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 int sched_thermal_decay_shift;
82 static int __init setup_sched_thermal_decay_shift(char *str)
86 if (kstrtoint(str, 0, &_shift))
87 pr_warn("Unable to set scheduler thermal pressure decay shift parameter\n");
89 sched_thermal_decay_shift = clamp(_shift, 0, 10);
92 __setup("sched_thermal_decay_shift=", setup_sched_thermal_decay_shift);
96 * For asym packing, by default the lower numbered CPU has higher priority.
98 int __weak arch_asym_cpu_priority(int cpu)
104 * The margin used when comparing utilization with CPU capacity.
108 #define fits_capacity(cap, max) ((cap) * 1280 < (max) * 1024)
111 * The margin used when comparing CPU capacities.
112 * is 'cap1' noticeably greater than 'cap2'
116 #define capacity_greater(cap1, cap2) ((cap1) * 1024 > (cap2) * 1078)
119 #ifdef CONFIG_CFS_BANDWIDTH
121 * Amount of runtime to allocate from global (tg) to local (per-cfs_rq) pool
122 * each time a cfs_rq requests quota.
124 * Note: in the case that the slice exceeds the runtime remaining (either due
125 * to consumption or the quota being specified to be smaller than the slice)
126 * we will always only issue the remaining available time.
128 * (default: 5 msec, units: microseconds)
130 static unsigned int sysctl_sched_cfs_bandwidth_slice = 5000UL;
133 #ifdef CONFIG_NUMA_BALANCING
134 /* Restrict the NUMA promotion throughput (MB/s) for each target node. */
135 static unsigned int sysctl_numa_balancing_promote_rate_limit = 65536;
139 static struct ctl_table sched_fair_sysctls[] = {
140 #ifdef CONFIG_CFS_BANDWIDTH
142 .procname = "sched_cfs_bandwidth_slice_us",
143 .data = &sysctl_sched_cfs_bandwidth_slice,
144 .maxlen = sizeof(unsigned int),
146 .proc_handler = proc_dointvec_minmax,
147 .extra1 = SYSCTL_ONE,
150 #ifdef CONFIG_NUMA_BALANCING
152 .procname = "numa_balancing_promote_rate_limit_MBps",
153 .data = &sysctl_numa_balancing_promote_rate_limit,
154 .maxlen = sizeof(unsigned int),
156 .proc_handler = proc_dointvec_minmax,
157 .extra1 = SYSCTL_ZERO,
159 #endif /* CONFIG_NUMA_BALANCING */
163 static int __init sched_fair_sysctl_init(void)
165 register_sysctl_init("kernel", sched_fair_sysctls);
168 late_initcall(sched_fair_sysctl_init);
171 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
177 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
183 static inline void update_load_set(struct load_weight *lw, unsigned long w)
190 * Increase the granularity value when there are more CPUs,
191 * because with more CPUs the 'effective latency' as visible
192 * to users decreases. But the relationship is not linear,
193 * so pick a second-best guess by going with the log2 of the
196 * This idea comes from the SD scheduler of Con Kolivas:
198 static unsigned int get_update_sysctl_factor(void)
200 unsigned int cpus = min_t(unsigned int, num_online_cpus(), 8);
203 switch (sysctl_sched_tunable_scaling) {
204 case SCHED_TUNABLESCALING_NONE:
207 case SCHED_TUNABLESCALING_LINEAR:
210 case SCHED_TUNABLESCALING_LOG:
212 factor = 1 + ilog2(cpus);
219 static void update_sysctl(void)
221 unsigned int factor = get_update_sysctl_factor();
223 #define SET_SYSCTL(name) \
224 (sysctl_##name = (factor) * normalized_sysctl_##name)
225 SET_SYSCTL(sched_base_slice);
229 void __init sched_init_granularity(void)
234 #define WMULT_CONST (~0U)
235 #define WMULT_SHIFT 32
237 static void __update_inv_weight(struct load_weight *lw)
241 if (likely(lw->inv_weight))
244 w = scale_load_down(lw->weight);
246 if (BITS_PER_LONG > 32 && unlikely(w >= WMULT_CONST))
248 else if (unlikely(!w))
249 lw->inv_weight = WMULT_CONST;
251 lw->inv_weight = WMULT_CONST / w;
255 * delta_exec * weight / lw.weight
257 * (delta_exec * (weight * lw->inv_weight)) >> WMULT_SHIFT
259 * Either weight := NICE_0_LOAD and lw \e sched_prio_to_wmult[], in which case
260 * we're guaranteed shift stays positive because inv_weight is guaranteed to
261 * fit 32 bits, and NICE_0_LOAD gives another 10 bits; therefore shift >= 22.
263 * Or, weight =< lw.weight (because lw.weight is the runqueue weight), thus
264 * weight/lw.weight <= 1, and therefore our shift will also be positive.
266 static u64 __calc_delta(u64 delta_exec, unsigned long weight, struct load_weight *lw)
268 u64 fact = scale_load_down(weight);
269 u32 fact_hi = (u32)(fact >> 32);
270 int shift = WMULT_SHIFT;
273 __update_inv_weight(lw);
275 if (unlikely(fact_hi)) {
281 fact = mul_u32_u32(fact, lw->inv_weight);
283 fact_hi = (u32)(fact >> 32);
290 return mul_u64_u32_shr(delta_exec, fact, shift);
296 static inline u64 calc_delta_fair(u64 delta, struct sched_entity *se)
298 if (unlikely(se->load.weight != NICE_0_LOAD))
299 delta = __calc_delta(delta, NICE_0_LOAD, &se->load);
304 const struct sched_class fair_sched_class;
306 /**************************************************************
307 * CFS operations on generic schedulable entities:
310 #ifdef CONFIG_FAIR_GROUP_SCHED
312 /* Walk up scheduling entities hierarchy */
313 #define for_each_sched_entity(se) \
314 for (; se; se = se->parent)
316 static inline bool list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
318 struct rq *rq = rq_of(cfs_rq);
319 int cpu = cpu_of(rq);
322 return rq->tmp_alone_branch == &rq->leaf_cfs_rq_list;
327 * Ensure we either appear before our parent (if already
328 * enqueued) or force our parent to appear after us when it is
329 * enqueued. The fact that we always enqueue bottom-up
330 * reduces this to two cases and a special case for the root
331 * cfs_rq. Furthermore, it also means that we will always reset
332 * tmp_alone_branch either when the branch is connected
333 * to a tree or when we reach the top of the tree
335 if (cfs_rq->tg->parent &&
336 cfs_rq->tg->parent->cfs_rq[cpu]->on_list) {
338 * If parent is already on the list, we add the child
339 * just before. Thanks to circular linked property of
340 * the list, this means to put the child at the tail
341 * of the list that starts by parent.
343 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
344 &(cfs_rq->tg->parent->cfs_rq[cpu]->leaf_cfs_rq_list));
346 * The branch is now connected to its tree so we can
347 * reset tmp_alone_branch to the beginning of the
350 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
354 if (!cfs_rq->tg->parent) {
356 * cfs rq without parent should be put
357 * at the tail of the list.
359 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
360 &rq->leaf_cfs_rq_list);
362 * We have reach the top of a tree so we can reset
363 * tmp_alone_branch to the beginning of the list.
365 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
370 * The parent has not already been added so we want to
371 * make sure that it will be put after us.
372 * tmp_alone_branch points to the begin of the branch
373 * where we will add parent.
375 list_add_rcu(&cfs_rq->leaf_cfs_rq_list, rq->tmp_alone_branch);
377 * update tmp_alone_branch to points to the new begin
380 rq->tmp_alone_branch = &cfs_rq->leaf_cfs_rq_list;
384 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
386 if (cfs_rq->on_list) {
387 struct rq *rq = rq_of(cfs_rq);
390 * With cfs_rq being unthrottled/throttled during an enqueue,
391 * it can happen the tmp_alone_branch points the a leaf that
392 * we finally want to del. In this case, tmp_alone_branch moves
393 * to the prev element but it will point to rq->leaf_cfs_rq_list
394 * at the end of the enqueue.
396 if (rq->tmp_alone_branch == &cfs_rq->leaf_cfs_rq_list)
397 rq->tmp_alone_branch = cfs_rq->leaf_cfs_rq_list.prev;
399 list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
404 static inline void assert_list_leaf_cfs_rq(struct rq *rq)
406 SCHED_WARN_ON(rq->tmp_alone_branch != &rq->leaf_cfs_rq_list);
409 /* Iterate thr' all leaf cfs_rq's on a runqueue */
410 #define for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) \
411 list_for_each_entry_safe(cfs_rq, pos, &rq->leaf_cfs_rq_list, \
414 /* Do the two (enqueued) entities belong to the same group ? */
415 static inline struct cfs_rq *
416 is_same_group(struct sched_entity *se, struct sched_entity *pse)
418 if (se->cfs_rq == pse->cfs_rq)
424 static inline struct sched_entity *parent_entity(const struct sched_entity *se)
430 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
432 int se_depth, pse_depth;
435 * preemption test can be made between sibling entities who are in the
436 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
437 * both tasks until we find their ancestors who are siblings of common
441 /* First walk up until both entities are at same depth */
442 se_depth = (*se)->depth;
443 pse_depth = (*pse)->depth;
445 while (se_depth > pse_depth) {
447 *se = parent_entity(*se);
450 while (pse_depth > se_depth) {
452 *pse = parent_entity(*pse);
455 while (!is_same_group(*se, *pse)) {
456 *se = parent_entity(*se);
457 *pse = parent_entity(*pse);
461 static int tg_is_idle(struct task_group *tg)
466 static int cfs_rq_is_idle(struct cfs_rq *cfs_rq)
468 return cfs_rq->idle > 0;
471 static int se_is_idle(struct sched_entity *se)
473 if (entity_is_task(se))
474 return task_has_idle_policy(task_of(se));
475 return cfs_rq_is_idle(group_cfs_rq(se));
478 #else /* !CONFIG_FAIR_GROUP_SCHED */
480 #define for_each_sched_entity(se) \
481 for (; se; se = NULL)
483 static inline bool list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
488 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
492 static inline void assert_list_leaf_cfs_rq(struct rq *rq)
496 #define for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) \
497 for (cfs_rq = &rq->cfs, pos = NULL; cfs_rq; cfs_rq = pos)
499 static inline struct sched_entity *parent_entity(struct sched_entity *se)
505 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
509 static inline int tg_is_idle(struct task_group *tg)
514 static int cfs_rq_is_idle(struct cfs_rq *cfs_rq)
519 static int se_is_idle(struct sched_entity *se)
524 #endif /* CONFIG_FAIR_GROUP_SCHED */
526 static __always_inline
527 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec);
529 /**************************************************************
530 * Scheduling class tree data structure manipulation methods:
533 static inline u64 max_vruntime(u64 max_vruntime, u64 vruntime)
535 s64 delta = (s64)(vruntime - max_vruntime);
537 max_vruntime = vruntime;
542 static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime)
544 s64 delta = (s64)(vruntime - min_vruntime);
546 min_vruntime = vruntime;
551 static inline bool entity_before(const struct sched_entity *a,
552 const struct sched_entity *b)
555 * Tiebreak on vruntime seems unnecessary since it can
558 return (s64)(a->deadline - b->deadline) < 0;
561 static inline s64 entity_key(struct cfs_rq *cfs_rq, struct sched_entity *se)
563 return (s64)(se->vruntime - cfs_rq->min_vruntime);
566 #define __node_2_se(node) \
567 rb_entry((node), struct sched_entity, run_node)
570 * Compute virtual time from the per-task service numbers:
572 * Fair schedulers conserve lag:
576 * Where lag_i is given by:
578 * lag_i = S - s_i = w_i * (V - v_i)
580 * Where S is the ideal service time and V is it's virtual time counterpart.
584 * \Sum w_i * (V - v_i) = 0
585 * \Sum w_i * V - w_i * v_i = 0
587 * From which we can solve an expression for V in v_i (which we have in
590 * \Sum v_i * w_i \Sum v_i * w_i
591 * V = -------------- = --------------
594 * Specifically, this is the weighted average of all entity virtual runtimes.
596 * [[ NOTE: this is only equal to the ideal scheduler under the condition
597 * that join/leave operations happen at lag_i = 0, otherwise the
598 * virtual time has non-continguous motion equivalent to:
602 * Also see the comment in place_entity() that deals with this. ]]
604 * However, since v_i is u64, and the multiplcation could easily overflow
605 * transform it into a relative form that uses smaller quantities:
607 * Substitute: v_i == (v_i - v0) + v0
609 * \Sum ((v_i - v0) + v0) * w_i \Sum (v_i - v0) * w_i
610 * V = ---------------------------- = --------------------- + v0
613 * Which we track using:
615 * v0 := cfs_rq->min_vruntime
616 * \Sum (v_i - v0) * w_i := cfs_rq->avg_vruntime
617 * \Sum w_i := cfs_rq->avg_load
619 * Since min_vruntime is a monotonic increasing variable that closely tracks
620 * the per-task service, these deltas: (v_i - v), will be in the order of the
621 * maximal (virtual) lag induced in the system due to quantisation.
623 * Also, we use scale_load_down() to reduce the size.
625 * As measured, the max (key * weight) value was ~44 bits for a kernel build.
628 avg_vruntime_add(struct cfs_rq *cfs_rq, struct sched_entity *se)
630 unsigned long weight = scale_load_down(se->load.weight);
631 s64 key = entity_key(cfs_rq, se);
633 cfs_rq->avg_vruntime += key * weight;
634 cfs_rq->avg_load += weight;
638 avg_vruntime_sub(struct cfs_rq *cfs_rq, struct sched_entity *se)
640 unsigned long weight = scale_load_down(se->load.weight);
641 s64 key = entity_key(cfs_rq, se);
643 cfs_rq->avg_vruntime -= key * weight;
644 cfs_rq->avg_load -= weight;
648 void avg_vruntime_update(struct cfs_rq *cfs_rq, s64 delta)
651 * v' = v + d ==> avg_vruntime' = avg_runtime - d*avg_load
653 cfs_rq->avg_vruntime -= cfs_rq->avg_load * delta;
657 * Specifically: avg_runtime() + 0 must result in entity_eligible() := true
658 * For this to be so, the result of this function must have a left bias.
660 u64 avg_vruntime(struct cfs_rq *cfs_rq)
662 struct sched_entity *curr = cfs_rq->curr;
663 s64 avg = cfs_rq->avg_vruntime;
664 long load = cfs_rq->avg_load;
666 if (curr && curr->on_rq) {
667 unsigned long weight = scale_load_down(curr->load.weight);
669 avg += entity_key(cfs_rq, curr) * weight;
674 /* sign flips effective floor / ceil */
677 avg = div_s64(avg, load);
680 return cfs_rq->min_vruntime + avg;
684 * lag_i = S - s_i = w_i * (V - v_i)
686 * However, since V is approximated by the weighted average of all entities it
687 * is possible -- by addition/removal/reweight to the tree -- to move V around
688 * and end up with a larger lag than we started with.
690 * Limit this to either double the slice length with a minimum of TICK_NSEC
691 * since that is the timing granularity.
693 * EEVDF gives the following limit for a steady state system:
695 * -r_max < lag < max(r_max, q)
697 * XXX could add max_slice to the augmented data to track this.
699 static void update_entity_lag(struct cfs_rq *cfs_rq, struct sched_entity *se)
703 SCHED_WARN_ON(!se->on_rq);
704 lag = avg_vruntime(cfs_rq) - se->vruntime;
706 limit = calc_delta_fair(max_t(u64, 2*se->slice, TICK_NSEC), se);
707 se->vlag = clamp(lag, -limit, limit);
711 * Entity is eligible once it received less service than it ought to have,
714 * lag_i = S - s_i = w_i*(V - v_i)
716 * lag_i >= 0 -> V >= v_i
719 * V = ------------------ + v
722 * lag_i >= 0 -> \Sum (v_i - v)*w_i >= (v_i - v)*(\Sum w_i)
724 * Note: using 'avg_vruntime() > se->vruntime' is inacurate due
725 * to the loss in precision caused by the division.
727 static int vruntime_eligible(struct cfs_rq *cfs_rq, u64 vruntime)
729 struct sched_entity *curr = cfs_rq->curr;
730 s64 avg = cfs_rq->avg_vruntime;
731 long load = cfs_rq->avg_load;
733 if (curr && curr->on_rq) {
734 unsigned long weight = scale_load_down(curr->load.weight);
736 avg += entity_key(cfs_rq, curr) * weight;
740 return avg >= (s64)(vruntime - cfs_rq->min_vruntime) * load;
743 int entity_eligible(struct cfs_rq *cfs_rq, struct sched_entity *se)
745 return vruntime_eligible(cfs_rq, se->vruntime);
748 static u64 __update_min_vruntime(struct cfs_rq *cfs_rq, u64 vruntime)
750 u64 min_vruntime = cfs_rq->min_vruntime;
752 * open coded max_vruntime() to allow updating avg_vruntime
754 s64 delta = (s64)(vruntime - min_vruntime);
756 avg_vruntime_update(cfs_rq, delta);
757 min_vruntime = vruntime;
762 static void update_min_vruntime(struct cfs_rq *cfs_rq)
764 struct sched_entity *se = __pick_root_entity(cfs_rq);
765 struct sched_entity *curr = cfs_rq->curr;
766 u64 vruntime = cfs_rq->min_vruntime;
770 vruntime = curr->vruntime;
777 vruntime = se->min_vruntime;
779 vruntime = min_vruntime(vruntime, se->min_vruntime);
782 /* ensure we never gain time by being placed backwards. */
783 u64_u32_store(cfs_rq->min_vruntime,
784 __update_min_vruntime(cfs_rq, vruntime));
787 static inline bool __entity_less(struct rb_node *a, const struct rb_node *b)
789 return entity_before(__node_2_se(a), __node_2_se(b));
792 #define vruntime_gt(field, lse, rse) ({ (s64)((lse)->field - (rse)->field) > 0; })
794 static inline void __min_vruntime_update(struct sched_entity *se, struct rb_node *node)
797 struct sched_entity *rse = __node_2_se(node);
798 if (vruntime_gt(min_vruntime, se, rse))
799 se->min_vruntime = rse->min_vruntime;
804 * se->min_vruntime = min(se->vruntime, {left,right}->min_vruntime)
806 static inline bool min_vruntime_update(struct sched_entity *se, bool exit)
808 u64 old_min_vruntime = se->min_vruntime;
809 struct rb_node *node = &se->run_node;
811 se->min_vruntime = se->vruntime;
812 __min_vruntime_update(se, node->rb_right);
813 __min_vruntime_update(se, node->rb_left);
815 return se->min_vruntime == old_min_vruntime;
818 RB_DECLARE_CALLBACKS(static, min_vruntime_cb, struct sched_entity,
819 run_node, min_vruntime, min_vruntime_update);
822 * Enqueue an entity into the rb-tree:
824 static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
826 avg_vruntime_add(cfs_rq, se);
827 se->min_vruntime = se->vruntime;
828 rb_add_augmented_cached(&se->run_node, &cfs_rq->tasks_timeline,
829 __entity_less, &min_vruntime_cb);
832 static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
834 rb_erase_augmented_cached(&se->run_node, &cfs_rq->tasks_timeline,
836 avg_vruntime_sub(cfs_rq, se);
839 struct sched_entity *__pick_root_entity(struct cfs_rq *cfs_rq)
841 struct rb_node *root = cfs_rq->tasks_timeline.rb_root.rb_node;
846 return __node_2_se(root);
849 struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq)
851 struct rb_node *left = rb_first_cached(&cfs_rq->tasks_timeline);
856 return __node_2_se(left);
860 * Earliest Eligible Virtual Deadline First
862 * In order to provide latency guarantees for different request sizes
863 * EEVDF selects the best runnable task from two criteria:
865 * 1) the task must be eligible (must be owed service)
867 * 2) from those tasks that meet 1), we select the one
868 * with the earliest virtual deadline.
870 * We can do this in O(log n) time due to an augmented RB-tree. The
871 * tree keeps the entries sorted on deadline, but also functions as a
872 * heap based on the vruntime by keeping:
874 * se->min_vruntime = min(se->vruntime, se->{left,right}->min_vruntime)
876 * Which allows tree pruning through eligibility.
878 static struct sched_entity *pick_eevdf(struct cfs_rq *cfs_rq)
880 struct rb_node *node = cfs_rq->tasks_timeline.rb_root.rb_node;
881 struct sched_entity *se = __pick_first_entity(cfs_rq);
882 struct sched_entity *curr = cfs_rq->curr;
883 struct sched_entity *best = NULL;
886 * We can safely skip eligibility check if there is only one entity
887 * in this cfs_rq, saving some cycles.
889 if (cfs_rq->nr_running == 1)
890 return curr && curr->on_rq ? curr : se;
892 if (curr && (!curr->on_rq || !entity_eligible(cfs_rq, curr)))
896 * Once selected, run a task until it either becomes non-eligible or
897 * until it gets a new slice. See the HACK in set_next_entity().
899 if (sched_feat(RUN_TO_PARITY) && curr && curr->vlag == curr->deadline)
902 /* Pick the leftmost entity if it's eligible */
903 if (se && entity_eligible(cfs_rq, se)) {
908 /* Heap search for the EEVD entity */
910 struct rb_node *left = node->rb_left;
913 * Eligible entities in left subtree are always better
914 * choices, since they have earlier deadlines.
916 if (left && vruntime_eligible(cfs_rq,
917 __node_2_se(left)->min_vruntime)) {
922 se = __node_2_se(node);
925 * The left subtree either is empty or has no eligible
926 * entity, so check the current node since it is the one
927 * with earliest deadline that might be eligible.
929 if (entity_eligible(cfs_rq, se)) {
934 node = node->rb_right;
937 if (!best || (curr && entity_before(curr, best)))
943 #ifdef CONFIG_SCHED_DEBUG
944 struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
946 struct rb_node *last = rb_last(&cfs_rq->tasks_timeline.rb_root);
951 return __node_2_se(last);
954 /**************************************************************
955 * Scheduling class statistics methods:
958 int sched_update_scaling(void)
960 unsigned int factor = get_update_sysctl_factor();
962 #define WRT_SYSCTL(name) \
963 (normalized_sysctl_##name = sysctl_##name / (factor))
964 WRT_SYSCTL(sched_base_slice);
972 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se);
975 * XXX: strictly: vd_i += N*r_i/w_i such that: vd_i > ve_i
976 * this is probably good enough.
978 static void update_deadline(struct cfs_rq *cfs_rq, struct sched_entity *se)
980 if ((s64)(se->vruntime - se->deadline) < 0)
984 * For EEVDF the virtual time slope is determined by w_i (iow.
985 * nice) while the request time r_i is determined by
986 * sysctl_sched_base_slice.
988 se->slice = sysctl_sched_base_slice;
991 * EEVDF: vd_i = ve_i + r_i / w_i
993 se->deadline = se->vruntime + calc_delta_fair(se->slice, se);
996 * The task has consumed its request, reschedule.
998 if (cfs_rq->nr_running > 1) {
999 resched_curr(rq_of(cfs_rq));
1000 clear_buddies(cfs_rq, se);
1007 static int select_idle_sibling(struct task_struct *p, int prev_cpu, int cpu);
1008 static unsigned long task_h_load(struct task_struct *p);
1009 static unsigned long capacity_of(int cpu);
1011 /* Give new sched_entity start runnable values to heavy its load in infant time */
1012 void init_entity_runnable_average(struct sched_entity *se)
1014 struct sched_avg *sa = &se->avg;
1016 memset(sa, 0, sizeof(*sa));
1019 * Tasks are initialized with full load to be seen as heavy tasks until
1020 * they get a chance to stabilize to their real load level.
1021 * Group entities are initialized with zero load to reflect the fact that
1022 * nothing has been attached to the task group yet.
1024 if (entity_is_task(se))
1025 sa->load_avg = scale_load_down(se->load.weight);
1027 /* when this task enqueue'ed, it will contribute to its cfs_rq's load_avg */
1031 * With new tasks being created, their initial util_avgs are extrapolated
1032 * based on the cfs_rq's current util_avg:
1034 * util_avg = cfs_rq->util_avg / (cfs_rq->load_avg + 1) * se.load.weight
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->load.weight;
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 * The pages in slow memory node should be migrated according
1835 * to hot/cold instead of private/shared.
1837 if (sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING &&
1838 !node_is_toptier(src_nid)) {
1839 struct pglist_data *pgdat;
1840 unsigned long rate_limit;
1841 unsigned int latency, th, def_th;
1843 pgdat = NODE_DATA(dst_nid);
1844 if (pgdat_free_space_enough(pgdat)) {
1845 /* workload changed, reset hot threshold */
1846 pgdat->nbp_threshold = 0;
1850 def_th = sysctl_numa_balancing_hot_threshold;
1851 rate_limit = sysctl_numa_balancing_promote_rate_limit << \
1853 numa_promotion_adjust_threshold(pgdat, rate_limit, def_th);
1855 th = pgdat->nbp_threshold ? : def_th;
1856 latency = numa_hint_fault_latency(folio);
1860 return !numa_promotion_rate_limit(pgdat, rate_limit,
1861 folio_nr_pages(folio));
1864 this_cpupid = cpu_pid_to_cpupid(dst_cpu, current->pid);
1865 last_cpupid = folio_xchg_last_cpupid(folio, this_cpupid);
1867 if (!(sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING) &&
1868 !node_is_toptier(src_nid) && !cpupid_valid(last_cpupid))
1872 * Allow first faults or private faults to migrate immediately early in
1873 * the lifetime of a task. The magic number 4 is based on waiting for
1874 * two full passes of the "multi-stage node selection" test that is
1877 if ((p->numa_preferred_nid == NUMA_NO_NODE || p->numa_scan_seq <= 4) &&
1878 (cpupid_pid_unset(last_cpupid) || cpupid_match_pid(p, last_cpupid)))
1882 * Multi-stage node selection is used in conjunction with a periodic
1883 * migration fault to build a temporal task<->page relation. By using
1884 * a two-stage filter we remove short/unlikely relations.
1886 * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate
1887 * a task's usage of a particular page (n_p) per total usage of this
1888 * page (n_t) (in a given time-span) to a probability.
1890 * Our periodic faults will sample this probability and getting the
1891 * same result twice in a row, given these samples are fully
1892 * independent, is then given by P(n)^2, provided our sample period
1893 * is sufficiently short compared to the usage pattern.
1895 * This quadric squishes small probabilities, making it less likely we
1896 * act on an unlikely task<->page relation.
1898 if (!cpupid_pid_unset(last_cpupid) &&
1899 cpupid_to_nid(last_cpupid) != dst_nid)
1902 /* Always allow migrate on private faults */
1903 if (cpupid_match_pid(p, last_cpupid))
1906 /* A shared fault, but p->numa_group has not been set up yet. */
1911 * Destination node is much more heavily used than the source
1912 * node? Allow migration.
1914 if (group_faults_cpu(ng, dst_nid) > group_faults_cpu(ng, src_nid) *
1915 ACTIVE_NODE_FRACTION)
1919 * Distribute memory according to CPU & memory use on each node,
1920 * with 3/4 hysteresis to avoid unnecessary memory migrations:
1922 * faults_cpu(dst) 3 faults_cpu(src)
1923 * --------------- * - > ---------------
1924 * faults_mem(dst) 4 faults_mem(src)
1926 return group_faults_cpu(ng, dst_nid) * group_faults(p, src_nid) * 3 >
1927 group_faults_cpu(ng, src_nid) * group_faults(p, dst_nid) * 4;
1931 * 'numa_type' describes the node at the moment of load balancing.
1934 /* The node has spare capacity that can be used to run more tasks. */
1937 * The node is fully used and the tasks don't compete for more CPU
1938 * cycles. Nevertheless, some tasks might wait before running.
1942 * The node is overloaded and can't provide expected CPU cycles to all
1948 /* Cached statistics for all CPUs within a node */
1951 unsigned long runnable;
1953 /* Total compute capacity of CPUs on a node */
1954 unsigned long compute_capacity;
1955 unsigned int nr_running;
1956 unsigned int weight;
1957 enum numa_type node_type;
1961 struct task_numa_env {
1962 struct task_struct *p;
1964 int src_cpu, src_nid;
1965 int dst_cpu, dst_nid;
1968 struct numa_stats src_stats, dst_stats;
1973 struct task_struct *best_task;
1978 static unsigned long cpu_load(struct rq *rq);
1979 static unsigned long cpu_runnable(struct rq *rq);
1982 numa_type numa_classify(unsigned int imbalance_pct,
1983 struct numa_stats *ns)
1985 if ((ns->nr_running > ns->weight) &&
1986 (((ns->compute_capacity * 100) < (ns->util * imbalance_pct)) ||
1987 ((ns->compute_capacity * imbalance_pct) < (ns->runnable * 100))))
1988 return node_overloaded;
1990 if ((ns->nr_running < ns->weight) ||
1991 (((ns->compute_capacity * 100) > (ns->util * imbalance_pct)) &&
1992 ((ns->compute_capacity * imbalance_pct) > (ns->runnable * 100))))
1993 return node_has_spare;
1995 return node_fully_busy;
1998 #ifdef CONFIG_SCHED_SMT
1999 /* Forward declarations of select_idle_sibling helpers */
2000 static inline bool test_idle_cores(int cpu);
2001 static inline int numa_idle_core(int idle_core, int cpu)
2003 if (!static_branch_likely(&sched_smt_present) ||
2004 idle_core >= 0 || !test_idle_cores(cpu))
2008 * Prefer cores instead of packing HT siblings
2009 * and triggering future load balancing.
2011 if (is_core_idle(cpu))
2017 static inline int numa_idle_core(int idle_core, int cpu)
2024 * Gather all necessary information to make NUMA balancing placement
2025 * decisions that are compatible with standard load balancer. This
2026 * borrows code and logic from update_sg_lb_stats but sharing a
2027 * common implementation is impractical.
2029 static void update_numa_stats(struct task_numa_env *env,
2030 struct numa_stats *ns, int nid,
2033 int cpu, idle_core = -1;
2035 memset(ns, 0, sizeof(*ns));
2039 for_each_cpu(cpu, cpumask_of_node(nid)) {
2040 struct rq *rq = cpu_rq(cpu);
2042 ns->load += cpu_load(rq);
2043 ns->runnable += cpu_runnable(rq);
2044 ns->util += cpu_util_cfs(cpu);
2045 ns->nr_running += rq->cfs.h_nr_running;
2046 ns->compute_capacity += capacity_of(cpu);
2048 if (find_idle && idle_core < 0 && !rq->nr_running && idle_cpu(cpu)) {
2049 if (READ_ONCE(rq->numa_migrate_on) ||
2050 !cpumask_test_cpu(cpu, env->p->cpus_ptr))
2053 if (ns->idle_cpu == -1)
2056 idle_core = numa_idle_core(idle_core, cpu);
2061 ns->weight = cpumask_weight(cpumask_of_node(nid));
2063 ns->node_type = numa_classify(env->imbalance_pct, ns);
2066 ns->idle_cpu = idle_core;
2069 static void task_numa_assign(struct task_numa_env *env,
2070 struct task_struct *p, long imp)
2072 struct rq *rq = cpu_rq(env->dst_cpu);
2074 /* Check if run-queue part of active NUMA balance. */
2075 if (env->best_cpu != env->dst_cpu && xchg(&rq->numa_migrate_on, 1)) {
2077 int start = env->dst_cpu;
2079 /* Find alternative idle CPU. */
2080 for_each_cpu_wrap(cpu, cpumask_of_node(env->dst_nid), start + 1) {
2081 if (cpu == env->best_cpu || !idle_cpu(cpu) ||
2082 !cpumask_test_cpu(cpu, env->p->cpus_ptr)) {
2087 rq = cpu_rq(env->dst_cpu);
2088 if (!xchg(&rq->numa_migrate_on, 1))
2092 /* Failed to find an alternative idle CPU */
2098 * Clear previous best_cpu/rq numa-migrate flag, since task now
2099 * found a better CPU to move/swap.
2101 if (env->best_cpu != -1 && env->best_cpu != env->dst_cpu) {
2102 rq = cpu_rq(env->best_cpu);
2103 WRITE_ONCE(rq->numa_migrate_on, 0);
2107 put_task_struct(env->best_task);
2112 env->best_imp = imp;
2113 env->best_cpu = env->dst_cpu;
2116 static bool load_too_imbalanced(long src_load, long dst_load,
2117 struct task_numa_env *env)
2120 long orig_src_load, orig_dst_load;
2121 long src_capacity, dst_capacity;
2124 * The load is corrected for the CPU capacity available on each node.
2127 * ------------ vs ---------
2128 * src_capacity dst_capacity
2130 src_capacity = env->src_stats.compute_capacity;
2131 dst_capacity = env->dst_stats.compute_capacity;
2133 imb = abs(dst_load * src_capacity - src_load * dst_capacity);
2135 orig_src_load = env->src_stats.load;
2136 orig_dst_load = env->dst_stats.load;
2138 old_imb = abs(orig_dst_load * src_capacity - orig_src_load * dst_capacity);
2140 /* Would this change make things worse? */
2141 return (imb > old_imb);
2145 * Maximum NUMA importance can be 1998 (2*999);
2146 * SMALLIMP @ 30 would be close to 1998/64.
2147 * Used to deter task migration.
2152 * This checks if the overall compute and NUMA accesses of the system would
2153 * be improved if the source tasks was migrated to the target dst_cpu taking
2154 * into account that it might be best if task running on the dst_cpu should
2155 * be exchanged with the source task
2157 static bool task_numa_compare(struct task_numa_env *env,
2158 long taskimp, long groupimp, bool maymove)
2160 struct numa_group *cur_ng, *p_ng = deref_curr_numa_group(env->p);
2161 struct rq *dst_rq = cpu_rq(env->dst_cpu);
2162 long imp = p_ng ? groupimp : taskimp;
2163 struct task_struct *cur;
2164 long src_load, dst_load;
2165 int dist = env->dist;
2168 bool stopsearch = false;
2170 if (READ_ONCE(dst_rq->numa_migrate_on))
2174 cur = rcu_dereference(dst_rq->curr);
2175 if (cur && ((cur->flags & PF_EXITING) || is_idle_task(cur)))
2179 * Because we have preemption enabled we can get migrated around and
2180 * end try selecting ourselves (current == env->p) as a swap candidate.
2182 if (cur == env->p) {
2188 if (maymove && moveimp >= env->best_imp)
2194 /* Skip this swap candidate if cannot move to the source cpu. */
2195 if (!cpumask_test_cpu(env->src_cpu, cur->cpus_ptr))
2199 * Skip this swap candidate if it is not moving to its preferred
2200 * node and the best task is.
2202 if (env->best_task &&
2203 env->best_task->numa_preferred_nid == env->src_nid &&
2204 cur->numa_preferred_nid != env->src_nid) {
2209 * "imp" is the fault differential for the source task between the
2210 * source and destination node. Calculate the total differential for
2211 * the source task and potential destination task. The more negative
2212 * the value is, the more remote accesses that would be expected to
2213 * be incurred if the tasks were swapped.
2215 * If dst and source tasks are in the same NUMA group, or not
2216 * in any group then look only at task weights.
2218 cur_ng = rcu_dereference(cur->numa_group);
2219 if (cur_ng == p_ng) {
2221 * Do not swap within a group or between tasks that have
2222 * no group if there is spare capacity. Swapping does
2223 * not address the load imbalance and helps one task at
2224 * the cost of punishing another.
2226 if (env->dst_stats.node_type == node_has_spare)
2229 imp = taskimp + task_weight(cur, env->src_nid, dist) -
2230 task_weight(cur, env->dst_nid, dist);
2232 * Add some hysteresis to prevent swapping the
2233 * tasks within a group over tiny differences.
2239 * Compare the group weights. If a task is all by itself
2240 * (not part of a group), use the task weight instead.
2243 imp += group_weight(cur, env->src_nid, dist) -
2244 group_weight(cur, env->dst_nid, dist);
2246 imp += task_weight(cur, env->src_nid, dist) -
2247 task_weight(cur, env->dst_nid, dist);
2250 /* Discourage picking a task already on its preferred node */
2251 if (cur->numa_preferred_nid == env->dst_nid)
2255 * Encourage picking a task that moves to its preferred node.
2256 * This potentially makes imp larger than it's maximum of
2257 * 1998 (see SMALLIMP and task_weight for why) but in this
2258 * case, it does not matter.
2260 if (cur->numa_preferred_nid == env->src_nid)
2263 if (maymove && moveimp > imp && moveimp > env->best_imp) {
2270 * Prefer swapping with a task moving to its preferred node over a
2273 if (env->best_task && cur->numa_preferred_nid == env->src_nid &&
2274 env->best_task->numa_preferred_nid != env->src_nid) {
2279 * If the NUMA importance is less than SMALLIMP,
2280 * task migration might only result in ping pong
2281 * of tasks and also hurt performance due to cache
2284 if (imp < SMALLIMP || imp <= env->best_imp + SMALLIMP / 2)
2288 * In the overloaded case, try and keep the load balanced.
2290 load = task_h_load(env->p) - task_h_load(cur);
2294 dst_load = env->dst_stats.load + load;
2295 src_load = env->src_stats.load - load;
2297 if (load_too_imbalanced(src_load, dst_load, env))
2301 /* Evaluate an idle CPU for a task numa move. */
2303 int cpu = env->dst_stats.idle_cpu;
2305 /* Nothing cached so current CPU went idle since the search. */
2310 * If the CPU is no longer truly idle and the previous best CPU
2311 * is, keep using it.
2313 if (!idle_cpu(cpu) && env->best_cpu >= 0 &&
2314 idle_cpu(env->best_cpu)) {
2315 cpu = env->best_cpu;
2321 task_numa_assign(env, cur, imp);
2324 * If a move to idle is allowed because there is capacity or load
2325 * balance improves then stop the search. While a better swap
2326 * candidate may exist, a search is not free.
2328 if (maymove && !cur && env->best_cpu >= 0 && idle_cpu(env->best_cpu))
2332 * If a swap candidate must be identified and the current best task
2333 * moves its preferred node then stop the search.
2335 if (!maymove && env->best_task &&
2336 env->best_task->numa_preferred_nid == env->src_nid) {
2345 static void task_numa_find_cpu(struct task_numa_env *env,
2346 long taskimp, long groupimp)
2348 bool maymove = false;
2352 * If dst node has spare capacity, then check if there is an
2353 * imbalance that would be overruled by the load balancer.
2355 if (env->dst_stats.node_type == node_has_spare) {
2356 unsigned int imbalance;
2357 int src_running, dst_running;
2360 * Would movement cause an imbalance? Note that if src has
2361 * more running tasks that the imbalance is ignored as the
2362 * move improves the imbalance from the perspective of the
2363 * CPU load balancer.
2365 src_running = env->src_stats.nr_running - 1;
2366 dst_running = env->dst_stats.nr_running + 1;
2367 imbalance = max(0, dst_running - src_running);
2368 imbalance = adjust_numa_imbalance(imbalance, dst_running,
2371 /* Use idle CPU if there is no imbalance */
2374 if (env->dst_stats.idle_cpu >= 0) {
2375 env->dst_cpu = env->dst_stats.idle_cpu;
2376 task_numa_assign(env, NULL, 0);
2381 long src_load, dst_load, load;
2383 * If the improvement from just moving env->p direction is better
2384 * than swapping tasks around, check if a move is possible.
2386 load = task_h_load(env->p);
2387 dst_load = env->dst_stats.load + load;
2388 src_load = env->src_stats.load - load;
2389 maymove = !load_too_imbalanced(src_load, dst_load, env);
2392 for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) {
2393 /* Skip this CPU if the source task cannot migrate */
2394 if (!cpumask_test_cpu(cpu, env->p->cpus_ptr))
2398 if (task_numa_compare(env, taskimp, groupimp, maymove))
2403 static int task_numa_migrate(struct task_struct *p)
2405 struct task_numa_env env = {
2408 .src_cpu = task_cpu(p),
2409 .src_nid = task_node(p),
2411 .imbalance_pct = 112,
2417 unsigned long taskweight, groupweight;
2418 struct sched_domain *sd;
2419 long taskimp, groupimp;
2420 struct numa_group *ng;
2425 * Pick the lowest SD_NUMA domain, as that would have the smallest
2426 * imbalance and would be the first to start moving tasks about.
2428 * And we want to avoid any moving of tasks about, as that would create
2429 * random movement of tasks -- counter the numa conditions we're trying
2433 sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu));
2435 env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2;
2436 env.imb_numa_nr = sd->imb_numa_nr;
2441 * Cpusets can break the scheduler domain tree into smaller
2442 * balance domains, some of which do not cross NUMA boundaries.
2443 * Tasks that are "trapped" in such domains cannot be migrated
2444 * elsewhere, so there is no point in (re)trying.
2446 if (unlikely(!sd)) {
2447 sched_setnuma(p, task_node(p));
2451 env.dst_nid = p->numa_preferred_nid;
2452 dist = env.dist = node_distance(env.src_nid, env.dst_nid);
2453 taskweight = task_weight(p, env.src_nid, dist);
2454 groupweight = group_weight(p, env.src_nid, dist);
2455 update_numa_stats(&env, &env.src_stats, env.src_nid, false);
2456 taskimp = task_weight(p, env.dst_nid, dist) - taskweight;
2457 groupimp = group_weight(p, env.dst_nid, dist) - groupweight;
2458 update_numa_stats(&env, &env.dst_stats, env.dst_nid, true);
2460 /* Try to find a spot on the preferred nid. */
2461 task_numa_find_cpu(&env, taskimp, groupimp);
2464 * Look at other nodes in these cases:
2465 * - there is no space available on the preferred_nid
2466 * - the task is part of a numa_group that is interleaved across
2467 * multiple NUMA nodes; in order to better consolidate the group,
2468 * we need to check other locations.
2470 ng = deref_curr_numa_group(p);
2471 if (env.best_cpu == -1 || (ng && ng->active_nodes > 1)) {
2472 for_each_node_state(nid, N_CPU) {
2473 if (nid == env.src_nid || nid == p->numa_preferred_nid)
2476 dist = node_distance(env.src_nid, env.dst_nid);
2477 if (sched_numa_topology_type == NUMA_BACKPLANE &&
2479 taskweight = task_weight(p, env.src_nid, dist);
2480 groupweight = group_weight(p, env.src_nid, dist);
2483 /* Only consider nodes where both task and groups benefit */
2484 taskimp = task_weight(p, nid, dist) - taskweight;
2485 groupimp = group_weight(p, nid, dist) - groupweight;
2486 if (taskimp < 0 && groupimp < 0)
2491 update_numa_stats(&env, &env.dst_stats, env.dst_nid, true);
2492 task_numa_find_cpu(&env, taskimp, groupimp);
2497 * If the task is part of a workload that spans multiple NUMA nodes,
2498 * and is migrating into one of the workload's active nodes, remember
2499 * this node as the task's preferred numa node, so the workload can
2501 * A task that migrated to a second choice node will be better off
2502 * trying for a better one later. Do not set the preferred node here.
2505 if (env.best_cpu == -1)
2508 nid = cpu_to_node(env.best_cpu);
2510 if (nid != p->numa_preferred_nid)
2511 sched_setnuma(p, nid);
2514 /* No better CPU than the current one was found. */
2515 if (env.best_cpu == -1) {
2516 trace_sched_stick_numa(p, env.src_cpu, NULL, -1);
2520 best_rq = cpu_rq(env.best_cpu);
2521 if (env.best_task == NULL) {
2522 ret = migrate_task_to(p, env.best_cpu);
2523 WRITE_ONCE(best_rq->numa_migrate_on, 0);
2525 trace_sched_stick_numa(p, env.src_cpu, NULL, env.best_cpu);
2529 ret = migrate_swap(p, env.best_task, env.best_cpu, env.src_cpu);
2530 WRITE_ONCE(best_rq->numa_migrate_on, 0);
2533 trace_sched_stick_numa(p, env.src_cpu, env.best_task, env.best_cpu);
2534 put_task_struct(env.best_task);
2538 /* Attempt to migrate a task to a CPU on the preferred node. */
2539 static void numa_migrate_preferred(struct task_struct *p)
2541 unsigned long interval = HZ;
2543 /* This task has no NUMA fault statistics yet */
2544 if (unlikely(p->numa_preferred_nid == NUMA_NO_NODE || !p->numa_faults))
2547 /* Periodically retry migrating the task to the preferred node */
2548 interval = min(interval, msecs_to_jiffies(p->numa_scan_period) / 16);
2549 p->numa_migrate_retry = jiffies + interval;
2551 /* Success if task is already running on preferred CPU */
2552 if (task_node(p) == p->numa_preferred_nid)
2555 /* Otherwise, try migrate to a CPU on the preferred node */
2556 task_numa_migrate(p);
2560 * Find out how many nodes the workload is actively running on. Do this by
2561 * tracking the nodes from which NUMA hinting faults are triggered. This can
2562 * be different from the set of nodes where the workload's memory is currently
2565 static void numa_group_count_active_nodes(struct numa_group *numa_group)
2567 unsigned long faults, max_faults = 0;
2568 int nid, active_nodes = 0;
2570 for_each_node_state(nid, N_CPU) {
2571 faults = group_faults_cpu(numa_group, nid);
2572 if (faults > max_faults)
2573 max_faults = faults;
2576 for_each_node_state(nid, N_CPU) {
2577 faults = group_faults_cpu(numa_group, nid);
2578 if (faults * ACTIVE_NODE_FRACTION > max_faults)
2582 numa_group->max_faults_cpu = max_faults;
2583 numa_group->active_nodes = active_nodes;
2587 * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
2588 * increments. The more local the fault statistics are, the higher the scan
2589 * period will be for the next scan window. If local/(local+remote) ratio is
2590 * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS)
2591 * the scan period will decrease. Aim for 70% local accesses.
2593 #define NUMA_PERIOD_SLOTS 10
2594 #define NUMA_PERIOD_THRESHOLD 7
2597 * Increase the scan period (slow down scanning) if the majority of
2598 * our memory is already on our local node, or if the majority of
2599 * the page accesses are shared with other processes.
2600 * Otherwise, decrease the scan period.
2602 static void update_task_scan_period(struct task_struct *p,
2603 unsigned long shared, unsigned long private)
2605 unsigned int period_slot;
2606 int lr_ratio, ps_ratio;
2609 unsigned long remote = p->numa_faults_locality[0];
2610 unsigned long local = p->numa_faults_locality[1];
2613 * If there were no record hinting faults then either the task is
2614 * completely idle or all activity is in areas that are not of interest
2615 * to automatic numa balancing. Related to that, if there were failed
2616 * migration then it implies we are migrating too quickly or the local
2617 * node is overloaded. In either case, scan slower
2619 if (local + shared == 0 || p->numa_faults_locality[2]) {
2620 p->numa_scan_period = min(p->numa_scan_period_max,
2621 p->numa_scan_period << 1);
2623 p->mm->numa_next_scan = jiffies +
2624 msecs_to_jiffies(p->numa_scan_period);
2630 * Prepare to scale scan period relative to the current period.
2631 * == NUMA_PERIOD_THRESHOLD scan period stays the same
2632 * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
2633 * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
2635 period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS);
2636 lr_ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote);
2637 ps_ratio = (private * NUMA_PERIOD_SLOTS) / (private + shared);
2639 if (ps_ratio >= NUMA_PERIOD_THRESHOLD) {
2641 * Most memory accesses are local. There is no need to
2642 * do fast NUMA scanning, since memory is already local.
2644 int slot = ps_ratio - NUMA_PERIOD_THRESHOLD;
2647 diff = slot * period_slot;
2648 } else if (lr_ratio >= NUMA_PERIOD_THRESHOLD) {
2650 * Most memory accesses are shared with other tasks.
2651 * There is no point in continuing fast NUMA scanning,
2652 * since other tasks may just move the memory elsewhere.
2654 int slot = lr_ratio - NUMA_PERIOD_THRESHOLD;
2657 diff = slot * period_slot;
2660 * Private memory faults exceed (SLOTS-THRESHOLD)/SLOTS,
2661 * yet they are not on the local NUMA node. Speed up
2662 * NUMA scanning to get the memory moved over.
2664 int ratio = max(lr_ratio, ps_ratio);
2665 diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot;
2668 p->numa_scan_period = clamp(p->numa_scan_period + diff,
2669 task_scan_min(p), task_scan_max(p));
2670 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
2674 * Get the fraction of time the task has been running since the last
2675 * NUMA placement cycle. The scheduler keeps similar statistics, but
2676 * decays those on a 32ms period, which is orders of magnitude off
2677 * from the dozens-of-seconds NUMA balancing period. Use the scheduler
2678 * stats only if the task is so new there are no NUMA statistics yet.
2680 static u64 numa_get_avg_runtime(struct task_struct *p, u64 *period)
2682 u64 runtime, delta, now;
2683 /* Use the start of this time slice to avoid calculations. */
2684 now = p->se.exec_start;
2685 runtime = p->se.sum_exec_runtime;
2687 if (p->last_task_numa_placement) {
2688 delta = runtime - p->last_sum_exec_runtime;
2689 *period = now - p->last_task_numa_placement;
2691 /* Avoid time going backwards, prevent potential divide error: */
2692 if (unlikely((s64)*period < 0))
2695 delta = p->se.avg.load_sum;
2696 *period = LOAD_AVG_MAX;
2699 p->last_sum_exec_runtime = runtime;
2700 p->last_task_numa_placement = now;
2706 * Determine the preferred nid for a task in a numa_group. This needs to
2707 * be done in a way that produces consistent results with group_weight,
2708 * otherwise workloads might not converge.
2710 static int preferred_group_nid(struct task_struct *p, int nid)
2715 /* Direct connections between all NUMA nodes. */
2716 if (sched_numa_topology_type == NUMA_DIRECT)
2720 * On a system with glueless mesh NUMA topology, group_weight
2721 * scores nodes according to the number of NUMA hinting faults on
2722 * both the node itself, and on nearby nodes.
2724 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
2725 unsigned long score, max_score = 0;
2726 int node, max_node = nid;
2728 dist = sched_max_numa_distance;
2730 for_each_node_state(node, N_CPU) {
2731 score = group_weight(p, node, dist);
2732 if (score > max_score) {
2741 * Finding the preferred nid in a system with NUMA backplane
2742 * interconnect topology is more involved. The goal is to locate
2743 * tasks from numa_groups near each other in the system, and
2744 * untangle workloads from different sides of the system. This requires
2745 * searching down the hierarchy of node groups, recursively searching
2746 * inside the highest scoring group of nodes. The nodemask tricks
2747 * keep the complexity of the search down.
2749 nodes = node_states[N_CPU];
2750 for (dist = sched_max_numa_distance; dist > LOCAL_DISTANCE; dist--) {
2751 unsigned long max_faults = 0;
2752 nodemask_t max_group = NODE_MASK_NONE;
2755 /* Are there nodes at this distance from each other? */
2756 if (!find_numa_distance(dist))
2759 for_each_node_mask(a, nodes) {
2760 unsigned long faults = 0;
2761 nodemask_t this_group;
2762 nodes_clear(this_group);
2764 /* Sum group's NUMA faults; includes a==b case. */
2765 for_each_node_mask(b, nodes) {
2766 if (node_distance(a, b) < dist) {
2767 faults += group_faults(p, b);
2768 node_set(b, this_group);
2769 node_clear(b, nodes);
2773 /* Remember the top group. */
2774 if (faults > max_faults) {
2775 max_faults = faults;
2776 max_group = this_group;
2778 * subtle: at the smallest distance there is
2779 * just one node left in each "group", the
2780 * winner is the preferred nid.
2785 /* Next round, evaluate the nodes within max_group. */
2793 static void task_numa_placement(struct task_struct *p)
2795 int seq, nid, max_nid = NUMA_NO_NODE;
2796 unsigned long max_faults = 0;
2797 unsigned long fault_types[2] = { 0, 0 };
2798 unsigned long total_faults;
2799 u64 runtime, period;
2800 spinlock_t *group_lock = NULL;
2801 struct numa_group *ng;
2804 * The p->mm->numa_scan_seq field gets updated without
2805 * exclusive access. Use READ_ONCE() here to ensure
2806 * that the field is read in a single access:
2808 seq = READ_ONCE(p->mm->numa_scan_seq);
2809 if (p->numa_scan_seq == seq)
2811 p->numa_scan_seq = seq;
2812 p->numa_scan_period_max = task_scan_max(p);
2814 total_faults = p->numa_faults_locality[0] +
2815 p->numa_faults_locality[1];
2816 runtime = numa_get_avg_runtime(p, &period);
2818 /* If the task is part of a group prevent parallel updates to group stats */
2819 ng = deref_curr_numa_group(p);
2821 group_lock = &ng->lock;
2822 spin_lock_irq(group_lock);
2825 /* Find the node with the highest number of faults */
2826 for_each_online_node(nid) {
2827 /* Keep track of the offsets in numa_faults array */
2828 int mem_idx, membuf_idx, cpu_idx, cpubuf_idx;
2829 unsigned long faults = 0, group_faults = 0;
2832 for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) {
2833 long diff, f_diff, f_weight;
2835 mem_idx = task_faults_idx(NUMA_MEM, nid, priv);
2836 membuf_idx = task_faults_idx(NUMA_MEMBUF, nid, priv);
2837 cpu_idx = task_faults_idx(NUMA_CPU, nid, priv);
2838 cpubuf_idx = task_faults_idx(NUMA_CPUBUF, nid, priv);
2840 /* Decay existing window, copy faults since last scan */
2841 diff = p->numa_faults[membuf_idx] - p->numa_faults[mem_idx] / 2;
2842 fault_types[priv] += p->numa_faults[membuf_idx];
2843 p->numa_faults[membuf_idx] = 0;
2846 * Normalize the faults_from, so all tasks in a group
2847 * count according to CPU use, instead of by the raw
2848 * number of faults. Tasks with little runtime have
2849 * little over-all impact on throughput, and thus their
2850 * faults are less important.
2852 f_weight = div64_u64(runtime << 16, period + 1);
2853 f_weight = (f_weight * p->numa_faults[cpubuf_idx]) /
2855 f_diff = f_weight - p->numa_faults[cpu_idx] / 2;
2856 p->numa_faults[cpubuf_idx] = 0;
2858 p->numa_faults[mem_idx] += diff;
2859 p->numa_faults[cpu_idx] += f_diff;
2860 faults += p->numa_faults[mem_idx];
2861 p->total_numa_faults += diff;
2864 * safe because we can only change our own group
2866 * mem_idx represents the offset for a given
2867 * nid and priv in a specific region because it
2868 * is at the beginning of the numa_faults array.
2870 ng->faults[mem_idx] += diff;
2871 ng->faults[cpu_idx] += f_diff;
2872 ng->total_faults += diff;
2873 group_faults += ng->faults[mem_idx];
2878 if (faults > max_faults) {
2879 max_faults = faults;
2882 } else if (group_faults > max_faults) {
2883 max_faults = group_faults;
2888 /* Cannot migrate task to CPU-less node */
2889 max_nid = numa_nearest_node(max_nid, N_CPU);
2892 numa_group_count_active_nodes(ng);
2893 spin_unlock_irq(group_lock);
2894 max_nid = preferred_group_nid(p, max_nid);
2898 /* Set the new preferred node */
2899 if (max_nid != p->numa_preferred_nid)
2900 sched_setnuma(p, max_nid);
2903 update_task_scan_period(p, fault_types[0], fault_types[1]);
2906 static inline int get_numa_group(struct numa_group *grp)
2908 return refcount_inc_not_zero(&grp->refcount);
2911 static inline void put_numa_group(struct numa_group *grp)
2913 if (refcount_dec_and_test(&grp->refcount))
2914 kfree_rcu(grp, rcu);
2917 static void task_numa_group(struct task_struct *p, int cpupid, int flags,
2920 struct numa_group *grp, *my_grp;
2921 struct task_struct *tsk;
2923 int cpu = cpupid_to_cpu(cpupid);
2926 if (unlikely(!deref_curr_numa_group(p))) {
2927 unsigned int size = sizeof(struct numa_group) +
2928 NR_NUMA_HINT_FAULT_STATS *
2929 nr_node_ids * sizeof(unsigned long);
2931 grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
2935 refcount_set(&grp->refcount, 1);
2936 grp->active_nodes = 1;
2937 grp->max_faults_cpu = 0;
2938 spin_lock_init(&grp->lock);
2941 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2942 grp->faults[i] = p->numa_faults[i];
2944 grp->total_faults = p->total_numa_faults;
2947 rcu_assign_pointer(p->numa_group, grp);
2951 tsk = READ_ONCE(cpu_rq(cpu)->curr);
2953 if (!cpupid_match_pid(tsk, cpupid))
2956 grp = rcu_dereference(tsk->numa_group);
2960 my_grp = deref_curr_numa_group(p);
2965 * Only join the other group if its bigger; if we're the bigger group,
2966 * the other task will join us.
2968 if (my_grp->nr_tasks > grp->nr_tasks)
2972 * Tie-break on the grp address.
2974 if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp)
2977 /* Always join threads in the same process. */
2978 if (tsk->mm == current->mm)
2981 /* Simple filter to avoid false positives due to PID collisions */
2982 if (flags & TNF_SHARED)
2985 /* Update priv based on whether false sharing was detected */
2988 if (join && !get_numa_group(grp))
2996 WARN_ON_ONCE(irqs_disabled());
2997 double_lock_irq(&my_grp->lock, &grp->lock);
2999 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) {
3000 my_grp->faults[i] -= p->numa_faults[i];
3001 grp->faults[i] += p->numa_faults[i];
3003 my_grp->total_faults -= p->total_numa_faults;
3004 grp->total_faults += p->total_numa_faults;
3009 spin_unlock(&my_grp->lock);
3010 spin_unlock_irq(&grp->lock);
3012 rcu_assign_pointer(p->numa_group, grp);
3014 put_numa_group(my_grp);
3023 * Get rid of NUMA statistics associated with a task (either current or dead).
3024 * If @final is set, the task is dead and has reached refcount zero, so we can
3025 * safely free all relevant data structures. Otherwise, there might be
3026 * concurrent reads from places like load balancing and procfs, and we should
3027 * reset the data back to default state without freeing ->numa_faults.
3029 void task_numa_free(struct task_struct *p, bool final)
3031 /* safe: p either is current or is being freed by current */
3032 struct numa_group *grp = rcu_dereference_raw(p->numa_group);
3033 unsigned long *numa_faults = p->numa_faults;
3034 unsigned long flags;
3041 spin_lock_irqsave(&grp->lock, flags);
3042 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
3043 grp->faults[i] -= p->numa_faults[i];
3044 grp->total_faults -= p->total_numa_faults;
3047 spin_unlock_irqrestore(&grp->lock, flags);
3048 RCU_INIT_POINTER(p->numa_group, NULL);
3049 put_numa_group(grp);
3053 p->numa_faults = NULL;
3056 p->total_numa_faults = 0;
3057 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
3063 * Got a PROT_NONE fault for a page on @node.
3065 void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags)
3067 struct task_struct *p = current;
3068 bool migrated = flags & TNF_MIGRATED;
3069 int cpu_node = task_node(current);
3070 int local = !!(flags & TNF_FAULT_LOCAL);
3071 struct numa_group *ng;
3074 if (!static_branch_likely(&sched_numa_balancing))
3077 /* for example, ksmd faulting in a user's mm */
3082 * NUMA faults statistics are unnecessary for the slow memory
3083 * node for memory tiering mode.
3085 if (!node_is_toptier(mem_node) &&
3086 (sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING ||
3087 !cpupid_valid(last_cpupid)))
3090 /* Allocate buffer to track faults on a per-node basis */
3091 if (unlikely(!p->numa_faults)) {
3092 int size = sizeof(*p->numa_faults) *
3093 NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids;
3095 p->numa_faults = kzalloc(size, GFP_KERNEL|__GFP_NOWARN);
3096 if (!p->numa_faults)
3099 p->total_numa_faults = 0;
3100 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
3104 * First accesses are treated as private, otherwise consider accesses
3105 * to be private if the accessing pid has not changed
3107 if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) {
3110 priv = cpupid_match_pid(p, last_cpupid);
3111 if (!priv && !(flags & TNF_NO_GROUP))
3112 task_numa_group(p, last_cpupid, flags, &priv);
3116 * If a workload spans multiple NUMA nodes, a shared fault that
3117 * occurs wholly within the set of nodes that the workload is
3118 * actively using should be counted as local. This allows the
3119 * scan rate to slow down when a workload has settled down.
3121 ng = deref_curr_numa_group(p);
3122 if (!priv && !local && ng && ng->active_nodes > 1 &&
3123 numa_is_active_node(cpu_node, ng) &&
3124 numa_is_active_node(mem_node, ng))
3128 * Retry to migrate task to preferred node periodically, in case it
3129 * previously failed, or the scheduler moved us.
3131 if (time_after(jiffies, p->numa_migrate_retry)) {
3132 task_numa_placement(p);
3133 numa_migrate_preferred(p);
3137 p->numa_pages_migrated += pages;
3138 if (flags & TNF_MIGRATE_FAIL)
3139 p->numa_faults_locality[2] += pages;
3141 p->numa_faults[task_faults_idx(NUMA_MEMBUF, mem_node, priv)] += pages;
3142 p->numa_faults[task_faults_idx(NUMA_CPUBUF, cpu_node, priv)] += pages;
3143 p->numa_faults_locality[local] += pages;
3146 static void reset_ptenuma_scan(struct task_struct *p)
3149 * We only did a read acquisition of the mmap sem, so
3150 * p->mm->numa_scan_seq is written to without exclusive access
3151 * and the update is not guaranteed to be atomic. That's not
3152 * much of an issue though, since this is just used for
3153 * statistical sampling. Use READ_ONCE/WRITE_ONCE, which are not
3154 * expensive, to avoid any form of compiler optimizations:
3156 WRITE_ONCE(p->mm->numa_scan_seq, READ_ONCE(p->mm->numa_scan_seq) + 1);
3157 p->mm->numa_scan_offset = 0;
3160 static bool vma_is_accessed(struct mm_struct *mm, struct vm_area_struct *vma)
3164 * Allow unconditional access first two times, so that all the (pages)
3165 * of VMAs get prot_none fault introduced irrespective of accesses.
3166 * This is also done to avoid any side effect of task scanning
3167 * amplifying the unfairness of disjoint set of VMAs' access.
3169 if ((READ_ONCE(current->mm->numa_scan_seq) - vma->numab_state->start_scan_seq) < 2)
3172 pids = vma->numab_state->pids_active[0] | vma->numab_state->pids_active[1];
3173 if (test_bit(hash_32(current->pid, ilog2(BITS_PER_LONG)), &pids))
3177 * Complete a scan that has already started regardless of PID access, or
3178 * some VMAs may never be scanned in multi-threaded applications:
3180 if (mm->numa_scan_offset > vma->vm_start) {
3181 trace_sched_skip_vma_numa(mm, vma, NUMAB_SKIP_IGNORE_PID);
3188 #define VMA_PID_RESET_PERIOD (4 * sysctl_numa_balancing_scan_delay)
3191 * The expensive part of numa migration is done from task_work context.
3192 * Triggered from task_tick_numa().
3194 static void task_numa_work(struct callback_head *work)
3196 unsigned long migrate, next_scan, now = jiffies;
3197 struct task_struct *p = current;
3198 struct mm_struct *mm = p->mm;
3199 u64 runtime = p->se.sum_exec_runtime;
3200 struct vm_area_struct *vma;
3201 unsigned long start, end;
3202 unsigned long nr_pte_updates = 0;
3203 long pages, virtpages;
3204 struct vma_iterator vmi;
3205 bool vma_pids_skipped;
3206 bool vma_pids_forced = false;
3208 SCHED_WARN_ON(p != container_of(work, struct task_struct, numa_work));
3212 * Who cares about NUMA placement when they're dying.
3214 * NOTE: make sure not to dereference p->mm before this check,
3215 * exit_task_work() happens _after_ exit_mm() so we could be called
3216 * without p->mm even though we still had it when we enqueued this
3219 if (p->flags & PF_EXITING)
3222 if (!mm->numa_next_scan) {
3223 mm->numa_next_scan = now +
3224 msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
3228 * Enforce maximal scan/migration frequency..
3230 migrate = mm->numa_next_scan;
3231 if (time_before(now, migrate))
3234 if (p->numa_scan_period == 0) {
3235 p->numa_scan_period_max = task_scan_max(p);
3236 p->numa_scan_period = task_scan_start(p);
3239 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
3240 if (!try_cmpxchg(&mm->numa_next_scan, &migrate, next_scan))
3244 * Delay this task enough that another task of this mm will likely win
3245 * the next time around.
3247 p->node_stamp += 2 * TICK_NSEC;
3249 pages = sysctl_numa_balancing_scan_size;
3250 pages <<= 20 - PAGE_SHIFT; /* MB in pages */
3251 virtpages = pages * 8; /* Scan up to this much virtual space */
3256 if (!mmap_read_trylock(mm))
3260 * VMAs are skipped if the current PID has not trapped a fault within
3261 * the VMA recently. Allow scanning to be forced if there is no
3262 * suitable VMA remaining.
3264 vma_pids_skipped = false;
3267 start = mm->numa_scan_offset;
3268 vma_iter_init(&vmi, mm, start);
3269 vma = vma_next(&vmi);
3271 reset_ptenuma_scan(p);
3273 vma_iter_set(&vmi, start);
3274 vma = vma_next(&vmi);
3278 if (!vma_migratable(vma) || !vma_policy_mof(vma) ||
3279 is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_MIXEDMAP)) {
3280 trace_sched_skip_vma_numa(mm, vma, NUMAB_SKIP_UNSUITABLE);
3285 * Shared library pages mapped by multiple processes are not
3286 * migrated as it is expected they are cache replicated. Avoid
3287 * hinting faults in read-only file-backed mappings or the vdso
3288 * as migrating the pages will be of marginal benefit.
3291 (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ))) {
3292 trace_sched_skip_vma_numa(mm, vma, NUMAB_SKIP_SHARED_RO);
3297 * Skip inaccessible VMAs to avoid any confusion between
3298 * PROT_NONE and NUMA hinting ptes
3300 if (!vma_is_accessible(vma)) {
3301 trace_sched_skip_vma_numa(mm, vma, NUMAB_SKIP_INACCESSIBLE);
3305 /* Initialise new per-VMA NUMAB state. */
3306 if (!vma->numab_state) {
3307 vma->numab_state = kzalloc(sizeof(struct vma_numab_state),
3309 if (!vma->numab_state)
3312 vma->numab_state->start_scan_seq = mm->numa_scan_seq;
3314 vma->numab_state->next_scan = now +
3315 msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
3317 /* Reset happens after 4 times scan delay of scan start */
3318 vma->numab_state->pids_active_reset = vma->numab_state->next_scan +
3319 msecs_to_jiffies(VMA_PID_RESET_PERIOD);
3322 * Ensure prev_scan_seq does not match numa_scan_seq,
3323 * to prevent VMAs being skipped prematurely on the
3326 vma->numab_state->prev_scan_seq = mm->numa_scan_seq - 1;
3330 * Scanning the VMA's of short lived tasks add more overhead. So
3331 * delay the scan for new VMAs.
3333 if (mm->numa_scan_seq && time_before(jiffies,
3334 vma->numab_state->next_scan)) {
3335 trace_sched_skip_vma_numa(mm, vma, NUMAB_SKIP_SCAN_DELAY);
3339 /* RESET access PIDs regularly for old VMAs. */
3340 if (mm->numa_scan_seq &&
3341 time_after(jiffies, vma->numab_state->pids_active_reset)) {
3342 vma->numab_state->pids_active_reset = vma->numab_state->pids_active_reset +
3343 msecs_to_jiffies(VMA_PID_RESET_PERIOD);
3344 vma->numab_state->pids_active[0] = READ_ONCE(vma->numab_state->pids_active[1]);
3345 vma->numab_state->pids_active[1] = 0;
3348 /* Do not rescan VMAs twice within the same sequence. */
3349 if (vma->numab_state->prev_scan_seq == mm->numa_scan_seq) {
3350 mm->numa_scan_offset = vma->vm_end;
3351 trace_sched_skip_vma_numa(mm, vma, NUMAB_SKIP_SEQ_COMPLETED);
3356 * Do not scan the VMA if task has not accessed it, unless no other
3357 * VMA candidate exists.
3359 if (!vma_pids_forced && !vma_is_accessed(mm, vma)) {
3360 vma_pids_skipped = true;
3361 trace_sched_skip_vma_numa(mm, vma, NUMAB_SKIP_PID_INACTIVE);
3366 start = max(start, vma->vm_start);
3367 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
3368 end = min(end, vma->vm_end);
3369 nr_pte_updates = change_prot_numa(vma, start, end);
3372 * Try to scan sysctl_numa_balancing_size worth of
3373 * hpages that have at least one present PTE that
3374 * is not already pte-numa. If the VMA contains
3375 * areas that are unused or already full of prot_numa
3376 * PTEs, scan up to virtpages, to skip through those
3380 pages -= (end - start) >> PAGE_SHIFT;
3381 virtpages -= (end - start) >> PAGE_SHIFT;
3384 if (pages <= 0 || virtpages <= 0)
3388 } while (end != vma->vm_end);
3390 /* VMA scan is complete, do not scan until next sequence. */
3391 vma->numab_state->prev_scan_seq = mm->numa_scan_seq;
3394 * Only force scan within one VMA at a time, to limit the
3395 * cost of scanning a potentially uninteresting VMA.
3397 if (vma_pids_forced)
3399 } for_each_vma(vmi, vma);
3402 * If no VMAs are remaining and VMAs were skipped due to the PID
3403 * not accessing the VMA previously, then force a scan to ensure
3406 if (!vma && !vma_pids_forced && vma_pids_skipped) {
3407 vma_pids_forced = true;
3413 * It is possible to reach the end of the VMA list but the last few
3414 * VMAs are not guaranteed to the vma_migratable. If they are not, we
3415 * would find the !migratable VMA on the next scan but not reset the
3416 * scanner to the start so check it now.
3419 mm->numa_scan_offset = start;
3421 reset_ptenuma_scan(p);
3422 mmap_read_unlock(mm);
3425 * Make sure tasks use at least 32x as much time to run other code
3426 * than they used here, to limit NUMA PTE scanning overhead to 3% max.
3427 * Usually update_task_scan_period slows down scanning enough; on an
3428 * overloaded system we need to limit overhead on a per task basis.
3430 if (unlikely(p->se.sum_exec_runtime != runtime)) {
3431 u64 diff = p->se.sum_exec_runtime - runtime;
3432 p->node_stamp += 32 * diff;
3436 void init_numa_balancing(unsigned long clone_flags, struct task_struct *p)
3439 struct mm_struct *mm = p->mm;
3442 mm_users = atomic_read(&mm->mm_users);
3443 if (mm_users == 1) {
3444 mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
3445 mm->numa_scan_seq = 0;
3449 p->numa_scan_seq = mm ? mm->numa_scan_seq : 0;
3450 p->numa_scan_period = sysctl_numa_balancing_scan_delay;
3451 p->numa_migrate_retry = 0;
3452 /* Protect against double add, see task_tick_numa and task_numa_work */
3453 p->numa_work.next = &p->numa_work;
3454 p->numa_faults = NULL;
3455 p->numa_pages_migrated = 0;
3456 p->total_numa_faults = 0;
3457 RCU_INIT_POINTER(p->numa_group, NULL);
3458 p->last_task_numa_placement = 0;
3459 p->last_sum_exec_runtime = 0;
3461 init_task_work(&p->numa_work, task_numa_work);
3463 /* New address space, reset the preferred nid */
3464 if (!(clone_flags & CLONE_VM)) {
3465 p->numa_preferred_nid = NUMA_NO_NODE;
3470 * New thread, keep existing numa_preferred_nid which should be copied
3471 * already by arch_dup_task_struct but stagger when scans start.
3476 delay = min_t(unsigned int, task_scan_max(current),
3477 current->numa_scan_period * mm_users * NSEC_PER_MSEC);
3478 delay += 2 * TICK_NSEC;
3479 p->node_stamp = delay;
3484 * Drive the periodic memory faults..
3486 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
3488 struct callback_head *work = &curr->numa_work;
3492 * We don't care about NUMA placement if we don't have memory.
3494 if (!curr->mm || (curr->flags & (PF_EXITING | PF_KTHREAD)) || work->next != work)
3498 * Using runtime rather than walltime has the dual advantage that
3499 * we (mostly) drive the selection from busy threads and that the
3500 * task needs to have done some actual work before we bother with
3503 now = curr->se.sum_exec_runtime;
3504 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
3506 if (now > curr->node_stamp + period) {
3507 if (!curr->node_stamp)
3508 curr->numa_scan_period = task_scan_start(curr);
3509 curr->node_stamp += period;
3511 if (!time_before(jiffies, curr->mm->numa_next_scan))
3512 task_work_add(curr, work, TWA_RESUME);
3516 static void update_scan_period(struct task_struct *p, int new_cpu)
3518 int src_nid = cpu_to_node(task_cpu(p));
3519 int dst_nid = cpu_to_node(new_cpu);
3521 if (!static_branch_likely(&sched_numa_balancing))
3524 if (!p->mm || !p->numa_faults || (p->flags & PF_EXITING))
3527 if (src_nid == dst_nid)
3531 * Allow resets if faults have been trapped before one scan
3532 * has completed. This is most likely due to a new task that
3533 * is pulled cross-node due to wakeups or load balancing.
3535 if (p->numa_scan_seq) {
3537 * Avoid scan adjustments if moving to the preferred
3538 * node or if the task was not previously running on
3539 * the preferred node.
3541 if (dst_nid == p->numa_preferred_nid ||
3542 (p->numa_preferred_nid != NUMA_NO_NODE &&
3543 src_nid != p->numa_preferred_nid))
3547 p->numa_scan_period = task_scan_start(p);
3551 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
3555 static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p)
3559 static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p)
3563 static inline void update_scan_period(struct task_struct *p, int new_cpu)
3567 #endif /* CONFIG_NUMA_BALANCING */
3570 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
3572 update_load_add(&cfs_rq->load, se->load.weight);
3574 if (entity_is_task(se)) {
3575 struct rq *rq = rq_of(cfs_rq);
3577 account_numa_enqueue(rq, task_of(se));
3578 list_add(&se->group_node, &rq->cfs_tasks);
3581 cfs_rq->nr_running++;
3583 cfs_rq->idle_nr_running++;
3587 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
3589 update_load_sub(&cfs_rq->load, se->load.weight);
3591 if (entity_is_task(se)) {
3592 account_numa_dequeue(rq_of(cfs_rq), task_of(se));
3593 list_del_init(&se->group_node);
3596 cfs_rq->nr_running--;
3598 cfs_rq->idle_nr_running--;
3602 * Signed add and clamp on underflow.
3604 * Explicitly do a load-store to ensure the intermediate value never hits
3605 * memory. This allows lockless observations without ever seeing the negative
3608 #define add_positive(_ptr, _val) do { \
3609 typeof(_ptr) ptr = (_ptr); \
3610 typeof(_val) val = (_val); \
3611 typeof(*ptr) res, var = READ_ONCE(*ptr); \
3615 if (val < 0 && res > var) \
3618 WRITE_ONCE(*ptr, res); \
3622 * Unsigned subtract and clamp on underflow.
3624 * Explicitly do a load-store to ensure the intermediate value never hits
3625 * memory. This allows lockless observations without ever seeing the negative
3628 #define sub_positive(_ptr, _val) do { \
3629 typeof(_ptr) ptr = (_ptr); \
3630 typeof(*ptr) val = (_val); \
3631 typeof(*ptr) res, var = READ_ONCE(*ptr); \
3635 WRITE_ONCE(*ptr, res); \
3639 * Remove and clamp on negative, from a local variable.
3641 * A variant of sub_positive(), which does not use explicit load-store
3642 * and is thus optimized for local variable updates.
3644 #define lsub_positive(_ptr, _val) do { \
3645 typeof(_ptr) ptr = (_ptr); \
3646 *ptr -= min_t(typeof(*ptr), *ptr, _val); \
3651 enqueue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3653 cfs_rq->avg.load_avg += se->avg.load_avg;
3654 cfs_rq->avg.load_sum += se_weight(se) * se->avg.load_sum;
3658 dequeue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3660 sub_positive(&cfs_rq->avg.load_avg, se->avg.load_avg);
3661 sub_positive(&cfs_rq->avg.load_sum, se_weight(se) * se->avg.load_sum);
3662 /* See update_cfs_rq_load_avg() */
3663 cfs_rq->avg.load_sum = max_t(u32, cfs_rq->avg.load_sum,
3664 cfs_rq->avg.load_avg * PELT_MIN_DIVIDER);
3668 enqueue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
3670 dequeue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
3673 static void reweight_eevdf(struct cfs_rq *cfs_rq, struct sched_entity *se,
3674 unsigned long weight)
3676 unsigned long old_weight = se->load.weight;
3677 u64 avruntime = avg_vruntime(cfs_rq);
3684 * COROLLARY #1: The virtual runtime of the entity needs to be
3685 * adjusted if re-weight at !0-lag point.
3687 * Proof: For contradiction assume this is not true, so we can
3688 * re-weight without changing vruntime at !0-lag point.
3690 * Weight VRuntime Avg-VRuntime
3694 * Since lag needs to be preserved through re-weight:
3696 * lag = (V - v)*w = (V'- v')*w', where v = v'
3697 * ==> V' = (V - v)*w/w' + v (1)
3699 * Let W be the total weight of the entities before reweight,
3700 * since V' is the new weighted average of entities:
3702 * V' = (WV + w'v - wv) / (W + w' - w) (2)
3704 * by using (1) & (2) we obtain:
3706 * (WV + w'v - wv) / (W + w' - w) = (V - v)*w/w' + v
3707 * ==> (WV-Wv+Wv+w'v-wv)/(W+w'-w) = (V - v)*w/w' + v
3708 * ==> (WV - Wv)/(W + w' - w) + v = (V - v)*w/w' + v
3709 * ==> (V - v)*W/(W + w' - w) = (V - v)*w/w' (3)
3711 * Since we are doing at !0-lag point which means V != v, we
3714 * ==> W / (W + w' - w) = w / w'
3715 * ==> Ww' = Ww + ww' - ww
3716 * ==> W * (w' - w) = w * (w' - w)
3717 * ==> W = w (re-weight indicates w' != w)
3719 * So the cfs_rq contains only one entity, hence vruntime of
3720 * the entity @v should always equal to the cfs_rq's weighted
3721 * average vruntime @V, which means we will always re-weight
3722 * at 0-lag point, thus breach assumption. Proof completed.
3725 * COROLLARY #2: Re-weight does NOT affect weighted average
3726 * vruntime of all the entities.
3728 * Proof: According to corollary #1, Eq. (1) should be:
3730 * (V - v)*w = (V' - v')*w'
3731 * ==> v' = V' - (V - v)*w/w' (4)
3733 * According to the weighted average formula, we have:
3735 * V' = (WV - wv + w'v') / (W - w + w')
3736 * = (WV - wv + w'(V' - (V - v)w/w')) / (W - w + w')
3737 * = (WV - wv + w'V' - Vw + wv) / (W - w + w')
3738 * = (WV + w'V' - Vw) / (W - w + w')
3740 * ==> V'*(W - w + w') = WV + w'V' - Vw
3741 * ==> V' * (W - w) = (W - w) * V (5)
3743 * If the entity is the only one in the cfs_rq, then reweight
3744 * always occurs at 0-lag point, so V won't change. Or else
3745 * there are other entities, hence W != w, then Eq. (5) turns
3746 * into V' = V. So V won't change in either case, proof done.
3749 * So according to corollary #1 & #2, the effect of re-weight
3750 * on vruntime should be:
3752 * v' = V' - (V - v) * w / w' (4)
3753 * = V - (V - v) * w / w'
3757 if (avruntime != se->vruntime) {
3758 vlag = (s64)(avruntime - se->vruntime);
3759 vlag = div_s64(vlag * old_weight, weight);
3760 se->vruntime = avruntime - vlag;
3767 * When the weight changes, the virtual time slope changes and
3768 * we should adjust the relative virtual deadline accordingly.
3770 * d' = v' + (d - v)*w/w'
3771 * = V' - (V - v)*w/w' + (d - v)*w/w'
3772 * = V - (V - v)*w/w' + (d - v)*w/w'
3773 * = V + (d - V)*w/w'
3775 vslice = (s64)(se->deadline - avruntime);
3776 vslice = div_s64(vslice * old_weight, weight);
3777 se->deadline = avruntime + vslice;
3780 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
3781 unsigned long weight)
3783 bool curr = cfs_rq->curr == se;
3786 /* commit outstanding execution time */
3788 update_curr(cfs_rq);
3790 __dequeue_entity(cfs_rq, se);
3791 update_load_sub(&cfs_rq->load, se->load.weight);
3793 dequeue_load_avg(cfs_rq, se);
3797 * Because we keep se->vlag = V - v_i, while: lag_i = w_i*(V - v_i),
3798 * we need to scale se->vlag when w_i changes.
3800 se->vlag = div_s64(se->vlag * se->load.weight, weight);
3802 reweight_eevdf(cfs_rq, se, weight);
3805 update_load_set(&se->load, weight);
3809 u32 divider = get_pelt_divider(&se->avg);
3811 se->avg.load_avg = div_u64(se_weight(se) * se->avg.load_sum, divider);
3815 enqueue_load_avg(cfs_rq, se);
3817 update_load_add(&cfs_rq->load, se->load.weight);
3819 __enqueue_entity(cfs_rq, se);
3822 * The entity's vruntime has been adjusted, so let's check
3823 * whether the rq-wide min_vruntime needs updated too. Since
3824 * the calculations above require stable min_vruntime rather
3825 * than up-to-date one, we do the update at the end of the
3828 update_min_vruntime(cfs_rq);
3832 void reweight_task(struct task_struct *p, int prio)
3834 struct sched_entity *se = &p->se;
3835 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3836 struct load_weight *load = &se->load;
3837 unsigned long weight = scale_load(sched_prio_to_weight[prio]);
3839 reweight_entity(cfs_rq, se, weight);
3840 load->inv_weight = sched_prio_to_wmult[prio];
3843 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
3845 #ifdef CONFIG_FAIR_GROUP_SCHED
3848 * All this does is approximate the hierarchical proportion which includes that
3849 * global sum we all love to hate.
3851 * That is, the weight of a group entity, is the proportional share of the
3852 * group weight based on the group runqueue weights. That is:
3854 * tg->weight * grq->load.weight
3855 * ge->load.weight = ----------------------------- (1)
3856 * \Sum grq->load.weight
3858 * Now, because computing that sum is prohibitively expensive to compute (been
3859 * there, done that) we approximate it with this average stuff. The average
3860 * moves slower and therefore the approximation is cheaper and more stable.
3862 * So instead of the above, we substitute:
3864 * grq->load.weight -> grq->avg.load_avg (2)
3866 * which yields the following:
3868 * tg->weight * grq->avg.load_avg
3869 * ge->load.weight = ------------------------------ (3)
3872 * Where: tg->load_avg ~= \Sum grq->avg.load_avg
3874 * That is shares_avg, and it is right (given the approximation (2)).
3876 * The problem with it is that because the average is slow -- it was designed
3877 * to be exactly that of course -- this leads to transients in boundary
3878 * conditions. In specific, the case where the group was idle and we start the
3879 * one task. It takes time for our CPU's grq->avg.load_avg to build up,
3880 * yielding bad latency etc..
3882 * Now, in that special case (1) reduces to:
3884 * tg->weight * grq->load.weight
3885 * ge->load.weight = ----------------------------- = tg->weight (4)
3888 * That is, the sum collapses because all other CPUs are idle; the UP scenario.
3890 * So what we do is modify our approximation (3) to approach (4) in the (near)
3895 * tg->weight * grq->load.weight
3896 * --------------------------------------------------- (5)
3897 * tg->load_avg - grq->avg.load_avg + grq->load.weight
3899 * But because grq->load.weight can drop to 0, resulting in a divide by zero,
3900 * we need to use grq->avg.load_avg as its lower bound, which then gives:
3903 * tg->weight * grq->load.weight
3904 * ge->load.weight = ----------------------------- (6)
3909 * tg_load_avg' = tg->load_avg - grq->avg.load_avg +
3910 * max(grq->load.weight, grq->avg.load_avg)
3912 * And that is shares_weight and is icky. In the (near) UP case it approaches
3913 * (4) while in the normal case it approaches (3). It consistently
3914 * overestimates the ge->load.weight and therefore:
3916 * \Sum ge->load.weight >= tg->weight
3920 static long calc_group_shares(struct cfs_rq *cfs_rq)
3922 long tg_weight, tg_shares, load, shares;
3923 struct task_group *tg = cfs_rq->tg;
3925 tg_shares = READ_ONCE(tg->shares);
3927 load = max(scale_load_down(cfs_rq->load.weight), cfs_rq->avg.load_avg);
3929 tg_weight = atomic_long_read(&tg->load_avg);
3931 /* Ensure tg_weight >= load */
3932 tg_weight -= cfs_rq->tg_load_avg_contrib;
3935 shares = (tg_shares * load);
3937 shares /= tg_weight;
3940 * MIN_SHARES has to be unscaled here to support per-CPU partitioning
3941 * of a group with small tg->shares value. It is a floor value which is
3942 * assigned as a minimum load.weight to the sched_entity representing
3943 * the group on a CPU.
3945 * E.g. on 64-bit for a group with tg->shares of scale_load(15)=15*1024
3946 * on an 8-core system with 8 tasks each runnable on one CPU shares has
3947 * to be 15*1024*1/8=1920 instead of scale_load(MIN_SHARES)=2*1024. In
3948 * case no task is runnable on a CPU MIN_SHARES=2 should be returned
3951 return clamp_t(long, shares, MIN_SHARES, tg_shares);
3953 #endif /* CONFIG_SMP */
3956 * Recomputes the group entity based on the current state of its group
3959 static void update_cfs_group(struct sched_entity *se)
3961 struct cfs_rq *gcfs_rq = group_cfs_rq(se);
3967 if (throttled_hierarchy(gcfs_rq))
3971 shares = READ_ONCE(gcfs_rq->tg->shares);
3973 shares = calc_group_shares(gcfs_rq);
3975 if (unlikely(se->load.weight != shares))
3976 reweight_entity(cfs_rq_of(se), se, shares);
3979 #else /* CONFIG_FAIR_GROUP_SCHED */
3980 static inline void update_cfs_group(struct sched_entity *se)
3983 #endif /* CONFIG_FAIR_GROUP_SCHED */
3985 static inline void cfs_rq_util_change(struct cfs_rq *cfs_rq, int flags)
3987 struct rq *rq = rq_of(cfs_rq);
3989 if (&rq->cfs == cfs_rq) {
3991 * There are a few boundary cases this might miss but it should
3992 * get called often enough that that should (hopefully) not be
3995 * It will not get called when we go idle, because the idle
3996 * thread is a different class (!fair), nor will the utilization
3997 * number include things like RT tasks.
3999 * As is, the util number is not freq-invariant (we'd have to
4000 * implement arch_scale_freq_capacity() for that).
4002 * See cpu_util_cfs().
4004 cpufreq_update_util(rq, flags);
4009 static inline bool load_avg_is_decayed(struct sched_avg *sa)
4017 if (sa->runnable_sum)
4021 * _avg must be null when _sum are null because _avg = _sum / divider
4022 * Make sure that rounding and/or propagation of PELT values never
4025 SCHED_WARN_ON(sa->load_avg ||
4032 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
4034 return u64_u32_load_copy(cfs_rq->avg.last_update_time,
4035 cfs_rq->last_update_time_copy);
4037 #ifdef CONFIG_FAIR_GROUP_SCHED
4039 * Because list_add_leaf_cfs_rq always places a child cfs_rq on the list
4040 * immediately before a parent cfs_rq, and cfs_rqs are removed from the list
4041 * bottom-up, we only have to test whether the cfs_rq before us on the list
4043 * If cfs_rq is not on the list, test whether a child needs its to be added to
4044 * connect a branch to the tree * (see list_add_leaf_cfs_rq() for details).
4046 static inline bool child_cfs_rq_on_list(struct cfs_rq *cfs_rq)
4048 struct cfs_rq *prev_cfs_rq;
4049 struct list_head *prev;
4051 if (cfs_rq->on_list) {
4052 prev = cfs_rq->leaf_cfs_rq_list.prev;
4054 struct rq *rq = rq_of(cfs_rq);
4056 prev = rq->tmp_alone_branch;
4059 prev_cfs_rq = container_of(prev, struct cfs_rq, leaf_cfs_rq_list);
4061 return (prev_cfs_rq->tg->parent == cfs_rq->tg);
4064 static inline bool cfs_rq_is_decayed(struct cfs_rq *cfs_rq)
4066 if (cfs_rq->load.weight)
4069 if (!load_avg_is_decayed(&cfs_rq->avg))
4072 if (child_cfs_rq_on_list(cfs_rq))
4079 * update_tg_load_avg - update the tg's load avg
4080 * @cfs_rq: the cfs_rq whose avg changed
4082 * This function 'ensures': tg->load_avg := \Sum tg->cfs_rq[]->avg.load.
4083 * However, because tg->load_avg is a global value there are performance
4086 * In order to avoid having to look at the other cfs_rq's, we use a
4087 * differential update where we store the last value we propagated. This in
4088 * turn allows skipping updates if the differential is 'small'.
4090 * Updating tg's load_avg is necessary before update_cfs_share().
4092 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq)
4098 * No need to update load_avg for root_task_group as it is not used.
4100 if (cfs_rq->tg == &root_task_group)
4103 /* rq has been offline and doesn't contribute to the share anymore: */
4104 if (!cpu_active(cpu_of(rq_of(cfs_rq))))
4108 * For migration heavy workloads, access to tg->load_avg can be
4109 * unbound. Limit the update rate to at most once per ms.
4111 now = sched_clock_cpu(cpu_of(rq_of(cfs_rq)));
4112 if (now - cfs_rq->last_update_tg_load_avg < NSEC_PER_MSEC)
4115 delta = cfs_rq->avg.load_avg - cfs_rq->tg_load_avg_contrib;
4116 if (abs(delta) > cfs_rq->tg_load_avg_contrib / 64) {
4117 atomic_long_add(delta, &cfs_rq->tg->load_avg);
4118 cfs_rq->tg_load_avg_contrib = cfs_rq->avg.load_avg;
4119 cfs_rq->last_update_tg_load_avg = now;
4123 static inline void clear_tg_load_avg(struct cfs_rq *cfs_rq)
4129 * No need to update load_avg for root_task_group, as it is not used.
4131 if (cfs_rq->tg == &root_task_group)
4134 now = sched_clock_cpu(cpu_of(rq_of(cfs_rq)));
4135 delta = 0 - cfs_rq->tg_load_avg_contrib;
4136 atomic_long_add(delta, &cfs_rq->tg->load_avg);
4137 cfs_rq->tg_load_avg_contrib = 0;
4138 cfs_rq->last_update_tg_load_avg = now;
4141 /* CPU offline callback: */
4142 static void __maybe_unused clear_tg_offline_cfs_rqs(struct rq *rq)
4144 struct task_group *tg;
4146 lockdep_assert_rq_held(rq);
4149 * The rq clock has already been updated in
4150 * set_rq_offline(), so we should skip updating
4151 * the rq clock again in unthrottle_cfs_rq().
4153 rq_clock_start_loop_update(rq);
4156 list_for_each_entry_rcu(tg, &task_groups, list) {
4157 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
4159 clear_tg_load_avg(cfs_rq);
4163 rq_clock_stop_loop_update(rq);
4167 * Called within set_task_rq() right before setting a task's CPU. The
4168 * caller only guarantees p->pi_lock is held; no other assumptions,
4169 * including the state of rq->lock, should be made.
4171 void set_task_rq_fair(struct sched_entity *se,
4172 struct cfs_rq *prev, struct cfs_rq *next)
4174 u64 p_last_update_time;
4175 u64 n_last_update_time;
4177 if (!sched_feat(ATTACH_AGE_LOAD))
4181 * We are supposed to update the task to "current" time, then its up to
4182 * date and ready to go to new CPU/cfs_rq. But we have difficulty in
4183 * getting what current time is, so simply throw away the out-of-date
4184 * time. This will result in the wakee task is less decayed, but giving
4185 * the wakee more load sounds not bad.
4187 if (!(se->avg.last_update_time && prev))
4190 p_last_update_time = cfs_rq_last_update_time(prev);
4191 n_last_update_time = cfs_rq_last_update_time(next);
4193 __update_load_avg_blocked_se(p_last_update_time, se);
4194 se->avg.last_update_time = n_last_update_time;
4198 * When on migration a sched_entity joins/leaves the PELT hierarchy, we need to
4199 * propagate its contribution. The key to this propagation is the invariant
4200 * that for each group:
4202 * ge->avg == grq->avg (1)
4204 * _IFF_ we look at the pure running and runnable sums. Because they
4205 * represent the very same entity, just at different points in the hierarchy.
4207 * Per the above update_tg_cfs_util() and update_tg_cfs_runnable() are trivial
4208 * and simply copies the running/runnable sum over (but still wrong, because
4209 * the group entity and group rq do not have their PELT windows aligned).
4211 * However, update_tg_cfs_load() is more complex. So we have:
4213 * ge->avg.load_avg = ge->load.weight * ge->avg.runnable_avg (2)
4215 * And since, like util, the runnable part should be directly transferable,
4216 * the following would _appear_ to be the straight forward approach:
4218 * grq->avg.load_avg = grq->load.weight * grq->avg.runnable_avg (3)
4220 * And per (1) we have:
4222 * ge->avg.runnable_avg == grq->avg.runnable_avg
4226 * ge->load.weight * grq->avg.load_avg
4227 * ge->avg.load_avg = ----------------------------------- (4)
4230 * Except that is wrong!
4232 * Because while for entities historical weight is not important and we
4233 * really only care about our future and therefore can consider a pure
4234 * runnable sum, runqueues can NOT do this.
4236 * We specifically want runqueues to have a load_avg that includes
4237 * historical weights. Those represent the blocked load, the load we expect
4238 * to (shortly) return to us. This only works by keeping the weights as
4239 * integral part of the sum. We therefore cannot decompose as per (3).
4241 * Another reason this doesn't work is that runnable isn't a 0-sum entity.
4242 * Imagine a rq with 2 tasks that each are runnable 2/3 of the time. Then the
4243 * rq itself is runnable anywhere between 2/3 and 1 depending on how the
4244 * runnable section of these tasks overlap (or not). If they were to perfectly
4245 * align the rq as a whole would be runnable 2/3 of the time. If however we
4246 * always have at least 1 runnable task, the rq as a whole is always runnable.
4248 * So we'll have to approximate.. :/
4250 * Given the constraint:
4252 * ge->avg.running_sum <= ge->avg.runnable_sum <= LOAD_AVG_MAX
4254 * We can construct a rule that adds runnable to a rq by assuming minimal
4257 * On removal, we'll assume each task is equally runnable; which yields:
4259 * grq->avg.runnable_sum = grq->avg.load_sum / grq->load.weight
4261 * XXX: only do this for the part of runnable > running ?
4265 update_tg_cfs_util(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq)
4267 long delta_sum, delta_avg = gcfs_rq->avg.util_avg - se->avg.util_avg;
4268 u32 new_sum, divider;
4270 /* Nothing to update */
4275 * cfs_rq->avg.period_contrib can be used for both cfs_rq and se.
4276 * See ___update_load_avg() for details.
4278 divider = get_pelt_divider(&cfs_rq->avg);
4281 /* Set new sched_entity's utilization */
4282 se->avg.util_avg = gcfs_rq->avg.util_avg;
4283 new_sum = se->avg.util_avg * divider;
4284 delta_sum = (long)new_sum - (long)se->avg.util_sum;
4285 se->avg.util_sum = new_sum;
4287 /* Update parent cfs_rq utilization */
4288 add_positive(&cfs_rq->avg.util_avg, delta_avg);
4289 add_positive(&cfs_rq->avg.util_sum, delta_sum);
4291 /* See update_cfs_rq_load_avg() */
4292 cfs_rq->avg.util_sum = max_t(u32, cfs_rq->avg.util_sum,
4293 cfs_rq->avg.util_avg * PELT_MIN_DIVIDER);
4297 update_tg_cfs_runnable(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq)
4299 long delta_sum, delta_avg = gcfs_rq->avg.runnable_avg - se->avg.runnable_avg;
4300 u32 new_sum, divider;
4302 /* Nothing to update */
4307 * cfs_rq->avg.period_contrib can be used for both cfs_rq and se.
4308 * See ___update_load_avg() for details.
4310 divider = get_pelt_divider(&cfs_rq->avg);
4312 /* Set new sched_entity's runnable */
4313 se->avg.runnable_avg = gcfs_rq->avg.runnable_avg;
4314 new_sum = se->avg.runnable_avg * divider;
4315 delta_sum = (long)new_sum - (long)se->avg.runnable_sum;
4316 se->avg.runnable_sum = new_sum;
4318 /* Update parent cfs_rq runnable */
4319 add_positive(&cfs_rq->avg.runnable_avg, delta_avg);
4320 add_positive(&cfs_rq->avg.runnable_sum, delta_sum);
4321 /* See update_cfs_rq_load_avg() */
4322 cfs_rq->avg.runnable_sum = max_t(u32, cfs_rq->avg.runnable_sum,
4323 cfs_rq->avg.runnable_avg * PELT_MIN_DIVIDER);
4327 update_tg_cfs_load(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq)
4329 long delta_avg, running_sum, runnable_sum = gcfs_rq->prop_runnable_sum;
4330 unsigned long load_avg;
4338 gcfs_rq->prop_runnable_sum = 0;
4341 * cfs_rq->avg.period_contrib can be used for both cfs_rq and se.
4342 * See ___update_load_avg() for details.
4344 divider = get_pelt_divider(&cfs_rq->avg);
4346 if (runnable_sum >= 0) {
4348 * Add runnable; clip at LOAD_AVG_MAX. Reflects that until
4349 * the CPU is saturated running == runnable.
4351 runnable_sum += se->avg.load_sum;
4352 runnable_sum = min_t(long, runnable_sum, divider);
4355 * Estimate the new unweighted runnable_sum of the gcfs_rq by
4356 * assuming all tasks are equally runnable.
4358 if (scale_load_down(gcfs_rq->load.weight)) {
4359 load_sum = div_u64(gcfs_rq->avg.load_sum,
4360 scale_load_down(gcfs_rq->load.weight));
4363 /* But make sure to not inflate se's runnable */
4364 runnable_sum = min(se->avg.load_sum, load_sum);
4368 * runnable_sum can't be lower than running_sum
4369 * Rescale running sum to be in the same range as runnable sum
4370 * running_sum is in [0 : LOAD_AVG_MAX << SCHED_CAPACITY_SHIFT]
4371 * runnable_sum is in [0 : LOAD_AVG_MAX]
4373 running_sum = se->avg.util_sum >> SCHED_CAPACITY_SHIFT;
4374 runnable_sum = max(runnable_sum, running_sum);
4376 load_sum = se_weight(se) * runnable_sum;
4377 load_avg = div_u64(load_sum, divider);
4379 delta_avg = load_avg - se->avg.load_avg;
4383 delta_sum = load_sum - (s64)se_weight(se) * se->avg.load_sum;
4385 se->avg.load_sum = runnable_sum;
4386 se->avg.load_avg = load_avg;
4387 add_positive(&cfs_rq->avg.load_avg, delta_avg);
4388 add_positive(&cfs_rq->avg.load_sum, delta_sum);
4389 /* See update_cfs_rq_load_avg() */
4390 cfs_rq->avg.load_sum = max_t(u32, cfs_rq->avg.load_sum,
4391 cfs_rq->avg.load_avg * PELT_MIN_DIVIDER);
4394 static inline void add_tg_cfs_propagate(struct cfs_rq *cfs_rq, long runnable_sum)
4396 cfs_rq->propagate = 1;
4397 cfs_rq->prop_runnable_sum += runnable_sum;
4400 /* Update task and its cfs_rq load average */
4401 static inline int propagate_entity_load_avg(struct sched_entity *se)
4403 struct cfs_rq *cfs_rq, *gcfs_rq;
4405 if (entity_is_task(se))
4408 gcfs_rq = group_cfs_rq(se);
4409 if (!gcfs_rq->propagate)
4412 gcfs_rq->propagate = 0;
4414 cfs_rq = cfs_rq_of(se);
4416 add_tg_cfs_propagate(cfs_rq, gcfs_rq->prop_runnable_sum);
4418 update_tg_cfs_util(cfs_rq, se, gcfs_rq);
4419 update_tg_cfs_runnable(cfs_rq, se, gcfs_rq);
4420 update_tg_cfs_load(cfs_rq, se, gcfs_rq);
4422 trace_pelt_cfs_tp(cfs_rq);
4423 trace_pelt_se_tp(se);
4429 * Check if we need to update the load and the utilization of a blocked
4432 static inline bool skip_blocked_update(struct sched_entity *se)
4434 struct cfs_rq *gcfs_rq = group_cfs_rq(se);
4437 * If sched_entity still have not zero load or utilization, we have to
4440 if (se->avg.load_avg || se->avg.util_avg)
4444 * If there is a pending propagation, we have to update the load and
4445 * the utilization of the sched_entity:
4447 if (gcfs_rq->propagate)
4451 * Otherwise, the load and the utilization of the sched_entity is
4452 * already zero and there is no pending propagation, so it will be a
4453 * waste of time to try to decay it:
4458 #else /* CONFIG_FAIR_GROUP_SCHED */
4460 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq) {}
4462 static inline void clear_tg_offline_cfs_rqs(struct rq *rq) {}
4464 static inline int propagate_entity_load_avg(struct sched_entity *se)
4469 static inline void add_tg_cfs_propagate(struct cfs_rq *cfs_rq, long runnable_sum) {}
4471 #endif /* CONFIG_FAIR_GROUP_SCHED */
4473 #ifdef CONFIG_NO_HZ_COMMON
4474 static inline void migrate_se_pelt_lag(struct sched_entity *se)
4476 u64 throttled = 0, now, lut;
4477 struct cfs_rq *cfs_rq;
4481 if (load_avg_is_decayed(&se->avg))
4484 cfs_rq = cfs_rq_of(se);
4488 is_idle = is_idle_task(rcu_dereference(rq->curr));
4492 * The lag estimation comes with a cost we don't want to pay all the
4493 * time. Hence, limiting to the case where the source CPU is idle and
4494 * we know we are at the greatest risk to have an outdated clock.
4500 * Estimated "now" is: last_update_time + cfs_idle_lag + rq_idle_lag, where:
4502 * last_update_time (the cfs_rq's last_update_time)
4503 * = cfs_rq_clock_pelt()@cfs_rq_idle
4504 * = rq_clock_pelt()@cfs_rq_idle
4505 * - cfs->throttled_clock_pelt_time@cfs_rq_idle
4507 * cfs_idle_lag (delta between rq's update and cfs_rq's update)
4508 * = rq_clock_pelt()@rq_idle - rq_clock_pelt()@cfs_rq_idle
4510 * rq_idle_lag (delta between now and rq's update)
4511 * = sched_clock_cpu() - rq_clock()@rq_idle
4513 * We can then write:
4515 * now = rq_clock_pelt()@rq_idle - cfs->throttled_clock_pelt_time +
4516 * sched_clock_cpu() - rq_clock()@rq_idle
4518 * rq_clock_pelt()@rq_idle is rq->clock_pelt_idle
4519 * rq_clock()@rq_idle is rq->clock_idle
4520 * cfs->throttled_clock_pelt_time@cfs_rq_idle
4521 * is cfs_rq->throttled_pelt_idle
4524 #ifdef CONFIG_CFS_BANDWIDTH
4525 throttled = u64_u32_load(cfs_rq->throttled_pelt_idle);
4526 /* The clock has been stopped for throttling */
4527 if (throttled == U64_MAX)
4530 now = u64_u32_load(rq->clock_pelt_idle);
4532 * Paired with _update_idle_rq_clock_pelt(). It ensures at the worst case
4533 * is observed the old clock_pelt_idle value and the new clock_idle,
4534 * which lead to an underestimation. The opposite would lead to an
4538 lut = cfs_rq_last_update_time(cfs_rq);
4543 * cfs_rq->avg.last_update_time is more recent than our
4544 * estimation, let's use it.
4548 now += sched_clock_cpu(cpu_of(rq)) - u64_u32_load(rq->clock_idle);
4550 __update_load_avg_blocked_se(now, se);
4553 static void migrate_se_pelt_lag(struct sched_entity *se) {}
4557 * update_cfs_rq_load_avg - update the cfs_rq's load/util averages
4558 * @now: current time, as per cfs_rq_clock_pelt()
4559 * @cfs_rq: cfs_rq to update
4561 * The cfs_rq avg is the direct sum of all its entities (blocked and runnable)
4562 * avg. The immediate corollary is that all (fair) tasks must be attached.
4564 * cfs_rq->avg is used for task_h_load() and update_cfs_share() for example.
4566 * Return: true if the load decayed or we removed load.
4568 * Since both these conditions indicate a changed cfs_rq->avg.load we should
4569 * call update_tg_load_avg() when this function returns true.
4572 update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq)
4574 unsigned long removed_load = 0, removed_util = 0, removed_runnable = 0;
4575 struct sched_avg *sa = &cfs_rq->avg;
4578 if (cfs_rq->removed.nr) {
4580 u32 divider = get_pelt_divider(&cfs_rq->avg);
4582 raw_spin_lock(&cfs_rq->removed.lock);
4583 swap(cfs_rq->removed.util_avg, removed_util);
4584 swap(cfs_rq->removed.load_avg, removed_load);
4585 swap(cfs_rq->removed.runnable_avg, removed_runnable);
4586 cfs_rq->removed.nr = 0;
4587 raw_spin_unlock(&cfs_rq->removed.lock);
4590 sub_positive(&sa->load_avg, r);
4591 sub_positive(&sa->load_sum, r * divider);
4592 /* See sa->util_sum below */
4593 sa->load_sum = max_t(u32, sa->load_sum, sa->load_avg * PELT_MIN_DIVIDER);
4596 sub_positive(&sa->util_avg, r);
4597 sub_positive(&sa->util_sum, r * divider);
4599 * Because of rounding, se->util_sum might ends up being +1 more than
4600 * cfs->util_sum. Although this is not a problem by itself, detaching
4601 * a lot of tasks with the rounding problem between 2 updates of
4602 * util_avg (~1ms) can make cfs->util_sum becoming null whereas
4603 * cfs_util_avg is not.
4604 * Check that util_sum is still above its lower bound for the new
4605 * util_avg. Given that period_contrib might have moved since the last
4606 * sync, we are only sure that util_sum must be above or equal to
4607 * util_avg * minimum possible divider
4609 sa->util_sum = max_t(u32, sa->util_sum, sa->util_avg * PELT_MIN_DIVIDER);
4611 r = removed_runnable;
4612 sub_positive(&sa->runnable_avg, r);
4613 sub_positive(&sa->runnable_sum, r * divider);
4614 /* See sa->util_sum above */
4615 sa->runnable_sum = max_t(u32, sa->runnable_sum,
4616 sa->runnable_avg * PELT_MIN_DIVIDER);
4619 * removed_runnable is the unweighted version of removed_load so we
4620 * can use it to estimate removed_load_sum.
4622 add_tg_cfs_propagate(cfs_rq,
4623 -(long)(removed_runnable * divider) >> SCHED_CAPACITY_SHIFT);
4628 decayed |= __update_load_avg_cfs_rq(now, cfs_rq);
4629 u64_u32_store_copy(sa->last_update_time,
4630 cfs_rq->last_update_time_copy,
4631 sa->last_update_time);
4636 * attach_entity_load_avg - attach this entity to its cfs_rq load avg
4637 * @cfs_rq: cfs_rq to attach to
4638 * @se: sched_entity to attach
4640 * Must call update_cfs_rq_load_avg() before this, since we rely on
4641 * cfs_rq->avg.last_update_time being current.
4643 static void attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
4646 * cfs_rq->avg.period_contrib can be used for both cfs_rq and se.
4647 * See ___update_load_avg() for details.
4649 u32 divider = get_pelt_divider(&cfs_rq->avg);
4652 * When we attach the @se to the @cfs_rq, we must align the decay
4653 * window because without that, really weird and wonderful things can
4658 se->avg.last_update_time = cfs_rq->avg.last_update_time;
4659 se->avg.period_contrib = cfs_rq->avg.period_contrib;
4662 * Hell(o) Nasty stuff.. we need to recompute _sum based on the new
4663 * period_contrib. This isn't strictly correct, but since we're
4664 * entirely outside of the PELT hierarchy, nobody cares if we truncate
4667 se->avg.util_sum = se->avg.util_avg * divider;
4669 se->avg.runnable_sum = se->avg.runnable_avg * divider;
4671 se->avg.load_sum = se->avg.load_avg * divider;
4672 if (se_weight(se) < se->avg.load_sum)
4673 se->avg.load_sum = div_u64(se->avg.load_sum, se_weight(se));
4675 se->avg.load_sum = 1;
4677 enqueue_load_avg(cfs_rq, se);
4678 cfs_rq->avg.util_avg += se->avg.util_avg;
4679 cfs_rq->avg.util_sum += se->avg.util_sum;
4680 cfs_rq->avg.runnable_avg += se->avg.runnable_avg;
4681 cfs_rq->avg.runnable_sum += se->avg.runnable_sum;
4683 add_tg_cfs_propagate(cfs_rq, se->avg.load_sum);
4685 cfs_rq_util_change(cfs_rq, 0);
4687 trace_pelt_cfs_tp(cfs_rq);
4691 * detach_entity_load_avg - detach this entity from its cfs_rq load avg
4692 * @cfs_rq: cfs_rq to detach from
4693 * @se: sched_entity to detach
4695 * Must call update_cfs_rq_load_avg() before this, since we rely on
4696 * cfs_rq->avg.last_update_time being current.
4698 static void detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
4700 dequeue_load_avg(cfs_rq, se);
4701 sub_positive(&cfs_rq->avg.util_avg, se->avg.util_avg);
4702 sub_positive(&cfs_rq->avg.util_sum, se->avg.util_sum);
4703 /* See update_cfs_rq_load_avg() */
4704 cfs_rq->avg.util_sum = max_t(u32, cfs_rq->avg.util_sum,
4705 cfs_rq->avg.util_avg * PELT_MIN_DIVIDER);
4707 sub_positive(&cfs_rq->avg.runnable_avg, se->avg.runnable_avg);
4708 sub_positive(&cfs_rq->avg.runnable_sum, se->avg.runnable_sum);
4709 /* See update_cfs_rq_load_avg() */
4710 cfs_rq->avg.runnable_sum = max_t(u32, cfs_rq->avg.runnable_sum,
4711 cfs_rq->avg.runnable_avg * PELT_MIN_DIVIDER);
4713 add_tg_cfs_propagate(cfs_rq, -se->avg.load_sum);
4715 cfs_rq_util_change(cfs_rq, 0);
4717 trace_pelt_cfs_tp(cfs_rq);
4721 * Optional action to be done while updating the load average
4723 #define UPDATE_TG 0x1
4724 #define SKIP_AGE_LOAD 0x2
4725 #define DO_ATTACH 0x4
4726 #define DO_DETACH 0x8
4728 /* Update task and its cfs_rq load average */
4729 static inline void update_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
4731 u64 now = cfs_rq_clock_pelt(cfs_rq);
4735 * Track task load average for carrying it to new CPU after migrated, and
4736 * track group sched_entity load average for task_h_load calc in migration
4738 if (se->avg.last_update_time && !(flags & SKIP_AGE_LOAD))
4739 __update_load_avg_se(now, cfs_rq, se);
4741 decayed = update_cfs_rq_load_avg(now, cfs_rq);
4742 decayed |= propagate_entity_load_avg(se);
4744 if (!se->avg.last_update_time && (flags & DO_ATTACH)) {
4747 * DO_ATTACH means we're here from enqueue_entity().
4748 * !last_update_time means we've passed through
4749 * migrate_task_rq_fair() indicating we migrated.
4751 * IOW we're enqueueing a task on a new CPU.
4753 attach_entity_load_avg(cfs_rq, se);
4754 update_tg_load_avg(cfs_rq);
4756 } else if (flags & DO_DETACH) {
4758 * DO_DETACH means we're here from dequeue_entity()
4759 * and we are migrating task out of the CPU.
4761 detach_entity_load_avg(cfs_rq, se);
4762 update_tg_load_avg(cfs_rq);
4763 } else if (decayed) {
4764 cfs_rq_util_change(cfs_rq, 0);
4766 if (flags & UPDATE_TG)
4767 update_tg_load_avg(cfs_rq);
4772 * Synchronize entity load avg of dequeued entity without locking
4775 static void sync_entity_load_avg(struct sched_entity *se)
4777 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4778 u64 last_update_time;
4780 last_update_time = cfs_rq_last_update_time(cfs_rq);
4781 __update_load_avg_blocked_se(last_update_time, se);
4785 * Task first catches up with cfs_rq, and then subtract
4786 * itself from the cfs_rq (task must be off the queue now).
4788 static void remove_entity_load_avg(struct sched_entity *se)
4790 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4791 unsigned long flags;
4794 * tasks cannot exit without having gone through wake_up_new_task() ->
4795 * enqueue_task_fair() which will have added things to the cfs_rq,
4796 * so we can remove unconditionally.
4799 sync_entity_load_avg(se);
4801 raw_spin_lock_irqsave(&cfs_rq->removed.lock, flags);
4802 ++cfs_rq->removed.nr;
4803 cfs_rq->removed.util_avg += se->avg.util_avg;
4804 cfs_rq->removed.load_avg += se->avg.load_avg;
4805 cfs_rq->removed.runnable_avg += se->avg.runnable_avg;
4806 raw_spin_unlock_irqrestore(&cfs_rq->removed.lock, flags);
4809 static inline unsigned long cfs_rq_runnable_avg(struct cfs_rq *cfs_rq)
4811 return cfs_rq->avg.runnable_avg;
4814 static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq)
4816 return cfs_rq->avg.load_avg;
4819 static int newidle_balance(struct rq *this_rq, struct rq_flags *rf);
4821 static inline unsigned long task_util(struct task_struct *p)
4823 return READ_ONCE(p->se.avg.util_avg);
4826 static inline unsigned long task_runnable(struct task_struct *p)
4828 return READ_ONCE(p->se.avg.runnable_avg);
4831 static inline unsigned long _task_util_est(struct task_struct *p)
4833 return READ_ONCE(p->se.avg.util_est) & ~UTIL_AVG_UNCHANGED;
4836 static inline unsigned long task_util_est(struct task_struct *p)
4838 return max(task_util(p), _task_util_est(p));
4841 static inline void util_est_enqueue(struct cfs_rq *cfs_rq,
4842 struct task_struct *p)
4844 unsigned int enqueued;
4846 if (!sched_feat(UTIL_EST))
4849 /* Update root cfs_rq's estimated utilization */
4850 enqueued = cfs_rq->avg.util_est;
4851 enqueued += _task_util_est(p);
4852 WRITE_ONCE(cfs_rq->avg.util_est, enqueued);
4854 trace_sched_util_est_cfs_tp(cfs_rq);
4857 static inline void util_est_dequeue(struct cfs_rq *cfs_rq,
4858 struct task_struct *p)
4860 unsigned int enqueued;
4862 if (!sched_feat(UTIL_EST))
4865 /* Update root cfs_rq's estimated utilization */
4866 enqueued = cfs_rq->avg.util_est;
4867 enqueued -= min_t(unsigned int, enqueued, _task_util_est(p));
4868 WRITE_ONCE(cfs_rq->avg.util_est, enqueued);
4870 trace_sched_util_est_cfs_tp(cfs_rq);
4873 #define UTIL_EST_MARGIN (SCHED_CAPACITY_SCALE / 100)
4875 static inline void util_est_update(struct cfs_rq *cfs_rq,
4876 struct task_struct *p,
4879 unsigned int ewma, dequeued, last_ewma_diff;
4881 if (!sched_feat(UTIL_EST))
4885 * Skip update of task's estimated utilization when the task has not
4886 * yet completed an activation, e.g. being migrated.
4891 /* Get current estimate of utilization */
4892 ewma = READ_ONCE(p->se.avg.util_est);
4895 * If the PELT values haven't changed since enqueue time,
4896 * skip the util_est update.
4898 if (ewma & UTIL_AVG_UNCHANGED)
4901 /* Get utilization at dequeue */
4902 dequeued = task_util(p);
4905 * Reset EWMA on utilization increases, the moving average is used only
4906 * to smooth utilization decreases.
4908 if (ewma <= dequeued) {
4914 * Skip update of task's estimated utilization when its members are
4915 * already ~1% close to its last activation value.
4917 last_ewma_diff = ewma - dequeued;
4918 if (last_ewma_diff < UTIL_EST_MARGIN)
4922 * To avoid overestimation of actual task utilization, skip updates if
4923 * we cannot grant there is idle time in this CPU.
4925 if (dequeued > arch_scale_cpu_capacity(cpu_of(rq_of(cfs_rq))))
4929 * To avoid underestimate of task utilization, skip updates of EWMA if
4930 * we cannot grant that thread got all CPU time it wanted.
4932 if ((dequeued + UTIL_EST_MARGIN) < task_runnable(p))
4937 * Update Task's estimated utilization
4939 * When *p completes an activation we can consolidate another sample
4940 * of the task size. This is done by using this value to update the
4941 * Exponential Weighted Moving Average (EWMA):
4943 * ewma(t) = w * task_util(p) + (1-w) * ewma(t-1)
4944 * = w * task_util(p) + ewma(t-1) - w * ewma(t-1)
4945 * = w * (task_util(p) - ewma(t-1)) + ewma(t-1)
4946 * = w * ( -last_ewma_diff ) + ewma(t-1)
4947 * = w * (-last_ewma_diff + ewma(t-1) / w)
4949 * Where 'w' is the weight of new samples, which is configured to be
4950 * 0.25, thus making w=1/4 ( >>= UTIL_EST_WEIGHT_SHIFT)
4952 ewma <<= UTIL_EST_WEIGHT_SHIFT;
4953 ewma -= last_ewma_diff;
4954 ewma >>= UTIL_EST_WEIGHT_SHIFT;
4956 ewma |= UTIL_AVG_UNCHANGED;
4957 WRITE_ONCE(p->se.avg.util_est, ewma);
4959 trace_sched_util_est_se_tp(&p->se);
4962 static inline int util_fits_cpu(unsigned long util,
4963 unsigned long uclamp_min,
4964 unsigned long uclamp_max,
4967 unsigned long capacity_orig, capacity_orig_thermal;
4968 unsigned long capacity = capacity_of(cpu);
4969 bool fits, uclamp_max_fits;
4972 * Check if the real util fits without any uclamp boost/cap applied.
4974 fits = fits_capacity(util, capacity);
4976 if (!uclamp_is_used())
4980 * We must use arch_scale_cpu_capacity() for comparing against uclamp_min and
4981 * uclamp_max. We only care about capacity pressure (by using
4982 * capacity_of()) for comparing against the real util.
4984 * If a task is boosted to 1024 for example, we don't want a tiny
4985 * pressure to skew the check whether it fits a CPU or not.
4987 * Similarly if a task is capped to arch_scale_cpu_capacity(little_cpu), it
4988 * should fit a little cpu even if there's some pressure.
4990 * Only exception is for thermal pressure since it has a direct impact
4991 * on available OPP of the system.
4993 * We honour it for uclamp_min only as a drop in performance level
4994 * could result in not getting the requested minimum performance level.
4996 * For uclamp_max, we can tolerate a drop in performance level as the
4997 * goal is to cap the task. So it's okay if it's getting less.
4999 capacity_orig = arch_scale_cpu_capacity(cpu);
5000 capacity_orig_thermal = capacity_orig - arch_scale_thermal_pressure(cpu);
5003 * We want to force a task to fit a cpu as implied by uclamp_max.
5004 * But we do have some corner cases to cater for..
5010 * |_ _ _ _ _ _ _ _ _ ___ _ _ _ | _ | _ _ _ _ _ uclamp_max
5013 * | | | | | | | (util somewhere in this region)
5016 * +----------------------------------------
5019 * In the above example if a task is capped to a specific performance
5020 * point, y, then when:
5022 * * util = 80% of x then it does not fit on cpu0 and should migrate
5024 * * util = 80% of y then it is forced to fit on cpu1 to honour
5025 * uclamp_max request.
5027 * which is what we're enforcing here. A task always fits if
5028 * uclamp_max <= capacity_orig. But when uclamp_max > capacity_orig,
5029 * the normal upmigration rules should withhold still.
5031 * Only exception is when we are on max capacity, then we need to be
5032 * careful not to block overutilized state. This is so because:
5034 * 1. There's no concept of capping at max_capacity! We can't go
5035 * beyond this performance level anyway.
5036 * 2. The system is being saturated when we're operating near
5037 * max capacity, it doesn't make sense to block overutilized.
5039 uclamp_max_fits = (capacity_orig == SCHED_CAPACITY_SCALE) && (uclamp_max == SCHED_CAPACITY_SCALE);
5040 uclamp_max_fits = !uclamp_max_fits && (uclamp_max <= capacity_orig);
5041 fits = fits || uclamp_max_fits;
5046 * | ___ (region a, capped, util >= uclamp_max)
5048 * |_ _ _ _ _ _ _ _ _ ___ _ _ _ | _ | _ _ _ _ _ uclamp_max
5050 * | ___ | | | | (region b, uclamp_min <= util <= uclamp_max)
5051 * |_ _ _|_ _|_ _ _ _| _ | _ _ _| _ | _ _ _ _ _ uclamp_min
5053 * | | | | | | | (region c, boosted, util < uclamp_min)
5054 * +----------------------------------------
5057 * a) If util > uclamp_max, then we're capped, we don't care about
5058 * actual fitness value here. We only care if uclamp_max fits
5059 * capacity without taking margin/pressure into account.
5060 * See comment above.
5062 * b) If uclamp_min <= util <= uclamp_max, then the normal
5063 * fits_capacity() rules apply. Except we need to ensure that we
5064 * enforce we remain within uclamp_max, see comment above.
5066 * c) If util < uclamp_min, then we are boosted. Same as (b) but we
5067 * need to take into account the boosted value fits the CPU without
5068 * taking margin/pressure into account.
5070 * Cases (a) and (b) are handled in the 'fits' variable already. We
5071 * just need to consider an extra check for case (c) after ensuring we
5072 * handle the case uclamp_min > uclamp_max.
5074 uclamp_min = min(uclamp_min, uclamp_max);
5075 if (fits && (util < uclamp_min) && (uclamp_min > capacity_orig_thermal))
5081 static inline int task_fits_cpu(struct task_struct *p, int cpu)
5083 unsigned long uclamp_min = uclamp_eff_value(p, UCLAMP_MIN);
5084 unsigned long uclamp_max = uclamp_eff_value(p, UCLAMP_MAX);
5085 unsigned long util = task_util_est(p);
5087 * Return true only if the cpu fully fits the task requirements, which
5088 * include the utilization but also the performance hints.
5090 return (util_fits_cpu(util, uclamp_min, uclamp_max, cpu) > 0);
5093 static inline void update_misfit_status(struct task_struct *p, struct rq *rq)
5095 if (!sched_asym_cpucap_active())
5098 if (!p || p->nr_cpus_allowed == 1) {
5099 rq->misfit_task_load = 0;
5103 if (task_fits_cpu(p, cpu_of(rq))) {
5104 rq->misfit_task_load = 0;
5109 * Make sure that misfit_task_load will not be null even if
5110 * task_h_load() returns 0.
5112 rq->misfit_task_load = max_t(unsigned long, task_h_load(p), 1);
5115 #else /* CONFIG_SMP */
5117 static inline bool cfs_rq_is_decayed(struct cfs_rq *cfs_rq)
5119 return !cfs_rq->nr_running;
5122 #define UPDATE_TG 0x0
5123 #define SKIP_AGE_LOAD 0x0
5124 #define DO_ATTACH 0x0
5125 #define DO_DETACH 0x0
5127 static inline void update_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int not_used1)
5129 cfs_rq_util_change(cfs_rq, 0);
5132 static inline void remove_entity_load_avg(struct sched_entity *se) {}
5135 attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
5137 detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
5139 static inline int newidle_balance(struct rq *rq, struct rq_flags *rf)
5145 util_est_enqueue(struct cfs_rq *cfs_rq, struct task_struct *p) {}
5148 util_est_dequeue(struct cfs_rq *cfs_rq, struct task_struct *p) {}
5151 util_est_update(struct cfs_rq *cfs_rq, struct task_struct *p,
5153 static inline void update_misfit_status(struct task_struct *p, struct rq *rq) {}
5155 #endif /* CONFIG_SMP */
5158 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
5160 u64 vslice, vruntime = avg_vruntime(cfs_rq);
5163 se->slice = sysctl_sched_base_slice;
5164 vslice = calc_delta_fair(se->slice, se);
5167 * Due to how V is constructed as the weighted average of entities,
5168 * adding tasks with positive lag, or removing tasks with negative lag
5169 * will move 'time' backwards, this can screw around with the lag of
5172 * EEVDF: placement strategy #1 / #2
5174 if (sched_feat(PLACE_LAG) && cfs_rq->nr_running) {
5175 struct sched_entity *curr = cfs_rq->curr;
5181 * If we want to place a task and preserve lag, we have to
5182 * consider the effect of the new entity on the weighted
5183 * average and compensate for this, otherwise lag can quickly
5186 * Lag is defined as:
5188 * lag_i = S - s_i = w_i * (V - v_i)
5190 * To avoid the 'w_i' term all over the place, we only track
5193 * vl_i = V - v_i <=> v_i = V - vl_i
5195 * And we take V to be the weighted average of all v:
5197 * V = (\Sum w_j*v_j) / W
5199 * Where W is: \Sum w_j
5201 * Then, the weighted average after adding an entity with lag
5204 * V' = (\Sum w_j*v_j + w_i*v_i) / (W + w_i)
5205 * = (W*V + w_i*(V - vl_i)) / (W + w_i)
5206 * = (W*V + w_i*V - w_i*vl_i) / (W + w_i)
5207 * = (V*(W + w_i) - w_i*l) / (W + w_i)
5208 * = V - w_i*vl_i / (W + w_i)
5210 * And the actual lag after adding an entity with vl_i is:
5213 * = V - w_i*vl_i / (W + w_i) - (V - vl_i)
5214 * = vl_i - w_i*vl_i / (W + w_i)
5216 * Which is strictly less than vl_i. So in order to preserve lag
5217 * we should inflate the lag before placement such that the
5218 * effective lag after placement comes out right.
5220 * As such, invert the above relation for vl'_i to get the vl_i
5221 * we need to use such that the lag after placement is the lag
5222 * we computed before dequeue.
5224 * vl'_i = vl_i - w_i*vl_i / (W + w_i)
5225 * = ((W + w_i)*vl_i - w_i*vl_i) / (W + w_i)
5227 * (W + w_i)*vl'_i = (W + w_i)*vl_i - w_i*vl_i
5230 * vl_i = (W + w_i)*vl'_i / W
5232 load = cfs_rq->avg_load;
5233 if (curr && curr->on_rq)
5234 load += scale_load_down(curr->load.weight);
5236 lag *= load + scale_load_down(se->load.weight);
5237 if (WARN_ON_ONCE(!load))
5239 lag = div_s64(lag, load);
5242 se->vruntime = vruntime - lag;
5245 * When joining the competition; the exisiting tasks will be,
5246 * on average, halfway through their slice, as such start tasks
5247 * off with half a slice to ease into the competition.
5249 if (sched_feat(PLACE_DEADLINE_INITIAL) && (flags & ENQUEUE_INITIAL))
5253 * EEVDF: vd_i = ve_i + r_i/w_i
5255 se->deadline = se->vruntime + vslice;
5258 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
5259 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq);
5261 static inline bool cfs_bandwidth_used(void);
5264 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
5266 bool curr = cfs_rq->curr == se;
5269 * If we're the current task, we must renormalise before calling
5273 place_entity(cfs_rq, se, flags);
5275 update_curr(cfs_rq);
5278 * When enqueuing a sched_entity, we must:
5279 * - Update loads to have both entity and cfs_rq synced with now.
5280 * - For group_entity, update its runnable_weight to reflect the new
5281 * h_nr_running of its group cfs_rq.
5282 * - For group_entity, update its weight to reflect the new share of
5284 * - Add its new weight to cfs_rq->load.weight
5286 update_load_avg(cfs_rq, se, UPDATE_TG | DO_ATTACH);
5287 se_update_runnable(se);
5289 * XXX update_load_avg() above will have attached us to the pelt sum;
5290 * but update_cfs_group() here will re-adjust the weight and have to
5291 * undo/redo all that. Seems wasteful.
5293 update_cfs_group(se);
5296 * XXX now that the entity has been re-weighted, and it's lag adjusted,
5297 * we can place the entity.
5300 place_entity(cfs_rq, se, flags);
5302 account_entity_enqueue(cfs_rq, se);
5304 /* Entity has migrated, no longer consider this task hot */
5305 if (flags & ENQUEUE_MIGRATED)
5308 check_schedstat_required();
5309 update_stats_enqueue_fair(cfs_rq, se, flags);
5311 __enqueue_entity(cfs_rq, se);
5314 if (cfs_rq->nr_running == 1) {
5315 check_enqueue_throttle(cfs_rq);
5316 if (!throttled_hierarchy(cfs_rq)) {
5317 list_add_leaf_cfs_rq(cfs_rq);
5319 #ifdef CONFIG_CFS_BANDWIDTH
5320 struct rq *rq = rq_of(cfs_rq);
5322 if (cfs_rq_throttled(cfs_rq) && !cfs_rq->throttled_clock)
5323 cfs_rq->throttled_clock = rq_clock(rq);
5324 if (!cfs_rq->throttled_clock_self)
5325 cfs_rq->throttled_clock_self = rq_clock(rq);
5331 static void __clear_buddies_next(struct sched_entity *se)
5333 for_each_sched_entity(se) {
5334 struct cfs_rq *cfs_rq = cfs_rq_of(se);
5335 if (cfs_rq->next != se)
5338 cfs_rq->next = NULL;
5342 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
5344 if (cfs_rq->next == se)
5345 __clear_buddies_next(se);
5348 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
5351 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
5353 int action = UPDATE_TG;
5355 if (entity_is_task(se) && task_on_rq_migrating(task_of(se)))
5356 action |= DO_DETACH;
5359 * Update run-time statistics of the 'current'.
5361 update_curr(cfs_rq);
5364 * When dequeuing a sched_entity, we must:
5365 * - Update loads to have both entity and cfs_rq synced with now.
5366 * - For group_entity, update its runnable_weight to reflect the new
5367 * h_nr_running of its group cfs_rq.
5368 * - Subtract its previous weight from cfs_rq->load.weight.
5369 * - For group entity, update its weight to reflect the new share
5370 * of its group cfs_rq.
5372 update_load_avg(cfs_rq, se, action);
5373 se_update_runnable(se);
5375 update_stats_dequeue_fair(cfs_rq, se, flags);
5377 clear_buddies(cfs_rq, se);
5379 update_entity_lag(cfs_rq, se);
5380 if (se != cfs_rq->curr)
5381 __dequeue_entity(cfs_rq, se);
5383 account_entity_dequeue(cfs_rq, se);
5385 /* return excess runtime on last dequeue */
5386 return_cfs_rq_runtime(cfs_rq);
5388 update_cfs_group(se);
5391 * Now advance min_vruntime if @se was the entity holding it back,
5392 * except when: DEQUEUE_SAVE && !DEQUEUE_MOVE, in this case we'll be
5393 * put back on, and if we advance min_vruntime, we'll be placed back
5394 * further than we started -- ie. we'll be penalized.
5396 if ((flags & (DEQUEUE_SAVE | DEQUEUE_MOVE)) != DEQUEUE_SAVE)
5397 update_min_vruntime(cfs_rq);
5399 if (cfs_rq->nr_running == 0)
5400 update_idle_cfs_rq_clock_pelt(cfs_rq);
5404 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
5406 clear_buddies(cfs_rq, se);
5408 /* 'current' is not kept within the tree. */
5411 * Any task has to be enqueued before it get to execute on
5412 * a CPU. So account for the time it spent waiting on the
5415 update_stats_wait_end_fair(cfs_rq, se);
5416 __dequeue_entity(cfs_rq, se);
5417 update_load_avg(cfs_rq, se, UPDATE_TG);
5419 * HACK, stash a copy of deadline at the point of pick in vlag,
5420 * which isn't used until dequeue.
5422 se->vlag = se->deadline;
5425 update_stats_curr_start(cfs_rq, se);
5429 * Track our maximum slice length, if the CPU's load is at
5430 * least twice that of our own weight (i.e. dont track it
5431 * when there are only lesser-weight tasks around):
5433 if (schedstat_enabled() &&
5434 rq_of(cfs_rq)->cfs.load.weight >= 2*se->load.weight) {
5435 struct sched_statistics *stats;
5437 stats = __schedstats_from_se(se);
5438 __schedstat_set(stats->slice_max,
5439 max((u64)stats->slice_max,
5440 se->sum_exec_runtime - se->prev_sum_exec_runtime));
5443 se->prev_sum_exec_runtime = se->sum_exec_runtime;
5447 * Pick the next process, keeping these things in mind, in this order:
5448 * 1) keep things fair between processes/task groups
5449 * 2) pick the "next" process, since someone really wants that to run
5450 * 3) pick the "last" process, for cache locality
5451 * 4) do not run the "skip" process, if something else is available
5453 static struct sched_entity *
5454 pick_next_entity(struct cfs_rq *cfs_rq)
5457 * Enabling NEXT_BUDDY will affect latency but not fairness.
5459 if (sched_feat(NEXT_BUDDY) &&
5460 cfs_rq->next && entity_eligible(cfs_rq, cfs_rq->next))
5461 return cfs_rq->next;
5463 return pick_eevdf(cfs_rq);
5466 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
5468 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
5471 * If still on the runqueue then deactivate_task()
5472 * was not called and update_curr() has to be done:
5475 update_curr(cfs_rq);
5477 /* throttle cfs_rqs exceeding runtime */
5478 check_cfs_rq_runtime(cfs_rq);
5481 update_stats_wait_start_fair(cfs_rq, prev);
5482 /* Put 'current' back into the tree. */
5483 __enqueue_entity(cfs_rq, prev);
5484 /* in !on_rq case, update occurred at dequeue */
5485 update_load_avg(cfs_rq, prev, 0);
5487 cfs_rq->curr = NULL;
5491 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
5494 * Update run-time statistics of the 'current'.
5496 update_curr(cfs_rq);
5499 * Ensure that runnable average is periodically updated.
5501 update_load_avg(cfs_rq, curr, UPDATE_TG);
5502 update_cfs_group(curr);
5504 #ifdef CONFIG_SCHED_HRTICK
5506 * queued ticks are scheduled to match the slice, so don't bother
5507 * validating it and just reschedule.
5510 resched_curr(rq_of(cfs_rq));
5514 * don't let the period tick interfere with the hrtick preemption
5516 if (!sched_feat(DOUBLE_TICK) &&
5517 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
5523 /**************************************************
5524 * CFS bandwidth control machinery
5527 #ifdef CONFIG_CFS_BANDWIDTH
5529 #ifdef CONFIG_JUMP_LABEL
5530 static struct static_key __cfs_bandwidth_used;
5532 static inline bool cfs_bandwidth_used(void)
5534 return static_key_false(&__cfs_bandwidth_used);
5537 void cfs_bandwidth_usage_inc(void)
5539 static_key_slow_inc_cpuslocked(&__cfs_bandwidth_used);
5542 void cfs_bandwidth_usage_dec(void)
5544 static_key_slow_dec_cpuslocked(&__cfs_bandwidth_used);
5546 #else /* CONFIG_JUMP_LABEL */
5547 static bool cfs_bandwidth_used(void)
5552 void cfs_bandwidth_usage_inc(void) {}
5553 void cfs_bandwidth_usage_dec(void) {}
5554 #endif /* CONFIG_JUMP_LABEL */
5557 * default period for cfs group bandwidth.
5558 * default: 0.1s, units: nanoseconds
5560 static inline u64 default_cfs_period(void)
5562 return 100000000ULL;
5565 static inline u64 sched_cfs_bandwidth_slice(void)
5567 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
5571 * Replenish runtime according to assigned quota. We use sched_clock_cpu
5572 * directly instead of rq->clock to avoid adding additional synchronization
5575 * requires cfs_b->lock
5577 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
5581 if (unlikely(cfs_b->quota == RUNTIME_INF))
5584 cfs_b->runtime += cfs_b->quota;
5585 runtime = cfs_b->runtime_snap - cfs_b->runtime;
5587 cfs_b->burst_time += runtime;
5591 cfs_b->runtime = min(cfs_b->runtime, cfs_b->quota + cfs_b->burst);
5592 cfs_b->runtime_snap = cfs_b->runtime;
5595 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
5597 return &tg->cfs_bandwidth;
5600 /* returns 0 on failure to allocate runtime */
5601 static int __assign_cfs_rq_runtime(struct cfs_bandwidth *cfs_b,
5602 struct cfs_rq *cfs_rq, u64 target_runtime)
5604 u64 min_amount, amount = 0;
5606 lockdep_assert_held(&cfs_b->lock);
5608 /* note: this is a positive sum as runtime_remaining <= 0 */
5609 min_amount = target_runtime - cfs_rq->runtime_remaining;
5611 if (cfs_b->quota == RUNTIME_INF)
5612 amount = min_amount;
5614 start_cfs_bandwidth(cfs_b);
5616 if (cfs_b->runtime > 0) {
5617 amount = min(cfs_b->runtime, min_amount);
5618 cfs_b->runtime -= amount;
5623 cfs_rq->runtime_remaining += amount;
5625 return cfs_rq->runtime_remaining > 0;
5628 /* returns 0 on failure to allocate runtime */
5629 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
5631 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
5634 raw_spin_lock(&cfs_b->lock);
5635 ret = __assign_cfs_rq_runtime(cfs_b, cfs_rq, sched_cfs_bandwidth_slice());
5636 raw_spin_unlock(&cfs_b->lock);
5641 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
5643 /* dock delta_exec before expiring quota (as it could span periods) */
5644 cfs_rq->runtime_remaining -= delta_exec;
5646 if (likely(cfs_rq->runtime_remaining > 0))
5649 if (cfs_rq->throttled)
5652 * if we're unable to extend our runtime we resched so that the active
5653 * hierarchy can be throttled
5655 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
5656 resched_curr(rq_of(cfs_rq));
5659 static __always_inline
5660 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
5662 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
5665 __account_cfs_rq_runtime(cfs_rq, delta_exec);
5668 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
5670 return cfs_bandwidth_used() && cfs_rq->throttled;
5673 /* check whether cfs_rq, or any parent, is throttled */
5674 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
5676 return cfs_bandwidth_used() && cfs_rq->throttle_count;
5680 * Ensure that neither of the group entities corresponding to src_cpu or
5681 * dest_cpu are members of a throttled hierarchy when performing group
5682 * load-balance operations.
5684 static inline int throttled_lb_pair(struct task_group *tg,
5685 int src_cpu, int dest_cpu)
5687 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
5689 src_cfs_rq = tg->cfs_rq[src_cpu];
5690 dest_cfs_rq = tg->cfs_rq[dest_cpu];
5692 return throttled_hierarchy(src_cfs_rq) ||
5693 throttled_hierarchy(dest_cfs_rq);
5696 static int tg_unthrottle_up(struct task_group *tg, void *data)
5698 struct rq *rq = data;
5699 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
5701 cfs_rq->throttle_count--;
5702 if (!cfs_rq->throttle_count) {
5703 cfs_rq->throttled_clock_pelt_time += rq_clock_pelt(rq) -
5704 cfs_rq->throttled_clock_pelt;
5706 /* Add cfs_rq with load or one or more already running entities to the list */
5707 if (!cfs_rq_is_decayed(cfs_rq))
5708 list_add_leaf_cfs_rq(cfs_rq);
5710 if (cfs_rq->throttled_clock_self) {
5711 u64 delta = rq_clock(rq) - cfs_rq->throttled_clock_self;
5713 cfs_rq->throttled_clock_self = 0;
5715 if (SCHED_WARN_ON((s64)delta < 0))
5718 cfs_rq->throttled_clock_self_time += delta;
5725 static int tg_throttle_down(struct task_group *tg, void *data)
5727 struct rq *rq = data;
5728 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
5730 /* group is entering throttled state, stop time */
5731 if (!cfs_rq->throttle_count) {
5732 cfs_rq->throttled_clock_pelt = rq_clock_pelt(rq);
5733 list_del_leaf_cfs_rq(cfs_rq);
5735 SCHED_WARN_ON(cfs_rq->throttled_clock_self);
5736 if (cfs_rq->nr_running)
5737 cfs_rq->throttled_clock_self = rq_clock(rq);
5739 cfs_rq->throttle_count++;
5744 static bool throttle_cfs_rq(struct cfs_rq *cfs_rq)
5746 struct rq *rq = rq_of(cfs_rq);
5747 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
5748 struct sched_entity *se;
5749 long task_delta, idle_task_delta, dequeue = 1;
5751 raw_spin_lock(&cfs_b->lock);
5752 /* This will start the period timer if necessary */
5753 if (__assign_cfs_rq_runtime(cfs_b, cfs_rq, 1)) {
5755 * We have raced with bandwidth becoming available, and if we
5756 * actually throttled the timer might not unthrottle us for an
5757 * entire period. We additionally needed to make sure that any
5758 * subsequent check_cfs_rq_runtime calls agree not to throttle
5759 * us, as we may commit to do cfs put_prev+pick_next, so we ask
5760 * for 1ns of runtime rather than just check cfs_b.
5764 list_add_tail_rcu(&cfs_rq->throttled_list,
5765 &cfs_b->throttled_cfs_rq);
5767 raw_spin_unlock(&cfs_b->lock);
5770 return false; /* Throttle no longer required. */
5772 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
5774 /* freeze hierarchy runnable averages while throttled */
5776 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
5779 task_delta = cfs_rq->h_nr_running;
5780 idle_task_delta = cfs_rq->idle_h_nr_running;
5781 for_each_sched_entity(se) {
5782 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
5783 /* throttled entity or throttle-on-deactivate */
5787 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
5789 if (cfs_rq_is_idle(group_cfs_rq(se)))
5790 idle_task_delta = cfs_rq->h_nr_running;
5792 qcfs_rq->h_nr_running -= task_delta;
5793 qcfs_rq->idle_h_nr_running -= idle_task_delta;
5795 if (qcfs_rq->load.weight) {
5796 /* Avoid re-evaluating load for this entity: */
5797 se = parent_entity(se);
5802 for_each_sched_entity(se) {
5803 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
5804 /* throttled entity or throttle-on-deactivate */
5808 update_load_avg(qcfs_rq, se, 0);
5809 se_update_runnable(se);
5811 if (cfs_rq_is_idle(group_cfs_rq(se)))
5812 idle_task_delta = cfs_rq->h_nr_running;
5814 qcfs_rq->h_nr_running -= task_delta;
5815 qcfs_rq->idle_h_nr_running -= idle_task_delta;
5818 /* At this point se is NULL and we are at root level*/
5819 sub_nr_running(rq, task_delta);
5823 * Note: distribution will already see us throttled via the
5824 * throttled-list. rq->lock protects completion.
5826 cfs_rq->throttled = 1;
5827 SCHED_WARN_ON(cfs_rq->throttled_clock);
5828 if (cfs_rq->nr_running)
5829 cfs_rq->throttled_clock = rq_clock(rq);
5833 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
5835 struct rq *rq = rq_of(cfs_rq);
5836 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
5837 struct sched_entity *se;
5838 long task_delta, idle_task_delta;
5840 se = cfs_rq->tg->se[cpu_of(rq)];
5842 cfs_rq->throttled = 0;
5844 update_rq_clock(rq);
5846 raw_spin_lock(&cfs_b->lock);
5847 if (cfs_rq->throttled_clock) {
5848 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
5849 cfs_rq->throttled_clock = 0;
5851 list_del_rcu(&cfs_rq->throttled_list);
5852 raw_spin_unlock(&cfs_b->lock);
5854 /* update hierarchical throttle state */
5855 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
5857 if (!cfs_rq->load.weight) {
5858 if (!cfs_rq->on_list)
5861 * Nothing to run but something to decay (on_list)?
5862 * Complete the branch.
5864 for_each_sched_entity(se) {
5865 if (list_add_leaf_cfs_rq(cfs_rq_of(se)))
5868 goto unthrottle_throttle;
5871 task_delta = cfs_rq->h_nr_running;
5872 idle_task_delta = cfs_rq->idle_h_nr_running;
5873 for_each_sched_entity(se) {
5874 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
5878 enqueue_entity(qcfs_rq, se, ENQUEUE_WAKEUP);
5880 if (cfs_rq_is_idle(group_cfs_rq(se)))
5881 idle_task_delta = cfs_rq->h_nr_running;
5883 qcfs_rq->h_nr_running += task_delta;
5884 qcfs_rq->idle_h_nr_running += idle_task_delta;
5886 /* end evaluation on encountering a throttled cfs_rq */
5887 if (cfs_rq_throttled(qcfs_rq))
5888 goto unthrottle_throttle;
5891 for_each_sched_entity(se) {
5892 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
5894 update_load_avg(qcfs_rq, se, UPDATE_TG);
5895 se_update_runnable(se);
5897 if (cfs_rq_is_idle(group_cfs_rq(se)))
5898 idle_task_delta = cfs_rq->h_nr_running;
5900 qcfs_rq->h_nr_running += task_delta;
5901 qcfs_rq->idle_h_nr_running += idle_task_delta;
5903 /* end evaluation on encountering a throttled cfs_rq */
5904 if (cfs_rq_throttled(qcfs_rq))
5905 goto unthrottle_throttle;
5908 /* At this point se is NULL and we are at root level*/
5909 add_nr_running(rq, task_delta);
5911 unthrottle_throttle:
5912 assert_list_leaf_cfs_rq(rq);
5914 /* Determine whether we need to wake up potentially idle CPU: */
5915 if (rq->curr == rq->idle && rq->cfs.nr_running)
5920 static void __cfsb_csd_unthrottle(void *arg)
5922 struct cfs_rq *cursor, *tmp;
5923 struct rq *rq = arg;
5929 * Iterating over the list can trigger several call to
5930 * update_rq_clock() in unthrottle_cfs_rq().
5931 * Do it once and skip the potential next ones.
5933 update_rq_clock(rq);
5934 rq_clock_start_loop_update(rq);
5937 * Since we hold rq lock we're safe from concurrent manipulation of
5938 * the CSD list. However, this RCU critical section annotates the
5939 * fact that we pair with sched_free_group_rcu(), so that we cannot
5940 * race with group being freed in the window between removing it
5941 * from the list and advancing to the next entry in the list.
5945 list_for_each_entry_safe(cursor, tmp, &rq->cfsb_csd_list,
5946 throttled_csd_list) {
5947 list_del_init(&cursor->throttled_csd_list);
5949 if (cfs_rq_throttled(cursor))
5950 unthrottle_cfs_rq(cursor);
5955 rq_clock_stop_loop_update(rq);
5959 static inline void __unthrottle_cfs_rq_async(struct cfs_rq *cfs_rq)
5961 struct rq *rq = rq_of(cfs_rq);
5964 if (rq == this_rq()) {
5965 unthrottle_cfs_rq(cfs_rq);
5969 /* Already enqueued */
5970 if (SCHED_WARN_ON(!list_empty(&cfs_rq->throttled_csd_list)))
5973 first = list_empty(&rq->cfsb_csd_list);
5974 list_add_tail(&cfs_rq->throttled_csd_list, &rq->cfsb_csd_list);
5976 smp_call_function_single_async(cpu_of(rq), &rq->cfsb_csd);
5979 static inline void __unthrottle_cfs_rq_async(struct cfs_rq *cfs_rq)
5981 unthrottle_cfs_rq(cfs_rq);
5985 static void unthrottle_cfs_rq_async(struct cfs_rq *cfs_rq)
5987 lockdep_assert_rq_held(rq_of(cfs_rq));
5989 if (SCHED_WARN_ON(!cfs_rq_throttled(cfs_rq) ||
5990 cfs_rq->runtime_remaining <= 0))
5993 __unthrottle_cfs_rq_async(cfs_rq);
5996 static bool distribute_cfs_runtime(struct cfs_bandwidth *cfs_b)
5998 int this_cpu = smp_processor_id();
5999 u64 runtime, remaining = 1;
6000 bool throttled = false;
6001 struct cfs_rq *cfs_rq, *tmp;
6004 LIST_HEAD(local_unthrottle);
6007 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
6016 rq_lock_irqsave(rq, &rf);
6017 if (!cfs_rq_throttled(cfs_rq))
6020 /* Already queued for async unthrottle */
6021 if (!list_empty(&cfs_rq->throttled_csd_list))
6024 /* By the above checks, this should never be true */
6025 SCHED_WARN_ON(cfs_rq->runtime_remaining > 0);
6027 raw_spin_lock(&cfs_b->lock);
6028 runtime = -cfs_rq->runtime_remaining + 1;
6029 if (runtime > cfs_b->runtime)
6030 runtime = cfs_b->runtime;
6031 cfs_b->runtime -= runtime;
6032 remaining = cfs_b->runtime;
6033 raw_spin_unlock(&cfs_b->lock);
6035 cfs_rq->runtime_remaining += runtime;
6037 /* we check whether we're throttled above */
6038 if (cfs_rq->runtime_remaining > 0) {
6039 if (cpu_of(rq) != this_cpu) {
6040 unthrottle_cfs_rq_async(cfs_rq);
6043 * We currently only expect to be unthrottling
6044 * a single cfs_rq locally.
6046 SCHED_WARN_ON(!list_empty(&local_unthrottle));
6047 list_add_tail(&cfs_rq->throttled_csd_list,
6055 rq_unlock_irqrestore(rq, &rf);
6058 list_for_each_entry_safe(cfs_rq, tmp, &local_unthrottle,
6059 throttled_csd_list) {
6060 struct rq *rq = rq_of(cfs_rq);
6062 rq_lock_irqsave(rq, &rf);
6064 list_del_init(&cfs_rq->throttled_csd_list);
6066 if (cfs_rq_throttled(cfs_rq))
6067 unthrottle_cfs_rq(cfs_rq);
6069 rq_unlock_irqrestore(rq, &rf);
6071 SCHED_WARN_ON(!list_empty(&local_unthrottle));
6079 * Responsible for refilling a task_group's bandwidth and unthrottling its
6080 * cfs_rqs as appropriate. If there has been no activity within the last
6081 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
6082 * used to track this state.
6084 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun, unsigned long flags)
6088 /* no need to continue the timer with no bandwidth constraint */
6089 if (cfs_b->quota == RUNTIME_INF)
6090 goto out_deactivate;
6092 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
6093 cfs_b->nr_periods += overrun;
6095 /* Refill extra burst quota even if cfs_b->idle */
6096 __refill_cfs_bandwidth_runtime(cfs_b);
6099 * idle depends on !throttled (for the case of a large deficit), and if
6100 * we're going inactive then everything else can be deferred
6102 if (cfs_b->idle && !throttled)
6103 goto out_deactivate;
6106 /* mark as potentially idle for the upcoming period */
6111 /* account preceding periods in which throttling occurred */
6112 cfs_b->nr_throttled += overrun;
6115 * This check is repeated as we release cfs_b->lock while we unthrottle.
6117 while (throttled && cfs_b->runtime > 0) {
6118 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
6119 /* we can't nest cfs_b->lock while distributing bandwidth */
6120 throttled = distribute_cfs_runtime(cfs_b);
6121 raw_spin_lock_irqsave(&cfs_b->lock, flags);
6125 * While we are ensured activity in the period following an
6126 * unthrottle, this also covers the case in which the new bandwidth is
6127 * insufficient to cover the existing bandwidth deficit. (Forcing the
6128 * timer to remain active while there are any throttled entities.)
6138 /* a cfs_rq won't donate quota below this amount */
6139 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
6140 /* minimum remaining period time to redistribute slack quota */
6141 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
6142 /* how long we wait to gather additional slack before distributing */
6143 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
6146 * Are we near the end of the current quota period?
6148 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
6149 * hrtimer base being cleared by hrtimer_start. In the case of
6150 * migrate_hrtimers, base is never cleared, so we are fine.
6152 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
6154 struct hrtimer *refresh_timer = &cfs_b->period_timer;
6157 /* if the call-back is running a quota refresh is already occurring */
6158 if (hrtimer_callback_running(refresh_timer))
6161 /* is a quota refresh about to occur? */
6162 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
6163 if (remaining < (s64)min_expire)
6169 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
6171 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
6173 /* if there's a quota refresh soon don't bother with slack */
6174 if (runtime_refresh_within(cfs_b, min_left))
6177 /* don't push forwards an existing deferred unthrottle */
6178 if (cfs_b->slack_started)
6180 cfs_b->slack_started = true;
6182 hrtimer_start(&cfs_b->slack_timer,
6183 ns_to_ktime(cfs_bandwidth_slack_period),
6187 /* we know any runtime found here is valid as update_curr() precedes return */
6188 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
6190 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
6191 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
6193 if (slack_runtime <= 0)
6196 raw_spin_lock(&cfs_b->lock);
6197 if (cfs_b->quota != RUNTIME_INF) {
6198 cfs_b->runtime += slack_runtime;
6200 /* we are under rq->lock, defer unthrottling using a timer */
6201 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
6202 !list_empty(&cfs_b->throttled_cfs_rq))
6203 start_cfs_slack_bandwidth(cfs_b);
6205 raw_spin_unlock(&cfs_b->lock);
6207 /* even if it's not valid for return we don't want to try again */
6208 cfs_rq->runtime_remaining -= slack_runtime;
6211 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
6213 if (!cfs_bandwidth_used())
6216 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
6219 __return_cfs_rq_runtime(cfs_rq);
6223 * This is done with a timer (instead of inline with bandwidth return) since
6224 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
6226 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
6228 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
6229 unsigned long flags;
6231 /* confirm we're still not at a refresh boundary */
6232 raw_spin_lock_irqsave(&cfs_b->lock, flags);
6233 cfs_b->slack_started = false;
6235 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
6236 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
6240 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice)
6241 runtime = cfs_b->runtime;
6243 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
6248 distribute_cfs_runtime(cfs_b);
6252 * When a group wakes up we want to make sure that its quota is not already
6253 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
6254 * runtime as update_curr() throttling can not trigger until it's on-rq.
6256 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
6258 if (!cfs_bandwidth_used())
6261 /* an active group must be handled by the update_curr()->put() path */
6262 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
6265 /* ensure the group is not already throttled */
6266 if (cfs_rq_throttled(cfs_rq))
6269 /* update runtime allocation */
6270 account_cfs_rq_runtime(cfs_rq, 0);
6271 if (cfs_rq->runtime_remaining <= 0)
6272 throttle_cfs_rq(cfs_rq);
6275 static void sync_throttle(struct task_group *tg, int cpu)
6277 struct cfs_rq *pcfs_rq, *cfs_rq;
6279 if (!cfs_bandwidth_used())
6285 cfs_rq = tg->cfs_rq[cpu];
6286 pcfs_rq = tg->parent->cfs_rq[cpu];
6288 cfs_rq->throttle_count = pcfs_rq->throttle_count;
6289 cfs_rq->throttled_clock_pelt = rq_clock_pelt(cpu_rq(cpu));
6292 /* conditionally throttle active cfs_rq's from put_prev_entity() */
6293 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
6295 if (!cfs_bandwidth_used())
6298 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
6302 * it's possible for a throttled entity to be forced into a running
6303 * state (e.g. set_curr_task), in this case we're finished.
6305 if (cfs_rq_throttled(cfs_rq))
6308 return throttle_cfs_rq(cfs_rq);
6311 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
6313 struct cfs_bandwidth *cfs_b =
6314 container_of(timer, struct cfs_bandwidth, slack_timer);
6316 do_sched_cfs_slack_timer(cfs_b);
6318 return HRTIMER_NORESTART;
6321 extern const u64 max_cfs_quota_period;
6323 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
6325 struct cfs_bandwidth *cfs_b =
6326 container_of(timer, struct cfs_bandwidth, period_timer);
6327 unsigned long flags;
6332 raw_spin_lock_irqsave(&cfs_b->lock, flags);
6334 overrun = hrtimer_forward_now(timer, cfs_b->period);
6338 idle = do_sched_cfs_period_timer(cfs_b, overrun, flags);
6341 u64 new, old = ktime_to_ns(cfs_b->period);
6344 * Grow period by a factor of 2 to avoid losing precision.
6345 * Precision loss in the quota/period ratio can cause __cfs_schedulable
6349 if (new < max_cfs_quota_period) {
6350 cfs_b->period = ns_to_ktime(new);
6354 pr_warn_ratelimited(
6355 "cfs_period_timer[cpu%d]: period too short, scaling up (new cfs_period_us = %lld, cfs_quota_us = %lld)\n",
6357 div_u64(new, NSEC_PER_USEC),
6358 div_u64(cfs_b->quota, NSEC_PER_USEC));
6360 pr_warn_ratelimited(
6361 "cfs_period_timer[cpu%d]: period too short, but cannot scale up without losing precision (cfs_period_us = %lld, cfs_quota_us = %lld)\n",
6363 div_u64(old, NSEC_PER_USEC),
6364 div_u64(cfs_b->quota, NSEC_PER_USEC));
6367 /* reset count so we don't come right back in here */
6372 cfs_b->period_active = 0;
6373 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
6375 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
6378 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b, struct cfs_bandwidth *parent)
6380 raw_spin_lock_init(&cfs_b->lock);
6382 cfs_b->quota = RUNTIME_INF;
6383 cfs_b->period = ns_to_ktime(default_cfs_period());
6385 cfs_b->hierarchical_quota = parent ? parent->hierarchical_quota : RUNTIME_INF;
6387 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
6388 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED);
6389 cfs_b->period_timer.function = sched_cfs_period_timer;
6391 /* Add a random offset so that timers interleave */
6392 hrtimer_set_expires(&cfs_b->period_timer,
6393 get_random_u32_below(cfs_b->period));
6394 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
6395 cfs_b->slack_timer.function = sched_cfs_slack_timer;
6396 cfs_b->slack_started = false;
6399 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
6401 cfs_rq->runtime_enabled = 0;
6402 INIT_LIST_HEAD(&cfs_rq->throttled_list);
6403 INIT_LIST_HEAD(&cfs_rq->throttled_csd_list);
6406 void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
6408 lockdep_assert_held(&cfs_b->lock);
6410 if (cfs_b->period_active)
6413 cfs_b->period_active = 1;
6414 hrtimer_forward_now(&cfs_b->period_timer, cfs_b->period);
6415 hrtimer_start_expires(&cfs_b->period_timer, HRTIMER_MODE_ABS_PINNED);
6418 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
6420 int __maybe_unused i;
6422 /* init_cfs_bandwidth() was not called */
6423 if (!cfs_b->throttled_cfs_rq.next)
6426 hrtimer_cancel(&cfs_b->period_timer);
6427 hrtimer_cancel(&cfs_b->slack_timer);
6430 * It is possible that we still have some cfs_rq's pending on a CSD
6431 * list, though this race is very rare. In order for this to occur, we
6432 * must have raced with the last task leaving the group while there
6433 * exist throttled cfs_rq(s), and the period_timer must have queued the
6434 * CSD item but the remote cpu has not yet processed it. To handle this,
6435 * we can simply flush all pending CSD work inline here. We're
6436 * guaranteed at this point that no additional cfs_rq of this group can
6440 for_each_possible_cpu(i) {
6441 struct rq *rq = cpu_rq(i);
6442 unsigned long flags;
6444 if (list_empty(&rq->cfsb_csd_list))
6447 local_irq_save(flags);
6448 __cfsb_csd_unthrottle(rq);
6449 local_irq_restore(flags);
6455 * Both these CPU hotplug callbacks race against unregister_fair_sched_group()
6457 * The race is harmless, since modifying bandwidth settings of unhooked group
6458 * bits doesn't do much.
6461 /* cpu online callback */
6462 static void __maybe_unused update_runtime_enabled(struct rq *rq)
6464 struct task_group *tg;
6466 lockdep_assert_rq_held(rq);
6469 list_for_each_entry_rcu(tg, &task_groups, list) {
6470 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
6471 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
6473 raw_spin_lock(&cfs_b->lock);
6474 cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF;
6475 raw_spin_unlock(&cfs_b->lock);
6480 /* cpu offline callback */
6481 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
6483 struct task_group *tg;
6485 lockdep_assert_rq_held(rq);
6488 * The rq clock has already been updated in the
6489 * set_rq_offline(), so we should skip updating
6490 * the rq clock again in unthrottle_cfs_rq().
6492 rq_clock_start_loop_update(rq);
6495 list_for_each_entry_rcu(tg, &task_groups, list) {
6496 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
6498 if (!cfs_rq->runtime_enabled)
6502 * clock_task is not advancing so we just need to make sure
6503 * there's some valid quota amount
6505 cfs_rq->runtime_remaining = 1;
6507 * Offline rq is schedulable till CPU is completely disabled
6508 * in take_cpu_down(), so we prevent new cfs throttling here.
6510 cfs_rq->runtime_enabled = 0;
6512 if (cfs_rq_throttled(cfs_rq))
6513 unthrottle_cfs_rq(cfs_rq);
6517 rq_clock_stop_loop_update(rq);
6520 bool cfs_task_bw_constrained(struct task_struct *p)
6522 struct cfs_rq *cfs_rq = task_cfs_rq(p);
6524 if (!cfs_bandwidth_used())
6527 if (cfs_rq->runtime_enabled ||
6528 tg_cfs_bandwidth(cfs_rq->tg)->hierarchical_quota != RUNTIME_INF)
6534 #ifdef CONFIG_NO_HZ_FULL
6535 /* called from pick_next_task_fair() */
6536 static void sched_fair_update_stop_tick(struct rq *rq, struct task_struct *p)
6538 int cpu = cpu_of(rq);
6540 if (!sched_feat(HZ_BW) || !cfs_bandwidth_used())
6543 if (!tick_nohz_full_cpu(cpu))
6546 if (rq->nr_running != 1)
6550 * We know there is only one task runnable and we've just picked it. The
6551 * normal enqueue path will have cleared TICK_DEP_BIT_SCHED if we will
6552 * be otherwise able to stop the tick. Just need to check if we are using
6553 * bandwidth control.
6555 if (cfs_task_bw_constrained(p))
6556 tick_nohz_dep_set_cpu(cpu, TICK_DEP_BIT_SCHED);
6560 #else /* CONFIG_CFS_BANDWIDTH */
6562 static inline bool cfs_bandwidth_used(void)
6567 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {}
6568 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; }
6569 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
6570 static inline void sync_throttle(struct task_group *tg, int cpu) {}
6571 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
6573 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
6578 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
6583 static inline int throttled_lb_pair(struct task_group *tg,
6584 int src_cpu, int dest_cpu)
6589 #ifdef CONFIG_FAIR_GROUP_SCHED
6590 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b, struct cfs_bandwidth *parent) {}
6591 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
6594 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
6598 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
6599 static inline void update_runtime_enabled(struct rq *rq) {}
6600 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
6601 #ifdef CONFIG_CGROUP_SCHED
6602 bool cfs_task_bw_constrained(struct task_struct *p)
6607 #endif /* CONFIG_CFS_BANDWIDTH */
6609 #if !defined(CONFIG_CFS_BANDWIDTH) || !defined(CONFIG_NO_HZ_FULL)
6610 static inline void sched_fair_update_stop_tick(struct rq *rq, struct task_struct *p) {}
6613 /**************************************************
6614 * CFS operations on tasks:
6617 #ifdef CONFIG_SCHED_HRTICK
6618 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
6620 struct sched_entity *se = &p->se;
6622 SCHED_WARN_ON(task_rq(p) != rq);
6624 if (rq->cfs.h_nr_running > 1) {
6625 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
6626 u64 slice = se->slice;
6627 s64 delta = slice - ran;
6630 if (task_current(rq, p))
6634 hrtick_start(rq, delta);
6639 * called from enqueue/dequeue and updates the hrtick when the
6640 * current task is from our class and nr_running is low enough
6643 static void hrtick_update(struct rq *rq)
6645 struct task_struct *curr = rq->curr;
6647 if (!hrtick_enabled_fair(rq) || curr->sched_class != &fair_sched_class)
6650 hrtick_start_fair(rq, curr);
6652 #else /* !CONFIG_SCHED_HRTICK */
6654 hrtick_start_fair(struct rq *rq, struct task_struct *p)
6658 static inline void hrtick_update(struct rq *rq)
6664 static inline bool cpu_overutilized(int cpu)
6666 unsigned long rq_util_min = uclamp_rq_get(cpu_rq(cpu), UCLAMP_MIN);
6667 unsigned long rq_util_max = uclamp_rq_get(cpu_rq(cpu), UCLAMP_MAX);
6669 /* Return true only if the utilization doesn't fit CPU's capacity */
6670 return !util_fits_cpu(cpu_util_cfs(cpu), rq_util_min, rq_util_max, cpu);
6673 static inline void update_overutilized_status(struct rq *rq)
6675 if (!READ_ONCE(rq->rd->overutilized) && cpu_overutilized(rq->cpu)) {
6676 WRITE_ONCE(rq->rd->overutilized, SG_OVERUTILIZED);
6677 trace_sched_overutilized_tp(rq->rd, SG_OVERUTILIZED);
6681 static inline void update_overutilized_status(struct rq *rq) { }
6684 /* Runqueue only has SCHED_IDLE tasks enqueued */
6685 static int sched_idle_rq(struct rq *rq)
6687 return unlikely(rq->nr_running == rq->cfs.idle_h_nr_running &&
6692 static int sched_idle_cpu(int cpu)
6694 return sched_idle_rq(cpu_rq(cpu));
6699 * The enqueue_task method is called before nr_running is
6700 * increased. Here we update the fair scheduling stats and
6701 * then put the task into the rbtree:
6704 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
6706 struct cfs_rq *cfs_rq;
6707 struct sched_entity *se = &p->se;
6708 int idle_h_nr_running = task_has_idle_policy(p);
6709 int task_new = !(flags & ENQUEUE_WAKEUP);
6712 * The code below (indirectly) updates schedutil which looks at
6713 * the cfs_rq utilization to select a frequency.
6714 * Let's add the task's estimated utilization to the cfs_rq's
6715 * estimated utilization, before we update schedutil.
6717 util_est_enqueue(&rq->cfs, p);
6720 * If in_iowait is set, the code below may not trigger any cpufreq
6721 * utilization updates, so do it here explicitly with the IOWAIT flag
6725 cpufreq_update_util(rq, SCHED_CPUFREQ_IOWAIT);
6727 for_each_sched_entity(se) {
6730 cfs_rq = cfs_rq_of(se);
6731 enqueue_entity(cfs_rq, se, flags);
6733 cfs_rq->h_nr_running++;
6734 cfs_rq->idle_h_nr_running += idle_h_nr_running;
6736 if (cfs_rq_is_idle(cfs_rq))
6737 idle_h_nr_running = 1;
6739 /* end evaluation on encountering a throttled cfs_rq */
6740 if (cfs_rq_throttled(cfs_rq))
6741 goto enqueue_throttle;
6743 flags = ENQUEUE_WAKEUP;
6746 for_each_sched_entity(se) {
6747 cfs_rq = cfs_rq_of(se);
6749 update_load_avg(cfs_rq, se, UPDATE_TG);
6750 se_update_runnable(se);
6751 update_cfs_group(se);
6753 cfs_rq->h_nr_running++;
6754 cfs_rq->idle_h_nr_running += idle_h_nr_running;
6756 if (cfs_rq_is_idle(cfs_rq))
6757 idle_h_nr_running = 1;
6759 /* end evaluation on encountering a throttled cfs_rq */
6760 if (cfs_rq_throttled(cfs_rq))
6761 goto enqueue_throttle;
6764 /* At this point se is NULL and we are at root level*/
6765 add_nr_running(rq, 1);
6768 * Since new tasks are assigned an initial util_avg equal to
6769 * half of the spare capacity of their CPU, tiny tasks have the
6770 * ability to cross the overutilized threshold, which will
6771 * result in the load balancer ruining all the task placement
6772 * done by EAS. As a way to mitigate that effect, do not account
6773 * for the first enqueue operation of new tasks during the
6774 * overutilized flag detection.
6776 * A better way of solving this problem would be to wait for
6777 * the PELT signals of tasks to converge before taking them
6778 * into account, but that is not straightforward to implement,
6779 * and the following generally works well enough in practice.
6782 update_overutilized_status(rq);
6785 assert_list_leaf_cfs_rq(rq);
6790 static void set_next_buddy(struct sched_entity *se);
6793 * The dequeue_task method is called before nr_running is
6794 * decreased. We remove the task from the rbtree and
6795 * update the fair scheduling stats:
6797 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
6799 struct cfs_rq *cfs_rq;
6800 struct sched_entity *se = &p->se;
6801 int task_sleep = flags & DEQUEUE_SLEEP;
6802 int idle_h_nr_running = task_has_idle_policy(p);
6803 bool was_sched_idle = sched_idle_rq(rq);
6805 util_est_dequeue(&rq->cfs, p);
6807 for_each_sched_entity(se) {
6808 cfs_rq = cfs_rq_of(se);
6809 dequeue_entity(cfs_rq, se, flags);
6811 cfs_rq->h_nr_running--;
6812 cfs_rq->idle_h_nr_running -= idle_h_nr_running;
6814 if (cfs_rq_is_idle(cfs_rq))
6815 idle_h_nr_running = 1;
6817 /* end evaluation on encountering a throttled cfs_rq */
6818 if (cfs_rq_throttled(cfs_rq))
6819 goto dequeue_throttle;
6821 /* Don't dequeue parent if it has other entities besides us */
6822 if (cfs_rq->load.weight) {
6823 /* Avoid re-evaluating load for this entity: */
6824 se = parent_entity(se);
6826 * Bias pick_next to pick a task from this cfs_rq, as
6827 * p is sleeping when it is within its sched_slice.
6829 if (task_sleep && se && !throttled_hierarchy(cfs_rq))
6833 flags |= DEQUEUE_SLEEP;
6836 for_each_sched_entity(se) {
6837 cfs_rq = cfs_rq_of(se);
6839 update_load_avg(cfs_rq, se, UPDATE_TG);
6840 se_update_runnable(se);
6841 update_cfs_group(se);
6843 cfs_rq->h_nr_running--;
6844 cfs_rq->idle_h_nr_running -= idle_h_nr_running;
6846 if (cfs_rq_is_idle(cfs_rq))
6847 idle_h_nr_running = 1;
6849 /* end evaluation on encountering a throttled cfs_rq */
6850 if (cfs_rq_throttled(cfs_rq))
6851 goto dequeue_throttle;
6855 /* At this point se is NULL and we are at root level*/
6856 sub_nr_running(rq, 1);
6858 /* balance early to pull high priority tasks */
6859 if (unlikely(!was_sched_idle && sched_idle_rq(rq)))
6860 rq->next_balance = jiffies;
6863 util_est_update(&rq->cfs, p, task_sleep);
6869 /* Working cpumask for: load_balance, load_balance_newidle. */
6870 static DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
6871 static DEFINE_PER_CPU(cpumask_var_t, select_rq_mask);
6872 static DEFINE_PER_CPU(cpumask_var_t, should_we_balance_tmpmask);
6874 #ifdef CONFIG_NO_HZ_COMMON
6877 cpumask_var_t idle_cpus_mask;
6879 int has_blocked; /* Idle CPUS has blocked load */
6880 int needs_update; /* Newly idle CPUs need their next_balance collated */
6881 unsigned long next_balance; /* in jiffy units */
6882 unsigned long next_blocked; /* Next update of blocked load in jiffies */
6883 } nohz ____cacheline_aligned;
6885 #endif /* CONFIG_NO_HZ_COMMON */
6887 static unsigned long cpu_load(struct rq *rq)
6889 return cfs_rq_load_avg(&rq->cfs);
6893 * cpu_load_without - compute CPU load without any contributions from *p
6894 * @cpu: the CPU which load is requested
6895 * @p: the task which load should be discounted
6897 * The load of a CPU is defined by the load of tasks currently enqueued on that
6898 * CPU as well as tasks which are currently sleeping after an execution on that
6901 * This method returns the load of the specified CPU by discounting the load of
6902 * the specified task, whenever the task is currently contributing to the CPU
6905 static unsigned long cpu_load_without(struct rq *rq, struct task_struct *p)
6907 struct cfs_rq *cfs_rq;
6910 /* Task has no contribution or is new */
6911 if (cpu_of(rq) != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time))
6912 return cpu_load(rq);
6915 load = READ_ONCE(cfs_rq->avg.load_avg);
6917 /* Discount task's util from CPU's util */
6918 lsub_positive(&load, task_h_load(p));
6923 static unsigned long cpu_runnable(struct rq *rq)
6925 return cfs_rq_runnable_avg(&rq->cfs);
6928 static unsigned long cpu_runnable_without(struct rq *rq, struct task_struct *p)
6930 struct cfs_rq *cfs_rq;
6931 unsigned int runnable;
6933 /* Task has no contribution or is new */
6934 if (cpu_of(rq) != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time))
6935 return cpu_runnable(rq);
6938 runnable = READ_ONCE(cfs_rq->avg.runnable_avg);
6940 /* Discount task's runnable from CPU's runnable */
6941 lsub_positive(&runnable, p->se.avg.runnable_avg);
6946 static unsigned long capacity_of(int cpu)
6948 return cpu_rq(cpu)->cpu_capacity;
6951 static void record_wakee(struct task_struct *p)
6954 * Only decay a single time; tasks that have less then 1 wakeup per
6955 * jiffy will not have built up many flips.
6957 if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) {
6958 current->wakee_flips >>= 1;
6959 current->wakee_flip_decay_ts = jiffies;
6962 if (current->last_wakee != p) {
6963 current->last_wakee = p;
6964 current->wakee_flips++;
6969 * Detect M:N waker/wakee relationships via a switching-frequency heuristic.
6971 * A waker of many should wake a different task than the one last awakened
6972 * at a frequency roughly N times higher than one of its wakees.
6974 * In order to determine whether we should let the load spread vs consolidating
6975 * to shared cache, we look for a minimum 'flip' frequency of llc_size in one
6976 * partner, and a factor of lls_size higher frequency in the other.
6978 * With both conditions met, we can be relatively sure that the relationship is
6979 * non-monogamous, with partner count exceeding socket size.
6981 * Waker/wakee being client/server, worker/dispatcher, interrupt source or
6982 * whatever is irrelevant, spread criteria is apparent partner count exceeds
6985 static int wake_wide(struct task_struct *p)
6987 unsigned int master = current->wakee_flips;
6988 unsigned int slave = p->wakee_flips;
6989 int factor = __this_cpu_read(sd_llc_size);
6992 swap(master, slave);
6993 if (slave < factor || master < slave * factor)
6999 * The purpose of wake_affine() is to quickly determine on which CPU we can run
7000 * soonest. For the purpose of speed we only consider the waking and previous
7003 * wake_affine_idle() - only considers 'now', it check if the waking CPU is
7004 * cache-affine and is (or will be) idle.
7006 * wake_affine_weight() - considers the weight to reflect the average
7007 * scheduling latency of the CPUs. This seems to work
7008 * for the overloaded case.
7011 wake_affine_idle(int this_cpu, int prev_cpu, int sync)
7014 * If this_cpu is idle, it implies the wakeup is from interrupt
7015 * context. Only allow the move if cache is shared. Otherwise an
7016 * interrupt intensive workload could force all tasks onto one
7017 * node depending on the IO topology or IRQ affinity settings.
7019 * If the prev_cpu is idle and cache affine then avoid a migration.
7020 * There is no guarantee that the cache hot data from an interrupt
7021 * is more important than cache hot data on the prev_cpu and from
7022 * a cpufreq perspective, it's better to have higher utilisation
7025 if (available_idle_cpu(this_cpu) && cpus_share_cache(this_cpu, prev_cpu))
7026 return available_idle_cpu(prev_cpu) ? prev_cpu : this_cpu;
7028 if (sync && cpu_rq(this_cpu)->nr_running == 1)
7031 if (available_idle_cpu(prev_cpu))
7034 return nr_cpumask_bits;
7038 wake_affine_weight(struct sched_domain *sd, struct task_struct *p,
7039 int this_cpu, int prev_cpu, int sync)
7041 s64 this_eff_load, prev_eff_load;
7042 unsigned long task_load;
7044 this_eff_load = cpu_load(cpu_rq(this_cpu));
7047 unsigned long current_load = task_h_load(current);
7049 if (current_load > this_eff_load)
7052 this_eff_load -= current_load;
7055 task_load = task_h_load(p);
7057 this_eff_load += task_load;
7058 if (sched_feat(WA_BIAS))
7059 this_eff_load *= 100;
7060 this_eff_load *= capacity_of(prev_cpu);
7062 prev_eff_load = cpu_load(cpu_rq(prev_cpu));
7063 prev_eff_load -= task_load;
7064 if (sched_feat(WA_BIAS))
7065 prev_eff_load *= 100 + (sd->imbalance_pct - 100) / 2;
7066 prev_eff_load *= capacity_of(this_cpu);
7069 * If sync, adjust the weight of prev_eff_load such that if
7070 * prev_eff == this_eff that select_idle_sibling() will consider
7071 * stacking the wakee on top of the waker if no other CPU is
7077 return this_eff_load < prev_eff_load ? this_cpu : nr_cpumask_bits;
7080 static int wake_affine(struct sched_domain *sd, struct task_struct *p,
7081 int this_cpu, int prev_cpu, int sync)
7083 int target = nr_cpumask_bits;
7085 if (sched_feat(WA_IDLE))
7086 target = wake_affine_idle(this_cpu, prev_cpu, sync);
7088 if (sched_feat(WA_WEIGHT) && target == nr_cpumask_bits)
7089 target = wake_affine_weight(sd, p, this_cpu, prev_cpu, sync);
7091 schedstat_inc(p->stats.nr_wakeups_affine_attempts);
7092 if (target != this_cpu)
7095 schedstat_inc(sd->ttwu_move_affine);
7096 schedstat_inc(p->stats.nr_wakeups_affine);
7100 static struct sched_group *
7101 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu);
7104 * find_idlest_group_cpu - find the idlest CPU among the CPUs in the group.
7107 find_idlest_group_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
7109 unsigned long load, min_load = ULONG_MAX;
7110 unsigned int min_exit_latency = UINT_MAX;
7111 u64 latest_idle_timestamp = 0;
7112 int least_loaded_cpu = this_cpu;
7113 int shallowest_idle_cpu = -1;
7116 /* Check if we have any choice: */
7117 if (group->group_weight == 1)
7118 return cpumask_first(sched_group_span(group));
7120 /* Traverse only the allowed CPUs */
7121 for_each_cpu_and(i, sched_group_span(group), p->cpus_ptr) {
7122 struct rq *rq = cpu_rq(i);
7124 if (!sched_core_cookie_match(rq, p))
7127 if (sched_idle_cpu(i))
7130 if (available_idle_cpu(i)) {
7131 struct cpuidle_state *idle = idle_get_state(rq);
7132 if (idle && idle->exit_latency < min_exit_latency) {
7134 * We give priority to a CPU whose idle state
7135 * has the smallest exit latency irrespective
7136 * of any idle timestamp.
7138 min_exit_latency = idle->exit_latency;
7139 latest_idle_timestamp = rq->idle_stamp;
7140 shallowest_idle_cpu = i;
7141 } else if ((!idle || idle->exit_latency == min_exit_latency) &&
7142 rq->idle_stamp > latest_idle_timestamp) {
7144 * If equal or no active idle state, then
7145 * the most recently idled CPU might have
7148 latest_idle_timestamp = rq->idle_stamp;
7149 shallowest_idle_cpu = i;
7151 } else if (shallowest_idle_cpu == -1) {
7152 load = cpu_load(cpu_rq(i));
7153 if (load < min_load) {
7155 least_loaded_cpu = i;
7160 return shallowest_idle_cpu != -1 ? shallowest_idle_cpu : least_loaded_cpu;
7163 static inline int find_idlest_cpu(struct sched_domain *sd, struct task_struct *p,
7164 int cpu, int prev_cpu, int sd_flag)
7168 if (!cpumask_intersects(sched_domain_span(sd), p->cpus_ptr))
7172 * We need task's util for cpu_util_without, sync it up to
7173 * prev_cpu's last_update_time.
7175 if (!(sd_flag & SD_BALANCE_FORK))
7176 sync_entity_load_avg(&p->se);
7179 struct sched_group *group;
7180 struct sched_domain *tmp;
7183 if (!(sd->flags & sd_flag)) {
7188 group = find_idlest_group(sd, p, cpu);
7194 new_cpu = find_idlest_group_cpu(group, p, cpu);
7195 if (new_cpu == cpu) {
7196 /* Now try balancing at a lower domain level of 'cpu': */
7201 /* Now try balancing at a lower domain level of 'new_cpu': */
7203 weight = sd->span_weight;
7205 for_each_domain(cpu, tmp) {
7206 if (weight <= tmp->span_weight)
7208 if (tmp->flags & sd_flag)
7216 static inline int __select_idle_cpu(int cpu, struct task_struct *p)
7218 if ((available_idle_cpu(cpu) || sched_idle_cpu(cpu)) &&
7219 sched_cpu_cookie_match(cpu_rq(cpu), p))
7225 #ifdef CONFIG_SCHED_SMT
7226 DEFINE_STATIC_KEY_FALSE(sched_smt_present);
7227 EXPORT_SYMBOL_GPL(sched_smt_present);
7229 static inline void set_idle_cores(int cpu, int val)
7231 struct sched_domain_shared *sds;
7233 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
7235 WRITE_ONCE(sds->has_idle_cores, val);
7238 static inline bool test_idle_cores(int cpu)
7240 struct sched_domain_shared *sds;
7242 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
7244 return READ_ONCE(sds->has_idle_cores);
7250 * Scans the local SMT mask to see if the entire core is idle, and records this
7251 * information in sd_llc_shared->has_idle_cores.
7253 * Since SMT siblings share all cache levels, inspecting this limited remote
7254 * state should be fairly cheap.
7256 void __update_idle_core(struct rq *rq)
7258 int core = cpu_of(rq);
7262 if (test_idle_cores(core))
7265 for_each_cpu(cpu, cpu_smt_mask(core)) {
7269 if (!available_idle_cpu(cpu))
7273 set_idle_cores(core, 1);
7279 * Scan the entire LLC domain for idle cores; this dynamically switches off if
7280 * there are no idle cores left in the system; tracked through
7281 * sd_llc->shared->has_idle_cores and enabled through update_idle_core() above.
7283 static int select_idle_core(struct task_struct *p, int core, struct cpumask *cpus, int *idle_cpu)
7288 for_each_cpu(cpu, cpu_smt_mask(core)) {
7289 if (!available_idle_cpu(cpu)) {
7291 if (*idle_cpu == -1) {
7292 if (sched_idle_cpu(cpu) && cpumask_test_cpu(cpu, cpus)) {
7300 if (*idle_cpu == -1 && cpumask_test_cpu(cpu, cpus))
7307 cpumask_andnot(cpus, cpus, cpu_smt_mask(core));
7312 * Scan the local SMT mask for idle CPUs.
7314 static int select_idle_smt(struct task_struct *p, struct sched_domain *sd, int target)
7318 for_each_cpu_and(cpu, cpu_smt_mask(target), p->cpus_ptr) {
7322 * Check if the CPU is in the LLC scheduling domain of @target.
7323 * Due to isolcpus, there is no guarantee that all the siblings are in the domain.
7325 if (!cpumask_test_cpu(cpu, sched_domain_span(sd)))
7327 if (available_idle_cpu(cpu) || sched_idle_cpu(cpu))
7334 #else /* CONFIG_SCHED_SMT */
7336 static inline void set_idle_cores(int cpu, int val)
7340 static inline bool test_idle_cores(int cpu)
7345 static inline int select_idle_core(struct task_struct *p, int core, struct cpumask *cpus, int *idle_cpu)
7347 return __select_idle_cpu(core, p);
7350 static inline int select_idle_smt(struct task_struct *p, struct sched_domain *sd, int target)
7355 #endif /* CONFIG_SCHED_SMT */
7358 * Scan the LLC domain for idle CPUs; this is dynamically regulated by
7359 * comparing the average scan cost (tracked in sd->avg_scan_cost) against the
7360 * average idle time for this rq (as found in rq->avg_idle).
7362 static int select_idle_cpu(struct task_struct *p, struct sched_domain *sd, bool has_idle_core, int target)
7364 struct cpumask *cpus = this_cpu_cpumask_var_ptr(select_rq_mask);
7365 int i, cpu, idle_cpu = -1, nr = INT_MAX;
7366 struct sched_domain_shared *sd_share;
7368 cpumask_and(cpus, sched_domain_span(sd), p->cpus_ptr);
7370 if (sched_feat(SIS_UTIL)) {
7371 sd_share = rcu_dereference(per_cpu(sd_llc_shared, target));
7373 /* because !--nr is the condition to stop scan */
7374 nr = READ_ONCE(sd_share->nr_idle_scan) + 1;
7375 /* overloaded LLC is unlikely to have idle cpu/core */
7381 if (static_branch_unlikely(&sched_cluster_active)) {
7382 struct sched_group *sg = sd->groups;
7384 if (sg->flags & SD_CLUSTER) {
7385 for_each_cpu_wrap(cpu, sched_group_span(sg), target + 1) {
7386 if (!cpumask_test_cpu(cpu, cpus))
7389 if (has_idle_core) {
7390 i = select_idle_core(p, cpu, cpus, &idle_cpu);
7391 if ((unsigned int)i < nr_cpumask_bits)
7396 idle_cpu = __select_idle_cpu(cpu, p);
7397 if ((unsigned int)idle_cpu < nr_cpumask_bits)
7401 cpumask_andnot(cpus, cpus, sched_group_span(sg));
7405 for_each_cpu_wrap(cpu, cpus, target + 1) {
7406 if (has_idle_core) {
7407 i = select_idle_core(p, cpu, cpus, &idle_cpu);
7408 if ((unsigned int)i < nr_cpumask_bits)
7414 idle_cpu = __select_idle_cpu(cpu, p);
7415 if ((unsigned int)idle_cpu < nr_cpumask_bits)
7421 set_idle_cores(target, false);
7427 * Scan the asym_capacity domain for idle CPUs; pick the first idle one on which
7428 * the task fits. If no CPU is big enough, but there are idle ones, try to
7429 * maximize capacity.
7432 select_idle_capacity(struct task_struct *p, struct sched_domain *sd, int target)
7434 unsigned long task_util, util_min, util_max, best_cap = 0;
7435 int fits, best_fits = 0;
7436 int cpu, best_cpu = -1;
7437 struct cpumask *cpus;
7439 cpus = this_cpu_cpumask_var_ptr(select_rq_mask);
7440 cpumask_and(cpus, sched_domain_span(sd), p->cpus_ptr);
7442 task_util = task_util_est(p);
7443 util_min = uclamp_eff_value(p, UCLAMP_MIN);
7444 util_max = uclamp_eff_value(p, UCLAMP_MAX);
7446 for_each_cpu_wrap(cpu, cpus, target) {
7447 unsigned long cpu_cap = capacity_of(cpu);
7449 if (!available_idle_cpu(cpu) && !sched_idle_cpu(cpu))
7452 fits = util_fits_cpu(task_util, util_min, util_max, cpu);
7454 /* This CPU fits with all requirements */
7458 * Only the min performance hint (i.e. uclamp_min) doesn't fit.
7459 * Look for the CPU with best capacity.
7462 cpu_cap = arch_scale_cpu_capacity(cpu) - thermal_load_avg(cpu_rq(cpu));
7465 * First, select CPU which fits better (-1 being better than 0).
7466 * Then, select the one with best capacity at same level.
7468 if ((fits < best_fits) ||
7469 ((fits == best_fits) && (cpu_cap > best_cap))) {
7479 static inline bool asym_fits_cpu(unsigned long util,
7480 unsigned long util_min,
7481 unsigned long util_max,
7484 if (sched_asym_cpucap_active())
7486 * Return true only if the cpu fully fits the task requirements
7487 * which include the utilization and the performance hints.
7489 return (util_fits_cpu(util, util_min, util_max, cpu) > 0);
7495 * Try and locate an idle core/thread in the LLC cache domain.
7497 static int select_idle_sibling(struct task_struct *p, int prev, int target)
7499 bool has_idle_core = false;
7500 struct sched_domain *sd;
7501 unsigned long task_util, util_min, util_max;
7502 int i, recent_used_cpu, prev_aff = -1;
7505 * On asymmetric system, update task utilization because we will check
7506 * that the task fits with cpu's capacity.
7508 if (sched_asym_cpucap_active()) {
7509 sync_entity_load_avg(&p->se);
7510 task_util = task_util_est(p);
7511 util_min = uclamp_eff_value(p, UCLAMP_MIN);
7512 util_max = uclamp_eff_value(p, UCLAMP_MAX);
7516 * per-cpu select_rq_mask usage
7518 lockdep_assert_irqs_disabled();
7520 if ((available_idle_cpu(target) || sched_idle_cpu(target)) &&
7521 asym_fits_cpu(task_util, util_min, util_max, target))
7525 * If the previous CPU is cache affine and idle, don't be stupid:
7527 if (prev != target && cpus_share_cache(prev, target) &&
7528 (available_idle_cpu(prev) || sched_idle_cpu(prev)) &&
7529 asym_fits_cpu(task_util, util_min, util_max, prev)) {
7531 if (!static_branch_unlikely(&sched_cluster_active) ||
7532 cpus_share_resources(prev, target))
7539 * Allow a per-cpu kthread to stack with the wakee if the
7540 * kworker thread and the tasks previous CPUs are the same.
7541 * The assumption is that the wakee queued work for the
7542 * per-cpu kthread that is now complete and the wakeup is
7543 * essentially a sync wakeup. An obvious example of this
7544 * pattern is IO completions.
7546 if (is_per_cpu_kthread(current) &&
7548 prev == smp_processor_id() &&
7549 this_rq()->nr_running <= 1 &&
7550 asym_fits_cpu(task_util, util_min, util_max, prev)) {
7554 /* Check a recently used CPU as a potential idle candidate: */
7555 recent_used_cpu = p->recent_used_cpu;
7556 p->recent_used_cpu = prev;
7557 if (recent_used_cpu != prev &&
7558 recent_used_cpu != target &&
7559 cpus_share_cache(recent_used_cpu, target) &&
7560 (available_idle_cpu(recent_used_cpu) || sched_idle_cpu(recent_used_cpu)) &&
7561 cpumask_test_cpu(recent_used_cpu, p->cpus_ptr) &&
7562 asym_fits_cpu(task_util, util_min, util_max, recent_used_cpu)) {
7564 if (!static_branch_unlikely(&sched_cluster_active) ||
7565 cpus_share_resources(recent_used_cpu, target))
7566 return recent_used_cpu;
7569 recent_used_cpu = -1;
7573 * For asymmetric CPU capacity systems, our domain of interest is
7574 * sd_asym_cpucapacity rather than sd_llc.
7576 if (sched_asym_cpucap_active()) {
7577 sd = rcu_dereference(per_cpu(sd_asym_cpucapacity, target));
7579 * On an asymmetric CPU capacity system where an exclusive
7580 * cpuset defines a symmetric island (i.e. one unique
7581 * capacity_orig value through the cpuset), the key will be set
7582 * but the CPUs within that cpuset will not have a domain with
7583 * SD_ASYM_CPUCAPACITY. These should follow the usual symmetric
7587 i = select_idle_capacity(p, sd, target);
7588 return ((unsigned)i < nr_cpumask_bits) ? i : target;
7592 sd = rcu_dereference(per_cpu(sd_llc, target));
7596 if (sched_smt_active()) {
7597 has_idle_core = test_idle_cores(target);
7599 if (!has_idle_core && cpus_share_cache(prev, target)) {
7600 i = select_idle_smt(p, sd, prev);
7601 if ((unsigned int)i < nr_cpumask_bits)
7606 i = select_idle_cpu(p, sd, has_idle_core, target);
7607 if ((unsigned)i < nr_cpumask_bits)
7611 * For cluster machines which have lower sharing cache like L2 or
7612 * LLC Tag, we tend to find an idle CPU in the target's cluster
7613 * first. But prev_cpu or recent_used_cpu may also be a good candidate,
7614 * use them if possible when no idle CPU found in select_idle_cpu().
7616 if ((unsigned int)prev_aff < nr_cpumask_bits)
7618 if ((unsigned int)recent_used_cpu < nr_cpumask_bits)
7619 return recent_used_cpu;
7625 * cpu_util() - Estimates the amount of CPU capacity used by CFS tasks.
7626 * @cpu: the CPU to get the utilization for
7627 * @p: task for which the CPU utilization should be predicted or NULL
7628 * @dst_cpu: CPU @p migrates to, -1 if @p moves from @cpu or @p == NULL
7629 * @boost: 1 to enable boosting, otherwise 0
7631 * The unit of the return value must be the same as the one of CPU capacity
7632 * so that CPU utilization can be compared with CPU capacity.
7634 * CPU utilization is the sum of running time of runnable tasks plus the
7635 * recent utilization of currently non-runnable tasks on that CPU.
7636 * It represents the amount of CPU capacity currently used by CFS tasks in
7637 * the range [0..max CPU capacity] with max CPU capacity being the CPU
7638 * capacity at f_max.
7640 * The estimated CPU utilization is defined as the maximum between CPU
7641 * utilization and sum of the estimated utilization of the currently
7642 * runnable tasks on that CPU. It preserves a utilization "snapshot" of
7643 * previously-executed tasks, which helps better deduce how busy a CPU will
7644 * be when a long-sleeping task wakes up. The contribution to CPU utilization
7645 * of such a task would be significantly decayed at this point of time.
7647 * Boosted CPU utilization is defined as max(CPU runnable, CPU utilization).
7648 * CPU contention for CFS tasks can be detected by CPU runnable > CPU
7649 * utilization. Boosting is implemented in cpu_util() so that internal
7650 * users (e.g. EAS) can use it next to external users (e.g. schedutil),
7651 * latter via cpu_util_cfs_boost().
7653 * CPU utilization can be higher than the current CPU capacity
7654 * (f_curr/f_max * max CPU capacity) or even the max CPU capacity because
7655 * of rounding errors as well as task migrations or wakeups of new tasks.
7656 * CPU utilization has to be capped to fit into the [0..max CPU capacity]
7657 * range. Otherwise a group of CPUs (CPU0 util = 121% + CPU1 util = 80%)
7658 * could be seen as over-utilized even though CPU1 has 20% of spare CPU
7659 * capacity. CPU utilization is allowed to overshoot current CPU capacity
7660 * though since this is useful for predicting the CPU capacity required
7661 * after task migrations (scheduler-driven DVFS).
7663 * Return: (Boosted) (estimated) utilization for the specified CPU.
7665 static unsigned long
7666 cpu_util(int cpu, struct task_struct *p, int dst_cpu, int boost)
7668 struct cfs_rq *cfs_rq = &cpu_rq(cpu)->cfs;
7669 unsigned long util = READ_ONCE(cfs_rq->avg.util_avg);
7670 unsigned long runnable;
7673 runnable = READ_ONCE(cfs_rq->avg.runnable_avg);
7674 util = max(util, runnable);
7678 * If @dst_cpu is -1 or @p migrates from @cpu to @dst_cpu remove its
7679 * contribution. If @p migrates from another CPU to @cpu add its
7680 * contribution. In all the other cases @cpu is not impacted by the
7681 * migration so its util_avg is already correct.
7683 if (p && task_cpu(p) == cpu && dst_cpu != cpu)
7684 lsub_positive(&util, task_util(p));
7685 else if (p && task_cpu(p) != cpu && dst_cpu == cpu)
7686 util += task_util(p);
7688 if (sched_feat(UTIL_EST)) {
7689 unsigned long util_est;
7691 util_est = READ_ONCE(cfs_rq->avg.util_est);
7694 * During wake-up @p isn't enqueued yet and doesn't contribute
7695 * to any cpu_rq(cpu)->cfs.avg.util_est.
7696 * If @dst_cpu == @cpu add it to "simulate" cpu_util after @p
7697 * has been enqueued.
7699 * During exec (@dst_cpu = -1) @p is enqueued and does
7700 * contribute to cpu_rq(cpu)->cfs.util_est.
7701 * Remove it to "simulate" cpu_util without @p's contribution.
7703 * Despite the task_on_rq_queued(@p) check there is still a
7704 * small window for a possible race when an exec
7705 * select_task_rq_fair() races with LB's detach_task().
7709 * p->on_rq = TASK_ON_RQ_MIGRATING;
7710 * -------------------------------- A
7712 * dequeue_task_fair() + Race Time
7713 * util_est_dequeue() /
7714 * -------------------------------- B
7716 * The additional check "current == p" is required to further
7717 * reduce the race window.
7720 util_est += _task_util_est(p);
7721 else if (p && unlikely(task_on_rq_queued(p) || current == p))
7722 lsub_positive(&util_est, _task_util_est(p));
7724 util = max(util, util_est);
7727 return min(util, arch_scale_cpu_capacity(cpu));
7730 unsigned long cpu_util_cfs(int cpu)
7732 return cpu_util(cpu, NULL, -1, 0);
7735 unsigned long cpu_util_cfs_boost(int cpu)
7737 return cpu_util(cpu, NULL, -1, 1);
7741 * cpu_util_without: compute cpu utilization without any contributions from *p
7742 * @cpu: the CPU which utilization is requested
7743 * @p: the task which utilization should be discounted
7745 * The utilization of a CPU is defined by the utilization of tasks currently
7746 * enqueued on that CPU as well as tasks which are currently sleeping after an
7747 * execution on that CPU.
7749 * This method returns the utilization of the specified CPU by discounting the
7750 * utilization of the specified task, whenever the task is currently
7751 * contributing to the CPU utilization.
7753 static unsigned long cpu_util_without(int cpu, struct task_struct *p)
7755 /* Task has no contribution or is new */
7756 if (cpu != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time))
7759 return cpu_util(cpu, p, -1, 0);
7763 * energy_env - Utilization landscape for energy estimation.
7764 * @task_busy_time: Utilization contribution by the task for which we test the
7765 * placement. Given by eenv_task_busy_time().
7766 * @pd_busy_time: Utilization of the whole perf domain without the task
7767 * contribution. Given by eenv_pd_busy_time().
7768 * @cpu_cap: Maximum CPU capacity for the perf domain.
7769 * @pd_cap: Entire perf domain capacity. (pd->nr_cpus * cpu_cap).
7772 unsigned long task_busy_time;
7773 unsigned long pd_busy_time;
7774 unsigned long cpu_cap;
7775 unsigned long pd_cap;
7779 * Compute the task busy time for compute_energy(). This time cannot be
7780 * injected directly into effective_cpu_util() because of the IRQ scaling.
7781 * The latter only makes sense with the most recent CPUs where the task has
7784 static inline void eenv_task_busy_time(struct energy_env *eenv,
7785 struct task_struct *p, int prev_cpu)
7787 unsigned long busy_time, max_cap = arch_scale_cpu_capacity(prev_cpu);
7788 unsigned long irq = cpu_util_irq(cpu_rq(prev_cpu));
7790 if (unlikely(irq >= max_cap))
7791 busy_time = max_cap;
7793 busy_time = scale_irq_capacity(task_util_est(p), irq, max_cap);
7795 eenv->task_busy_time = busy_time;
7799 * Compute the perf_domain (PD) busy time for compute_energy(). Based on the
7800 * utilization for each @pd_cpus, it however doesn't take into account
7801 * clamping since the ratio (utilization / cpu_capacity) is already enough to
7802 * scale the EM reported power consumption at the (eventually clamped)
7805 * The contribution of the task @p for which we want to estimate the
7806 * energy cost is removed (by cpu_util()) and must be calculated
7807 * separately (see eenv_task_busy_time). This ensures:
7809 * - A stable PD utilization, no matter which CPU of that PD we want to place
7812 * - A fair comparison between CPUs as the task contribution (task_util())
7813 * will always be the same no matter which CPU utilization we rely on
7814 * (util_avg or util_est).
7816 * Set @eenv busy time for the PD that spans @pd_cpus. This busy time can't
7817 * exceed @eenv->pd_cap.
7819 static inline void eenv_pd_busy_time(struct energy_env *eenv,
7820 struct cpumask *pd_cpus,
7821 struct task_struct *p)
7823 unsigned long busy_time = 0;
7826 for_each_cpu(cpu, pd_cpus) {
7827 unsigned long util = cpu_util(cpu, p, -1, 0);
7829 busy_time += effective_cpu_util(cpu, util, NULL, NULL);
7832 eenv->pd_busy_time = min(eenv->pd_cap, busy_time);
7836 * Compute the maximum utilization for compute_energy() when the task @p
7837 * is placed on the cpu @dst_cpu.
7839 * Returns the maximum utilization among @eenv->cpus. This utilization can't
7840 * exceed @eenv->cpu_cap.
7842 static inline unsigned long
7843 eenv_pd_max_util(struct energy_env *eenv, struct cpumask *pd_cpus,
7844 struct task_struct *p, int dst_cpu)
7846 unsigned long max_util = 0;
7849 for_each_cpu(cpu, pd_cpus) {
7850 struct task_struct *tsk = (cpu == dst_cpu) ? p : NULL;
7851 unsigned long util = cpu_util(cpu, p, dst_cpu, 1);
7852 unsigned long eff_util, min, max;
7855 * Performance domain frequency: utilization clamping
7856 * must be considered since it affects the selection
7857 * of the performance domain frequency.
7858 * NOTE: in case RT tasks are running, by default the
7859 * FREQUENCY_UTIL's utilization can be max OPP.
7861 eff_util = effective_cpu_util(cpu, util, &min, &max);
7863 /* Task's uclamp can modify min and max value */
7864 if (tsk && uclamp_is_used()) {
7865 min = max(min, uclamp_eff_value(p, UCLAMP_MIN));
7868 * If there is no active max uclamp constraint,
7869 * directly use task's one, otherwise keep max.
7871 if (uclamp_rq_is_idle(cpu_rq(cpu)))
7872 max = uclamp_eff_value(p, UCLAMP_MAX);
7874 max = max(max, uclamp_eff_value(p, UCLAMP_MAX));
7877 eff_util = sugov_effective_cpu_perf(cpu, eff_util, min, max);
7878 max_util = max(max_util, eff_util);
7881 return min(max_util, eenv->cpu_cap);
7885 * compute_energy(): Use the Energy Model to estimate the energy that @pd would
7886 * consume for a given utilization landscape @eenv. When @dst_cpu < 0, the task
7887 * contribution is ignored.
7889 static inline unsigned long
7890 compute_energy(struct energy_env *eenv, struct perf_domain *pd,
7891 struct cpumask *pd_cpus, struct task_struct *p, int dst_cpu)
7893 unsigned long max_util = eenv_pd_max_util(eenv, pd_cpus, p, dst_cpu);
7894 unsigned long busy_time = eenv->pd_busy_time;
7895 unsigned long energy;
7898 busy_time = min(eenv->pd_cap, busy_time + eenv->task_busy_time);
7900 energy = em_cpu_energy(pd->em_pd, max_util, busy_time, eenv->cpu_cap);
7902 trace_sched_compute_energy_tp(p, dst_cpu, energy, max_util, busy_time);
7908 * find_energy_efficient_cpu(): Find most energy-efficient target CPU for the
7909 * waking task. find_energy_efficient_cpu() looks for the CPU with maximum
7910 * spare capacity in each performance domain and uses it as a potential
7911 * candidate to execute the task. Then, it uses the Energy Model to figure
7912 * out which of the CPU candidates is the most energy-efficient.
7914 * The rationale for this heuristic is as follows. In a performance domain,
7915 * all the most energy efficient CPU candidates (according to the Energy
7916 * Model) are those for which we'll request a low frequency. When there are
7917 * several CPUs for which the frequency request will be the same, we don't
7918 * have enough data to break the tie between them, because the Energy Model
7919 * only includes active power costs. With this model, if we assume that
7920 * frequency requests follow utilization (e.g. using schedutil), the CPU with
7921 * the maximum spare capacity in a performance domain is guaranteed to be among
7922 * the best candidates of the performance domain.
7924 * In practice, it could be preferable from an energy standpoint to pack
7925 * small tasks on a CPU in order to let other CPUs go in deeper idle states,
7926 * but that could also hurt our chances to go cluster idle, and we have no
7927 * ways to tell with the current Energy Model if this is actually a good
7928 * idea or not. So, find_energy_efficient_cpu() basically favors
7929 * cluster-packing, and spreading inside a cluster. That should at least be
7930 * a good thing for latency, and this is consistent with the idea that most
7931 * of the energy savings of EAS come from the asymmetry of the system, and
7932 * not so much from breaking the tie between identical CPUs. That's also the
7933 * reason why EAS is enabled in the topology code only for systems where
7934 * SD_ASYM_CPUCAPACITY is set.
7936 * NOTE: Forkees are not accepted in the energy-aware wake-up path because
7937 * they don't have any useful utilization data yet and it's not possible to
7938 * forecast their impact on energy consumption. Consequently, they will be
7939 * placed by find_idlest_cpu() on the least loaded CPU, which might turn out
7940 * to be energy-inefficient in some use-cases. The alternative would be to
7941 * bias new tasks towards specific types of CPUs first, or to try to infer
7942 * their util_avg from the parent task, but those heuristics could hurt
7943 * other use-cases too. So, until someone finds a better way to solve this,
7944 * let's keep things simple by re-using the existing slow path.
7946 static int find_energy_efficient_cpu(struct task_struct *p, int prev_cpu)
7948 struct cpumask *cpus = this_cpu_cpumask_var_ptr(select_rq_mask);
7949 unsigned long prev_delta = ULONG_MAX, best_delta = ULONG_MAX;
7950 unsigned long p_util_min = uclamp_is_used() ? uclamp_eff_value(p, UCLAMP_MIN) : 0;
7951 unsigned long p_util_max = uclamp_is_used() ? uclamp_eff_value(p, UCLAMP_MAX) : 1024;
7952 struct root_domain *rd = this_rq()->rd;
7953 int cpu, best_energy_cpu, target = -1;
7954 int prev_fits = -1, best_fits = -1;
7955 unsigned long best_thermal_cap = 0;
7956 unsigned long prev_thermal_cap = 0;
7957 struct sched_domain *sd;
7958 struct perf_domain *pd;
7959 struct energy_env eenv;
7962 pd = rcu_dereference(rd->pd);
7963 if (!pd || READ_ONCE(rd->overutilized))
7967 * Energy-aware wake-up happens on the lowest sched_domain starting
7968 * from sd_asym_cpucapacity spanning over this_cpu and prev_cpu.
7970 sd = rcu_dereference(*this_cpu_ptr(&sd_asym_cpucapacity));
7971 while (sd && !cpumask_test_cpu(prev_cpu, sched_domain_span(sd)))
7978 sync_entity_load_avg(&p->se);
7979 if (!task_util_est(p) && p_util_min == 0)
7982 eenv_task_busy_time(&eenv, p, prev_cpu);
7984 for (; pd; pd = pd->next) {
7985 unsigned long util_min = p_util_min, util_max = p_util_max;
7986 unsigned long cpu_cap, cpu_thermal_cap, util;
7987 long prev_spare_cap = -1, max_spare_cap = -1;
7988 unsigned long rq_util_min, rq_util_max;
7989 unsigned long cur_delta, base_energy;
7990 int max_spare_cap_cpu = -1;
7991 int fits, max_fits = -1;
7993 cpumask_and(cpus, perf_domain_span(pd), cpu_online_mask);
7995 if (cpumask_empty(cpus))
7998 /* Account thermal pressure for the energy estimation */
7999 cpu = cpumask_first(cpus);
8000 cpu_thermal_cap = arch_scale_cpu_capacity(cpu);
8001 cpu_thermal_cap -= arch_scale_thermal_pressure(cpu);
8003 eenv.cpu_cap = cpu_thermal_cap;
8006 for_each_cpu(cpu, cpus) {
8007 struct rq *rq = cpu_rq(cpu);
8009 eenv.pd_cap += cpu_thermal_cap;
8011 if (!cpumask_test_cpu(cpu, sched_domain_span(sd)))
8014 if (!cpumask_test_cpu(cpu, p->cpus_ptr))
8017 util = cpu_util(cpu, p, cpu, 0);
8018 cpu_cap = capacity_of(cpu);
8021 * Skip CPUs that cannot satisfy the capacity request.
8022 * IOW, placing the task there would make the CPU
8023 * overutilized. Take uclamp into account to see how
8024 * much capacity we can get out of the CPU; this is
8025 * aligned with sched_cpu_util().
8027 if (uclamp_is_used() && !uclamp_rq_is_idle(rq)) {
8029 * Open code uclamp_rq_util_with() except for
8030 * the clamp() part. Ie: apply max aggregation
8031 * only. util_fits_cpu() logic requires to
8032 * operate on non clamped util but must use the
8033 * max-aggregated uclamp_{min, max}.
8035 rq_util_min = uclamp_rq_get(rq, UCLAMP_MIN);
8036 rq_util_max = uclamp_rq_get(rq, UCLAMP_MAX);
8038 util_min = max(rq_util_min, p_util_min);
8039 util_max = max(rq_util_max, p_util_max);
8042 fits = util_fits_cpu(util, util_min, util_max, cpu);
8046 lsub_positive(&cpu_cap, util);
8048 if (cpu == prev_cpu) {
8049 /* Always use prev_cpu as a candidate. */
8050 prev_spare_cap = cpu_cap;
8052 } else if ((fits > max_fits) ||
8053 ((fits == max_fits) && ((long)cpu_cap > max_spare_cap))) {
8055 * Find the CPU with the maximum spare capacity
8056 * among the remaining CPUs in the performance
8059 max_spare_cap = cpu_cap;
8060 max_spare_cap_cpu = cpu;
8065 if (max_spare_cap_cpu < 0 && prev_spare_cap < 0)
8068 eenv_pd_busy_time(&eenv, cpus, p);
8069 /* Compute the 'base' energy of the pd, without @p */
8070 base_energy = compute_energy(&eenv, pd, cpus, p, -1);
8072 /* Evaluate the energy impact of using prev_cpu. */
8073 if (prev_spare_cap > -1) {
8074 prev_delta = compute_energy(&eenv, pd, cpus, p,
8076 /* CPU utilization has changed */
8077 if (prev_delta < base_energy)
8079 prev_delta -= base_energy;
8080 prev_thermal_cap = cpu_thermal_cap;
8081 best_delta = min(best_delta, prev_delta);
8084 /* Evaluate the energy impact of using max_spare_cap_cpu. */
8085 if (max_spare_cap_cpu >= 0 && max_spare_cap > prev_spare_cap) {
8086 /* Current best energy cpu fits better */
8087 if (max_fits < best_fits)
8091 * Both don't fit performance hint (i.e. uclamp_min)
8092 * but best energy cpu has better capacity.
8094 if ((max_fits < 0) &&
8095 (cpu_thermal_cap <= best_thermal_cap))
8098 cur_delta = compute_energy(&eenv, pd, cpus, p,
8100 /* CPU utilization has changed */
8101 if (cur_delta < base_energy)
8103 cur_delta -= base_energy;
8106 * Both fit for the task but best energy cpu has lower
8109 if ((max_fits > 0) && (best_fits > 0) &&
8110 (cur_delta >= best_delta))
8113 best_delta = cur_delta;
8114 best_energy_cpu = max_spare_cap_cpu;
8115 best_fits = max_fits;
8116 best_thermal_cap = cpu_thermal_cap;
8121 if ((best_fits > prev_fits) ||
8122 ((best_fits > 0) && (best_delta < prev_delta)) ||
8123 ((best_fits < 0) && (best_thermal_cap > prev_thermal_cap)))
8124 target = best_energy_cpu;
8135 * select_task_rq_fair: Select target runqueue for the waking task in domains
8136 * that have the relevant SD flag set. In practice, this is SD_BALANCE_WAKE,
8137 * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
8139 * Balances load by selecting the idlest CPU in the idlest group, or under
8140 * certain conditions an idle sibling CPU if the domain has SD_WAKE_AFFINE set.
8142 * Returns the target CPU number.
8145 select_task_rq_fair(struct task_struct *p, int prev_cpu, int wake_flags)
8147 int sync = (wake_flags & WF_SYNC) && !(current->flags & PF_EXITING);
8148 struct sched_domain *tmp, *sd = NULL;
8149 int cpu = smp_processor_id();
8150 int new_cpu = prev_cpu;
8151 int want_affine = 0;
8152 /* SD_flags and WF_flags share the first nibble */
8153 int sd_flag = wake_flags & 0xF;
8156 * required for stable ->cpus_allowed
8158 lockdep_assert_held(&p->pi_lock);
8159 if (wake_flags & WF_TTWU) {
8162 if ((wake_flags & WF_CURRENT_CPU) &&
8163 cpumask_test_cpu(cpu, p->cpus_ptr))
8166 if (sched_energy_enabled()) {
8167 new_cpu = find_energy_efficient_cpu(p, prev_cpu);
8173 want_affine = !wake_wide(p) && cpumask_test_cpu(cpu, p->cpus_ptr);
8177 for_each_domain(cpu, tmp) {
8179 * If both 'cpu' and 'prev_cpu' are part of this domain,
8180 * cpu is a valid SD_WAKE_AFFINE target.
8182 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
8183 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
8184 if (cpu != prev_cpu)
8185 new_cpu = wake_affine(tmp, p, cpu, prev_cpu, sync);
8187 sd = NULL; /* Prefer wake_affine over balance flags */
8192 * Usually only true for WF_EXEC and WF_FORK, as sched_domains
8193 * usually do not have SD_BALANCE_WAKE set. That means wakeup
8194 * will usually go to the fast path.
8196 if (tmp->flags & sd_flag)
8198 else if (!want_affine)
8204 new_cpu = find_idlest_cpu(sd, p, cpu, prev_cpu, sd_flag);
8205 } else if (wake_flags & WF_TTWU) { /* XXX always ? */
8207 new_cpu = select_idle_sibling(p, prev_cpu, new_cpu);
8215 * Called immediately before a task is migrated to a new CPU; task_cpu(p) and
8216 * cfs_rq_of(p) references at time of call are still valid and identify the
8217 * previous CPU. The caller guarantees p->pi_lock or task_rq(p)->lock is held.
8219 static void migrate_task_rq_fair(struct task_struct *p, int new_cpu)
8221 struct sched_entity *se = &p->se;
8223 if (!task_on_rq_migrating(p)) {
8224 remove_entity_load_avg(se);
8227 * Here, the task's PELT values have been updated according to
8228 * the current rq's clock. But if that clock hasn't been
8229 * updated in a while, a substantial idle time will be missed,
8230 * leading to an inflation after wake-up on the new rq.
8232 * Estimate the missing time from the cfs_rq last_update_time
8233 * and update sched_avg to improve the PELT continuity after
8236 migrate_se_pelt_lag(se);
8239 /* Tell new CPU we are migrated */
8240 se->avg.last_update_time = 0;
8242 update_scan_period(p, new_cpu);
8245 static void task_dead_fair(struct task_struct *p)
8247 remove_entity_load_avg(&p->se);
8251 balance_fair(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
8256 return newidle_balance(rq, rf) != 0;
8258 #endif /* CONFIG_SMP */
8260 static void set_next_buddy(struct sched_entity *se)
8262 for_each_sched_entity(se) {
8263 if (SCHED_WARN_ON(!se->on_rq))
8267 cfs_rq_of(se)->next = se;
8272 * Preempt the current task with a newly woken task if needed:
8274 static void check_preempt_wakeup_fair(struct rq *rq, struct task_struct *p, int wake_flags)
8276 struct task_struct *curr = rq->curr;
8277 struct sched_entity *se = &curr->se, *pse = &p->se;
8278 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
8279 int cse_is_idle, pse_is_idle;
8281 if (unlikely(se == pse))
8285 * This is possible from callers such as attach_tasks(), in which we
8286 * unconditionally wakeup_preempt() after an enqueue (which may have
8287 * lead to a throttle). This both saves work and prevents false
8288 * next-buddy nomination below.
8290 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
8293 if (sched_feat(NEXT_BUDDY) && !(wake_flags & WF_FORK)) {
8294 set_next_buddy(pse);
8298 * We can come here with TIF_NEED_RESCHED already set from new task
8301 * Note: this also catches the edge-case of curr being in a throttled
8302 * group (e.g. via set_curr_task), since update_curr() (in the
8303 * enqueue of curr) will have resulted in resched being set. This
8304 * prevents us from potentially nominating it as a false LAST_BUDDY
8307 if (test_tsk_need_resched(curr))
8310 /* Idle tasks are by definition preempted by non-idle tasks. */
8311 if (unlikely(task_has_idle_policy(curr)) &&
8312 likely(!task_has_idle_policy(p)))
8316 * Batch and idle tasks do not preempt non-idle tasks (their preemption
8317 * is driven by the tick):
8319 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
8322 find_matching_se(&se, &pse);
8325 cse_is_idle = se_is_idle(se);
8326 pse_is_idle = se_is_idle(pse);
8329 * Preempt an idle group in favor of a non-idle group (and don't preempt
8330 * in the inverse case).
8332 if (cse_is_idle && !pse_is_idle)
8334 if (cse_is_idle != pse_is_idle)
8337 cfs_rq = cfs_rq_of(se);
8338 update_curr(cfs_rq);
8341 * XXX pick_eevdf(cfs_rq) != se ?
8343 if (pick_eevdf(cfs_rq) == pse)
8353 static struct task_struct *pick_task_fair(struct rq *rq)
8355 struct sched_entity *se;
8356 struct cfs_rq *cfs_rq;
8360 if (!cfs_rq->nr_running)
8364 struct sched_entity *curr = cfs_rq->curr;
8366 /* When we pick for a remote RQ, we'll not have done put_prev_entity() */
8369 update_curr(cfs_rq);
8373 if (unlikely(check_cfs_rq_runtime(cfs_rq)))
8377 se = pick_next_entity(cfs_rq);
8378 cfs_rq = group_cfs_rq(se);
8385 struct task_struct *
8386 pick_next_task_fair(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
8388 struct cfs_rq *cfs_rq = &rq->cfs;
8389 struct sched_entity *se;
8390 struct task_struct *p;
8394 if (!sched_fair_runnable(rq))
8397 #ifdef CONFIG_FAIR_GROUP_SCHED
8398 if (!prev || prev->sched_class != &fair_sched_class)
8402 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
8403 * likely that a next task is from the same cgroup as the current.
8405 * Therefore attempt to avoid putting and setting the entire cgroup
8406 * hierarchy, only change the part that actually changes.
8410 struct sched_entity *curr = cfs_rq->curr;
8413 * Since we got here without doing put_prev_entity() we also
8414 * have to consider cfs_rq->curr. If it is still a runnable
8415 * entity, update_curr() will update its vruntime, otherwise
8416 * forget we've ever seen it.
8420 update_curr(cfs_rq);
8425 * This call to check_cfs_rq_runtime() will do the
8426 * throttle and dequeue its entity in the parent(s).
8427 * Therefore the nr_running test will indeed
8430 if (unlikely(check_cfs_rq_runtime(cfs_rq))) {
8433 if (!cfs_rq->nr_running)
8440 se = pick_next_entity(cfs_rq);
8441 cfs_rq = group_cfs_rq(se);
8447 * Since we haven't yet done put_prev_entity and if the selected task
8448 * is a different task than we started out with, try and touch the
8449 * least amount of cfs_rqs.
8452 struct sched_entity *pse = &prev->se;
8454 while (!(cfs_rq = is_same_group(se, pse))) {
8455 int se_depth = se->depth;
8456 int pse_depth = pse->depth;
8458 if (se_depth <= pse_depth) {
8459 put_prev_entity(cfs_rq_of(pse), pse);
8460 pse = parent_entity(pse);
8462 if (se_depth >= pse_depth) {
8463 set_next_entity(cfs_rq_of(se), se);
8464 se = parent_entity(se);
8468 put_prev_entity(cfs_rq, pse);
8469 set_next_entity(cfs_rq, se);
8476 put_prev_task(rq, prev);
8479 se = pick_next_entity(cfs_rq);
8480 set_next_entity(cfs_rq, se);
8481 cfs_rq = group_cfs_rq(se);
8486 done: __maybe_unused;
8489 * Move the next running task to the front of
8490 * the list, so our cfs_tasks list becomes MRU
8493 list_move(&p->se.group_node, &rq->cfs_tasks);
8496 if (hrtick_enabled_fair(rq))
8497 hrtick_start_fair(rq, p);
8499 update_misfit_status(p, rq);
8500 sched_fair_update_stop_tick(rq, p);
8508 new_tasks = newidle_balance(rq, rf);
8511 * Because newidle_balance() releases (and re-acquires) rq->lock, it is
8512 * possible for any higher priority task to appear. In that case we
8513 * must re-start the pick_next_entity() loop.
8522 * rq is about to be idle, check if we need to update the
8523 * lost_idle_time of clock_pelt
8525 update_idle_rq_clock_pelt(rq);
8530 static struct task_struct *__pick_next_task_fair(struct rq *rq)
8532 return pick_next_task_fair(rq, NULL, NULL);
8536 * Account for a descheduled task:
8538 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
8540 struct sched_entity *se = &prev->se;
8541 struct cfs_rq *cfs_rq;
8543 for_each_sched_entity(se) {
8544 cfs_rq = cfs_rq_of(se);
8545 put_prev_entity(cfs_rq, se);
8550 * sched_yield() is very simple
8552 static void yield_task_fair(struct rq *rq)
8554 struct task_struct *curr = rq->curr;
8555 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
8556 struct sched_entity *se = &curr->se;
8559 * Are we the only task in the tree?
8561 if (unlikely(rq->nr_running == 1))
8564 clear_buddies(cfs_rq, se);
8566 update_rq_clock(rq);
8568 * Update run-time statistics of the 'current'.
8570 update_curr(cfs_rq);
8572 * Tell update_rq_clock() that we've just updated,
8573 * so we don't do microscopic update in schedule()
8574 * and double the fastpath cost.
8576 rq_clock_skip_update(rq);
8578 se->deadline += calc_delta_fair(se->slice, se);
8581 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p)
8583 struct sched_entity *se = &p->se;
8585 /* throttled hierarchies are not runnable */
8586 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
8589 /* Tell the scheduler that we'd really like pse to run next. */
8592 yield_task_fair(rq);
8598 /**************************************************
8599 * Fair scheduling class load-balancing methods.
8603 * The purpose of load-balancing is to achieve the same basic fairness the
8604 * per-CPU scheduler provides, namely provide a proportional amount of compute
8605 * time to each task. This is expressed in the following equation:
8607 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
8609 * Where W_i,n is the n-th weight average for CPU i. The instantaneous weight
8610 * W_i,0 is defined as:
8612 * W_i,0 = \Sum_j w_i,j (2)
8614 * Where w_i,j is the weight of the j-th runnable task on CPU i. This weight
8615 * is derived from the nice value as per sched_prio_to_weight[].
8617 * The weight average is an exponential decay average of the instantaneous
8620 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
8622 * C_i is the compute capacity of CPU i, typically it is the
8623 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
8624 * can also include other factors [XXX].
8626 * To achieve this balance we define a measure of imbalance which follows
8627 * directly from (1):
8629 * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4)
8631 * We them move tasks around to minimize the imbalance. In the continuous
8632 * function space it is obvious this converges, in the discrete case we get
8633 * a few fun cases generally called infeasible weight scenarios.
8636 * - infeasible weights;
8637 * - local vs global optima in the discrete case. ]
8642 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
8643 * for all i,j solution, we create a tree of CPUs that follows the hardware
8644 * topology where each level pairs two lower groups (or better). This results
8645 * in O(log n) layers. Furthermore we reduce the number of CPUs going up the
8646 * tree to only the first of the previous level and we decrease the frequency
8647 * of load-balance at each level inv. proportional to the number of CPUs in
8653 * \Sum { --- * --- * 2^i } = O(n) (5)
8655 * `- size of each group
8656 * | | `- number of CPUs doing load-balance
8658 * `- sum over all levels
8660 * Coupled with a limit on how many tasks we can migrate every balance pass,
8661 * this makes (5) the runtime complexity of the balancer.
8663 * An important property here is that each CPU is still (indirectly) connected
8664 * to every other CPU in at most O(log n) steps:
8666 * The adjacency matrix of the resulting graph is given by:
8669 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
8672 * And you'll find that:
8674 * A^(log_2 n)_i,j != 0 for all i,j (7)
8676 * Showing there's indeed a path between every CPU in at most O(log n) steps.
8677 * The task movement gives a factor of O(m), giving a convergence complexity
8680 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
8685 * In order to avoid CPUs going idle while there's still work to do, new idle
8686 * balancing is more aggressive and has the newly idle CPU iterate up the domain
8687 * tree itself instead of relying on other CPUs to bring it work.
8689 * This adds some complexity to both (5) and (8) but it reduces the total idle
8697 * Cgroups make a horror show out of (2), instead of a simple sum we get:
8700 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
8705 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
8707 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on CPU i.
8709 * The big problem is S_k, its a global sum needed to compute a local (W_i)
8712 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
8713 * rewrite all of this once again.]
8716 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
8718 enum fbq_type { regular, remote, all };
8721 * 'group_type' describes the group of CPUs at the moment of load balancing.
8723 * The enum is ordered by pulling priority, with the group with lowest priority
8724 * first so the group_type can simply be compared when selecting the busiest
8725 * group. See update_sd_pick_busiest().
8728 /* The group has spare capacity that can be used to run more tasks. */
8729 group_has_spare = 0,
8731 * The group is fully used and the tasks don't compete for more CPU
8732 * cycles. Nevertheless, some tasks might wait before running.
8736 * One task doesn't fit with CPU's capacity and must be migrated to a
8737 * more powerful CPU.
8741 * Balance SMT group that's fully busy. Can benefit from migration
8742 * a task on SMT with busy sibling to another CPU on idle core.
8746 * SD_ASYM_PACKING only: One local CPU with higher capacity is available,
8747 * and the task should be migrated to it instead of running on the
8752 * The tasks' affinity constraints previously prevented the scheduler
8753 * from balancing the load across the system.
8757 * The CPU is overloaded and can't provide expected CPU cycles to all
8763 enum migration_type {
8770 #define LBF_ALL_PINNED 0x01
8771 #define LBF_NEED_BREAK 0x02
8772 #define LBF_DST_PINNED 0x04
8773 #define LBF_SOME_PINNED 0x08
8774 #define LBF_ACTIVE_LB 0x10
8777 struct sched_domain *sd;
8785 struct cpumask *dst_grpmask;
8787 enum cpu_idle_type idle;
8789 /* The set of CPUs under consideration for load-balancing */
8790 struct cpumask *cpus;
8795 unsigned int loop_break;
8796 unsigned int loop_max;
8798 enum fbq_type fbq_type;
8799 enum migration_type migration_type;
8800 struct list_head tasks;
8804 * Is this task likely cache-hot:
8806 static int task_hot(struct task_struct *p, struct lb_env *env)
8810 lockdep_assert_rq_held(env->src_rq);
8812 if (p->sched_class != &fair_sched_class)
8815 if (unlikely(task_has_idle_policy(p)))
8818 /* SMT siblings share cache */
8819 if (env->sd->flags & SD_SHARE_CPUCAPACITY)
8823 * Buddy candidates are cache hot:
8825 if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running &&
8826 (&p->se == cfs_rq_of(&p->se)->next))
8829 if (sysctl_sched_migration_cost == -1)
8833 * Don't migrate task if the task's cookie does not match
8834 * with the destination CPU's core cookie.
8836 if (!sched_core_cookie_match(cpu_rq(env->dst_cpu), p))
8839 if (sysctl_sched_migration_cost == 0)
8842 delta = rq_clock_task(env->src_rq) - p->se.exec_start;
8844 return delta < (s64)sysctl_sched_migration_cost;
8847 #ifdef CONFIG_NUMA_BALANCING
8849 * Returns 1, if task migration degrades locality
8850 * Returns 0, if task migration improves locality i.e migration preferred.
8851 * Returns -1, if task migration is not affected by locality.
8853 static int migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
8855 struct numa_group *numa_group = rcu_dereference(p->numa_group);
8856 unsigned long src_weight, dst_weight;
8857 int src_nid, dst_nid, dist;
8859 if (!static_branch_likely(&sched_numa_balancing))
8862 if (!p->numa_faults || !(env->sd->flags & SD_NUMA))
8865 src_nid = cpu_to_node(env->src_cpu);
8866 dst_nid = cpu_to_node(env->dst_cpu);
8868 if (src_nid == dst_nid)
8871 /* Migrating away from the preferred node is always bad. */
8872 if (src_nid == p->numa_preferred_nid) {
8873 if (env->src_rq->nr_running > env->src_rq->nr_preferred_running)
8879 /* Encourage migration to the preferred node. */
8880 if (dst_nid == p->numa_preferred_nid)
8883 /* Leaving a core idle is often worse than degrading locality. */
8884 if (env->idle == CPU_IDLE)
8887 dist = node_distance(src_nid, dst_nid);
8889 src_weight = group_weight(p, src_nid, dist);
8890 dst_weight = group_weight(p, dst_nid, dist);
8892 src_weight = task_weight(p, src_nid, dist);
8893 dst_weight = task_weight(p, dst_nid, dist);
8896 return dst_weight < src_weight;
8900 static inline int migrate_degrades_locality(struct task_struct *p,
8908 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
8911 int can_migrate_task(struct task_struct *p, struct lb_env *env)
8915 lockdep_assert_rq_held(env->src_rq);
8918 * We do not migrate tasks that are:
8919 * 1) throttled_lb_pair, or
8920 * 2) cannot be migrated to this CPU due to cpus_ptr, or
8921 * 3) running (obviously), or
8922 * 4) are cache-hot on their current CPU.
8924 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
8927 /* Disregard pcpu kthreads; they are where they need to be. */
8928 if (kthread_is_per_cpu(p))
8931 if (!cpumask_test_cpu(env->dst_cpu, p->cpus_ptr)) {
8934 schedstat_inc(p->stats.nr_failed_migrations_affine);
8936 env->flags |= LBF_SOME_PINNED;
8939 * Remember if this task can be migrated to any other CPU in
8940 * our sched_group. We may want to revisit it if we couldn't
8941 * meet load balance goals by pulling other tasks on src_cpu.
8943 * Avoid computing new_dst_cpu
8945 * - if we have already computed one in current iteration
8946 * - if it's an active balance
8948 if (env->idle == CPU_NEWLY_IDLE ||
8949 env->flags & (LBF_DST_PINNED | LBF_ACTIVE_LB))
8952 /* Prevent to re-select dst_cpu via env's CPUs: */
8953 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
8954 if (cpumask_test_cpu(cpu, p->cpus_ptr)) {
8955 env->flags |= LBF_DST_PINNED;
8956 env->new_dst_cpu = cpu;
8964 /* Record that we found at least one task that could run on dst_cpu */
8965 env->flags &= ~LBF_ALL_PINNED;
8967 if (task_on_cpu(env->src_rq, p)) {
8968 schedstat_inc(p->stats.nr_failed_migrations_running);
8973 * Aggressive migration if:
8975 * 2) destination numa is preferred
8976 * 3) task is cache cold, or
8977 * 4) too many balance attempts have failed.
8979 if (env->flags & LBF_ACTIVE_LB)
8982 tsk_cache_hot = migrate_degrades_locality(p, env);
8983 if (tsk_cache_hot == -1)
8984 tsk_cache_hot = task_hot(p, env);
8986 if (tsk_cache_hot <= 0 ||
8987 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
8988 if (tsk_cache_hot == 1) {
8989 schedstat_inc(env->sd->lb_hot_gained[env->idle]);
8990 schedstat_inc(p->stats.nr_forced_migrations);
8995 schedstat_inc(p->stats.nr_failed_migrations_hot);
9000 * detach_task() -- detach the task for the migration specified in env
9002 static void detach_task(struct task_struct *p, struct lb_env *env)
9004 lockdep_assert_rq_held(env->src_rq);
9006 deactivate_task(env->src_rq, p, DEQUEUE_NOCLOCK);
9007 set_task_cpu(p, env->dst_cpu);
9011 * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
9012 * part of active balancing operations within "domain".
9014 * Returns a task if successful and NULL otherwise.
9016 static struct task_struct *detach_one_task(struct lb_env *env)
9018 struct task_struct *p;
9020 lockdep_assert_rq_held(env->src_rq);
9022 list_for_each_entry_reverse(p,
9023 &env->src_rq->cfs_tasks, se.group_node) {
9024 if (!can_migrate_task(p, env))
9027 detach_task(p, env);
9030 * Right now, this is only the second place where
9031 * lb_gained[env->idle] is updated (other is detach_tasks)
9032 * so we can safely collect stats here rather than
9033 * inside detach_tasks().
9035 schedstat_inc(env->sd->lb_gained[env->idle]);
9042 * detach_tasks() -- tries to detach up to imbalance load/util/tasks from
9043 * busiest_rq, as part of a balancing operation within domain "sd".
9045 * Returns number of detached tasks if successful and 0 otherwise.
9047 static int detach_tasks(struct lb_env *env)
9049 struct list_head *tasks = &env->src_rq->cfs_tasks;
9050 unsigned long util, load;
9051 struct task_struct *p;
9054 lockdep_assert_rq_held(env->src_rq);
9057 * Source run queue has been emptied by another CPU, clear
9058 * LBF_ALL_PINNED flag as we will not test any task.
9060 if (env->src_rq->nr_running <= 1) {
9061 env->flags &= ~LBF_ALL_PINNED;
9065 if (env->imbalance <= 0)
9068 while (!list_empty(tasks)) {
9070 * We don't want to steal all, otherwise we may be treated likewise,
9071 * which could at worst lead to a livelock crash.
9073 if (env->idle != CPU_NOT_IDLE && env->src_rq->nr_running <= 1)
9078 * We've more or less seen every task there is, call it quits
9079 * unless we haven't found any movable task yet.
9081 if (env->loop > env->loop_max &&
9082 !(env->flags & LBF_ALL_PINNED))
9085 /* take a breather every nr_migrate tasks */
9086 if (env->loop > env->loop_break) {
9087 env->loop_break += SCHED_NR_MIGRATE_BREAK;
9088 env->flags |= LBF_NEED_BREAK;
9092 p = list_last_entry(tasks, struct task_struct, se.group_node);
9094 if (!can_migrate_task(p, env))
9097 switch (env->migration_type) {
9100 * Depending of the number of CPUs and tasks and the
9101 * cgroup hierarchy, task_h_load() can return a null
9102 * value. Make sure that env->imbalance decreases
9103 * otherwise detach_tasks() will stop only after
9104 * detaching up to loop_max tasks.
9106 load = max_t(unsigned long, task_h_load(p), 1);
9108 if (sched_feat(LB_MIN) &&
9109 load < 16 && !env->sd->nr_balance_failed)
9113 * Make sure that we don't migrate too much load.
9114 * Nevertheless, let relax the constraint if
9115 * scheduler fails to find a good waiting task to
9118 if (shr_bound(load, env->sd->nr_balance_failed) > env->imbalance)
9121 env->imbalance -= load;
9125 util = task_util_est(p);
9127 if (shr_bound(util, env->sd->nr_balance_failed) > env->imbalance)
9130 env->imbalance -= util;
9137 case migrate_misfit:
9138 /* This is not a misfit task */
9139 if (task_fits_cpu(p, env->src_cpu))
9146 detach_task(p, env);
9147 list_add(&p->se.group_node, &env->tasks);
9151 #ifdef CONFIG_PREEMPTION
9153 * NEWIDLE balancing is a source of latency, so preemptible
9154 * kernels will stop after the first task is detached to minimize
9155 * the critical section.
9157 if (env->idle == CPU_NEWLY_IDLE)
9162 * We only want to steal up to the prescribed amount of
9165 if (env->imbalance <= 0)
9170 list_move(&p->se.group_node, tasks);
9174 * Right now, this is one of only two places we collect this stat
9175 * so we can safely collect detach_one_task() stats here rather
9176 * than inside detach_one_task().
9178 schedstat_add(env->sd->lb_gained[env->idle], detached);
9184 * attach_task() -- attach the task detached by detach_task() to its new rq.
9186 static void attach_task(struct rq *rq, struct task_struct *p)
9188 lockdep_assert_rq_held(rq);
9190 WARN_ON_ONCE(task_rq(p) != rq);
9191 activate_task(rq, p, ENQUEUE_NOCLOCK);
9192 wakeup_preempt(rq, p, 0);
9196 * attach_one_task() -- attaches the task returned from detach_one_task() to
9199 static void attach_one_task(struct rq *rq, struct task_struct *p)
9204 update_rq_clock(rq);
9210 * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
9213 static void attach_tasks(struct lb_env *env)
9215 struct list_head *tasks = &env->tasks;
9216 struct task_struct *p;
9219 rq_lock(env->dst_rq, &rf);
9220 update_rq_clock(env->dst_rq);
9222 while (!list_empty(tasks)) {
9223 p = list_first_entry(tasks, struct task_struct, se.group_node);
9224 list_del_init(&p->se.group_node);
9226 attach_task(env->dst_rq, p);
9229 rq_unlock(env->dst_rq, &rf);
9232 #ifdef CONFIG_NO_HZ_COMMON
9233 static inline bool cfs_rq_has_blocked(struct cfs_rq *cfs_rq)
9235 if (cfs_rq->avg.load_avg)
9238 if (cfs_rq->avg.util_avg)
9244 static inline bool others_have_blocked(struct rq *rq)
9246 if (cpu_util_rt(rq))
9249 if (cpu_util_dl(rq))
9252 if (thermal_load_avg(rq))
9255 if (cpu_util_irq(rq))
9261 static inline void update_blocked_load_tick(struct rq *rq)
9263 WRITE_ONCE(rq->last_blocked_load_update_tick, jiffies);
9266 static inline void update_blocked_load_status(struct rq *rq, bool has_blocked)
9269 rq->has_blocked_load = 0;
9272 static inline bool cfs_rq_has_blocked(struct cfs_rq *cfs_rq) { return false; }
9273 static inline bool others_have_blocked(struct rq *rq) { return false; }
9274 static inline void update_blocked_load_tick(struct rq *rq) {}
9275 static inline void update_blocked_load_status(struct rq *rq, bool has_blocked) {}
9278 static bool __update_blocked_others(struct rq *rq, bool *done)
9280 const struct sched_class *curr_class;
9281 u64 now = rq_clock_pelt(rq);
9282 unsigned long thermal_pressure;
9286 * update_load_avg() can call cpufreq_update_util(). Make sure that RT,
9287 * DL and IRQ signals have been updated before updating CFS.
9289 curr_class = rq->curr->sched_class;
9291 thermal_pressure = arch_scale_thermal_pressure(cpu_of(rq));
9293 decayed = update_rt_rq_load_avg(now, rq, curr_class == &rt_sched_class) |
9294 update_dl_rq_load_avg(now, rq, curr_class == &dl_sched_class) |
9295 update_thermal_load_avg(rq_clock_thermal(rq), rq, thermal_pressure) |
9296 update_irq_load_avg(rq, 0);
9298 if (others_have_blocked(rq))
9304 #ifdef CONFIG_FAIR_GROUP_SCHED
9306 static bool __update_blocked_fair(struct rq *rq, bool *done)
9308 struct cfs_rq *cfs_rq, *pos;
9309 bool decayed = false;
9310 int cpu = cpu_of(rq);
9313 * Iterates the task_group tree in a bottom up fashion, see
9314 * list_add_leaf_cfs_rq() for details.
9316 for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) {
9317 struct sched_entity *se;
9319 if (update_cfs_rq_load_avg(cfs_rq_clock_pelt(cfs_rq), cfs_rq)) {
9320 update_tg_load_avg(cfs_rq);
9322 if (cfs_rq->nr_running == 0)
9323 update_idle_cfs_rq_clock_pelt(cfs_rq);
9325 if (cfs_rq == &rq->cfs)
9329 /* Propagate pending load changes to the parent, if any: */
9330 se = cfs_rq->tg->se[cpu];
9331 if (se && !skip_blocked_update(se))
9332 update_load_avg(cfs_rq_of(se), se, UPDATE_TG);
9335 * There can be a lot of idle CPU cgroups. Don't let fully
9336 * decayed cfs_rqs linger on the list.
9338 if (cfs_rq_is_decayed(cfs_rq))
9339 list_del_leaf_cfs_rq(cfs_rq);
9341 /* Don't need periodic decay once load/util_avg are null */
9342 if (cfs_rq_has_blocked(cfs_rq))
9350 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
9351 * This needs to be done in a top-down fashion because the load of a child
9352 * group is a fraction of its parents load.
9354 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
9356 struct rq *rq = rq_of(cfs_rq);
9357 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
9358 unsigned long now = jiffies;
9361 if (cfs_rq->last_h_load_update == now)
9364 WRITE_ONCE(cfs_rq->h_load_next, NULL);
9365 for_each_sched_entity(se) {
9366 cfs_rq = cfs_rq_of(se);
9367 WRITE_ONCE(cfs_rq->h_load_next, se);
9368 if (cfs_rq->last_h_load_update == now)
9373 cfs_rq->h_load = cfs_rq_load_avg(cfs_rq);
9374 cfs_rq->last_h_load_update = now;
9377 while ((se = READ_ONCE(cfs_rq->h_load_next)) != NULL) {
9378 load = cfs_rq->h_load;
9379 load = div64_ul(load * se->avg.load_avg,
9380 cfs_rq_load_avg(cfs_rq) + 1);
9381 cfs_rq = group_cfs_rq(se);
9382 cfs_rq->h_load = load;
9383 cfs_rq->last_h_load_update = now;
9387 static unsigned long task_h_load(struct task_struct *p)
9389 struct cfs_rq *cfs_rq = task_cfs_rq(p);
9391 update_cfs_rq_h_load(cfs_rq);
9392 return div64_ul(p->se.avg.load_avg * cfs_rq->h_load,
9393 cfs_rq_load_avg(cfs_rq) + 1);
9396 static bool __update_blocked_fair(struct rq *rq, bool *done)
9398 struct cfs_rq *cfs_rq = &rq->cfs;
9401 decayed = update_cfs_rq_load_avg(cfs_rq_clock_pelt(cfs_rq), cfs_rq);
9402 if (cfs_rq_has_blocked(cfs_rq))
9408 static unsigned long task_h_load(struct task_struct *p)
9410 return p->se.avg.load_avg;
9414 static void update_blocked_averages(int cpu)
9416 bool decayed = false, done = true;
9417 struct rq *rq = cpu_rq(cpu);
9420 rq_lock_irqsave(rq, &rf);
9421 update_blocked_load_tick(rq);
9422 update_rq_clock(rq);
9424 decayed |= __update_blocked_others(rq, &done);
9425 decayed |= __update_blocked_fair(rq, &done);
9427 update_blocked_load_status(rq, !done);
9429 cpufreq_update_util(rq, 0);
9430 rq_unlock_irqrestore(rq, &rf);
9433 /********** Helpers for find_busiest_group ************************/
9436 * sg_lb_stats - stats of a sched_group required for load_balancing
9438 struct sg_lb_stats {
9439 unsigned long avg_load; /*Avg load across the CPUs of the group */
9440 unsigned long group_load; /* Total load over the CPUs of the group */
9441 unsigned long group_capacity;
9442 unsigned long group_util; /* Total utilization over the CPUs of the group */
9443 unsigned long group_runnable; /* Total runnable time over the CPUs of the group */
9444 unsigned int sum_nr_running; /* Nr of tasks running in the group */
9445 unsigned int sum_h_nr_running; /* Nr of CFS tasks running in the group */
9446 unsigned int idle_cpus;
9447 unsigned int group_weight;
9448 enum group_type group_type;
9449 unsigned int group_asym_packing; /* Tasks should be moved to preferred CPU */
9450 unsigned int group_smt_balance; /* Task on busy SMT be moved */
9451 unsigned long group_misfit_task_load; /* A CPU has a task too big for its capacity */
9452 #ifdef CONFIG_NUMA_BALANCING
9453 unsigned int nr_numa_running;
9454 unsigned int nr_preferred_running;
9459 * sd_lb_stats - Structure to store the statistics of a sched_domain
9460 * during load balancing.
9462 struct sd_lb_stats {
9463 struct sched_group *busiest; /* Busiest group in this sd */
9464 struct sched_group *local; /* Local group in this sd */
9465 unsigned long total_load; /* Total load of all groups in sd */
9466 unsigned long total_capacity; /* Total capacity of all groups in sd */
9467 unsigned long avg_load; /* Average load across all groups in sd */
9468 unsigned int prefer_sibling; /* tasks should go to sibling first */
9470 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
9471 struct sg_lb_stats local_stat; /* Statistics of the local group */
9474 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
9477 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
9478 * local_stat because update_sg_lb_stats() does a full clear/assignment.
9479 * We must however set busiest_stat::group_type and
9480 * busiest_stat::idle_cpus to the worst busiest group because
9481 * update_sd_pick_busiest() reads these before assignment.
9483 *sds = (struct sd_lb_stats){
9487 .total_capacity = 0UL,
9489 .idle_cpus = UINT_MAX,
9490 .group_type = group_has_spare,
9495 static unsigned long scale_rt_capacity(int cpu)
9497 struct rq *rq = cpu_rq(cpu);
9498 unsigned long max = arch_scale_cpu_capacity(cpu);
9499 unsigned long used, free;
9502 irq = cpu_util_irq(rq);
9504 if (unlikely(irq >= max))
9508 * avg_rt.util_avg and avg_dl.util_avg track binary signals
9509 * (running and not running) with weights 0 and 1024 respectively.
9510 * avg_thermal.load_avg tracks thermal pressure and the weighted
9511 * average uses the actual delta max capacity(load).
9513 used = cpu_util_rt(rq);
9514 used += cpu_util_dl(rq);
9515 used += thermal_load_avg(rq);
9517 if (unlikely(used >= max))
9522 return scale_irq_capacity(free, irq, max);
9525 static void update_cpu_capacity(struct sched_domain *sd, int cpu)
9527 unsigned long capacity = scale_rt_capacity(cpu);
9528 struct sched_group *sdg = sd->groups;
9533 cpu_rq(cpu)->cpu_capacity = capacity;
9534 trace_sched_cpu_capacity_tp(cpu_rq(cpu));
9536 sdg->sgc->capacity = capacity;
9537 sdg->sgc->min_capacity = capacity;
9538 sdg->sgc->max_capacity = capacity;
9541 void update_group_capacity(struct sched_domain *sd, int cpu)
9543 struct sched_domain *child = sd->child;
9544 struct sched_group *group, *sdg = sd->groups;
9545 unsigned long capacity, min_capacity, max_capacity;
9546 unsigned long interval;
9548 interval = msecs_to_jiffies(sd->balance_interval);
9549 interval = clamp(interval, 1UL, max_load_balance_interval);
9550 sdg->sgc->next_update = jiffies + interval;
9553 update_cpu_capacity(sd, cpu);
9558 min_capacity = ULONG_MAX;
9561 if (child->flags & SD_OVERLAP) {
9563 * SD_OVERLAP domains cannot assume that child groups
9564 * span the current group.
9567 for_each_cpu(cpu, sched_group_span(sdg)) {
9568 unsigned long cpu_cap = capacity_of(cpu);
9570 capacity += cpu_cap;
9571 min_capacity = min(cpu_cap, min_capacity);
9572 max_capacity = max(cpu_cap, max_capacity);
9576 * !SD_OVERLAP domains can assume that child groups
9577 * span the current group.
9580 group = child->groups;
9582 struct sched_group_capacity *sgc = group->sgc;
9584 capacity += sgc->capacity;
9585 min_capacity = min(sgc->min_capacity, min_capacity);
9586 max_capacity = max(sgc->max_capacity, max_capacity);
9587 group = group->next;
9588 } while (group != child->groups);
9591 sdg->sgc->capacity = capacity;
9592 sdg->sgc->min_capacity = min_capacity;
9593 sdg->sgc->max_capacity = max_capacity;
9597 * Check whether the capacity of the rq has been noticeably reduced by side
9598 * activity. The imbalance_pct is used for the threshold.
9599 * Return true is the capacity is reduced
9602 check_cpu_capacity(struct rq *rq, struct sched_domain *sd)
9604 return ((rq->cpu_capacity * sd->imbalance_pct) <
9605 (arch_scale_cpu_capacity(cpu_of(rq)) * 100));
9609 * Check whether a rq has a misfit task and if it looks like we can actually
9610 * help that task: we can migrate the task to a CPU of higher capacity, or
9611 * the task's current CPU is heavily pressured.
9613 static inline int check_misfit_status(struct rq *rq, struct sched_domain *sd)
9615 return rq->misfit_task_load &&
9616 (arch_scale_cpu_capacity(rq->cpu) < rq->rd->max_cpu_capacity ||
9617 check_cpu_capacity(rq, sd));
9621 * Group imbalance indicates (and tries to solve) the problem where balancing
9622 * groups is inadequate due to ->cpus_ptr constraints.
9624 * Imagine a situation of two groups of 4 CPUs each and 4 tasks each with a
9625 * cpumask covering 1 CPU of the first group and 3 CPUs of the second group.
9628 * { 0 1 2 3 } { 4 5 6 7 }
9631 * If we were to balance group-wise we'd place two tasks in the first group and
9632 * two tasks in the second group. Clearly this is undesired as it will overload
9633 * cpu 3 and leave one of the CPUs in the second group unused.
9635 * The current solution to this issue is detecting the skew in the first group
9636 * by noticing the lower domain failed to reach balance and had difficulty
9637 * moving tasks due to affinity constraints.
9639 * When this is so detected; this group becomes a candidate for busiest; see
9640 * update_sd_pick_busiest(). And calculate_imbalance() and
9641 * find_busiest_group() avoid some of the usual balance conditions to allow it
9642 * to create an effective group imbalance.
9644 * This is a somewhat tricky proposition since the next run might not find the
9645 * group imbalance and decide the groups need to be balanced again. A most
9646 * subtle and fragile situation.
9649 static inline int sg_imbalanced(struct sched_group *group)
9651 return group->sgc->imbalance;
9655 * group_has_capacity returns true if the group has spare capacity that could
9656 * be used by some tasks.
9657 * We consider that a group has spare capacity if the number of task is
9658 * smaller than the number of CPUs or if the utilization is lower than the
9659 * available capacity for CFS tasks.
9660 * For the latter, we use a threshold to stabilize the state, to take into
9661 * account the variance of the tasks' load and to return true if the available
9662 * capacity in meaningful for the load balancer.
9663 * As an example, an available capacity of 1% can appear but it doesn't make
9664 * any benefit for the load balance.
9667 group_has_capacity(unsigned int imbalance_pct, struct sg_lb_stats *sgs)
9669 if (sgs->sum_nr_running < sgs->group_weight)
9672 if ((sgs->group_capacity * imbalance_pct) <
9673 (sgs->group_runnable * 100))
9676 if ((sgs->group_capacity * 100) >
9677 (sgs->group_util * imbalance_pct))
9684 * group_is_overloaded returns true if the group has more tasks than it can
9686 * group_is_overloaded is not equals to !group_has_capacity because a group
9687 * with the exact right number of tasks, has no more spare capacity but is not
9688 * overloaded so both group_has_capacity and group_is_overloaded return
9692 group_is_overloaded(unsigned int imbalance_pct, struct sg_lb_stats *sgs)
9694 if (sgs->sum_nr_running <= sgs->group_weight)
9697 if ((sgs->group_capacity * 100) <
9698 (sgs->group_util * imbalance_pct))
9701 if ((sgs->group_capacity * imbalance_pct) <
9702 (sgs->group_runnable * 100))
9709 group_type group_classify(unsigned int imbalance_pct,
9710 struct sched_group *group,
9711 struct sg_lb_stats *sgs)
9713 if (group_is_overloaded(imbalance_pct, sgs))
9714 return group_overloaded;
9716 if (sg_imbalanced(group))
9717 return group_imbalanced;
9719 if (sgs->group_asym_packing)
9720 return group_asym_packing;
9722 if (sgs->group_smt_balance)
9723 return group_smt_balance;
9725 if (sgs->group_misfit_task_load)
9726 return group_misfit_task;
9728 if (!group_has_capacity(imbalance_pct, sgs))
9729 return group_fully_busy;
9731 return group_has_spare;
9735 * sched_use_asym_prio - Check whether asym_packing priority must be used
9736 * @sd: The scheduling domain of the load balancing
9739 * Always use CPU priority when balancing load between SMT siblings. When
9740 * balancing load between cores, it is not sufficient that @cpu is idle. Only
9741 * use CPU priority if the whole core is idle.
9743 * Returns: True if the priority of @cpu must be followed. False otherwise.
9745 static bool sched_use_asym_prio(struct sched_domain *sd, int cpu)
9747 if (!sched_smt_active())
9750 return sd->flags & SD_SHARE_CPUCAPACITY || is_core_idle(cpu);
9754 * sched_asym - Check if the destination CPU can do asym_packing load balance
9755 * @env: The load balancing environment
9756 * @sgs: Load-balancing statistics of the candidate busiest group
9757 * @group: The candidate busiest group
9759 * @env::dst_cpu can do asym_packing if it has higher priority than the
9760 * preferred CPU of @group.
9762 * SMT is a special case. If we are balancing load between cores, @env::dst_cpu
9763 * can do asym_packing balance only if all its SMT siblings are idle. Also, it
9764 * can only do it if @group is an SMT group and has exactly on busy CPU. Larger
9765 * imbalances in the number of CPUS are dealt with in find_busiest_group().
9767 * If we are balancing load within an SMT core, or at PKG domain level, always
9770 * Return: true if @env::dst_cpu can do with asym_packing load balance. False
9774 sched_asym(struct lb_env *env, struct sg_lb_stats *sgs, struct sched_group *group)
9776 /* Ensure that the whole local core is idle, if applicable. */
9777 if (!sched_use_asym_prio(env->sd, env->dst_cpu))
9781 * CPU priorities does not make sense for SMT cores with more than one
9784 if (group->flags & SD_SHARE_CPUCAPACITY) {
9785 if (sgs->group_weight - sgs->idle_cpus != 1)
9789 return sched_asym_prefer(env->dst_cpu, group->asym_prefer_cpu);
9792 /* One group has more than one SMT CPU while the other group does not */
9793 static inline bool smt_vs_nonsmt_groups(struct sched_group *sg1,
9794 struct sched_group *sg2)
9799 return (sg1->flags & SD_SHARE_CPUCAPACITY) !=
9800 (sg2->flags & SD_SHARE_CPUCAPACITY);
9803 static inline bool smt_balance(struct lb_env *env, struct sg_lb_stats *sgs,
9804 struct sched_group *group)
9806 if (env->idle == CPU_NOT_IDLE)
9810 * For SMT source group, it is better to move a task
9811 * to a CPU that doesn't have multiple tasks sharing its CPU capacity.
9812 * Note that if a group has a single SMT, SD_SHARE_CPUCAPACITY
9815 if (group->flags & SD_SHARE_CPUCAPACITY &&
9816 sgs->sum_h_nr_running > 1)
9822 static inline long sibling_imbalance(struct lb_env *env,
9823 struct sd_lb_stats *sds,
9824 struct sg_lb_stats *busiest,
9825 struct sg_lb_stats *local)
9827 int ncores_busiest, ncores_local;
9830 if (env->idle == CPU_NOT_IDLE || !busiest->sum_nr_running)
9833 ncores_busiest = sds->busiest->cores;
9834 ncores_local = sds->local->cores;
9836 if (ncores_busiest == ncores_local) {
9837 imbalance = busiest->sum_nr_running;
9838 lsub_positive(&imbalance, local->sum_nr_running);
9842 /* Balance such that nr_running/ncores ratio are same on both groups */
9843 imbalance = ncores_local * busiest->sum_nr_running;
9844 lsub_positive(&imbalance, ncores_busiest * local->sum_nr_running);
9845 /* Normalize imbalance and do rounding on normalization */
9846 imbalance = 2 * imbalance + ncores_local + ncores_busiest;
9847 imbalance /= ncores_local + ncores_busiest;
9849 /* Take advantage of resource in an empty sched group */
9850 if (imbalance <= 1 && local->sum_nr_running == 0 &&
9851 busiest->sum_nr_running > 1)
9858 sched_reduced_capacity(struct rq *rq, struct sched_domain *sd)
9861 * When there is more than 1 task, the group_overloaded case already
9862 * takes care of cpu with reduced capacity
9864 if (rq->cfs.h_nr_running != 1)
9867 return check_cpu_capacity(rq, sd);
9871 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
9872 * @env: The load balancing environment.
9873 * @sds: Load-balancing data with statistics of the local group.
9874 * @group: sched_group whose statistics are to be updated.
9875 * @sgs: variable to hold the statistics for this group.
9876 * @sg_status: Holds flag indicating the status of the sched_group
9878 static inline void update_sg_lb_stats(struct lb_env *env,
9879 struct sd_lb_stats *sds,
9880 struct sched_group *group,
9881 struct sg_lb_stats *sgs,
9884 int i, nr_running, local_group;
9886 memset(sgs, 0, sizeof(*sgs));
9888 local_group = group == sds->local;
9890 for_each_cpu_and(i, sched_group_span(group), env->cpus) {
9891 struct rq *rq = cpu_rq(i);
9892 unsigned long load = cpu_load(rq);
9894 sgs->group_load += load;
9895 sgs->group_util += cpu_util_cfs(i);
9896 sgs->group_runnable += cpu_runnable(rq);
9897 sgs->sum_h_nr_running += rq->cfs.h_nr_running;
9899 nr_running = rq->nr_running;
9900 sgs->sum_nr_running += nr_running;
9903 *sg_status |= SG_OVERLOAD;
9905 if (cpu_overutilized(i))
9906 *sg_status |= SG_OVERUTILIZED;
9908 #ifdef CONFIG_NUMA_BALANCING
9909 sgs->nr_numa_running += rq->nr_numa_running;
9910 sgs->nr_preferred_running += rq->nr_preferred_running;
9913 * No need to call idle_cpu() if nr_running is not 0
9915 if (!nr_running && idle_cpu(i)) {
9917 /* Idle cpu can't have misfit task */
9924 if (env->sd->flags & SD_ASYM_CPUCAPACITY) {
9925 /* Check for a misfit task on the cpu */
9926 if (sgs->group_misfit_task_load < rq->misfit_task_load) {
9927 sgs->group_misfit_task_load = rq->misfit_task_load;
9928 *sg_status |= SG_OVERLOAD;
9930 } else if ((env->idle != CPU_NOT_IDLE) &&
9931 sched_reduced_capacity(rq, env->sd)) {
9932 /* Check for a task running on a CPU with reduced capacity */
9933 if (sgs->group_misfit_task_load < load)
9934 sgs->group_misfit_task_load = load;
9938 sgs->group_capacity = group->sgc->capacity;
9940 sgs->group_weight = group->group_weight;
9942 /* Check if dst CPU is idle and preferred to this group */
9943 if (!local_group && env->sd->flags & SD_ASYM_PACKING &&
9944 env->idle != CPU_NOT_IDLE && sgs->sum_h_nr_running &&
9945 sched_asym(env, sgs, group)) {
9946 sgs->group_asym_packing = 1;
9949 /* Check for loaded SMT group to be balanced to dst CPU */
9950 if (!local_group && smt_balance(env, sgs, group))
9951 sgs->group_smt_balance = 1;
9953 sgs->group_type = group_classify(env->sd->imbalance_pct, group, sgs);
9955 /* Computing avg_load makes sense only when group is overloaded */
9956 if (sgs->group_type == group_overloaded)
9957 sgs->avg_load = (sgs->group_load * SCHED_CAPACITY_SCALE) /
9958 sgs->group_capacity;
9962 * update_sd_pick_busiest - return 1 on busiest group
9963 * @env: The load balancing environment.
9964 * @sds: sched_domain statistics
9965 * @sg: sched_group candidate to be checked for being the busiest
9966 * @sgs: sched_group statistics
9968 * Determine if @sg is a busier group than the previously selected
9971 * Return: %true if @sg is a busier group than the previously selected
9972 * busiest group. %false otherwise.
9974 static bool update_sd_pick_busiest(struct lb_env *env,
9975 struct sd_lb_stats *sds,
9976 struct sched_group *sg,
9977 struct sg_lb_stats *sgs)
9979 struct sg_lb_stats *busiest = &sds->busiest_stat;
9981 /* Make sure that there is at least one task to pull */
9982 if (!sgs->sum_h_nr_running)
9986 * Don't try to pull misfit tasks we can't help.
9987 * We can use max_capacity here as reduction in capacity on some
9988 * CPUs in the group should either be possible to resolve
9989 * internally or be covered by avg_load imbalance (eventually).
9991 if ((env->sd->flags & SD_ASYM_CPUCAPACITY) &&
9992 (sgs->group_type == group_misfit_task) &&
9993 (!capacity_greater(capacity_of(env->dst_cpu), sg->sgc->max_capacity) ||
9994 sds->local_stat.group_type != group_has_spare))
9997 if (sgs->group_type > busiest->group_type)
10000 if (sgs->group_type < busiest->group_type)
10004 * The candidate and the current busiest group are the same type of
10005 * group. Let check which one is the busiest according to the type.
10008 switch (sgs->group_type) {
10009 case group_overloaded:
10010 /* Select the overloaded group with highest avg_load. */
10011 return sgs->avg_load > busiest->avg_load;
10013 case group_imbalanced:
10015 * Select the 1st imbalanced group as we don't have any way to
10016 * choose one more than another.
10020 case group_asym_packing:
10021 /* Prefer to move from lowest priority CPU's work */
10022 return sched_asym_prefer(sds->busiest->asym_prefer_cpu, sg->asym_prefer_cpu);
10024 case group_misfit_task:
10026 * If we have more than one misfit sg go with the biggest
10029 return sgs->group_misfit_task_load > busiest->group_misfit_task_load;
10031 case group_smt_balance:
10033 * Check if we have spare CPUs on either SMT group to
10034 * choose has spare or fully busy handling.
10036 if (sgs->idle_cpus != 0 || busiest->idle_cpus != 0)
10041 case group_fully_busy:
10043 * Select the fully busy group with highest avg_load. In
10044 * theory, there is no need to pull task from such kind of
10045 * group because tasks have all compute capacity that they need
10046 * but we can still improve the overall throughput by reducing
10047 * contention when accessing shared HW resources.
10049 * XXX for now avg_load is not computed and always 0 so we
10050 * select the 1st one, except if @sg is composed of SMT
10054 if (sgs->avg_load < busiest->avg_load)
10057 if (sgs->avg_load == busiest->avg_load) {
10059 * SMT sched groups need more help than non-SMT groups.
10060 * If @sg happens to also be SMT, either choice is good.
10062 if (sds->busiest->flags & SD_SHARE_CPUCAPACITY)
10068 case group_has_spare:
10070 * Do not pick sg with SMT CPUs over sg with pure CPUs,
10071 * as we do not want to pull task off SMT core with one task
10072 * and make the core idle.
10074 if (smt_vs_nonsmt_groups(sds->busiest, sg)) {
10075 if (sg->flags & SD_SHARE_CPUCAPACITY && sgs->sum_h_nr_running <= 1)
10083 * Select not overloaded group with lowest number of idle cpus
10084 * and highest number of running tasks. We could also compare
10085 * the spare capacity which is more stable but it can end up
10086 * that the group has less spare capacity but finally more idle
10087 * CPUs which means less opportunity to pull tasks.
10089 if (sgs->idle_cpus > busiest->idle_cpus)
10091 else if ((sgs->idle_cpus == busiest->idle_cpus) &&
10092 (sgs->sum_nr_running <= busiest->sum_nr_running))
10099 * Candidate sg has no more than one task per CPU and has higher
10100 * per-CPU capacity. Migrating tasks to less capable CPUs may harm
10101 * throughput. Maximize throughput, power/energy consequences are not
10104 if ((env->sd->flags & SD_ASYM_CPUCAPACITY) &&
10105 (sgs->group_type <= group_fully_busy) &&
10106 (capacity_greater(sg->sgc->min_capacity, capacity_of(env->dst_cpu))))
10112 #ifdef CONFIG_NUMA_BALANCING
10113 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
10115 if (sgs->sum_h_nr_running > sgs->nr_numa_running)
10117 if (sgs->sum_h_nr_running > sgs->nr_preferred_running)
10122 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
10124 if (rq->nr_running > rq->nr_numa_running)
10126 if (rq->nr_running > rq->nr_preferred_running)
10131 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
10136 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
10140 #endif /* CONFIG_NUMA_BALANCING */
10143 struct sg_lb_stats;
10146 * task_running_on_cpu - return 1 if @p is running on @cpu.
10149 static unsigned int task_running_on_cpu(int cpu, struct task_struct *p)
10151 /* Task has no contribution or is new */
10152 if (cpu != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time))
10155 if (task_on_rq_queued(p))
10162 * idle_cpu_without - would a given CPU be idle without p ?
10163 * @cpu: the processor on which idleness is tested.
10164 * @p: task which should be ignored.
10166 * Return: 1 if the CPU would be idle. 0 otherwise.
10168 static int idle_cpu_without(int cpu, struct task_struct *p)
10170 struct rq *rq = cpu_rq(cpu);
10172 if (rq->curr != rq->idle && rq->curr != p)
10176 * rq->nr_running can't be used but an updated version without the
10177 * impact of p on cpu must be used instead. The updated nr_running
10178 * be computed and tested before calling idle_cpu_without().
10181 if (rq->ttwu_pending)
10188 * update_sg_wakeup_stats - Update sched_group's statistics for wakeup.
10189 * @sd: The sched_domain level to look for idlest group.
10190 * @group: sched_group whose statistics are to be updated.
10191 * @sgs: variable to hold the statistics for this group.
10192 * @p: The task for which we look for the idlest group/CPU.
10194 static inline void update_sg_wakeup_stats(struct sched_domain *sd,
10195 struct sched_group *group,
10196 struct sg_lb_stats *sgs,
10197 struct task_struct *p)
10201 memset(sgs, 0, sizeof(*sgs));
10203 /* Assume that task can't fit any CPU of the group */
10204 if (sd->flags & SD_ASYM_CPUCAPACITY)
10205 sgs->group_misfit_task_load = 1;
10207 for_each_cpu(i, sched_group_span(group)) {
10208 struct rq *rq = cpu_rq(i);
10209 unsigned int local;
10211 sgs->group_load += cpu_load_without(rq, p);
10212 sgs->group_util += cpu_util_without(i, p);
10213 sgs->group_runnable += cpu_runnable_without(rq, p);
10214 local = task_running_on_cpu(i, p);
10215 sgs->sum_h_nr_running += rq->cfs.h_nr_running - local;
10217 nr_running = rq->nr_running - local;
10218 sgs->sum_nr_running += nr_running;
10221 * No need to call idle_cpu_without() if nr_running is not 0
10223 if (!nr_running && idle_cpu_without(i, p))
10226 /* Check if task fits in the CPU */
10227 if (sd->flags & SD_ASYM_CPUCAPACITY &&
10228 sgs->group_misfit_task_load &&
10229 task_fits_cpu(p, i))
10230 sgs->group_misfit_task_load = 0;
10234 sgs->group_capacity = group->sgc->capacity;
10236 sgs->group_weight = group->group_weight;
10238 sgs->group_type = group_classify(sd->imbalance_pct, group, sgs);
10241 * Computing avg_load makes sense only when group is fully busy or
10244 if (sgs->group_type == group_fully_busy ||
10245 sgs->group_type == group_overloaded)
10246 sgs->avg_load = (sgs->group_load * SCHED_CAPACITY_SCALE) /
10247 sgs->group_capacity;
10250 static bool update_pick_idlest(struct sched_group *idlest,
10251 struct sg_lb_stats *idlest_sgs,
10252 struct sched_group *group,
10253 struct sg_lb_stats *sgs)
10255 if (sgs->group_type < idlest_sgs->group_type)
10258 if (sgs->group_type > idlest_sgs->group_type)
10262 * The candidate and the current idlest group are the same type of
10263 * group. Let check which one is the idlest according to the type.
10266 switch (sgs->group_type) {
10267 case group_overloaded:
10268 case group_fully_busy:
10269 /* Select the group with lowest avg_load. */
10270 if (idlest_sgs->avg_load <= sgs->avg_load)
10274 case group_imbalanced:
10275 case group_asym_packing:
10276 case group_smt_balance:
10277 /* Those types are not used in the slow wakeup path */
10280 case group_misfit_task:
10281 /* Select group with the highest max capacity */
10282 if (idlest->sgc->max_capacity >= group->sgc->max_capacity)
10286 case group_has_spare:
10287 /* Select group with most idle CPUs */
10288 if (idlest_sgs->idle_cpus > sgs->idle_cpus)
10291 /* Select group with lowest group_util */
10292 if (idlest_sgs->idle_cpus == sgs->idle_cpus &&
10293 idlest_sgs->group_util <= sgs->group_util)
10303 * find_idlest_group() finds and returns the least busy CPU group within the
10306 * Assumes p is allowed on at least one CPU in sd.
10308 static struct sched_group *
10309 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
10311 struct sched_group *idlest = NULL, *local = NULL, *group = sd->groups;
10312 struct sg_lb_stats local_sgs, tmp_sgs;
10313 struct sg_lb_stats *sgs;
10314 unsigned long imbalance;
10315 struct sg_lb_stats idlest_sgs = {
10316 .avg_load = UINT_MAX,
10317 .group_type = group_overloaded,
10323 /* Skip over this group if it has no CPUs allowed */
10324 if (!cpumask_intersects(sched_group_span(group),
10328 /* Skip over this group if no cookie matched */
10329 if (!sched_group_cookie_match(cpu_rq(this_cpu), p, group))
10332 local_group = cpumask_test_cpu(this_cpu,
10333 sched_group_span(group));
10342 update_sg_wakeup_stats(sd, group, sgs, p);
10344 if (!local_group && update_pick_idlest(idlest, &idlest_sgs, group, sgs)) {
10349 } while (group = group->next, group != sd->groups);
10352 /* There is no idlest group to push tasks to */
10356 /* The local group has been skipped because of CPU affinity */
10361 * If the local group is idler than the selected idlest group
10362 * don't try and push the task.
10364 if (local_sgs.group_type < idlest_sgs.group_type)
10368 * If the local group is busier than the selected idlest group
10369 * try and push the task.
10371 if (local_sgs.group_type > idlest_sgs.group_type)
10374 switch (local_sgs.group_type) {
10375 case group_overloaded:
10376 case group_fully_busy:
10378 /* Calculate allowed imbalance based on load */
10379 imbalance = scale_load_down(NICE_0_LOAD) *
10380 (sd->imbalance_pct-100) / 100;
10383 * When comparing groups across NUMA domains, it's possible for
10384 * the local domain to be very lightly loaded relative to the
10385 * remote domains but "imbalance" skews the comparison making
10386 * remote CPUs look much more favourable. When considering
10387 * cross-domain, add imbalance to the load on the remote node
10388 * and consider staying local.
10391 if ((sd->flags & SD_NUMA) &&
10392 ((idlest_sgs.avg_load + imbalance) >= local_sgs.avg_load))
10396 * If the local group is less loaded than the selected
10397 * idlest group don't try and push any tasks.
10399 if (idlest_sgs.avg_load >= (local_sgs.avg_load + imbalance))
10402 if (100 * local_sgs.avg_load <= sd->imbalance_pct * idlest_sgs.avg_load)
10406 case group_imbalanced:
10407 case group_asym_packing:
10408 case group_smt_balance:
10409 /* Those type are not used in the slow wakeup path */
10412 case group_misfit_task:
10413 /* Select group with the highest max capacity */
10414 if (local->sgc->max_capacity >= idlest->sgc->max_capacity)
10418 case group_has_spare:
10420 if (sd->flags & SD_NUMA) {
10421 int imb_numa_nr = sd->imb_numa_nr;
10422 #ifdef CONFIG_NUMA_BALANCING
10425 * If there is spare capacity at NUMA, try to select
10426 * the preferred node
10428 if (cpu_to_node(this_cpu) == p->numa_preferred_nid)
10431 idlest_cpu = cpumask_first(sched_group_span(idlest));
10432 if (cpu_to_node(idlest_cpu) == p->numa_preferred_nid)
10434 #endif /* CONFIG_NUMA_BALANCING */
10436 * Otherwise, keep the task close to the wakeup source
10437 * and improve locality if the number of running tasks
10438 * would remain below threshold where an imbalance is
10439 * allowed while accounting for the possibility the
10440 * task is pinned to a subset of CPUs. If there is a
10441 * real need of migration, periodic load balance will
10444 if (p->nr_cpus_allowed != NR_CPUS) {
10445 struct cpumask *cpus = this_cpu_cpumask_var_ptr(select_rq_mask);
10447 cpumask_and(cpus, sched_group_span(local), p->cpus_ptr);
10448 imb_numa_nr = min(cpumask_weight(cpus), sd->imb_numa_nr);
10451 imbalance = abs(local_sgs.idle_cpus - idlest_sgs.idle_cpus);
10452 if (!adjust_numa_imbalance(imbalance,
10453 local_sgs.sum_nr_running + 1,
10458 #endif /* CONFIG_NUMA */
10461 * Select group with highest number of idle CPUs. We could also
10462 * compare the utilization which is more stable but it can end
10463 * up that the group has less spare capacity but finally more
10464 * idle CPUs which means more opportunity to run task.
10466 if (local_sgs.idle_cpus >= idlest_sgs.idle_cpus)
10474 static void update_idle_cpu_scan(struct lb_env *env,
10475 unsigned long sum_util)
10477 struct sched_domain_shared *sd_share;
10478 int llc_weight, pct;
10481 * Update the number of CPUs to scan in LLC domain, which could
10482 * be used as a hint in select_idle_cpu(). The update of sd_share
10483 * could be expensive because it is within a shared cache line.
10484 * So the write of this hint only occurs during periodic load
10485 * balancing, rather than CPU_NEWLY_IDLE, because the latter
10486 * can fire way more frequently than the former.
10488 if (!sched_feat(SIS_UTIL) || env->idle == CPU_NEWLY_IDLE)
10491 llc_weight = per_cpu(sd_llc_size, env->dst_cpu);
10492 if (env->sd->span_weight != llc_weight)
10495 sd_share = rcu_dereference(per_cpu(sd_llc_shared, env->dst_cpu));
10500 * The number of CPUs to search drops as sum_util increases, when
10501 * sum_util hits 85% or above, the scan stops.
10502 * The reason to choose 85% as the threshold is because this is the
10503 * imbalance_pct(117) when a LLC sched group is overloaded.
10505 * let y = SCHED_CAPACITY_SCALE - p * x^2 [1]
10506 * and y'= y / SCHED_CAPACITY_SCALE
10508 * x is the ratio of sum_util compared to the CPU capacity:
10509 * x = sum_util / (llc_weight * SCHED_CAPACITY_SCALE)
10510 * y' is the ratio of CPUs to be scanned in the LLC domain,
10511 * and the number of CPUs to scan is calculated by:
10513 * nr_scan = llc_weight * y' [2]
10515 * When x hits the threshold of overloaded, AKA, when
10516 * x = 100 / pct, y drops to 0. According to [1],
10517 * p should be SCHED_CAPACITY_SCALE * pct^2 / 10000
10519 * Scale x by SCHED_CAPACITY_SCALE:
10520 * x' = sum_util / llc_weight; [3]
10522 * and finally [1] becomes:
10523 * y = SCHED_CAPACITY_SCALE -
10524 * x'^2 * pct^2 / (10000 * SCHED_CAPACITY_SCALE) [4]
10529 do_div(x, llc_weight);
10532 pct = env->sd->imbalance_pct;
10533 tmp = x * x * pct * pct;
10534 do_div(tmp, 10000 * SCHED_CAPACITY_SCALE);
10535 tmp = min_t(long, tmp, SCHED_CAPACITY_SCALE);
10536 y = SCHED_CAPACITY_SCALE - tmp;
10540 do_div(y, SCHED_CAPACITY_SCALE);
10541 if ((int)y != sd_share->nr_idle_scan)
10542 WRITE_ONCE(sd_share->nr_idle_scan, (int)y);
10546 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
10547 * @env: The load balancing environment.
10548 * @sds: variable to hold the statistics for this sched_domain.
10551 static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
10553 struct sched_group *sg = env->sd->groups;
10554 struct sg_lb_stats *local = &sds->local_stat;
10555 struct sg_lb_stats tmp_sgs;
10556 unsigned long sum_util = 0;
10560 struct sg_lb_stats *sgs = &tmp_sgs;
10563 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_span(sg));
10568 if (env->idle != CPU_NEWLY_IDLE ||
10569 time_after_eq(jiffies, sg->sgc->next_update))
10570 update_group_capacity(env->sd, env->dst_cpu);
10573 update_sg_lb_stats(env, sds, sg, sgs, &sg_status);
10575 if (!local_group && update_sd_pick_busiest(env, sds, sg, sgs)) {
10577 sds->busiest_stat = *sgs;
10580 /* Now, start updating sd_lb_stats */
10581 sds->total_load += sgs->group_load;
10582 sds->total_capacity += sgs->group_capacity;
10584 sum_util += sgs->group_util;
10586 } while (sg != env->sd->groups);
10589 * Indicate that the child domain of the busiest group prefers tasks
10590 * go to a child's sibling domains first. NB the flags of a sched group
10591 * are those of the child domain.
10594 sds->prefer_sibling = !!(sds->busiest->flags & SD_PREFER_SIBLING);
10597 if (env->sd->flags & SD_NUMA)
10598 env->fbq_type = fbq_classify_group(&sds->busiest_stat);
10600 if (!env->sd->parent) {
10601 struct root_domain *rd = env->dst_rq->rd;
10603 /* update overload indicator if we are at root domain */
10604 WRITE_ONCE(rd->overload, sg_status & SG_OVERLOAD);
10606 /* Update over-utilization (tipping point, U >= 0) indicator */
10607 WRITE_ONCE(rd->overutilized, sg_status & SG_OVERUTILIZED);
10608 trace_sched_overutilized_tp(rd, sg_status & SG_OVERUTILIZED);
10609 } else if (sg_status & SG_OVERUTILIZED) {
10610 struct root_domain *rd = env->dst_rq->rd;
10612 WRITE_ONCE(rd->overutilized, SG_OVERUTILIZED);
10613 trace_sched_overutilized_tp(rd, SG_OVERUTILIZED);
10616 update_idle_cpu_scan(env, sum_util);
10620 * calculate_imbalance - Calculate the amount of imbalance present within the
10621 * groups of a given sched_domain during load balance.
10622 * @env: load balance environment
10623 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
10625 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
10627 struct sg_lb_stats *local, *busiest;
10629 local = &sds->local_stat;
10630 busiest = &sds->busiest_stat;
10632 if (busiest->group_type == group_misfit_task) {
10633 if (env->sd->flags & SD_ASYM_CPUCAPACITY) {
10634 /* Set imbalance to allow misfit tasks to be balanced. */
10635 env->migration_type = migrate_misfit;
10636 env->imbalance = 1;
10639 * Set load imbalance to allow moving task from cpu
10640 * with reduced capacity.
10642 env->migration_type = migrate_load;
10643 env->imbalance = busiest->group_misfit_task_load;
10648 if (busiest->group_type == group_asym_packing) {
10650 * In case of asym capacity, we will try to migrate all load to
10651 * the preferred CPU.
10653 env->migration_type = migrate_task;
10654 env->imbalance = busiest->sum_h_nr_running;
10658 if (busiest->group_type == group_smt_balance) {
10659 /* Reduce number of tasks sharing CPU capacity */
10660 env->migration_type = migrate_task;
10661 env->imbalance = 1;
10665 if (busiest->group_type == group_imbalanced) {
10667 * In the group_imb case we cannot rely on group-wide averages
10668 * to ensure CPU-load equilibrium, try to move any task to fix
10669 * the imbalance. The next load balance will take care of
10670 * balancing back the system.
10672 env->migration_type = migrate_task;
10673 env->imbalance = 1;
10678 * Try to use spare capacity of local group without overloading it or
10679 * emptying busiest.
10681 if (local->group_type == group_has_spare) {
10682 if ((busiest->group_type > group_fully_busy) &&
10683 !(env->sd->flags & SD_SHARE_PKG_RESOURCES)) {
10685 * If busiest is overloaded, try to fill spare
10686 * capacity. This might end up creating spare capacity
10687 * in busiest or busiest still being overloaded but
10688 * there is no simple way to directly compute the
10689 * amount of load to migrate in order to balance the
10692 env->migration_type = migrate_util;
10693 env->imbalance = max(local->group_capacity, local->group_util) -
10697 * In some cases, the group's utilization is max or even
10698 * higher than capacity because of migrations but the
10699 * local CPU is (newly) idle. There is at least one
10700 * waiting task in this overloaded busiest group. Let's
10703 if (env->idle != CPU_NOT_IDLE && env->imbalance == 0) {
10704 env->migration_type = migrate_task;
10705 env->imbalance = 1;
10711 if (busiest->group_weight == 1 || sds->prefer_sibling) {
10713 * When prefer sibling, evenly spread running tasks on
10716 env->migration_type = migrate_task;
10717 env->imbalance = sibling_imbalance(env, sds, busiest, local);
10721 * If there is no overload, we just want to even the number of
10724 env->migration_type = migrate_task;
10725 env->imbalance = max_t(long, 0,
10726 (local->idle_cpus - busiest->idle_cpus));
10730 /* Consider allowing a small imbalance between NUMA groups */
10731 if (env->sd->flags & SD_NUMA) {
10732 env->imbalance = adjust_numa_imbalance(env->imbalance,
10733 local->sum_nr_running + 1,
10734 env->sd->imb_numa_nr);
10738 /* Number of tasks to move to restore balance */
10739 env->imbalance >>= 1;
10745 * Local is fully busy but has to take more load to relieve the
10748 if (local->group_type < group_overloaded) {
10750 * Local will become overloaded so the avg_load metrics are
10754 local->avg_load = (local->group_load * SCHED_CAPACITY_SCALE) /
10755 local->group_capacity;
10758 * If the local group is more loaded than the selected
10759 * busiest group don't try to pull any tasks.
10761 if (local->avg_load >= busiest->avg_load) {
10762 env->imbalance = 0;
10766 sds->avg_load = (sds->total_load * SCHED_CAPACITY_SCALE) /
10767 sds->total_capacity;
10770 * If the local group is more loaded than the average system
10771 * load, don't try to pull any tasks.
10773 if (local->avg_load >= sds->avg_load) {
10774 env->imbalance = 0;
10781 * Both group are or will become overloaded and we're trying to get all
10782 * the CPUs to the average_load, so we don't want to push ourselves
10783 * above the average load, nor do we wish to reduce the max loaded CPU
10784 * below the average load. At the same time, we also don't want to
10785 * reduce the group load below the group capacity. Thus we look for
10786 * the minimum possible imbalance.
10788 env->migration_type = migrate_load;
10789 env->imbalance = min(
10790 (busiest->avg_load - sds->avg_load) * busiest->group_capacity,
10791 (sds->avg_load - local->avg_load) * local->group_capacity
10792 ) / SCHED_CAPACITY_SCALE;
10795 /******* find_busiest_group() helpers end here *********************/
10798 * Decision matrix according to the local and busiest group type:
10800 * busiest \ local has_spare fully_busy misfit asym imbalanced overloaded
10801 * has_spare nr_idle balanced N/A N/A balanced balanced
10802 * fully_busy nr_idle nr_idle N/A N/A balanced balanced
10803 * misfit_task force N/A N/A N/A N/A N/A
10804 * asym_packing force force N/A N/A force force
10805 * imbalanced force force N/A N/A force force
10806 * overloaded force force N/A N/A force avg_load
10808 * N/A : Not Applicable because already filtered while updating
10810 * balanced : The system is balanced for these 2 groups.
10811 * force : Calculate the imbalance as load migration is probably needed.
10812 * avg_load : Only if imbalance is significant enough.
10813 * nr_idle : dst_cpu is not busy and the number of idle CPUs is quite
10814 * different in groups.
10818 * find_busiest_group - Returns the busiest group within the sched_domain
10819 * if there is an imbalance.
10820 * @env: The load balancing environment.
10822 * Also calculates the amount of runnable load which should be moved
10823 * to restore balance.
10825 * Return: - The busiest group if imbalance exists.
10827 static struct sched_group *find_busiest_group(struct lb_env *env)
10829 struct sg_lb_stats *local, *busiest;
10830 struct sd_lb_stats sds;
10832 init_sd_lb_stats(&sds);
10835 * Compute the various statistics relevant for load balancing at
10838 update_sd_lb_stats(env, &sds);
10840 /* There is no busy sibling group to pull tasks from */
10844 busiest = &sds.busiest_stat;
10846 /* Misfit tasks should be dealt with regardless of the avg load */
10847 if (busiest->group_type == group_misfit_task)
10848 goto force_balance;
10850 if (sched_energy_enabled()) {
10851 struct root_domain *rd = env->dst_rq->rd;
10853 if (rcu_dereference(rd->pd) && !READ_ONCE(rd->overutilized))
10857 /* ASYM feature bypasses nice load balance check */
10858 if (busiest->group_type == group_asym_packing)
10859 goto force_balance;
10862 * If the busiest group is imbalanced the below checks don't
10863 * work because they assume all things are equal, which typically
10864 * isn't true due to cpus_ptr constraints and the like.
10866 if (busiest->group_type == group_imbalanced)
10867 goto force_balance;
10869 local = &sds.local_stat;
10871 * If the local group is busier than the selected busiest group
10872 * don't try and pull any tasks.
10874 if (local->group_type > busiest->group_type)
10878 * When groups are overloaded, use the avg_load to ensure fairness
10881 if (local->group_type == group_overloaded) {
10883 * If the local group is more loaded than the selected
10884 * busiest group don't try to pull any tasks.
10886 if (local->avg_load >= busiest->avg_load)
10889 /* XXX broken for overlapping NUMA groups */
10890 sds.avg_load = (sds.total_load * SCHED_CAPACITY_SCALE) /
10891 sds.total_capacity;
10894 * Don't pull any tasks if this group is already above the
10895 * domain average load.
10897 if (local->avg_load >= sds.avg_load)
10901 * If the busiest group is more loaded, use imbalance_pct to be
10904 if (100 * busiest->avg_load <=
10905 env->sd->imbalance_pct * local->avg_load)
10910 * Try to move all excess tasks to a sibling domain of the busiest
10911 * group's child domain.
10913 if (sds.prefer_sibling && local->group_type == group_has_spare &&
10914 sibling_imbalance(env, &sds, busiest, local) > 1)
10915 goto force_balance;
10917 if (busiest->group_type != group_overloaded) {
10918 if (env->idle == CPU_NOT_IDLE) {
10920 * If the busiest group is not overloaded (and as a
10921 * result the local one too) but this CPU is already
10922 * busy, let another idle CPU try to pull task.
10927 if (busiest->group_type == group_smt_balance &&
10928 smt_vs_nonsmt_groups(sds.local, sds.busiest)) {
10929 /* Let non SMT CPU pull from SMT CPU sharing with sibling */
10930 goto force_balance;
10933 if (busiest->group_weight > 1 &&
10934 local->idle_cpus <= (busiest->idle_cpus + 1)) {
10936 * If the busiest group is not overloaded
10937 * and there is no imbalance between this and busiest
10938 * group wrt idle CPUs, it is balanced. The imbalance
10939 * becomes significant if the diff is greater than 1
10940 * otherwise we might end up to just move the imbalance
10941 * on another group. Of course this applies only if
10942 * there is more than 1 CPU per group.
10947 if (busiest->sum_h_nr_running == 1) {
10949 * busiest doesn't have any tasks waiting to run
10956 /* Looks like there is an imbalance. Compute it */
10957 calculate_imbalance(env, &sds);
10958 return env->imbalance ? sds.busiest : NULL;
10961 env->imbalance = 0;
10966 * find_busiest_queue - find the busiest runqueue among the CPUs in the group.
10968 static struct rq *find_busiest_queue(struct lb_env *env,
10969 struct sched_group *group)
10971 struct rq *busiest = NULL, *rq;
10972 unsigned long busiest_util = 0, busiest_load = 0, busiest_capacity = 1;
10973 unsigned int busiest_nr = 0;
10976 for_each_cpu_and(i, sched_group_span(group), env->cpus) {
10977 unsigned long capacity, load, util;
10978 unsigned int nr_running;
10982 rt = fbq_classify_rq(rq);
10985 * We classify groups/runqueues into three groups:
10986 * - regular: there are !numa tasks
10987 * - remote: there are numa tasks that run on the 'wrong' node
10988 * - all: there is no distinction
10990 * In order to avoid migrating ideally placed numa tasks,
10991 * ignore those when there's better options.
10993 * If we ignore the actual busiest queue to migrate another
10994 * task, the next balance pass can still reduce the busiest
10995 * queue by moving tasks around inside the node.
10997 * If we cannot move enough load due to this classification
10998 * the next pass will adjust the group classification and
10999 * allow migration of more tasks.
11001 * Both cases only affect the total convergence complexity.
11003 if (rt > env->fbq_type)
11006 nr_running = rq->cfs.h_nr_running;
11010 capacity = capacity_of(i);
11013 * For ASYM_CPUCAPACITY domains, don't pick a CPU that could
11014 * eventually lead to active_balancing high->low capacity.
11015 * Higher per-CPU capacity is considered better than balancing
11018 if (env->sd->flags & SD_ASYM_CPUCAPACITY &&
11019 !capacity_greater(capacity_of(env->dst_cpu), capacity) &&
11024 * Make sure we only pull tasks from a CPU of lower priority
11025 * when balancing between SMT siblings.
11027 * If balancing between cores, let lower priority CPUs help
11028 * SMT cores with more than one busy sibling.
11030 if ((env->sd->flags & SD_ASYM_PACKING) &&
11031 sched_use_asym_prio(env->sd, i) &&
11032 sched_asym_prefer(i, env->dst_cpu) &&
11036 switch (env->migration_type) {
11039 * When comparing with load imbalance, use cpu_load()
11040 * which is not scaled with the CPU capacity.
11042 load = cpu_load(rq);
11044 if (nr_running == 1 && load > env->imbalance &&
11045 !check_cpu_capacity(rq, env->sd))
11049 * For the load comparisons with the other CPUs,
11050 * consider the cpu_load() scaled with the CPU
11051 * capacity, so that the load can be moved away
11052 * from the CPU that is potentially running at a
11055 * Thus we're looking for max(load_i / capacity_i),
11056 * crosswise multiplication to rid ourselves of the
11057 * division works out to:
11058 * load_i * capacity_j > load_j * capacity_i;
11059 * where j is our previous maximum.
11061 if (load * busiest_capacity > busiest_load * capacity) {
11062 busiest_load = load;
11063 busiest_capacity = capacity;
11069 util = cpu_util_cfs_boost(i);
11072 * Don't try to pull utilization from a CPU with one
11073 * running task. Whatever its utilization, we will fail
11076 if (nr_running <= 1)
11079 if (busiest_util < util) {
11080 busiest_util = util;
11086 if (busiest_nr < nr_running) {
11087 busiest_nr = nr_running;
11092 case migrate_misfit:
11094 * For ASYM_CPUCAPACITY domains with misfit tasks we
11095 * simply seek the "biggest" misfit task.
11097 if (rq->misfit_task_load > busiest_load) {
11098 busiest_load = rq->misfit_task_load;
11111 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
11112 * so long as it is large enough.
11114 #define MAX_PINNED_INTERVAL 512
11117 asym_active_balance(struct lb_env *env)
11120 * ASYM_PACKING needs to force migrate tasks from busy but lower
11121 * priority CPUs in order to pack all tasks in the highest priority
11122 * CPUs. When done between cores, do it only if the whole core if the
11123 * whole core is idle.
11125 * If @env::src_cpu is an SMT core with busy siblings, let
11126 * the lower priority @env::dst_cpu help it. Do not follow
11129 return env->idle != CPU_NOT_IDLE && (env->sd->flags & SD_ASYM_PACKING) &&
11130 sched_use_asym_prio(env->sd, env->dst_cpu) &&
11131 (sched_asym_prefer(env->dst_cpu, env->src_cpu) ||
11132 !sched_use_asym_prio(env->sd, env->src_cpu));
11136 imbalanced_active_balance(struct lb_env *env)
11138 struct sched_domain *sd = env->sd;
11141 * The imbalanced case includes the case of pinned tasks preventing a fair
11142 * distribution of the load on the system but also the even distribution of the
11143 * threads on a system with spare capacity
11145 if ((env->migration_type == migrate_task) &&
11146 (sd->nr_balance_failed > sd->cache_nice_tries+2))
11152 static int need_active_balance(struct lb_env *env)
11154 struct sched_domain *sd = env->sd;
11156 if (asym_active_balance(env))
11159 if (imbalanced_active_balance(env))
11163 * The dst_cpu is idle and the src_cpu CPU has only 1 CFS task.
11164 * It's worth migrating the task if the src_cpu's capacity is reduced
11165 * because of other sched_class or IRQs if more capacity stays
11166 * available on dst_cpu.
11168 if ((env->idle != CPU_NOT_IDLE) &&
11169 (env->src_rq->cfs.h_nr_running == 1)) {
11170 if ((check_cpu_capacity(env->src_rq, sd)) &&
11171 (capacity_of(env->src_cpu)*sd->imbalance_pct < capacity_of(env->dst_cpu)*100))
11175 if (env->migration_type == migrate_misfit)
11181 static int active_load_balance_cpu_stop(void *data);
11183 static int should_we_balance(struct lb_env *env)
11185 struct cpumask *swb_cpus = this_cpu_cpumask_var_ptr(should_we_balance_tmpmask);
11186 struct sched_group *sg = env->sd->groups;
11187 int cpu, idle_smt = -1;
11190 * Ensure the balancing environment is consistent; can happen
11191 * when the softirq triggers 'during' hotplug.
11193 if (!cpumask_test_cpu(env->dst_cpu, env->cpus))
11197 * In the newly idle case, we will allow all the CPUs
11198 * to do the newly idle load balance.
11200 * However, we bail out if we already have tasks or a wakeup pending,
11201 * to optimize wakeup latency.
11203 if (env->idle == CPU_NEWLY_IDLE) {
11204 if (env->dst_rq->nr_running > 0 || env->dst_rq->ttwu_pending)
11209 cpumask_copy(swb_cpus, group_balance_mask(sg));
11210 /* Try to find first idle CPU */
11211 for_each_cpu_and(cpu, swb_cpus, env->cpus) {
11212 if (!idle_cpu(cpu))
11216 * Don't balance to idle SMT in busy core right away when
11217 * balancing cores, but remember the first idle SMT CPU for
11218 * later consideration. Find CPU on an idle core first.
11220 if (!(env->sd->flags & SD_SHARE_CPUCAPACITY) && !is_core_idle(cpu)) {
11221 if (idle_smt == -1)
11224 * If the core is not idle, and first SMT sibling which is
11225 * idle has been found, then its not needed to check other
11226 * SMT siblings for idleness:
11228 #ifdef CONFIG_SCHED_SMT
11229 cpumask_andnot(swb_cpus, swb_cpus, cpu_smt_mask(cpu));
11235 * Are we the first idle core in a non-SMT domain or higher,
11236 * or the first idle CPU in a SMT domain?
11238 return cpu == env->dst_cpu;
11241 /* Are we the first idle CPU with busy siblings? */
11242 if (idle_smt != -1)
11243 return idle_smt == env->dst_cpu;
11245 /* Are we the first CPU of this group ? */
11246 return group_balance_cpu(sg) == env->dst_cpu;
11250 * Check this_cpu to ensure it is balanced within domain. Attempt to move
11251 * tasks if there is an imbalance.
11253 static int load_balance(int this_cpu, struct rq *this_rq,
11254 struct sched_domain *sd, enum cpu_idle_type idle,
11255 int *continue_balancing)
11257 int ld_moved, cur_ld_moved, active_balance = 0;
11258 struct sched_domain *sd_parent = sd->parent;
11259 struct sched_group *group;
11260 struct rq *busiest;
11261 struct rq_flags rf;
11262 struct cpumask *cpus = this_cpu_cpumask_var_ptr(load_balance_mask);
11263 struct lb_env env = {
11265 .dst_cpu = this_cpu,
11267 .dst_grpmask = group_balance_mask(sd->groups),
11269 .loop_break = SCHED_NR_MIGRATE_BREAK,
11272 .tasks = LIST_HEAD_INIT(env.tasks),
11275 cpumask_and(cpus, sched_domain_span(sd), cpu_active_mask);
11277 schedstat_inc(sd->lb_count[idle]);
11280 if (!should_we_balance(&env)) {
11281 *continue_balancing = 0;
11285 group = find_busiest_group(&env);
11287 schedstat_inc(sd->lb_nobusyg[idle]);
11291 busiest = find_busiest_queue(&env, group);
11293 schedstat_inc(sd->lb_nobusyq[idle]);
11297 WARN_ON_ONCE(busiest == env.dst_rq);
11299 schedstat_add(sd->lb_imbalance[idle], env.imbalance);
11301 env.src_cpu = busiest->cpu;
11302 env.src_rq = busiest;
11305 /* Clear this flag as soon as we find a pullable task */
11306 env.flags |= LBF_ALL_PINNED;
11307 if (busiest->nr_running > 1) {
11309 * Attempt to move tasks. If find_busiest_group has found
11310 * an imbalance but busiest->nr_running <= 1, the group is
11311 * still unbalanced. ld_moved simply stays zero, so it is
11312 * correctly treated as an imbalance.
11314 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
11317 rq_lock_irqsave(busiest, &rf);
11318 update_rq_clock(busiest);
11321 * cur_ld_moved - load moved in current iteration
11322 * ld_moved - cumulative load moved across iterations
11324 cur_ld_moved = detach_tasks(&env);
11327 * We've detached some tasks from busiest_rq. Every
11328 * task is masked "TASK_ON_RQ_MIGRATING", so we can safely
11329 * unlock busiest->lock, and we are able to be sure
11330 * that nobody can manipulate the tasks in parallel.
11331 * See task_rq_lock() family for the details.
11334 rq_unlock(busiest, &rf);
11336 if (cur_ld_moved) {
11337 attach_tasks(&env);
11338 ld_moved += cur_ld_moved;
11341 local_irq_restore(rf.flags);
11343 if (env.flags & LBF_NEED_BREAK) {
11344 env.flags &= ~LBF_NEED_BREAK;
11345 /* Stop if we tried all running tasks */
11346 if (env.loop < busiest->nr_running)
11351 * Revisit (affine) tasks on src_cpu that couldn't be moved to
11352 * us and move them to an alternate dst_cpu in our sched_group
11353 * where they can run. The upper limit on how many times we
11354 * iterate on same src_cpu is dependent on number of CPUs in our
11357 * This changes load balance semantics a bit on who can move
11358 * load to a given_cpu. In addition to the given_cpu itself
11359 * (or a ilb_cpu acting on its behalf where given_cpu is
11360 * nohz-idle), we now have balance_cpu in a position to move
11361 * load to given_cpu. In rare situations, this may cause
11362 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
11363 * _independently_ and at _same_ time to move some load to
11364 * given_cpu) causing excess load to be moved to given_cpu.
11365 * This however should not happen so much in practice and
11366 * moreover subsequent load balance cycles should correct the
11367 * excess load moved.
11369 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
11371 /* Prevent to re-select dst_cpu via env's CPUs */
11372 __cpumask_clear_cpu(env.dst_cpu, env.cpus);
11374 env.dst_rq = cpu_rq(env.new_dst_cpu);
11375 env.dst_cpu = env.new_dst_cpu;
11376 env.flags &= ~LBF_DST_PINNED;
11378 env.loop_break = SCHED_NR_MIGRATE_BREAK;
11381 * Go back to "more_balance" rather than "redo" since we
11382 * need to continue with same src_cpu.
11388 * We failed to reach balance because of affinity.
11391 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
11393 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0)
11394 *group_imbalance = 1;
11397 /* All tasks on this runqueue were pinned by CPU affinity */
11398 if (unlikely(env.flags & LBF_ALL_PINNED)) {
11399 __cpumask_clear_cpu(cpu_of(busiest), cpus);
11401 * Attempting to continue load balancing at the current
11402 * sched_domain level only makes sense if there are
11403 * active CPUs remaining as possible busiest CPUs to
11404 * pull load from which are not contained within the
11405 * destination group that is receiving any migrated
11408 if (!cpumask_subset(cpus, env.dst_grpmask)) {
11410 env.loop_break = SCHED_NR_MIGRATE_BREAK;
11413 goto out_all_pinned;
11418 schedstat_inc(sd->lb_failed[idle]);
11420 * Increment the failure counter only on periodic balance.
11421 * We do not want newidle balance, which can be very
11422 * frequent, pollute the failure counter causing
11423 * excessive cache_hot migrations and active balances.
11425 if (idle != CPU_NEWLY_IDLE)
11426 sd->nr_balance_failed++;
11428 if (need_active_balance(&env)) {
11429 unsigned long flags;
11431 raw_spin_rq_lock_irqsave(busiest, flags);
11434 * Don't kick the active_load_balance_cpu_stop,
11435 * if the curr task on busiest CPU can't be
11436 * moved to this_cpu:
11438 if (!cpumask_test_cpu(this_cpu, busiest->curr->cpus_ptr)) {
11439 raw_spin_rq_unlock_irqrestore(busiest, flags);
11440 goto out_one_pinned;
11443 /* Record that we found at least one task that could run on this_cpu */
11444 env.flags &= ~LBF_ALL_PINNED;
11447 * ->active_balance synchronizes accesses to
11448 * ->active_balance_work. Once set, it's cleared
11449 * only after active load balance is finished.
11451 if (!busiest->active_balance) {
11452 busiest->active_balance = 1;
11453 busiest->push_cpu = this_cpu;
11454 active_balance = 1;
11458 raw_spin_rq_unlock_irqrestore(busiest, flags);
11459 if (active_balance) {
11460 stop_one_cpu_nowait(cpu_of(busiest),
11461 active_load_balance_cpu_stop, busiest,
11462 &busiest->active_balance_work);
11467 sd->nr_balance_failed = 0;
11470 if (likely(!active_balance) || need_active_balance(&env)) {
11471 /* We were unbalanced, so reset the balancing interval */
11472 sd->balance_interval = sd->min_interval;
11479 * We reach balance although we may have faced some affinity
11480 * constraints. Clear the imbalance flag only if other tasks got
11481 * a chance to move and fix the imbalance.
11483 if (sd_parent && !(env.flags & LBF_ALL_PINNED)) {
11484 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
11486 if (*group_imbalance)
11487 *group_imbalance = 0;
11492 * We reach balance because all tasks are pinned at this level so
11493 * we can't migrate them. Let the imbalance flag set so parent level
11494 * can try to migrate them.
11496 schedstat_inc(sd->lb_balanced[idle]);
11498 sd->nr_balance_failed = 0;
11504 * newidle_balance() disregards balance intervals, so we could
11505 * repeatedly reach this code, which would lead to balance_interval
11506 * skyrocketing in a short amount of time. Skip the balance_interval
11507 * increase logic to avoid that.
11509 if (env.idle == CPU_NEWLY_IDLE)
11512 /* tune up the balancing interval */
11513 if ((env.flags & LBF_ALL_PINNED &&
11514 sd->balance_interval < MAX_PINNED_INTERVAL) ||
11515 sd->balance_interval < sd->max_interval)
11516 sd->balance_interval *= 2;
11521 static inline unsigned long
11522 get_sd_balance_interval(struct sched_domain *sd, int cpu_busy)
11524 unsigned long interval = sd->balance_interval;
11527 interval *= sd->busy_factor;
11529 /* scale ms to jiffies */
11530 interval = msecs_to_jiffies(interval);
11533 * Reduce likelihood of busy balancing at higher domains racing with
11534 * balancing at lower domains by preventing their balancing periods
11535 * from being multiples of each other.
11540 interval = clamp(interval, 1UL, max_load_balance_interval);
11546 update_next_balance(struct sched_domain *sd, unsigned long *next_balance)
11548 unsigned long interval, next;
11550 /* used by idle balance, so cpu_busy = 0 */
11551 interval = get_sd_balance_interval(sd, 0);
11552 next = sd->last_balance + interval;
11554 if (time_after(*next_balance, next))
11555 *next_balance = next;
11559 * active_load_balance_cpu_stop is run by the CPU stopper. It pushes
11560 * running tasks off the busiest CPU onto idle CPUs. It requires at
11561 * least 1 task to be running on each physical CPU where possible, and
11562 * avoids physical / logical imbalances.
11564 static int active_load_balance_cpu_stop(void *data)
11566 struct rq *busiest_rq = data;
11567 int busiest_cpu = cpu_of(busiest_rq);
11568 int target_cpu = busiest_rq->push_cpu;
11569 struct rq *target_rq = cpu_rq(target_cpu);
11570 struct sched_domain *sd;
11571 struct task_struct *p = NULL;
11572 struct rq_flags rf;
11574 rq_lock_irq(busiest_rq, &rf);
11576 * Between queueing the stop-work and running it is a hole in which
11577 * CPUs can become inactive. We should not move tasks from or to
11580 if (!cpu_active(busiest_cpu) || !cpu_active(target_cpu))
11583 /* Make sure the requested CPU hasn't gone down in the meantime: */
11584 if (unlikely(busiest_cpu != smp_processor_id() ||
11585 !busiest_rq->active_balance))
11588 /* Is there any task to move? */
11589 if (busiest_rq->nr_running <= 1)
11593 * This condition is "impossible", if it occurs
11594 * we need to fix it. Originally reported by
11595 * Bjorn Helgaas on a 128-CPU setup.
11597 WARN_ON_ONCE(busiest_rq == target_rq);
11599 /* Search for an sd spanning us and the target CPU. */
11601 for_each_domain(target_cpu, sd) {
11602 if (cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
11607 struct lb_env env = {
11609 .dst_cpu = target_cpu,
11610 .dst_rq = target_rq,
11611 .src_cpu = busiest_rq->cpu,
11612 .src_rq = busiest_rq,
11614 .flags = LBF_ACTIVE_LB,
11617 schedstat_inc(sd->alb_count);
11618 update_rq_clock(busiest_rq);
11620 p = detach_one_task(&env);
11622 schedstat_inc(sd->alb_pushed);
11623 /* Active balancing done, reset the failure counter. */
11624 sd->nr_balance_failed = 0;
11626 schedstat_inc(sd->alb_failed);
11631 busiest_rq->active_balance = 0;
11632 rq_unlock(busiest_rq, &rf);
11635 attach_one_task(target_rq, p);
11637 local_irq_enable();
11642 static DEFINE_SPINLOCK(balancing);
11645 * Scale the max load_balance interval with the number of CPUs in the system.
11646 * This trades load-balance latency on larger machines for less cross talk.
11648 void update_max_interval(void)
11650 max_load_balance_interval = HZ*num_online_cpus()/10;
11653 static inline bool update_newidle_cost(struct sched_domain *sd, u64 cost)
11655 if (cost > sd->max_newidle_lb_cost) {
11657 * Track max cost of a domain to make sure to not delay the
11658 * next wakeup on the CPU.
11660 sd->max_newidle_lb_cost = cost;
11661 sd->last_decay_max_lb_cost = jiffies;
11662 } else if (time_after(jiffies, sd->last_decay_max_lb_cost + HZ)) {
11664 * Decay the newidle max times by ~1% per second to ensure that
11665 * it is not outdated and the current max cost is actually
11668 sd->max_newidle_lb_cost = (sd->max_newidle_lb_cost * 253) / 256;
11669 sd->last_decay_max_lb_cost = jiffies;
11678 * It checks each scheduling domain to see if it is due to be balanced,
11679 * and initiates a balancing operation if so.
11681 * Balancing parameters are set up in init_sched_domains.
11683 static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle)
11685 int continue_balancing = 1;
11687 int busy = idle != CPU_IDLE && !sched_idle_cpu(cpu);
11688 unsigned long interval;
11689 struct sched_domain *sd;
11690 /* Earliest time when we have to do rebalance again */
11691 unsigned long next_balance = jiffies + 60*HZ;
11692 int update_next_balance = 0;
11693 int need_serialize, need_decay = 0;
11697 for_each_domain(cpu, sd) {
11699 * Decay the newidle max times here because this is a regular
11700 * visit to all the domains.
11702 need_decay = update_newidle_cost(sd, 0);
11703 max_cost += sd->max_newidle_lb_cost;
11706 * Stop the load balance at this level. There is another
11707 * CPU in our sched group which is doing load balancing more
11710 if (!continue_balancing) {
11716 interval = get_sd_balance_interval(sd, busy);
11718 need_serialize = sd->flags & SD_SERIALIZE;
11719 if (need_serialize) {
11720 if (!spin_trylock(&balancing))
11724 if (time_after_eq(jiffies, sd->last_balance + interval)) {
11725 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
11727 * The LBF_DST_PINNED logic could have changed
11728 * env->dst_cpu, so we can't know our idle
11729 * state even if we migrated tasks. Update it.
11731 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
11732 busy = idle != CPU_IDLE && !sched_idle_cpu(cpu);
11734 sd->last_balance = jiffies;
11735 interval = get_sd_balance_interval(sd, busy);
11737 if (need_serialize)
11738 spin_unlock(&balancing);
11740 if (time_after(next_balance, sd->last_balance + interval)) {
11741 next_balance = sd->last_balance + interval;
11742 update_next_balance = 1;
11747 * Ensure the rq-wide value also decays but keep it at a
11748 * reasonable floor to avoid funnies with rq->avg_idle.
11750 rq->max_idle_balance_cost =
11751 max((u64)sysctl_sched_migration_cost, max_cost);
11756 * next_balance will be updated only when there is a need.
11757 * When the cpu is attached to null domain for ex, it will not be
11760 if (likely(update_next_balance))
11761 rq->next_balance = next_balance;
11765 static inline int on_null_domain(struct rq *rq)
11767 return unlikely(!rcu_dereference_sched(rq->sd));
11770 #ifdef CONFIG_NO_HZ_COMMON
11772 * NOHZ idle load balancing (ILB) details:
11774 * - When one of the busy CPUs notices that there may be an idle rebalancing
11775 * needed, they will kick the idle load balancer, which then does idle
11776 * load balancing for all the idle CPUs.
11778 * - HK_TYPE_MISC CPUs are used for this task, because HK_TYPE_SCHED is not set
11781 static inline int find_new_ilb(void)
11783 const struct cpumask *hk_mask;
11786 hk_mask = housekeeping_cpumask(HK_TYPE_MISC);
11788 for_each_cpu_and(ilb_cpu, nohz.idle_cpus_mask, hk_mask) {
11790 if (ilb_cpu == smp_processor_id())
11793 if (idle_cpu(ilb_cpu))
11801 * Kick a CPU to do the NOHZ balancing, if it is time for it, via a cross-CPU
11802 * SMP function call (IPI).
11804 * We pick the first idle CPU in the HK_TYPE_MISC housekeeping set (if there is one).
11806 static void kick_ilb(unsigned int flags)
11811 * Increase nohz.next_balance only when if full ilb is triggered but
11812 * not if we only update stats.
11814 if (flags & NOHZ_BALANCE_KICK)
11815 nohz.next_balance = jiffies+1;
11817 ilb_cpu = find_new_ilb();
11822 * Access to rq::nohz_csd is serialized by NOHZ_KICK_MASK; he who sets
11823 * the first flag owns it; cleared by nohz_csd_func().
11825 flags = atomic_fetch_or(flags, nohz_flags(ilb_cpu));
11826 if (flags & NOHZ_KICK_MASK)
11830 * This way we generate an IPI on the target CPU which
11831 * is idle, and the softirq performing NOHZ idle load balancing
11832 * will be run before returning from the IPI.
11834 smp_call_function_single_async(ilb_cpu, &cpu_rq(ilb_cpu)->nohz_csd);
11838 * Current decision point for kicking the idle load balancer in the presence
11839 * of idle CPUs in the system.
11841 static void nohz_balancer_kick(struct rq *rq)
11843 unsigned long now = jiffies;
11844 struct sched_domain_shared *sds;
11845 struct sched_domain *sd;
11846 int nr_busy, i, cpu = rq->cpu;
11847 unsigned int flags = 0;
11849 if (unlikely(rq->idle_balance))
11853 * We may be recently in ticked or tickless idle mode. At the first
11854 * busy tick after returning from idle, we will update the busy stats.
11856 nohz_balance_exit_idle(rq);
11859 * None are in tickless mode and hence no need for NOHZ idle load
11862 if (likely(!atomic_read(&nohz.nr_cpus)))
11865 if (READ_ONCE(nohz.has_blocked) &&
11866 time_after(now, READ_ONCE(nohz.next_blocked)))
11867 flags = NOHZ_STATS_KICK;
11869 if (time_before(now, nohz.next_balance))
11872 if (rq->nr_running >= 2) {
11873 flags = NOHZ_STATS_KICK | NOHZ_BALANCE_KICK;
11879 sd = rcu_dereference(rq->sd);
11882 * If there's a runnable CFS task and the current CPU has reduced
11883 * capacity, kick the ILB to see if there's a better CPU to run on:
11885 if (rq->cfs.h_nr_running >= 1 && check_cpu_capacity(rq, sd)) {
11886 flags = NOHZ_STATS_KICK | NOHZ_BALANCE_KICK;
11891 sd = rcu_dereference(per_cpu(sd_asym_packing, cpu));
11894 * When ASYM_PACKING; see if there's a more preferred CPU
11895 * currently idle; in which case, kick the ILB to move tasks
11898 * When balancing betwen cores, all the SMT siblings of the
11899 * preferred CPU must be idle.
11901 for_each_cpu_and(i, sched_domain_span(sd), nohz.idle_cpus_mask) {
11902 if (sched_use_asym_prio(sd, i) &&
11903 sched_asym_prefer(i, cpu)) {
11904 flags = NOHZ_STATS_KICK | NOHZ_BALANCE_KICK;
11910 sd = rcu_dereference(per_cpu(sd_asym_cpucapacity, cpu));
11913 * When ASYM_CPUCAPACITY; see if there's a higher capacity CPU
11914 * to run the misfit task on.
11916 if (check_misfit_status(rq, sd)) {
11917 flags = NOHZ_STATS_KICK | NOHZ_BALANCE_KICK;
11922 * For asymmetric systems, we do not want to nicely balance
11923 * cache use, instead we want to embrace asymmetry and only
11924 * ensure tasks have enough CPU capacity.
11926 * Skip the LLC logic because it's not relevant in that case.
11931 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
11934 * If there is an imbalance between LLC domains (IOW we could
11935 * increase the overall cache utilization), we need a less-loaded LLC
11936 * domain to pull some load from. Likewise, we may need to spread
11937 * load within the current LLC domain (e.g. packed SMT cores but
11938 * other CPUs are idle). We can't really know from here how busy
11939 * the others are - so just get a NOHZ balance going if it looks
11940 * like this LLC domain has tasks we could move.
11942 nr_busy = atomic_read(&sds->nr_busy_cpus);
11944 flags = NOHZ_STATS_KICK | NOHZ_BALANCE_KICK;
11951 if (READ_ONCE(nohz.needs_update))
11952 flags |= NOHZ_NEXT_KICK;
11958 static void set_cpu_sd_state_busy(int cpu)
11960 struct sched_domain *sd;
11963 sd = rcu_dereference(per_cpu(sd_llc, cpu));
11965 if (!sd || !sd->nohz_idle)
11969 atomic_inc(&sd->shared->nr_busy_cpus);
11974 void nohz_balance_exit_idle(struct rq *rq)
11976 SCHED_WARN_ON(rq != this_rq());
11978 if (likely(!rq->nohz_tick_stopped))
11981 rq->nohz_tick_stopped = 0;
11982 cpumask_clear_cpu(rq->cpu, nohz.idle_cpus_mask);
11983 atomic_dec(&nohz.nr_cpus);
11985 set_cpu_sd_state_busy(rq->cpu);
11988 static void set_cpu_sd_state_idle(int cpu)
11990 struct sched_domain *sd;
11993 sd = rcu_dereference(per_cpu(sd_llc, cpu));
11995 if (!sd || sd->nohz_idle)
11999 atomic_dec(&sd->shared->nr_busy_cpus);
12005 * This routine will record that the CPU is going idle with tick stopped.
12006 * This info will be used in performing idle load balancing in the future.
12008 void nohz_balance_enter_idle(int cpu)
12010 struct rq *rq = cpu_rq(cpu);
12012 SCHED_WARN_ON(cpu != smp_processor_id());
12014 /* If this CPU is going down, then nothing needs to be done: */
12015 if (!cpu_active(cpu))
12018 /* Spare idle load balancing on CPUs that don't want to be disturbed: */
12019 if (!housekeeping_cpu(cpu, HK_TYPE_SCHED))
12023 * Can be set safely without rq->lock held
12024 * If a clear happens, it will have evaluated last additions because
12025 * rq->lock is held during the check and the clear
12027 rq->has_blocked_load = 1;
12030 * The tick is still stopped but load could have been added in the
12031 * meantime. We set the nohz.has_blocked flag to trig a check of the
12032 * *_avg. The CPU is already part of nohz.idle_cpus_mask so the clear
12033 * of nohz.has_blocked can only happen after checking the new load
12035 if (rq->nohz_tick_stopped)
12038 /* If we're a completely isolated CPU, we don't play: */
12039 if (on_null_domain(rq))
12042 rq->nohz_tick_stopped = 1;
12044 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
12045 atomic_inc(&nohz.nr_cpus);
12048 * Ensures that if nohz_idle_balance() fails to observe our
12049 * @idle_cpus_mask store, it must observe the @has_blocked
12050 * and @needs_update stores.
12052 smp_mb__after_atomic();
12054 set_cpu_sd_state_idle(cpu);
12056 WRITE_ONCE(nohz.needs_update, 1);
12059 * Each time a cpu enter idle, we assume that it has blocked load and
12060 * enable the periodic update of the load of idle cpus
12062 WRITE_ONCE(nohz.has_blocked, 1);
12065 static bool update_nohz_stats(struct rq *rq)
12067 unsigned int cpu = rq->cpu;
12069 if (!rq->has_blocked_load)
12072 if (!cpumask_test_cpu(cpu, nohz.idle_cpus_mask))
12075 if (!time_after(jiffies, READ_ONCE(rq->last_blocked_load_update_tick)))
12078 update_blocked_averages(cpu);
12080 return rq->has_blocked_load;
12084 * Internal function that runs load balance for all idle cpus. The load balance
12085 * can be a simple update of blocked load or a complete load balance with
12086 * tasks movement depending of flags.
12088 static void _nohz_idle_balance(struct rq *this_rq, unsigned int flags)
12090 /* Earliest time when we have to do rebalance again */
12091 unsigned long now = jiffies;
12092 unsigned long next_balance = now + 60*HZ;
12093 bool has_blocked_load = false;
12094 int update_next_balance = 0;
12095 int this_cpu = this_rq->cpu;
12099 SCHED_WARN_ON((flags & NOHZ_KICK_MASK) == NOHZ_BALANCE_KICK);
12102 * We assume there will be no idle load after this update and clear
12103 * the has_blocked flag. If a cpu enters idle in the mean time, it will
12104 * set the has_blocked flag and trigger another update of idle load.
12105 * Because a cpu that becomes idle, is added to idle_cpus_mask before
12106 * setting the flag, we are sure to not clear the state and not
12107 * check the load of an idle cpu.
12109 * Same applies to idle_cpus_mask vs needs_update.
12111 if (flags & NOHZ_STATS_KICK)
12112 WRITE_ONCE(nohz.has_blocked, 0);
12113 if (flags & NOHZ_NEXT_KICK)
12114 WRITE_ONCE(nohz.needs_update, 0);
12117 * Ensures that if we miss the CPU, we must see the has_blocked
12118 * store from nohz_balance_enter_idle().
12123 * Start with the next CPU after this_cpu so we will end with this_cpu and let a
12124 * chance for other idle cpu to pull load.
12126 for_each_cpu_wrap(balance_cpu, nohz.idle_cpus_mask, this_cpu+1) {
12127 if (!idle_cpu(balance_cpu))
12131 * If this CPU gets work to do, stop the load balancing
12132 * work being done for other CPUs. Next load
12133 * balancing owner will pick it up.
12135 if (need_resched()) {
12136 if (flags & NOHZ_STATS_KICK)
12137 has_blocked_load = true;
12138 if (flags & NOHZ_NEXT_KICK)
12139 WRITE_ONCE(nohz.needs_update, 1);
12143 rq = cpu_rq(balance_cpu);
12145 if (flags & NOHZ_STATS_KICK)
12146 has_blocked_load |= update_nohz_stats(rq);
12149 * If time for next balance is due,
12152 if (time_after_eq(jiffies, rq->next_balance)) {
12153 struct rq_flags rf;
12155 rq_lock_irqsave(rq, &rf);
12156 update_rq_clock(rq);
12157 rq_unlock_irqrestore(rq, &rf);
12159 if (flags & NOHZ_BALANCE_KICK)
12160 rebalance_domains(rq, CPU_IDLE);
12163 if (time_after(next_balance, rq->next_balance)) {
12164 next_balance = rq->next_balance;
12165 update_next_balance = 1;
12170 * next_balance will be updated only when there is a need.
12171 * When the CPU is attached to null domain for ex, it will not be
12174 if (likely(update_next_balance))
12175 nohz.next_balance = next_balance;
12177 if (flags & NOHZ_STATS_KICK)
12178 WRITE_ONCE(nohz.next_blocked,
12179 now + msecs_to_jiffies(LOAD_AVG_PERIOD));
12182 /* There is still blocked load, enable periodic update */
12183 if (has_blocked_load)
12184 WRITE_ONCE(nohz.has_blocked, 1);
12188 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
12189 * rebalancing for all the cpus for whom scheduler ticks are stopped.
12191 static bool nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
12193 unsigned int flags = this_rq->nohz_idle_balance;
12198 this_rq->nohz_idle_balance = 0;
12200 if (idle != CPU_IDLE)
12203 _nohz_idle_balance(this_rq, flags);
12209 * Check if we need to directly run the ILB for updating blocked load before
12210 * entering idle state. Here we run ILB directly without issuing IPIs.
12212 * Note that when this function is called, the tick may not yet be stopped on
12213 * this CPU yet. nohz.idle_cpus_mask is updated only when tick is stopped and
12214 * cleared on the next busy tick. In other words, nohz.idle_cpus_mask updates
12215 * don't align with CPUs enter/exit idle to avoid bottlenecks due to high idle
12216 * entry/exit rate (usec). So it is possible that _nohz_idle_balance() is
12217 * called from this function on (this) CPU that's not yet in the mask. That's
12218 * OK because the goal of nohz_run_idle_balance() is to run ILB only for
12219 * updating the blocked load of already idle CPUs without waking up one of
12220 * those idle CPUs and outside the preempt disable / irq off phase of the local
12221 * cpu about to enter idle, because it can take a long time.
12223 void nohz_run_idle_balance(int cpu)
12225 unsigned int flags;
12227 flags = atomic_fetch_andnot(NOHZ_NEWILB_KICK, nohz_flags(cpu));
12230 * Update the blocked load only if no SCHED_SOFTIRQ is about to happen
12231 * (ie NOHZ_STATS_KICK set) and will do the same.
12233 if ((flags == NOHZ_NEWILB_KICK) && !need_resched())
12234 _nohz_idle_balance(cpu_rq(cpu), NOHZ_STATS_KICK);
12237 static void nohz_newidle_balance(struct rq *this_rq)
12239 int this_cpu = this_rq->cpu;
12242 * This CPU doesn't want to be disturbed by scheduler
12245 if (!housekeeping_cpu(this_cpu, HK_TYPE_SCHED))
12248 /* Will wake up very soon. No time for doing anything else*/
12249 if (this_rq->avg_idle < sysctl_sched_migration_cost)
12252 /* Don't need to update blocked load of idle CPUs*/
12253 if (!READ_ONCE(nohz.has_blocked) ||
12254 time_before(jiffies, READ_ONCE(nohz.next_blocked)))
12258 * Set the need to trigger ILB in order to update blocked load
12259 * before entering idle state.
12261 atomic_or(NOHZ_NEWILB_KICK, nohz_flags(this_cpu));
12264 #else /* !CONFIG_NO_HZ_COMMON */
12265 static inline void nohz_balancer_kick(struct rq *rq) { }
12267 static inline bool nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
12272 static inline void nohz_newidle_balance(struct rq *this_rq) { }
12273 #endif /* CONFIG_NO_HZ_COMMON */
12276 * newidle_balance is called by schedule() if this_cpu is about to become
12277 * idle. Attempts to pull tasks from other CPUs.
12280 * < 0 - we released the lock and there are !fair tasks present
12281 * 0 - failed, no new tasks
12282 * > 0 - success, new (fair) tasks present
12284 static int newidle_balance(struct rq *this_rq, struct rq_flags *rf)
12286 unsigned long next_balance = jiffies + HZ;
12287 int this_cpu = this_rq->cpu;
12288 u64 t0, t1, curr_cost = 0;
12289 struct sched_domain *sd;
12290 int pulled_task = 0;
12292 update_misfit_status(NULL, this_rq);
12295 * There is a task waiting to run. No need to search for one.
12296 * Return 0; the task will be enqueued when switching to idle.
12298 if (this_rq->ttwu_pending)
12302 * We must set idle_stamp _before_ calling idle_balance(), such that we
12303 * measure the duration of idle_balance() as idle time.
12305 this_rq->idle_stamp = rq_clock(this_rq);
12308 * Do not pull tasks towards !active CPUs...
12310 if (!cpu_active(this_cpu))
12314 * This is OK, because current is on_cpu, which avoids it being picked
12315 * for load-balance and preemption/IRQs are still disabled avoiding
12316 * further scheduler activity on it and we're being very careful to
12317 * re-start the picking loop.
12319 rq_unpin_lock(this_rq, rf);
12322 sd = rcu_dereference_check_sched_domain(this_rq->sd);
12324 if (!READ_ONCE(this_rq->rd->overload) ||
12325 (sd && this_rq->avg_idle < sd->max_newidle_lb_cost)) {
12328 update_next_balance(sd, &next_balance);
12335 raw_spin_rq_unlock(this_rq);
12337 t0 = sched_clock_cpu(this_cpu);
12338 update_blocked_averages(this_cpu);
12341 for_each_domain(this_cpu, sd) {
12342 int continue_balancing = 1;
12345 update_next_balance(sd, &next_balance);
12347 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost)
12350 if (sd->flags & SD_BALANCE_NEWIDLE) {
12352 pulled_task = load_balance(this_cpu, this_rq,
12353 sd, CPU_NEWLY_IDLE,
12354 &continue_balancing);
12356 t1 = sched_clock_cpu(this_cpu);
12357 domain_cost = t1 - t0;
12358 update_newidle_cost(sd, domain_cost);
12360 curr_cost += domain_cost;
12365 * Stop searching for tasks to pull if there are
12366 * now runnable tasks on this rq.
12368 if (pulled_task || this_rq->nr_running > 0 ||
12369 this_rq->ttwu_pending)
12374 raw_spin_rq_lock(this_rq);
12376 if (curr_cost > this_rq->max_idle_balance_cost)
12377 this_rq->max_idle_balance_cost = curr_cost;
12380 * While browsing the domains, we released the rq lock, a task could
12381 * have been enqueued in the meantime. Since we're not going idle,
12382 * pretend we pulled a task.
12384 if (this_rq->cfs.h_nr_running && !pulled_task)
12387 /* Is there a task of a high priority class? */
12388 if (this_rq->nr_running != this_rq->cfs.h_nr_running)
12392 /* Move the next balance forward */
12393 if (time_after(this_rq->next_balance, next_balance))
12394 this_rq->next_balance = next_balance;
12397 this_rq->idle_stamp = 0;
12399 nohz_newidle_balance(this_rq);
12401 rq_repin_lock(this_rq, rf);
12403 return pulled_task;
12407 * run_rebalance_domains is triggered when needed from the scheduler tick.
12408 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
12410 static __latent_entropy void run_rebalance_domains(struct softirq_action *h)
12412 struct rq *this_rq = this_rq();
12413 enum cpu_idle_type idle = this_rq->idle_balance ?
12414 CPU_IDLE : CPU_NOT_IDLE;
12417 * If this CPU has a pending nohz_balance_kick, then do the
12418 * balancing on behalf of the other idle CPUs whose ticks are
12419 * stopped. Do nohz_idle_balance *before* rebalance_domains to
12420 * give the idle CPUs a chance to load balance. Else we may
12421 * load balance only within the local sched_domain hierarchy
12422 * and abort nohz_idle_balance altogether if we pull some load.
12424 if (nohz_idle_balance(this_rq, idle))
12427 /* normal load balance */
12428 update_blocked_averages(this_rq->cpu);
12429 rebalance_domains(this_rq, idle);
12433 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
12435 void trigger_load_balance(struct rq *rq)
12438 * Don't need to rebalance while attached to NULL domain or
12439 * runqueue CPU is not active
12441 if (unlikely(on_null_domain(rq) || !cpu_active(cpu_of(rq))))
12444 if (time_after_eq(jiffies, rq->next_balance))
12445 raise_softirq(SCHED_SOFTIRQ);
12447 nohz_balancer_kick(rq);
12450 static void rq_online_fair(struct rq *rq)
12454 update_runtime_enabled(rq);
12457 static void rq_offline_fair(struct rq *rq)
12461 /* Ensure any throttled groups are reachable by pick_next_task */
12462 unthrottle_offline_cfs_rqs(rq);
12464 /* Ensure that we remove rq contribution to group share: */
12465 clear_tg_offline_cfs_rqs(rq);
12468 #endif /* CONFIG_SMP */
12470 #ifdef CONFIG_SCHED_CORE
12472 __entity_slice_used(struct sched_entity *se, int min_nr_tasks)
12474 u64 rtime = se->sum_exec_runtime - se->prev_sum_exec_runtime;
12475 u64 slice = se->slice;
12477 return (rtime * min_nr_tasks > slice);
12480 #define MIN_NR_TASKS_DURING_FORCEIDLE 2
12481 static inline void task_tick_core(struct rq *rq, struct task_struct *curr)
12483 if (!sched_core_enabled(rq))
12487 * If runqueue has only one task which used up its slice and
12488 * if the sibling is forced idle, then trigger schedule to
12489 * give forced idle task a chance.
12491 * sched_slice() considers only this active rq and it gets the
12492 * whole slice. But during force idle, we have siblings acting
12493 * like a single runqueue and hence we need to consider runnable
12494 * tasks on this CPU and the forced idle CPU. Ideally, we should
12495 * go through the forced idle rq, but that would be a perf hit.
12496 * We can assume that the forced idle CPU has at least
12497 * MIN_NR_TASKS_DURING_FORCEIDLE - 1 tasks and use that to check
12498 * if we need to give up the CPU.
12500 if (rq->core->core_forceidle_count && rq->cfs.nr_running == 1 &&
12501 __entity_slice_used(&curr->se, MIN_NR_TASKS_DURING_FORCEIDLE))
12506 * se_fi_update - Update the cfs_rq->min_vruntime_fi in a CFS hierarchy if needed.
12508 static void se_fi_update(const struct sched_entity *se, unsigned int fi_seq,
12511 for_each_sched_entity(se) {
12512 struct cfs_rq *cfs_rq = cfs_rq_of(se);
12515 if (cfs_rq->forceidle_seq == fi_seq)
12517 cfs_rq->forceidle_seq = fi_seq;
12520 cfs_rq->min_vruntime_fi = cfs_rq->min_vruntime;
12524 void task_vruntime_update(struct rq *rq, struct task_struct *p, bool in_fi)
12526 struct sched_entity *se = &p->se;
12528 if (p->sched_class != &fair_sched_class)
12531 se_fi_update(se, rq->core->core_forceidle_seq, in_fi);
12534 bool cfs_prio_less(const struct task_struct *a, const struct task_struct *b,
12537 struct rq *rq = task_rq(a);
12538 const struct sched_entity *sea = &a->se;
12539 const struct sched_entity *seb = &b->se;
12540 struct cfs_rq *cfs_rqa;
12541 struct cfs_rq *cfs_rqb;
12544 SCHED_WARN_ON(task_rq(b)->core != rq->core);
12546 #ifdef CONFIG_FAIR_GROUP_SCHED
12548 * Find an se in the hierarchy for tasks a and b, such that the se's
12549 * are immediate siblings.
12551 while (sea->cfs_rq->tg != seb->cfs_rq->tg) {
12552 int sea_depth = sea->depth;
12553 int seb_depth = seb->depth;
12555 if (sea_depth >= seb_depth)
12556 sea = parent_entity(sea);
12557 if (sea_depth <= seb_depth)
12558 seb = parent_entity(seb);
12561 se_fi_update(sea, rq->core->core_forceidle_seq, in_fi);
12562 se_fi_update(seb, rq->core->core_forceidle_seq, in_fi);
12564 cfs_rqa = sea->cfs_rq;
12565 cfs_rqb = seb->cfs_rq;
12567 cfs_rqa = &task_rq(a)->cfs;
12568 cfs_rqb = &task_rq(b)->cfs;
12572 * Find delta after normalizing se's vruntime with its cfs_rq's
12573 * min_vruntime_fi, which would have been updated in prior calls
12574 * to se_fi_update().
12576 delta = (s64)(sea->vruntime - seb->vruntime) +
12577 (s64)(cfs_rqb->min_vruntime_fi - cfs_rqa->min_vruntime_fi);
12582 static int task_is_throttled_fair(struct task_struct *p, int cpu)
12584 struct cfs_rq *cfs_rq;
12586 #ifdef CONFIG_FAIR_GROUP_SCHED
12587 cfs_rq = task_group(p)->cfs_rq[cpu];
12589 cfs_rq = &cpu_rq(cpu)->cfs;
12591 return throttled_hierarchy(cfs_rq);
12594 static inline void task_tick_core(struct rq *rq, struct task_struct *curr) {}
12598 * scheduler tick hitting a task of our scheduling class.
12600 * NOTE: This function can be called remotely by the tick offload that
12601 * goes along full dynticks. Therefore no local assumption can be made
12602 * and everything must be accessed through the @rq and @curr passed in
12605 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
12607 struct cfs_rq *cfs_rq;
12608 struct sched_entity *se = &curr->se;
12610 for_each_sched_entity(se) {
12611 cfs_rq = cfs_rq_of(se);
12612 entity_tick(cfs_rq, se, queued);
12615 if (static_branch_unlikely(&sched_numa_balancing))
12616 task_tick_numa(rq, curr);
12618 update_misfit_status(curr, rq);
12619 update_overutilized_status(task_rq(curr));
12621 task_tick_core(rq, curr);
12625 * called on fork with the child task as argument from the parent's context
12626 * - child not yet on the tasklist
12627 * - preemption disabled
12629 static void task_fork_fair(struct task_struct *p)
12631 struct sched_entity *se = &p->se, *curr;
12632 struct cfs_rq *cfs_rq;
12633 struct rq *rq = this_rq();
12634 struct rq_flags rf;
12637 update_rq_clock(rq);
12639 cfs_rq = task_cfs_rq(current);
12640 curr = cfs_rq->curr;
12642 update_curr(cfs_rq);
12643 place_entity(cfs_rq, se, ENQUEUE_INITIAL);
12644 rq_unlock(rq, &rf);
12648 * Priority of the task has changed. Check to see if we preempt
12649 * the current task.
12652 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
12654 if (!task_on_rq_queued(p))
12657 if (rq->cfs.nr_running == 1)
12661 * Reschedule if we are currently running on this runqueue and
12662 * our priority decreased, or if we are not currently running on
12663 * this runqueue and our priority is higher than the current's
12665 if (task_current(rq, p)) {
12666 if (p->prio > oldprio)
12669 wakeup_preempt(rq, p, 0);
12672 #ifdef CONFIG_FAIR_GROUP_SCHED
12674 * Propagate the changes of the sched_entity across the tg tree to make it
12675 * visible to the root
12677 static void propagate_entity_cfs_rq(struct sched_entity *se)
12679 struct cfs_rq *cfs_rq = cfs_rq_of(se);
12681 if (cfs_rq_throttled(cfs_rq))
12684 if (!throttled_hierarchy(cfs_rq))
12685 list_add_leaf_cfs_rq(cfs_rq);
12687 /* Start to propagate at parent */
12690 for_each_sched_entity(se) {
12691 cfs_rq = cfs_rq_of(se);
12693 update_load_avg(cfs_rq, se, UPDATE_TG);
12695 if (cfs_rq_throttled(cfs_rq))
12698 if (!throttled_hierarchy(cfs_rq))
12699 list_add_leaf_cfs_rq(cfs_rq);
12703 static void propagate_entity_cfs_rq(struct sched_entity *se) { }
12706 static void detach_entity_cfs_rq(struct sched_entity *se)
12708 struct cfs_rq *cfs_rq = cfs_rq_of(se);
12712 * In case the task sched_avg hasn't been attached:
12713 * - A forked task which hasn't been woken up by wake_up_new_task().
12714 * - A task which has been woken up by try_to_wake_up() but is
12715 * waiting for actually being woken up by sched_ttwu_pending().
12717 if (!se->avg.last_update_time)
12721 /* Catch up with the cfs_rq and remove our load when we leave */
12722 update_load_avg(cfs_rq, se, 0);
12723 detach_entity_load_avg(cfs_rq, se);
12724 update_tg_load_avg(cfs_rq);
12725 propagate_entity_cfs_rq(se);
12728 static void attach_entity_cfs_rq(struct sched_entity *se)
12730 struct cfs_rq *cfs_rq = cfs_rq_of(se);
12732 /* Synchronize entity with its cfs_rq */
12733 update_load_avg(cfs_rq, se, sched_feat(ATTACH_AGE_LOAD) ? 0 : SKIP_AGE_LOAD);
12734 attach_entity_load_avg(cfs_rq, se);
12735 update_tg_load_avg(cfs_rq);
12736 propagate_entity_cfs_rq(se);
12739 static void detach_task_cfs_rq(struct task_struct *p)
12741 struct sched_entity *se = &p->se;
12743 detach_entity_cfs_rq(se);
12746 static void attach_task_cfs_rq(struct task_struct *p)
12748 struct sched_entity *se = &p->se;
12750 attach_entity_cfs_rq(se);
12753 static void switched_from_fair(struct rq *rq, struct task_struct *p)
12755 detach_task_cfs_rq(p);
12758 static void switched_to_fair(struct rq *rq, struct task_struct *p)
12760 attach_task_cfs_rq(p);
12762 if (task_on_rq_queued(p)) {
12764 * We were most likely switched from sched_rt, so
12765 * kick off the schedule if running, otherwise just see
12766 * if we can still preempt the current task.
12768 if (task_current(rq, p))
12771 wakeup_preempt(rq, p, 0);
12775 /* Account for a task changing its policy or group.
12777 * This routine is mostly called to set cfs_rq->curr field when a task
12778 * migrates between groups/classes.
12780 static void set_next_task_fair(struct rq *rq, struct task_struct *p, bool first)
12782 struct sched_entity *se = &p->se;
12785 if (task_on_rq_queued(p)) {
12787 * Move the next running task to the front of the list, so our
12788 * cfs_tasks list becomes MRU one.
12790 list_move(&se->group_node, &rq->cfs_tasks);
12794 for_each_sched_entity(se) {
12795 struct cfs_rq *cfs_rq = cfs_rq_of(se);
12797 set_next_entity(cfs_rq, se);
12798 /* ensure bandwidth has been allocated on our new cfs_rq */
12799 account_cfs_rq_runtime(cfs_rq, 0);
12803 void init_cfs_rq(struct cfs_rq *cfs_rq)
12805 cfs_rq->tasks_timeline = RB_ROOT_CACHED;
12806 u64_u32_store(cfs_rq->min_vruntime, (u64)(-(1LL << 20)));
12808 raw_spin_lock_init(&cfs_rq->removed.lock);
12812 #ifdef CONFIG_FAIR_GROUP_SCHED
12813 static void task_change_group_fair(struct task_struct *p)
12816 * We couldn't detach or attach a forked task which
12817 * hasn't been woken up by wake_up_new_task().
12819 if (READ_ONCE(p->__state) == TASK_NEW)
12822 detach_task_cfs_rq(p);
12825 /* Tell se's cfs_rq has been changed -- migrated */
12826 p->se.avg.last_update_time = 0;
12828 set_task_rq(p, task_cpu(p));
12829 attach_task_cfs_rq(p);
12832 void free_fair_sched_group(struct task_group *tg)
12836 for_each_possible_cpu(i) {
12838 kfree(tg->cfs_rq[i]);
12847 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
12849 struct sched_entity *se;
12850 struct cfs_rq *cfs_rq;
12853 tg->cfs_rq = kcalloc(nr_cpu_ids, sizeof(cfs_rq), GFP_KERNEL);
12856 tg->se = kcalloc(nr_cpu_ids, sizeof(se), GFP_KERNEL);
12860 tg->shares = NICE_0_LOAD;
12862 init_cfs_bandwidth(tg_cfs_bandwidth(tg), tg_cfs_bandwidth(parent));
12864 for_each_possible_cpu(i) {
12865 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
12866 GFP_KERNEL, cpu_to_node(i));
12870 se = kzalloc_node(sizeof(struct sched_entity_stats),
12871 GFP_KERNEL, cpu_to_node(i));
12875 init_cfs_rq(cfs_rq);
12876 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
12877 init_entity_runnable_average(se);
12888 void online_fair_sched_group(struct task_group *tg)
12890 struct sched_entity *se;
12891 struct rq_flags rf;
12895 for_each_possible_cpu(i) {
12898 rq_lock_irq(rq, &rf);
12899 update_rq_clock(rq);
12900 attach_entity_cfs_rq(se);
12901 sync_throttle(tg, i);
12902 rq_unlock_irq(rq, &rf);
12906 void unregister_fair_sched_group(struct task_group *tg)
12908 unsigned long flags;
12912 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
12914 for_each_possible_cpu(cpu) {
12916 remove_entity_load_avg(tg->se[cpu]);
12919 * Only empty task groups can be destroyed; so we can speculatively
12920 * check on_list without danger of it being re-added.
12922 if (!tg->cfs_rq[cpu]->on_list)
12927 raw_spin_rq_lock_irqsave(rq, flags);
12928 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
12929 raw_spin_rq_unlock_irqrestore(rq, flags);
12933 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
12934 struct sched_entity *se, int cpu,
12935 struct sched_entity *parent)
12937 struct rq *rq = cpu_rq(cpu);
12941 init_cfs_rq_runtime(cfs_rq);
12943 tg->cfs_rq[cpu] = cfs_rq;
12946 /* se could be NULL for root_task_group */
12951 se->cfs_rq = &rq->cfs;
12954 se->cfs_rq = parent->my_q;
12955 se->depth = parent->depth + 1;
12959 /* guarantee group entities always have weight */
12960 update_load_set(&se->load, NICE_0_LOAD);
12961 se->parent = parent;
12964 static DEFINE_MUTEX(shares_mutex);
12966 static int __sched_group_set_shares(struct task_group *tg, unsigned long shares)
12970 lockdep_assert_held(&shares_mutex);
12973 * We can't change the weight of the root cgroup.
12978 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
12980 if (tg->shares == shares)
12983 tg->shares = shares;
12984 for_each_possible_cpu(i) {
12985 struct rq *rq = cpu_rq(i);
12986 struct sched_entity *se = tg->se[i];
12987 struct rq_flags rf;
12989 /* Propagate contribution to hierarchy */
12990 rq_lock_irqsave(rq, &rf);
12991 update_rq_clock(rq);
12992 for_each_sched_entity(se) {
12993 update_load_avg(cfs_rq_of(se), se, UPDATE_TG);
12994 update_cfs_group(se);
12996 rq_unlock_irqrestore(rq, &rf);
13002 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
13006 mutex_lock(&shares_mutex);
13007 if (tg_is_idle(tg))
13010 ret = __sched_group_set_shares(tg, shares);
13011 mutex_unlock(&shares_mutex);
13016 int sched_group_set_idle(struct task_group *tg, long idle)
13020 if (tg == &root_task_group)
13023 if (idle < 0 || idle > 1)
13026 mutex_lock(&shares_mutex);
13028 if (tg->idle == idle) {
13029 mutex_unlock(&shares_mutex);
13035 for_each_possible_cpu(i) {
13036 struct rq *rq = cpu_rq(i);
13037 struct sched_entity *se = tg->se[i];
13038 struct cfs_rq *parent_cfs_rq, *grp_cfs_rq = tg->cfs_rq[i];
13039 bool was_idle = cfs_rq_is_idle(grp_cfs_rq);
13040 long idle_task_delta;
13041 struct rq_flags rf;
13043 rq_lock_irqsave(rq, &rf);
13045 grp_cfs_rq->idle = idle;
13046 if (WARN_ON_ONCE(was_idle == cfs_rq_is_idle(grp_cfs_rq)))
13050 parent_cfs_rq = cfs_rq_of(se);
13051 if (cfs_rq_is_idle(grp_cfs_rq))
13052 parent_cfs_rq->idle_nr_running++;
13054 parent_cfs_rq->idle_nr_running--;
13057 idle_task_delta = grp_cfs_rq->h_nr_running -
13058 grp_cfs_rq->idle_h_nr_running;
13059 if (!cfs_rq_is_idle(grp_cfs_rq))
13060 idle_task_delta *= -1;
13062 for_each_sched_entity(se) {
13063 struct cfs_rq *cfs_rq = cfs_rq_of(se);
13068 cfs_rq->idle_h_nr_running += idle_task_delta;
13070 /* Already accounted at parent level and above. */
13071 if (cfs_rq_is_idle(cfs_rq))
13076 rq_unlock_irqrestore(rq, &rf);
13079 /* Idle groups have minimum weight. */
13080 if (tg_is_idle(tg))
13081 __sched_group_set_shares(tg, scale_load(WEIGHT_IDLEPRIO));
13083 __sched_group_set_shares(tg, NICE_0_LOAD);
13085 mutex_unlock(&shares_mutex);
13089 #endif /* CONFIG_FAIR_GROUP_SCHED */
13092 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
13094 struct sched_entity *se = &task->se;
13095 unsigned int rr_interval = 0;
13098 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
13101 if (rq->cfs.load.weight)
13102 rr_interval = NS_TO_JIFFIES(se->slice);
13104 return rr_interval;
13108 * All the scheduling class methods:
13110 DEFINE_SCHED_CLASS(fair) = {
13112 .enqueue_task = enqueue_task_fair,
13113 .dequeue_task = dequeue_task_fair,
13114 .yield_task = yield_task_fair,
13115 .yield_to_task = yield_to_task_fair,
13117 .wakeup_preempt = check_preempt_wakeup_fair,
13119 .pick_next_task = __pick_next_task_fair,
13120 .put_prev_task = put_prev_task_fair,
13121 .set_next_task = set_next_task_fair,
13124 .balance = balance_fair,
13125 .pick_task = pick_task_fair,
13126 .select_task_rq = select_task_rq_fair,
13127 .migrate_task_rq = migrate_task_rq_fair,
13129 .rq_online = rq_online_fair,
13130 .rq_offline = rq_offline_fair,
13132 .task_dead = task_dead_fair,
13133 .set_cpus_allowed = set_cpus_allowed_common,
13136 .task_tick = task_tick_fair,
13137 .task_fork = task_fork_fair,
13139 .prio_changed = prio_changed_fair,
13140 .switched_from = switched_from_fair,
13141 .switched_to = switched_to_fair,
13143 .get_rr_interval = get_rr_interval_fair,
13145 .update_curr = update_curr_fair,
13147 #ifdef CONFIG_FAIR_GROUP_SCHED
13148 .task_change_group = task_change_group_fair,
13151 #ifdef CONFIG_SCHED_CORE
13152 .task_is_throttled = task_is_throttled_fair,
13155 #ifdef CONFIG_UCLAMP_TASK
13156 .uclamp_enabled = 1,
13160 #ifdef CONFIG_SCHED_DEBUG
13161 void print_cfs_stats(struct seq_file *m, int cpu)
13163 struct cfs_rq *cfs_rq, *pos;
13166 for_each_leaf_cfs_rq_safe(cpu_rq(cpu), cfs_rq, pos)
13167 print_cfs_rq(m, cpu, cfs_rq);
13171 #ifdef CONFIG_NUMA_BALANCING
13172 void show_numa_stats(struct task_struct *p, struct seq_file *m)
13175 unsigned long tsf = 0, tpf = 0, gsf = 0, gpf = 0;
13176 struct numa_group *ng;
13179 ng = rcu_dereference(p->numa_group);
13180 for_each_online_node(node) {
13181 if (p->numa_faults) {
13182 tsf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 0)];
13183 tpf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 1)];
13186 gsf = ng->faults[task_faults_idx(NUMA_MEM, node, 0)],
13187 gpf = ng->faults[task_faults_idx(NUMA_MEM, node, 1)];
13189 print_numa_stats(m, node, tsf, tpf, gsf, gpf);
13193 #endif /* CONFIG_NUMA_BALANCING */
13194 #endif /* CONFIG_SCHED_DEBUG */
13196 __init void init_sched_fair_class(void)
13201 for_each_possible_cpu(i) {
13202 zalloc_cpumask_var_node(&per_cpu(load_balance_mask, i), GFP_KERNEL, cpu_to_node(i));
13203 zalloc_cpumask_var_node(&per_cpu(select_rq_mask, i), GFP_KERNEL, cpu_to_node(i));
13204 zalloc_cpumask_var_node(&per_cpu(should_we_balance_tmpmask, i),
13205 GFP_KERNEL, cpu_to_node(i));
13207 #ifdef CONFIG_CFS_BANDWIDTH
13208 INIT_CSD(&cpu_rq(i)->cfsb_csd, __cfsb_csd_unthrottle, cpu_rq(i));
13209 INIT_LIST_HEAD(&cpu_rq(i)->cfsb_csd_list);
13213 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
13215 #ifdef CONFIG_NO_HZ_COMMON
13216 nohz.next_balance = jiffies;
13217 nohz.next_blocked = jiffies;
13218 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);