runnable(se);
5916 if (cfs_rq_is_idle(group_cfs_rq(se)))
5917 idle_task_delta = cfs_rq->h_nr_running;
5919 qcfs_rq->h_nr_running += task_delta;
5920 qcfs_rq->idle_h_nr_running += idle_task_delta;
5922 /* end evaluation on encountering a throttled cfs_rq */
5923 if (cfs_rq_throttled(qcfs_rq))
5924 goto unthrottle_throttle;
5927 /* At this point se is NULL and we are at root level*/
5928 add_nr_running(rq, task_delta);
5930 unthrottle_throttle:
5931 assert_list_leaf_cfs_rq(rq);
5933 /* Determine whether we need to wake up potentially idle CPU: */
5934 if (rq->curr == rq->idle && rq->cfs.nr_running)
5939 static void __cfsb_csd_unthrottle(void *arg)
5941 struct cfs_rq *cursor, *tmp;
5942 struct rq *rq = arg;
5948 * Iterating over the list can trigger several call to
5949 * update_rq_clock() in unthrottle_cfs_rq().
5950 * Do it once and skip the potential next ones.
5952 update_rq_clock(rq);
5953 rq_clock_start_loop_update(rq);
5956 * Since we hold rq lock we're safe from concurrent manipulation of
5957 * the CSD list. However, this RCU critical section annotates the
5958 * fact that we pair with sched_free_group_rcu(), so that we cannot
5959 * race with group being freed in the window between removing it
5960 * from the list and advancing to the next entry in the list.
5964 list_for_each_entry_safe(cursor, tmp, &rq->cfsb_csd_list,
5965 throttled_csd_list) {
5966 list_del_init(&cursor->throttled_csd_list);
5968 if (cfs_rq_throttled(cursor))
5969 unthrottle_cfs_rq(cursor);
5974 rq_clock_stop_loop_update(rq);
5978 static inline void __unthrottle_cfs_rq_async(struct cfs_rq *cfs_rq)
5980 struct rq *rq = rq_of(cfs_rq);
5983 if (rq == this_rq()) {
5984 unthrottle_cfs_rq(cfs_rq);
5988 /* Already enqueued */
5989 if (SCHED_WARN_ON(!list_empty(&cfs_rq->throttled_csd_list)))
5992 first = list_empty(&rq->cfsb_csd_list);
5993 list_add_tail(&cfs_rq->throttled_csd_list, &rq->cfsb_csd_list);
5995 smp_call_function_single_async(cpu_of(rq), &rq->cfsb_csd);
5998 static inline void __unthrottle_cfs_rq_async(struct cfs_rq *cfs_rq)
6000 unthrottle_cfs_rq(cfs_rq);
6004 static void unthrottle_cfs_rq_async(struct cfs_rq *cfs_rq)
6006 lockdep_assert_rq_held(rq_of(cfs_rq));
6008 if (SCHED_WARN_ON(!cfs_rq_throttled(cfs_rq) ||
6009 cfs_rq->runtime_remaining <= 0))
6012 __unthrottle_cfs_rq_async(cfs_rq);
6015 static bool distribute_cfs_runtime(struct cfs_bandwidth *cfs_b)
6017 int this_cpu = smp_processor_id();
6018 u64 runtime, remaining = 1;
6019 bool throttled = false;
6020 struct cfs_rq *cfs_rq, *tmp;
6023 LIST_HEAD(local_unthrottle);
6026 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
6035 rq_lock_irqsave(rq, &rf);
6036 if (!cfs_rq_throttled(cfs_rq))
6039 /* Already queued for async unthrottle */
6040 if (!list_empty(&cfs_rq->throttled_csd_list))
6043 /* By the above checks, this should never be true */
6044 SCHED_WARN_ON(cfs_rq->runtime_remaining > 0);
6046 raw_spin_lock(&cfs_b->lock);
6047 runtime = -cfs_rq->runtime_remaining + 1;
6048 if (runtime > cfs_b->runtime)
6049 runtime = cfs_b->runtime;
6050 cfs_b->runtime -= runtime;
6051 remaining = cfs_b->runtime;
6052 raw_spin_unlock(&cfs_b->lock);
6054 cfs_rq->runtime_remaining += runtime;
6056 /* we check whether we're throttled above */
6057 if (cfs_rq->runtime_remaining > 0) {
6058 if (cpu_of(rq) != this_cpu) {
6059 unthrottle_cfs_rq_async(cfs_rq);
6062 * We currently only expect to be unthrottling
6063 * a single cfs_rq locally.
6065 SCHED_WARN_ON(!list_empty(&local_unthrottle));
6066 list_add_tail(&cfs_rq->throttled_csd_list,
6074 rq_unlock_irqrestore(rq, &rf);
6077 list_for_each_entry_safe(cfs_rq, tmp, &local_unthrottle,
6078 throttled_csd_list) {
6079 struct rq *rq = rq_of(cfs_rq);
6081 rq_lock_irqsave(rq, &rf);
6083 list_del_init(&cfs_rq->throttled_csd_list);
6085 if (cfs_rq_throttled(cfs_rq))
6086 unthrottle_cfs_rq(cfs_rq);
6088 rq_unlock_irqrestore(rq, &rf);
6090 SCHED_WARN_ON(!list_empty(&local_unthrottle));
6098 * Responsible for refilling a task_group's bandwidth and unthrottling its
6099 * cfs_rqs as appropriate. If there has been no activity within the last
6100 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
6101 * used to track this state.
6103 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun, unsigned long flags)
6107 /* no need to continue the timer with no bandwidth constraint */
6108 if (cfs_b->quota == RUNTIME_INF)
6109 goto out_deactivate;
6111 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
6112 cfs_b->nr_periods += overrun;
6114 /* Refill extra burst quota even if cfs_b->idle */
6115 __refill_cfs_bandwidth_runtime(cfs_b);
6118 * idle depends on !throttled (for the case of a large deficit), and if
6119 * we're going inactive then everything else can be deferred
6121 if (cfs_b->idle && !throttled)
6122 goto out_deactivate;
6125 /* mark as potentially idle for the upcoming period */
6130 /* account preceding periods in which throttling occurred */
6131 cfs_b->nr_throttled += overrun;
6134 * This check is repeated as we release cfs_b->lock while we unthrottle.
6136 while (throttled && cfs_b->runtime > 0) {
6137 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
6138 /* we can't nest cfs_b->lock while distributing bandwidth */
6139 throttled = distribute_cfs_runtime(cfs_b);
6140 raw_spin_lock_irqsave(&cfs_b->lock, flags);
6144 * While we are ensured activity in the period following an
6145 * unthrottle, this also covers the case in which the new bandwidth is
6146 * insufficient to cover the existing bandwidth deficit. (Forcing the
6147 * timer to remain active while there are any throttled entities.)
6157 /* a cfs_rq won't donate quota below this amount */
6158 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
6159 /* minimum remaining period time to redistribute slack quota */
6160 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
6161 /* how long we wait to gather additional slack before distributing */
6162 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
6165 * Are we near the end of the current quota period?
6167 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
6168 * hrtimer base being cleared by hrtimer_start. In the case of
6169 * migrate_hrtimers, base is never cleared, so we are fine.
6171 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
6173 struct hrtimer *refresh_timer = &cfs_b->period_timer;
6176 /* if the call-back is running a quota refresh is already occurring */
6177 if (hrtimer_callback_running(refresh_timer))
6180 /* is a quota refresh about to occur? */
6181 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
6182 if (remaining < (s64)min_expire)
6188 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
6190 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
6192 /* if there's a quota refresh soon don't bother with slack */
6193 if (runtime_refresh_within(cfs_b, min_left))
6196 /* don't push forwards an existing deferred unthrottle */
6197 if (cfs_b->slack_started)
6199 cfs_b->slack_started = true;
6201 hrtimer_start(&cfs_b->slack_timer,
6202 ns_to_ktime(cfs_bandwidth_slack_period),
6206 /* we know any runtime found here is valid as update_curr() precedes return */
6207 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
6209 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
6210 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
6212 if (slack_runtime <= 0)
6215 raw_spin_lock(&cfs_b->lock);
6216 if (cfs_b->quota != RUNTIME_INF) {
6217 cfs_b->runtime += slack_runtime;
6219 /* we are under rq->lock, defer unthrottling using a timer */
6220 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
6221 !list_empty(&cfs_b->throttled_cfs_rq))
6222 start_cfs_slack_bandwidth(cfs_b);
6224 raw_spin_unlock(&cfs_b->lock);
6226 /* even if it's not valid for return we don't want to try again */
6227 cfs_rq->runtime_remaining -= slack_runtime;
6230 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
6232 if (!cfs_bandwidth_used())
6235 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
6238 __return_cfs_rq_runtime(cfs_rq);
6242 * This is done with a timer (instead of inline with bandwidth return) since
6243 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
6245 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
6247 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
6248 unsigned long flags;
6250 /* confirm we're still not at a refresh boundary */
6251 raw_spin_lock_irqsave(&cfs_b->lock, flags);
6252 cfs_b->slack_started = false;
6254 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
6255 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
6259 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice)
6260 runtime = cfs_b->runtime;
6262 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
6267 distribute_cfs_runtime(cfs_b);
6271 * When a group wakes up we want to make sure that its quota is not already
6272 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
6273 * runtime as update_curr() throttling can not trigger until it's on-rq.
6275 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
6277 if (!cfs_bandwidth_used())
6280 /* an active group must be handled by the update_curr()->put() path */
6281 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
6284 /* ensure the group is not already throttled */
6285 if (cfs_rq_throttled(cfs_rq))
6288 /* update runtime allocation */
6289 account_cfs_rq_runtime(cfs_rq, 0);
6290 if (cfs_rq->runtime_remaining <= 0)
6291 throttle_cfs_rq(cfs_rq);
6294 static void sync_throttle(struct task_group *tg, int cpu)
6296 struct cfs_rq *pcfs_rq, *cfs_rq;
6298 if (!cfs_bandwidth_used())
6304 cfs_rq = tg->cfs_rq[cpu];
6305 pcfs_rq = tg->parent->cfs_rq[cpu];
6307 cfs_rq->throttle_count = pcfs_rq->throttle_count;
6308 cfs_rq->throttled_clock_pelt = rq_clock_pelt(cpu_rq(cpu));
6311 /* conditionally throttle active cfs_rq's from put_prev_entity() */
6312 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
6314 if (!cfs_bandwidth_used())
6317 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
6321 * it's possible for a throttled entity to be forced into a running
6322 * state (e.g. set_curr_task), in this case we're finished.
6324 if (cfs_rq_throttled(cfs_rq))
6327 return throttle_cfs_rq(cfs_rq);
6330 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
6332 struct cfs_bandwidth *cfs_b =
6333 container_of(timer, struct cfs_bandwidth, slack_timer);
6335 do_sched_cfs_slack_timer(cfs_b);
6337 return HRTIMER_NORESTART;
6340 extern const u64 max_cfs_quota_period;
6342 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
6344 struct cfs_bandwidth *cfs_b =
6345 container_of(timer, struct cfs_bandwidth, period_timer);
6346 unsigned long flags;
6351 raw_spin_lock_irqsave(&cfs_b->lock, flags);
6353 overrun = hrtimer_forward_now(timer, cfs_b->period);
6357 idle = do_sched_cfs_period_timer(cfs_b, overrun, flags);
6360 u64 new, old = ktime_to_ns(cfs_b->period);
6363 * Grow period by a factor of 2 to avoid losing precision.
6364 * Precision loss in the quota/period ratio can cause __cfs_schedulable
6368 if (new < max_cfs_quota_period) {
6369 cfs_b->period = ns_to_ktime(new);
6373 pr_warn_ratelimited(
6374 "cfs_period_timer[cpu%d]: period too short, scaling up (new cfs_period_us = %lld, cfs_quota_us = %lld)\n",
6376 div_u64(new, NSEC_PER_USEC),
6377 div_u64(cfs_b->quota, NSEC_PER_USEC));
6379 pr_warn_ratelimited(
6380 "cfs_period_timer[cpu%d]: period too short, but cannot scale up without losing precision (cfs_period_us = %lld, cfs_quota_us = %lld)\n",
6382 div_u64(old, NSEC_PER_USEC),
6383 div_u64(cfs_b->quota, NSEC_PER_USEC));
6386 /* reset count so we don't come right back in here */
6391 cfs_b->period_active = 0;
6392 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
6394 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
6397 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b, struct cfs_bandwidth *parent)
6399 raw_spin_lock_init(&cfs_b->lock);
6401 cfs_b->quota = RUNTIME_INF;
6402 cfs_b->period = ns_to_ktime(default_cfs_period());
6404 cfs_b->hierarchical_quota = parent ? parent->hierarchical_quota : RUNTIME_INF;
6406 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
6407 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED);
6408 cfs_b->period_timer.function = sched_cfs_period_timer;
6410 /* Add a random offset so that timers interleave */
6411 hrtimer_set_expires(&cfs_b->period_timer,
6412 get_random_u32_below(cfs_b->period));
6413 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
6414 cfs_b->slack_timer.function = sched_cfs_slack_timer;
6415 cfs_b->slack_started = false;
6418 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
6420 cfs_rq->runtime_enabled = 0;
6421 INIT_LIST_HEAD(&cfs_rq->throttled_list);
6422 INIT_LIST_HEAD(&cfs_rq->throttled_csd_list);
6425 void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
6427 lockdep_assert_held(&cfs_b->lock);
6429 if (cfs_b->period_active)
6432 cfs_b->period_active = 1;
6433 hrtimer_forward_now(&cfs_b->period_timer, cfs_b->period);
6434 hrtimer_start_expires(&cfs_b->period_timer, HRTIMER_MODE_ABS_PINNED);
6437 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
6439 int __maybe_unused i;
6441 /* init_cfs_bandwidth() was not called */
6442 if (!cfs_b->throttled_cfs_rq.next)
6445 hrtimer_cancel(&cfs_b->period_timer);
6446 hrtimer_cancel(&cfs_b->slack_timer);
6449 * It is possible that we still have some cfs_rq's pending on a CSD
6450 * list, though this race is very rare. In order for this to occur, we
6451 * must have raced with the last task leaving the group while there
6452 * exist throttled cfs_rq(s), and the period_timer must have queued the
6453 * CSD item but the remote cpu has not yet processed it. To handle this,
6454 * we can simply flush all pending CSD work inline here. We're
6455 * guaranteed at this point that no additional cfs_rq of this group can
6459 for_each_possible_cpu(i) {
6460 struct rq *rq = cpu_rq(i);
6461 unsigned long flags;
6463 if (list_empty(&rq->cfsb_csd_list))
6466 local_irq_save(flags);
6467 __cfsb_csd_unthrottle(rq);
6468 local_irq_restore(flags);
6474 * Both these CPU hotplug callbacks race against unregister_fair_sched_group()
6476 * The race is harmless, since modifying bandwidth settings of unhooked group
6477 * bits doesn't do much.
6480 /* cpu online callback */
6481 static void __maybe_unused update_runtime_enabled(struct rq *rq)
6483 struct task_group *tg;
6485 lockdep_assert_rq_held(rq);
6488 list_for_each_entry_rcu(tg, &task_groups, list) {
6489 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
6490 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
6492 raw_spin_lock(&cfs_b->lock);
6493 cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF;
6494 raw_spin_unlock(&cfs_b->lock);
6499 /* cpu offline callback */
6500 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
6502 struct task_group *tg;
6504 lockdep_assert_rq_held(rq);
6507 * The rq clock has already been updated in the
6508 * set_rq_offline(), so we should skip updating
6509 * the rq clock again in unthrottle_cfs_rq().
6511 rq_clock_start_loop_update(rq);
6514 list_for_each_entry_rcu(tg, &task_groups, list) {
6515 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
6517 if (!cfs_rq->runtime_enabled)
6521 * clock_task is not advancing so we just need to make sure
6522 * there's some valid quota amount
6524 cfs_rq->runtime_remaining = 1;
6526 * Offline rq is schedulable till CPU is completely disabled
6527 * in take_cpu_down(), so we prevent new cfs throttling here.
6529 cfs_rq->runtime_enabled = 0;
6531 if (cfs_rq_throttled(cfs_rq))
6532 unthrottle_cfs_rq(cfs_rq);
6536 rq_clock_stop_loop_update(rq);
6539 bool cfs_task_bw_constrained(struct task_struct *p)
6541 struct cfs_rq *cfs_rq = task_cfs_rq(p);
6543 if (!cfs_bandwidth_used())
6546 if (cfs_rq->runtime_enabled ||
6547 tg_cfs_bandwidth(cfs_rq->tg)->hierarchical_quota != RUNTIME_INF)
6553 #ifdef CONFIG_NO_HZ_FULL
6554 /* called from pick_next_task_fair() */
6555 static void sched_fair_update_stop_tick(struct rq *rq, struct task_struct *p)
6557 int cpu = cpu_of(rq);
6559 if (!sched_feat(HZ_BW) || !cfs_bandwidth_used())
6562 if (!tick_nohz_full_cpu(cpu))
6565 if (rq->nr_running != 1)
6569 * We know there is only one task runnable and we've just picked it. The
6570 * normal enqueue path will have cleared TICK_DEP_BIT_SCHED if we will
6571 * be otherwise able to stop the tick. Just need to check if we are using
6572 * bandwidth control.
6574 if (cfs_task_bw_constrained(p))
6575 tick_nohz_dep_set_cpu(cpu, TICK_DEP_BIT_SCHED);
6579 #else /* CONFIG_CFS_BANDWIDTH */
6581 static inline bool cfs_bandwidth_used(void)
6586 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {}
6587 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; }
6588 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
6589 static inline void sync_throttle(struct task_group *tg, int cpu) {}
6590 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
6592 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
6597 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
6602 static inline int throttled_lb_pair(struct task_group *tg,
6603 int src_cpu, int dest_cpu)
6608 #ifdef CONFIG_FAIR_GROUP_SCHED
6609 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b, struct cfs_bandwidth *parent) {}
6610 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
6613 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
6617 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
6618 static inline void update_runtime_enabled(struct rq *rq) {}
6619 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
6620 #ifdef CONFIG_CGROUP_SCHED
6621 bool cfs_task_bw_constrained(struct task_struct *p)
6626 #endif /* CONFIG_CFS_BANDWIDTH */
6628 #if !defined(CONFIG_CFS_BANDWIDTH) || !defined(CONFIG_NO_HZ_FULL)
6629 static inline void sched_fair_update_stop_tick(struct rq *rq, struct task_struct *p) {}
6632 /**************************************************
6633 * CFS operations on tasks:
6636 #ifdef CONFIG_SCHED_HRTICK
6637 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
6639 struct sched_entity *se = &p->se;
6641 SCHED_WARN_ON(task_rq(p) != rq);
6643 if (rq->cfs.h_nr_running > 1) {
6644 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
6645 u64 slice = se->slice;
6646 s64 delta = slice - ran;
6649 if (task_current(rq, p))
6653 hrtick_start(rq, delta);
6658 * called from enqueue/dequeue and updates the hrtick when the
6659 * current task is from our class and nr_running is low enough
6662 static void hrtick_update(struct rq *rq)
6664 struct task_struct *curr = rq->curr;
6666 if (!hrtick_enabled_fair(rq) || curr->sched_class != &fair_sched_class)
6669 hrtick_start_fair(rq, curr);
6671 #else /* !CONFIG_SCHED_HRTICK */
6673 hrtick_start_fair(struct rq *rq, struct task_struct *p)
6677 static inline void hrtick_update(struct rq *rq)
6683 static inline bool cpu_overutilized(int cpu)
6685 unsigned long rq_util_min, rq_util_max;
6687 if (!sched_energy_enabled())
6690 rq_util_min = uclamp_rq_get(cpu_rq(cpu), UCLAMP_MIN);
6691 rq_util_max = uclamp_rq_get(cpu_rq(cpu), UCLAMP_MAX);
6693 /* Return true only if the utilization doesn't fit CPU's capacity */
6694 return !util_fits_cpu(cpu_util_cfs(cpu), rq_util_min, rq_util_max, cpu);
6698 * overutilized value make sense only if EAS is enabled
6700 static inline bool is_rd_overutilized(struct root_domain *rd)
6702 return !sched_energy_enabled() || READ_ONCE(rd->overutilized);
6705 static inline void set_rd_overutilized(struct root_domain *rd, bool flag)
6707 if (!sched_energy_enabled())
6710 WRITE_ONCE(rd->overutilized, flag);
6711 trace_sched_overutilized_tp(rd, flag);
6714 static inline void check_update_overutilized_status(struct rq *rq)
6717 * overutilized field is used for load balancing decisions only
6718 * if energy aware scheduler is being used
6721 if (!is_rd_overutilized(rq->rd) && cpu_overutilized(rq->cpu))
6722 set_rd_overutilized(rq->rd, 1);
6725 static inline void check_update_overutilized_status(struct rq *rq) { }
6728 /* Runqueue only has SCHED_IDLE tasks enqueued */
6729 static int sched_idle_rq(struct rq *rq)
6731 return unlikely(rq->nr_running == rq->cfs.idle_h_nr_running &&
6736 static int sched_idle_cpu(int cpu)
6738 return sched_idle_rq(cpu_rq(cpu));
6743 * The enqueue_task method is called before nr_running is
6744 * increased. Here we update the fair scheduling stats and
6745 * then put the task into the rbtree:
6748 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
6750 struct cfs_rq *cfs_rq;
6751 struct sched_entity *se = &p->se;
6752 int idle_h_nr_running = task_has_idle_policy(p);
6753 int task_new = !(flags & ENQUEUE_WAKEUP);
6756 * The code below (indirectly) updates schedutil which looks at
6757 * the cfs_rq utilization to select a frequency.
6758 * Let's add the task's estimated utilization to the cfs_rq's
6759 * estimated utilization, before we update schedutil.
6761 util_est_enqueue(&rq->cfs, p);
6764 * If in_iowait is set, the code below may not trigger any cpufreq
6765 * utilization updates, so do it here explicitly with the IOWAIT flag
6769 cpufreq_update_util(rq, SCHED_CPUFREQ_IOWAIT);
6771 for_each_sched_entity(se) {
6774 cfs_rq = cfs_rq_of(se);
6775 enqueue_entity(cfs_rq, se, flags);
6777 cfs_rq->h_nr_running++;
6778 cfs_rq->idle_h_nr_running += idle_h_nr_running;
6780 if (cfs_rq_is_idle(cfs_rq))
6781 idle_h_nr_running = 1;
6783 /* end evaluation on encountering a throttled cfs_rq */
6784 if (cfs_rq_throttled(cfs_rq))
6785 goto enqueue_throttle;
6787 flags = ENQUEUE_WAKEUP;
6790 for_each_sched_entity(se) {
6791 cfs_rq = cfs_rq_of(se);
6793 update_load_avg(cfs_rq, se, UPDATE_TG);
6794 se_update_runnable(se);
6795 update_cfs_group(se);
6797 cfs_rq->h_nr_running++;
6798 cfs_rq->idle_h_nr_running += idle_h_nr_running;
6800 if (cfs_rq_is_idle(cfs_rq))
6801 idle_h_nr_running = 1;
6803 /* end evaluation on encountering a throttled cfs_rq */
6804 if (cfs_rq_throttled(cfs_rq))
6805 goto enqueue_throttle;
6808 /* At this point se is NULL and we are at root level*/
6809 add_nr_running(rq, 1);
6812 * Since new tasks are assigned an initial util_avg equal to
6813 * half of the spare capacity of their CPU, tiny tasks have the
6814 * ability to cross the overutilized threshold, which will
6815 * result in the load balancer ruining all the task placement
6816 * done by EAS. As a way to mitigate that effect, do not account
6817 * for the first enqueue operation of new tasks during the
6818 * overutilized flag detection.
6820 * A better way of solving this problem would be to wait for
6821 * the PELT signals of tasks to converge before taking them
6822 * into account, but that is not straightforward to implement,
6823 * and the following generally works well enough in practice.
6826 check_update_overutilized_status(rq);
6829 assert_list_leaf_cfs_rq(rq);
6834 static void set_next_buddy(struct sched_entity *se);
6837 * The dequeue_task method is called before nr_running is
6838 * decreased. We remove the task from the rbtree and
6839 * update the fair scheduling stats:
6841 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
6843 struct cfs_rq *cfs_rq;
6844 struct sched_entity *se = &p->se;
6845 int task_sleep = flags & DEQUEUE_SLEEP;
6846 int idle_h_nr_running = task_has_idle_policy(p);
6847 bool was_sched_idle = sched_idle_rq(rq);
6849 util_est_dequeue(&rq->cfs, p);
6851 for_each_sched_entity(se) {
6852 cfs_rq = cfs_rq_of(se);
6853 dequeue_entity(cfs_rq, se, flags);
6855 cfs_rq->h_nr_running--;
6856 cfs_rq->idle_h_nr_running -= idle_h_nr_running;
6858 if (cfs_rq_is_idle(cfs_rq))
6859 idle_h_nr_running = 1;
6861 /* end evaluation on encountering a throttled cfs_rq */
6862 if (cfs_rq_throttled(cfs_rq))
6863 goto dequeue_throttle;
6865 /* Don't dequeue parent if it has other entities besides us */
6866 if (cfs_rq->load.weight) {
6867 /* Avoid re-evaluating load for this entity: */
6868 se = parent_entity(se);
6870 * Bias pick_next to pick a task from this cfs_rq, as
6871 * p is sleeping when it is within its sched_slice.
6873 if (task_sleep && se && !throttled_hierarchy(cfs_rq))
6877 flags |= DEQUEUE_SLEEP;
6880 for_each_sched_entity(se) {
6881 cfs_rq = cfs_rq_of(se);
6883 update_load_avg(cfs_rq, se, UPDATE_TG);
6884 se_update_runnable(se);
6885 update_cfs_group(se);
6887 cfs_rq->h_nr_running--;
6888 cfs_rq->idle_h_nr_running -= idle_h_nr_running;
6890 if (cfs_rq_is_idle(cfs_rq))
6891 idle_h_nr_running = 1;
6893 /* end evaluation on encountering a throttled cfs_rq */
6894 if (cfs_rq_throttled(cfs_rq))
6895 goto dequeue_throttle;
6899 /* At this point se is NULL and we are at root level*/
6900 sub_nr_running(rq, 1);
6902 /* balance early to pull high priority tasks */
6903 if (unlikely(!was_sched_idle && sched_idle_rq(rq)))
6904 rq->next_balance = jiffies;
6907 util_est_update(&rq->cfs, p, task_sleep);
6913 /* Working cpumask for: sched_balance_rq(), sched_balance_newidle(). */
6914 static DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
6915 static DEFINE_PER_CPU(cpumask_var_t, select_rq_mask);
6916 static DEFINE_PER_CPU(cpumask_var_t, should_we_balance_tmpmask);
6918 #ifdef CONFIG_NO_HZ_COMMON
6921 cpumask_var_t idle_cpus_mask;
6923 int has_blocked; /* Idle CPUS has blocked load */
6924 int needs_update; /* Newly idle CPUs need their next_balance collated */
6925 unsigned long next_balance; /* in jiffy units */
6926 unsigned long next_blocked; /* Next update of blocked load in jiffies */
6927 } nohz ____cacheline_aligned;
6929 #endif /* CONFIG_NO_HZ_COMMON */
6931 static unsigned long cpu_load(struct rq *rq)
6933 return cfs_rq_load_avg(&rq->cfs);
6937 * cpu_load_without - compute CPU load without any contributions from *p
6938 * @cpu: the CPU which load is requested
6939 * @p: the task which load should be discounted
6941 * The load of a CPU is defined by the load of tasks currently enqueued on that
6942 * CPU as well as tasks which are currently sleeping after an execution on that
6945 * This method returns the load of the specified CPU by discounting the load of
6946 * the specified task, whenever the task is currently contributing to the CPU
6949 static unsigned long cpu_load_without(struct rq *rq, struct task_struct *p)
6951 struct cfs_rq *cfs_rq;
6954 /* Task has no contribution or is new */
6955 if (cpu_of(rq) != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time))
6956 return cpu_load(rq);
6959 load = READ_ONCE(cfs_rq->avg.load_avg);
6961 /* Discount task's util from CPU's util */
6962 lsub_positive(&load, task_h_load(p));
6967 static unsigned long cpu_runnable(struct rq *rq)
6969 return cfs_rq_runnable_avg(&rq->cfs);
6972 static unsigned long cpu_runnable_without(struct rq *rq, struct task_struct *p)
6974 struct cfs_rq *cfs_rq;
6975 unsigned int runnable;
6977 /* Task has no contribution or is new */
6978 if (cpu_of(rq) != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time))
6979 return cpu_runnable(rq);
6982 runnable = READ_ONCE(cfs_rq->avg.runnable_avg);
6984 /* Discount task's runnable from CPU's runnable */
6985 lsub_positive(&runnable, p->se.avg.runnable_avg);
6990 static unsigned long capacity_of(int cpu)
6992 return cpu_rq(cpu)->cpu_capacity;
6995 static void record_wakee(struct task_struct *p)
6998 * Only decay a single time; tasks that have less then 1 wakeup per
6999 * jiffy will not have built up many flips.
7001 if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) {
7002 current->wakee_flips >>= 1;
7003 current->wakee_flip_decay_ts = jiffies;
7006 if (current->last_wakee != p) {
7007 current->last_wakee = p;
7008 current->wakee_flips++;
7013 * Detect M:N waker/wakee relationships via a switching-frequency heuristic.
7015 * A waker of many should wake a different task than the one last awakened
7016 * at a frequency roughly N times higher than one of its wakees.
7018 * In order to determine whether we should let the load spread vs consolidating
7019 * to shared cache, we look for a minimum 'flip' frequency of llc_size in one
7020 * partner, and a factor of lls_size higher frequency in the other.
7022 * With both conditions met, we can be relatively sure that the relationship is
7023 * non-monogamous, with partner count exceeding socket size.
7025 * Waker/wakee being client/server, worker/dispatcher, interrupt source or
7026 * whatever is irrelevant, spread criteria is apparent partner count exceeds
7029 static int wake_wide(struct task_struct *p)
7031 unsigned int master = current->wakee_flips;
7032 unsigned int slave = p->wakee_flips;
7033 int factor = __this_cpu_read(sd_llc_size);
7036 swap(master, slave);
7037 if (slave < factor || master < slave * factor)
7043 * The purpose of wake_affine() is to quickly determine on which CPU we can run
7044 * soonest. For the purpose of speed we only consider the waking and previous
7047 * wake_affine_idle() - only considers 'now', it check if the waking CPU is
7048 * cache-affine and is (or will be) idle.
7050 * wake_affine_weight() - considers the weight to reflect the average
7051 * scheduling latency of the CPUs. This seems to work
7052 * for the overloaded case.
7055 wake_affine_idle(int this_cpu, int prev_cpu, int sync)
7058 * If this_cpu is idle, it implies the wakeup is from interrupt
7059 * context. Only allow the move if cache is shared. Otherwise an
7060 * interrupt intensive workload could force all tasks onto one
7061 * node depending on the IO topology or IRQ affinity settings.
7063 * If the prev_cpu is idle and cache affine then avoid a migration.
7064 * There is no guarantee that the cache hot data from an interrupt
7065 * is more important than cache hot data on the prev_cpu and from
7066 * a cpufreq perspective, it's better to have higher utilisation
7069 if (available_idle_cpu(this_cpu) && cpus_share_cache(this_cpu, prev_cpu))
7070 return available_idle_cpu(prev_cpu) ? prev_cpu : this_cpu;
7072 if (sync && cpu_rq(this_cpu)->nr_running == 1)
7075 if (available_idle_cpu(prev_cpu))
7078 return nr_cpumask_bits;
7082 wake_affine_weight(struct sched_domain *sd, struct task_struct *p,
7083 int this_cpu, int prev_cpu, int sync)
7085 s64 this_eff_load, prev_eff_load;
7086 unsigned long task_load;
7088 this_eff_load = cpu_load(cpu_rq(this_cpu));
7091 unsigned long current_load = task_h_load(current);
7093 if (current_load > this_eff_load)
7096 this_eff_load -= current_load;
7099 task_load = task_h_load(p);
7101 this_eff_load += task_load;
7102 if (sched_feat(WA_BIAS))
7103 this_eff_load *= 100;
7104 this_eff_load *= capacity_of(prev_cpu);
7106 prev_eff_load = cpu_load(cpu_rq(prev_cpu));
7107 prev_eff_load -= task_load;
7108 if (sched_feat(WA_BIAS))
7109 prev_eff_load *= 100 + (sd->imbalance_pct - 100) / 2;
7110 prev_eff_load *= capacity_of(this_cpu);
7113 * If sync, adjust the weight of prev_eff_load such that if
7114 * prev_eff == this_eff that select_idle_sibling() will consider
7115 * stacking the wakee on top of the waker if no other CPU is
7121 return this_eff_load < prev_eff_load ? this_cpu : nr_cpumask_bits;
7124 static int wake_affine(struct sched_domain *sd, struct task_struct *p,
7125 int this_cpu, int prev_cpu, int sync)
7127 int target = nr_cpumask_bits;
7129 if (sched_feat(WA_IDLE))
7130 target = wake_affine_idle(this_cpu, prev_cpu, sync);
7132 if (sched_feat(WA_WEIGHT) && target == nr_cpumask_bits)
7133 target = wake_affine_weight(sd, p, this_cpu, prev_cpu, sync);
7135 schedstat_inc(p->stats.nr_wakeups_affine_attempts);
7136 if (target != this_cpu)
7139 schedstat_inc(sd->ttwu_move_affine);
7140 schedstat_inc(p->stats.nr_wakeups_affine);
7144 static struct sched_group *
7145 sched_balance_find_dst_group(struct sched_domain *sd, struct task_struct *p, int this_cpu);
7148 * sched_balance_find_dst_group_cpu - find the idlest CPU among the CPUs in the group.
7151 sched_balance_find_dst_group_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
7153 unsigned long load, min_load = ULONG_MAX;
7154 unsigned int min_exit_latency = UINT_MAX;
7155 u64 latest_idle_timestamp = 0;
7156 int least_loaded_cpu = this_cpu;
7157 int shallowest_idle_cpu = -1;
7160 /* Check if we have any choice: */
7161 if (group->group_weight == 1)
7162 return cpumask_first(sched_group_span(group));
7164 /* Traverse only the allowed CPUs */
7165 for_each_cpu_and(i, sched_group_span(group), p->cpus_ptr) {
7166 struct rq *rq = cpu_rq(i);
7168 if (!sched_core_cookie_match(rq, p))
7171 if (sched_idle_cpu(i))
7174 if (available_idle_cpu(i)) {
7175 struct cpuidle_state *idle = idle_get_state(rq);
7176 if (idle && idle->exit_latency < min_exit_latency) {
7178 * We give priority to a CPU whose idle state
7179 * has the smallest exit latency irrespective
7180 * of any idle timestamp.
7182 min_exit_latency = idle->exit_latency;
7183 latest_idle_timestamp = rq->idle_stamp;
7184 shallowest_idle_cpu = i;
7185 } else if ((!idle || idle->exit_latency == min_exit_latency) &&
7186 rq->idle_stamp > latest_idle_timestamp) {
7188 * If equal or no active idle state, then
7189 * the most recently idled CPU might have
7192 latest_idle_timestamp = rq->idle_stamp;
7193 shallowest_idle_cpu = i;
7195 } else if (shallowest_idle_cpu == -1) {
7196 load = cpu_load(cpu_rq(i));
7197 if (load < min_load) {
7199 least_loaded_cpu = i;
7204 return shallowest_idle_cpu != -1 ? shallowest_idle_cpu : least_loaded_cpu;
7207 static inline int sched_balance_find_dst_cpu(struct sched_domain *sd, struct task_struct *p,
7208 int cpu, int prev_cpu, int sd_flag)
7212 if (!cpumask_intersects(sched_domain_span(sd), p->cpus_ptr))
7216 * We need task's util for cpu_util_without, sync it up to
7217 * prev_cpu's last_update_time.
7219 if (!(sd_flag & SD_BALANCE_FORK))
7220 sync_entity_load_avg(&p->se);
7223 struct sched_group *group;
7224 struct sched_domain *tmp;
7227 if (!(sd->flags & sd_flag)) {
7232 group = sched_balance_find_dst_group(sd, p, cpu);
7238 new_cpu = sched_balance_find_dst_group_cpu(group, p, cpu);
7239 if (new_cpu == cpu) {
7240 /* Now try balancing at a lower domain level of 'cpu': */
7245 /* Now try balancing at a lower domain level of 'new_cpu': */
7247 weight = sd->span_weight;
7249 for_each_domain(cpu, tmp) {
7250 if (weight <= tmp->span_weight)
7252 if (tmp->flags & sd_flag)
7260 static inline int __select_idle_cpu(int cpu, struct task_struct *p)
7262 if ((available_idle_cpu(cpu) || sched_idle_cpu(cpu)) &&
7263 sched_cpu_cookie_match(cpu_rq(cpu), p))
7269 #ifdef CONFIG_SCHED_SMT
7270 DEFINE_STATIC_KEY_FALSE(sched_smt_present);
7271 EXPORT_SYMBOL_GPL(sched_smt_present);
7273 static inline void set_idle_cores(int cpu, int val)
7275 struct sched_domain_shared *sds;
7277 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
7279 WRITE_ONCE(sds->has_idle_cores, val);
7282 static inline bool test_idle_cores(int cpu)
7284 struct sched_domain_shared *sds;
7286 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
7288 return READ_ONCE(sds->has_idle_cores);
7294 * Scans the local SMT mask to see if the entire core is idle, and records this
7295 * information in sd_llc_shared->has_idle_cores.
7297 * Since SMT siblings share all cache levels, inspecting this limited remote
7298 * state should be fairly cheap.
7300 void __update_idle_core(struct rq *rq)
7302 int core = cpu_of(rq);
7306 if (test_idle_cores(core))
7309 for_each_cpu(cpu, cpu_smt_mask(core)) {
7313 if (!available_idle_cpu(cpu))
7317 set_idle_cores(core, 1);
7323 * Scan the entire LLC domain for idle cores; this dynamically switches off if
7324 * there are no idle cores left in the system; tracked through
7325 * sd_llc->shared->has_idle_cores and enabled through update_idle_core() above.
7327 static int select_idle_core(struct task_struct *p, int core, struct cpumask *cpus, int *idle_cpu)
7332 for_each_cpu(cpu, cpu_smt_mask(core)) {
7333 if (!available_idle_cpu(cpu)) {
7335 if (*idle_cpu == -1) {
7336 if (sched_idle_cpu(cpu) && cpumask_test_cpu(cpu, cpus)) {
7344 if (*idle_cpu == -1 && cpumask_test_cpu(cpu, cpus))
7351 cpumask_andnot(cpus, cpus, cpu_smt_mask(core));
7356 * Scan the local SMT mask for idle CPUs.
7358 static int select_idle_smt(struct task_struct *p, struct sched_domain *sd, int target)
7362 for_each_cpu_and(cpu, cpu_smt_mask(target), p->cpus_ptr) {
7366 * Check if the CPU is in the LLC scheduling domain of @target.
7367 * Due to isolcpus, there is no guarantee that all the siblings are in the domain.
7369 if (!cpumask_test_cpu(cpu, sched_domain_span(sd)))
7371 if (available_idle_cpu(cpu) || sched_idle_cpu(cpu))
7378 #else /* CONFIG_SCHED_SMT */
7380 static inline void set_idle_cores(int cpu, int val)
7384 static inline bool test_idle_cores(int cpu)
7389 static inline int select_idle_core(struct task_struct *p, int core, struct cpumask *cpus, int *idle_cpu)
7391 return __select_idle_cpu(core, p);
7394 static inline int select_idle_smt(struct task_struct *p, struct sched_domain *sd, int target)
7399 #endif /* CONFIG_SCHED_SMT */
7402 * Scan the LLC domain for idle CPUs; this is dynamically regulated by
7403 * comparing the average scan cost (tracked in sd->avg_scan_cost) against the
7404 * average idle time for this rq (as found in rq->avg_idle).
7406 static int select_idle_cpu(struct task_struct *p, struct sched_domain *sd, bool has_idle_core, int target)
7408 struct cpumask *cpus = this_cpu_cpumask_var_ptr(select_rq_mask);
7409 int i, cpu, idle_cpu = -1, nr = INT_MAX;
7410 struct sched_domain_shared *sd_share;
7412 cpumask_and(cpus, sched_domain_span(sd), p->cpus_ptr);
7414 if (sched_feat(SIS_UTIL)) {
7415 sd_share = rcu_dereference(per_cpu(sd_llc_shared, target));
7417 /* because !--nr is the condition to stop scan */
7418 nr = READ_ONCE(sd_share->nr_idle_scan) + 1;
7419 /* overloaded LLC is unlikely to have idle cpu/core */
7425 if (static_branch_unlikely(&sched_cluster_active)) {
7426 struct sched_group *sg = sd->groups;
7428 if (sg->flags & SD_CLUSTER) {
7429 for_each_cpu_wrap(cpu, sched_group_span(sg), target + 1) {
7430 if (!cpumask_test_cpu(cpu, cpus))
7433 if (has_idle_core) {
7434 i = select_idle_core(p, cpu, cpus, &idle_cpu);
7435 if ((unsigned int)i < nr_cpumask_bits)
7440 idle_cpu = __select_idle_cpu(cpu, p);
7441 if ((unsigned int)idle_cpu < nr_cpumask_bits)
7445 cpumask_andnot(cpus, cpus, sched_group_span(sg));
7449 for_each_cpu_wrap(cpu, cpus, target + 1) {
7450 if (has_idle_core) {
7451 i = select_idle_core(p, cpu, cpus, &idle_cpu);
7452 if ((unsigned int)i < nr_cpumask_bits)
7458 idle_cpu = __select_idle_cpu(cpu, p);
7459 if ((unsigned int)idle_cpu < nr_cpumask_bits)
7465 set_idle_cores(target, false);
7471 * Scan the asym_capacity domain for idle CPUs; pick the first idle one on which
7472 * the task fits. If no CPU is big enough, but there are idle ones, try to
7473 * maximize capacity.
7476 select_idle_capacity(struct task_struct *p, struct sched_domain *sd, int target)
7478 unsigned long task_util, util_min, util_max, best_cap = 0;
7479 int fits, best_fits = 0;
7480 int cpu, best_cpu = -1;
7481 struct cpumask *cpus;
7483 cpus = this_cpu_cpumask_var_ptr(select_rq_mask);
7484 cpumask_and(cpus, sched_domain_span(sd), p->cpus_ptr);
7486 task_util = task_util_est(p);
7487 util_min = uclamp_eff_value(p, UCLAMP_MIN);
7488 util_max = uclamp_eff_value(p, UCLAMP_MAX);
7490 for_each_cpu_wrap(cpu, cpus, target) {
7491 unsigned long cpu_cap = capacity_of(cpu);
7493 if (!available_idle_cpu(cpu) && !sched_idle_cpu(cpu))
7496 fits = util_fits_cpu(task_util, util_min, util_max, cpu);
7498 /* This CPU fits with all requirements */
7502 * Only the min performance hint (i.e. uclamp_min) doesn't fit.
7503 * Look for the CPU with best capacity.
7506 cpu_cap = get_actual_cpu_capacity(cpu);
7509 * First, select CPU which fits better (-1 being better than 0).
7510 * Then, select the one with best capacity at same level.
7512 if ((fits < best_fits) ||
7513 ((fits == best_fits) && (cpu_cap > best_cap))) {
7523 static inline bool asym_fits_cpu(unsigned long util,
7524 unsigned long util_min,
7525 unsigned long util_max,
7528 if (sched_asym_cpucap_active())
7530 * Return true only if the cpu fully fits the task requirements
7531 * which include the utilization and the performance hints.
7533 return (util_fits_cpu(util, util_min, util_max, cpu) > 0);
7539 * Try and locate an idle core/thread in the LLC cache domain.
7541 static int select_idle_sibling(struct task_struct *p, int prev, int target)
7543 bool has_idle_core = false;
7544 struct sched_domain *sd;
7545 unsigned long task_util, util_min, util_max;
7546 int i, recent_used_cpu, prev_aff = -1;
7549 * On asymmetric system, update task utilization because we will check
7550 * that the task fits with CPU's capacity.
7552 if (sched_asym_cpucap_active()) {
7553 sync_entity_load_avg(&p->se);
7554 task_util = task_util_est(p);
7555 util_min = uclamp_eff_value(p, UCLAMP_MIN);
7556 util_max = uclamp_eff_value(p, UCLAMP_MAX);
7560 * per-cpu select_rq_mask usage
7562 lockdep_assert_irqs_disabled();
7564 if ((available_idle_cpu(target) || sched_idle_cpu(target)) &&
7565 asym_fits_cpu(task_util, util_min, util_max, target))
7569 * If the previous CPU is cache affine and idle, don't be stupid:
7571 if (prev != target && cpus_share_cache(prev, target) &&
7572 (available_idle_cpu(prev) || sched_idle_cpu(prev)) &&
7573 asym_fits_cpu(task_util, util_min, util_max, prev)) {
7575 if (!static_branch_unlikely(&sched_cluster_active) ||
7576 cpus_share_resources(prev, target))
7583 * Allow a per-cpu kthread to stack with the wakee if the
7584 * kworker thread and the tasks previous CPUs are the same.
7585 * The assumption is that the wakee queued work for the
7586 * per-cpu kthread that is now complete and the wakeup is
7587 * essentially a sync wakeup. An obvious example of this
7588 * pattern is IO completions.
7590 if (is_per_cpu_kthread(current) &&
7592 prev == smp_processor_id() &&
7593 this_rq()->nr_running <= 1 &&
7594 asym_fits_cpu(task_util, util_min, util_max, prev)) {
7598 /* Check a recently used CPU as a potential idle candidate: */
7599 recent_used_cpu = p->recent_used_cpu;
7600 p->recent_used_cpu = prev;
7601 if (recent_used_cpu != prev &&
7602 recent_used_cpu != target &&
7603 cpus_share_cache(recent_used_cpu, target) &&
7604 (available_idle_cpu(recent_used_cpu) || sched_idle_cpu(recent_used_cpu)) &&
7605 cpumask_test_cpu(recent_used_cpu, p->cpus_ptr) &&
7606 asym_fits_cpu(task_util, util_min, util_max, recent_used_cpu)) {
7608 if (!static_branch_unlikely(&sched_cluster_active) ||
7609 cpus_share_resources(recent_used_cpu, target))
7610 return recent_used_cpu;
7613 recent_used_cpu = -1;
7617 * For asymmetric CPU capacity systems, our domain of interest is
7618 * sd_asym_cpucapacity rather than sd_llc.
7620 if (sched_asym_cpucap_active()) {
7621 sd = rcu_dereference(per_cpu(sd_asym_cpucapacity, target));
7623 * On an asymmetric CPU capacity system where an exclusive
7624 * cpuset defines a symmetric island (i.e. one unique
7625 * capacity_orig value through the cpuset), the key will be set
7626 * but the CPUs within that cpuset will not have a domain with
7627 * SD_ASYM_CPUCAPACITY. These should follow the usual symmetric
7631 i = select_idle_capacity(p, sd, target);
7632 return ((unsigned)i < nr_cpumask_bits) ? i : target;
7636 sd = rcu_dereference(per_cpu(sd_llc, target));
7640 if (sched_smt_active()) {
7641 has_idle_core = test_idle_cores(target);
7643 if (!has_idle_core && cpus_share_cache(prev, target)) {
7644 i = select_idle_smt(p, sd, prev);
7645 if ((unsigned int)i < nr_cpumask_bits)
7650 i = select_idle_cpu(p, sd, has_idle_core, target);
7651 if ((unsigned)i < nr_cpumask_bits)
7655 * For cluster machines which have lower sharing cache like L2 or
7656 * LLC Tag, we tend to find an idle CPU in the target's cluster
7657 * first. But prev_cpu or recent_used_cpu may also be a good candidate,
7658 * use them if possible when no idle CPU found in select_idle_cpu().
7660 if ((unsigned int)prev_aff < nr_cpumask_bits)
7662 if ((unsigned int)recent_used_cpu < nr_cpumask_bits)
7663 return recent_used_cpu;
7669 * cpu_util() - Estimates the amount of CPU capacity used by CFS tasks.
7670 * @cpu: the CPU to get the utilization for
7671 * @p: task for which the CPU utilization should be predicted or NULL
7672 * @dst_cpu: CPU @p migrates to, -1 if @p moves from @cpu or @p == NULL
7673 * @boost: 1 to enable boosting, otherwise 0
7675 * The unit of the return value must be the same as the one of CPU capacity
7676 * so that CPU utilization can be compared with CPU capacity.
7678 * CPU utilization is the sum of running time of runnable tasks plus the
7679 * recent utilization of currently non-runnable tasks on that CPU.
7680 * It represents the amount of CPU capacity currently used by CFS tasks in
7681 * the range [0..max CPU capacity] with max CPU capacity being the CPU
7682 * capacity at f_max.
7684 * The estimated CPU utilization is defined as the maximum between CPU
7685 * utilization and sum of the estimated utilization of the currently
7686 * runnable tasks on that CPU. It preserves a utilization "snapshot" of
7687 * previously-executed tasks, which helps better deduce how busy a CPU will
7688 * be when a long-sleeping task wakes up. The contribution to CPU utilization
7689 * of such a task would be significantly decayed at this point of time.
7691 * Boosted CPU utilization is defined as max(CPU runnable, CPU utilization).
7692 * CPU contention for CFS tasks can be detected by CPU runnable > CPU
7693 * utilization. Boosting is implemented in cpu_util() so that internal
7694 * users (e.g. EAS) can use it next to external users (e.g. schedutil),
7695 * latter via cpu_util_cfs_boost().
7697 * CPU utilization can be higher than the current CPU capacity
7698 * (f_curr/f_max * max CPU capacity) or even the max CPU capacity because
7699 * of rounding errors as well as task migrations or wakeups of new tasks.
7700 * CPU utilization has to be capped to fit into the [0..max CPU capacity]
7701 * range. Otherwise a group of CPUs (CPU0 util = 121% + CPU1 util = 80%)
7702 * could be seen as over-utilized even though CPU1 has 20% of spare CPU
7703 * capacity. CPU utilization is allowed to overshoot current CPU capacity
7704 * though since this is useful for predicting the CPU capacity required
7705 * after task migrations (scheduler-driven DVFS).
7707 * Return: (Boosted) (estimated) utilization for the specified CPU.
7709 static unsigned long
7710 cpu_util(int cpu, struct task_struct *p, int dst_cpu, int boost)
7712 struct cfs_rq *cfs_rq = &cpu_rq(cpu)->cfs;
7713 unsigned long util = READ_ONCE(cfs_rq->avg.util_avg);
7714 unsigned long runnable;
7717 runnable = READ_ONCE(cfs_rq->avg.runnable_avg);
7718 util = max(util, runnable);
7722 * If @dst_cpu is -1 or @p migrates from @cpu to @dst_cpu remove its
7723 * contribution. If @p migrates from another CPU to @cpu add its
7724 * contribution. In all the other cases @cpu is not impacted by the
7725 * migration so its util_avg is already correct.
7727 if (p && task_cpu(p) == cpu && dst_cpu != cpu)
7728 lsub_positive(&util, task_util(p));
7729 else if (p && task_cpu(p) != cpu && dst_cpu == cpu)
7730 util += task_util(p);
7732 if (sched_feat(UTIL_EST)) {
7733 unsigned long util_est;
7735 util_est = READ_ONCE(cfs_rq->avg.util_est);
7738 * During wake-up @p isn't enqueued yet and doesn't contribute
7739 * to any cpu_rq(cpu)->cfs.avg.util_est.
7740 * If @dst_cpu == @cpu add it to "simulate" cpu_util after @p
7741 * has been enqueued.
7743 * During exec (@dst_cpu = -1) @p is enqueued and does
7744 * contribute to cpu_rq(cpu)->cfs.util_est.
7745 * Remove it to "simulate" cpu_util without @p's contribution.
7747 * Despite the task_on_rq_queued(@p) check there is still a
7748 * small window for a possible race when an exec
7749 * select_task_rq_fair() races with LB's detach_task().
7753 * p->on_rq = TASK_ON_RQ_MIGRATING;
7754 * -------------------------------- A
7756 * dequeue_task_fair() + Race Time
7757 * util_est_dequeue() /
7758 * -------------------------------- B
7760 * The additional check "current == p" is required to further
7761 * reduce the race window.
7764 util_est += _task_util_est(p);
7765 else if (p && unlikely(task_on_rq_queued(p) || current == p))
7766 lsub_positive(&util_est, _task_util_est(p));
7768 util = max(util, util_est);
7771 return min(util, arch_scale_cpu_capacity(cpu));
7774 unsigned long cpu_util_cfs(int cpu)
7776 return cpu_util(cpu, NULL, -1, 0);
7779 unsigned long cpu_util_cfs_boost(int cpu)
7781 return cpu_util(cpu, NULL, -1, 1);
7785 * cpu_util_without: compute cpu utilization without any contributions from *p
7786 * @cpu: the CPU which utilization is requested
7787 * @p: the task which utilization should be discounted
7789 * The utilization of a CPU is defined by the utilization of tasks currently
7790 * enqueued on that CPU as well as tasks which are currently sleeping after an
7791 * execution on that CPU.
7793 * This method returns the utilization of the specified CPU by discounting the
7794 * utilization of the specified task, whenever the task is currently
7795 * contributing to the CPU utilization.
7797 static unsigned long cpu_util_without(int cpu, struct task_struct *p)
7799 /* Task has no contribution or is new */
7800 if (cpu != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time))
7803 return cpu_util(cpu, p, -1, 0);
7807 * energy_env - Utilization landscape for energy estimation.
7808 * @task_busy_time: Utilization contribution by the task for which we test the
7809 * placement. Given by eenv_task_busy_time().
7810 * @pd_busy_time: Utilization of the whole perf domain without the task
7811 * contribution. Given by eenv_pd_busy_time().
7812 * @cpu_cap: Maximum CPU capacity for the perf domain.
7813 * @pd_cap: Entire perf domain capacity. (pd->nr_cpus * cpu_cap).
7816 unsigned long task_busy_time;
7817 unsigned long pd_busy_time;
7818 unsigned long cpu_cap;
7819 unsigned long pd_cap;
7823 * Compute the task busy time for compute_energy(). This time cannot be
7824 * injected directly into effective_cpu_util() because of the IRQ scaling.
7825 * The latter only makes sense with the most recent CPUs where the task has
7828 static inline void eenv_task_busy_time(struct energy_env *eenv,
7829 struct task_struct *p, int prev_cpu)
7831 unsigned long busy_time, max_cap = arch_scale_cpu_capacity(prev_cpu);
7832 unsigned long irq = cpu_util_irq(cpu_rq(prev_cpu));
7834 if (unlikely(irq >= max_cap))
7835 busy_time = max_cap;
7837 busy_time = scale_irq_capacity(task_util_est(p), irq, max_cap);
7839 eenv->task_busy_time = busy_time;
7843 * Compute the perf_domain (PD) busy time for compute_energy(). Based on the
7844 * utilization for each @pd_cpus, it however doesn't take into account
7845 * clamping since the ratio (utilization / cpu_capacity) is already enough to
7846 * scale the EM reported power consumption at the (eventually clamped)
7849 * The contribution of the task @p for which we want to estimate the
7850 * energy cost is removed (by cpu_util()) and must be calculated
7851 * separately (see eenv_task_busy_time). This ensures:
7853 * - A stable PD utilization, no matter which CPU of that PD we want to place
7856 * - A fair comparison between CPUs as the task contribution (task_util())
7857 * will always be the same no matter which CPU utilization we rely on
7858 * (util_avg or util_est).
7860 * Set @eenv busy time for the PD that spans @pd_cpus. This busy time can't
7861 * exceed @eenv->pd_cap.
7863 static inline void eenv_pd_busy_time(struct energy_env *eenv,
7864 struct cpumask *pd_cpus,
7865 struct task_struct *p)
7867 unsigned long busy_time = 0;
7870 for_each_cpu(cpu, pd_cpus) {
7871 unsigned long util = cpu_util(cpu, p, -1, 0);
7873 busy_time += effective_cpu_util(cpu, util, NULL, NULL);
7876 eenv->pd_busy_time = min(eenv->pd_cap, busy_time);
7880 * Compute the maximum utilization for compute_energy() when the task @p
7881 * is placed on the cpu @dst_cpu.
7883 * Returns the maximum utilization among @eenv->cpus. This utilization can't
7884 * exceed @eenv->cpu_cap.
7886 static inline unsigned long
7887 eenv_pd_max_util(struct energy_env *eenv, struct cpumask *pd_cpus,
7888 struct task_struct *p, int dst_cpu)
7890 unsigned long max_util = 0;
7893 for_each_cpu(cpu, pd_cpus) {
7894 struct task_struct *tsk = (cpu == dst_cpu) ? p : NULL;
7895 unsigned long util = cpu_util(cpu, p, dst_cpu, 1);
7896 unsigned long eff_util, min, max;
7899 * Performance domain frequency: utilization clamping
7900 * must be considered since it affects the selection
7901 * of the performance domain frequency.
7902 * NOTE: in case RT tasks are running, by default the min
7903 * utilization can be max OPP.
7905 eff_util = effective_cpu_util(cpu, util, &min, &max);
7907 /* Task's uclamp can modify min and max value */
7908 if (tsk && uclamp_is_used()) {
7909 min = max(min, uclamp_eff_value(p, UCLAMP_MIN));
7912 * If there is no active max uclamp constraint,
7913 * directly use task's one, otherwise keep max.
7915 if (uclamp_rq_is_idle(cpu_rq(cpu)))
7916 max = uclamp_eff_value(p, UCLAMP_MAX);
7918 max = max(max, uclamp_eff_value(p, UCLAMP_MAX));
7921 eff_util = sugov_effective_cpu_perf(cpu, eff_util, min, max);
7922 max_util = max(max_util, eff_util);
7925 return min(max_util, eenv->cpu_cap);
7929 * compute_energy(): Use the Energy Model to estimate the energy that @pd would
7930 * consume for a given utilization landscape @eenv. When @dst_cpu < 0, the task
7931 * contribution is ignored.
7933 static inline unsigned long
7934 compute_energy(struct energy_env *eenv, struct perf_domain *pd,
7935 struct cpumask *pd_cpus, struct task_struct *p, int dst_cpu)
7937 unsigned long max_util = eenv_pd_max_util(eenv, pd_cpus, p, dst_cpu);
7938 unsigned long busy_time = eenv->pd_busy_time;
7939 unsigned long energy;
7942 busy_time = min(eenv->pd_cap, busy_time + eenv->task_busy_time);
7944 energy = em_cpu_energy(pd->em_pd, max_util, busy_time, eenv->cpu_cap);
7946 trace_sched_compute_energy_tp(p, dst_cpu, energy, max_util, busy_time);
7952 * find_energy_efficient_cpu(): Find most energy-efficient target CPU for the
7953 * waking task. find_energy_efficient_cpu() looks for the CPU with maximum
7954 * spare capacity in each performance domain and uses it as a potential
7955 * candidate to execute the task. Then, it uses the Energy Model to figure
7956 * out which of the CPU candidates is the most energy-efficient.
7958 * The rationale for this heuristic is as follows. In a performance domain,
7959 * all the most energy efficient CPU candidates (according to the Energy
7960 * Model) are those for which we'll request a low frequency. When there are
7961 * several CPUs for which the frequency request will be the same, we don't
7962 * have enough data to break the tie between them, because the Energy Model
7963 * only includes active power costs. With this model, if we assume that
7964 * frequency requests follow utilization (e.g. using schedutil), the CPU with
7965 * the maximum spare capacity in a performance domain is guaranteed to be among
7966 * the best candidates of the performance domain.
7968 * In practice, it could be preferable from an energy standpoint to pack
7969 * small tasks on a CPU in order to let other CPUs go in deeper idle states,
7970 * but that could also hurt our chances to go cluster idle, and we have no
7971 * ways to tell with the current Energy Model if this is actually a good
7972 * idea or not. So, find_energy_efficient_cpu() basically favors
7973 * cluster-packing, and spreading inside a cluster. That should at least be
7974 * a good thing for latency, and this is consistent with the idea that most
7975 * of the energy savings of EAS come from the asymmetry of the system, and
7976 * not so much from breaking the tie between identical CPUs. That's also the
7977 * reason why EAS is enabled in the topology code only for systems where
7978 * SD_ASYM_CPUCAPACITY is set.
7980 * NOTE: Forkees are not accepted in the energy-aware wake-up path because
7981 * they don't have any useful utilization data yet and it's not possible to
7982 * forecast their impact on energy consumption. Consequently, they will be
7983 * placed by sched_balance_find_dst_cpu() on the least loaded CPU, which might turn out
7984 * to be energy-inefficient in some use-cases. The alternative would be to
7985 * bias new tasks towards specific types of CPUs first, or to try to infer
7986 * their util_avg from the parent task, but those heuristics could hurt
7987 * other use-cases too. So, until someone finds a better way to solve this,
7988 * let's keep things simple by re-using the existing slow path.
7990 static int find_energy_efficient_cpu(struct task_struct *p, int prev_cpu)
7992 struct cpumask *cpus = this_cpu_cpumask_var_ptr(select_rq_mask);
7993 unsigned long prev_delta = ULONG_MAX, best_delta = ULONG_MAX;
7994 unsigned long p_util_min = uclamp_is_used() ? uclamp_eff_value(p, UCLAMP_MIN) : 0;
7995 unsigned long p_util_max = uclamp_is_used() ? uclamp_eff_value(p, UCLAMP_MAX) : 1024;
7996 struct root_domain *rd = this_rq()->rd;
7997 int cpu, best_energy_cpu, target = -1;
7998 int prev_fits = -1, best_fits = -1;
7999 unsigned long best_actual_cap = 0;
8000 unsigned long prev_actual_cap = 0;
8001 struct sched_domain *sd;
8002 struct perf_domain *pd;
8003 struct energy_env eenv;
8006 pd = rcu_dereference(rd->pd);
8011 * Energy-aware wake-up happens on the lowest sched_domain starting
8012 * from sd_asym_cpucapacity spanning over this_cpu and prev_cpu.
8014 sd = rcu_dereference(*this_cpu_ptr(&sd_asym_cpucapacity));
8015 while (sd && !cpumask_test_cpu(prev_cpu, sched_domain_span(sd)))
8022 sync_entity_load_avg(&p->se);
8023 if (!task_util_est(p) && p_util_min == 0)
8026 eenv_task_busy_time(&eenv, p, prev_cpu);
8028 for (; pd; pd = pd->next) {
8029 unsigned long util_min = p_util_min, util_max = p_util_max;
8030 unsigned long cpu_cap, cpu_actual_cap, util;
8031 long prev_spare_cap = -1, max_spare_cap = -1;
8032 unsigned long rq_util_min, rq_util_max;
8033 unsigned long cur_delta, base_energy;
8034 int max_spare_cap_cpu = -1;
8035 int fits, max_fits = -1;
8037 cpumask_and(cpus, perf_domain_span(pd), cpu_online_mask);
8039 if (cpumask_empty(cpus))
8042 /* Account external pressure for the energy estimation */
8043 cpu = cpumask_first(cpus);
8044 cpu_actual_cap = get_actual_cpu_capacity(cpu);
8046 eenv.cpu_cap = cpu_actual_cap;
8049 for_each_cpu(cpu, cpus) {
8050 struct rq *rq = cpu_rq(cpu);
8052 eenv.pd_cap += cpu_actual_cap;
8054 if (!cpumask_test_cpu(cpu, sched_domain_span(sd)))
8057 if (!cpumask_test_cpu(cpu, p->cpus_ptr))
8060 util = cpu_util(cpu, p, cpu, 0);
8061 cpu_cap = capacity_of(cpu);
8064 * Skip CPUs that cannot satisfy the capacity request.
8065 * IOW, placing the task there would make the CPU
8066 * overutilized. Take uclamp into account to see how
8067 * much capacity we can get out of the CPU; this is
8068 * aligned with sched_cpu_util().
8070 if (uclamp_is_used() && !uclamp_rq_is_idle(rq)) {
8072 * Open code uclamp_rq_util_with() except for
8073 * the clamp() part. I.e.: apply max aggregation
8074 * only. util_fits_cpu() logic requires to
8075 * operate on non clamped util but must use the
8076 * max-aggregated uclamp_{min, max}.
8078 rq_util_min = uclamp_rq_get(rq, UCLAMP_MIN);
8079 rq_util_max = uclamp_rq_get(rq, UCLAMP_MAX);
8081 util_min = max(rq_util_min, p_util_min);
8082 util_max = max(rq_util_max, p_util_max);
8085 fits = util_fits_cpu(util, util_min, util_max, cpu);
8089 lsub_positive(&cpu_cap, util);
8091 if (cpu == prev_cpu) {
8092 /* Always use prev_cpu as a candidate. */
8093 prev_spare_cap = cpu_cap;
8095 } else if ((fits > max_fits) ||
8096 ((fits == max_fits) && ((long)cpu_cap > max_spare_cap))) {
8098 * Find the CPU with the maximum spare capacity
8099 * among the remaining CPUs in the performance
8102 max_spare_cap = cpu_cap;
8103 max_spare_cap_cpu = cpu;
8108 if (max_spare_cap_cpu < 0 && prev_spare_cap < 0)
8111 eenv_pd_busy_time(&eenv, cpus, p);
8112 /* Compute the 'base' energy of the pd, without @p */
8113 base_energy = compute_energy(&eenv, pd, cpus, p, -1);
8115 /* Evaluate the energy impact of using prev_cpu. */
8116 if (prev_spare_cap > -1) {
8117 prev_delta = compute_energy(&eenv, pd, cpus, p,
8119 /* CPU utilization has changed */
8120 if (prev_delta < base_energy)
8122 prev_delta -= base_energy;
8123 prev_actual_cap = cpu_actual_cap;
8124 best_delta = min(best_delta, prev_delta);
8127 /* Evaluate the energy impact of using max_spare_cap_cpu. */
8128 if (max_spare_cap_cpu >= 0 && max_spare_cap > prev_spare_cap) {
8129 /* Current best energy cpu fits better */
8130 if (max_fits < best_fits)
8134 * Both don't fit performance hint (i.e. uclamp_min)
8135 * but best energy cpu has better capacity.
8137 if ((max_fits < 0) &&
8138 (cpu_actual_cap <= best_actual_cap))
8141 cur_delta = compute_energy(&eenv, pd, cpus, p,
8143 /* CPU utilization has changed */
8144 if (cur_delta < base_energy)
8146 cur_delta -= base_energy;
8149 * Both fit for the task but best energy cpu has lower
8152 if ((max_fits > 0) && (best_fits > 0) &&
8153 (cur_delta >= best_delta))
8156 best_delta = cur_delta;
8157 best_energy_cpu = max_spare_cap_cpu;
8158 best_fits = max_fits;
8159 best_actual_cap = cpu_actual_cap;
8164 if ((best_fits > prev_fits) ||
8165 ((best_fits > 0) && (best_delta < prev_delta)) ||
8166 ((best_fits < 0) && (best_actual_cap > prev_actual_cap)))
8167 target = best_energy_cpu;
8178 * select_task_rq_fair: Select target runqueue for the waking task in domains
8179 * that have the relevant SD flag set. In practice, this is SD_BALANCE_WAKE,
8180 * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
8182 * Balances load by selecting the idlest CPU in the idlest group, or under
8183 * certain conditions an idle sibling CPU if the domain has SD_WAKE_AFFINE set.
8185 * Returns the target CPU number.
8188 select_task_rq_fair(struct task_struct *p, int prev_cpu, int wake_flags)
8190 int sync = (wake_flags & WF_SYNC) && !(current->flags & PF_EXITING);
8191 struct sched_domain *tmp, *sd = NULL;
8192 int cpu = smp_processor_id();
8193 int new_cpu = prev_cpu;
8194 int want_affine = 0;
8195 /* SD_flags and WF_flags share the first nibble */
8196 int sd_flag = wake_flags & 0xF;
8199 * required for stable ->cpus_allowed
8201 lockdep_assert_held(&p->pi_lock);
8202 if (wake_flags & WF_TTWU) {
8205 if ((wake_flags & WF_CURRENT_CPU) &&
8206 cpumask_test_cpu(cpu, p->cpus_ptr))
8209 if (!is_rd_overutilized(this_rq()->rd)) {
8210 new_cpu = find_energy_efficient_cpu(p, prev_cpu);
8216 want_affine = !wake_wide(p) && cpumask_test_cpu(cpu, p->cpus_ptr);
8220 for_each_domain(cpu, tmp) {
8222 * If both 'cpu' and 'prev_cpu' are part of this domain,
8223 * cpu is a valid SD_WAKE_AFFINE target.
8225 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
8226 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
8227 if (cpu != prev_cpu)
8228 new_cpu = wake_affine(tmp, p, cpu, prev_cpu, sync);
8230 sd = NULL; /* Prefer wake_affine over balance flags */
8235 * Usually only true for WF_EXEC and WF_FORK, as sched_domains
8236 * usually do not have SD_BALANCE_WAKE set. That means wakeup
8237 * will usually go to the fast path.
8239 if (tmp->flags & sd_flag)
8241 else if (!want_affine)
8247 new_cpu = sched_balance_find_dst_cpu(sd, p, cpu, prev_cpu, sd_flag);
8248 } else if (wake_flags & WF_TTWU) { /* XXX always ? */
8250 new_cpu = select_idle_sibling(p, prev_cpu, new_cpu);
8258 * Called immediately before a task is migrated to a new CPU; task_cpu(p) and
8259 * cfs_rq_of(p) references at time of call are still valid and identify the
8260 * previous CPU. The caller guarantees p->pi_lock or task_rq(p)->lock is held.
8262 static void migrate_task_rq_fair(struct task_struct *p, int new_cpu)
8264 struct sched_entity *se = &p->se;
8266 if (!task_on_rq_migrating(p)) {
8267 remove_entity_load_avg(se);
8270 * Here, the task's PELT values have been updated according to
8271 * the current rq's clock. But if that clock hasn't been
8272 * updated in a while, a substantial idle time will be missed,
8273 * leading to an inflation after wake-up on the new rq.
8275 * Estimate the missing time from the cfs_rq last_update_time
8276 * and update sched_avg to improve the PELT continuity after
8279 migrate_se_pelt_lag(se);
8282 /* Tell new CPU we are migrated */
8283 se->avg.last_update_time = 0;
8285 update_scan_period(p, new_cpu);
8288 static void task_dead_fair(struct task_struct *p)
8290 remove_entity_load_avg(&p->se);
8294 * Set the max capacity the task is allowed to run at for misfit detection.
8296 static void set_task_max_allowed_capacity(struct task_struct *p)
8298 struct asym_cap_data *entry;
8300 if (!sched_asym_cpucap_active())
8304 list_for_each_entry_rcu(entry, &asym_cap_list, link) {
8307 cpumask = cpu_capacity_span(entry);
8308 if (!cpumask_intersects(p->cpus_ptr, cpumask))
8311 p->max_allowed_capacity = entry->capacity;
8317 static void set_cpus_allowed_fair(struct task_struct *p, struct affinity_context *ctx)
8319 set_cpus_allowed_common(p, ctx);
8320 set_task_max_allowed_capacity(p);
8324 balance_fair(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
8329 return sched_balance_newidle(rq, rf) != 0;
8332 static inline void set_task_max_allowed_capacity(struct task_struct *p) {}
8333 #endif /* CONFIG_SMP */
8335 static void set_next_buddy(struct sched_entity *se)
8337 for_each_sched_entity(se) {
8338 if (SCHED_WARN_ON(!se->on_rq))
8342 cfs_rq_of(se)->next = se;
8347 * Preempt the current task with a newly woken task if needed:
8349 static void check_preempt_wakeup_fair(struct rq *rq, struct task_struct *p, int wake_flags)
8351 struct task_struct *curr = rq->curr;
8352 struct sched_entity *se = &curr->se, *pse = &p->se;
8353 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
8354 int cse_is_idle, pse_is_idle;
8356 if (unlikely(se == pse))
8360 * This is possible from callers such as attach_tasks(), in which we
8361 * unconditionally wakeup_preempt() after an enqueue (which may have
8362 * lead to a throttle). This both saves work and prevents false
8363 * next-buddy nomination below.
8365 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
8368 if (sched_feat(NEXT_BUDDY) && !(wake_flags & WF_FORK)) {
8369 set_next_buddy(pse);
8373 * We can come here with TIF_NEED_RESCHED already set from new task
8376 * Note: this also catches the edge-case of curr being in a throttled
8377 * group (e.g. via set_curr_task), since update_curr() (in the
8378 * enqueue of curr) will have resulted in resched being set. This
8379 * prevents us from potentially nominating it as a false LAST_BUDDY
8382 if (test_tsk_need_resched(curr))
8385 /* Idle tasks are by definition preempted by non-idle tasks. */
8386 if (unlikely(task_has_idle_policy(curr)) &&
8387 likely(!task_has_idle_policy(p)))
8391 * Batch and idle tasks do not preempt non-idle tasks (their preemption
8392 * is driven by the tick):
8394 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
8397 find_matching_se(&se, &pse);
8400 cse_is_idle = se_is_idle(se);
8401 pse_is_idle = se_is_idle(pse);
8404 * Preempt an idle group in favor of a non-idle group (and don't preempt
8405 * in the inverse case).
8407 if (cse_is_idle && !pse_is_idle)
8409 if (cse_is_idle != pse_is_idle)
8412 cfs_rq = cfs_rq_of(se);
8413 update_curr(cfs_rq);
8416 * XXX pick_eevdf(cfs_rq) != se ?
8418 if (pick_eevdf(cfs_rq) == pse)
8428 static struct task_struct *pick_task_fair(struct rq *rq)
8430 struct sched_entity *se;
8431 struct cfs_rq *cfs_rq;
8435 if (!cfs_rq->nr_running)
8439 struct sched_entity *curr = cfs_rq->curr;
8441 /* When we pick for a remote RQ, we'll not have done put_prev_entity() */
8444 update_curr(cfs_rq);
8448 if (unlikely(check_cfs_rq_runtime(cfs_rq)))
8452 se = pick_next_entity(cfs_rq);
8453 cfs_rq = group_cfs_rq(se);
8460 struct task_struct *
8461 pick_next_task_fair(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
8463 struct cfs_rq *cfs_rq = &rq->cfs;
8464 struct sched_entity *se;
8465 struct task_struct *p;
8469 if (!sched_fair_runnable(rq))
8472 #ifdef CONFIG_FAIR_GROUP_SCHED
8473 if (!prev || prev->sched_class != &fair_sched_class)
8477 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
8478 * likely that a next task is from the same cgroup as the current.
8480 * Therefore attempt to avoid putting and setting the entire cgroup
8481 * hierarchy, only change the part that actually changes.
8485 struct sched_entity *curr = cfs_rq->curr;
8488 * Since we got here without doing put_prev_entity() we also
8489 * have to consider cfs_rq->curr. If it is still a runnable
8490 * entity, update_curr() will update its vruntime, otherwise
8491 * forget we've ever seen it.
8495 update_curr(cfs_rq);
8500 * This call to check_cfs_rq_runtime() will do the
8501 * throttle and dequeue its entity in the parent(s).
8502 * Therefore the nr_running test will indeed
8505 if (unlikely(check_cfs_rq_runtime(cfs_rq))) {
8508 if (!cfs_rq->nr_running)
8515 se = pick_next_entity(cfs_rq);
8516 cfs_rq = group_cfs_rq(se);
8522 * Since we haven't yet done put_prev_entity and if the selected task
8523 * is a different task than we started out with, try and touch the
8524 * least amount of cfs_rqs.
8527 struct sched_entity *pse = &prev->se;
8529 while (!(cfs_rq = is_same_group(se, pse))) {
8530 int se_depth = se->depth;
8531 int pse_depth = pse->depth;
8533 if (se_depth <= pse_depth) {
8534 put_prev_entity(cfs_rq_of(pse), pse);
8535 pse = parent_entity(pse);
8537 if (se_depth >= pse_depth) {
8538 set_next_entity(cfs_rq_of(se), se);
8539 se = parent_entity(se);
8543 put_prev_entity(cfs_rq, pse);
8544 set_next_entity(cfs_rq, se);
8551 put_prev_task(rq, prev);
8554 se = pick_next_entity(cfs_rq);
8555 set_next_entity(cfs_rq, se);
8556 cfs_rq = group_cfs_rq(se);
8561 done: __maybe_unused;
8564 * Move the next running task to the front of
8565 * the list, so our cfs_tasks list becomes MRU
8568 list_move(&p->se.group_node, &rq->cfs_tasks);
8571 if (hrtick_enabled_fair(rq))
8572 hrtick_start_fair(rq, p);
8574 update_misfit_status(p, rq);
8575 sched_fair_update_stop_tick(rq, p);
8583 new_tasks = sched_balance_newidle(rq, rf);
8586 * Because sched_balance_newidle() releases (and re-acquires) rq->lock, it is
8587 * possible for any higher priority task to appear. In that case we
8588 * must re-start the pick_next_entity() loop.
8597 * rq is about to be idle, check if we need to update the
8598 * lost_idle_time of clock_pelt
8600 update_idle_rq_clock_pelt(rq);
8605 static struct task_struct *__pick_next_task_fair(struct rq *rq)
8607 return pick_next_task_fair(rq, NULL, NULL);
8611 * Account for a descheduled task:
8613 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
8615 struct sched_entity *se = &prev->se;
8616 struct cfs_rq *cfs_rq;
8618 for_each_sched_entity(se) {
8619 cfs_rq = cfs_rq_of(se);
8620 put_prev_entity(cfs_rq, se);
8625 * sched_yield() is very simple
8627 static void yield_task_fair(struct rq *rq)
8629 struct task_struct *curr = rq->curr;
8630 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
8631 struct sched_entity *se = &curr->se;
8634 * Are we the only task in the tree?
8636 if (unlikely(rq->nr_running == 1))
8639 clear_buddies(cfs_rq, se);
8641 update_rq_clock(rq);
8643 * Update run-time statistics of the 'current'.
8645 update_curr(cfs_rq);
8647 * Tell update_rq_clock() that we've just updated,
8648 * so we don't do microscopic update in schedule()
8649 * and double the fastpath cost.
8651 rq_clock_skip_update(rq);
8653 se->deadline += calc_delta_fair(se->slice, se);
8656 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p)
8658 struct sched_entity *se = &p->se;
8660 /* throttled hierarchies are not runnable */
8661 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
8664 /* Tell the scheduler that we'd really like se to run next. */
8667 yield_task_fair(rq);
8673 /**************************************************
8674 * Fair scheduling class load-balancing methods.
8678 * The purpose of load-balancing is to achieve the same basic fairness the
8679 * per-CPU scheduler provides, namely provide a proportional amount of compute
8680 * time to each task. This is expressed in the following equation:
8682 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
8684 * Where W_i,n is the n-th weight average for CPU i. The instantaneous weight
8685 * W_i,0 is defined as:
8687 * W_i,0 = \Sum_j w_i,j (2)
8689 * Where w_i,j is the weight of the j-th runnable task on CPU i. This weight
8690 * is derived from the nice value as per sched_prio_to_weight[].
8692 * The weight average is an exponential decay average of the instantaneous
8695 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
8697 * C_i is the compute capacity of CPU i, typically it is the
8698 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
8699 * can also include other factors [XXX].
8701 * To achieve this balance we define a measure of imbalance which follows
8702 * directly from (1):
8704 * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4)
8706 * We them move tasks around to minimize the imbalance. In the continuous
8707 * function space it is obvious this converges, in the discrete case we get
8708 * a few fun cases generally called infeasible weight scenarios.
8711 * - infeasible weights;
8712 * - local vs global optima in the discrete case. ]
8717 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
8718 * for all i,j solution, we create a tree of CPUs that follows the hardware
8719 * topology where each level pairs two lower groups (or better). This results
8720 * in O(log n) layers. Furthermore we reduce the number of CPUs going up the
8721 * tree to only the first of the previous level and we decrease the frequency
8722 * of load-balance at each level inv. proportional to the number of CPUs in
8728 * \Sum { --- * --- * 2^i } = O(n) (5)
8730 * `- size of each group
8731 * | | `- number of CPUs doing load-balance
8733 * `- sum over all levels
8735 * Coupled with a limit on how many tasks we can migrate every balance pass,
8736 * this makes (5) the runtime complexity of the balancer.
8738 * An important property here is that each CPU is still (indirectly) connected
8739 * to every other CPU in at most O(log n) steps:
8741 * The adjacency matrix of the resulting graph is given by:
8744 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
8747 * And you'll find that:
8749 * A^(log_2 n)_i,j != 0 for all i,j (7)
8751 * Showing there's indeed a path between every CPU in at most O(log n) steps.
8752 * The task movement gives a factor of O(m), giving a convergence complexity
8755 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
8760 * In order to avoid CPUs going idle while there's still work to do, new idle
8761 * balancing is more aggressive and has the newly idle CPU iterate up the domain
8762 * tree itself instead of relying on other CPUs to bring it work.
8764 * This adds some complexity to both (5) and (8) but it reduces the total idle
8772 * Cgroups make a horror show out of (2), instead of a simple sum we get:
8775 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
8780 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
8782 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on CPU i.
8784 * The big problem is S_k, its a global sum needed to compute a local (W_i)
8787 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
8788 * rewrite all of this once again.]
8791 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
8793 enum fbq_type { regular, remote, all };
8796 * 'group_type' describes the group of CPUs at the moment of load balancing.
8798 * The enum is ordered by pulling priority, with the group with lowest priority
8799 * first so the group_type can simply be compared when selecting the busiest
8800 * group. See update_sd_pick_busiest().
8803 /* The group has spare capacity that can be used to run more tasks. */
8804 group_has_spare = 0,
8806 * The group is fully used and the tasks don't compete for more CPU
8807 * cycles. Nevertheless, some tasks might wait before running.
8811 * One task doesn't fit with CPU's capacity and must be migrated to a
8812 * more powerful CPU.
8816 * Balance SMT group that's fully busy. Can benefit from migration
8817 * a task on SMT with busy sibling to another CPU on idle core.
8821 * SD_ASYM_PACKING only: One local CPU with higher capacity is available,
8822 * and the task should be migrated to it instead of running on the
8827 * The tasks' affinity constraints previously prevented the scheduler
8828 * from balancing the load across the system.
8832 * The CPU is overloaded and can't provide expected CPU cycles to all
8838 enum migration_type {
8845 #define LBF_ALL_PINNED 0x01
8846 #define LBF_NEED_BREAK 0x02
8847 #define LBF_DST_PINNED 0x04
8848 #define LBF_SOME_PINNED 0x08
8849 #define LBF_ACTIVE_LB 0x10
8852 struct sched_domain *sd;
8860 struct cpumask *dst_grpmask;
8862 enum cpu_idle_type idle;
8864 /* The set of CPUs under consideration for load-balancing */
8865 struct cpumask *cpus;
8870 unsigned int loop_break;
8871 unsigned int loop_max;
8873 enum fbq_type fbq_type;
8874 enum migration_type migration_type;
8875 struct list_head tasks;
8879 * Is this task likely cache-hot:
8881 static int task_hot(struct task_struct *p, struct lb_env *env)
8885 lockdep_assert_rq_held(env->src_rq);
8887 if (p->sched_class != &fair_sched_class)
8890 if (unlikely(task_has_idle_policy(p)))
8893 /* SMT siblings share cache */
8894 if (env->sd->flags & SD_SHARE_CPUCAPACITY)
8898 * Buddy candidates are cache hot:
8900 if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running &&
8901 (&p->se == cfs_rq_of(&p->se)->next))
8904 if (sysctl_sched_migration_cost == -1)
8908 * Don't migrate task if the task's cookie does not match
8909 * with the destination CPU's core cookie.
8911 if (!sched_core_cookie_match(cpu_rq(env->dst_cpu), p))
8914 if (sysctl_sched_migration_cost == 0)
8917 delta = rq_clock_task(env->src_rq) - p->se.exec_start;
8919 return delta < (s64)sysctl_sched_migration_cost;
8922 #ifdef CONFIG_NUMA_BALANCING
8924 * Returns 1, if task migration degrades locality
8925 * Returns 0, if task migration improves locality i.e migration preferred.
8926 * Returns -1, if task migration is not affected by locality.
8928 static int migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
8930 struct numa_group *numa_group = rcu_dereference(p->numa_group);
8931 unsigned long src_weight, dst_weight;
8932 int src_nid, dst_nid, dist;
8934 if (!static_branch_likely(&sched_numa_balancing))
8937 if (!p->numa_faults || !(env->sd->flags & SD_NUMA))
8940 src_nid = cpu_to_node(env->src_cpu);
8941 dst_nid = cpu_to_node(env->dst_cpu);
8943 if (src_nid == dst_nid)
8946 /* Migrating away from the preferred node is always bad. */
8947 if (src_nid == p->numa_preferred_nid) {
8948 if (env->src_rq->nr_running > env->src_rq->nr_preferred_running)
8954 /* Encourage migration to the preferred node. */
8955 if (dst_nid == p->numa_preferred_nid)
8958 /* Leaving a core idle is often worse than degrading locality. */
8959 if (env->idle == CPU_IDLE)
8962 dist = node_distance(src_nid, dst_nid);
8964 src_weight = group_weight(p, src_nid, dist);
8965 dst_weight = group_weight(p, dst_nid, dist);
8967 src_weight = task_weight(p, src_nid, dist);
8968 dst_weight = task_weight(p, dst_nid, dist);
8971 return dst_weight < src_weight;
8975 static inline int migrate_degrades_locality(struct task_struct *p,
8983 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
8986 int can_migrate_task(struct task_struct *p, struct lb_env *env)
8990 lockdep_assert_rq_held(env->src_rq);
8993 * We do not migrate tasks that are:
8994 * 1) throttled_lb_pair, or
8995 * 2) cannot be migrated to this CPU due to cpus_ptr, or
8996 * 3) running (obviously), or
8997 * 4) are cache-hot on their current CPU.
8999 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
9002 /* Disregard percpu kthreads; they are where they need to be. */
9003 if (kthread_is_per_cpu(p))
9006 if (!cpumask_test_cpu(env->dst_cpu, p->cpus_ptr)) {
9009 schedstat_inc(p->stats.nr_failed_migrations_affine);
9011 env->flags |= LBF_SOME_PINNED;
9014 * Remember if this task can be migrated to any other CPU in
9015 * our sched_group. We may want to revisit it if we couldn't
9016 * meet load balance goals by pulling other tasks on src_cpu.
9018 * Avoid computing new_dst_cpu
9020 * - if we have already computed one in current iteration
9021 * - if it's an active balance
9023 if (env->idle == CPU_NEWLY_IDLE ||
9024 env->flags & (LBF_DST_PINNED | LBF_ACTIVE_LB))
9027 /* Prevent to re-select dst_cpu via env's CPUs: */
9028 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
9029 if (cpumask_test_cpu(cpu, p->cpus_ptr)) {
9030 env->flags |= LBF_DST_PINNED;
9031 env->new_dst_cpu = cpu;
9039 /* Record that we found at least one task that could run on dst_cpu */
9040 env->flags &= ~LBF_ALL_PINNED;
9042 if (task_on_cpu(env->src_rq, p)) {
9043 schedstat_inc(p->stats.nr_failed_migrations_running);
9048 * Aggressive migration if:
9050 * 2) destination numa is preferred
9051 * 3) task is cache cold, or
9052 * 4) too many balance attempts have failed.
9054 if (env->flags & LBF_ACTIVE_LB)
9057 tsk_cache_hot = migrate_degrades_locality(p, env);
9058 if (tsk_cache_hot == -1)
9059 tsk_cache_hot = task_hot(p, env);
9061 if (tsk_cache_hot <= 0 ||
9062 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
9063 if (tsk_cache_hot == 1) {
9064 schedstat_inc(env->sd->lb_hot_gained[env->idle]);
9065 schedstat_inc(p->stats.nr_forced_migrations);
9070 schedstat_inc(p->stats.nr_failed_migrations_hot);
9075 * detach_task() -- detach the task for the migration specified in env
9077 static void detach_task(struct task_struct *p, struct lb_env *env)
9079 lockdep_assert_rq_held(env->src_rq);
9081 deactivate_task(env->src_rq, p, DEQUEUE_NOCLOCK);
9082 set_task_cpu(p, env->dst_cpu);
9086 * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
9087 * part of active balancing operations within "domain".
9089 * Returns a task if successful and NULL otherwise.
9091 static struct task_struct *detach_one_task(struct lb_env *env)
9093 struct task_struct *p;
9095 lockdep_assert_rq_held(env->src_rq);
9097 list_for_each_entry_reverse(p,
9098 &env->src_rq->cfs_tasks, se.group_node) {
9099 if (!can_migrate_task(p, env))
9102 detach_task(p, env);
9105 * Right now, this is only the second place where
9106 * lb_gained[env->idle] is updated (other is detach_tasks)
9107 * so we can safely collect stats here rather than
9108 * inside detach_tasks().
9110 schedstat_inc(env->sd->lb_gained[env->idle]);
9117 * detach_tasks() -- tries to detach up to imbalance load/util/tasks from
9118 * busiest_rq, as part of a balancing operation within domain "sd".
9120 * Returns number of detached tasks if successful and 0 otherwise.
9122 static int detach_tasks(struct lb_env *env)
9124 struct list_head *tasks = &env->src_rq->cfs_tasks;
9125 unsigned long util, load;
9126 struct task_struct *p;
9129 lockdep_assert_rq_held(env->src_rq);
9132 * Source run queue has been emptied by another CPU, clear
9133 * LBF_ALL_PINNED flag as we will not test any task.
9135 if (env->src_rq->nr_running <= 1) {
9136 env->flags &= ~LBF_ALL_PINNED;
9140 if (env->imbalance <= 0)
9143 while (!list_empty(tasks)) {
9145 * We don't want to steal all, otherwise we may be treated likewise,
9146 * which could at worst lead to a livelock crash.
9148 if (env->idle && env->src_rq->nr_running <= 1)
9153 * We've more or less seen every task there is, call it quits
9154 * unless we haven't found any movable task yet.
9156 if (env->loop > env->loop_max &&
9157 !(env->flags & LBF_ALL_PINNED))
9160 /* take a breather every nr_migrate tasks */
9161 if (env->loop > env->loop_break) {
9162 env->loop_break += SCHED_NR_MIGRATE_BREAK;
9163 env->flags |= LBF_NEED_BREAK;
9167 p = list_last_entry(tasks, struct task_struct, se.group_node);
9169 if (!can_migrate_task(p, env))
9172 switch (env->migration_type) {
9175 * Depending of the number of CPUs and tasks and the
9176 * cgroup hierarchy, task_h_load() can return a null
9177 * value. Make sure that env->imbalance decreases
9178 * otherwise detach_tasks() will stop only after
9179 * detaching up to loop_max tasks.
9181 load = max_t(unsigned long, task_h_load(p), 1);
9183 if (sched_feat(LB_MIN) &&
9184 load < 16 && !env->sd->nr_balance_failed)
9188 * Make sure that we don't migrate too much load.
9189 * Nevertheless, let relax the constraint if
9190 * scheduler fails to find a good waiting task to
9193 if (shr_bound(load, env->sd->nr_balance_failed) > env->imbalance)
9196 env->imbalance -= load;
9200 util = task_util_est(p);
9202 if (shr_bound(util, env->sd->nr_balance_failed) > env->imbalance)
9205 env->imbalance -= util;
9212 case migrate_misfit:
9213 /* This is not a misfit task */
9214 if (task_fits_cpu(p, env->src_cpu))
9221 detach_task(p, env);
9222 list_add(&p->se.group_node, &env->tasks);
9226 #ifdef CONFIG_PREEMPTION
9228 * NEWIDLE balancing is a source of latency, so preemptible
9229 * kernels will stop after the first task is detached to minimize
9230 * the critical section.
9232 if (env->idle == CPU_NEWLY_IDLE)
9237 * We only want to steal up to the prescribed amount of
9240 if (env->imbalance <= 0)
9245 list_move(&p->se.group_node, tasks);
9249 * Right now, this is one of only two places we collect this stat
9250 * so we can safely collect detach_one_task() stats here rather
9251 * than inside detach_one_task().
9253 schedstat_add(env->sd->lb_gained[env->idle], detached);
9259 * attach_task() -- attach the task detached by detach_task() to its new rq.
9261 static void attach_task(struct rq *rq, struct task_struct *p)
9263 lockdep_assert_rq_held(rq);
9265 WARN_ON_ONCE(task_rq(p) != rq);
9266 activate_task(rq, p, ENQUEUE_NOCLOCK);
9267 wakeup_preempt(rq, p, 0);
9271 * attach_one_task() -- attaches the task returned from detach_one_task() to
9274 static void attach_one_task(struct rq *rq, struct task_struct *p)
9279 update_rq_clock(rq);
9285 * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
9288 static void attach_tasks(struct lb_env *env)
9290 struct list_head *tasks = &env->tasks;
9291 struct task_struct *p;
9294 rq_lock(env->dst_rq, &rf);
9295 update_rq_clock(env->dst_rq);
9297 while (!list_empty(tasks)) {
9298 p = list_first_entry(tasks, struct task_struct, se.group_node);
9299 list_del_init(&p->se.group_node);
9301 attach_task(env->dst_rq, p);
9304 rq_unlock(env->dst_rq, &rf);
9307 #ifdef CONFIG_NO_HZ_COMMON
9308 static inline bool cfs_rq_has_blocked(struct cfs_rq *cfs_rq)
9310 if (cfs_rq->avg.load_avg)
9313 if (cfs_rq->avg.util_avg)
9319 static inline bool others_have_blocked(struct rq *rq)
9321 if (cpu_util_rt(rq))
9324 if (cpu_util_dl(rq))
9327 if (hw_load_avg(rq))
9330 if (cpu_util_irq(rq))
9336 static inline void update_blocked_load_tick(struct rq *rq)
9338 WRITE_ONCE(rq->last_blocked_load_update_tick, jiffies);
9341 static inline void update_blocked_load_status(struct rq *rq, bool has_blocked)
9344 rq->has_blocked_load = 0;
9347 static inline bool cfs_rq_has_blocked(struct cfs_rq *cfs_rq) { return false; }
9348 static inline bool others_have_blocked(struct rq *rq) { return false; }
9349 static inline void update_blocked_load_tick(struct rq *rq) {}
9350 static inline void update_blocked_load_status(struct rq *rq, bool has_blocked) {}
9353 static bool __update_blocked_others(struct rq *rq, bool *done)
9355 const struct sched_class *curr_class;
9356 u64 now = rq_clock_pelt(rq);
9357 unsigned long hw_pressure;
9361 * update_load_avg() can call cpufreq_update_util(). Make sure that RT,
9362 * DL and IRQ signals have been updated before updating CFS.
9364 curr_class = rq->curr->sched_class;
9366 hw_pressure = arch_scale_hw_pressure(cpu_of(rq));
9368 decayed = update_rt_rq_load_avg(now, rq, curr_class == &rt_sched_class) |
9369 update_dl_rq_load_avg(now, rq, curr_class == &dl_sched_class) |
9370 update_hw_load_avg(now, rq, hw_pressure) |
9371 update_irq_load_avg(rq, 0);
9373 if (others_have_blocked(rq))
9379 #ifdef CONFIG_FAIR_GROUP_SCHED
9381 static bool __update_blocked_fair(struct rq *rq, bool *done)
9383 struct cfs_rq *cfs_rq, *pos;
9384 bool decayed = false;
9385 int cpu = cpu_of(rq);
9388 * Iterates the task_group tree in a bottom up fashion, see
9389 * list_add_leaf_cfs_rq() for details.
9391 for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) {
9392 struct sched_entity *se;
9394 if (update_cfs_rq_load_avg(cfs_rq_clock_pelt(cfs_rq), cfs_rq)) {
9395 update_tg_load_avg(cfs_rq);
9397 if (cfs_rq->nr_running == 0)
9398 update_idle_cfs_rq_clock_pelt(cfs_rq);
9400 if (cfs_rq == &rq->cfs)
9404 /* Propagate pending load changes to the parent, if any: */
9405 se = cfs_rq->tg->se[cpu];
9406 if (se && !skip_blocked_update(se))
9407 update_load_avg(cfs_rq_of(se), se, UPDATE_TG);
9410 * There can be a lot of idle CPU cgroups. Don't let fully
9411 * decayed cfs_rqs linger on the list.
9413 if (cfs_rq_is_decayed(cfs_rq))
9414 list_del_leaf_cfs_rq(cfs_rq);
9416 /* Don't need periodic decay once load/util_avg are null */
9417 if (cfs_rq_has_blocked(cfs_rq))
9425 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
9426 * This needs to be done in a top-down fashion because the load of a child
9427 * group is a fraction of its parents load.
9429 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
9431 struct rq *rq = rq_of(cfs_rq);
9432 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
9433 unsigned long now = jiffies;
9436 if (cfs_rq->last_h_load_update == now)
9439 WRITE_ONCE(cfs_rq->h_load_next, NULL);
9440 for_each_sched_entity(se) {
9441 cfs_rq = cfs_rq_of(se);
9442 WRITE_ONCE(cfs_rq->h_load_next, se);
9443 if (cfs_rq->last_h_load_update == now)
9448 cfs_rq->h_load = cfs_rq_load_avg(cfs_rq);
9449 cfs_rq->last_h_load_update = now;
9452 while ((se = READ_ONCE(cfs_rq->h_load_next)) != NULL) {
9453 load = cfs_rq->h_load;
9454 load = div64_ul(load * se->avg.load_avg,
9455 cfs_rq_load_avg(cfs_rq) + 1);
9456 cfs_rq = group_cfs_rq(se);
9457 cfs_rq->h_load = load;
9458 cfs_rq->last_h_load_update = now;
9462 static unsigned long task_h_load(struct task_struct *p)
9464 struct cfs_rq *cfs_rq = task_cfs_rq(p);
9466 update_cfs_rq_h_load(cfs_rq);
9467 return div64_ul(p->se.avg.load_avg * cfs_rq->h_load,
9468 cfs_rq_load_avg(cfs_rq) + 1);
9471 static bool __update_blocked_fair(struct rq *rq, bool *done)
9473 struct cfs_rq *cfs_rq = &rq->cfs;
9476 decayed = update_cfs_rq_load_avg(cfs_rq_clock_pelt(cfs_rq), cfs_rq);
9477 if (cfs_rq_has_blocked(cfs_rq))
9483 static unsigned long task_h_load(struct task_struct *p)
9485 return p->se.avg.load_avg;
9489 static void sched_balance_update_blocked_averages(int cpu)
9491 bool decayed = false, done = true;
9492 struct rq *rq = cpu_rq(cpu);
9495 rq_lock_irqsave(rq, &rf);
9496 update_blocked_load_tick(rq);
9497 update_rq_clock(rq);
9499 decayed |= __update_blocked_others(rq, &done);
9500 decayed |= __update_blocked_fair(rq, &done);
9502 update_blocked_load_status(rq, !done);
9504 cpufreq_update_util(rq, 0);
9505 rq_unlock_irqrestore(rq, &rf);
9508 /********** Helpers for sched_balance_find_src_group ************************/
9511 * sg_lb_stats - stats of a sched_group required for load-balancing:
9513 struct sg_lb_stats {
9514 unsigned long avg_load; /* Avg load over the CPUs of the group */
9515 unsigned long group_load; /* Total load over the CPUs of the group */
9516 unsigned long group_capacity; /* Capacity over the CPUs of the group */
9517 unsigned long group_util; /* Total utilization over the CPUs of the group */
9518 unsigned long group_runnable; /* Total runnable time over the CPUs of the group */
9519 unsigned int sum_nr_running; /* Nr of all tasks running in the group */
9520 unsigned int sum_h_nr_running; /* Nr of CFS tasks running in the group */
9521 unsigned int idle_cpus; /* Nr of idle CPUs in the group */
9522 unsigned int group_weight;
9523 enum group_type group_type;
9524 unsigned int group_asym_packing; /* Tasks should be moved to preferred CPU */
9525 unsigned int group_smt_balance; /* Task on busy SMT be moved */
9526 unsigned long group_misfit_task_load; /* A CPU has a task too big for its capacity */
9527 #ifdef CONFIG_NUMA_BALANCING
9528 unsigned int nr_numa_running;
9529 unsigned int nr_preferred_running;
9534 * sd_lb_stats - stats of a sched_domain required for load-balancing:
9536 struct sd_lb_stats {
9537 struct sched_group *busiest; /* Busiest group in this sd */
9538 struct sched_group *local; /* Local group in this sd */
9539 unsigned long total_load; /* Total load of all groups in sd */
9540 unsigned long total_capacity; /* Total capacity of all groups in sd */
9541 unsigned long avg_load; /* Average load across all groups in sd */
9542 unsigned int prefer_sibling; /* Tasks should go to sibling first */
9544 struct sg_lb_stats busiest_stat; /* Statistics of the busiest group */
9545 struct sg_lb_stats local_stat; /* Statistics of the local group */
9548 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
9551 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
9552 * local_stat because update_sg_lb_stats() does a full clear/assignment.
9553 * We must however set busiest_stat::group_type and
9554 * busiest_stat::idle_cpus to the worst busiest group because
9555 * update_sd_pick_busiest() reads these before assignment.
9557 *sds = (struct sd_lb_stats){
9561 .total_capacity = 0UL,
9563 .idle_cpus = UINT_MAX,
9564 .group_type = group_has_spare,
9569 static unsigned long scale_rt_capacity(int cpu)
9571 unsigned long max = get_actual_cpu_capacity(cpu);
9572 struct rq *rq = cpu_rq(cpu);
9573 unsigned long used, free;
9576 irq = cpu_util_irq(rq);
9578 if (unlikely(irq >= max))
9582 * avg_rt.util_avg and avg_dl.util_avg track binary signals
9583 * (running and not running) with weights 0 and 1024 respectively.
9585 used = cpu_util_rt(rq);
9586 used += cpu_util_dl(rq);
9588 if (unlikely(used >= max))
9593 return scale_irq_capacity(free, irq, max);
9596 static void update_cpu_capacity(struct sched_domain *sd, int cpu)
9598 unsigned long capacity = scale_rt_capacity(cpu);
9599 struct sched_group *sdg = sd->groups;
9604 cpu_rq(cpu)->cpu_capacity = capacity;
9605 trace_sched_cpu_capacity_tp(cpu_rq(cpu));
9607 sdg->sgc->capacity = capacity;
9608 sdg->sgc->min_capacity = capacity;
9609 sdg->sgc->max_capacity = capacity;
9612 void update_group_capacity(struct sched_domain *sd, int cpu)
9614 struct sched_domain *child = sd->child;
9615 struct sched_group *group, *sdg = sd->groups;
9616 unsigned long capacity, min_capacity, max_capacity;
9617 unsigned long interval;
9619 interval = msecs_to_jiffies(sd->balance_interval);
9620 interval = clamp(interval, 1UL, max_load_balance_interval);
9621 sdg->sgc->next_update = jiffies + interval;
9624 update_cpu_capacity(sd, cpu);
9629 min_capacity = ULONG_MAX;
9632 if (child->flags & SD_OVERLAP) {
9634 * SD_OVERLAP domains cannot assume that child groups
9635 * span the current group.
9638 for_each_cpu(cpu, sched_group_span(sdg)) {
9639 unsigned long cpu_cap = capacity_of(cpu);
9641 capacity += cpu_cap;
9642 min_capacity = min(cpu_cap, min_capacity);
9643 max_capacity = max(cpu_cap, max_capacity);
9647 * !SD_OVERLAP domains can assume that child groups
9648 * span the current group.
9651 group = child->groups;
9653 struct sched_group_capacity *sgc = group->sgc;
9655 capacity += sgc->capacity;
9656 min_capacity = min(sgc->min_capacity, min_capacity);
9657 max_capacity = max(sgc->max_capacity, max_capacity);
9658 group = group->next;
9659 } while (group != child->groups);
9662 sdg->sgc->capacity = capacity;
9663 sdg->sgc->min_capacity = min_capacity;
9664 sdg->sgc->max_capacity = max_capacity;
9668 * Check whether the capacity of the rq has been noticeably reduced by side
9669 * activity. The imbalance_pct is used for the threshold.
9670 * Return true is the capacity is reduced
9673 check_cpu_capacity(struct rq *rq, struct sched_domain *sd)
9675 return ((rq->cpu_capacity * sd->imbalance_pct) <
9676 (arch_scale_cpu_capacity(cpu_of(rq)) * 100));
9679 /* Check if the rq has a misfit task */
9680 static inline bool check_misfit_status(struct rq *rq)
9682 return rq->misfit_task_load;
9686 * Group imbalance indicates (and tries to solve) the problem where balancing
9687 * groups is inadequate due to ->cpus_ptr constraints.
9689 * Imagine a situation of two groups of 4 CPUs each and 4 tasks each with a
9690 * cpumask covering 1 CPU of the first group and 3 CPUs of the second group.
9693 * { 0 1 2 3 } { 4 5 6 7 }
9696 * If we were to balance group-wise we'd place two tasks in the first group and
9697 * two tasks in the second group. Clearly this is undesired as it will overload
9698 * cpu 3 and leave one of the CPUs in the second group unused.
9700 * The current solution to this issue is detecting the skew in the first group
9701 * by noticing the lower domain failed to reach balance and had difficulty
9702 * moving tasks due to affinity constraints.
9704 * When this is so detected; this group becomes a candidate for busiest; see
9705 * update_sd_pick_busiest(). And calculate_imbalance() and
9706 * sched_balance_find_src_group() avoid some of the usual balance conditions to allow it
9707 * to create an effective group imbalance.
9709 * This is a somewhat tricky proposition since the next run might not find the
9710 * group imbalance and decide the groups need to be balanced again. A most
9711 * subtle and fragile situation.
9714 static inline int sg_imbalanced(struct sched_group *group)
9716 return group->sgc->imbalance;
9720 * group_has_capacity returns true if the group has spare capacity that could
9721 * be used by some tasks.
9722 * We consider that a group has spare capacity if the number of task is
9723 * smaller than the number of CPUs or if the utilization is lower than the
9724 * available capacity for CFS tasks.
9725 * For the latter, we use a threshold to stabilize the state, to take into
9726 * account the variance of the tasks' load and to return true if the available
9727 * capacity in meaningful for the load balancer.
9728 * As an example, an available capacity of 1% can appear but it doesn't make
9729 * any benefit for the load balance.
9732 group_has_capacity(unsigned int imbalance_pct, struct sg_lb_stats *sgs)
9734 if (sgs->sum_nr_running < sgs->group_weight)
9737 if ((sgs->group_capacity * imbalance_pct) <
9738 (sgs->group_runnable * 100))
9741 if ((sgs->group_capacity * 100) >
9742 (sgs->group_util * imbalance_pct))
9749 * group_is_overloaded returns true if the group has more tasks than it can
9751 * group_is_overloaded is not equals to !group_has_capacity because a group
9752 * with the exact right number of tasks, has no more spare capacity but is not
9753 * overloaded so both group_has_capacity and group_is_overloaded return
9757 group_is_overloaded(unsigned int imbalance_pct, struct sg_lb_stats *sgs)
9759 if (sgs->sum_nr_running <= sgs->group_weight)
9762 if ((sgs->group_capacity * 100) <
9763 (sgs->group_util * imbalance_pct))
9766 if ((sgs->group_capacity * imbalance_pct) <
9767 (sgs->group_runnable * 100))
9774 group_type group_classify(unsigned int imbalance_pct,
9775 struct sched_group *group,
9776 struct sg_lb_stats *sgs)
9778 if (group_is_overloaded(imbalance_pct, sgs))
9779 return group_overloaded;
9781 if (sg_imbalanced(group))
9782 return group_imbalanced;
9784 if (sgs->group_asym_packing)
9785 return group_asym_packing;
9787 if (sgs->group_smt_balance)
9788 return group_smt_balance;
9790 if (sgs->group_misfit_task_load)
9791 return group_misfit_task;
9793 if (!group_has_capacity(imbalance_pct, sgs))
9794 return group_fully_busy;
9796 return group_has_spare;
9800 * sched_use_asym_prio - Check whether asym_packing priority must be used
9801 * @sd: The scheduling domain of the load balancing
9804 * Always use CPU priority when balancing load between SMT siblings. When
9805 * balancing load between cores, it is not sufficient that @cpu is idle. Only
9806 * use CPU priority if the whole core is idle.
9808 * Returns: True if the priority of @cpu must be followed. False otherwise.
9810 static bool sched_use_asym_prio(struct sched_domain *sd, int cpu)
9812 if (!(sd->flags & SD_ASYM_PACKING))
9815 if (!sched_smt_active())
9818 return sd->flags & SD_SHARE_CPUCAPACITY || is_core_idle(cpu);
9821 static inline bool sched_asym(struct sched_domain *sd, int dst_cpu, int src_cpu)
9824 * First check if @dst_cpu can do asym_packing load balance. Only do it
9825 * if it has higher priority than @src_cpu.
9827 return sched_use_asym_prio(sd, dst_cpu) &&
9828 sched_asym_prefer(dst_cpu, src_cpu);
9832 * sched_group_asym - Check if the destination CPU can do asym_packing balance
9833 * @env: The load balancing environment
9834 * @sgs: Load-balancing statistics of the candidate busiest group
9835 * @group: The candidate busiest group
9837 * @env::dst_cpu can do asym_packing if it has higher priority than the
9838 * preferred CPU of @group.
9840 * Return: true if @env::dst_cpu can do with asym_packing load balance. False
9844 sched_group_asym(struct lb_env *env, struct sg_lb_stats *sgs, struct sched_group *group)
9847 * CPU priorities do not make sense for SMT cores with more than one
9850 if ((group->flags & SD_SHARE_CPUCAPACITY) &&
9851 (sgs->group_weight - sgs->idle_cpus != 1))
9854 return sched_asym(env->sd, env->dst_cpu, group->asym_prefer_cpu);
9857 /* One group has more than one SMT CPU while the other group does not */
9858 static inline bool smt_vs_nonsmt_groups(struct sched_group *sg1,
9859 struct sched_group *sg2)
9864 return (sg1->flags & SD_SHARE_CPUCAPACITY) !=
9865 (sg2->flags & SD_SHARE_CPUCAPACITY);
9868 static inline bool smt_balance(struct lb_env *env, struct sg_lb_stats *sgs,
9869 struct sched_group *group)
9875 * For SMT source group, it is better to move a task
9876 * to a CPU that doesn't have multiple tasks sharing its CPU capacity.
9877 * Note that if a group has a single SMT, SD_SHARE_CPUCAPACITY
9880 if (group->flags & SD_SHARE_CPUCAPACITY &&
9881 sgs->sum_h_nr_running > 1)
9887 static inline long sibling_imbalance(struct lb_env *env,
9888 struct sd_lb_stats *sds,
9889 struct sg_lb_stats *busiest,
9890 struct sg_lb_stats *local)
9892 int ncores_busiest, ncores_local;
9895 if (!env->idle || !busiest->sum_nr_running)
9898 ncores_busiest = sds->busiest->cores;
9899 ncores_local = sds->local->cores;
9901 if (ncores_busiest == ncores_local) {
9902 imbalance = busiest->sum_nr_running;
9903 lsub_positive(&imbalance, local->sum_nr_running);
9907 /* Balance such that nr_running/ncores ratio are same on both groups */
9908 imbalance = ncores_local * busiest->sum_nr_running;
9909 lsub_positive(&imbalance, ncores_busiest * local->sum_nr_running);
9910 /* Normalize imbalance and do rounding on normalization */
9911 imbalance = 2 * imbalance + ncores_local + ncores_busiest;
9912 imbalance /= ncores_local + ncores_busiest;
9914 /* Take advantage of resource in an empty sched group */
9915 if (imbalance <= 1 && local->sum_nr_running == 0 &&
9916 busiest->sum_nr_running > 1)
9923 sched_reduced_capacity(struct rq *rq, struct sched_domain *sd)
9926 * When there is more than 1 task, the group_overloaded case already
9927 * takes care of cpu with reduced capacity
9929 if (rq->cfs.h_nr_running != 1)
9932 return check_cpu_capacity(rq, sd);
9936 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
9937 * @env: The load balancing environment.
9938 * @sds: Load-balancing data with statistics of the local group.
9939 * @group: sched_group whose statistics are to be updated.
9940 * @sgs: variable to hold the statistics for this group.
9941 * @sg_overloaded: sched_group is overloaded
9942 * @sg_overutilized: sched_group is overutilized
9944 static inline void update_sg_lb_stats(struct lb_env *env,
9945 struct sd_lb_stats *sds,
9946 struct sched_group *group,
9947 struct sg_lb_stats *sgs,
9948 bool *sg_overloaded,
9949 bool *sg_overutilized)
9951 int i, nr_running, local_group;
9953 memset(sgs, 0, sizeof(*sgs));
9955 local_group = group == sds->local;
9957 for_each_cpu_and(i, sched_group_span(group), env->cpus) {
9958 struct rq *rq = cpu_rq(i);
9959 unsigned long load = cpu_load(rq);
9961 sgs->group_load += load;
9962 sgs->group_util += cpu_util_cfs(i);
9963 sgs->group_runnable += cpu_runnable(rq);
9964 sgs->sum_h_nr_running += rq->cfs.h_nr_running;
9966 nr_running = rq->nr_running;
9967 sgs->sum_nr_running += nr_running;
9972 if (cpu_overutilized(i))
9973 *sg_overutilized = 1;
9975 #ifdef CONFIG_NUMA_BALANCING
9976 sgs->nr_numa_running += rq->nr_numa_running;
9977 sgs->nr_preferred_running += rq->nr_preferred_running;
9980 * No need to call idle_cpu() if nr_running is not 0
9982 if (!nr_running && idle_cpu(i)) {
9984 /* Idle cpu can't have misfit task */
9991 if (env->sd->flags & SD_ASYM_CPUCAPACITY) {
9992 /* Check for a misfit task on the cpu */
9993 if (sgs->group_misfit_task_load < rq->misfit_task_load) {
9994 sgs->group_misfit_task_load = rq->misfit_task_load;
9997 } else if (env->idle && sched_reduced_capacity(rq, env->sd)) {
9998 /* Check for a task running on a CPU with reduced capacity */
9999 if (sgs->group_misfit_task_load < load)
10000 sgs->group_misfit_task_load = load;
10004 sgs->group_capacity = group->sgc->capacity;
10006 sgs->group_weight = group->group_weight;
10008 /* Check if dst CPU is idle and preferred to this group */
10009 if (!local_group && env->idle && sgs->sum_h_nr_running &&
10010 sched_group_asym(env, sgs, group))
10011 sgs->group_asym_packing = 1;
10013 /* Check for loaded SMT group to be balanced to dst CPU */
10014 if (!local_group && smt_balance(env, sgs, group))
10015 sgs->group_smt_balance = 1;
10017 sgs->group_type = group_classify(env->sd->imbalance_pct, group, sgs);
10019 /* Computing avg_load makes sense only when group is overloaded */
10020 if (sgs->group_type == group_overloaded)
10021 sgs->avg_load = (sgs->group_load * SCHED_CAPACITY_SCALE) /
10022 sgs->group_capacity;
10026 * update_sd_pick_busiest - return 1 on busiest group
10027 * @env: The load balancing environment.
10028 * @sds: sched_domain statistics
10029 * @sg: sched_group candidate to be checked for being the busiest
10030 * @sgs: sched_group statistics
10032 * Determine if @sg is a busier group than the previously selected
10035 * Return: %true if @sg is a busier group than the previously selected
10036 * busiest group. %false otherwise.
10038 static bool update_sd_pick_busiest(struct lb_env *env,
10039 struct sd_lb_stats *sds,
10040 struct sched_group *sg,
10041 struct sg_lb_stats *sgs)
10043 struct sg_lb_stats *busiest = &sds->busiest_stat;
10045 /* Make sure that there is at least one task to pull */
10046 if (!sgs->sum_h_nr_running)
10050 * Don't try to pull misfit tasks we can't help.
10051 * We can use max_capacity here as reduction in capacity on some
10052 * CPUs in the group should either be possible to resolve
10053 * internally or be covered by avg_load imbalance (eventually).
10055 if ((env->sd->flags & SD_ASYM_CPUCAPACITY) &&
10056 (sgs->group_type == group_misfit_task) &&
10057 (!capacity_greater(capacity_of(env->dst_cpu), sg->sgc->max_capacity) ||
10058 sds->local_stat.group_type != group_has_spare))
10061 if (sgs->group_type > busiest->group_type)
10064 if (sgs->group_type < busiest->group_type)
10068 * The candidate and the current busiest group are the same type of
10069 * group. Let check which one is the busiest according to the type.
10072 switch (sgs->group_type) {
10073 case group_overloaded:
10074 /* Select the overloaded group with highest avg_load. */
10075 return sgs->avg_load > busiest->avg_load;
10077 case group_imbalanced:
10079 * Select the 1st imbalanced group as we don't have any way to
10080 * choose one more than another.
10084 case group_asym_packing:
10085 /* Prefer to move from lowest priority CPU's work */
10086 return sched_asym_prefer(sds->busiest->asym_prefer_cpu, sg->asym_prefer_cpu);
10088 case group_misfit_task:
10090 * If we have more than one misfit sg go with the biggest
10093 return sgs->group_misfit_task_load > busiest->group_misfit_task_load;
10095 case group_smt_balance:
10097 * Check if we have spare CPUs on either SMT group to
10098 * choose has spare or fully busy handling.
10100 if (sgs->idle_cpus != 0 || busiest->idle_cpus != 0)
10105 case group_fully_busy:
10107 * Select the fully busy group with highest avg_load. In
10108 * theory, there is no need to pull task from such kind of
10109 * group because tasks have all compute capacity that they need
10110 * but we can still improve the overall throughput by reducing
10111 * contention when accessing shared HW resources.
10113 * XXX for now avg_load is not computed and always 0 so we
10114 * select the 1st one, except if @sg is composed of SMT
10118 if (sgs->avg_load < busiest->avg_load)
10121 if (sgs->avg_load == busiest->avg_load) {
10123 * SMT sched groups need more help than non-SMT groups.
10124 * If @sg happens to also be SMT, either choice is good.
10126 if (sds->busiest->flags & SD_SHARE_CPUCAPACITY)
10132 case group_has_spare:
10134 * Do not pick sg with SMT CPUs over sg with pure CPUs,
10135 * as we do not want to pull task off SMT core with one task
10136 * and make the core idle.
10138 if (smt_vs_nonsmt_groups(sds->busiest, sg)) {
10139 if (sg->flags & SD_SHARE_CPUCAPACITY && sgs->sum_h_nr_running <= 1)
10147 * Select not overloaded group with lowest number of idle CPUs
10148 * and highest number of running tasks. We could also compare
10149 * the spare capacity which is more stable but it can end up
10150 * that the group has less spare capacity but finally more idle
10151 * CPUs which means less opportunity to pull tasks.
10153 if (sgs->idle_cpus > busiest->idle_cpus)
10155 else if ((sgs->idle_cpus == busiest->idle_cpus) &&
10156 (sgs->sum_nr_running <= busiest->sum_nr_running))
10163 * Candidate sg has no more than one task per CPU and has higher
10164 * per-CPU capacity. Migrating tasks to less capable CPUs may harm
10165 * throughput. Maximize throughput, power/energy consequences are not
10168 if ((env->sd->flags & SD_ASYM_CPUCAPACITY) &&
10169 (sgs->group_type <= group_fully_busy) &&
10170 (capacity_greater(sg->sgc->min_capacity, capacity_of(env->dst_cpu))))
10176 #ifdef CONFIG_NUMA_BALANCING
10177 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
10179 if (sgs->sum_h_nr_running > sgs->nr_numa_running)
10181 if (sgs->sum_h_nr_running > sgs->nr_preferred_running)
10186 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
10188 if (rq->nr_running > rq->nr_numa_running)
10190 if (rq->nr_running > rq->nr_preferred_running)
10195 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
10200 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
10204 #endif /* CONFIG_NUMA_BALANCING */
10207 struct sg_lb_stats;
10210 * task_running_on_cpu - return 1 if @p is running on @cpu.
10213 static unsigned int task_running_on_cpu(int cpu, struct task_struct *p)
10215 /* Task has no contribution or is new */
10216 if (cpu != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time))
10219 if (task_on_rq_queued(p))
10226 * idle_cpu_without - would a given CPU be idle without p ?
10227 * @cpu: the processor on which idleness is tested.
10228 * @p: task which should be ignored.
10230 * Return: 1 if the CPU would be idle. 0 otherwise.
10232 static int idle_cpu_without(int cpu, struct task_struct *p)
10234 struct rq *rq = cpu_rq(cpu);
10236 if (rq->curr != rq->idle && rq->curr != p)
10240 * rq->nr_running can't be used but an updated version without the
10241 * impact of p on cpu must be used instead. The updated nr_running
10242 * be computed and tested before calling idle_cpu_without().
10245 if (rq->ttwu_pending)
10252 * update_sg_wakeup_stats - Update sched_group's statistics for wakeup.
10253 * @sd: The sched_domain level to look for idlest group.
10254 * @group: sched_group whose statistics are to be updated.
10255 * @sgs: variable to hold the statistics for this group.
10256 * @p: The task for which we look for the idlest group/CPU.
10258 static inline void update_sg_wakeup_stats(struct sched_domain *sd,
10259 struct sched_group *group,
10260 struct sg_lb_stats *sgs,
10261 struct task_struct *p)
10265 memset(sgs, 0, sizeof(*sgs));
10267 /* Assume that task can't fit any CPU of the group */
10268 if (sd->flags & SD_ASYM_CPUCAPACITY)
10269 sgs->group_misfit_task_load = 1;
10271 for_each_cpu(i, sched_group_span(group)) {
10272 struct rq *rq = cpu_rq(i);
10273 unsigned int local;
10275 sgs->group_load += cpu_load_without(rq, p);
10276 sgs->group_util += cpu_util_without(i, p);
10277 sgs->group_runnable += cpu_runnable_without(rq, p);
10278 local = task_running_on_cpu(i, p);
10279 sgs->sum_h_nr_running += rq->cfs.h_nr_running - local;
10281 nr_running = rq->nr_running - local;
10282 sgs->sum_nr_running += nr_running;
10285 * No need to call idle_cpu_without() if nr_running is not 0
10287 if (!nr_running && idle_cpu_without(i, p))
10290 /* Check if task fits in the CPU */
10291 if (sd->flags & SD_ASYM_CPUCAPACITY &&
10292 sgs->group_misfit_task_load &&
10293 task_fits_cpu(p, i))
10294 sgs->group_misfit_task_load = 0;
10298 sgs->group_capacity = group->sgc->capacity;
10300 sgs->group_weight = group->group_weight;
10302 sgs->group_type = group_classify(sd->imbalance_pct, group, sgs);
10305 * Computing avg_load makes sense only when group is fully busy or
10308 if (sgs->group_type == group_fully_busy ||
10309 sgs->group_type == group_overloaded)
10310 sgs->avg_load = (sgs->group_load * SCHED_CAPACITY_SCALE) /
10311 sgs->group_capacity;
10314 static bool update_pick_idlest(struct sched_group *idlest,
10315 struct sg_lb_stats *idlest_sgs,
10316 struct sched_group *group,
10317 struct sg_lb_stats *sgs)
10319 if (sgs->group_type < idlest_sgs->group_type)
10322 if (sgs->group_type > idlest_sgs->group_type)
10326 * The candidate and the current idlest group are the same type of
10327 * group. Let check which one is the idlest according to the type.
10330 switch (sgs->group_type) {
10331 case group_overloaded:
10332 case group_fully_busy:
10333 /* Select the group with lowest avg_load. */
10334 if (idlest_sgs->avg_load <= sgs->avg_load)
10338 case group_imbalanced:
10339 case group_asym_packing:
10340 case group_smt_balance:
10341 /* Those types are not used in the slow wakeup path */
10344 case group_misfit_task:
10345 /* Select group with the highest max capacity */
10346 if (idlest->sgc->max_capacity >= group->sgc->max_capacity)
10350 case group_has_spare:
10351 /* Select group with most idle CPUs */
10352 if (idlest_sgs->idle_cpus > sgs->idle_cpus)
10355 /* Select group with lowest group_util */
10356 if (idlest_sgs->idle_cpus == sgs->idle_cpus &&
10357 idlest_sgs->group_util <= sgs->group_util)
10367 * sched_balance_find_dst_group() finds and returns the least busy CPU group within the
10370 * Assumes p is allowed on at least one CPU in sd.
10372 static struct sched_group *
10373 sched_balance_find_dst_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
10375 struct sched_group *idlest = NULL, *local = NULL, *group = sd->groups;
10376 struct sg_lb_stats local_sgs, tmp_sgs;
10377 struct sg_lb_stats *sgs;
10378 unsigned long imbalance;
10379 struct sg_lb_stats idlest_sgs = {
10380 .avg_load = UINT_MAX,
10381 .group_type = group_overloaded,
10387 /* Skip over this group if it has no CPUs allowed */
10388 if (!cpumask_intersects(sched_group_span(group),
10392 /* Skip over this group if no cookie matched */
10393 if (!sched_group_cookie_match(cpu_rq(this_cpu), p, group))
10396 local_group = cpumask_test_cpu(this_cpu,
10397 sched_group_span(group));
10406 update_sg_wakeup_stats(sd, group, sgs, p);
10408 if (!local_group && update_pick_idlest(idlest, &idlest_sgs, group, sgs)) {
10413 } while (group = group->next, group != sd->groups);
10416 /* There is no idlest group to push tasks to */
10420 /* The local group has been skipped because of CPU affinity */
10425 * If the local group is idler than the selected idlest group
10426 * don't try and push the task.
10428 if (local_sgs.group_type < idlest_sgs.group_type)
10432 * If the local group is busier than the selected idlest group
10433 * try and push the task.
10435 if (local_sgs.group_type > idlest_sgs.group_type)
10438 switch (local_sgs.group_type) {
10439 case group_overloaded:
10440 case group_fully_busy:
10442 /* Calculate allowed imbalance based on load */
10443 imbalance = scale_load_down(NICE_0_LOAD) *
10444 (sd->imbalance_pct-100) / 100;
10447 * When comparing groups across NUMA domains, it's possible for
10448 * the local domain to be very lightly loaded relative to the
10449 * remote domains but "imbalance" skews the comparison making
10450 * remote CPUs look much more favourable. When considering
10451 * cross-domain, add imbalance to the load on the remote node
10452 * and consider staying local.
10455 if ((sd->flags & SD_NUMA) &&
10456 ((idlest_sgs.avg_load + imbalance) >= local_sgs.avg_load))
10460 * If the local group is less loaded than the selected
10461 * idlest group don't try and push any tasks.
10463 if (idlest_sgs.avg_load >= (local_sgs.avg_load + imbalance))
10466 if (100 * local_sgs.avg_load <= sd->imbalance_pct * idlest_sgs.avg_load)
10470 case group_imbalanced:
10471 case group_asym_packing:
10472 case group_smt_balance:
10473 /* Those type are not used in the slow wakeup path */
10476 case group_misfit_task:
10477 /* Select group with the highest max capacity */
10478 if (local->sgc->max_capacity >= idlest->sgc->max_capacity)
10482 case group_has_spare:
10484 if (sd->flags & SD_NUMA) {
10485 int imb_numa_nr = sd->imb_numa_nr;
10486 #ifdef CONFIG_NUMA_BALANCING
10489 * If there is spare capacity at NUMA, try to select
10490 * the preferred node
10492 if (cpu_to_node(this_cpu) == p->numa_preferred_nid)
10495 idlest_cpu = cpumask_first(sched_group_span(idlest));
10496 if (cpu_to_node(idlest_cpu) == p->numa_preferred_nid)
10498 #endif /* CONFIG_NUMA_BALANCING */
10500 * Otherwise, keep the task close to the wakeup source
10501 * and improve locality if the number of running tasks
10502 * would remain below threshold where an imbalance is
10503 * allowed while accounting for the possibility the
10504 * task is pinned to a subset of CPUs. If there is a
10505 * real need of migration, periodic load balance will
10508 if (p->nr_cpus_allowed != NR_CPUS) {
10509 struct cpumask *cpus = this_cpu_cpumask_var_ptr(select_rq_mask);
10511 cpumask_and(cpus, sched_group_span(local), p->cpus_ptr);
10512 imb_numa_nr = min(cpumask_weight(cpus), sd->imb_numa_nr);
10515 imbalance = abs(local_sgs.idle_cpus - idlest_sgs.idle_cpus);
10516 if (!adjust_numa_imbalance(imbalance,
10517 local_sgs.sum_nr_running + 1,
10522 #endif /* CONFIG_NUMA */
10525 * Select group with highest number of idle CPUs. We could also
10526 * compare the utilization which is more stable but it can end
10527 * up that the group has less spare capacity but finally more
10528 * idle CPUs which means more opportunity to run task.
10530 if (local_sgs.idle_cpus >= idlest_sgs.idle_cpus)
10538 static void update_idle_cpu_scan(struct lb_env *env,
10539 unsigned long sum_util)
10541 struct sched_domain_shared *sd_share;
10542 int llc_weight, pct;
10545 * Update the number of CPUs to scan in LLC domain, which could
10546 * be used as a hint in select_idle_cpu(). The update of sd_share
10547 * could be expensive because it is within a shared cache line.
10548 * So the write of this hint only occurs during periodic load
10549 * balancing, rather than CPU_NEWLY_IDLE, because the latter
10550 * can fire way more frequently than the former.
10552 if (!sched_feat(SIS_UTIL) || env->idle == CPU_NEWLY_IDLE)
10555 llc_weight = per_cpu(sd_llc_size, env->dst_cpu);
10556 if (env->sd->span_weight != llc_weight)
10559 sd_share = rcu_dereference(per_cpu(sd_llc_shared, env->dst_cpu));
10564 * The number of CPUs to search drops as sum_util increases, when
10565 * sum_util hits 85% or above, the scan stops.
10566 * The reason to choose 85% as the threshold is because this is the
10567 * imbalance_pct(117) when a LLC sched group is overloaded.
10569 * let y = SCHED_CAPACITY_SCALE - p * x^2 [1]
10570 * and y'= y / SCHED_CAPACITY_SCALE
10572 * x is the ratio of sum_util compared to the CPU capacity:
10573 * x = sum_util / (llc_weight * SCHED_CAPACITY_SCALE)
10574 * y' is the ratio of CPUs to be scanned in the LLC domain,
10575 * and the number of CPUs to scan is calculated by:
10577 * nr_scan = llc_weight * y' [2]
10579 * When x hits the threshold of overloaded, AKA, when
10580 * x = 100 / pct, y drops to 0. According to [1],
10581 * p should be SCHED_CAPACITY_SCALE * pct^2 / 10000
10583 * Scale x by SCHED_CAPACITY_SCALE:
10584 * x' = sum_util / llc_weight; [3]
10586 * and finally [1] becomes:
10587 * y = SCHED_CAPACITY_SCALE -
10588 * x'^2 * pct^2 / (10000 * SCHED_CAPACITY_SCALE) [4]
10593 do_div(x, llc_weight);
10596 pct = env->sd->imbalance_pct;
10597 tmp = x * x * pct * pct;
10598 do_div(tmp, 10000 * SCHED_CAPACITY_SCALE);
10599 tmp = min_t(long, tmp, SCHED_CAPACITY_SCALE);
10600 y = SCHED_CAPACITY_SCALE - tmp;
10604 do_div(y, SCHED_CAPACITY_SCALE);
10605 if ((int)y != sd_share->nr_idle_scan)
10606 WRITE_ONCE(sd_share->nr_idle_scan, (int)y);
10610 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
10611 * @env: The load balancing environment.
10612 * @sds: variable to hold the statistics for this sched_domain.
10615 static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
10617 struct sched_group *sg = env->sd->groups;
10618 struct sg_lb_stats *local = &sds->local_stat;
10619 struct sg_lb_stats tmp_sgs;
10620 unsigned long sum_util = 0;
10621 bool sg_overloaded = 0, sg_overutilized = 0;
10624 struct sg_lb_stats *sgs = &tmp_sgs;
10627 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_span(sg));
10632 if (env->idle != CPU_NEWLY_IDLE ||
10633 time_after_eq(jiffies, sg->sgc->next_update))
10634 update_group_capacity(env->sd, env->dst_cpu);
10637 update_sg_lb_stats(env, sds, sg, sgs, &sg_overloaded, &sg_overutilized);
10639 if (!local_group && update_sd_pick_busiest(env, sds, sg, sgs)) {
10641 sds->busiest_stat = *sgs;
10644 /* Now, start updating sd_lb_stats */
10645 sds->total_load += sgs->group_load;
10646 sds->total_capacity += sgs->group_capacity;
10648 sum_util += sgs->group_util;
10650 } while (sg != env->sd->groups);
10653 * Indicate that the child domain of the busiest group prefers tasks
10654 * go to a child's sibling domains first. NB the flags of a sched group
10655 * are those of the child domain.
10658 sds->prefer_sibling = !!(sds->busiest->flags & SD_PREFER_SIBLING);
10661 if (env->sd->flags & SD_NUMA)
10662 env->fbq_type = fbq_classify_group(&sds->busiest_stat);
10664 if (!env->sd->parent) {
10665 /* update overload indicator if we are at root domain */
10666 set_rd_overloaded(env->dst_rq->rd, sg_overloaded);
10668 /* Update over-utilization (tipping point, U >= 0) indicator */
10669 set_rd_overutilized(env->dst_rq->rd, sg_overutilized);
10670 } else if (sg_overutilized) {
10671 set_rd_overutilized(env->dst_rq->rd, sg_overutilized);
10674 update_idle_cpu_scan(env, sum_util);
10678 * calculate_imbalance - Calculate the amount of imbalance present within the
10679 * groups of a given sched_domain during load balance.
10680 * @env: load balance environment
10681 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
10683 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
10685 struct sg_lb_stats *local, *busiest;
10687 local = &sds->local_stat;
10688 busiest = &sds->busiest_stat;
10690 if (busiest->group_type == group_misfit_task) {
10691 if (env->sd->flags & SD_ASYM_CPUCAPACITY) {
10692 /* Set imbalance to allow misfit tasks to be balanced. */
10693 env->migration_type = migrate_misfit;
10694 env->imbalance = 1;
10697 * Set load imbalance to allow moving task from cpu
10698 * with reduced capacity.
10700 env->migration_type = migrate_load;
10701 env->imbalance = busiest->group_misfit_task_load;
10706 if (busiest->group_type == group_asym_packing) {
10708 * In case of asym capacity, we will try to migrate all load to
10709 * the preferred CPU.
10711 env->migration_type = migrate_task;
10712 env->imbalance = busiest->sum_h_nr_running;
10716 if (busiest->group_type == group_smt_balance) {
10717 /* Reduce number of tasks sharing CPU capacity */
10718 env->migration_type = migrate_task;
10719 env->imbalance = 1;
10723 if (busiest->group_type == group_imbalanced) {
10725 * In the group_imb case we cannot rely on group-wide averages
10726 * to ensure CPU-load equilibrium, try to move any task to fix
10727 * the imbalance. The next load balance will take care of
10728 * balancing back the system.
10730 env->migration_type = migrate_task;
10731 env->imbalance = 1;
10736 * Try to use spare capacity of local group without overloading it or
10737 * emptying busiest.
10739 if (local->group_type == group_has_spare) {
10740 if ((busiest->group_type > group_fully_busy) &&
10741 !(env->sd->flags & SD_SHARE_LLC)) {
10743 * If busiest is overloaded, try to fill spare
10744 * capacity. This might end up creating spare capacity
10745 * in busiest or busiest still being overloaded but
10746 * there is no simple way to directly compute the
10747 * amount of load to migrate in order to balance the
10750 env->migration_type = migrate_util;
10751 env->imbalance = max(local->group_capacity, local->group_util) -
10755 * In some cases, the group's utilization is max or even
10756 * higher than capacity because of migrations but the
10757 * local CPU is (newly) idle. There is at least one
10758 * waiting task in this overloaded busiest group. Let's
10761 if (env->idle && env->imbalance == 0) {
10762 env->migration_type = migrate_task;
10763 env->imbalance = 1;
10769 if (busiest->group_weight == 1 || sds->prefer_sibling) {
10771 * When prefer sibling, evenly spread running tasks on
10774 env->migration_type = migrate_task;
10775 env->imbalance = sibling_imbalance(env, sds, busiest, local);
10779 * If there is no overload, we just want to even the number of
10782 env->migration_type = migrate_task;
10783 env->imbalance = max_t(long, 0,
10784 (local->idle_cpus - busiest->idle_cpus));
10788 /* Consider allowing a small imbalance between NUMA groups */
10789 if (env->sd->flags & SD_NUMA) {
10790 env->imbalance = adjust_numa_imbalance(env->imbalance,
10791 local->sum_nr_running + 1,
10792 env->sd->imb_numa_nr);
10796 /* Number of tasks to move to restore balance */
10797 env->imbalance >>= 1;
10803 * Local is fully busy but has to take more load to relieve the
10806 if (local->group_type < group_overloaded) {
10808 * Local will become overloaded so the avg_load metrics are
10812 local->avg_load = (local->group_load * SCHED_CAPACITY_SCALE) /
10813 local->group_capacity;
10816 * If the local group is more loaded than the selected
10817 * busiest group don't try to pull any tasks.
10819 if (local->avg_load >= busiest->avg_load) {
10820 env->imbalance = 0;
10824 sds->avg_load = (sds->total_load * SCHED_CAPACITY_SCALE) /
10825 sds->total_capacity;
10828 * If the local group is more loaded than the average system
10829 * load, don't try to pull any tasks.
10831 if (local->avg_load >= sds->avg_load) {
10832 env->imbalance = 0;
10839 * Both group are or will become overloaded and we're trying to get all
10840 * the CPUs to the average_load, so we don't want to push ourselves
10841 * above the average load, nor do we wish to reduce the max loaded CPU
10842 * below the average load. At the same time, we also don't want to
10843 * reduce the group load below the group capacity. Thus we look for
10844 * the minimum possible imbalance.
10846 env->migration_type = migrate_load;
10847 env->imbalance = min(
10848 (busiest->avg_load - sds->avg_load) * busiest->group_capacity,
10849 (sds->avg_load - local->avg_load) * local->group_capacity
10850 ) / SCHED_CAPACITY_SCALE;
10853 /******* sched_balance_find_src_group() helpers end here *********************/
10856 * Decision matrix according to the local and busiest group type:
10858 * busiest \ local has_spare fully_busy misfit asym imbalanced overloaded
10859 * has_spare nr_idle balanced N/A N/A balanced balanced
10860 * fully_busy nr_idle nr_idle N/A N/A balanced balanced
10861 * misfit_task force N/A N/A N/A N/A N/A
10862 * asym_packing force force N/A N/A force force
10863 * imbalanced force force N/A N/A force force
10864 * overloaded force force N/A N/A force avg_load
10866 * N/A : Not Applicable because already filtered while updating
10868 * balanced : The system is balanced for these 2 groups.
10869 * force : Calculate the imbalance as load migration is probably needed.
10870 * avg_load : Only if imbalance is significant enough.
10871 * nr_idle : dst_cpu is not busy and the number of idle CPUs is quite
10872 * different in groups.
10876 * sched_balance_find_src_group - Returns the busiest group within the sched_domain
10877 * if there is an imbalance.
10878 * @env: The load balancing environment.
10880 * Also calculates the amount of runnable load which should be moved
10881 * to restore balance.
10883 * Return: - The busiest group if imbalance exists.
10885 static struct sched_group *sched_balance_find_src_group(struct lb_env *env)
10887 struct sg_lb_stats *local, *busiest;
10888 struct sd_lb_stats sds;
10890 init_sd_lb_stats(&sds);
10893 * Compute the various statistics relevant for load balancing at
10896 update_sd_lb_stats(env, &sds);
10898 /* There is no busy sibling group to pull tasks from */
10902 busiest = &sds.busiest_stat;
10904 /* Misfit tasks should be dealt with regardless of the avg load */
10905 if (busiest->group_type == group_misfit_task)
10906 goto force_balance;
10908 if (!is_rd_overutilized(env->dst_rq->rd) &&
10909 rcu_dereference(env->dst_rq->rd->pd))
10912 /* ASYM feature bypasses nice load balance check */
10913 if (busiest->group_type == group_asym_packing)
10914 goto force_balance;
10917 * If the busiest group is imbalanced the below checks don't
10918 * work because they assume all things are equal, which typically
10919 * isn't true due to cpus_ptr constraints and the like.
10921 if (busiest->group_type == group_imbalanced)
10922 goto force_balance;
10924 local = &sds.local_stat;
10926 * If the local group is busier than the selected busiest group
10927 * don't try and pull any tasks.
10929 if (local->group_type > busiest->group_type)
10933 * When groups are overloaded, use the avg_load to ensure fairness
10936 if (local->group_type == group_overloaded) {
10938 * If the local group is more loaded than the selected
10939 * busiest group don't try to pull any tasks.
10941 if (local->avg_load >= busiest->avg_load)
10944 /* XXX broken for overlapping NUMA groups */
10945 sds.avg_load = (sds.total_load * SCHED_CAPACITY_SCALE) /
10946 sds.total_capacity;
10949 * Don't pull any tasks if this group is already above the
10950 * domain average load.
10952 if (local->avg_load >= sds.avg_load)
10956 * If the busiest group is more loaded, use imbalance_pct to be
10959 if (100 * busiest->avg_load <=
10960 env->sd->imbalance_pct * local->avg_load)
10965 * Try to move all excess tasks to a sibling domain of the busiest
10966 * group's child domain.
10968 if (sds.prefer_sibling && local->group_type == group_has_spare &&
10969 sibling_imbalance(env, &sds, busiest, local) > 1)
10970 goto force_balance;
10972 if (busiest->group_type != group_overloaded) {
10975 * If the busiest group is not overloaded (and as a
10976 * result the local one too) but this CPU is already
10977 * busy, let another idle CPU try to pull task.
10982 if (busiest->group_type == group_smt_balance &&
10983 smt_vs_nonsmt_groups(sds.local, sds.busiest)) {
10984 /* Let non SMT CPU pull from SMT CPU sharing with sibling */
10985 goto force_balance;
10988 if (busiest->group_weight > 1 &&
10989 local->idle_cpus <= (busiest->idle_cpus + 1)) {
10991 * If the busiest group is not overloaded
10992 * and there is no imbalance between this and busiest
10993 * group wrt idle CPUs, it is balanced. The imbalance
10994 * becomes significant if the diff is greater than 1
10995 * otherwise we might end up to just move the imbalance
10996 * on another group. Of course this applies only if
10997 * there is more than 1 CPU per group.
11002 if (busiest->sum_h_nr_running == 1) {
11004 * busiest doesn't have any tasks waiting to run
11011 /* Looks like there is an imbalance. Compute it */
11012 calculate_imbalance(env, &sds);
11013 return env->imbalance ? sds.busiest : NULL;
11016 env->imbalance = 0;
11021 * sched_balance_find_src_rq - find the busiest runqueue among the CPUs in the group.
11023 static struct rq *sched_balance_find_src_rq(struct lb_env *env,
11024 struct sched_group *group)
11026 struct rq *busiest = NULL, *rq;
11027 unsigned long busiest_util = 0, busiest_load = 0, busiest_capacity = 1;
11028 unsigned int busiest_nr = 0;
11031 for_each_cpu_and(i, sched_group_span(group), env->cpus) {
11032 unsigned long capacity, load, util;
11033 unsigned int nr_running;
11037 rt = fbq_classify_rq(rq);
11040 * We classify groups/runqueues into three groups:
11041 * - regular: there are !numa tasks
11042 * - remote: there are numa tasks that run on the 'wrong' node
11043 * - all: there is no distinction
11045 * In order to avoid migrating ideally placed numa tasks,
11046 * ignore those when there's better options.
11048 * If we ignore the actual busiest queue to migrate another
11049 * task, the next balance pass can still reduce the busiest
11050 * queue by moving tasks around inside the node.
11052 * If we cannot move enough load due to this classification
11053 * the next pass will adjust the group classification and
11054 * allow migration of more tasks.
11056 * Both cases only affect the total convergence complexity.
11058 if (rt > env->fbq_type)
11061 nr_running = rq->cfs.h_nr_running;
11065 capacity = capacity_of(i);
11068 * For ASYM_CPUCAPACITY domains, don't pick a CPU that could
11069 * eventually lead to active_balancing high->low capacity.
11070 * Higher per-CPU capacity is considered better than balancing
11073 if (env->sd->flags & SD_ASYM_CPUCAPACITY &&
11074 !capacity_greater(capacity_of(env->dst_cpu), capacity) &&
11079 * Make sure we only pull tasks from a CPU of lower priority
11080 * when balancing between SMT siblings.
11082 * If balancing between cores, let lower priority CPUs help
11083 * SMT cores with more than one busy sibling.
11085 if (sched_asym(env->sd, i, env->dst_cpu) && nr_running == 1)
11088 switch (env->migration_type) {
11091 * When comparing with load imbalance, use cpu_load()
11092 * which is not scaled with the CPU capacity.
11094 load = cpu_load(rq);
11096 if (nr_running == 1 && load > env->imbalance &&
11097 !check_cpu_capacity(rq, env->sd))
11101 * For the load comparisons with the other CPUs,
11102 * consider the cpu_load() scaled with the CPU
11103 * capacity, so that the load can be moved away
11104 * from the CPU that is potentially running at a
11107 * Thus we're looking for max(load_i / capacity_i),
11108 * crosswise multiplication to rid ourselves of the
11109 * division works out to:
11110 * load_i * capacity_j > load_j * capacity_i;
11111 * where j is our previous maximum.
11113 if (load * busiest_capacity > busiest_load * capacity) {
11114 busiest_load = load;
11115 busiest_capacity = capacity;
11121 util = cpu_util_cfs_boost(i);
11124 * Don't try to pull utilization from a CPU with one
11125 * running task. Whatever its utilization, we will fail
11128 if (nr_running <= 1)
11131 if (busiest_util < util) {
11132 busiest_util = util;
11138 if (busiest_nr < nr_running) {
11139 busiest_nr = nr_running;
11144 case migrate_misfit:
11146 * For ASYM_CPUCAPACITY domains with misfit tasks we
11147 * simply seek the "biggest" misfit task.
11149 if (rq->misfit_task_load > busiest_load) {
11150 busiest_load = rq->misfit_task_load;
11163 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
11164 * so long as it is large enough.
11166 #define MAX_PINNED_INTERVAL 512
11169 asym_active_balance(struct lb_env *env)
11172 * ASYM_PACKING needs to force migrate tasks from busy but lower
11173 * priority CPUs in order to pack all tasks in the highest priority
11174 * CPUs. When done between cores, do it only if the whole core if the
11175 * whole core is idle.
11177 * If @env::src_cpu is an SMT core with busy siblings, let
11178 * the lower priority @env::dst_cpu help it. Do not follow
11181 return env->idle && sched_use_asym_prio(env->sd, env->dst_cpu) &&
11182 (sched_asym_prefer(env->dst_cpu, env->src_cpu) ||
11183 !sched_use_asym_prio(env->sd, env->src_cpu));
11187 imbalanced_active_balance(struct lb_env *env)
11189 struct sched_domain *sd = env->sd;
11192 * The imbalanced case includes the case of pinned tasks preventing a fair
11193 * distribution of the load on the system but also the even distribution of the
11194 * threads on a system with spare capacity
11196 if ((env->migration_type == migrate_task) &&
11197 (sd->nr_balance_failed > sd->cache_nice_tries+2))
11203 static int need_active_balance(struct lb_env *env)
11205 struct sched_domain *sd = env->sd;
11207 if (asym_active_balance(env))
11210 if (imbalanced_active_balance(env))
11214 * The dst_cpu is idle and the src_cpu CPU has only 1 CFS task.
11215 * It's worth migrating the task if the src_cpu's capacity is reduced
11216 * because of other sched_class or IRQs if more capacity stays
11217 * available on dst_cpu.
11220 (env->src_rq->cfs.h_nr_running == 1)) {
11221 if ((check_cpu_capacity(env->src_rq, sd)) &&
11222 (capacity_of(env->src_cpu)*sd->imbalance_pct < capacity_of(env->dst_cpu)*100))
11226 if (env->migration_type == migrate_misfit)
11232 static int active_load_balance_cpu_stop(void *data);
11234 static int should_we_balance(struct lb_env *env)
11236 struct cpumask *swb_cpus = this_cpu_cpumask_var_ptr(should_we_balance_tmpmask);
11237 struct sched_group *sg = env->sd->groups;
11238 int cpu, idle_smt = -1;
11241 * Ensure the balancing environment is consistent; can happen
11242 * when the softirq triggers 'during' hotplug.
11244 if (!cpumask_test_cpu(env->dst_cpu, env->cpus))
11248 * In the newly idle case, we will allow all the CPUs
11249 * to do the newly idle load balance.
11251 * However, we bail out if we already have tasks or a wakeup pending,
11252 * to optimize wakeup latency.
11254 if (env->idle == CPU_NEWLY_IDLE) {
11255 if (env->dst_rq->nr_running > 0 || env->dst_rq->ttwu_pending)
11260 cpumask_copy(swb_cpus, group_balance_mask(sg));
11261 /* Try to find first idle CPU */
11262 for_each_cpu_and(cpu, swb_cpus, env->cpus) {
11263 if (!idle_cpu(cpu))
11267 * Don't balance to idle SMT in busy core right away when
11268 * balancing cores, but remember the first idle SMT CPU for
11269 * later consideration. Find CPU on an idle core first.
11271 if (!(env->sd->flags & SD_SHARE_CPUCAPACITY) && !is_core_idle(cpu)) {
11272 if (idle_smt == -1)
11275 * If the core is not idle, and first SMT sibling which is
11276 * idle has been found, then its not needed to check other
11277 * SMT siblings for idleness:
11279 #ifdef CONFIG_SCHED_SMT
11280 cpumask_andnot(swb_cpus, swb_cpus, cpu_smt_mask(cpu));
11286 * Are we the first idle core in a non-SMT domain or higher,
11287 * or the first idle CPU in a SMT domain?
11289 return cpu == env->dst_cpu;
11292 /* Are we the first idle CPU with busy siblings? */
11293 if (idle_smt != -1)
11294 return idle_smt == env->dst_cpu;
11296 /* Are we the first CPU of this group ? */
11297 return group_balance_cpu(sg) == env->dst_cpu;
11301 * Check this_cpu to ensure it is balanced within domain. Attempt to move
11302 * tasks if there is an imbalance.
11304 static int sched_balance_rq(int this_cpu, struct rq *this_rq,
11305 struct sched_domain *sd, enum cpu_idle_type idle,
11306 int *continue_balancing)
11308 int ld_moved, cur_ld_moved, active_balance = 0;
11309 struct sched_domain *sd_parent = sd->parent;
11310 struct sched_group *group;
11311 struct rq *busiest;
11312 struct rq_flags rf;
11313 struct cpumask *cpus = this_cpu_cpumask_var_ptr(load_balance_mask);
11314 struct lb_env env = {
11316 .dst_cpu = this_cpu,
11318 .dst_grpmask = group_balance_mask(sd->groups),
11320 .loop_break = SCHED_NR_MIGRATE_BREAK,
11323 .tasks = LIST_HEAD_INIT(env.tasks),
11326 cpumask_and(cpus, sched_domain_span(sd), cpu_active_mask);
11328 schedstat_inc(sd->lb_count[idle]);
11331 if (!should_we_balance(&env)) {
11332 *continue_balancing = 0;
11336 group = sched_balance_find_src_group(&env);
11338 schedstat_inc(sd->lb_nobusyg[idle]);
11342 busiest = sched_balance_find_src_rq(&env, group);
11344 schedstat_inc(sd->lb_nobusyq[idle]);
11348 WARN_ON_ONCE(busiest == env.dst_rq);
11350 schedstat_add(sd->lb_imbalance[idle], env.imbalance);
11352 env.src_cpu = busiest->cpu;
11353 env.src_rq = busiest;
11356 /* Clear this flag as soon as we find a pullable task */
11357 env.flags |= LBF_ALL_PINNED;
11358 if (busiest->nr_running > 1) {
11360 * Attempt to move tasks. If sched_balance_find_src_group has found
11361 * an imbalance but busiest->nr_running <= 1, the group is
11362 * still unbalanced. ld_moved simply stays zero, so it is
11363 * correctly treated as an imbalance.
11365 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
11368 rq_lock_irqsave(busiest, &rf);
11369 update_rq_clock(busiest);
11372 * cur_ld_moved - load moved in current iteration
11373 * ld_moved - cumulative load moved across iterations
11375 cur_ld_moved = detach_tasks(&env);
11378 * We've detached some tasks from busiest_rq. Every
11379 * task is masked "TASK_ON_RQ_MIGRATING", so we can safely
11380 * unlock busiest->lock, and we are able to be sure
11381 * that nobody can manipulate the tasks in parallel.
11382 * See task_rq_lock() family for the details.
11385 rq_unlock(busiest, &rf);
11387 if (cur_ld_moved) {
11388 attach_tasks(&env);
11389 ld_moved += cur_ld_moved;
11392 local_irq_restore(rf.flags);
11394 if (env.flags & LBF_NEED_BREAK) {
11395 env.flags &= ~LBF_NEED_BREAK;
11396 /* Stop if we tried all running tasks */
11397 if (env.loop < busiest->nr_running)
11402 * Revisit (affine) tasks on src_cpu that couldn't be moved to
11403 * us and move them to an alternate dst_cpu in our sched_group
11404 * where they can run. The upper limit on how many times we
11405 * iterate on same src_cpu is dependent on number of CPUs in our
11408 * This changes load balance semantics a bit on who can move
11409 * load to a given_cpu. In addition to the given_cpu itself
11410 * (or a ilb_cpu acting on its behalf where given_cpu is
11411 * nohz-idle), we now have balance_cpu in a position to move
11412 * load to given_cpu. In rare situations, this may cause
11413 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
11414 * _independently_ and at _same_ time to move some load to
11415 * given_cpu) causing excess load to be moved to given_cpu.
11416 * This however should not happen so much in practice and
11417 * moreover subsequent load balance cycles should correct the
11418 * excess load moved.
11420 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
11422 /* Prevent to re-select dst_cpu via env's CPUs */
11423 __cpumask_clear_cpu(env.dst_cpu, env.cpus);
11425 env.dst_rq = cpu_rq(env.new_dst_cpu);
11426 env.dst_cpu = env.new_dst_cpu;
11427 env.flags &= ~LBF_DST_PINNED;
11429 env.loop_break = SCHED_NR_MIGRATE_BREAK;
11432 * Go back to "more_balance" rather than "redo" since we
11433 * need to continue with same src_cpu.
11439 * We failed to reach balance because of affinity.
11442 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
11444 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0)
11445 *group_imbalance = 1;
11448 /* All tasks on this runqueue were pinned by CPU affinity */
11449 if (unlikely(env.flags & LBF_ALL_PINNED)) {
11450 __cpumask_clear_cpu(cpu_of(busiest), cpus);
11452 * Attempting to continue load balancing at the current
11453 * sched_domain level only makes sense if there are
11454 * active CPUs remaining as possible busiest CPUs to
11455 * pull load from which are not contained within the
11456 * destination group that is receiving any migrated
11459 if (!cpumask_subset(cpus, env.dst_grpmask)) {
11461 env.loop_break = SCHED_NR_MIGRATE_BREAK;
11464 goto out_all_pinned;
11469 schedstat_inc(sd->lb_failed[idle]);
11471 * Increment the failure counter only on periodic balance.
11472 * We do not want newidle balance, which can be very
11473 * frequent, pollute the failure counter causing
11474 * excessive cache_hot migrations and active balances.
11476 * Similarly for migration_misfit which is not related to
11477 * load/util migration, don't pollute nr_balance_failed.
11479 if (idle != CPU_NEWLY_IDLE &&
11480 env.migration_type != migrate_misfit)
11481 sd->nr_balance_failed++;
11483 if (need_active_balance(&env)) {
11484 unsigned long flags;
11486 raw_spin_rq_lock_irqsave(busiest, flags);
11489 * Don't kick the active_load_balance_cpu_stop,
11490 * if the curr task on busiest CPU can't be
11491 * moved to this_cpu:
11493 if (!cpumask_test_cpu(this_cpu, busiest->curr->cpus_ptr)) {
11494 raw_spin_rq_unlock_irqrestore(busiest, flags);
11495 goto out_one_pinned;
11498 /* Record that we found at least one task that could run on this_cpu */
11499 env.flags &= ~LBF_ALL_PINNED;
11502 * ->active_balance synchronizes accesses to
11503 * ->active_balance_work. Once set, it's cleared
11504 * only after active load balance is finished.
11506 if (!busiest->active_balance) {
11507 busiest->active_balance = 1;
11508 busiest->push_cpu = this_cpu;
11509 active_balance = 1;
11513 raw_spin_rq_unlock_irqrestore(busiest, flags);
11514 if (active_balance) {
11515 stop_one_cpu_nowait(cpu_of(busiest),
11516 active_load_balance_cpu_stop, busiest,
11517 &busiest->active_balance_work);
11522 sd->nr_balance_failed = 0;
11525 if (likely(!active_balance) || need_active_balance(&env)) {
11526 /* We were unbalanced, so reset the balancing interval */
11527 sd->balance_interval = sd->min_interval;
11534 * We reach balance although we may have faced some affinity
11535 * constraints. Clear the imbalance flag only if other tasks got
11536 * a chance to move and fix the imbalance.
11538 if (sd_parent && !(env.flags & LBF_ALL_PINNED)) {
11539 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
11541 if (*group_imbalance)
11542 *group_imbalance = 0;
11547 * We reach balance because all tasks are pinned at this level so
11548 * we can't migrate them. Let the imbalance flag set so parent level
11549 * can try to migrate them.
11551 schedstat_inc(sd->lb_balanced[idle]);
11553 sd->nr_balance_failed = 0;
11559 * sched_balance_newidle() disregards balance intervals, so we could
11560 * repeatedly reach this code, which would lead to balance_interval
11561 * skyrocketing in a short amount of time. Skip the balance_interval
11562 * increase logic to avoid that.
11564 * Similarly misfit migration which is not necessarily an indication of
11565 * the system being busy and requires lb to backoff to let it settle
11568 if (env.idle == CPU_NEWLY_IDLE ||
11569 env.migration_type == migrate_misfit)
11572 /* tune up the balancing interval */
11573 if ((env.flags & LBF_ALL_PINNED &&
11574 sd->balance_interval < MAX_PINNED_INTERVAL) ||
11575 sd->balance_interval < sd->max_interval)
11576 sd->balance_interval *= 2;
11581 static inline unsigned long
11582 get_sd_balance_interval(struct sched_domain *sd, int cpu_busy)
11584 unsigned long interval = sd->balance_interval;
11587 interval *= sd->busy_factor;
11589 /* scale ms to jiffies */
11590 interval = msecs_to_jiffies(interval);
11593 * Reduce likelihood of busy balancing at higher domains racing with
11594 * balancing at lower domains by preventing their balancing periods
11595 * from being multiples of each other.
11600 interval = clamp(interval, 1UL, max_load_balance_interval);
11606 update_next_balance(struct sched_domain *sd, unsigned long *next_balance)
11608 unsigned long interval, next;
11610 /* used by idle balance, so cpu_busy = 0 */
11611 interval = get_sd_balance_interval(sd, 0);
11612 next = sd->last_balance + interval;
11614 if (time_after(*next_balance, next))
11615 *next_balance = next;
11619 * active_load_balance_cpu_stop is run by the CPU stopper. It pushes
11620 * running tasks off the busiest CPU onto idle CPUs. It requires at
11621 * least 1 task to be running on each physical CPU where possible, and
11622 * avoids physical / logical imbalances.
11624 static int active_load_balance_cpu_stop(void *data)
11626 struct rq *busiest_rq = data;
11627 int busiest_cpu = cpu_of(busiest_rq);
11628 int target_cpu = busiest_rq->push_cpu;
11629 struct rq *target_rq = cpu_rq(target_cpu);
11630 struct sched_domain *sd;
11631 struct task_struct *p = NULL;
11632 struct rq_flags rf;
11634 rq_lock_irq(busiest_rq, &rf);
11636 * Between queueing the stop-work and running it is a hole in which
11637 * CPUs can become inactive. We should not move tasks from or to
11640 if (!cpu_active(busiest_cpu) || !cpu_active(target_cpu))
11643 /* Make sure the requested CPU hasn't gone down in the meantime: */
11644 if (unlikely(busiest_cpu != smp_processor_id() ||
11645 !busiest_rq->active_balance))
11648 /* Is there any task to move? */
11649 if (busiest_rq->nr_running <= 1)
11653 * This condition is "impossible", if it occurs
11654 * we need to fix it. Originally reported by
11655 * Bjorn Helgaas on a 128-CPU setup.
11657 WARN_ON_ONCE(busiest_rq == target_rq);
11659 /* Search for an sd spanning us and the target CPU. */
11661 for_each_domain(target_cpu, sd) {
11662 if (cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
11667 struct lb_env env = {
11669 .dst_cpu = target_cpu,
11670 .dst_rq = target_rq,
11671 .src_cpu = busiest_rq->cpu,
11672 .src_rq = busiest_rq,
11674 .flags = LBF_ACTIVE_LB,
11677 schedstat_inc(sd->alb_count);
11678 update_rq_clock(busiest_rq);
11680 p = detach_one_task(&env);
11682 schedstat_inc(sd->alb_pushed);
11683 /* Active balancing done, reset the failure counter. */
11684 sd->nr_balance_failed = 0;
11686 schedstat_inc(sd->alb_failed);
11691 busiest_rq->active_balance = 0;
11692 rq_unlock(busiest_rq, &rf);
11695 attach_one_task(target_rq, p);
11697 local_irq_enable();
11703 * This flag serializes load-balancing passes over large domains
11704 * (above the NODE topology level) - only one load-balancing instance
11705 * may run at a time, to reduce overhead on very large systems with
11706 * lots of CPUs and large NUMA distances.
11708 * - Note that load-balancing passes triggered while another one
11709 * is executing are skipped and not re-tried.
11711 * - Also note that this does not serialize rebalance_domains()
11712 * execution, as non-SD_SERIALIZE domains will still be
11713 * load-balanced in parallel.
11715 static atomic_t sched_balance_running = ATOMIC_INIT(0);
11718 * Scale the max sched_balance_rq interval with the number of CPUs in the system.
11719 * This trades load-balance latency on larger machines for less cross talk.
11721 void update_max_interval(void)
11723 max_load_balance_interval = HZ*num_online_cpus()/10;
11726 static inline bool update_newidle_cost(struct sched_domain *sd, u64 cost)
11728 if (cost > sd->max_newidle_lb_cost) {
11730 * Track max cost of a domain to make sure to not delay the
11731 * next wakeup on the CPU.
11733 sd->max_newidle_lb_cost = cost;
11734 sd->last_decay_max_lb_cost = jiffies;
11735 } else if (time_after(jiffies, sd->last_decay_max_lb_cost + HZ)) {
11737 * Decay the newidle max times by ~1% per second to ensure that
11738 * it is not outdated and the current max cost is actually
11741 sd->max_newidle_lb_cost = (sd->max_newidle_lb_cost * 253) / 256;
11742 sd->last_decay_max_lb_cost = jiffies;
11751 * It checks each scheduling domain to see if it is due to be balanced,
11752 * and initiates a balancing operation if so.
11754 * Balancing parameters are set up in init_sched_domains.
11756 static void sched_balance_domains(struct rq *rq, enum cpu_idle_type idle)
11758 int continue_balancing = 1;
11760 int busy = idle != CPU_IDLE && !sched_idle_cpu(cpu);
11761 unsigned long interval;
11762 struct sched_domain *sd;
11763 /* Earliest time when we have to do rebalance again */
11764 unsigned long next_balance = jiffies + 60*HZ;
11765 int update_next_balance = 0;
11766 int need_serialize, need_decay = 0;
11770 for_each_domain(cpu, sd) {
11772 * Decay the newidle max times here because this is a regular
11773 * visit to all the domains.
11775 need_decay = update_newidle_cost(sd, 0);
11776 max_cost += sd->max_newidle_lb_cost;
11779 * Stop the load balance at this level. There is another
11780 * CPU in our sched group which is doing load balancing more
11783 if (!continue_balancing) {
11789 interval = get_sd_balance_interval(sd, busy);
11791 need_serialize = sd->flags & SD_SERIALIZE;
11792 if (need_serialize) {
11793 if (atomic_cmpxchg_acquire(&sched_balance_running, 0, 1))
11797 if (time_after_eq(jiffies, sd->last_balance + interval)) {
11798 if (sched_balance_rq(cpu, rq, sd, idle, &continue_balancing)) {
11800 * The LBF_DST_PINNED logic could have changed
11801 * env->dst_cpu, so we can't know our idle
11802 * state even if we migrated tasks. Update it.
11804 idle = idle_cpu(cpu);
11805 busy = !idle && !sched_idle_cpu(cpu);
11807 sd->last_balance = jiffies;
11808 interval = get_sd_balance_interval(sd, busy);
11810 if (need_serialize)
11811 atomic_set_release(&sched_balance_running, 0);
11813 if (time_after(next_balance, sd->last_balance + interval)) {
11814 next_balance = sd->last_balance + interval;
11815 update_next_balance = 1;
11820 * Ensure the rq-wide value also decays but keep it at a
11821 * reasonable floor to avoid funnies with rq->avg_idle.
11823 rq->max_idle_balance_cost =
11824 max((u64)sysctl_sched_migration_cost, max_cost);
11829 * next_balance will be updated only when there is a need.
11830 * When the cpu is attached to null domain for ex, it will not be
11833 if (likely(update_next_balance))
11834 rq->next_balance = next_balance;
11838 static inline int on_null_domain(struct rq *rq)
11840 return unlikely(!rcu_dereference_sched(rq->sd));
11843 #ifdef CONFIG_NO_HZ_COMMON
11845 * NOHZ idle load balancing (ILB) details:
11847 * - When one of the busy CPUs notices that there may be an idle rebalancing
11848 * needed, they will kick the idle load balancer, which then does idle
11849 * load balancing for all the idle CPUs.
11851 * - HK_TYPE_MISC CPUs are used for this task, because HK_TYPE_SCHED is not set
11854 static inline int find_new_ilb(void)
11856 const struct cpumask *hk_mask;
11859 hk_mask = housekeeping_cpumask(HK_TYPE_MISC);
11861 for_each_cpu_and(ilb_cpu, nohz.idle_cpus_mask, hk_mask) {
11863 if (ilb_cpu == smp_processor_id())
11866 if (idle_cpu(ilb_cpu))
11874 * Kick a CPU to do the NOHZ balancing, if it is time for it, via a cross-CPU
11875 * SMP function call (IPI).
11877 * We pick the first idle CPU in the HK_TYPE_MISC housekeeping set (if there is one).
11879 static void kick_ilb(unsigned int flags)
11884 * Increase nohz.next_balance only when if full ilb is triggered but
11885 * not if we only update stats.
11887 if (flags & NOHZ_BALANCE_KICK)
11888 nohz.next_balance = jiffies+1;
11890 ilb_cpu = find_new_ilb();
11895 * Access to rq::nohz_csd is serialized by NOHZ_KICK_MASK; he who sets
11896 * the first flag owns it; cleared by nohz_csd_func().
11898 flags = atomic_fetch_or(flags, nohz_flags(ilb_cpu));
11899 if (flags & NOHZ_KICK_MASK)
11903 * This way we generate an IPI on the target CPU which
11904 * is idle, and the softirq performing NOHZ idle load balancing
11905 * will be run before returning from the IPI.
11907 smp_call_function_single_async(ilb_cpu, &cpu_rq(ilb_cpu)->nohz_csd);
11911 * Current decision point for kicking the idle load balancer in the presence
11912 * of idle CPUs in the system.
11914 static void nohz_balancer_kick(struct rq *rq)
11916 unsigned long now = jiffies;
11917 struct sched_domain_shared *sds;
11918 struct sched_domain *sd;
11919 int nr_busy, i, cpu = rq->cpu;
11920 unsigned int flags = 0;
11922 if (unlikely(rq->idle_balance))
11926 * We may be recently in ticked or tickless idle mode. At the first
11927 * busy tick after returning from idle, we will update the busy stats.
11929 nohz_balance_exit_idle(rq);
11932 * None are in tickless mode and hence no need for NOHZ idle load
11935 if (likely(!atomic_read(&nohz.nr_cpus)))
11938 if (READ_ONCE(nohz.has_blocked) &&
11939 time_after(now, READ_ONCE(nohz.next_blocked)))
11940 flags = NOHZ_STATS_KICK;
11942 if (time_before(now, nohz.next_balance))
11945 if (rq->nr_running >= 2) {
11946 flags = NOHZ_STATS_KICK | NOHZ_BALANCE_KICK;
11952 sd = rcu_dereference(rq->sd);
11955 * If there's a runnable CFS task and the current CPU has reduced
11956 * capacity, kick the ILB to see if there's a better CPU to run on:
11958 if (rq->cfs.h_nr_running >= 1 && check_cpu_capacity(rq, sd)) {
11959 flags = NOHZ_STATS_KICK | NOHZ_BALANCE_KICK;
11964 sd = rcu_dereference(per_cpu(sd_asym_packing, cpu));
11967 * When ASYM_PACKING; see if there's a more preferred CPU
11968 * currently idle; in which case, kick the ILB to move tasks
11971 * When balancing between cores, all the SMT siblings of the
11972 * preferred CPU must be idle.
11974 for_each_cpu_and(i, sched_domain_span(sd), nohz.idle_cpus_mask) {
11975 if (sched_asym(sd, i, cpu)) {
11976 flags = NOHZ_STATS_KICK | NOHZ_BALANCE_KICK;
11982 sd = rcu_dereference(per_cpu(sd_asym_cpucapacity, cpu));
11985 * When ASYM_CPUCAPACITY; see if there's a higher capacity CPU
11986 * to run the misfit task on.
11988 if (check_misfit_status(rq)) {
11989 flags = NOHZ_STATS_KICK | NOHZ_BALANCE_KICK;
11994 * For asymmetric systems, we do not want to nicely balance
11995 * cache use, instead we want to embrace asymmetry and only
11996 * ensure tasks have enough CPU capacity.
11998 * Skip the LLC logic because it's not relevant in that case.
12003 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
12006 * If there is an imbalance between LLC domains (IOW we could
12007 * increase the overall cache utilization), we need a less-loaded LLC
12008 * domain to pull some load from. Likewise, we may need to spread
12009 * load within the current LLC domain (e.g. packed SMT cores but
12010 * other CPUs are idle). We can't really know from here how busy
12011 * the others are - so just get a NOHZ balance going if it looks
12012 * like this LLC domain has tasks we could move.
12014 nr_busy = atomic_read(&sds->nr_busy_cpus);
12016 flags = NOHZ_STATS_KICK | NOHZ_BALANCE_KICK;
12023 if (READ_ONCE(nohz.needs_update))
12024 flags |= NOHZ_NEXT_KICK;
12030 static void set_cpu_sd_state_busy(int cpu)
12032 struct sched_domain *sd;
12035 sd = rcu_dereference(per_cpu(sd_llc, cpu));
12037 if (!sd || !sd->nohz_idle)
12041 atomic_inc(&sd->shared->nr_busy_cpus);
12046 void nohz_balance_exit_idle(struct rq *rq)
12048 SCHED_WARN_ON(rq != this_rq());
12050 if (likely(!rq->nohz_tick_stopped))
12053 rq->nohz_tick_stopped = 0;
12054 cpumask_clear_cpu(rq->cpu, nohz.idle_cpus_mask);
12055 atomic_dec(&nohz.nr_cpus);
12057 set_cpu_sd_state_busy(rq->cpu);
12060 static void set_cpu_sd_state_idle(int cpu)
12062 struct sched_domain *sd;
12065 sd = rcu_dereference(per_cpu(sd_llc, cpu));
12067 if (!sd || sd->nohz_idle)
12071 atomic_dec(&sd->shared->nr_busy_cpus);
12077 * This routine will record that the CPU is going idle with tick stopped.
12078 * This info will be used in performing idle load balancing in the future.
12080 void nohz_balance_enter_idle(int cpu)
12082 struct rq *rq = cpu_rq(cpu);
12084 SCHED_WARN_ON(cpu != smp_processor_id());
12086 /* If this CPU is going down, then nothing needs to be done: */
12087 if (!cpu_active(cpu))
12090 /* Spare idle load balancing on CPUs that don't want to be disturbed: */
12091 if (!housekeeping_cpu(cpu, HK_TYPE_SCHED))
12095 * Can be set safely without rq->lock held
12096 * If a clear happens, it will have evaluated last additions because
12097 * rq->lock is held during the check and the clear
12099 rq->has_blocked_load = 1;
12102 * The tick is still stopped but load could have been added in the
12103 * meantime. We set the nohz.has_blocked flag to trig a check of the
12104 * *_avg. The CPU is already part of nohz.idle_cpus_mask so the clear
12105 * of nohz.has_blocked can only happen after checking the new load
12107 if (rq->nohz_tick_stopped)
12110 /* If we're a completely isolated CPU, we don't play: */
12111 if (on_null_domain(rq))
12114 rq->nohz_tick_stopped = 1;
12116 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
12117 atomic_inc(&nohz.nr_cpus);
12120 * Ensures that if nohz_idle_balance() fails to observe our
12121 * @idle_cpus_mask store, it must observe the @has_blocked
12122 * and @needs_update stores.
12124 smp_mb__after_atomic();
12126 set_cpu_sd_state_idle(cpu);
12128 WRITE_ONCE(nohz.needs_update, 1);
12131 * Each time a cpu enter idle, we assume that it has blocked load and
12132 * enable the periodic update of the load of idle CPUs
12134 WRITE_ONCE(nohz.has_blocked, 1);
12137 static bool update_nohz_stats(struct rq *rq)
12139 unsigned int cpu = rq->cpu;
12141 if (!rq->has_blocked_load)
12144 if (!cpumask_test_cpu(cpu, nohz.idle_cpus_mask))
12147 if (!time_after(jiffies, READ_ONCE(rq->last_blocked_load_update_tick)))
12150 sched_balance_update_blocked_averages(cpu);
12152 return rq->has_blocked_load;
12156 * Internal function that runs load balance for all idle CPUs. The load balance
12157 * can be a simple update of blocked load or a complete load balance with
12158 * tasks movement depending of flags.
12160 static void _nohz_idle_balance(struct rq *this_rq, unsigned int flags)
12162 /* Earliest time when we have to do rebalance again */
12163 unsigned long now = jiffies;
12164 unsigned long next_balance = now + 60*HZ;
12165 bool has_blocked_load = false;
12166 int update_next_balance = 0;
12167 int this_cpu = this_rq->cpu;
12171 SCHED_WARN_ON((flags & NOHZ_KICK_MASK) == NOHZ_BALANCE_KICK);
12174 * We assume there will be no idle load after this update and clear
12175 * the has_blocked flag. If a cpu enters idle in the mean time, it will
12176 * set the has_blocked flag and trigger another update of idle load.
12177 * Because a cpu that becomes idle, is added to idle_cpus_mask before
12178 * setting the flag, we are sure to not clear the state and not
12179 * check the load of an idle cpu.
12181 * Same applies to idle_cpus_mask vs needs_update.
12183 if (flags & NOHZ_STATS_KICK)
12184 WRITE_ONCE(nohz.has_blocked, 0);
12185 if (flags & NOHZ_NEXT_KICK)
12186 WRITE_ONCE(nohz.needs_update, 0);
12189 * Ensures that if we miss the CPU, we must see the has_blocked
12190 * store from nohz_balance_enter_idle().
12195 * Start with the next CPU after this_cpu so we will end with this_cpu and let a
12196 * chance for other idle cpu to pull load.
12198 for_each_cpu_wrap(balance_cpu, nohz.idle_cpus_mask, this_cpu+1) {
12199 if (!idle_cpu(balance_cpu))
12203 * If this CPU gets work to do, stop the load balancing
12204 * work being done for other CPUs. Next load
12205 * balancing owner will pick it up.
12207 if (need_resched()) {
12208 if (flags & NOHZ_STATS_KICK)
12209 has_blocked_load = true;
12210 if (flags & NOHZ_NEXT_KICK)
12211 WRITE_ONCE(nohz.needs_update, 1);
12215 rq = cpu_rq(balance_cpu);
12217 if (flags & NOHZ_STATS_KICK)
12218 has_blocked_load |= update_nohz_stats(rq);
12221 * If time for next balance is due,
12224 if (time_after_eq(jiffies, rq->next_balance)) {
12225 struct rq_flags rf;
12227 rq_lock_irqsave(rq, &rf);
12228 update_rq_clock(rq);
12229 rq_unlock_irqrestore(rq, &rf);
12231 if (flags & NOHZ_BALANCE_KICK)
12232 sched_balance_domains(rq, CPU_IDLE);
12235 if (time_after(next_balance, rq->next_balance)) {
12236 next_balance = rq->next_balance;
12237 update_next_balance = 1;
12242 * next_balance will be updated only when there is a need.
12243 * When the CPU is attached to null domain for ex, it will not be
12246 if (likely(update_next_balance))
12247 nohz.next_balance = next_balance;
12249 if (flags & NOHZ_STATS_KICK)
12250 WRITE_ONCE(nohz.next_blocked,
12251 now + msecs_to_jiffies(LOAD_AVG_PERIOD));
12254 /* There is still blocked load, enable periodic update */
12255 if (has_blocked_load)
12256 WRITE_ONCE(nohz.has_blocked, 1);
12260 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
12261 * rebalancing for all the CPUs for whom scheduler ticks are stopped.
12263 static bool nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
12265 unsigned int flags = this_rq->nohz_idle_balance;
12270 this_rq->nohz_idle_balance = 0;
12272 if (idle != CPU_IDLE)
12275 _nohz_idle_balance(this_rq, flags);
12281 * Check if we need to directly run the ILB for updating blocked load before
12282 * entering idle state. Here we run ILB directly without issuing IPIs.
12284 * Note that when this function is called, the tick may not yet be stopped on
12285 * this CPU yet. nohz.idle_cpus_mask is updated only when tick is stopped and
12286 * cleared on the next busy tick. In other words, nohz.idle_cpus_mask updates
12287 * don't align with CPUs enter/exit idle to avoid bottlenecks due to high idle
12288 * entry/exit rate (usec). So it is possible that _nohz_idle_balance() is
12289 * called from this function on (this) CPU that's not yet in the mask. That's
12290 * OK because the goal of nohz_run_idle_balance() is to run ILB only for
12291 * updating the blocked load of already idle CPUs without waking up one of
12292 * those idle CPUs and outside the preempt disable / IRQ off phase of the local
12293 * cpu about to enter idle, because it can take a long time.
12295 void nohz_run_idle_balance(int cpu)
12297 unsigned int flags;
12299 flags = atomic_fetch_andnot(NOHZ_NEWILB_KICK, nohz_flags(cpu));
12302 * Update the blocked load only if no SCHED_SOFTIRQ is about to happen
12303 * (i.e. NOHZ_STATS_KICK set) and will do the same.
12305 if ((flags == NOHZ_NEWILB_KICK) && !need_resched())
12306 _nohz_idle_balance(cpu_rq(cpu), NOHZ_STATS_KICK);
12309 static void nohz_newidle_balance(struct rq *this_rq)
12311 int this_cpu = this_rq->cpu;
12314 * This CPU doesn't want to be disturbed by scheduler
12317 if (!housekeeping_cpu(this_cpu, HK_TYPE_SCHED))
12320 /* Will wake up very soon. No time for doing anything else*/
12321 if (this_rq->avg_idle < sysctl_sched_migration_cost)
12324 /* Don't need to update blocked load of idle CPUs*/
12325 if (!READ_ONCE(nohz.has_blocked) ||
12326 time_before(jiffies, READ_ONCE(nohz.next_blocked)))
12330 * Set the need to trigger ILB in order to update blocked load
12331 * before entering idle state.
12333 atomic_or(NOHZ_NEWILB_KICK, nohz_flags(this_cpu));
12336 #else /* !CONFIG_NO_HZ_COMMON */
12337 static inline void nohz_balancer_kick(struct rq *rq) { }
12339 static inline bool nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
12344 static inline void nohz_newidle_balance(struct rq *this_rq) { }
12345 #endif /* CONFIG_NO_HZ_COMMON */
12348 * sched_balance_newidle is called by schedule() if this_cpu is about to become
12349 * idle. Attempts to pull tasks from other CPUs.
12352 * < 0 - we released the lock and there are !fair tasks present
12353 * 0 - failed, no new tasks
12354 * > 0 - success, new (fair) tasks present
12356 static int sched_balance_newidle(struct rq *this_rq, struct rq_flags *rf)
12358 unsigned long next_balance = jiffies + HZ;
12359 int this_cpu = this_rq->cpu;
12360 int continue_balancing = 1;
12361 u64 t0, t1, curr_cost = 0;
12362 struct sched_domain *sd;
12363 int pulled_task = 0;
12365 update_misfit_status(NULL, this_rq);
12368 * There is a task waiting to run. No need to search for one.
12369 * Return 0; the task will be enqueued when switching to idle.
12371 if (this_rq->ttwu_pending)
12375 * We must set idle_stamp _before_ calling sched_balance_rq()
12376 * for CPU_NEWLY_IDLE, such that we measure the this duration
12379 this_rq->idle_stamp = rq_clock(this_rq);
12382 * Do not pull tasks towards !active CPUs...
12384 if (!cpu_active(this_cpu))
12388 * This is OK, because current is on_cpu, which avoids it being picked
12389 * for load-balance and preemption/IRQs are still disabled avoiding
12390 * further scheduler activity on it and we're being very careful to
12391 * re-start the picking loop.
12393 rq_unpin_lock(this_rq, rf);
12396 sd = rcu_dereference_check_sched_domain(this_rq->sd);
12398 if (!get_rd_overloaded(this_rq->rd) ||
12399 (sd && this_rq->avg_idle < sd->max_newidle_lb_cost)) {
12402 update_next_balance(sd, &next_balance);
12409 raw_spin_rq_unlock(this_rq);
12411 t0 = sched_clock_cpu(this_cpu);
12412 sched_balance_update_blocked_averages(this_cpu);
12415 for_each_domain(this_cpu, sd) {
12418 update_next_balance(sd, &next_balance);
12420 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost)
12423 if (sd->flags & SD_BALANCE_NEWIDLE) {
12425 pulled_task = sched_balance_rq(this_cpu, this_rq,
12426 sd, CPU_NEWLY_IDLE,
12427 &continue_balancing);
12429 t1 = sched_clock_cpu(this_cpu);
12430 domain_cost = t1 - t0;
12431 update_newidle_cost(sd, domain_cost);
12433 curr_cost += domain_cost;
12438 * Stop searching for tasks to pull if there are
12439 * now runnable tasks on this rq.
12441 if (pulled_task || !continue_balancing)
12446 raw_spin_rq_lock(this_rq);
12448 if (curr_cost > this_rq->max_idle_balance_cost)
12449 this_rq->max_idle_balance_cost = curr_cost;
12452 * While browsing the domains, we released the rq lock, a task could
12453 * have been enqueued in the meantime. Since we're not going idle,
12454 * pretend we pulled a task.
12456 if (this_rq->cfs.h_nr_running && !pulled_task)
12459 /* Is there a task of a high priority class? */
12460 if (this_rq->nr_running != this_rq->cfs.h_nr_running)
12464 /* Move the next balance forward */
12465 if (time_after(this_rq->next_balance, next_balance))
12466 this_rq->next_balance = next_balance;
12469 this_rq->idle_stamp = 0;
12471 nohz_newidle_balance(this_rq);
12473 rq_repin_lock(this_rq, rf);
12475 return pulled_task;
12479 * This softirq handler is triggered via SCHED_SOFTIRQ from two places:
12481 * - directly from the local scheduler_tick() for periodic load balancing
12483 * - indirectly from a remote scheduler_tick() for NOHZ idle balancing
12484 * through the SMP cross-call nohz_csd_func()
12486 static __latent_entropy void sched_balance_softirq(struct softirq_action *h)
12488 struct rq *this_rq = this_rq();
12489 enum cpu_idle_type idle = this_rq->idle_balance;
12491 * If this CPU has a pending NOHZ_BALANCE_KICK, then do the
12492 * balancing on behalf of the other idle CPUs whose ticks are
12493 * stopped. Do nohz_idle_balance *before* sched_balance_domains to
12494 * give the idle CPUs a chance to load balance. Else we may
12495 * load balance only within the local sched_domain hierarchy
12496 * and abort nohz_idle_balance altogether if we pull some load.
12498 if (nohz_idle_balance(this_rq, idle))
12501 /* normal load balance */
12502 sched_balance_update_blocked_averages(this_rq->cpu);
12503 sched_balance_domains(this_rq, idle);
12507 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
12509 void sched_balance_trigger(struct rq *rq)
12512 * Don't need to rebalance while attached to NULL domain or
12513 * runqueue CPU is not active
12515 if (unlikely(on_null_domain(rq) || !cpu_active(cpu_of(rq))))
12518 if (time_after_eq(jiffies, rq->next_balance))
12519 raise_softirq(SCHED_SOFTIRQ);
12521 nohz_balancer_kick(rq);
12524 static void rq_online_fair(struct rq *rq)
12528 update_runtime_enabled(rq);
12531 static void rq_offline_fair(struct rq *rq)
12535 /* Ensure any throttled groups are reachable by pick_next_task */
12536 unthrottle_offline_cfs_rqs(rq);
12538 /* Ensure that we remove rq contribution to group share: */
12539 clear_tg_offline_cfs_rqs(rq);
12542 #endif /* CONFIG_SMP */
12544 #ifdef CONFIG_SCHED_CORE
12546 __entity_slice_used(struct sched_entity *se, int min_nr_tasks)
12548 u64 rtime = se->sum_exec_runtime - se->prev_sum_exec_runtime;
12549 u64 slice = se->slice;
12551 return (rtime * min_nr_tasks > slice);
12554 #define MIN_NR_TASKS_DURING_FORCEIDLE 2
12555 static inline void task_tick_core(struct rq *rq, struct task_struct *curr)
12557 if (!sched_core_enabled(rq))
12561 * If runqueue has only one task which used up its slice and
12562 * if the sibling is forced idle, then trigger schedule to
12563 * give forced idle task a chance.
12565 * sched_slice() considers only this active rq and it gets the
12566 * whole slice. But during force idle, we have siblings acting
12567 * like a single runqueue and hence we need to consider runnable
12568 * tasks on this CPU and the forced idle CPU. Ideally, we should
12569 * go through the forced idle rq, but that would be a perf hit.
12570 * We can assume that the forced idle CPU has at least
12571 * MIN_NR_TASKS_DURING_FORCEIDLE - 1 tasks and use that to check
12572 * if we need to give up the CPU.
12574 if (rq->core->core_forceidle_count && rq->cfs.nr_running == 1 &&
12575 __entity_slice_used(&curr->se, MIN_NR_TASKS_DURING_FORCEIDLE))
12580 * se_fi_update - Update the cfs_rq->min_vruntime_fi in a CFS hierarchy if needed.
12582 static void se_fi_update(const struct sched_entity *se, unsigned int fi_seq,
12585 for_each_sched_entity(se) {
12586 struct cfs_rq *cfs_rq = cfs_rq_of(se);
12589 if (cfs_rq->forceidle_seq == fi_seq)
12591 cfs_rq->forceidle_seq = fi_seq;
12594 cfs_rq->min_vruntime_fi = cfs_rq->min_vruntime;
12598 void task_vruntime_update(struct rq *rq, struct task_struct *p, bool in_fi)
12600 struct sched_entity *se = &p->se;
12602 if (p->sched_class != &fair_sched_class)
12605 se_fi_update(se, rq->core->core_forceidle_seq, in_fi);
12608 bool cfs_prio_less(const struct task_struct *a, const struct task_struct *b,
12611 struct rq *rq = task_rq(a);
12612 const struct sched_entity *sea = &a->se;
12613 const struct sched_entity *seb = &b->se;
12614 struct cfs_rq *cfs_rqa;
12615 struct cfs_rq *cfs_rqb;
12618 SCHED_WARN_ON(task_rq(b)->core != rq->core);
12620 #ifdef CONFIG_FAIR_GROUP_SCHED
12622 * Find an se in the hierarchy for tasks a and b, such that the se's
12623 * are immediate siblings.
12625 while (sea->cfs_rq->tg != seb->cfs_rq->tg) {
12626 int sea_depth = sea->depth;
12627 int seb_depth = seb->depth;
12629 if (sea_depth >= seb_depth)
12630 sea = parent_entity(sea);
12631 if (sea_depth <= seb_depth)
12632 seb = parent_entity(seb);
12635 se_fi_update(sea, rq->core->core_forceidle_seq, in_fi);
12636 se_fi_update(seb, rq->core->core_forceidle_seq, in_fi);
12638 cfs_rqa = sea->cfs_rq;
12639 cfs_rqb = seb->cfs_rq;
12641 cfs_rqa = &task_rq(a)->cfs;
12642 cfs_rqb = &task_rq(b)->cfs;
12646 * Find delta after normalizing se's vruntime with its cfs_rq's
12647 * min_vruntime_fi, which would have been updated in prior calls
12648 * to se_fi_update().
12650 delta = (s64)(sea->vruntime - seb->vruntime) +
12651 (s64)(cfs_rqb->min_vruntime_fi - cfs_rqa->min_vruntime_fi);
12656 static int task_is_throttled_fair(struct task_struct *p, int cpu)
12658 struct cfs_rq *cfs_rq;
12660 #ifdef CONFIG_FAIR_GROUP_SCHED
12661 cfs_rq = task_group(p)->cfs_rq[cpu];
12663 cfs_rq = &cpu_rq(cpu)->cfs;
12665 return throttled_hierarchy(cfs_rq);
12668 static inline void task_tick_core(struct rq *rq, struct task_struct *curr) {}
12672 * scheduler tick hitting a task of our scheduling class.
12674 * NOTE: This function can be called remotely by the tick offload that
12675 * goes along full dynticks. Therefore no local assumption can be made
12676 * and everything must be accessed through the @rq and @curr passed in
12679 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
12681 struct cfs_rq *cfs_rq;
12682 struct sched_entity *se = &curr->se;
12684 for_each_sched_entity(se) {
12685 cfs_rq = cfs_rq_of(se);
12686 entity_tick(cfs_rq, se, queued);
12689 if (static_branch_unlikely(&sched_numa_balancing))
12690 task_tick_numa(rq, curr);
12692 update_misfit_status(curr, rq);
12693 check_update_overutilized_status(task_rq(curr));
12695 task_tick_core(rq, curr);
12699 * called on fork with the child task as argument from the parent's context
12700 * - child not yet on the tasklist
12701 * - preemption disabled
12703 static void task_fork_fair(struct task_struct *p)
12705 struct sched_entity *se = &p->se, *curr;
12706 struct cfs_rq *cfs_rq;
12707 struct rq *rq = this_rq();
12708 struct rq_flags rf;
12711 update_rq_clock(rq);
12713 set_task_max_allowed_capacity(p);
12715 cfs_rq = task_cfs_rq(current);
12716 curr = cfs_rq->curr;
12718 update_curr(cfs_rq);
12719 place_entity(cfs_rq, se, ENQUEUE_INITIAL);
12720 rq_unlock(rq, &rf);
12724 * Priority of the task has changed. Check to see if we preempt
12725 * the current task.
12728 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
12730 if (!task_on_rq_queued(p))
12733 if (rq->cfs.nr_running == 1)
12737 * Reschedule if we are currently running on this runqueue and
12738 * our priority decreased, or if we are not currently running on
12739 * this runqueue and our priority is higher than the current's
12741 if (task_current(rq, p)) {
12742 if (p->prio > oldprio)
12745 wakeup_preempt(rq, p, 0);
12748 #ifdef CONFIG_FAIR_GROUP_SCHED
12750 * Propagate the changes of the sched_entity across the tg tree to make it
12751 * visible to the root
12753 static void propagate_entity_cfs_rq(struct sched_entity *se)
12755 struct cfs_rq *cfs_rq = cfs_rq_of(se);
12757 if (cfs_rq_throttled(cfs_rq))
12760 if (!throttled_hierarchy(cfs_rq))
12761 list_add_leaf_cfs_rq(cfs_rq);
12763 /* Start to propagate at parent */
12766 for_each_sched_entity(se) {
12767 cfs_rq = cfs_rq_of(se);
12769 update_load_avg(cfs_rq, se, UPDATE_TG);
12771 if (cfs_rq_throttled(cfs_rq))
12774 if (!throttled_hierarchy(cfs_rq))
12775 list_add_leaf_cfs_rq(cfs_rq);
12779 static void propagate_entity_cfs_rq(struct sched_entity *se) { }
12782 static void detach_entity_cfs_rq(struct sched_entity *se)
12784 struct cfs_rq *cfs_rq = cfs_rq_of(se);
12788 * In case the task sched_avg hasn't been attached:
12789 * - A forked task which hasn't been woken up by wake_up_new_task().
12790 * - A task which has been woken up by try_to_wake_up() but is
12791 * waiting for actually being woken up by sched_ttwu_pending().
12793 if (!se->avg.last_update_time)
12797 /* Catch up with the cfs_rq and remove our load when we leave */
12798 update_load_avg(cfs_rq, se, 0);
12799 detach_entity_load_avg(cfs_rq, se);
12800 update_tg_load_avg(cfs_rq);
12801 propagate_entity_cfs_rq(se);
12804 static void attach_entity_cfs_rq(struct sched_entity *se)
12806 struct cfs_rq *cfs_rq = cfs_rq_of(se);
12808 /* Synchronize entity with its cfs_rq */
12809 update_load_avg(cfs_rq, se, sched_feat(ATTACH_AGE_LOAD) ? 0 : SKIP_AGE_LOAD);
12810 attach_entity_load_avg(cfs_rq, se);
12811 update_tg_load_avg(cfs_rq);
12812 propagate_entity_cfs_rq(se);
12815 static void detach_task_cfs_rq(struct task_struct *p)
12817 struct sched_entity *se = &p->se;
12819 detach_entity_cfs_rq(se);
12822 static void attach_task_cfs_rq(struct task_struct *p)
12824 struct sched_entity *se = &p->se;
12826 attach_entity_cfs_rq(se);
12829 static void switched_from_fair(struct rq *rq, struct task_struct *p)
12831 detach_task_cfs_rq(p);
12834 static void switched_to_fair(struct rq *rq, struct task_struct *p)
12836 attach_task_cfs_rq(p);
12838 set_task_max_allowed_capacity(p);
12840 if (task_on_rq_queued(p)) {
12842 * We were most likely switched from sched_rt, so
12843 * kick off the schedule if running, otherwise just see
12844 * if we can still preempt the current task.
12846 if (task_current(rq, p))
12849 wakeup_preempt(rq, p, 0);
12853 /* Account for a task changing its policy or group.
12855 * This routine is mostly called to set cfs_rq->curr field when a task
12856 * migrates between groups/classes.
12858 static void set_next_task_fair(struct rq *rq, struct task_struct *p, bool first)
12860 struct sched_entity *se = &p->se;
12863 if (task_on_rq_queued(p)) {
12865 * Move the next running task to the front of the list, so our
12866 * cfs_tasks list becomes MRU one.
12868 list_move(&se->group_node, &rq->cfs_tasks);
12872 for_each_sched_entity(se) {
12873 struct cfs_rq *cfs_rq = cfs_rq_of(se);
12875 set_next_entity(cfs_rq, se);
12876 /* ensure bandwidth has been allocated on our new cfs_rq */
12877 account_cfs_rq_runtime(cfs_rq, 0);
12881 void init_cfs_rq(struct cfs_rq *cfs_rq)
12883 cfs_rq->tasks_timeline = RB_ROOT_CACHED;
12884 u64_u32_store(cfs_rq->min_vruntime, (u64)(-(1LL << 20)));
12886 raw_spin_lock_init(&cfs_rq->removed.lock);
12890 #ifdef CONFIG_FAIR_GROUP_SCHED
12891 static void task_change_group_fair(struct task_struct *p)
12894 * We couldn't detach or attach a forked task which
12895 * hasn't been woken up by wake_up_new_task().
12897 if (READ_ONCE(p->__state) == TASK_NEW)
12900 detach_task_cfs_rq(p);
12903 /* Tell se's cfs_rq has been changed -- migrated */
12904 p->se.avg.last_update_time = 0;
12906 set_task_rq(p, task_cpu(p));
12907 attach_task_cfs_rq(p);
12910 void free_fair_sched_group(struct task_group *tg)
12914 for_each_possible_cpu(i) {
12916 kfree(tg->cfs_rq[i]);
12925 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
12927 struct sched_entity *se;
12928 struct cfs_rq *cfs_rq;
12931 tg->cfs_rq = kcalloc(nr_cpu_ids, sizeof(cfs_rq), GFP_KERNEL);
12934 tg->se = kcalloc(nr_cpu_ids, sizeof(se), GFP_KERNEL);
12938 tg->shares = NICE_0_LOAD;
12940 init_cfs_bandwidth(tg_cfs_bandwidth(tg), tg_cfs_bandwidth(parent));
12942 for_each_possible_cpu(i) {
12943 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
12944 GFP_KERNEL, cpu_to_node(i));
12948 se = kzalloc_node(sizeof(struct sched_entity_stats),
12949 GFP_KERNEL, cpu_to_node(i));
12953 init_cfs_rq(cfs_rq);
12954 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
12955 init_entity_runnable_average(se);
12966 void online_fair_sched_group(struct task_group *tg)
12968 struct sched_entity *se;
12969 struct rq_flags rf;
12973 for_each_possible_cpu(i) {
12976 rq_lock_irq(rq, &rf);
12977 update_rq_clock(rq);
12978 attach_entity_cfs_rq(se);
12979 sync_throttle(tg, i);
12980 rq_unlock_irq(rq, &rf);
12984 void unregister_fair_sched_group(struct task_group *tg)
12986 unsigned long flags;
12990 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
12992 for_each_possible_cpu(cpu) {
12994 remove_entity_load_avg(tg->se[cpu]);
12997 * Only empty task groups can be destroyed; so we can speculatively
12998 * check on_list without danger of it being re-added.
13000 if (!tg->cfs_rq[cpu]->on_list)
13005 raw_spin_rq_lock_irqsave(rq, flags);
13006 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
13007 raw_spin_rq_unlock_irqrestore(rq, flags);
13011 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
13012 struct sched_entity *se, int cpu,
13013 struct sched_entity *parent)
13015 struct rq *rq = cpu_rq(cpu);
13019 init_cfs_rq_runtime(cfs_rq);
13021 tg->cfs_rq[cpu] = cfs_rq;
13024 /* se could be NULL for root_task_group */
13029 se->cfs_rq = &rq->cfs;
13032 se->cfs_rq = parent->my_q;
13033 se->depth = parent->depth + 1;
13037 /* guarantee group entities always have weight */
13038 update_load_set(&se->load, NICE_0_LOAD);
13039 se->parent = parent;
13042 static DEFINE_MUTEX(shares_mutex);
13044 static int __sched_group_set_shares(struct task_group *tg, unsigned long shares)
13048 lockdep_assert_held(&shares_mutex);
13051 * We can't change the weight of the root cgroup.
13056 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
13058 if (tg->shares == shares)
13061 tg->shares = shares;
13062 for_each_possible_cpu(i) {
13063 struct rq *rq = cpu_rq(i);
13064 struct sched_entity *se = tg->se[i];
13065 struct rq_flags rf;
13067 /* Propagate contribution to hierarchy */
13068 rq_lock_irqsave(rq, &rf);
13069 update_rq_clock(rq);
13070 for_each_sched_entity(se) {
13071 update_load_avg(cfs_rq_of(se), se, UPDATE_TG);
13072 update_cfs_group(se);
13074 rq_unlock_irqrestore(rq, &rf);
13080 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
13084 mutex_lock(&shares_mutex);
13085 if (tg_is_idle(tg))
13088 ret = __sched_group_set_shares(tg, shares);
13089 mutex_unlock(&shares_mutex);
13094 int sched_group_set_idle(struct task_group *tg, long idle)
13098 if (tg == &root_task_group)
13101 if (idle < 0 || idle > 1)
13104 mutex_lock(&shares_mutex);
13106 if (tg->idle == idle) {
13107 mutex_unlock(&shares_mutex);
13113 for_each_possible_cpu(i) {
13114 struct rq *rq = cpu_rq(i);
13115 struct sched_entity *se = tg->se[i];
13116 struct cfs_rq *parent_cfs_rq, *grp_cfs_rq = tg->cfs_rq[i];
13117 bool was_idle = cfs_rq_is_idle(grp_cfs_rq);
13118 long idle_task_delta;
13119 struct rq_flags rf;
13121 rq_lock_irqsave(rq, &rf);
13123 grp_cfs_rq->idle = idle;
13124 if (WARN_ON_ONCE(was_idle == cfs_rq_is_idle(grp_cfs_rq)))
13128 parent_cfs_rq = cfs_rq_of(se);
13129 if (cfs_rq_is_idle(grp_cfs_rq))
13130 parent_cfs_rq->idle_nr_running++;
13132 parent_cfs_rq->idle_nr_running--;
13135 idle_task_delta = grp_cfs_rq->h_nr_running -
13136 grp_cfs_rq->idle_h_nr_running;
13137 if (!cfs_rq_is_idle(grp_cfs_rq))
13138 idle_task_delta *= -1;
13140 for_each_sched_entity(se) {
13141 struct cfs_rq *cfs_rq = cfs_rq_of(se);
13146 cfs_rq->idle_h_nr_running += idle_task_delta;
13148 /* Already accounted at parent level and above. */
13149 if (cfs_rq_is_idle(cfs_rq))
13154 rq_unlock_irqrestore(rq, &rf);
13157 /* Idle groups have minimum weight. */
13158 if (tg_is_idle(tg))
13159 __sched_group_set_shares(tg, scale_load(WEIGHT_IDLEPRIO));
13161 __sched_group_set_shares(tg, NICE_0_LOAD);
13163 mutex_unlock(&shares_mutex);
13167 #endif /* CONFIG_FAIR_GROUP_SCHED */
13170 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
13172 struct sched_entity *se = &task->se;
13173 unsigned int rr_interval = 0;
13176 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
13179 if (rq->cfs.load.weight)
13180 rr_interval = NS_TO_JIFFIES(se->slice);
13182 return rr_interval;
13186 * All the scheduling class methods:
13188 DEFINE_SCHED_CLASS(fair) = {
13190 .enqueue_task = enqueue_task_fair,
13191 .dequeue_task = dequeue_task_fair,
13192 .yield_task = yield_task_fair,
13193 .yield_to_task = yield_to_task_fair,
13195 .wakeup_preempt = check_preempt_wakeup_fair,
13197 .pick_next_task = __pick_next_task_fair,
13198 .put_prev_task = put_prev_task_fair,
13199 .set_next_task = set_next_task_fair,
13202 .balance = balance_fair,
13203 .pick_task = pick_task_fair,
13204 .select_task_rq = select_task_rq_fair,
13205 .migrate_task_rq = migrate_task_rq_fair,
13207 .rq_online = rq_online_fair,
13208 .rq_offline = rq_offline_fair,
13210 .task_dead = task_dead_fair,
13211 .set_cpus_allowed = set_cpus_allowed_fair,
13214 .task_tick = task_tick_fair,
13215 .task_fork = task_fork_fair,
13217 .prio_changed = prio_changed_fair,
13218 .switched_from = switched_from_fair,
13219 .switched_to = switched_to_fair,
13221 .get_rr_interval = get_rr_interval_fair,
13223 .update_curr = update_curr_fair,
13225 #ifdef CONFIG_FAIR_GROUP_SCHED
13226 .task_change_group = task_change_group_fair,
13229 #ifdef CONFIG_SCHED_CORE
13230 .task_is_throttled = task_is_throttled_fair,
13233 #ifdef CONFIG_UCLAMP_TASK
13234 .uclamp_enabled = 1,
13238 #ifdef CONFIG_SCHED_DEBUG
13239 void print_cfs_stats(struct seq_file *m, int cpu)
13241 struct cfs_rq *cfs_rq, *pos;
13244 for_each_leaf_cfs_rq_safe(cpu_rq(cpu), cfs_rq, pos)
13245 print_cfs_rq(m, cpu, cfs_rq);
13249 #ifdef CONFIG_NUMA_BALANCING
13250 void show_numa_stats(struct task_struct *p, struct seq_file *m)
13253 unsigned long tsf = 0, tpf = 0, gsf = 0, gpf = 0;
13254 struct numa_group *ng;
13257 ng = rcu_dereference(p->numa_group);
13258 for_each_online_node(node) {
13259 if (p->numa_faults) {
13260 tsf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 0)];
13261 tpf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 1)];
13264 gsf = ng->faults[task_faults_idx(NUMA_MEM, node, 0)],
13265 gpf = ng->faults[task_faults_idx(NUMA_MEM, node, 1)];
13267 print_numa_stats(m, node, tsf, tpf, gsf, gpf);
13271 #endif /* CONFIG_NUMA_BALANCING */
13272 #endif /* CONFIG_SCHED_DEBUG */
13274 __init void init_sched_fair_class(void)
13279 for_each_possible_cpu(i) {
13280 zalloc_cpumask_var_node(&per_cpu(load_balance_mask, i), GFP_KERNEL, cpu_to_node(i));
13281 zalloc_cpumask_var_node(&per_cpu(select_rq_mask, i), GFP_KERNEL, cpu_to_node(i));
13282 zalloc_cpumask_var_node(&per_cpu(should_we_balance_tmpmask, i),
13283 GFP_KERNEL, cpu_to_node(i));
13285 #ifdef CONFIG_CFS_BANDWIDTH
13286 INIT_CSD(&cpu_rq(i)->cfsb_csd, __cfsb_csd_unthrottle, cpu_rq(i));
13287 INIT_LIST_HEAD(&cpu_rq(i)->cfsb_csd_list);
13291 open_softirq(SCHED_SOFTIRQ, sched_balance_softirq);
13293 #ifdef CONFIG_NO_HZ_COMMON
13294 nohz.next_balance = jiffies;
13295 nohz.next_blocked = jiffies;
13296 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);