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 s64 entity_lag(u64 avruntime, struct sched_entity *se)
703 vlag = avruntime - se->vruntime;
704 limit = calc_delta_fair(max_t(u64, 2*se->slice, TICK_NSEC), se);
706 return clamp(vlag, -limit, limit);
709 static void update_entity_lag(struct cfs_rq *cfs_rq, struct sched_entity *se)
711 SCHED_WARN_ON(!se->on_rq);
713 se->vlag = entity_lag(avg_vruntime(cfs_rq), se);
717 * Entity is eligible once it received less service than it ought to have,
720 * lag_i = S - s_i = w_i*(V - v_i)
722 * lag_i >= 0 -> V >= v_i
725 * V = ------------------ + v
728 * lag_i >= 0 -> \Sum (v_i - v)*w_i >= (v_i - v)*(\Sum w_i)
730 * Note: using 'avg_vruntime() > se->vruntime' is inacurate due
731 * to the loss in precision caused by the division.
733 static int vruntime_eligible(struct cfs_rq *cfs_rq, u64 vruntime)
735 struct sched_entity *curr = cfs_rq->curr;
736 s64 avg = cfs_rq->avg_vruntime;
737 long load = cfs_rq->avg_load;
739 if (curr && curr->on_rq) {
740 unsigned long weight = scale_load_down(curr->load.weight);
742 avg += entity_key(cfs_rq, curr) * weight;
746 return avg >= (s64)(vruntime - cfs_rq->min_vruntime) * load;
749 int entity_eligible(struct cfs_rq *cfs_rq, struct sched_entity *se)
751 return vruntime_eligible(cfs_rq, se->vruntime);
754 static u64 __update_min_vruntime(struct cfs_rq *cfs_rq, u64 vruntime)
756 u64 min_vruntime = cfs_rq->min_vruntime;
758 * open coded max_vruntime() to allow updating avg_vruntime
760 s64 delta = (s64)(vruntime - min_vruntime);
762 avg_vruntime_update(cfs_rq, delta);
763 min_vruntime = vruntime;
768 static void update_min_vruntime(struct cfs_rq *cfs_rq)
770 struct sched_entity *se = __pick_root_entity(cfs_rq);
771 struct sched_entity *curr = cfs_rq->curr;
772 u64 vruntime = cfs_rq->min_vruntime;
776 vruntime = curr->vruntime;
783 vruntime = se->min_vruntime;
785 vruntime = min_vruntime(vruntime, se->min_vruntime);
788 /* ensure we never gain time by being placed backwards. */
789 u64_u32_store(cfs_rq->min_vruntime,
790 __update_min_vruntime(cfs_rq, vruntime));
793 static inline bool __entity_less(struct rb_node *a, const struct rb_node *b)
795 return entity_before(__node_2_se(a), __node_2_se(b));
798 #define vruntime_gt(field, lse, rse) ({ (s64)((lse)->field - (rse)->field) > 0; })
800 static inline void __min_vruntime_update(struct sched_entity *se, struct rb_node *node)
803 struct sched_entity *rse = __node_2_se(node);
804 if (vruntime_gt(min_vruntime, se, rse))
805 se->min_vruntime = rse->min_vruntime;
810 * se->min_vruntime = min(se->vruntime, {left,right}->min_vruntime)
812 static inline bool min_vruntime_update(struct sched_entity *se, bool exit)
814 u64 old_min_vruntime = se->min_vruntime;
815 struct rb_node *node = &se->run_node;
817 se->min_vruntime = se->vruntime;
818 __min_vruntime_update(se, node->rb_right);
819 __min_vruntime_update(se, node->rb_left);
821 return se->min_vruntime == old_min_vruntime;
824 RB_DECLARE_CALLBACKS(static, min_vruntime_cb, struct sched_entity,
825 run_node, min_vruntime, min_vruntime_update);
828 * Enqueue an entity into the rb-tree:
830 static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
832 avg_vruntime_add(cfs_rq, se);
833 se->min_vruntime = se->vruntime;
834 rb_add_augmented_cached(&se->run_node, &cfs_rq->tasks_timeline,
835 __entity_less, &min_vruntime_cb);
838 static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
840 rb_erase_augmented_cached(&se->run_node, &cfs_rq->tasks_timeline,
842 avg_vruntime_sub(cfs_rq, se);
845 struct sched_entity *__pick_root_entity(struct cfs_rq *cfs_rq)
847 struct rb_node *root = cfs_rq->tasks_timeline.rb_root.rb_node;
852 return __node_2_se(root);
855 struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq)
857 struct rb_node *left = rb_first_cached(&cfs_rq->tasks_timeline);
862 return __node_2_se(left);
866 * Earliest Eligible Virtual Deadline First
868 * In order to provide latency guarantees for different request sizes
869 * EEVDF selects the best runnable task from two criteria:
871 * 1) the task must be eligible (must be owed service)
873 * 2) from those tasks that meet 1), we select the one
874 * with the earliest virtual deadline.
876 * We can do this in O(log n) time due to an augmented RB-tree. The
877 * tree keeps the entries sorted on deadline, but also functions as a
878 * heap based on the vruntime by keeping:
880 * se->min_vruntime = min(se->vruntime, se->{left,right}->min_vruntime)
882 * Which allows tree pruning through eligibility.
884 static struct sched_entity *pick_eevdf(struct cfs_rq *cfs_rq)
886 struct rb_node *node = cfs_rq->tasks_timeline.rb_root.rb_node;
887 struct sched_entity *se = __pick_first_entity(cfs_rq);
888 struct sched_entity *curr = cfs_rq->curr;
889 struct sched_entity *best = NULL;
892 * We can safely skip eligibility check if there is only one entity
893 * in this cfs_rq, saving some cycles.
895 if (cfs_rq->nr_running == 1)
896 return curr && curr->on_rq ? curr : se;
898 if (curr && (!curr->on_rq || !entity_eligible(cfs_rq, curr)))
902 * Once selected, run a task until it either becomes non-eligible or
903 * until it gets a new slice. See the HACK in set_next_entity().
905 if (sched_feat(RUN_TO_PARITY) && curr && curr->vlag == curr->deadline)
908 /* Pick the leftmost entity if it's eligible */
909 if (se && entity_eligible(cfs_rq, se)) {
914 /* Heap search for the EEVD entity */
916 struct rb_node *left = node->rb_left;
919 * Eligible entities in left subtree are always better
920 * choices, since they have earlier deadlines.
922 if (left && vruntime_eligible(cfs_rq,
923 __node_2_se(left)->min_vruntime)) {
928 se = __node_2_se(node);
931 * The left subtree either is empty or has no eligible
932 * entity, so check the current node since it is the one
933 * with earliest deadline that might be eligible.
935 if (entity_eligible(cfs_rq, se)) {
940 node = node->rb_right;
943 if (!best || (curr && entity_before(curr, best)))
949 #ifdef CONFIG_SCHED_DEBUG
950 struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
952 struct rb_node *last = rb_last(&cfs_rq->tasks_timeline.rb_root);
957 return __node_2_se(last);
960 /**************************************************************
961 * Scheduling class statistics methods:
964 int sched_update_scaling(void)
966 unsigned int factor = get_update_sysctl_factor();
968 #define WRT_SYSCTL(name) \
969 (normalized_sysctl_##name = sysctl_##name / (factor))
970 WRT_SYSCTL(sched_base_slice);
978 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se);
981 * XXX: strictly: vd_i += N*r_i/w_i such that: vd_i > ve_i
982 * this is probably good enough.
984 static void update_deadline(struct cfs_rq *cfs_rq, struct sched_entity *se)
986 if ((s64)(se->vruntime - se->deadline) < 0)
990 * For EEVDF the virtual time slope is determined by w_i (iow.
991 * nice) while the request time r_i is determined by
992 * sysctl_sched_base_slice.
994 se->slice = sysctl_sched_base_slice;
997 * EEVDF: vd_i = ve_i + r_i / w_i
999 se->deadline = se->vruntime + calc_delta_fair(se->slice, se);
1002 * The task has consumed its request, reschedule.
1004 if (cfs_rq->nr_running > 1) {
1005 resched_curr(rq_of(cfs_rq));
1006 clear_buddies(cfs_rq, se);
1013 static int select_idle_sibling(struct task_struct *p, int prev_cpu, int cpu);
1014 static unsigned long task_h_load(struct task_struct *p);
1015 static unsigned long capacity_of(int cpu);
1017 /* Give new sched_entity start runnable values to heavy its load in infant time */
1018 void init_entity_runnable_average(struct sched_entity *se)
1020 struct sched_avg *sa = &se->avg;
1022 memset(sa, 0, sizeof(*sa));
1025 * Tasks are initialized with full load to be seen as heavy tasks until
1026 * they get a chance to stabilize to their real load level.
1027 * Group entities are initialized with zero load to reflect the fact that
1028 * nothing has been attached to the task group yet.
1030 if (entity_is_task(se))
1031 sa->load_avg = scale_load_down(se->load.weight);
1033 /* when this task enqueue'ed, it will contribute to its cfs_rq's load_avg */
1037 * With new tasks being created, their initial util_avgs are extrapolated
1038 * based on the cfs_rq's current util_avg:
1040 * util_avg = cfs_rq->util_avg / (cfs_rq->load_avg + 1) * se.load.weight
1042 * However, in many cases, the above util_avg does not give a desired
1043 * value. Moreover, the sum of the util_avgs may be divergent, such
1044 * as when the series is a harmonic series.
1046 * To solve this problem, we also cap the util_avg of successive tasks to
1047 * only 1/2 of the left utilization budget:
1049 * util_avg_cap = (cpu_scale - cfs_rq->avg.util_avg) / 2^n
1051 * where n denotes the nth task and cpu_scale the CPU capacity.
1053 * For example, for a CPU with 1024 of capacity, a simplest series from
1054 * the beginning would be like:
1056 * task util_avg: 512, 256, 128, 64, 32, 16, 8, ...
1057 * cfs_rq util_avg: 512, 768, 896, 960, 992, 1008, 1016, ...
1059 * Finally, that extrapolated util_avg is clamped to the cap (util_avg_cap)
1060 * if util_avg > util_avg_cap.
1062 void post_init_entity_util_avg(struct task_struct *p)
1064 struct sched_entity *se = &p->se;
1065 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1066 struct sched_avg *sa = &se->avg;
1067 long cpu_scale = arch_scale_cpu_capacity(cpu_of(rq_of(cfs_rq)));
1068 long cap = (long)(cpu_scale - cfs_rq->avg.util_avg) / 2;
1070 if (p->sched_class != &fair_sched_class) {
1072 * For !fair tasks do:
1074 update_cfs_rq_load_avg(now, cfs_rq);
1075 attach_entity_load_avg(cfs_rq, se);
1076 switched_from_fair(rq, p);
1078 * such that the next switched_to_fair() has the
1081 se->avg.last_update_time = cfs_rq_clock_pelt(cfs_rq);
1086 if (cfs_rq->avg.util_avg != 0) {
1087 sa->util_avg = cfs_rq->avg.util_avg * se->load.weight;
1088 sa->util_avg /= (cfs_rq->avg.load_avg + 1);
1090 if (sa->util_avg > cap)
1097 sa->runnable_avg = sa->util_avg;
1100 #else /* !CONFIG_SMP */
1101 void init_entity_runnable_average(struct sched_entity *se)
1104 void post_init_entity_util_avg(struct task_struct *p)
1107 static void update_tg_load_avg(struct cfs_rq *cfs_rq)
1110 #endif /* CONFIG_SMP */
1112 static s64 update_curr_se(struct rq *rq, struct sched_entity *curr)
1114 u64 now = rq_clock_task(rq);
1117 delta_exec = now - curr->exec_start;
1118 if (unlikely(delta_exec <= 0))
1121 curr->exec_start = now;
1122 curr->sum_exec_runtime += delta_exec;
1124 if (schedstat_enabled()) {
1125 struct sched_statistics *stats;
1127 stats = __schedstats_from_se(curr);
1128 __schedstat_set(stats->exec_max,
1129 max(delta_exec, stats->exec_max));
1135 static inline void update_curr_task(struct task_struct *p, s64 delta_exec)
1137 trace_sched_stat_runtime(p, delta_exec);
1138 account_group_exec_runtime(p, delta_exec);
1139 cgroup_account_cputime(p, delta_exec);
1141 dl_server_update(p->dl_server, delta_exec);
1145 * Used by other classes to account runtime.
1147 s64 update_curr_common(struct rq *rq)
1149 struct task_struct *curr = rq->curr;
1152 delta_exec = update_curr_se(rq, &curr->se);
1153 if (likely(delta_exec > 0))
1154 update_curr_task(curr, delta_exec);
1160 * Update the current task's runtime statistics.
1162 static void update_curr(struct cfs_rq *cfs_rq)
1164 struct sched_entity *curr = cfs_rq->curr;
1167 if (unlikely(!curr))
1170 delta_exec = update_curr_se(rq_of(cfs_rq), curr);
1171 if (unlikely(delta_exec <= 0))
1174 curr->vruntime += calc_delta_fair(delta_exec, curr);
1175 update_deadline(cfs_rq, curr);
1176 update_min_vruntime(cfs_rq);
1178 if (entity_is_task(curr))
1179 update_curr_task(task_of(curr), delta_exec);
1181 account_cfs_rq_runtime(cfs_rq, delta_exec);
1184 static void update_curr_fair(struct rq *rq)
1186 update_curr(cfs_rq_of(&rq->curr->se));
1190 update_stats_wait_start_fair(struct cfs_rq *cfs_rq, struct sched_entity *se)
1192 struct sched_statistics *stats;
1193 struct task_struct *p = NULL;
1195 if (!schedstat_enabled())
1198 stats = __schedstats_from_se(se);
1200 if (entity_is_task(se))
1203 __update_stats_wait_start(rq_of(cfs_rq), p, stats);
1207 update_stats_wait_end_fair(struct cfs_rq *cfs_rq, struct sched_entity *se)
1209 struct sched_statistics *stats;
1210 struct task_struct *p = NULL;
1212 if (!schedstat_enabled())
1215 stats = __schedstats_from_se(se);
1218 * When the sched_schedstat changes from 0 to 1, some sched se
1219 * maybe already in the runqueue, the se->statistics.wait_start
1220 * will be 0.So it will let the delta wrong. We need to avoid this
1223 if (unlikely(!schedstat_val(stats->wait_start)))
1226 if (entity_is_task(se))
1229 __update_stats_wait_end(rq_of(cfs_rq), p, stats);
1233 update_stats_enqueue_sleeper_fair(struct cfs_rq *cfs_rq, struct sched_entity *se)
1235 struct sched_statistics *stats;
1236 struct task_struct *tsk = NULL;
1238 if (!schedstat_enabled())
1241 stats = __schedstats_from_se(se);
1243 if (entity_is_task(se))
1246 __update_stats_enqueue_sleeper(rq_of(cfs_rq), tsk, stats);
1250 * Task is being enqueued - update stats:
1253 update_stats_enqueue_fair(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
1255 if (!schedstat_enabled())
1259 * Are we enqueueing a waiting task? (for current tasks
1260 * a dequeue/enqueue event is a NOP)
1262 if (se != cfs_rq->curr)
1263 update_stats_wait_start_fair(cfs_rq, se);
1265 if (flags & ENQUEUE_WAKEUP)
1266 update_stats_enqueue_sleeper_fair(cfs_rq, se);
1270 update_stats_dequeue_fair(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
1273 if (!schedstat_enabled())
1277 * Mark the end of the wait period if dequeueing a
1280 if (se != cfs_rq->curr)
1281 update_stats_wait_end_fair(cfs_rq, se);
1283 if ((flags & DEQUEUE_SLEEP) && entity_is_task(se)) {
1284 struct task_struct *tsk = task_of(se);
1287 /* XXX racy against TTWU */
1288 state = READ_ONCE(tsk->__state);
1289 if (state & TASK_INTERRUPTIBLE)
1290 __schedstat_set(tsk->stats.sleep_start,
1291 rq_clock(rq_of(cfs_rq)));
1292 if (state & TASK_UNINTERRUPTIBLE)
1293 __schedstat_set(tsk->stats.block_start,
1294 rq_clock(rq_of(cfs_rq)));
1299 * We are picking a new current task - update its stats:
1302 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
1305 * We are starting a new run period:
1307 se->exec_start = rq_clock_task(rq_of(cfs_rq));
1310 /**************************************************
1311 * Scheduling class queueing methods:
1314 static inline bool is_core_idle(int cpu)
1316 #ifdef CONFIG_SCHED_SMT
1319 for_each_cpu(sibling, cpu_smt_mask(cpu)) {
1323 if (!idle_cpu(sibling))
1332 #define NUMA_IMBALANCE_MIN 2
1335 adjust_numa_imbalance(int imbalance, int dst_running, int imb_numa_nr)
1338 * Allow a NUMA imbalance if busy CPUs is less than the maximum
1339 * threshold. Above this threshold, individual tasks may be contending
1340 * for both memory bandwidth and any shared HT resources. This is an
1341 * approximation as the number of running tasks may not be related to
1342 * the number of busy CPUs due to sched_setaffinity.
1344 if (dst_running > imb_numa_nr)
1348 * Allow a small imbalance based on a simple pair of communicating
1349 * tasks that remain local when the destination is lightly loaded.
1351 if (imbalance <= NUMA_IMBALANCE_MIN)
1356 #endif /* CONFIG_NUMA */
1358 #ifdef CONFIG_NUMA_BALANCING
1360 * Approximate time to scan a full NUMA task in ms. The task scan period is
1361 * calculated based on the tasks virtual memory size and
1362 * numa_balancing_scan_size.
1364 unsigned int sysctl_numa_balancing_scan_period_min = 1000;
1365 unsigned int sysctl_numa_balancing_scan_period_max = 60000;
1367 /* Portion of address space to scan in MB */
1368 unsigned int sysctl_numa_balancing_scan_size = 256;
1370 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
1371 unsigned int sysctl_numa_balancing_scan_delay = 1000;
1373 /* The page with hint page fault latency < threshold in ms is considered hot */
1374 unsigned int sysctl_numa_balancing_hot_threshold = MSEC_PER_SEC;
1377 refcount_t refcount;
1379 spinlock_t lock; /* nr_tasks, tasks */
1384 struct rcu_head rcu;
1385 unsigned long total_faults;
1386 unsigned long max_faults_cpu;
1388 * faults[] array is split into two regions: faults_mem and faults_cpu.
1390 * Faults_cpu is used to decide whether memory should move
1391 * towards the CPU. As a consequence, these stats are weighted
1392 * more by CPU use than by memory faults.
1394 unsigned long faults[];
1398 * For functions that can be called in multiple contexts that permit reading
1399 * ->numa_group (see struct task_struct for locking rules).
1401 static struct numa_group *deref_task_numa_group(struct task_struct *p)
1403 return rcu_dereference_check(p->numa_group, p == current ||
1404 (lockdep_is_held(__rq_lockp(task_rq(p))) && !READ_ONCE(p->on_cpu)));
1407 static struct numa_group *deref_curr_numa_group(struct task_struct *p)
1409 return rcu_dereference_protected(p->numa_group, p == current);
1412 static inline unsigned long group_faults_priv(struct numa_group *ng);
1413 static inline unsigned long group_faults_shared(struct numa_group *ng);
1415 static unsigned int task_nr_scan_windows(struct task_struct *p)
1417 unsigned long rss = 0;
1418 unsigned long nr_scan_pages;
1421 * Calculations based on RSS as non-present and empty pages are skipped
1422 * by the PTE scanner and NUMA hinting faults should be trapped based
1425 nr_scan_pages = sysctl_numa_balancing_scan_size << (20 - PAGE_SHIFT);
1426 rss = get_mm_rss(p->mm);
1428 rss = nr_scan_pages;
1430 rss = round_up(rss, nr_scan_pages);
1431 return rss / nr_scan_pages;
1434 /* For sanity's sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */
1435 #define MAX_SCAN_WINDOW 2560
1437 static unsigned int task_scan_min(struct task_struct *p)
1439 unsigned int scan_size = READ_ONCE(sysctl_numa_balancing_scan_size);
1440 unsigned int scan, floor;
1441 unsigned int windows = 1;
1443 if (scan_size < MAX_SCAN_WINDOW)
1444 windows = MAX_SCAN_WINDOW / scan_size;
1445 floor = 1000 / windows;
1447 scan = sysctl_numa_balancing_scan_period_min / task_nr_scan_windows(p);
1448 return max_t(unsigned int, floor, scan);
1451 static unsigned int task_scan_start(struct task_struct *p)
1453 unsigned long smin = task_scan_min(p);
1454 unsigned long period = smin;
1455 struct numa_group *ng;
1457 /* Scale the maximum scan period with the amount of shared memory. */
1459 ng = rcu_dereference(p->numa_group);
1461 unsigned long shared = group_faults_shared(ng);
1462 unsigned long private = group_faults_priv(ng);
1464 period *= refcount_read(&ng->refcount);
1465 period *= shared + 1;
1466 period /= private + shared + 1;
1470 return max(smin, period);
1473 static unsigned int task_scan_max(struct task_struct *p)
1475 unsigned long smin = task_scan_min(p);
1477 struct numa_group *ng;
1479 /* Watch for min being lower than max due to floor calculations */
1480 smax = sysctl_numa_balancing_scan_period_max / task_nr_scan_windows(p);
1482 /* Scale the maximum scan period with the amount of shared memory. */
1483 ng = deref_curr_numa_group(p);
1485 unsigned long shared = group_faults_shared(ng);
1486 unsigned long private = group_faults_priv(ng);
1487 unsigned long period = smax;
1489 period *= refcount_read(&ng->refcount);
1490 period *= shared + 1;
1491 period /= private + shared + 1;
1493 smax = max(smax, period);
1496 return max(smin, smax);
1499 static void account_numa_enqueue(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 static void account_numa_dequeue(struct rq *rq, struct task_struct *p)
1507 rq->nr_numa_running -= (p->numa_preferred_nid != NUMA_NO_NODE);
1508 rq->nr_preferred_running -= (p->numa_preferred_nid == task_node(p));
1511 /* Shared or private faults. */
1512 #define NR_NUMA_HINT_FAULT_TYPES 2
1514 /* Memory and CPU locality */
1515 #define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2)
1517 /* Averaged statistics, and temporary buffers. */
1518 #define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2)
1520 pid_t task_numa_group_id(struct task_struct *p)
1522 struct numa_group *ng;
1526 ng = rcu_dereference(p->numa_group);
1535 * The averaged statistics, shared & private, memory & CPU,
1536 * occupy the first half of the array. The second half of the
1537 * array is for current counters, which are averaged into the
1538 * first set by task_numa_placement.
1540 static inline int task_faults_idx(enum numa_faults_stats s, int nid, int priv)
1542 return NR_NUMA_HINT_FAULT_TYPES * (s * nr_node_ids + nid) + priv;
1545 static inline unsigned long task_faults(struct task_struct *p, int nid)
1547 if (!p->numa_faults)
1550 return p->numa_faults[task_faults_idx(NUMA_MEM, nid, 0)] +
1551 p->numa_faults[task_faults_idx(NUMA_MEM, nid, 1)];
1554 static inline unsigned long group_faults(struct task_struct *p, int nid)
1556 struct numa_group *ng = deref_task_numa_group(p);
1561 return ng->faults[task_faults_idx(NUMA_MEM, nid, 0)] +
1562 ng->faults[task_faults_idx(NUMA_MEM, nid, 1)];
1565 static inline unsigned long group_faults_cpu(struct numa_group *group, int nid)
1567 return group->faults[task_faults_idx(NUMA_CPU, nid, 0)] +
1568 group->faults[task_faults_idx(NUMA_CPU, nid, 1)];
1571 static inline unsigned long group_faults_priv(struct numa_group *ng)
1573 unsigned long faults = 0;
1576 for_each_online_node(node) {
1577 faults += ng->faults[task_faults_idx(NUMA_MEM, node, 1)];
1583 static inline unsigned long group_faults_shared(struct numa_group *ng)
1585 unsigned long faults = 0;
1588 for_each_online_node(node) {
1589 faults += ng->faults[task_faults_idx(NUMA_MEM, node, 0)];
1596 * A node triggering more than 1/3 as many NUMA faults as the maximum is
1597 * considered part of a numa group's pseudo-interleaving set. Migrations
1598 * between these nodes are slowed down, to allow things to settle down.
1600 #define ACTIVE_NODE_FRACTION 3
1602 static bool numa_is_active_node(int nid, struct numa_group *ng)
1604 return group_faults_cpu(ng, nid) * ACTIVE_NODE_FRACTION > ng->max_faults_cpu;
1607 /* Handle placement on systems where not all nodes are directly connected. */
1608 static unsigned long score_nearby_nodes(struct task_struct *p, int nid,
1609 int lim_dist, bool task)
1611 unsigned long score = 0;
1615 * All nodes are directly connected, and the same distance
1616 * from each other. No need for fancy placement algorithms.
1618 if (sched_numa_topology_type == NUMA_DIRECT)
1621 /* sched_max_numa_distance may be changed in parallel. */
1622 max_dist = READ_ONCE(sched_max_numa_distance);
1624 * This code is called for each node, introducing N^2 complexity,
1625 * which should be ok given the number of nodes rarely exceeds 8.
1627 for_each_online_node(node) {
1628 unsigned long faults;
1629 int dist = node_distance(nid, node);
1632 * The furthest away nodes in the system are not interesting
1633 * for placement; nid was already counted.
1635 if (dist >= max_dist || node == nid)
1639 * On systems with a backplane NUMA topology, compare groups
1640 * of nodes, and move tasks towards the group with the most
1641 * memory accesses. When comparing two nodes at distance
1642 * "hoplimit", only nodes closer by than "hoplimit" are part
1643 * of each group. Skip other nodes.
1645 if (sched_numa_topology_type == NUMA_BACKPLANE && dist >= lim_dist)
1648 /* Add up the faults from nearby nodes. */
1650 faults = task_faults(p, node);
1652 faults = group_faults(p, node);
1655 * On systems with a glueless mesh NUMA topology, there are
1656 * no fixed "groups of nodes". Instead, nodes that are not
1657 * directly connected bounce traffic through intermediate
1658 * nodes; a numa_group can occupy any set of nodes.
1659 * The further away a node is, the less the faults count.
1660 * This seems to result in good task placement.
1662 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
1663 faults *= (max_dist - dist);
1664 faults /= (max_dist - LOCAL_DISTANCE);
1674 * These return the fraction of accesses done by a particular task, or
1675 * task group, on a particular numa node. The group weight is given a
1676 * larger multiplier, in order to group tasks together that are almost
1677 * evenly spread out between numa nodes.
1679 static inline unsigned long task_weight(struct task_struct *p, int nid,
1682 unsigned long faults, total_faults;
1684 if (!p->numa_faults)
1687 total_faults = p->total_numa_faults;
1692 faults = task_faults(p, nid);
1693 faults += score_nearby_nodes(p, nid, dist, true);
1695 return 1000 * faults / total_faults;
1698 static inline unsigned long group_weight(struct task_struct *p, int nid,
1701 struct numa_group *ng = deref_task_numa_group(p);
1702 unsigned long faults, total_faults;
1707 total_faults = ng->total_faults;
1712 faults = group_faults(p, nid);
1713 faults += score_nearby_nodes(p, nid, dist, false);
1715 return 1000 * faults / total_faults;
1719 * If memory tiering mode is enabled, cpupid of slow memory page is
1720 * used to record scan time instead of CPU and PID. When tiering mode
1721 * is disabled at run time, the scan time (in cpupid) will be
1722 * interpreted as CPU and PID. So CPU needs to be checked to avoid to
1723 * access out of array bound.
1725 static inline bool cpupid_valid(int cpupid)
1727 return cpupid_to_cpu(cpupid) < nr_cpu_ids;
1731 * For memory tiering mode, if there are enough free pages (more than
1732 * enough watermark defined here) in fast memory node, to take full
1733 * advantage of fast memory capacity, all recently accessed slow
1734 * memory pages will be migrated to fast memory node without
1735 * considering hot threshold.
1737 static bool pgdat_free_space_enough(struct pglist_data *pgdat)
1740 unsigned long enough_wmark;
1742 enough_wmark = max(1UL * 1024 * 1024 * 1024 >> PAGE_SHIFT,
1743 pgdat->node_present_pages >> 4);
1744 for (z = pgdat->nr_zones - 1; z >= 0; z--) {
1745 struct zone *zone = pgdat->node_zones + z;
1747 if (!populated_zone(zone))
1750 if (zone_watermark_ok(zone, 0,
1751 wmark_pages(zone, WMARK_PROMO) + enough_wmark,
1759 * For memory tiering mode, when page tables are scanned, the scan
1760 * time will be recorded in struct page in addition to make page
1761 * PROT_NONE for slow memory page. So when the page is accessed, in
1762 * hint page fault handler, the hint page fault latency is calculated
1765 * hint page fault latency = hint page fault time - scan time
1767 * The smaller the hint page fault latency, the higher the possibility
1768 * for the page to be hot.
1770 static int numa_hint_fault_latency(struct folio *folio)
1772 int last_time, time;
1774 time = jiffies_to_msecs(jiffies);
1775 last_time = folio_xchg_access_time(folio, time);
1777 return (time - last_time) & PAGE_ACCESS_TIME_MASK;
1781 * For memory tiering mode, too high promotion/demotion throughput may
1782 * hurt application latency. So we provide a mechanism to rate limit
1783 * the number of pages that are tried to be promoted.
1785 static bool numa_promotion_rate_limit(struct pglist_data *pgdat,
1786 unsigned long rate_limit, int nr)
1788 unsigned long nr_cand;
1789 unsigned int now, start;
1791 now = jiffies_to_msecs(jiffies);
1792 mod_node_page_state(pgdat, PGPROMOTE_CANDIDATE, nr);
1793 nr_cand = node_page_state(pgdat, PGPROMOTE_CANDIDATE);
1794 start = pgdat->nbp_rl_start;
1795 if (now - start > MSEC_PER_SEC &&
1796 cmpxchg(&pgdat->nbp_rl_start, start, now) == start)
1797 pgdat->nbp_rl_nr_cand = nr_cand;
1798 if (nr_cand - pgdat->nbp_rl_nr_cand >= rate_limit)
1803 #define NUMA_MIGRATION_ADJUST_STEPS 16
1805 static void numa_promotion_adjust_threshold(struct pglist_data *pgdat,
1806 unsigned long rate_limit,
1807 unsigned int ref_th)
1809 unsigned int now, start, th_period, unit_th, th;
1810 unsigned long nr_cand, ref_cand, diff_cand;
1812 now = jiffies_to_msecs(jiffies);
1813 th_period = sysctl_numa_balancing_scan_period_max;
1814 start = pgdat->nbp_th_start;
1815 if (now - start > th_period &&
1816 cmpxchg(&pgdat->nbp_th_start, start, now) == start) {
1817 ref_cand = rate_limit *
1818 sysctl_numa_balancing_scan_period_max / MSEC_PER_SEC;
1819 nr_cand = node_page_state(pgdat, PGPROMOTE_CANDIDATE);
1820 diff_cand = nr_cand - pgdat->nbp_th_nr_cand;
1821 unit_th = ref_th * 2 / NUMA_MIGRATION_ADJUST_STEPS;
1822 th = pgdat->nbp_threshold ? : ref_th;
1823 if (diff_cand > ref_cand * 11 / 10)
1824 th = max(th - unit_th, unit_th);
1825 else if (diff_cand < ref_cand * 9 / 10)
1826 th = min(th + unit_th, ref_th * 2);
1827 pgdat->nbp_th_nr_cand = nr_cand;
1828 pgdat->nbp_threshold = th;
1832 bool should_numa_migrate_memory(struct task_struct *p, struct folio *folio,
1833 int src_nid, int dst_cpu)
1835 struct numa_group *ng = deref_curr_numa_group(p);
1836 int dst_nid = cpu_to_node(dst_cpu);
1837 int last_cpupid, this_cpupid;
1840 * Cannot migrate to memoryless nodes.
1842 if (!node_state(dst_nid, N_MEMORY))
1846 * The pages in slow memory node should be migrated according
1847 * to hot/cold instead of private/shared.
1849 if (sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING &&
1850 !node_is_toptier(src_nid)) {
1851 struct pglist_data *pgdat;
1852 unsigned long rate_limit;
1853 unsigned int latency, th, def_th;
1855 pgdat = NODE_DATA(dst_nid);
1856 if (pgdat_free_space_enough(pgdat)) {
1857 /* workload changed, reset hot threshold */
1858 pgdat->nbp_threshold = 0;
1862 def_th = sysctl_numa_balancing_hot_threshold;
1863 rate_limit = sysctl_numa_balancing_promote_rate_limit << \
1865 numa_promotion_adjust_threshold(pgdat, rate_limit, def_th);
1867 th = pgdat->nbp_threshold ? : def_th;
1868 latency = numa_hint_fault_latency(folio);
1872 return !numa_promotion_rate_limit(pgdat, rate_limit,
1873 folio_nr_pages(folio));
1876 this_cpupid = cpu_pid_to_cpupid(dst_cpu, current->pid);
1877 last_cpupid = folio_xchg_last_cpupid(folio, this_cpupid);
1879 if (!(sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING) &&
1880 !node_is_toptier(src_nid) && !cpupid_valid(last_cpupid))
1884 * Allow first faults or private faults to migrate immediately early in
1885 * the lifetime of a task. The magic number 4 is based on waiting for
1886 * two full passes of the "multi-stage node selection" test that is
1889 if ((p->numa_preferred_nid == NUMA_NO_NODE || p->numa_scan_seq <= 4) &&
1890 (cpupid_pid_unset(last_cpupid) || cpupid_match_pid(p, last_cpupid)))
1894 * Multi-stage node selection is used in conjunction with a periodic
1895 * migration fault to build a temporal task<->page relation. By using
1896 * a two-stage filter we remove short/unlikely relations.
1898 * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate
1899 * a task's usage of a particular page (n_p) per total usage of this
1900 * page (n_t) (in a given time-span) to a probability.
1902 * Our periodic faults will sample this probability and getting the
1903 * same result twice in a row, given these samples are fully
1904 * independent, is then given by P(n)^2, provided our sample period
1905 * is sufficiently short compared to the usage pattern.
1907 * This quadric squishes small probabilities, making it less likely we
1908 * act on an unlikely task<->page relation.
1910 if (!cpupid_pid_unset(last_cpupid) &&
1911 cpupid_to_nid(last_cpupid) != dst_nid)
1914 /* Always allow migrate on private faults */
1915 if (cpupid_match_pid(p, last_cpupid))
1918 /* A shared fault, but p->numa_group has not been set up yet. */
1923 * Destination node is much more heavily used than the source
1924 * node? Allow migration.
1926 if (group_faults_cpu(ng, dst_nid) > group_faults_cpu(ng, src_nid) *
1927 ACTIVE_NODE_FRACTION)
1931 * Distribute memory according to CPU & memory use on each node,
1932 * with 3/4 hysteresis to avoid unnecessary memory migrations:
1934 * faults_cpu(dst) 3 faults_cpu(src)
1935 * --------------- * - > ---------------
1936 * faults_mem(dst) 4 faults_mem(src)
1938 return group_faults_cpu(ng, dst_nid) * group_faults(p, src_nid) * 3 >
1939 group_faults_cpu(ng, src_nid) * group_faults(p, dst_nid) * 4;
1943 * 'numa_type' describes the node at the moment of load balancing.
1946 /* The node has spare capacity that can be used to run more tasks. */
1949 * The node is fully used and the tasks don't compete for more CPU
1950 * cycles. Nevertheless, some tasks might wait before running.
1954 * The node is overloaded and can't provide expected CPU cycles to all
1960 /* Cached statistics for all CPUs within a node */
1963 unsigned long runnable;
1965 /* Total compute capacity of CPUs on a node */
1966 unsigned long compute_capacity;
1967 unsigned int nr_running;
1968 unsigned int weight;
1969 enum numa_type node_type;
1973 struct task_numa_env {
1974 struct task_struct *p;
1976 int src_cpu, src_nid;
1977 int dst_cpu, dst_nid;
1980 struct numa_stats src_stats, dst_stats;
1985 struct task_struct *best_task;
1990 static unsigned long cpu_load(struct rq *rq);
1991 static unsigned long cpu_runnable(struct rq *rq);
1994 numa_type numa_classify(unsigned int imbalance_pct,
1995 struct numa_stats *ns)
1997 if ((ns->nr_running > ns->weight) &&
1998 (((ns->compute_capacity * 100) < (ns->util * imbalance_pct)) ||
1999 ((ns->compute_capacity * imbalance_pct) < (ns->runnable * 100))))
2000 return node_overloaded;
2002 if ((ns->nr_running < ns->weight) ||
2003 (((ns->compute_capacity * 100) > (ns->util * imbalance_pct)) &&
2004 ((ns->compute_capacity * imbalance_pct) > (ns->runnable * 100))))
2005 return node_has_spare;
2007 return node_fully_busy;
2010 #ifdef CONFIG_SCHED_SMT
2011 /* Forward declarations of select_idle_sibling helpers */
2012 static inline bool test_idle_cores(int cpu);
2013 static inline int numa_idle_core(int idle_core, int cpu)
2015 if (!static_branch_likely(&sched_smt_present) ||
2016 idle_core >= 0 || !test_idle_cores(cpu))
2020 * Prefer cores instead of packing HT siblings
2021 * and triggering future load balancing.
2023 if (is_core_idle(cpu))
2029 static inline int numa_idle_core(int idle_core, int cpu)
2036 * Gather all necessary information to make NUMA balancing placement
2037 * decisions that are compatible with standard load balancer. This
2038 * borrows code and logic from update_sg_lb_stats but sharing a
2039 * common implementation is impractical.
2041 static void update_numa_stats(struct task_numa_env *env,
2042 struct numa_stats *ns, int nid,
2045 int cpu, idle_core = -1;
2047 memset(ns, 0, sizeof(*ns));
2051 for_each_cpu(cpu, cpumask_of_node(nid)) {
2052 struct rq *rq = cpu_rq(cpu);
2054 ns->load += cpu_load(rq);
2055 ns->runnable += cpu_runnable(rq);
2056 ns->util += cpu_util_cfs(cpu);
2057 ns->nr_running += rq->cfs.h_nr_running;
2058 ns->compute_capacity += capacity_of(cpu);
2060 if (find_idle && idle_core < 0 && !rq->nr_running && idle_cpu(cpu)) {
2061 if (READ_ONCE(rq->numa_migrate_on) ||
2062 !cpumask_test_cpu(cpu, env->p->cpus_ptr))
2065 if (ns->idle_cpu == -1)
2068 idle_core = numa_idle_core(idle_core, cpu);
2073 ns->weight = cpumask_weight(cpumask_of_node(nid));
2075 ns->node_type = numa_classify(env->imbalance_pct, ns);
2078 ns->idle_cpu = idle_core;
2081 static void task_numa_assign(struct task_numa_env *env,
2082 struct task_struct *p, long imp)
2084 struct rq *rq = cpu_rq(env->dst_cpu);
2086 /* Check if run-queue part of active NUMA balance. */
2087 if (env->best_cpu != env->dst_cpu && xchg(&rq->numa_migrate_on, 1)) {
2089 int start = env->dst_cpu;
2091 /* Find alternative idle CPU. */
2092 for_each_cpu_wrap(cpu, cpumask_of_node(env->dst_nid), start + 1) {
2093 if (cpu == env->best_cpu || !idle_cpu(cpu) ||
2094 !cpumask_test_cpu(cpu, env->p->cpus_ptr)) {
2099 rq = cpu_rq(env->dst_cpu);
2100 if (!xchg(&rq->numa_migrate_on, 1))
2104 /* Failed to find an alternative idle CPU */
2110 * Clear previous best_cpu/rq numa-migrate flag, since task now
2111 * found a better CPU to move/swap.
2113 if (env->best_cpu != -1 && env->best_cpu != env->dst_cpu) {
2114 rq = cpu_rq(env->best_cpu);
2115 WRITE_ONCE(rq->numa_migrate_on, 0);
2119 put_task_struct(env->best_task);
2124 env->best_imp = imp;
2125 env->best_cpu = env->dst_cpu;
2128 static bool load_too_imbalanced(long src_load, long dst_load,
2129 struct task_numa_env *env)
2132 long orig_src_load, orig_dst_load;
2133 long src_capacity, dst_capacity;
2136 * The load is corrected for the CPU capacity available on each node.
2139 * ------------ vs ---------
2140 * src_capacity dst_capacity
2142 src_capacity = env->src_stats.compute_capacity;
2143 dst_capacity = env->dst_stats.compute_capacity;
2145 imb = abs(dst_load * src_capacity - src_load * dst_capacity);
2147 orig_src_load = env->src_stats.load;
2148 orig_dst_load = env->dst_stats.load;
2150 old_imb = abs(orig_dst_load * src_capacity - orig_src_load * dst_capacity);
2152 /* Would this change make things worse? */
2153 return (imb > old_imb);
2157 * Maximum NUMA importance can be 1998 (2*999);
2158 * SMALLIMP @ 30 would be close to 1998/64.
2159 * Used to deter task migration.
2164 * This checks if the overall compute and NUMA accesses of the system would
2165 * be improved if the source tasks was migrated to the target dst_cpu taking
2166 * into account that it might be best if task running on the dst_cpu should
2167 * be exchanged with the source task
2169 static bool task_numa_compare(struct task_numa_env *env,
2170 long taskimp, long groupimp, bool maymove)
2172 struct numa_group *cur_ng, *p_ng = deref_curr_numa_group(env->p);
2173 struct rq *dst_rq = cpu_rq(env->dst_cpu);
2174 long imp = p_ng ? groupimp : taskimp;
2175 struct task_struct *cur;
2176 long src_load, dst_load;
2177 int dist = env->dist;
2180 bool stopsearch = false;
2182 if (READ_ONCE(dst_rq->numa_migrate_on))
2186 cur = rcu_dereference(dst_rq->curr);
2187 if (cur && ((cur->flags & PF_EXITING) || is_idle_task(cur)))
2191 * Because we have preemption enabled we can get migrated around and
2192 * end try selecting ourselves (current == env->p) as a swap candidate.
2194 if (cur == env->p) {
2200 if (maymove && moveimp >= env->best_imp)
2206 /* Skip this swap candidate if cannot move to the source cpu. */
2207 if (!cpumask_test_cpu(env->src_cpu, cur->cpus_ptr))
2211 * Skip this swap candidate if it is not moving to its preferred
2212 * node and the best task is.
2214 if (env->best_task &&
2215 env->best_task->numa_preferred_nid == env->src_nid &&
2216 cur->numa_preferred_nid != env->src_nid) {
2221 * "imp" is the fault differential for the source task between the
2222 * source and destination node. Calculate the total differential for
2223 * the source task and potential destination task. The more negative
2224 * the value is, the more remote accesses that would be expected to
2225 * be incurred if the tasks were swapped.
2227 * If dst and source tasks are in the same NUMA group, or not
2228 * in any group then look only at task weights.
2230 cur_ng = rcu_dereference(cur->numa_group);
2231 if (cur_ng == p_ng) {
2233 * Do not swap within a group or between tasks that have
2234 * no group if there is spare capacity. Swapping does
2235 * not address the load imbalance and helps one task at
2236 * the cost of punishing another.
2238 if (env->dst_stats.node_type == node_has_spare)
2241 imp = taskimp + task_weight(cur, env->src_nid, dist) -
2242 task_weight(cur, env->dst_nid, dist);
2244 * Add some hysteresis to prevent swapping the
2245 * tasks within a group over tiny differences.
2251 * Compare the group weights. If a task is all by itself
2252 * (not part of a group), use the task weight instead.
2255 imp += group_weight(cur, env->src_nid, dist) -
2256 group_weight(cur, env->dst_nid, dist);
2258 imp += task_weight(cur, env->src_nid, dist) -
2259 task_weight(cur, env->dst_nid, dist);
2262 /* Discourage picking a task already on its preferred node */
2263 if (cur->numa_preferred_nid == env->dst_nid)
2267 * Encourage picking a task that moves to its preferred node.
2268 * This potentially makes imp larger than it's maximum of
2269 * 1998 (see SMALLIMP and task_weight for why) but in this
2270 * case, it does not matter.
2272 if (cur->numa_preferred_nid == env->src_nid)
2275 if (maymove && moveimp > imp && moveimp > env->best_imp) {
2282 * Prefer swapping with a task moving to its preferred node over a
2285 if (env->best_task && cur->numa_preferred_nid == env->src_nid &&
2286 env->best_task->numa_preferred_nid != env->src_nid) {
2291 * If the NUMA importance is less than SMALLIMP,
2292 * task migration might only result in ping pong
2293 * of tasks and also hurt performance due to cache
2296 if (imp < SMALLIMP || imp <= env->best_imp + SMALLIMP / 2)
2300 * In the overloaded case, try and keep the load balanced.
2302 load = task_h_load(env->p) - task_h_load(cur);
2306 dst_load = env->dst_stats.load + load;
2307 src_load = env->src_stats.load - load;
2309 if (load_too_imbalanced(src_load, dst_load, env))
2313 /* Evaluate an idle CPU for a task numa move. */
2315 int cpu = env->dst_stats.idle_cpu;
2317 /* Nothing cached so current CPU went idle since the search. */
2322 * If the CPU is no longer truly idle and the previous best CPU
2323 * is, keep using it.
2325 if (!idle_cpu(cpu) && env->best_cpu >= 0 &&
2326 idle_cpu(env->best_cpu)) {
2327 cpu = env->best_cpu;
2333 task_numa_assign(env, cur, imp);
2336 * If a move to idle is allowed because there is capacity or load
2337 * balance improves then stop the search. While a better swap
2338 * candidate may exist, a search is not free.
2340 if (maymove && !cur && env->best_cpu >= 0 && idle_cpu(env->best_cpu))
2344 * If a swap candidate must be identified and the current best task
2345 * moves its preferred node then stop the search.
2347 if (!maymove && env->best_task &&
2348 env->best_task->numa_preferred_nid == env->src_nid) {
2357 static void task_numa_find_cpu(struct task_numa_env *env,
2358 long taskimp, long groupimp)
2360 bool maymove = false;
2364 * If dst node has spare capacity, then check if there is an
2365 * imbalance that would be overruled by the load balancer.
2367 if (env->dst_stats.node_type == node_has_spare) {
2368 unsigned int imbalance;
2369 int src_running, dst_running;
2372 * Would movement cause an imbalance? Note that if src has
2373 * more running tasks that the imbalance is ignored as the
2374 * move improves the imbalance from the perspective of the
2375 * CPU load balancer.
2377 src_running = env->src_stats.nr_running - 1;
2378 dst_running = env->dst_stats.nr_running + 1;
2379 imbalance = max(0, dst_running - src_running);
2380 imbalance = adjust_numa_imbalance(imbalance, dst_running,
2383 /* Use idle CPU if there is no imbalance */
2386 if (env->dst_stats.idle_cpu >= 0) {
2387 env->dst_cpu = env->dst_stats.idle_cpu;
2388 task_numa_assign(env, NULL, 0);
2393 long src_load, dst_load, load;
2395 * If the improvement from just moving env->p direction is better
2396 * than swapping tasks around, check if a move is possible.
2398 load = task_h_load(env->p);
2399 dst_load = env->dst_stats.load + load;
2400 src_load = env->src_stats.load - load;
2401 maymove = !load_too_imbalanced(src_load, dst_load, env);
2404 for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) {
2405 /* Skip this CPU if the source task cannot migrate */
2406 if (!cpumask_test_cpu(cpu, env->p->cpus_ptr))
2410 if (task_numa_compare(env, taskimp, groupimp, maymove))
2415 static int task_numa_migrate(struct task_struct *p)
2417 struct task_numa_env env = {
2420 .src_cpu = task_cpu(p),
2421 .src_nid = task_node(p),
2423 .imbalance_pct = 112,
2429 unsigned long taskweight, groupweight;
2430 struct sched_domain *sd;
2431 long taskimp, groupimp;
2432 struct numa_group *ng;
2437 * Pick the lowest SD_NUMA domain, as that would have the smallest
2438 * imbalance and would be the first to start moving tasks about.
2440 * And we want to avoid any moving of tasks about, as that would create
2441 * random movement of tasks -- counter the numa conditions we're trying
2445 sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu));
2447 env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2;
2448 env.imb_numa_nr = sd->imb_numa_nr;
2453 * Cpusets can break the scheduler domain tree into smaller
2454 * balance domains, some of which do not cross NUMA boundaries.
2455 * Tasks that are "trapped" in such domains cannot be migrated
2456 * elsewhere, so there is no point in (re)trying.
2458 if (unlikely(!sd)) {
2459 sched_setnuma(p, task_node(p));
2463 env.dst_nid = p->numa_preferred_nid;
2464 dist = env.dist = node_distance(env.src_nid, env.dst_nid);
2465 taskweight = task_weight(p, env.src_nid, dist);
2466 groupweight = group_weight(p, env.src_nid, dist);
2467 update_numa_stats(&env, &env.src_stats, env.src_nid, false);
2468 taskimp = task_weight(p, env.dst_nid, dist) - taskweight;
2469 groupimp = group_weight(p, env.dst_nid, dist) - groupweight;
2470 update_numa_stats(&env, &env.dst_stats, env.dst_nid, true);
2472 /* Try to find a spot on the preferred nid. */
2473 task_numa_find_cpu(&env, taskimp, groupimp);
2476 * Look at other nodes in these cases:
2477 * - there is no space available on the preferred_nid
2478 * - the task is part of a numa_group that is interleaved across
2479 * multiple NUMA nodes; in order to better consolidate the group,
2480 * we need to check other locations.
2482 ng = deref_curr_numa_group(p);
2483 if (env.best_cpu == -1 || (ng && ng->active_nodes > 1)) {
2484 for_each_node_state(nid, N_CPU) {
2485 if (nid == env.src_nid || nid == p->numa_preferred_nid)
2488 dist = node_distance(env.src_nid, env.dst_nid);
2489 if (sched_numa_topology_type == NUMA_BACKPLANE &&
2491 taskweight = task_weight(p, env.src_nid, dist);
2492 groupweight = group_weight(p, env.src_nid, dist);
2495 /* Only consider nodes where both task and groups benefit */
2496 taskimp = task_weight(p, nid, dist) - taskweight;
2497 groupimp = group_weight(p, nid, dist) - groupweight;
2498 if (taskimp < 0 && groupimp < 0)
2503 update_numa_stats(&env, &env.dst_stats, env.dst_nid, true);
2504 task_numa_find_cpu(&env, taskimp, groupimp);
2509 * If the task is part of a workload that spans multiple NUMA nodes,
2510 * and is migrating into one of the workload's active nodes, remember
2511 * this node as the task's preferred numa node, so the workload can
2513 * A task that migrated to a second choice node will be better off
2514 * trying for a better one later. Do not set the preferred node here.
2517 if (env.best_cpu == -1)
2520 nid = cpu_to_node(env.best_cpu);
2522 if (nid != p->numa_preferred_nid)
2523 sched_setnuma(p, nid);
2526 /* No better CPU than the current one was found. */
2527 if (env.best_cpu == -1) {
2528 trace_sched_stick_numa(p, env.src_cpu, NULL, -1);
2532 best_rq = cpu_rq(env.best_cpu);
2533 if (env.best_task == NULL) {
2534 ret = migrate_task_to(p, env.best_cpu);
2535 WRITE_ONCE(best_rq->numa_migrate_on, 0);
2537 trace_sched_stick_numa(p, env.src_cpu, NULL, env.best_cpu);
2541 ret = migrate_swap(p, env.best_task, env.best_cpu, env.src_cpu);
2542 WRITE_ONCE(best_rq->numa_migrate_on, 0);
2545 trace_sched_stick_numa(p, env.src_cpu, env.best_task, env.best_cpu);
2546 put_task_struct(env.best_task);
2550 /* Attempt to migrate a task to a CPU on the preferred node. */
2551 static void numa_migrate_preferred(struct task_struct *p)
2553 unsigned long interval = HZ;
2555 /* This task has no NUMA fault statistics yet */
2556 if (unlikely(p->numa_preferred_nid == NUMA_NO_NODE || !p->numa_faults))
2559 /* Periodically retry migrating the task to the preferred node */
2560 interval = min(interval, msecs_to_jiffies(p->numa_scan_period) / 16);
2561 p->numa_migrate_retry = jiffies + interval;
2563 /* Success if task is already running on preferred CPU */
2564 if (task_node(p) == p->numa_preferred_nid)
2567 /* Otherwise, try migrate to a CPU on the preferred node */
2568 task_numa_migrate(p);
2572 * Find out how many nodes the workload is actively running on. Do this by
2573 * tracking the nodes from which NUMA hinting faults are triggered. This can
2574 * be different from the set of nodes where the workload's memory is currently
2577 static void numa_group_count_active_nodes(struct numa_group *numa_group)
2579 unsigned long faults, max_faults = 0;
2580 int nid, active_nodes = 0;
2582 for_each_node_state(nid, N_CPU) {
2583 faults = group_faults_cpu(numa_group, nid);
2584 if (faults > max_faults)
2585 max_faults = faults;
2588 for_each_node_state(nid, N_CPU) {
2589 faults = group_faults_cpu(numa_group, nid);
2590 if (faults * ACTIVE_NODE_FRACTION > max_faults)
2594 numa_group->max_faults_cpu = max_faults;
2595 numa_group->active_nodes = active_nodes;
2599 * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
2600 * increments. The more local the fault statistics are, the higher the scan
2601 * period will be for the next scan window. If local/(local+remote) ratio is
2602 * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS)
2603 * the scan period will decrease. Aim for 70% local accesses.
2605 #define NUMA_PERIOD_SLOTS 10
2606 #define NUMA_PERIOD_THRESHOLD 7
2609 * Increase the scan period (slow down scanning) if the majority of
2610 * our memory is already on our local node, or if the majority of
2611 * the page accesses are shared with other processes.
2612 * Otherwise, decrease the scan period.
2614 static void update_task_scan_period(struct task_struct *p,
2615 unsigned long shared, unsigned long private)
2617 unsigned int period_slot;
2618 int lr_ratio, ps_ratio;
2621 unsigned long remote = p->numa_faults_locality[0];
2622 unsigned long local = p->numa_faults_locality[1];
2625 * If there were no record hinting faults then either the task is
2626 * completely idle or all activity is in areas that are not of interest
2627 * to automatic numa balancing. Related to that, if there were failed
2628 * migration then it implies we are migrating too quickly or the local
2629 * node is overloaded. In either case, scan slower
2631 if (local + shared == 0 || p->numa_faults_locality[2]) {
2632 p->numa_scan_period = min(p->numa_scan_period_max,
2633 p->numa_scan_period << 1);
2635 p->mm->numa_next_scan = jiffies +
2636 msecs_to_jiffies(p->numa_scan_period);
2642 * Prepare to scale scan period relative to the current period.
2643 * == NUMA_PERIOD_THRESHOLD scan period stays the same
2644 * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
2645 * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
2647 period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS);
2648 lr_ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote);
2649 ps_ratio = (private * NUMA_PERIOD_SLOTS) / (private + shared);
2651 if (ps_ratio >= NUMA_PERIOD_THRESHOLD) {
2653 * Most memory accesses are local. There is no need to
2654 * do fast NUMA scanning, since memory is already local.
2656 int slot = ps_ratio - NUMA_PERIOD_THRESHOLD;
2659 diff = slot * period_slot;
2660 } else if (lr_ratio >= NUMA_PERIOD_THRESHOLD) {
2662 * Most memory accesses are shared with other tasks.
2663 * There is no point in continuing fast NUMA scanning,
2664 * since other tasks may just move the memory elsewhere.
2666 int slot = lr_ratio - NUMA_PERIOD_THRESHOLD;
2669 diff = slot * period_slot;
2672 * Private memory faults exceed (SLOTS-THRESHOLD)/SLOTS,
2673 * yet they are not on the local NUMA node. Speed up
2674 * NUMA scanning to get the memory moved over.
2676 int ratio = max(lr_ratio, ps_ratio);
2677 diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot;
2680 p->numa_scan_period = clamp(p->numa_scan_period + diff,
2681 task_scan_min(p), task_scan_max(p));
2682 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
2686 * Get the fraction of time the task has been running since the last
2687 * NUMA placement cycle. The scheduler keeps similar statistics, but
2688 * decays those on a 32ms period, which is orders of magnitude off
2689 * from the dozens-of-seconds NUMA balancing period. Use the scheduler
2690 * stats only if the task is so new there are no NUMA statistics yet.
2692 static u64 numa_get_avg_runtime(struct task_struct *p, u64 *period)
2694 u64 runtime, delta, now;
2695 /* Use the start of this time slice to avoid calculations. */
2696 now = p->se.exec_start;
2697 runtime = p->se.sum_exec_runtime;
2699 if (p->last_task_numa_placement) {
2700 delta = runtime - p->last_sum_exec_runtime;
2701 *period = now - p->last_task_numa_placement;
2703 /* Avoid time going backwards, prevent potential divide error: */
2704 if (unlikely((s64)*period < 0))
2707 delta = p->se.avg.load_sum;
2708 *period = LOAD_AVG_MAX;
2711 p->last_sum_exec_runtime = runtime;
2712 p->last_task_numa_placement = now;
2718 * Determine the preferred nid for a task in a numa_group. This needs to
2719 * be done in a way that produces consistent results with group_weight,
2720 * otherwise workloads might not converge.
2722 static int preferred_group_nid(struct task_struct *p, int nid)
2727 /* Direct connections between all NUMA nodes. */
2728 if (sched_numa_topology_type == NUMA_DIRECT)
2732 * On a system with glueless mesh NUMA topology, group_weight
2733 * scores nodes according to the number of NUMA hinting faults on
2734 * both the node itself, and on nearby nodes.
2736 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
2737 unsigned long score, max_score = 0;
2738 int node, max_node = nid;
2740 dist = sched_max_numa_distance;
2742 for_each_node_state(node, N_CPU) {
2743 score = group_weight(p, node, dist);
2744 if (score > max_score) {
2753 * Finding the preferred nid in a system with NUMA backplane
2754 * interconnect topology is more involved. The goal is to locate
2755 * tasks from numa_groups near each other in the system, and
2756 * untangle workloads from different sides of the system. This requires
2757 * searching down the hierarchy of node groups, recursively searching
2758 * inside the highest scoring group of nodes. The nodemask tricks
2759 * keep the complexity of the search down.
2761 nodes = node_states[N_CPU];
2762 for (dist = sched_max_numa_distance; dist > LOCAL_DISTANCE; dist--) {
2763 unsigned long max_faults = 0;
2764 nodemask_t max_group = NODE_MASK_NONE;
2767 /* Are there nodes at this distance from each other? */
2768 if (!find_numa_distance(dist))
2771 for_each_node_mask(a, nodes) {
2772 unsigned long faults = 0;
2773 nodemask_t this_group;
2774 nodes_clear(this_group);
2776 /* Sum group's NUMA faults; includes a==b case. */
2777 for_each_node_mask(b, nodes) {
2778 if (node_distance(a, b) < dist) {
2779 faults += group_faults(p, b);
2780 node_set(b, this_group);
2781 node_clear(b, nodes);
2785 /* Remember the top group. */
2786 if (faults > max_faults) {
2787 max_faults = faults;
2788 max_group = this_group;
2790 * subtle: at the smallest distance there is
2791 * just one node left in each "group", the
2792 * winner is the preferred nid.
2797 /* Next round, evaluate the nodes within max_group. */
2805 static void task_numa_placement(struct task_struct *p)
2807 int seq, nid, max_nid = NUMA_NO_NODE;
2808 unsigned long max_faults = 0;
2809 unsigned long fault_types[2] = { 0, 0 };
2810 unsigned long total_faults;
2811 u64 runtime, period;
2812 spinlock_t *group_lock = NULL;
2813 struct numa_group *ng;
2816 * The p->mm->numa_scan_seq field gets updated without
2817 * exclusive access. Use READ_ONCE() here to ensure
2818 * that the field is read in a single access:
2820 seq = READ_ONCE(p->mm->numa_scan_seq);
2821 if (p->numa_scan_seq == seq)
2823 p->numa_scan_seq = seq;
2824 p->numa_scan_period_max = task_scan_max(p);
2826 total_faults = p->numa_faults_locality[0] +
2827 p->numa_faults_locality[1];
2828 runtime = numa_get_avg_runtime(p, &period);
2830 /* If the task is part of a group prevent parallel updates to group stats */
2831 ng = deref_curr_numa_group(p);
2833 group_lock = &ng->lock;
2834 spin_lock_irq(group_lock);
2837 /* Find the node with the highest number of faults */
2838 for_each_online_node(nid) {
2839 /* Keep track of the offsets in numa_faults array */
2840 int mem_idx, membuf_idx, cpu_idx, cpubuf_idx;
2841 unsigned long faults = 0, group_faults = 0;
2844 for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) {
2845 long diff, f_diff, f_weight;
2847 mem_idx = task_faults_idx(NUMA_MEM, nid, priv);
2848 membuf_idx = task_faults_idx(NUMA_MEMBUF, nid, priv);
2849 cpu_idx = task_faults_idx(NUMA_CPU, nid, priv);
2850 cpubuf_idx = task_faults_idx(NUMA_CPUBUF, nid, priv);
2852 /* Decay existing window, copy faults since last scan */
2853 diff = p->numa_faults[membuf_idx] - p->numa_faults[mem_idx] / 2;
2854 fault_types[priv] += p->numa_faults[membuf_idx];
2855 p->numa_faults[membuf_idx] = 0;
2858 * Normalize the faults_from, so all tasks in a group
2859 * count according to CPU use, instead of by the raw
2860 * number of faults. Tasks with little runtime have
2861 * little over-all impact on throughput, and thus their
2862 * faults are less important.
2864 f_weight = div64_u64(runtime << 16, period + 1);
2865 f_weight = (f_weight * p->numa_faults[cpubuf_idx]) /
2867 f_diff = f_weight - p->numa_faults[cpu_idx] / 2;
2868 p->numa_faults[cpubuf_idx] = 0;
2870 p->numa_faults[mem_idx] += diff;
2871 p->numa_faults[cpu_idx] += f_diff;
2872 faults += p->numa_faults[mem_idx];
2873 p->total_numa_faults += diff;
2876 * safe because we can only change our own group
2878 * mem_idx represents the offset for a given
2879 * nid and priv in a specific region because it
2880 * is at the beginning of the numa_faults array.
2882 ng->faults[mem_idx] += diff;
2883 ng->faults[cpu_idx] += f_diff;
2884 ng->total_faults += diff;
2885 group_faults += ng->faults[mem_idx];
2890 if (faults > max_faults) {
2891 max_faults = faults;
2894 } else if (group_faults > max_faults) {
2895 max_faults = group_faults;
2900 /* Cannot migrate task to CPU-less node */
2901 max_nid = numa_nearest_node(max_nid, N_CPU);
2904 numa_group_count_active_nodes(ng);
2905 spin_unlock_irq(group_lock);
2906 max_nid = preferred_group_nid(p, max_nid);
2910 /* Set the new preferred node */
2911 if (max_nid != p->numa_preferred_nid)
2912 sched_setnuma(p, max_nid);
2915 update_task_scan_period(p, fault_types[0], fault_types[1]);
2918 static inline int get_numa_group(struct numa_group *grp)
2920 return refcount_inc_not_zero(&grp->refcount);
2923 static inline void put_numa_group(struct numa_group *grp)
2925 if (refcount_dec_and_test(&grp->refcount))
2926 kfree_rcu(grp, rcu);
2929 static void task_numa_group(struct task_struct *p, int cpupid, int flags,
2932 struct numa_group *grp, *my_grp;
2933 struct task_struct *tsk;
2935 int cpu = cpupid_to_cpu(cpupid);
2938 if (unlikely(!deref_curr_numa_group(p))) {
2939 unsigned int size = sizeof(struct numa_group) +
2940 NR_NUMA_HINT_FAULT_STATS *
2941 nr_node_ids * sizeof(unsigned long);
2943 grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
2947 refcount_set(&grp->refcount, 1);
2948 grp->active_nodes = 1;
2949 grp->max_faults_cpu = 0;
2950 spin_lock_init(&grp->lock);
2953 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2954 grp->faults[i] = p->numa_faults[i];
2956 grp->total_faults = p->total_numa_faults;
2959 rcu_assign_pointer(p->numa_group, grp);
2963 tsk = READ_ONCE(cpu_rq(cpu)->curr);
2965 if (!cpupid_match_pid(tsk, cpupid))
2968 grp = rcu_dereference(tsk->numa_group);
2972 my_grp = deref_curr_numa_group(p);
2977 * Only join the other group if its bigger; if we're the bigger group,
2978 * the other task will join us.
2980 if (my_grp->nr_tasks > grp->nr_tasks)
2984 * Tie-break on the grp address.
2986 if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp)
2989 /* Always join threads in the same process. */
2990 if (tsk->mm == current->mm)
2993 /* Simple filter to avoid false positives due to PID collisions */
2994 if (flags & TNF_SHARED)
2997 /* Update priv based on whether false sharing was detected */
3000 if (join && !get_numa_group(grp))
3008 WARN_ON_ONCE(irqs_disabled());
3009 double_lock_irq(&my_grp->lock, &grp->lock);
3011 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) {
3012 my_grp->faults[i] -= p->numa_faults[i];
3013 grp->faults[i] += p->numa_faults[i];
3015 my_grp->total_faults -= p->total_numa_faults;
3016 grp->total_faults += p->total_numa_faults;
3021 spin_unlock(&my_grp->lock);
3022 spin_unlock_irq(&grp->lock);
3024 rcu_assign_pointer(p->numa_group, grp);
3026 put_numa_group(my_grp);
3035 * Get rid of NUMA statistics associated with a task (either current or dead).
3036 * If @final is set, the task is dead and has reached refcount zero, so we can
3037 * safely free all relevant data structures. Otherwise, there might be
3038 * concurrent reads from places like load balancing and procfs, and we should
3039 * reset the data back to default state without freeing ->numa_faults.
3041 void task_numa_free(struct task_struct *p, bool final)
3043 /* safe: p either is current or is being freed by current */
3044 struct numa_group *grp = rcu_dereference_raw(p->numa_group);
3045 unsigned long *numa_faults = p->numa_faults;
3046 unsigned long flags;
3053 spin_lock_irqsave(&grp->lock, flags);
3054 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
3055 grp->faults[i] -= p->numa_faults[i];
3056 grp->total_faults -= p->total_numa_faults;
3059 spin_unlock_irqrestore(&grp->lock, flags);
3060 RCU_INIT_POINTER(p->numa_group, NULL);
3061 put_numa_group(grp);
3065 p->numa_faults = NULL;
3068 p->total_numa_faults = 0;
3069 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
3075 * Got a PROT_NONE fault for a page on @node.
3077 void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags)
3079 struct task_struct *p = current;
3080 bool migrated = flags & TNF_MIGRATED;
3081 int cpu_node = task_node(current);
3082 int local = !!(flags & TNF_FAULT_LOCAL);
3083 struct numa_group *ng;
3086 if (!static_branch_likely(&sched_numa_balancing))
3089 /* for example, ksmd faulting in a user's mm */
3094 * NUMA faults statistics are unnecessary for the slow memory
3095 * node for memory tiering mode.
3097 if (!node_is_toptier(mem_node) &&
3098 (sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING ||
3099 !cpupid_valid(last_cpupid)))
3102 /* Allocate buffer to track faults on a per-node basis */
3103 if (unlikely(!p->numa_faults)) {
3104 int size = sizeof(*p->numa_faults) *
3105 NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids;
3107 p->numa_faults = kzalloc(size, GFP_KERNEL|__GFP_NOWARN);
3108 if (!p->numa_faults)
3111 p->total_numa_faults = 0;
3112 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
3116 * First accesses are treated as private, otherwise consider accesses
3117 * to be private if the accessing pid has not changed
3119 if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) {
3122 priv = cpupid_match_pid(p, last_cpupid);
3123 if (!priv && !(flags & TNF_NO_GROUP))
3124 task_numa_group(p, last_cpupid, flags, &priv);
3128 * If a workload spans multiple NUMA nodes, a shared fault that
3129 * occurs wholly within the set of nodes that the workload is
3130 * actively using should be counted as local. This allows the
3131 * scan rate to slow down when a workload has settled down.
3133 ng = deref_curr_numa_group(p);
3134 if (!priv && !local && ng && ng->active_nodes > 1 &&
3135 numa_is_active_node(cpu_node, ng) &&
3136 numa_is_active_node(mem_node, ng))
3140 * Retry to migrate task to preferred node periodically, in case it
3141 * previously failed, or the scheduler moved us.
3143 if (time_after(jiffies, p->numa_migrate_retry)) {
3144 task_numa_placement(p);
3145 numa_migrate_preferred(p);
3149 p->numa_pages_migrated += pages;
3150 if (flags & TNF_MIGRATE_FAIL)
3151 p->numa_faults_locality[2] += pages;
3153 p->numa_faults[task_faults_idx(NUMA_MEMBUF, mem_node, priv)] += pages;
3154 p->numa_faults[task_faults_idx(NUMA_CPUBUF, cpu_node, priv)] += pages;
3155 p->numa_faults_locality[local] += pages;
3158 static void reset_ptenuma_scan(struct task_struct *p)
3161 * We only did a read acquisition of the mmap sem, so
3162 * p->mm->numa_scan_seq is written to without exclusive access
3163 * and the update is not guaranteed to be atomic. That's not
3164 * much of an issue though, since this is just used for
3165 * statistical sampling. Use READ_ONCE/WRITE_ONCE, which are not
3166 * expensive, to avoid any form of compiler optimizations:
3168 WRITE_ONCE(p->mm->numa_scan_seq, READ_ONCE(p->mm->numa_scan_seq) + 1);
3169 p->mm->numa_scan_offset = 0;
3172 static bool vma_is_accessed(struct mm_struct *mm, struct vm_area_struct *vma)
3176 * Allow unconditional access first two times, so that all the (pages)
3177 * of VMAs get prot_none fault introduced irrespective of accesses.
3178 * This is also done to avoid any side effect of task scanning
3179 * amplifying the unfairness of disjoint set of VMAs' access.
3181 if ((READ_ONCE(current->mm->numa_scan_seq) - vma->numab_state->start_scan_seq) < 2)
3184 pids = vma->numab_state->pids_active[0] | vma->numab_state->pids_active[1];
3185 if (test_bit(hash_32(current->pid, ilog2(BITS_PER_LONG)), &pids))
3189 * Complete a scan that has already started regardless of PID access, or
3190 * some VMAs may never be scanned in multi-threaded applications:
3192 if (mm->numa_scan_offset > vma->vm_start) {
3193 trace_sched_skip_vma_numa(mm, vma, NUMAB_SKIP_IGNORE_PID);
3200 #define VMA_PID_RESET_PERIOD (4 * sysctl_numa_balancing_scan_delay)
3203 * The expensive part of numa migration is done from task_work context.
3204 * Triggered from task_tick_numa().
3206 static void task_numa_work(struct callback_head *work)
3208 unsigned long migrate, next_scan, now = jiffies;
3209 struct task_struct *p = current;
3210 struct mm_struct *mm = p->mm;
3211 u64 runtime = p->se.sum_exec_runtime;
3212 struct vm_area_struct *vma;
3213 unsigned long start, end;
3214 unsigned long nr_pte_updates = 0;
3215 long pages, virtpages;
3216 struct vma_iterator vmi;
3217 bool vma_pids_skipped;
3218 bool vma_pids_forced = false;
3220 SCHED_WARN_ON(p != container_of(work, struct task_struct, numa_work));
3224 * Who cares about NUMA placement when they're dying.
3226 * NOTE: make sure not to dereference p->mm before this check,
3227 * exit_task_work() happens _after_ exit_mm() so we could be called
3228 * without p->mm even though we still had it when we enqueued this
3231 if (p->flags & PF_EXITING)
3234 if (!mm->numa_next_scan) {
3235 mm->numa_next_scan = now +
3236 msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
3240 * Enforce maximal scan/migration frequency..
3242 migrate = mm->numa_next_scan;
3243 if (time_before(now, migrate))
3246 if (p->numa_scan_period == 0) {
3247 p->numa_scan_period_max = task_scan_max(p);
3248 p->numa_scan_period = task_scan_start(p);
3251 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
3252 if (!try_cmpxchg(&mm->numa_next_scan, &migrate, next_scan))
3256 * Delay this task enough that another task of this mm will likely win
3257 * the next time around.
3259 p->node_stamp += 2 * TICK_NSEC;
3261 pages = sysctl_numa_balancing_scan_size;
3262 pages <<= 20 - PAGE_SHIFT; /* MB in pages */
3263 virtpages = pages * 8; /* Scan up to this much virtual space */
3268 if (!mmap_read_trylock(mm))
3272 * VMAs are skipped if the current PID has not trapped a fault within
3273 * the VMA recently. Allow scanning to be forced if there is no
3274 * suitable VMA remaining.
3276 vma_pids_skipped = false;
3279 start = mm->numa_scan_offset;
3280 vma_iter_init(&vmi, mm, start);
3281 vma = vma_next(&vmi);
3283 reset_ptenuma_scan(p);
3285 vma_iter_set(&vmi, start);
3286 vma = vma_next(&vmi);
3290 if (!vma_migratable(vma) || !vma_policy_mof(vma) ||
3291 is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_MIXEDMAP)) {
3292 trace_sched_skip_vma_numa(mm, vma, NUMAB_SKIP_UNSUITABLE);
3297 * Shared library pages mapped by multiple processes are not
3298 * migrated as it is expected they are cache replicated. Avoid
3299 * hinting faults in read-only file-backed mappings or the vdso
3300 * as migrating the pages will be of marginal benefit.
3303 (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ))) {
3304 trace_sched_skip_vma_numa(mm, vma, NUMAB_SKIP_SHARED_RO);
3309 * Skip inaccessible VMAs to avoid any confusion between
3310 * PROT_NONE and NUMA hinting ptes
3312 if (!vma_is_accessible(vma)) {
3313 trace_sched_skip_vma_numa(mm, vma, NUMAB_SKIP_INACCESSIBLE);
3317 /* Initialise new per-VMA NUMAB state. */
3318 if (!vma->numab_state) {
3319 vma->numab_state = kzalloc(sizeof(struct vma_numab_state),
3321 if (!vma->numab_state)
3324 vma->numab_state->start_scan_seq = mm->numa_scan_seq;
3326 vma->numab_state->next_scan = now +
3327 msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
3329 /* Reset happens after 4 times scan delay of scan start */
3330 vma->numab_state->pids_active_reset = vma->numab_state->next_scan +
3331 msecs_to_jiffies(VMA_PID_RESET_PERIOD);
3334 * Ensure prev_scan_seq does not match numa_scan_seq,
3335 * to prevent VMAs being skipped prematurely on the
3338 vma->numab_state->prev_scan_seq = mm->numa_scan_seq - 1;
3342 * Scanning the VMA's of short lived tasks add more overhead. So
3343 * delay the scan for new VMAs.
3345 if (mm->numa_scan_seq && time_before(jiffies,
3346 vma->numab_state->next_scan)) {
3347 trace_sched_skip_vma_numa(mm, vma, NUMAB_SKIP_SCAN_DELAY);
3351 /* RESET access PIDs regularly for old VMAs. */
3352 if (mm->numa_scan_seq &&
3353 time_after(jiffies, vma->numab_state->pids_active_reset)) {
3354 vma->numab_state->pids_active_reset = vma->numab_state->pids_active_reset +
3355 msecs_to_jiffies(VMA_PID_RESET_PERIOD);
3356 vma->numab_state->pids_active[0] = READ_ONCE(vma->numab_state->pids_active[1]);
3357 vma->numab_state->pids_active[1] = 0;
3360 /* Do not rescan VMAs twice within the same sequence. */
3361 if (vma->numab_state->prev_scan_seq == mm->numa_scan_seq) {
3362 mm->numa_scan_offset = vma->vm_end;
3363 trace_sched_skip_vma_numa(mm, vma, NUMAB_SKIP_SEQ_COMPLETED);
3368 * Do not scan the VMA if task has not accessed it, unless no other
3369 * VMA candidate exists.
3371 if (!vma_pids_forced && !vma_is_accessed(mm, vma)) {
3372 vma_pids_skipped = true;
3373 trace_sched_skip_vma_numa(mm, vma, NUMAB_SKIP_PID_INACTIVE);
3378 start = max(start, vma->vm_start);
3379 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
3380 end = min(end, vma->vm_end);
3381 nr_pte_updates = change_prot_numa(vma, start, end);
3384 * Try to scan sysctl_numa_balancing_size worth of
3385 * hpages that have at least one present PTE that
3386 * is not already pte-numa. If the VMA contains
3387 * areas that are unused or already full of prot_numa
3388 * PTEs, scan up to virtpages, to skip through those
3392 pages -= (end - start) >> PAGE_SHIFT;
3393 virtpages -= (end - start) >> PAGE_SHIFT;
3396 if (pages <= 0 || virtpages <= 0)
3400 } while (end != vma->vm_end);
3402 /* VMA scan is complete, do not scan until next sequence. */
3403 vma->numab_state->prev_scan_seq = mm->numa_scan_seq;
3406 * Only force scan within one VMA at a time, to limit the
3407 * cost of scanning a potentially uninteresting VMA.
3409 if (vma_pids_forced)
3411 } for_each_vma(vmi, vma);
3414 * If no VMAs are remaining and VMAs were skipped due to the PID
3415 * not accessing the VMA previously, then force a scan to ensure
3418 if (!vma && !vma_pids_forced && vma_pids_skipped) {
3419 vma_pids_forced = true;
3425 * It is possible to reach the end of the VMA list but the last few
3426 * VMAs are not guaranteed to the vma_migratable. If they are not, we
3427 * would find the !migratable VMA on the next scan but not reset the
3428 * scanner to the start so check it now.
3431 mm->numa_scan_offset = start;
3433 reset_ptenuma_scan(p);
3434 mmap_read_unlock(mm);
3437 * Make sure tasks use at least 32x as much time to run other code
3438 * than they used here, to limit NUMA PTE scanning overhead to 3% max.
3439 * Usually update_task_scan_period slows down scanning enough; on an
3440 * overloaded system we need to limit overhead on a per task basis.
3442 if (unlikely(p->se.sum_exec_runtime != runtime)) {
3443 u64 diff = p->se.sum_exec_runtime - runtime;
3444 p->node_stamp += 32 * diff;
3448 void init_numa_balancing(unsigned long clone_flags, struct task_struct *p)
3451 struct mm_struct *mm = p->mm;
3454 mm_users = atomic_read(&mm->mm_users);
3455 if (mm_users == 1) {
3456 mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
3457 mm->numa_scan_seq = 0;
3461 p->numa_scan_seq = mm ? mm->numa_scan_seq : 0;
3462 p->numa_scan_period = sysctl_numa_balancing_scan_delay;
3463 p->numa_migrate_retry = 0;
3464 /* Protect against double add, see task_tick_numa and task_numa_work */
3465 p->numa_work.next = &p->numa_work;
3466 p->numa_faults = NULL;
3467 p->numa_pages_migrated = 0;
3468 p->total_numa_faults = 0;
3469 RCU_INIT_POINTER(p->numa_group, NULL);
3470 p->last_task_numa_placement = 0;
3471 p->last_sum_exec_runtime = 0;
3473 init_task_work(&p->numa_work, task_numa_work);
3475 /* New address space, reset the preferred nid */
3476 if (!(clone_flags & CLONE_VM)) {
3477 p->numa_preferred_nid = NUMA_NO_NODE;
3482 * New thread, keep existing numa_preferred_nid which should be copied
3483 * already by arch_dup_task_struct but stagger when scans start.
3488 delay = min_t(unsigned int, task_scan_max(current),
3489 current->numa_scan_period * mm_users * NSEC_PER_MSEC);
3490 delay += 2 * TICK_NSEC;
3491 p->node_stamp = delay;
3496 * Drive the periodic memory faults..
3498 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
3500 struct callback_head *work = &curr->numa_work;
3504 * We don't care about NUMA placement if we don't have memory.
3506 if (!curr->mm || (curr->flags & (PF_EXITING | PF_KTHREAD)) || work->next != work)
3510 * Using runtime rather than walltime has the dual advantage that
3511 * we (mostly) drive the selection from busy threads and that the
3512 * task needs to have done some actual work before we bother with
3515 now = curr->se.sum_exec_runtime;
3516 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
3518 if (now > curr->node_stamp + period) {
3519 if (!curr->node_stamp)
3520 curr->numa_scan_period = task_scan_start(curr);
3521 curr->node_stamp += period;
3523 if (!time_before(jiffies, curr->mm->numa_next_scan))
3524 task_work_add(curr, work, TWA_RESUME);
3528 static void update_scan_period(struct task_struct *p, int new_cpu)
3530 int src_nid = cpu_to_node(task_cpu(p));
3531 int dst_nid = cpu_to_node(new_cpu);
3533 if (!static_branch_likely(&sched_numa_balancing))
3536 if (!p->mm || !p->numa_faults || (p->flags & PF_EXITING))
3539 if (src_nid == dst_nid)
3543 * Allow resets if faults have been trapped before one scan
3544 * has completed. This is most likely due to a new task that
3545 * is pulled cross-node due to wakeups or load balancing.
3547 if (p->numa_scan_seq) {
3549 * Avoid scan adjustments if moving to the preferred
3550 * node or if the task was not previously running on
3551 * the preferred node.
3553 if (dst_nid == p->numa_preferred_nid ||
3554 (p->numa_preferred_nid != NUMA_NO_NODE &&
3555 src_nid != p->numa_preferred_nid))
3559 p->numa_scan_period = task_scan_start(p);
3563 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
3567 static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p)
3571 static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p)
3575 static inline void update_scan_period(struct task_struct *p, int new_cpu)
3579 #endif /* CONFIG_NUMA_BALANCING */
3582 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
3584 update_load_add(&cfs_rq->load, se->load.weight);
3586 if (entity_is_task(se)) {
3587 struct rq *rq = rq_of(cfs_rq);
3589 account_numa_enqueue(rq, task_of(se));
3590 list_add(&se->group_node, &rq->cfs_tasks);
3593 cfs_rq->nr_running++;
3595 cfs_rq->idle_nr_running++;
3599 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
3601 update_load_sub(&cfs_rq->load, se->load.weight);
3603 if (entity_is_task(se)) {
3604 account_numa_dequeue(rq_of(cfs_rq), task_of(se));
3605 list_del_init(&se->group_node);
3608 cfs_rq->nr_running--;
3610 cfs_rq->idle_nr_running--;
3614 * Signed add and clamp on underflow.
3616 * Explicitly do a load-store to ensure the intermediate value never hits
3617 * memory. This allows lockless observations without ever seeing the negative
3620 #define add_positive(_ptr, _val) do { \
3621 typeof(_ptr) ptr = (_ptr); \
3622 typeof(_val) val = (_val); \
3623 typeof(*ptr) res, var = READ_ONCE(*ptr); \
3627 if (val < 0 && res > var) \
3630 WRITE_ONCE(*ptr, res); \
3634 * Unsigned subtract and clamp on underflow.
3636 * Explicitly do a load-store to ensure the intermediate value never hits
3637 * memory. This allows lockless observations without ever seeing the negative
3640 #define sub_positive(_ptr, _val) do { \
3641 typeof(_ptr) ptr = (_ptr); \
3642 typeof(*ptr) val = (_val); \
3643 typeof(*ptr) res, var = READ_ONCE(*ptr); \
3647 WRITE_ONCE(*ptr, res); \
3651 * Remove and clamp on negative, from a local variable.
3653 * A variant of sub_positive(), which does not use explicit load-store
3654 * and is thus optimized for local variable updates.
3656 #define lsub_positive(_ptr, _val) do { \
3657 typeof(_ptr) ptr = (_ptr); \
3658 *ptr -= min_t(typeof(*ptr), *ptr, _val); \
3663 enqueue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3665 cfs_rq->avg.load_avg += se->avg.load_avg;
3666 cfs_rq->avg.load_sum += se_weight(se) * se->avg.load_sum;
3670 dequeue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3672 sub_positive(&cfs_rq->avg.load_avg, se->avg.load_avg);
3673 sub_positive(&cfs_rq->avg.load_sum, se_weight(se) * se->avg.load_sum);
3674 /* See update_cfs_rq_load_avg() */
3675 cfs_rq->avg.load_sum = max_t(u32, cfs_rq->avg.load_sum,
3676 cfs_rq->avg.load_avg * PELT_MIN_DIVIDER);
3680 enqueue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
3682 dequeue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
3685 static void reweight_eevdf(struct sched_entity *se, u64 avruntime,
3686 unsigned long weight)
3688 unsigned long old_weight = se->load.weight;
3695 * COROLLARY #1: The virtual runtime of the entity needs to be
3696 * adjusted if re-weight at !0-lag point.
3698 * Proof: For contradiction assume this is not true, so we can
3699 * re-weight without changing vruntime at !0-lag point.
3701 * Weight VRuntime Avg-VRuntime
3705 * Since lag needs to be preserved through re-weight:
3707 * lag = (V - v)*w = (V'- v')*w', where v = v'
3708 * ==> V' = (V - v)*w/w' + v (1)
3710 * Let W be the total weight of the entities before reweight,
3711 * since V' is the new weighted average of entities:
3713 * V' = (WV + w'v - wv) / (W + w' - w) (2)
3715 * by using (1) & (2) we obtain:
3717 * (WV + w'v - wv) / (W + w' - w) = (V - v)*w/w' + v
3718 * ==> (WV-Wv+Wv+w'v-wv)/(W+w'-w) = (V - v)*w/w' + v
3719 * ==> (WV - Wv)/(W + w' - w) + v = (V - v)*w/w' + v
3720 * ==> (V - v)*W/(W + w' - w) = (V - v)*w/w' (3)
3722 * Since we are doing at !0-lag point which means V != v, we
3725 * ==> W / (W + w' - w) = w / w'
3726 * ==> Ww' = Ww + ww' - ww
3727 * ==> W * (w' - w) = w * (w' - w)
3728 * ==> W = w (re-weight indicates w' != w)
3730 * So the cfs_rq contains only one entity, hence vruntime of
3731 * the entity @v should always equal to the cfs_rq's weighted
3732 * average vruntime @V, which means we will always re-weight
3733 * at 0-lag point, thus breach assumption. Proof completed.
3736 * COROLLARY #2: Re-weight does NOT affect weighted average
3737 * vruntime of all the entities.
3739 * Proof: According to corollary #1, Eq. (1) should be:
3741 * (V - v)*w = (V' - v')*w'
3742 * ==> v' = V' - (V - v)*w/w' (4)
3744 * According to the weighted average formula, we have:
3746 * V' = (WV - wv + w'v') / (W - w + w')
3747 * = (WV - wv + w'(V' - (V - v)w/w')) / (W - w + w')
3748 * = (WV - wv + w'V' - Vw + wv) / (W - w + w')
3749 * = (WV + w'V' - Vw) / (W - w + w')
3751 * ==> V'*(W - w + w') = WV + w'V' - Vw
3752 * ==> V' * (W - w) = (W - w) * V (5)
3754 * If the entity is the only one in the cfs_rq, then reweight
3755 * always occurs at 0-lag point, so V won't change. Or else
3756 * there are other entities, hence W != w, then Eq. (5) turns
3757 * into V' = V. So V won't change in either case, proof done.
3760 * So according to corollary #1 & #2, the effect of re-weight
3761 * on vruntime should be:
3763 * v' = V' - (V - v) * w / w' (4)
3764 * = V - (V - v) * w / w'
3768 if (avruntime != se->vruntime) {
3769 vlag = entity_lag(avruntime, se);
3770 vlag = div_s64(vlag * old_weight, weight);
3771 se->vruntime = avruntime - vlag;
3778 * When the weight changes, the virtual time slope changes and
3779 * we should adjust the relative virtual deadline accordingly.
3781 * d' = v' + (d - v)*w/w'
3782 * = V' - (V - v)*w/w' + (d - v)*w/w'
3783 * = V - (V - v)*w/w' + (d - v)*w/w'
3784 * = V + (d - V)*w/w'
3786 vslice = (s64)(se->deadline - avruntime);
3787 vslice = div_s64(vslice * old_weight, weight);
3788 se->deadline = avruntime + vslice;
3791 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
3792 unsigned long weight)
3794 bool curr = cfs_rq->curr == se;
3798 /* commit outstanding execution time */
3799 update_curr(cfs_rq);
3800 avruntime = avg_vruntime(cfs_rq);
3802 __dequeue_entity(cfs_rq, se);
3803 update_load_sub(&cfs_rq->load, se->load.weight);
3805 dequeue_load_avg(cfs_rq, se);
3808 reweight_eevdf(se, avruntime, weight);
3811 * Because we keep se->vlag = V - v_i, while: lag_i = w_i*(V - v_i),
3812 * we need to scale se->vlag when w_i changes.
3814 se->vlag = div_s64(se->vlag * se->load.weight, weight);
3817 update_load_set(&se->load, weight);
3821 u32 divider = get_pelt_divider(&se->avg);
3823 se->avg.load_avg = div_u64(se_weight(se) * se->avg.load_sum, divider);
3827 enqueue_load_avg(cfs_rq, se);
3829 update_load_add(&cfs_rq->load, se->load.weight);
3831 __enqueue_entity(cfs_rq, se);
3834 * The entity's vruntime has been adjusted, so let's check
3835 * whether the rq-wide min_vruntime needs updated too. Since
3836 * the calculations above require stable min_vruntime rather
3837 * than up-to-date one, we do the update at the end of the
3840 update_min_vruntime(cfs_rq);
3844 void reweight_task(struct task_struct *p, int prio)
3846 struct sched_entity *se = &p->se;
3847 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3848 struct load_weight *load = &se->load;
3849 unsigned long weight = scale_load(sched_prio_to_weight[prio]);
3851 reweight_entity(cfs_rq, se, weight);
3852 load->inv_weight = sched_prio_to_wmult[prio];
3855 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
3857 #ifdef CONFIG_FAIR_GROUP_SCHED
3860 * All this does is approximate the hierarchical proportion which includes that
3861 * global sum we all love to hate.
3863 * That is, the weight of a group entity, is the proportional share of the
3864 * group weight based on the group runqueue weights. That is:
3866 * tg->weight * grq->load.weight
3867 * ge->load.weight = ----------------------------- (1)
3868 * \Sum grq->load.weight
3870 * Now, because computing that sum is prohibitively expensive to compute (been
3871 * there, done that) we approximate it with this average stuff. The average
3872 * moves slower and therefore the approximation is cheaper and more stable.
3874 * So instead of the above, we substitute:
3876 * grq->load.weight -> grq->avg.load_avg (2)
3878 * which yields the following:
3880 * tg->weight * grq->avg.load_avg
3881 * ge->load.weight = ------------------------------ (3)
3884 * Where: tg->load_avg ~= \Sum grq->avg.load_avg
3886 * That is shares_avg, and it is right (given the approximation (2)).
3888 * The problem with it is that because the average is slow -- it was designed
3889 * to be exactly that of course -- this leads to transients in boundary
3890 * conditions. In specific, the case where the group was idle and we start the
3891 * one task. It takes time for our CPU's grq->avg.load_avg to build up,
3892 * yielding bad latency etc..
3894 * Now, in that special case (1) reduces to:
3896 * tg->weight * grq->load.weight
3897 * ge->load.weight = ----------------------------- = tg->weight (4)
3900 * That is, the sum collapses because all other CPUs are idle; the UP scenario.
3902 * So what we do is modify our approximation (3) to approach (4) in the (near)
3907 * tg->weight * grq->load.weight
3908 * --------------------------------------------------- (5)
3909 * tg->load_avg - grq->avg.load_avg + grq->load.weight
3911 * But because grq->load.weight can drop to 0, resulting in a divide by zero,
3912 * we need to use grq->avg.load_avg as its lower bound, which then gives:
3915 * tg->weight * grq->load.weight
3916 * ge->load.weight = ----------------------------- (6)
3921 * tg_load_avg' = tg->load_avg - grq->avg.load_avg +
3922 * max(grq->load.weight, grq->avg.load_avg)
3924 * And that is shares_weight and is icky. In the (near) UP case it approaches
3925 * (4) while in the normal case it approaches (3). It consistently
3926 * overestimates the ge->load.weight and therefore:
3928 * \Sum ge->load.weight >= tg->weight
3932 static long calc_group_shares(struct cfs_rq *cfs_rq)
3934 long tg_weight, tg_shares, load, shares;
3935 struct task_group *tg = cfs_rq->tg;
3937 tg_shares = READ_ONCE(tg->shares);
3939 load = max(scale_load_down(cfs_rq->load.weight), cfs_rq->avg.load_avg);
3941 tg_weight = atomic_long_read(&tg->load_avg);
3943 /* Ensure tg_weight >= load */
3944 tg_weight -= cfs_rq->tg_load_avg_contrib;
3947 shares = (tg_shares * load);
3949 shares /= tg_weight;
3952 * MIN_SHARES has to be unscaled here to support per-CPU partitioning
3953 * of a group with small tg->shares value. It is a floor value which is
3954 * assigned as a minimum load.weight to the sched_entity representing
3955 * the group on a CPU.
3957 * E.g. on 64-bit for a group with tg->shares of scale_load(15)=15*1024
3958 * on an 8-core system with 8 tasks each runnable on one CPU shares has
3959 * to be 15*1024*1/8=1920 instead of scale_load(MIN_SHARES)=2*1024. In
3960 * case no task is runnable on a CPU MIN_SHARES=2 should be returned
3963 return clamp_t(long, shares, MIN_SHARES, tg_shares);
3965 #endif /* CONFIG_SMP */
3968 * Recomputes the group entity based on the current state of its group
3971 static void update_cfs_group(struct sched_entity *se)
3973 struct cfs_rq *gcfs_rq = group_cfs_rq(se);
3979 if (throttled_hierarchy(gcfs_rq))
3983 shares = READ_ONCE(gcfs_rq->tg->shares);
3985 shares = calc_group_shares(gcfs_rq);
3987 if (unlikely(se->load.weight != shares))
3988 reweight_entity(cfs_rq_of(se), se, shares);
3991 #else /* CONFIG_FAIR_GROUP_SCHED */
3992 static inline void update_cfs_group(struct sched_entity *se)
3995 #endif /* CONFIG_FAIR_GROUP_SCHED */
3997 static inline void cfs_rq_util_change(struct cfs_rq *cfs_rq, int flags)
3999 struct rq *rq = rq_of(cfs_rq);
4001 if (&rq->cfs == cfs_rq) {
4003 * There are a few boundary cases this might miss but it should
4004 * get called often enough that that should (hopefully) not be
4007 * It will not get called when we go idle, because the idle
4008 * thread is a different class (!fair), nor will the utilization
4009 * number include things like RT tasks.
4011 * As is, the util number is not freq-invariant (we'd have to
4012 * implement arch_scale_freq_capacity() for that).
4014 * See cpu_util_cfs().
4016 cpufreq_update_util(rq, flags);
4021 static inline bool load_avg_is_decayed(struct sched_avg *sa)
4029 if (sa->runnable_sum)
4033 * _avg must be null when _sum are null because _avg = _sum / divider
4034 * Make sure that rounding and/or propagation of PELT values never
4037 SCHED_WARN_ON(sa->load_avg ||
4044 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
4046 return u64_u32_load_copy(cfs_rq->avg.last_update_time,
4047 cfs_rq->last_update_time_copy);
4049 #ifdef CONFIG_FAIR_GROUP_SCHED
4051 * Because list_add_leaf_cfs_rq always places a child cfs_rq on the list
4052 * immediately before a parent cfs_rq, and cfs_rqs are removed from the list
4053 * bottom-up, we only have to test whether the cfs_rq before us on the list
4055 * If cfs_rq is not on the list, test whether a child needs its to be added to
4056 * connect a branch to the tree * (see list_add_leaf_cfs_rq() for details).
4058 static inline bool child_cfs_rq_on_list(struct cfs_rq *cfs_rq)
4060 struct cfs_rq *prev_cfs_rq;
4061 struct list_head *prev;
4063 if (cfs_rq->on_list) {
4064 prev = cfs_rq->leaf_cfs_rq_list.prev;
4066 struct rq *rq = rq_of(cfs_rq);
4068 prev = rq->tmp_alone_branch;
4071 prev_cfs_rq = container_of(prev, struct cfs_rq, leaf_cfs_rq_list);
4073 return (prev_cfs_rq->tg->parent == cfs_rq->tg);
4076 static inline bool cfs_rq_is_decayed(struct cfs_rq *cfs_rq)
4078 if (cfs_rq->load.weight)
4081 if (!load_avg_is_decayed(&cfs_rq->avg))
4084 if (child_cfs_rq_on_list(cfs_rq))
4091 * update_tg_load_avg - update the tg's load avg
4092 * @cfs_rq: the cfs_rq whose avg changed
4094 * This function 'ensures': tg->load_avg := \Sum tg->cfs_rq[]->avg.load.
4095 * However, because tg->load_avg is a global value there are performance
4098 * In order to avoid having to look at the other cfs_rq's, we use a
4099 * differential update where we store the last value we propagated. This in
4100 * turn allows skipping updates if the differential is 'small'.
4102 * Updating tg's load_avg is necessary before update_cfs_share().
4104 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq)
4110 * No need to update load_avg for root_task_group as it is not used.
4112 if (cfs_rq->tg == &root_task_group)
4115 /* rq has been offline and doesn't contribute to the share anymore: */
4116 if (!cpu_active(cpu_of(rq_of(cfs_rq))))
4120 * For migration heavy workloads, access to tg->load_avg can be
4121 * unbound. Limit the update rate to at most once per ms.
4123 now = sched_clock_cpu(cpu_of(rq_of(cfs_rq)));
4124 if (now - cfs_rq->last_update_tg_load_avg < NSEC_PER_MSEC)
4127 delta = cfs_rq->avg.load_avg - cfs_rq->tg_load_avg_contrib;
4128 if (abs(delta) > cfs_rq->tg_load_avg_contrib / 64) {
4129 atomic_long_add(delta, &cfs_rq->tg->load_avg);
4130 cfs_rq->tg_load_avg_contrib = cfs_rq->avg.load_avg;
4131 cfs_rq->last_update_tg_load_avg = now;
4135 static inline void clear_tg_load_avg(struct cfs_rq *cfs_rq)
4141 * No need to update load_avg for root_task_group, as it is not used.
4143 if (cfs_rq->tg == &root_task_group)
4146 now = sched_clock_cpu(cpu_of(rq_of(cfs_rq)));
4147 delta = 0 - cfs_rq->tg_load_avg_contrib;
4148 atomic_long_add(delta, &cfs_rq->tg->load_avg);
4149 cfs_rq->tg_load_avg_contrib = 0;
4150 cfs_rq->last_update_tg_load_avg = now;
4153 /* CPU offline callback: */
4154 static void __maybe_unused clear_tg_offline_cfs_rqs(struct rq *rq)
4156 struct task_group *tg;
4158 lockdep_assert_rq_held(rq);
4161 * The rq clock has already been updated in
4162 * set_rq_offline(), so we should skip updating
4163 * the rq clock again in unthrottle_cfs_rq().
4165 rq_clock_start_loop_update(rq);
4168 list_for_each_entry_rcu(tg, &task_groups, list) {
4169 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
4171 clear_tg_load_avg(cfs_rq);
4175 rq_clock_stop_loop_update(rq);
4179 * Called within set_task_rq() right before setting a task's CPU. The
4180 * caller only guarantees p->pi_lock is held; no other assumptions,
4181 * including the state of rq->lock, should be made.
4183 void set_task_rq_fair(struct sched_entity *se,
4184 struct cfs_rq *prev, struct cfs_rq *next)
4186 u64 p_last_update_time;
4187 u64 n_last_update_time;
4189 if (!sched_feat(ATTACH_AGE_LOAD))
4193 * We are supposed to update the task to "current" time, then its up to
4194 * date and ready to go to new CPU/cfs_rq. But we have difficulty in
4195 * getting what current time is, so simply throw away the out-of-date
4196 * time. This will result in the wakee task is less decayed, but giving
4197 * the wakee more load sounds not bad.
4199 if (!(se->avg.last_update_time && prev))
4202 p_last_update_time = cfs_rq_last_update_time(prev);
4203 n_last_update_time = cfs_rq_last_update_time(next);
4205 __update_load_avg_blocked_se(p_last_update_time, se);
4206 se->avg.last_update_time = n_last_update_time;
4210 * When on migration a sched_entity joins/leaves the PELT hierarchy, we need to
4211 * propagate its contribution. The key to this propagation is the invariant
4212 * that for each group:
4214 * ge->avg == grq->avg (1)
4216 * _IFF_ we look at the pure running and runnable sums. Because they
4217 * represent the very same entity, just at different points in the hierarchy.
4219 * Per the above update_tg_cfs_util() and update_tg_cfs_runnable() are trivial
4220 * and simply copies the running/runnable sum over (but still wrong, because
4221 * the group entity and group rq do not have their PELT windows aligned).
4223 * However, update_tg_cfs_load() is more complex. So we have:
4225 * ge->avg.load_avg = ge->load.weight * ge->avg.runnable_avg (2)
4227 * And since, like util, the runnable part should be directly transferable,
4228 * the following would _appear_ to be the straight forward approach:
4230 * grq->avg.load_avg = grq->load.weight * grq->avg.runnable_avg (3)
4232 * And per (1) we have:
4234 * ge->avg.runnable_avg == grq->avg.runnable_avg
4238 * ge->load.weight * grq->avg.load_avg
4239 * ge->avg.load_avg = ----------------------------------- (4)
4242 * Except that is wrong!
4244 * Because while for entities historical weight is not important and we
4245 * really only care about our future and therefore can consider a pure
4246 * runnable sum, runqueues can NOT do this.
4248 * We specifically want runqueues to have a load_avg that includes
4249 * historical weights. Those represent the blocked load, the load we expect
4250 * to (shortly) return to us. This only works by keeping the weights as
4251 * integral part of the sum. We therefore cannot decompose as per (3).
4253 * Another reason this doesn't work is that runnable isn't a 0-sum entity.
4254 * Imagine a rq with 2 tasks that each are runnable 2/3 of the time. Then the
4255 * rq itself is runnable anywhere between 2/3 and 1 depending on how the
4256 * runnable section of these tasks overlap (or not). If they were to perfectly
4257 * align the rq as a whole would be runnable 2/3 of the time. If however we
4258 * always have at least 1 runnable task, the rq as a whole is always runnable.
4260 * So we'll have to approximate.. :/
4262 * Given the constraint:
4264 * ge->avg.running_sum <= ge->avg.runnable_sum <= LOAD_AVG_MAX
4266 * We can construct a rule that adds runnable to a rq by assuming minimal
4269 * On removal, we'll assume each task is equally runnable; which yields:
4271 * grq->avg.runnable_sum = grq->avg.load_sum / grq->load.weight
4273 * XXX: only do this for the part of runnable > running ?
4277 update_tg_cfs_util(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq)
4279 long delta_sum, delta_avg = gcfs_rq->avg.util_avg - se->avg.util_avg;
4280 u32 new_sum, divider;
4282 /* Nothing to update */
4287 * cfs_rq->avg.period_contrib can be used for both cfs_rq and se.
4288 * See ___update_load_avg() for details.
4290 divider = get_pelt_divider(&cfs_rq->avg);
4293 /* Set new sched_entity's utilization */
4294 se->avg.util_avg = gcfs_rq->avg.util_avg;
4295 new_sum = se->avg.util_avg * divider;
4296 delta_sum = (long)new_sum - (long)se->avg.util_sum;
4297 se->avg.util_sum = new_sum;
4299 /* Update parent cfs_rq utilization */
4300 add_positive(&cfs_rq->avg.util_avg, delta_avg);
4301 add_positive(&cfs_rq->avg.util_sum, delta_sum);
4303 /* See update_cfs_rq_load_avg() */
4304 cfs_rq->avg.util_sum = max_t(u32, cfs_rq->avg.util_sum,
4305 cfs_rq->avg.util_avg * PELT_MIN_DIVIDER);
4309 update_tg_cfs_runnable(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq)
4311 long delta_sum, delta_avg = gcfs_rq->avg.runnable_avg - se->avg.runnable_avg;
4312 u32 new_sum, divider;
4314 /* Nothing to update */
4319 * cfs_rq->avg.period_contrib can be used for both cfs_rq and se.
4320 * See ___update_load_avg() for details.
4322 divider = get_pelt_divider(&cfs_rq->avg);
4324 /* Set new sched_entity's runnable */
4325 se->avg.runnable_avg = gcfs_rq->avg.runnable_avg;
4326 new_sum = se->avg.runnable_avg * divider;
4327 delta_sum = (long)new_sum - (long)se->avg.runnable_sum;
4328 se->avg.runnable_sum = new_sum;
4330 /* Update parent cfs_rq runnable */
4331 add_positive(&cfs_rq->avg.runnable_avg, delta_avg);
4332 add_positive(&cfs_rq->avg.runnable_sum, delta_sum);
4333 /* See update_cfs_rq_load_avg() */
4334 cfs_rq->avg.runnable_sum = max_t(u32, cfs_rq->avg.runnable_sum,
4335 cfs_rq->avg.runnable_avg * PELT_MIN_DIVIDER);
4339 update_tg_cfs_load(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq)
4341 long delta_avg, running_sum, runnable_sum = gcfs_rq->prop_runnable_sum;
4342 unsigned long load_avg;
4350 gcfs_rq->prop_runnable_sum = 0;
4353 * cfs_rq->avg.period_contrib can be used for both cfs_rq and se.
4354 * See ___update_load_avg() for details.
4356 divider = get_pelt_divider(&cfs_rq->avg);
4358 if (runnable_sum >= 0) {
4360 * Add runnable; clip at LOAD_AVG_MAX. Reflects that until
4361 * the CPU is saturated running == runnable.
4363 runnable_sum += se->avg.load_sum;
4364 runnable_sum = min_t(long, runnable_sum, divider);
4367 * Estimate the new unweighted runnable_sum of the gcfs_rq by
4368 * assuming all tasks are equally runnable.
4370 if (scale_load_down(gcfs_rq->load.weight)) {
4371 load_sum = div_u64(gcfs_rq->avg.load_sum,
4372 scale_load_down(gcfs_rq->load.weight));
4375 /* But make sure to not inflate se's runnable */
4376 runnable_sum = min(se->avg.load_sum, load_sum);
4380 * runnable_sum can't be lower than running_sum
4381 * Rescale running sum to be in the same range as runnable sum
4382 * running_sum is in [0 : LOAD_AVG_MAX << SCHED_CAPACITY_SHIFT]
4383 * runnable_sum is in [0 : LOAD_AVG_MAX]
4385 running_sum = se->avg.util_sum >> SCHED_CAPACITY_SHIFT;
4386 runnable_sum = max(runnable_sum, running_sum);
4388 load_sum = se_weight(se) * runnable_sum;
4389 load_avg = div_u64(load_sum, divider);
4391 delta_avg = load_avg - se->avg.load_avg;
4395 delta_sum = load_sum - (s64)se_weight(se) * se->avg.load_sum;
4397 se->avg.load_sum = runnable_sum;
4398 se->avg.load_avg = load_avg;
4399 add_positive(&cfs_rq->avg.load_avg, delta_avg);
4400 add_positive(&cfs_rq->avg.load_sum, delta_sum);
4401 /* See update_cfs_rq_load_avg() */
4402 cfs_rq->avg.load_sum = max_t(u32, cfs_rq->avg.load_sum,
4403 cfs_rq->avg.load_avg * PELT_MIN_DIVIDER);
4406 static inline void add_tg_cfs_propagate(struct cfs_rq *cfs_rq, long runnable_sum)
4408 cfs_rq->propagate = 1;
4409 cfs_rq->prop_runnable_sum += runnable_sum;
4412 /* Update task and its cfs_rq load average */
4413 static inline int propagate_entity_load_avg(struct sched_entity *se)
4415 struct cfs_rq *cfs_rq, *gcfs_rq;
4417 if (entity_is_task(se))
4420 gcfs_rq = group_cfs_rq(se);
4421 if (!gcfs_rq->propagate)
4424 gcfs_rq->propagate = 0;
4426 cfs_rq = cfs_rq_of(se);
4428 add_tg_cfs_propagate(cfs_rq, gcfs_rq->prop_runnable_sum);
4430 update_tg_cfs_util(cfs_rq, se, gcfs_rq);
4431 update_tg_cfs_runnable(cfs_rq, se, gcfs_rq);
4432 update_tg_cfs_load(cfs_rq, se, gcfs_rq);
4434 trace_pelt_cfs_tp(cfs_rq);
4435 trace_pelt_se_tp(se);
4441 * Check if we need to update the load and the utilization of a blocked
4444 static inline bool skip_blocked_update(struct sched_entity *se)
4446 struct cfs_rq *gcfs_rq = group_cfs_rq(se);
4449 * If sched_entity still have not zero load or utilization, we have to
4452 if (se->avg.load_avg || se->avg.util_avg)
4456 * If there is a pending propagation, we have to update the load and
4457 * the utilization of the sched_entity:
4459 if (gcfs_rq->propagate)
4463 * Otherwise, the load and the utilization of the sched_entity is
4464 * already zero and there is no pending propagation, so it will be a
4465 * waste of time to try to decay it:
4470 #else /* CONFIG_FAIR_GROUP_SCHED */
4472 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq) {}
4474 static inline void clear_tg_offline_cfs_rqs(struct rq *rq) {}
4476 static inline int propagate_entity_load_avg(struct sched_entity *se)
4481 static inline void add_tg_cfs_propagate(struct cfs_rq *cfs_rq, long runnable_sum) {}
4483 #endif /* CONFIG_FAIR_GROUP_SCHED */
4485 #ifdef CONFIG_NO_HZ_COMMON
4486 static inline void migrate_se_pelt_lag(struct sched_entity *se)
4488 u64 throttled = 0, now, lut;
4489 struct cfs_rq *cfs_rq;
4493 if (load_avg_is_decayed(&se->avg))
4496 cfs_rq = cfs_rq_of(se);
4500 is_idle = is_idle_task(rcu_dereference(rq->curr));
4504 * The lag estimation comes with a cost we don't want to pay all the
4505 * time. Hence, limiting to the case where the source CPU is idle and
4506 * we know we are at the greatest risk to have an outdated clock.
4512 * Estimated "now" is: last_update_time + cfs_idle_lag + rq_idle_lag, where:
4514 * last_update_time (the cfs_rq's last_update_time)
4515 * = cfs_rq_clock_pelt()@cfs_rq_idle
4516 * = rq_clock_pelt()@cfs_rq_idle
4517 * - cfs->throttled_clock_pelt_time@cfs_rq_idle
4519 * cfs_idle_lag (delta between rq's update and cfs_rq's update)
4520 * = rq_clock_pelt()@rq_idle - rq_clock_pelt()@cfs_rq_idle
4522 * rq_idle_lag (delta between now and rq's update)
4523 * = sched_clock_cpu() - rq_clock()@rq_idle
4525 * We can then write:
4527 * now = rq_clock_pelt()@rq_idle - cfs->throttled_clock_pelt_time +
4528 * sched_clock_cpu() - rq_clock()@rq_idle
4530 * rq_clock_pelt()@rq_idle is rq->clock_pelt_idle
4531 * rq_clock()@rq_idle is rq->clock_idle
4532 * cfs->throttled_clock_pelt_time@cfs_rq_idle
4533 * is cfs_rq->throttled_pelt_idle
4536 #ifdef CONFIG_CFS_BANDWIDTH
4537 throttled = u64_u32_load(cfs_rq->throttled_pelt_idle);
4538 /* The clock has been stopped for throttling */
4539 if (throttled == U64_MAX)
4542 now = u64_u32_load(rq->clock_pelt_idle);
4544 * Paired with _update_idle_rq_clock_pelt(). It ensures at the worst case
4545 * is observed the old clock_pelt_idle value and the new clock_idle,
4546 * which lead to an underestimation. The opposite would lead to an
4550 lut = cfs_rq_last_update_time(cfs_rq);
4555 * cfs_rq->avg.last_update_time is more recent than our
4556 * estimation, let's use it.
4560 now += sched_clock_cpu(cpu_of(rq)) - u64_u32_load(rq->clock_idle);
4562 __update_load_avg_blocked_se(now, se);
4565 static void migrate_se_pelt_lag(struct sched_entity *se) {}
4569 * update_cfs_rq_load_avg - update the cfs_rq's load/util averages
4570 * @now: current time, as per cfs_rq_clock_pelt()
4571 * @cfs_rq: cfs_rq to update
4573 * The cfs_rq avg is the direct sum of all its entities (blocked and runnable)
4574 * avg. The immediate corollary is that all (fair) tasks must be attached.
4576 * cfs_rq->avg is used for task_h_load() and update_cfs_share() for example.
4578 * Return: true if the load decayed or we removed load.
4580 * Since both these conditions indicate a changed cfs_rq->avg.load we should
4581 * call update_tg_load_avg() when this function returns true.
4584 update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq)
4586 unsigned long removed_load = 0, removed_util = 0, removed_runnable = 0;
4587 struct sched_avg *sa = &cfs_rq->avg;
4590 if (cfs_rq->removed.nr) {
4592 u32 divider = get_pelt_divider(&cfs_rq->avg);
4594 raw_spin_lock(&cfs_rq->removed.lock);
4595 swap(cfs_rq->removed.util_avg, removed_util);
4596 swap(cfs_rq->removed.load_avg, removed_load);
4597 swap(cfs_rq->removed.runnable_avg, removed_runnable);
4598 cfs_rq->removed.nr = 0;
4599 raw_spin_unlock(&cfs_rq->removed.lock);
4602 sub_positive(&sa->load_avg, r);
4603 sub_positive(&sa->load_sum, r * divider);
4604 /* See sa->util_sum below */
4605 sa->load_sum = max_t(u32, sa->load_sum, sa->load_avg * PELT_MIN_DIVIDER);
4608 sub_positive(&sa->util_avg, r);
4609 sub_positive(&sa->util_sum, r * divider);
4611 * Because of rounding, se->util_sum might ends up being +1 more than
4612 * cfs->util_sum. Although this is not a problem by itself, detaching
4613 * a lot of tasks with the rounding problem between 2 updates of
4614 * util_avg (~1ms) can make cfs->util_sum becoming null whereas
4615 * cfs_util_avg is not.
4616 * Check that util_sum is still above its lower bound for the new
4617 * util_avg. Given that period_contrib might have moved since the last
4618 * sync, we are only sure that util_sum must be above or equal to
4619 * util_avg * minimum possible divider
4621 sa->util_sum = max_t(u32, sa->util_sum, sa->util_avg * PELT_MIN_DIVIDER);
4623 r = removed_runnable;
4624 sub_positive(&sa->runnable_avg, r);
4625 sub_positive(&sa->runnable_sum, r * divider);
4626 /* See sa->util_sum above */
4627 sa->runnable_sum = max_t(u32, sa->runnable_sum,
4628 sa->runnable_avg * PELT_MIN_DIVIDER);
4631 * removed_runnable is the unweighted version of removed_load so we
4632 * can use it to estimate removed_load_sum.
4634 add_tg_cfs_propagate(cfs_rq,
4635 -(long)(removed_runnable * divider) >> SCHED_CAPACITY_SHIFT);
4640 decayed |= __update_load_avg_cfs_rq(now, cfs_rq);
4641 u64_u32_store_copy(sa->last_update_time,
4642 cfs_rq->last_update_time_copy,
4643 sa->last_update_time);
4648 * attach_entity_load_avg - attach this entity to its cfs_rq load avg
4649 * @cfs_rq: cfs_rq to attach to
4650 * @se: sched_entity to attach
4652 * Must call update_cfs_rq_load_avg() before this, since we rely on
4653 * cfs_rq->avg.last_update_time being current.
4655 static void attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
4658 * cfs_rq->avg.period_contrib can be used for both cfs_rq and se.
4659 * See ___update_load_avg() for details.
4661 u32 divider = get_pelt_divider(&cfs_rq->avg);
4664 * When we attach the @se to the @cfs_rq, we must align the decay
4665 * window because without that, really weird and wonderful things can
4670 se->avg.last_update_time = cfs_rq->avg.last_update_time;
4671 se->avg.period_contrib = cfs_rq->avg.period_contrib;
4674 * Hell(o) Nasty stuff.. we need to recompute _sum based on the new
4675 * period_contrib. This isn't strictly correct, but since we're
4676 * entirely outside of the PELT hierarchy, nobody cares if we truncate
4679 se->avg.util_sum = se->avg.util_avg * divider;
4681 se->avg.runnable_sum = se->avg.runnable_avg * divider;
4683 se->avg.load_sum = se->avg.load_avg * divider;
4684 if (se_weight(se) < se->avg.load_sum)
4685 se->avg.load_sum = div_u64(se->avg.load_sum, se_weight(se));
4687 se->avg.load_sum = 1;
4689 enqueue_load_avg(cfs_rq, se);
4690 cfs_rq->avg.util_avg += se->avg.util_avg;
4691 cfs_rq->avg.util_sum += se->avg.util_sum;
4692 cfs_rq->avg.runnable_avg += se->avg.runnable_avg;
4693 cfs_rq->avg.runnable_sum += se->avg.runnable_sum;
4695 add_tg_cfs_propagate(cfs_rq, se->avg.load_sum);
4697 cfs_rq_util_change(cfs_rq, 0);
4699 trace_pelt_cfs_tp(cfs_rq);
4703 * detach_entity_load_avg - detach this entity from its cfs_rq load avg
4704 * @cfs_rq: cfs_rq to detach from
4705 * @se: sched_entity to detach
4707 * Must call update_cfs_rq_load_avg() before this, since we rely on
4708 * cfs_rq->avg.last_update_time being current.
4710 static void detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
4712 dequeue_load_avg(cfs_rq, se);
4713 sub_positive(&cfs_rq->avg.util_avg, se->avg.util_avg);
4714 sub_positive(&cfs_rq->avg.util_sum, se->avg.util_sum);
4715 /* See update_cfs_rq_load_avg() */
4716 cfs_rq->avg.util_sum = max_t(u32, cfs_rq->avg.util_sum,
4717 cfs_rq->avg.util_avg * PELT_MIN_DIVIDER);
4719 sub_positive(&cfs_rq->avg.runnable_avg, se->avg.runnable_avg);
4720 sub_positive(&cfs_rq->avg.runnable_sum, se->avg.runnable_sum);
4721 /* See update_cfs_rq_load_avg() */
4722 cfs_rq->avg.runnable_sum = max_t(u32, cfs_rq->avg.runnable_sum,
4723 cfs_rq->avg.runnable_avg * PELT_MIN_DIVIDER);
4725 add_tg_cfs_propagate(cfs_rq, -se->avg.load_sum);
4727 cfs_rq_util_change(cfs_rq, 0);
4729 trace_pelt_cfs_tp(cfs_rq);
4733 * Optional action to be done while updating the load average
4735 #define UPDATE_TG 0x1
4736 #define SKIP_AGE_LOAD 0x2
4737 #define DO_ATTACH 0x4
4738 #define DO_DETACH 0x8
4740 /* Update task and its cfs_rq load average */
4741 static inline void update_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
4743 u64 now = cfs_rq_clock_pelt(cfs_rq);
4747 * Track task load average for carrying it to new CPU after migrated, and
4748 * track group sched_entity load average for task_h_load calc in migration
4750 if (se->avg.last_update_time && !(flags & SKIP_AGE_LOAD))
4751 __update_load_avg_se(now, cfs_rq, se);
4753 decayed = update_cfs_rq_load_avg(now, cfs_rq);
4754 decayed |= propagate_entity_load_avg(se);
4756 if (!se->avg.last_update_time && (flags & DO_ATTACH)) {
4759 * DO_ATTACH means we're here from enqueue_entity().
4760 * !last_update_time means we've passed through
4761 * migrate_task_rq_fair() indicating we migrated.
4763 * IOW we're enqueueing a task on a new CPU.
4765 attach_entity_load_avg(cfs_rq, se);
4766 update_tg_load_avg(cfs_rq);
4768 } else if (flags & DO_DETACH) {
4770 * DO_DETACH means we're here from dequeue_entity()
4771 * and we are migrating task out of the CPU.
4773 detach_entity_load_avg(cfs_rq, se);
4774 update_tg_load_avg(cfs_rq);
4775 } else if (decayed) {
4776 cfs_rq_util_change(cfs_rq, 0);
4778 if (flags & UPDATE_TG)
4779 update_tg_load_avg(cfs_rq);
4784 * Synchronize entity load avg of dequeued entity without locking
4787 static void sync_entity_load_avg(struct sched_entity *se)
4789 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4790 u64 last_update_time;
4792 last_update_time = cfs_rq_last_update_time(cfs_rq);
4793 __update_load_avg_blocked_se(last_update_time, se);
4797 * Task first catches up with cfs_rq, and then subtract
4798 * itself from the cfs_rq (task must be off the queue now).
4800 static void remove_entity_load_avg(struct sched_entity *se)
4802 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4803 unsigned long flags;
4806 * tasks cannot exit without having gone through wake_up_new_task() ->
4807 * enqueue_task_fair() which will have added things to the cfs_rq,
4808 * so we can remove unconditionally.
4811 sync_entity_load_avg(se);
4813 raw_spin_lock_irqsave(&cfs_rq->removed.lock, flags);
4814 ++cfs_rq->removed.nr;
4815 cfs_rq->removed.util_avg += se->avg.util_avg;
4816 cfs_rq->removed.load_avg += se->avg.load_avg;
4817 cfs_rq->removed.runnable_avg += se->avg.runnable_avg;
4818 raw_spin_unlock_irqrestore(&cfs_rq->removed.lock, flags);
4821 static inline unsigned long cfs_rq_runnable_avg(struct cfs_rq *cfs_rq)
4823 return cfs_rq->avg.runnable_avg;
4826 static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq)
4828 return cfs_rq->avg.load_avg;
4831 static int newidle_balance(struct rq *this_rq, struct rq_flags *rf);
4833 static inline unsigned long task_util(struct task_struct *p)
4835 return READ_ONCE(p->se.avg.util_avg);
4838 static inline unsigned long task_runnable(struct task_struct *p)
4840 return READ_ONCE(p->se.avg.runnable_avg);
4843 static inline unsigned long _task_util_est(struct task_struct *p)
4845 return READ_ONCE(p->se.avg.util_est) & ~UTIL_AVG_UNCHANGED;
4848 static inline unsigned long task_util_est(struct task_struct *p)
4850 return max(task_util(p), _task_util_est(p));
4853 static inline void util_est_enqueue(struct cfs_rq *cfs_rq,
4854 struct task_struct *p)
4856 unsigned int enqueued;
4858 if (!sched_feat(UTIL_EST))
4861 /* Update root cfs_rq's estimated utilization */
4862 enqueued = cfs_rq->avg.util_est;
4863 enqueued += _task_util_est(p);
4864 WRITE_ONCE(cfs_rq->avg.util_est, enqueued);
4866 trace_sched_util_est_cfs_tp(cfs_rq);
4869 static inline void util_est_dequeue(struct cfs_rq *cfs_rq,
4870 struct task_struct *p)
4872 unsigned int enqueued;
4874 if (!sched_feat(UTIL_EST))
4877 /* Update root cfs_rq's estimated utilization */
4878 enqueued = cfs_rq->avg.util_est;
4879 enqueued -= min_t(unsigned int, enqueued, _task_util_est(p));
4880 WRITE_ONCE(cfs_rq->avg.util_est, enqueued);
4882 trace_sched_util_est_cfs_tp(cfs_rq);
4885 #define UTIL_EST_MARGIN (SCHED_CAPACITY_SCALE / 100)
4887 static inline void util_est_update(struct cfs_rq *cfs_rq,
4888 struct task_struct *p,
4891 unsigned int ewma, dequeued, last_ewma_diff;
4893 if (!sched_feat(UTIL_EST))
4897 * Skip update of task's estimated utilization when the task has not
4898 * yet completed an activation, e.g. being migrated.
4903 /* Get current estimate of utilization */
4904 ewma = READ_ONCE(p->se.avg.util_est);
4907 * If the PELT values haven't changed since enqueue time,
4908 * skip the util_est update.
4910 if (ewma & UTIL_AVG_UNCHANGED)
4913 /* Get utilization at dequeue */
4914 dequeued = task_util(p);
4917 * Reset EWMA on utilization increases, the moving average is used only
4918 * to smooth utilization decreases.
4920 if (ewma <= dequeued) {
4926 * Skip update of task's estimated utilization when its members are
4927 * already ~1% close to its last activation value.
4929 last_ewma_diff = ewma - dequeued;
4930 if (last_ewma_diff < UTIL_EST_MARGIN)
4934 * To avoid overestimation of actual task utilization, skip updates if
4935 * we cannot grant there is idle time in this CPU.
4937 if (dequeued > arch_scale_cpu_capacity(cpu_of(rq_of(cfs_rq))))
4941 * To avoid underestimate of task utilization, skip updates of EWMA if
4942 * we cannot grant that thread got all CPU time it wanted.
4944 if ((dequeued + UTIL_EST_MARGIN) < task_runnable(p))
4949 * Update Task's estimated utilization
4951 * When *p completes an activation we can consolidate another sample
4952 * of the task size. This is done by using this value to update the
4953 * Exponential Weighted Moving Average (EWMA):
4955 * ewma(t) = w * task_util(p) + (1-w) * ewma(t-1)
4956 * = w * task_util(p) + ewma(t-1) - w * ewma(t-1)
4957 * = w * (task_util(p) - ewma(t-1)) + ewma(t-1)
4958 * = w * ( -last_ewma_diff ) + ewma(t-1)
4959 * = w * (-last_ewma_diff + ewma(t-1) / w)
4961 * Where 'w' is the weight of new samples, which is configured to be
4962 * 0.25, thus making w=1/4 ( >>= UTIL_EST_WEIGHT_SHIFT)
4964 ewma <<= UTIL_EST_WEIGHT_SHIFT;
4965 ewma -= last_ewma_diff;
4966 ewma >>= UTIL_EST_WEIGHT_SHIFT;
4968 ewma |= UTIL_AVG_UNCHANGED;
4969 WRITE_ONCE(p->se.avg.util_est, ewma);
4971 trace_sched_util_est_se_tp(&p->se);
4974 static inline int util_fits_cpu(unsigned long util,
4975 unsigned long uclamp_min,
4976 unsigned long uclamp_max,
4979 unsigned long capacity_orig, capacity_orig_thermal;
4980 unsigned long capacity = capacity_of(cpu);
4981 bool fits, uclamp_max_fits;
4984 * Check if the real util fits without any uclamp boost/cap applied.
4986 fits = fits_capacity(util, capacity);
4988 if (!uclamp_is_used())
4992 * We must use arch_scale_cpu_capacity() for comparing against uclamp_min and
4993 * uclamp_max. We only care about capacity pressure (by using
4994 * capacity_of()) for comparing against the real util.
4996 * If a task is boosted to 1024 for example, we don't want a tiny
4997 * pressure to skew the check whether it fits a CPU or not.
4999 * Similarly if a task is capped to arch_scale_cpu_capacity(little_cpu), it
5000 * should fit a little cpu even if there's some pressure.
5002 * Only exception is for thermal pressure since it has a direct impact
5003 * on available OPP of the system.
5005 * We honour it for uclamp_min only as a drop in performance level
5006 * could result in not getting the requested minimum performance level.
5008 * For uclamp_max, we can tolerate a drop in performance level as the
5009 * goal is to cap the task. So it's okay if it's getting less.
5011 capacity_orig = arch_scale_cpu_capacity(cpu);
5012 capacity_orig_thermal = capacity_orig - arch_scale_thermal_pressure(cpu);
5015 * We want to force a task to fit a cpu as implied by uclamp_max.
5016 * But we do have some corner cases to cater for..
5022 * |_ _ _ _ _ _ _ _ _ ___ _ _ _ | _ | _ _ _ _ _ uclamp_max
5025 * | | | | | | | (util somewhere in this region)
5028 * +----------------------------------------
5031 * In the above example if a task is capped to a specific performance
5032 * point, y, then when:
5034 * * util = 80% of x then it does not fit on cpu0 and should migrate
5036 * * util = 80% of y then it is forced to fit on cpu1 to honour
5037 * uclamp_max request.
5039 * which is what we're enforcing here. A task always fits if
5040 * uclamp_max <= capacity_orig. But when uclamp_max > capacity_orig,
5041 * the normal upmigration rules should withhold still.
5043 * Only exception is when we are on max capacity, then we need to be
5044 * careful not to block overutilized state. This is so because:
5046 * 1. There's no concept of capping at max_capacity! We can't go
5047 * beyond this performance level anyway.
5048 * 2. The system is being saturated when we're operating near
5049 * max capacity, it doesn't make sense to block overutilized.
5051 uclamp_max_fits = (capacity_orig == SCHED_CAPACITY_SCALE) && (uclamp_max == SCHED_CAPACITY_SCALE);
5052 uclamp_max_fits = !uclamp_max_fits && (uclamp_max <= capacity_orig);
5053 fits = fits || uclamp_max_fits;
5058 * | ___ (region a, capped, util >= uclamp_max)
5060 * |_ _ _ _ _ _ _ _ _ ___ _ _ _ | _ | _ _ _ _ _ uclamp_max
5062 * | ___ | | | | (region b, uclamp_min <= util <= uclamp_max)
5063 * |_ _ _|_ _|_ _ _ _| _ | _ _ _| _ | _ _ _ _ _ uclamp_min
5065 * | | | | | | | (region c, boosted, util < uclamp_min)
5066 * +----------------------------------------
5069 * a) If util > uclamp_max, then we're capped, we don't care about
5070 * actual fitness value here. We only care if uclamp_max fits
5071 * capacity without taking margin/pressure into account.
5072 * See comment above.
5074 * b) If uclamp_min <= util <= uclamp_max, then the normal
5075 * fits_capacity() rules apply. Except we need to ensure that we
5076 * enforce we remain within uclamp_max, see comment above.
5078 * c) If util < uclamp_min, then we are boosted. Same as (b) but we
5079 * need to take into account the boosted value fits the CPU without
5080 * taking margin/pressure into account.
5082 * Cases (a) and (b) are handled in the 'fits' variable already. We
5083 * just need to consider an extra check for case (c) after ensuring we
5084 * handle the case uclamp_min > uclamp_max.
5086 uclamp_min = min(uclamp_min, uclamp_max);
5087 if (fits && (util < uclamp_min) && (uclamp_min > capacity_orig_thermal))
5093 static inline int task_fits_cpu(struct task_struct *p, int cpu)
5095 unsigned long uclamp_min = uclamp_eff_value(p, UCLAMP_MIN);
5096 unsigned long uclamp_max = uclamp_eff_value(p, UCLAMP_MAX);
5097 unsigned long util = task_util_est(p);
5099 * Return true only if the cpu fully fits the task requirements, which
5100 * include the utilization but also the performance hints.
5102 return (util_fits_cpu(util, uclamp_min, uclamp_max, cpu) > 0);
5105 static inline void update_misfit_status(struct task_struct *p, struct rq *rq)
5107 if (!sched_asym_cpucap_active())
5110 if (!p || p->nr_cpus_allowed == 1) {
5111 rq->misfit_task_load = 0;
5115 if (task_fits_cpu(p, cpu_of(rq))) {
5116 rq->misfit_task_load = 0;
5121 * Make sure that misfit_task_load will not be null even if
5122 * task_h_load() returns 0.
5124 rq->misfit_task_load = max_t(unsigned long, task_h_load(p), 1);
5127 #else /* CONFIG_SMP */
5129 static inline bool cfs_rq_is_decayed(struct cfs_rq *cfs_rq)
5131 return !cfs_rq->nr_running;
5134 #define UPDATE_TG 0x0
5135 #define SKIP_AGE_LOAD 0x0
5136 #define DO_ATTACH 0x0
5137 #define DO_DETACH 0x0
5139 static inline void update_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int not_used1)
5141 cfs_rq_util_change(cfs_rq, 0);
5144 static inline void remove_entity_load_avg(struct sched_entity *se) {}
5147 attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
5149 detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
5151 static inline int newidle_balance(struct rq *rq, struct rq_flags *rf)
5157 util_est_enqueue(struct cfs_rq *cfs_rq, struct task_struct *p) {}
5160 util_est_dequeue(struct cfs_rq *cfs_rq, struct task_struct *p) {}
5163 util_est_update(struct cfs_rq *cfs_rq, struct task_struct *p,
5165 static inline void update_misfit_status(struct task_struct *p, struct rq *rq) {}
5167 #endif /* CONFIG_SMP */
5170 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
5172 u64 vslice, vruntime = avg_vruntime(cfs_rq);
5175 se->slice = sysctl_sched_base_slice;
5176 vslice = calc_delta_fair(se->slice, se);
5179 * Due to how V is constructed as the weighted average of entities,
5180 * adding tasks with positive lag, or removing tasks with negative lag
5181 * will move 'time' backwards, this can screw around with the lag of
5184 * EEVDF: placement strategy #1 / #2
5186 if (sched_feat(PLACE_LAG) && cfs_rq->nr_running) {
5187 struct sched_entity *curr = cfs_rq->curr;
5193 * If we want to place a task and preserve lag, we have to
5194 * consider the effect of the new entity on the weighted
5195 * average and compensate for this, otherwise lag can quickly
5198 * Lag is defined as:
5200 * lag_i = S - s_i = w_i * (V - v_i)
5202 * To avoid the 'w_i' term all over the place, we only track
5205 * vl_i = V - v_i <=> v_i = V - vl_i
5207 * And we take V to be the weighted average of all v:
5209 * V = (\Sum w_j*v_j) / W
5211 * Where W is: \Sum w_j
5213 * Then, the weighted average after adding an entity with lag
5216 * V' = (\Sum w_j*v_j + w_i*v_i) / (W + w_i)
5217 * = (W*V + w_i*(V - vl_i)) / (W + w_i)
5218 * = (W*V + w_i*V - w_i*vl_i) / (W + w_i)
5219 * = (V*(W + w_i) - w_i*l) / (W + w_i)
5220 * = V - w_i*vl_i / (W + w_i)
5222 * And the actual lag after adding an entity with vl_i is:
5225 * = V - w_i*vl_i / (W + w_i) - (V - vl_i)
5226 * = vl_i - w_i*vl_i / (W + w_i)
5228 * Which is strictly less than vl_i. So in order to preserve lag
5229 * we should inflate the lag before placement such that the
5230 * effective lag after placement comes out right.
5232 * As such, invert the above relation for vl'_i to get the vl_i
5233 * we need to use such that the lag after placement is the lag
5234 * we computed before dequeue.
5236 * vl'_i = vl_i - w_i*vl_i / (W + w_i)
5237 * = ((W + w_i)*vl_i - w_i*vl_i) / (W + w_i)
5239 * (W + w_i)*vl'_i = (W + w_i)*vl_i - w_i*vl_i
5242 * vl_i = (W + w_i)*vl'_i / W
5244 load = cfs_rq->avg_load;
5245 if (curr && curr->on_rq)
5246 load += scale_load_down(curr->load.weight);
5248 lag *= load + scale_load_down(se->load.weight);
5249 if (WARN_ON_ONCE(!load))
5251 lag = div_s64(lag, load);
5254 se->vruntime = vruntime - lag;
5257 * When joining the competition; the exisiting tasks will be,
5258 * on average, halfway through their slice, as such start tasks
5259 * off with half a slice to ease into the competition.
5261 if (sched_feat(PLACE_DEADLINE_INITIAL) && (flags & ENQUEUE_INITIAL))
5265 * EEVDF: vd_i = ve_i + r_i/w_i
5267 se->deadline = se->vruntime + vslice;
5270 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
5271 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq);
5273 static inline bool cfs_bandwidth_used(void);
5276 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
5278 bool curr = cfs_rq->curr == se;
5281 * If we're the current task, we must renormalise before calling
5285 place_entity(cfs_rq, se, flags);
5287 update_curr(cfs_rq);
5290 * When enqueuing a sched_entity, we must:
5291 * - Update loads to have both entity and cfs_rq synced with now.
5292 * - For group_entity, update its runnable_weight to reflect the new
5293 * h_nr_running of its group cfs_rq.
5294 * - For group_entity, update its weight to reflect the new share of
5296 * - Add its new weight to cfs_rq->load.weight
5298 update_load_avg(cfs_rq, se, UPDATE_TG | DO_ATTACH);
5299 se_update_runnable(se);
5301 * XXX update_load_avg() above will have attached us to the pelt sum;
5302 * but update_cfs_group() here will re-adjust the weight and have to
5303 * undo/redo all that. Seems wasteful.
5305 update_cfs_group(se);
5308 * XXX now that the entity has been re-weighted, and it's lag adjusted,
5309 * we can place the entity.
5312 place_entity(cfs_rq, se, flags);
5314 account_entity_enqueue(cfs_rq, se);
5316 /* Entity has migrated, no longer consider this task hot */
5317 if (flags & ENQUEUE_MIGRATED)
5320 check_schedstat_required();
5321 update_stats_enqueue_fair(cfs_rq, se, flags);
5323 __enqueue_entity(cfs_rq, se);
5326 if (cfs_rq->nr_running == 1) {
5327 check_enqueue_throttle(cfs_rq);
5328 if (!throttled_hierarchy(cfs_rq)) {
5329 list_add_leaf_cfs_rq(cfs_rq);
5331 #ifdef CONFIG_CFS_BANDWIDTH
5332 struct rq *rq = rq_of(cfs_rq);
5334 if (cfs_rq_throttled(cfs_rq) && !cfs_rq->throttled_clock)
5335 cfs_rq->throttled_clock = rq_clock(rq);
5336 if (!cfs_rq->throttled_clock_self)
5337 cfs_rq->throttled_clock_self = rq_clock(rq);
5343 static void __clear_buddies_next(struct sched_entity *se)
5345 for_each_sched_entity(se) {
5346 struct cfs_rq *cfs_rq = cfs_rq_of(se);
5347 if (cfs_rq->next != se)
5350 cfs_rq->next = NULL;
5354 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
5356 if (cfs_rq->next == se)
5357 __clear_buddies_next(se);
5360 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
5363 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
5365 int action = UPDATE_TG;
5367 if (entity_is_task(se) && task_on_rq_migrating(task_of(se)))
5368 action |= DO_DETACH;
5371 * Update run-time statistics of the 'current'.
5373 update_curr(cfs_rq);
5376 * When dequeuing a sched_entity, we must:
5377 * - Update loads to have both entity and cfs_rq synced with now.
5378 * - For group_entity, update its runnable_weight to reflect the new
5379 * h_nr_running of its group cfs_rq.
5380 * - Subtract its previous weight from cfs_rq->load.weight.
5381 * - For group entity, update its weight to reflect the new share
5382 * of its group cfs_rq.
5384 update_load_avg(cfs_rq, se, action);
5385 se_update_runnable(se);
5387 update_stats_dequeue_fair(cfs_rq, se, flags);
5389 clear_buddies(cfs_rq, se);
5391 update_entity_lag(cfs_rq, se);
5392 if (se != cfs_rq->curr)
5393 __dequeue_entity(cfs_rq, se);
5395 account_entity_dequeue(cfs_rq, se);
5397 /* return excess runtime on last dequeue */
5398 return_cfs_rq_runtime(cfs_rq);
5400 update_cfs_group(se);
5403 * Now advance min_vruntime if @se was the entity holding it back,
5404 * except when: DEQUEUE_SAVE && !DEQUEUE_MOVE, in this case we'll be
5405 * put back on, and if we advance min_vruntime, we'll be placed back
5406 * further than we started -- ie. we'll be penalized.
5408 if ((flags & (DEQUEUE_SAVE | DEQUEUE_MOVE)) != DEQUEUE_SAVE)
5409 update_min_vruntime(cfs_rq);
5411 if (cfs_rq->nr_running == 0)
5412 update_idle_cfs_rq_clock_pelt(cfs_rq);
5416 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
5418 clear_buddies(cfs_rq, se);
5420 /* 'current' is not kept within the tree. */
5423 * Any task has to be enqueued before it get to execute on
5424 * a CPU. So account for the time it spent waiting on the
5427 update_stats_wait_end_fair(cfs_rq, se);
5428 __dequeue_entity(cfs_rq, se);
5429 update_load_avg(cfs_rq, se, UPDATE_TG);
5431 * HACK, stash a copy of deadline at the point of pick in vlag,
5432 * which isn't used until dequeue.
5434 se->vlag = se->deadline;
5437 update_stats_curr_start(cfs_rq, se);
5441 * Track our maximum slice length, if the CPU's load is at
5442 * least twice that of our own weight (i.e. dont track it
5443 * when there are only lesser-weight tasks around):
5445 if (schedstat_enabled() &&
5446 rq_of(cfs_rq)->cfs.load.weight >= 2*se->load.weight) {
5447 struct sched_statistics *stats;
5449 stats = __schedstats_from_se(se);
5450 __schedstat_set(stats->slice_max,
5451 max((u64)stats->slice_max,
5452 se->sum_exec_runtime - se->prev_sum_exec_runtime));
5455 se->prev_sum_exec_runtime = se->sum_exec_runtime;
5459 * Pick the next process, keeping these things in mind, in this order:
5460 * 1) keep things fair between processes/task groups
5461 * 2) pick the "next" process, since someone really wants that to run
5462 * 3) pick the "last" process, for cache locality
5463 * 4) do not run the "skip" process, if something else is available
5465 static struct sched_entity *
5466 pick_next_entity(struct cfs_rq *cfs_rq)
5469 * Enabling NEXT_BUDDY will affect latency but not fairness.
5471 if (sched_feat(NEXT_BUDDY) &&
5472 cfs_rq->next && entity_eligible(cfs_rq, cfs_rq->next))
5473 return cfs_rq->next;
5475 return pick_eevdf(cfs_rq);
5478 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
5480 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
5483 * If still on the runqueue then deactivate_task()
5484 * was not called and update_curr() has to be done:
5487 update_curr(cfs_rq);
5489 /* throttle cfs_rqs exceeding runtime */
5490 check_cfs_rq_runtime(cfs_rq);
5493 update_stats_wait_start_fair(cfs_rq, prev);
5494 /* Put 'current' back into the tree. */
5495 __enqueue_entity(cfs_rq, prev);
5496 /* in !on_rq case, update occurred at dequeue */
5497 update_load_avg(cfs_rq, prev, 0);
5499 cfs_rq->curr = NULL;
5503 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
5506 * Update run-time statistics of the 'current'.
5508 update_curr(cfs_rq);
5511 * Ensure that runnable average is periodically updated.
5513 update_load_avg(cfs_rq, curr, UPDATE_TG);
5514 update_cfs_group(curr);
5516 #ifdef CONFIG_SCHED_HRTICK
5518 * queued ticks are scheduled to match the slice, so don't bother
5519 * validating it and just reschedule.
5522 resched_curr(rq_of(cfs_rq));
5526 * don't let the period tick interfere with the hrtick preemption
5528 if (!sched_feat(DOUBLE_TICK) &&
5529 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
5535 /**************************************************
5536 * CFS bandwidth control machinery
5539 #ifdef CONFIG_CFS_BANDWIDTH
5541 #ifdef CONFIG_JUMP_LABEL
5542 static struct static_key __cfs_bandwidth_used;
5544 static inline bool cfs_bandwidth_used(void)
5546 return static_key_false(&__cfs_bandwidth_used);
5549 void cfs_bandwidth_usage_inc(void)
5551 static_key_slow_inc_cpuslocked(&__cfs_bandwidth_used);
5554 void cfs_bandwidth_usage_dec(void)
5556 static_key_slow_dec_cpuslocked(&__cfs_bandwidth_used);
5558 #else /* CONFIG_JUMP_LABEL */
5559 static bool cfs_bandwidth_used(void)
5564 void cfs_bandwidth_usage_inc(void) {}
5565 void cfs_bandwidth_usage_dec(void) {}
5566 #endif /* CONFIG_JUMP_LABEL */
5569 * default period for cfs group bandwidth.
5570 * default: 0.1s, units: nanoseconds
5572 static inline u64 default_cfs_period(void)
5574 return 100000000ULL;
5577 static inline u64 sched_cfs_bandwidth_slice(void)
5579 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
5583 * Replenish runtime according to assigned quota. We use sched_clock_cpu
5584 * directly instead of rq->clock to avoid adding additional synchronization
5587 * requires cfs_b->lock
5589 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
5593 if (unlikely(cfs_b->quota == RUNTIME_INF))
5596 cfs_b->runtime += cfs_b->quota;
5597 runtime = cfs_b->runtime_snap - cfs_b->runtime;
5599 cfs_b->burst_time += runtime;
5603 cfs_b->runtime = min(cfs_b->runtime, cfs_b->quota + cfs_b->burst);
5604 cfs_b->runtime_snap = cfs_b->runtime;
5607 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
5609 return &tg->cfs_bandwidth;
5612 /* returns 0 on failure to allocate runtime */
5613 static int __assign_cfs_rq_runtime(struct cfs_bandwidth *cfs_b,
5614 struct cfs_rq *cfs_rq, u64 target_runtime)
5616 u64 min_amount, amount = 0;
5618 lockdep_assert_held(&cfs_b->lock);
5620 /* note: this is a positive sum as runtime_remaining <= 0 */
5621 min_amount = target_runtime - cfs_rq->runtime_remaining;
5623 if (cfs_b->quota == RUNTIME_INF)
5624 amount = min_amount;
5626 start_cfs_bandwidth(cfs_b);
5628 if (cfs_b->runtime > 0) {
5629 amount = min(cfs_b->runtime, min_amount);
5630 cfs_b->runtime -= amount;
5635 cfs_rq->runtime_remaining += amount;
5637 return cfs_rq->runtime_remaining > 0;
5640 /* returns 0 on failure to allocate runtime */
5641 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
5643 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
5646 raw_spin_lock(&cfs_b->lock);
5647 ret = __assign_cfs_rq_runtime(cfs_b, cfs_rq, sched_cfs_bandwidth_slice());
5648 raw_spin_unlock(&cfs_b->lock);
5653 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
5655 /* dock delta_exec before expiring quota (as it could span periods) */
5656 cfs_rq->runtime_remaining -= delta_exec;
5658 if (likely(cfs_rq->runtime_remaining > 0))
5661 if (cfs_rq->throttled)
5664 * if we're unable to extend our runtime we resched so that the active
5665 * hierarchy can be throttled
5667 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
5668 resched_curr(rq_of(cfs_rq));
5671 static __always_inline
5672 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
5674 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
5677 __account_cfs_rq_runtime(cfs_rq, delta_exec);
5680 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
5682 return cfs_bandwidth_used() && cfs_rq->throttled;
5685 /* check whether cfs_rq, or any parent, is throttled */
5686 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
5688 return cfs_bandwidth_used() && cfs_rq->throttle_count;
5692 * Ensure that neither of the group entities corresponding to src_cpu or
5693 * dest_cpu are members of a throttled hierarchy when performing group
5694 * load-balance operations.
5696 static inline int throttled_lb_pair(struct task_group *tg,
5697 int src_cpu, int dest_cpu)
5699 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
5701 src_cfs_rq = tg->cfs_rq[src_cpu];
5702 dest_cfs_rq = tg->cfs_rq[dest_cpu];
5704 return throttled_hierarchy(src_cfs_rq) ||
5705 throttled_hierarchy(dest_cfs_rq);
5708 static int tg_unthrottle_up(struct task_group *tg, void *data)
5710 struct rq *rq = data;
5711 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
5713 cfs_rq->throttle_count--;
5714 if (!cfs_rq->throttle_count) {
5715 cfs_rq->throttled_clock_pelt_time += rq_clock_pelt(rq) -
5716 cfs_rq->throttled_clock_pelt;
5718 /* Add cfs_rq with load or one or more already running entities to the list */
5719 if (!cfs_rq_is_decayed(cfs_rq))
5720 list_add_leaf_cfs_rq(cfs_rq);
5722 if (cfs_rq->throttled_clock_self) {
5723 u64 delta = rq_clock(rq) - cfs_rq->throttled_clock_self;
5725 cfs_rq->throttled_clock_self = 0;
5727 if (SCHED_WARN_ON((s64)delta < 0))
5730 cfs_rq->throttled_clock_self_time += delta;
5737 static int tg_throttle_down(struct task_group *tg, void *data)
5739 struct rq *rq = data;
5740 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
5742 /* group is entering throttled state, stop time */
5743 if (!cfs_rq->throttle_count) {
5744 cfs_rq->throttled_clock_pelt = rq_clock_pelt(rq);
5745 list_del_leaf_cfs_rq(cfs_rq);
5747 SCHED_WARN_ON(cfs_rq->throttled_clock_self);
5748 if (cfs_rq->nr_running)
5749 cfs_rq->throttled_clock_self = rq_clock(rq);
5751 cfs_rq->throttle_count++;
5756 static bool throttle_cfs_rq(struct cfs_rq *cfs_rq)
5758 struct rq *rq = rq_of(cfs_rq);
5759 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
5760 struct sched_entity *se;
5761 long task_delta, idle_task_delta, dequeue = 1;
5763 raw_spin_lock(&cfs_b->lock);
5764 /* This will start the period timer if necessary */
5765 if (__assign_cfs_rq_runtime(cfs_b, cfs_rq, 1)) {
5767 * We have raced with bandwidth becoming available, and if we
5768 * actually throttled the timer might not unthrottle us for an
5769 * entire period. We additionally needed to make sure that any
5770 * subsequent check_cfs_rq_runtime calls agree not to throttle
5771 * us, as we may commit to do cfs put_prev+pick_next, so we ask
5772 * for 1ns of runtime rather than just check cfs_b.
5776 list_add_tail_rcu(&cfs_rq->throttled_list,
5777 &cfs_b->throttled_cfs_rq);
5779 raw_spin_unlock(&cfs_b->lock);
5782 return false; /* Throttle no longer required. */
5784 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
5786 /* freeze hierarchy runnable averages while throttled */
5788 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
5791 task_delta = cfs_rq->h_nr_running;
5792 idle_task_delta = cfs_rq->idle_h_nr_running;
5793 for_each_sched_entity(se) {
5794 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
5795 /* throttled entity or throttle-on-deactivate */
5799 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
5801 if (cfs_rq_is_idle(group_cfs_rq(se)))
5802 idle_task_delta = cfs_rq->h_nr_running;
5804 qcfs_rq->h_nr_running -= task_delta;
5805 qcfs_rq->idle_h_nr_running -= idle_task_delta;
5807 if (qcfs_rq->load.weight) {
5808 /* Avoid re-evaluating load for this entity: */
5809 se = parent_entity(se);
5814 for_each_sched_entity(se) {
5815 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
5816 /* throttled entity or throttle-on-deactivate */
5820 update_load_avg(qcfs_rq, se, 0);
5821 se_update_runnable(se);
5823 if (cfs_rq_is_idle(group_cfs_rq(se)))
5824 idle_task_delta = cfs_rq->h_nr_running;
5826 qcfs_rq->h_nr_running -= task_delta;
5827 qcfs_rq->idle_h_nr_running -= idle_task_delta;
5830 /* At this point se is NULL and we are at root level*/
5831 sub_nr_running(rq, task_delta);
5835 * Note: distribution will already see us throttled via the
5836 * throttled-list. rq->lock protects completion.
5838 cfs_rq->throttled = 1;
5839 SCHED_WARN_ON(cfs_rq->throttled_clock);
5840 if (cfs_rq->nr_running)
5841 cfs_rq->throttled_clock = rq_clock(rq);
5845 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
5847 struct rq *rq = rq_of(cfs_rq);
5848 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
5849 struct sched_entity *se;
5850 long task_delta, idle_task_delta;
5852 se = cfs_rq->tg->se[cpu_of(rq)];
5854 cfs_rq->throttled = 0;
5856 update_rq_clock(rq);
5858 raw_spin_lock(&cfs_b->lock);
5859 if (cfs_rq->throttled_clock) {
5860 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
5861 cfs_rq->throttled_clock = 0;
5863 list_del_rcu(&cfs_rq->throttled_list);
5864 raw_spin_unlock(&cfs_b->lock);
5866 /* update hierarchical throttle state */
5867 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
5869 if (!cfs_rq->load.weight) {
5870 if (!cfs_rq->on_list)
5873 * Nothing to run but something to decay (on_list)?
5874 * Complete the branch.
5876 for_each_sched_entity(se) {
5877 if (list_add_leaf_cfs_rq(cfs_rq_of(se)))
5880 goto unthrottle_throttle;
5883 task_delta = cfs_rq->h_nr_running;
5884 idle_task_delta = cfs_rq->idle_h_nr_running;
5885 for_each_sched_entity(se) {
5886 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
5890 enqueue_entity(qcfs_rq, se, ENQUEUE_WAKEUP);
5892 if (cfs_rq_is_idle(group_cfs_rq(se)))
5893 idle_task_delta = cfs_rq->h_nr_running;
5895 qcfs_rq->h_nr_running += task_delta;
5896 qcfs_rq->idle_h_nr_running += idle_task_delta;
5898 /* end evaluation on encountering a throttled cfs_rq */
5899 if (cfs_rq_throttled(qcfs_rq))
5900 goto unthrottle_throttle;
5903 for_each_sched_entity(se) {
5904 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
5906 update_load_avg(qcfs_rq, se, UPDATE_TG);
5907 se_update_runnable(se);
5909 if (cfs_rq_is_idle(group_cfs_rq(se)))
5910 idle_task_delta = cfs_rq->h_nr_running;
5912 qcfs_rq->h_nr_running += task_delta;
5913 qcfs_rq->idle_h_nr_running += idle_task_delta;
5915 /* end evaluation on encountering a throttled cfs_rq */
5916 if (cfs_rq_throttled(qcfs_rq))
5917 goto unthrottle_throttle;
5920 /* At this point se is NULL and we are at root level*/
5921 add_nr_running(rq, task_delta);
5923 unthrottle_throttle:
5924 assert_list_leaf_cfs_rq(rq);
5926 /* Determine whether we need to wake up potentially idle CPU: */
5927 if (rq->curr == rq->idle && rq->cfs.nr_running)
5932 static void __cfsb_csd_unthrottle(void *arg)
5934 struct cfs_rq *cursor, *tmp;
5935 struct rq *rq = arg;
5941 * Iterating over the list can trigger several call to
5942 * update_rq_clock() in unthrottle_cfs_rq().
5943 * Do it once and skip the potential next ones.
5945 update_rq_clock(rq);
5946 rq_clock_start_loop_update(rq);
5949 * Since we hold rq lock we're safe from concurrent manipulation of
5950 * the CSD list. However, this RCU critical section annotates the
5951 * fact that we pair with sched_free_group_rcu(), so that we cannot
5952 * race with group being freed in the window between removing it
5953 * from the list and advancing to the next entry in the list.
5957 list_for_each_entry_safe(cursor, tmp, &rq->cfsb_csd_list,
5958 throttled_csd_list) {
5959 list_del_init(&cursor->throttled_csd_list);
5961 if (cfs_rq_throttled(cursor))
5962 unthrottle_cfs_rq(cursor);
5967 rq_clock_stop_loop_update(rq);
5971 static inline void __unthrottle_cfs_rq_async(struct cfs_rq *cfs_rq)
5973 struct rq *rq = rq_of(cfs_rq);
5976 if (rq == this_rq()) {
5977 unthrottle_cfs_rq(cfs_rq);
5981 /* Already enqueued */
5982 if (SCHED_WARN_ON(!list_empty(&cfs_rq->throttled_csd_list)))
5985 first = list_empty(&rq->cfsb_csd_list);
5986 list_add_tail(&cfs_rq->throttled_csd_list, &rq->cfsb_csd_list);
5988 smp_call_function_single_async(cpu_of(rq), &rq->cfsb_csd);
5991 static inline void __unthrottle_cfs_rq_async(struct cfs_rq *cfs_rq)
5993 unthrottle_cfs_rq(cfs_rq);
5997 static void unthrottle_cfs_rq_async(struct cfs_rq *cfs_rq)
5999 lockdep_assert_rq_held(rq_of(cfs_rq));
6001 if (SCHED_WARN_ON(!cfs_rq_throttled(cfs_rq) ||
6002 cfs_rq->runtime_remaining <= 0))
6005 __unthrottle_cfs_rq_async(cfs_rq);
6008 static bool distribute_cfs_runtime(struct cfs_bandwidth *cfs_b)
6010 int this_cpu = smp_processor_id();
6011 u64 runtime, remaining = 1;
6012 bool throttled = false;
6013 struct cfs_rq *cfs_rq, *tmp;
6016 LIST_HEAD(local_unthrottle);
6019 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
6028 rq_lock_irqsave(rq, &rf);
6029 if (!cfs_rq_throttled(cfs_rq))
6032 /* Already queued for async unthrottle */
6033 if (!list_empty(&cfs_rq->throttled_csd_list))
6036 /* By the above checks, this should never be true */
6037 SCHED_WARN_ON(cfs_rq->runtime_remaining > 0);
6039 raw_spin_lock(&cfs_b->lock);
6040 runtime = -cfs_rq->runtime_remaining + 1;
6041 if (runtime > cfs_b->runtime)
6042 runtime = cfs_b->runtime;
6043 cfs_b->runtime -= runtime;
6044 remaining = cfs_b->runtime;
6045 raw_spin_unlock(&cfs_b->lock);
6047 cfs_rq->runtime_remaining += runtime;
6049 /* we check whether we're throttled above */
6050 if (cfs_rq->runtime_remaining > 0) {
6051 if (cpu_of(rq) != this_cpu) {
6052 unthrottle_cfs_rq_async(cfs_rq);
6055 * We currently only expect to be unthrottling
6056 * a single cfs_rq locally.
6058 SCHED_WARN_ON(!list_empty(&local_unthrottle));
6059 list_add_tail(&cfs_rq->throttled_csd_list,
6067 rq_unlock_irqrestore(rq, &rf);
6070 list_for_each_entry_safe(cfs_rq, tmp, &local_unthrottle,
6071 throttled_csd_list) {
6072 struct rq *rq = rq_of(cfs_rq);
6074 rq_lock_irqsave(rq, &rf);
6076 list_del_init(&cfs_rq->throttled_csd_list);
6078 if (cfs_rq_throttled(cfs_rq))
6079 unthrottle_cfs_rq(cfs_rq);
6081 rq_unlock_irqrestore(rq, &rf);
6083 SCHED_WARN_ON(!list_empty(&local_unthrottle));
6091 * Responsible for refilling a task_group's bandwidth and unthrottling its
6092 * cfs_rqs as appropriate. If there has been no activity within the last
6093 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
6094 * used to track this state.
6096 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun, unsigned long flags)
6100 /* no need to continue the timer with no bandwidth constraint */
6101 if (cfs_b->quota == RUNTIME_INF)
6102 goto out_deactivate;
6104 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
6105 cfs_b->nr_periods += overrun;
6107 /* Refill extra burst quota even if cfs_b->idle */
6108 __refill_cfs_bandwidth_runtime(cfs_b);
6111 * idle depends on !throttled (for the case of a large deficit), and if
6112 * we're going inactive then everything else can be deferred
6114 if (cfs_b->idle && !throttled)
6115 goto out_deactivate;
6118 /* mark as potentially idle for the upcoming period */
6123 /* account preceding periods in which throttling occurred */
6124 cfs_b->nr_throttled += overrun;
6127 * This check is repeated as we release cfs_b->lock while we unthrottle.
6129 while (throttled && cfs_b->runtime > 0) {
6130 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
6131 /* we can't nest cfs_b->lock while distributing bandwidth */
6132 throttled = distribute_cfs_runtime(cfs_b);
6133 raw_spin_lock_irqsave(&cfs_b->lock, flags);
6137 * While we are ensured activity in the period following an
6138 * unthrottle, this also covers the case in which the new bandwidth is
6139 * insufficient to cover the existing bandwidth deficit. (Forcing the
6140 * timer to remain active while there are any throttled entities.)
6150 /* a cfs_rq won't donate quota below this amount */
6151 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
6152 /* minimum remaining period time to redistribute slack quota */
6153 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
6154 /* how long we wait to gather additional slack before distributing */
6155 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
6158 * Are we near the end of the current quota period?
6160 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
6161 * hrtimer base being cleared by hrtimer_start. In the case of
6162 * migrate_hrtimers, base is never cleared, so we are fine.
6164 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
6166 struct hrtimer *refresh_timer = &cfs_b->period_timer;
6169 /* if the call-back is running a quota refresh is already occurring */
6170 if (hrtimer_callback_running(refresh_timer))
6173 /* is a quota refresh about to occur? */
6174 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
6175 if (remaining < (s64)min_expire)
6181 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
6183 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
6185 /* if there's a quota refresh soon don't bother with slack */
6186 if (runtime_refresh_within(cfs_b, min_left))
6189 /* don't push forwards an existing deferred unthrottle */
6190 if (cfs_b->slack_started)
6192 cfs_b->slack_started = true;
6194 hrtimer_start(&cfs_b->slack_timer,
6195 ns_to_ktime(cfs_bandwidth_slack_period),
6199 /* we know any runtime found here is valid as update_curr() precedes return */
6200 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
6202 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
6203 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
6205 if (slack_runtime <= 0)
6208 raw_spin_lock(&cfs_b->lock);
6209 if (cfs_b->quota != RUNTIME_INF) {
6210 cfs_b->runtime += slack_runtime;
6212 /* we are under rq->lock, defer unthrottling using a timer */
6213 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
6214 !list_empty(&cfs_b->throttled_cfs_rq))
6215 start_cfs_slack_bandwidth(cfs_b);
6217 raw_spin_unlock(&cfs_b->lock);
6219 /* even if it's not valid for return we don't want to try again */
6220 cfs_rq->runtime_remaining -= slack_runtime;
6223 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
6225 if (!cfs_bandwidth_used())
6228 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
6231 __return_cfs_rq_runtime(cfs_rq);
6235 * This is done with a timer (instead of inline with bandwidth return) since
6236 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
6238 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
6240 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
6241 unsigned long flags;
6243 /* confirm we're still not at a refresh boundary */
6244 raw_spin_lock_irqsave(&cfs_b->lock, flags);
6245 cfs_b->slack_started = false;
6247 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
6248 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
6252 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice)
6253 runtime = cfs_b->runtime;
6255 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
6260 distribute_cfs_runtime(cfs_b);
6264 * When a group wakes up we want to make sure that its quota is not already
6265 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
6266 * runtime as update_curr() throttling can not trigger until it's on-rq.
6268 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
6270 if (!cfs_bandwidth_used())
6273 /* an active group must be handled by the update_curr()->put() path */
6274 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
6277 /* ensure the group is not already throttled */
6278 if (cfs_rq_throttled(cfs_rq))
6281 /* update runtime allocation */
6282 account_cfs_rq_runtime(cfs_rq, 0);
6283 if (cfs_rq->runtime_remaining <= 0)
6284 throttle_cfs_rq(cfs_rq);
6287 static void sync_throttle(struct task_group *tg, int cpu)
6289 struct cfs_rq *pcfs_rq, *cfs_rq;
6291 if (!cfs_bandwidth_used())
6297 cfs_rq = tg->cfs_rq[cpu];
6298 pcfs_rq = tg->parent->cfs_rq[cpu];
6300 cfs_rq->throttle_count = pcfs_rq->throttle_count;
6301 cfs_rq->throttled_clock_pelt = rq_clock_pelt(cpu_rq(cpu));
6304 /* conditionally throttle active cfs_rq's from put_prev_entity() */
6305 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
6307 if (!cfs_bandwidth_used())
6310 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
6314 * it's possible for a throttled entity to be forced into a running
6315 * state (e.g. set_curr_task), in this case we're finished.
6317 if (cfs_rq_throttled(cfs_rq))
6320 return throttle_cfs_rq(cfs_rq);
6323 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
6325 struct cfs_bandwidth *cfs_b =
6326 container_of(timer, struct cfs_bandwidth, slack_timer);
6328 do_sched_cfs_slack_timer(cfs_b);
6330 return HRTIMER_NORESTART;
6333 extern const u64 max_cfs_quota_period;
6335 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
6337 struct cfs_bandwidth *cfs_b =
6338 container_of(timer, struct cfs_bandwidth, period_timer);
6339 unsigned long flags;
6344 raw_spin_lock_irqsave(&cfs_b->lock, flags);
6346 overrun = hrtimer_forward_now(timer, cfs_b->period);
6350 idle = do_sched_cfs_period_timer(cfs_b, overrun, flags);
6353 u64 new, old = ktime_to_ns(cfs_b->period);
6356 * Grow period by a factor of 2 to avoid losing precision.
6357 * Precision loss in the quota/period ratio can cause __cfs_schedulable
6361 if (new < max_cfs_quota_period) {
6362 cfs_b->period = ns_to_ktime(new);
6366 pr_warn_ratelimited(
6367 "cfs_period_timer[cpu%d]: period too short, scaling up (new cfs_period_us = %lld, cfs_quota_us = %lld)\n",
6369 div_u64(new, NSEC_PER_USEC),
6370 div_u64(cfs_b->quota, NSEC_PER_USEC));
6372 pr_warn_ratelimited(
6373 "cfs_period_timer[cpu%d]: period too short, but cannot scale up without losing precision (cfs_period_us = %lld, cfs_quota_us = %lld)\n",
6375 div_u64(old, NSEC_PER_USEC),
6376 div_u64(cfs_b->quota, NSEC_PER_USEC));
6379 /* reset count so we don't come right back in here */
6384 cfs_b->period_active = 0;
6385 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
6387 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
6390 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b, struct cfs_bandwidth *parent)
6392 raw_spin_lock_init(&cfs_b->lock);
6394 cfs_b->quota = RUNTIME_INF;
6395 cfs_b->period = ns_to_ktime(default_cfs_period());
6397 cfs_b->hierarchical_quota = parent ? parent->hierarchical_quota : RUNTIME_INF;
6399 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
6400 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED);
6401 cfs_b->period_timer.function = sched_cfs_period_timer;
6403 /* Add a random offset so that timers interleave */
6404 hrtimer_set_expires(&cfs_b->period_timer,
6405 get_random_u32_below(cfs_b->period));
6406 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
6407 cfs_b->slack_timer.function = sched_cfs_slack_timer;
6408 cfs_b->slack_started = false;
6411 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
6413 cfs_rq->runtime_enabled = 0;
6414 INIT_LIST_HEAD(&cfs_rq->throttled_list);
6415 INIT_LIST_HEAD(&cfs_rq->throttled_csd_list);
6418 void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
6420 lockdep_assert_held(&cfs_b->lock);
6422 if (cfs_b->period_active)
6425 cfs_b->period_active = 1;
6426 hrtimer_forward_now(&cfs_b->period_timer, cfs_b->period);
6427 hrtimer_start_expires(&cfs_b->period_timer, HRTIMER_MODE_ABS_PINNED);
6430 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
6432 int __maybe_unused i;
6434 /* init_cfs_bandwidth() was not called */
6435 if (!cfs_b->throttled_cfs_rq.next)
6438 hrtimer_cancel(&cfs_b->period_timer);
6439 hrtimer_cancel(&cfs_b->slack_timer);
6442 * It is possible that we still have some cfs_rq's pending on a CSD
6443 * list, though this race is very rare. In order for this to occur, we
6444 * must have raced with the last task leaving the group while there
6445 * exist throttled cfs_rq(s), and the period_timer must have queued the
6446 * CSD item but the remote cpu has not yet processed it. To handle this,
6447 * we can simply flush all pending CSD work inline here. We're
6448 * guaranteed at this point that no additional cfs_rq of this group can
6452 for_each_possible_cpu(i) {
6453 struct rq *rq = cpu_rq(i);
6454 unsigned long flags;
6456 if (list_empty(&rq->cfsb_csd_list))
6459 local_irq_save(flags);
6460 __cfsb_csd_unthrottle(rq);
6461 local_irq_restore(flags);
6467 * Both these CPU hotplug callbacks race against unregister_fair_sched_group()
6469 * The race is harmless, since modifying bandwidth settings of unhooked group
6470 * bits doesn't do much.
6473 /* cpu online callback */
6474 static void __maybe_unused update_runtime_enabled(struct rq *rq)
6476 struct task_group *tg;
6478 lockdep_assert_rq_held(rq);
6481 list_for_each_entry_rcu(tg, &task_groups, list) {
6482 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
6483 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
6485 raw_spin_lock(&cfs_b->lock);
6486 cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF;
6487 raw_spin_unlock(&cfs_b->lock);
6492 /* cpu offline callback */
6493 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
6495 struct task_group *tg;
6497 lockdep_assert_rq_held(rq);
6500 * The rq clock has already been updated in the
6501 * set_rq_offline(), so we should skip updating
6502 * the rq clock again in unthrottle_cfs_rq().
6504 rq_clock_start_loop_update(rq);
6507 list_for_each_entry_rcu(tg, &task_groups, list) {
6508 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
6510 if (!cfs_rq->runtime_enabled)
6514 * clock_task is not advancing so we just need to make sure
6515 * there's some valid quota amount
6517 cfs_rq->runtime_remaining = 1;
6519 * Offline rq is schedulable till CPU is completely disabled
6520 * in take_cpu_down(), so we prevent new cfs throttling here.
6522 cfs_rq->runtime_enabled = 0;
6524 if (cfs_rq_throttled(cfs_rq))
6525 unthrottle_cfs_rq(cfs_rq);
6529 rq_clock_stop_loop_update(rq);
6532 bool cfs_task_bw_constrained(struct task_struct *p)
6534 struct cfs_rq *cfs_rq = task_cfs_rq(p);
6536 if (!cfs_bandwidth_used())
6539 if (cfs_rq->runtime_enabled ||
6540 tg_cfs_bandwidth(cfs_rq->tg)->hierarchical_quota != RUNTIME_INF)
6546 #ifdef CONFIG_NO_HZ_FULL
6547 /* called from pick_next_task_fair() */
6548 static void sched_fair_update_stop_tick(struct rq *rq, struct task_struct *p)
6550 int cpu = cpu_of(rq);
6552 if (!sched_feat(HZ_BW) || !cfs_bandwidth_used())
6555 if (!tick_nohz_full_cpu(cpu))
6558 if (rq->nr_running != 1)
6562 * We know there is only one task runnable and we've just picked it. The
6563 * normal enqueue path will have cleared TICK_DEP_BIT_SCHED if we will
6564 * be otherwise able to stop the tick. Just need to check if we are using
6565 * bandwidth control.
6567 if (cfs_task_bw_constrained(p))
6568 tick_nohz_dep_set_cpu(cpu, TICK_DEP_BIT_SCHED);
6572 #else /* CONFIG_CFS_BANDWIDTH */
6574 static inline bool cfs_bandwidth_used(void)
6579 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {}
6580 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; }
6581 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
6582 static inline void sync_throttle(struct task_group *tg, int cpu) {}
6583 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
6585 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
6590 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
6595 static inline int throttled_lb_pair(struct task_group *tg,
6596 int src_cpu, int dest_cpu)
6601 #ifdef CONFIG_FAIR_GROUP_SCHED
6602 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b, struct cfs_bandwidth *parent) {}
6603 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
6606 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
6610 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
6611 static inline void update_runtime_enabled(struct rq *rq) {}
6612 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
6613 #ifdef CONFIG_CGROUP_SCHED
6614 bool cfs_task_bw_constrained(struct task_struct *p)
6619 #endif /* CONFIG_CFS_BANDWIDTH */
6621 #if !defined(CONFIG_CFS_BANDWIDTH) || !defined(CONFIG_NO_HZ_FULL)
6622 static inline void sched_fair_update_stop_tick(struct rq *rq, struct task_struct *p) {}
6625 /**************************************************
6626 * CFS operations on tasks:
6629 #ifdef CONFIG_SCHED_HRTICK
6630 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
6632 struct sched_entity *se = &p->se;
6634 SCHED_WARN_ON(task_rq(p) != rq);
6636 if (rq->cfs.h_nr_running > 1) {
6637 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
6638 u64 slice = se->slice;
6639 s64 delta = slice - ran;
6642 if (task_current(rq, p))
6646 hrtick_start(rq, delta);
6651 * called from enqueue/dequeue and updates the hrtick when the
6652 * current task is from our class and nr_running is low enough
6655 static void hrtick_update(struct rq *rq)
6657 struct task_struct *curr = rq->curr;
6659 if (!hrtick_enabled_fair(rq) || curr->sched_class != &fair_sched_class)
6662 hrtick_start_fair(rq, curr);
6664 #else /* !CONFIG_SCHED_HRTICK */
6666 hrtick_start_fair(struct rq *rq, struct task_struct *p)
6670 static inline void hrtick_update(struct rq *rq)
6676 static inline bool cpu_overutilized(int cpu)
6678 unsigned long rq_util_min = uclamp_rq_get(cpu_rq(cpu), UCLAMP_MIN);
6679 unsigned long rq_util_max = uclamp_rq_get(cpu_rq(cpu), UCLAMP_MAX);
6681 /* Return true only if the utilization doesn't fit CPU's capacity */
6682 return !util_fits_cpu(cpu_util_cfs(cpu), rq_util_min, rq_util_max, cpu);
6685 static inline void update_overutilized_status(struct rq *rq)
6687 if (!READ_ONCE(rq->rd->overutilized) && cpu_overutilized(rq->cpu)) {
6688 WRITE_ONCE(rq->rd->overutilized, SG_OVERUTILIZED);
6689 trace_sched_overutilized_tp(rq->rd, SG_OVERUTILIZED);
6693 static inline void update_overutilized_status(struct rq *rq) { }
6696 /* Runqueue only has SCHED_IDLE tasks enqueued */
6697 static int sched_idle_rq(struct rq *rq)
6699 return unlikely(rq->nr_running == rq->cfs.idle_h_nr_running &&
6704 static int sched_idle_cpu(int cpu)
6706 return sched_idle_rq(cpu_rq(cpu));
6711 * The enqueue_task method is called before nr_running is
6712 * increased. Here we update the fair scheduling stats and
6713 * then put the task into the rbtree:
6716 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
6718 struct cfs_rq *cfs_rq;
6719 struct sched_entity *se = &p->se;
6720 int idle_h_nr_running = task_has_idle_policy(p);
6721 int task_new = !(flags & ENQUEUE_WAKEUP);
6724 * The code below (indirectly) updates schedutil which looks at
6725 * the cfs_rq utilization to select a frequency.
6726 * Let's add the task's estimated utilization to the cfs_rq's
6727 * estimated utilization, before we update schedutil.
6729 util_est_enqueue(&rq->cfs, p);
6732 * If in_iowait is set, the code below may not trigger any cpufreq
6733 * utilization updates, so do it here explicitly with the IOWAIT flag
6737 cpufreq_update_util(rq, SCHED_CPUFREQ_IOWAIT);
6739 for_each_sched_entity(se) {
6742 cfs_rq = cfs_rq_of(se);
6743 enqueue_entity(cfs_rq, se, flags);
6745 cfs_rq->h_nr_running++;
6746 cfs_rq->idle_h_nr_running += idle_h_nr_running;
6748 if (cfs_rq_is_idle(cfs_rq))
6749 idle_h_nr_running = 1;
6751 /* end evaluation on encountering a throttled cfs_rq */
6752 if (cfs_rq_throttled(cfs_rq))
6753 goto enqueue_throttle;
6755 flags = ENQUEUE_WAKEUP;
6758 for_each_sched_entity(se) {
6759 cfs_rq = cfs_rq_of(se);
6761 update_load_avg(cfs_rq, se, UPDATE_TG);
6762 se_update_runnable(se);
6763 update_cfs_group(se);
6765 cfs_rq->h_nr_running++;
6766 cfs_rq->idle_h_nr_running += idle_h_nr_running;
6768 if (cfs_rq_is_idle(cfs_rq))
6769 idle_h_nr_running = 1;
6771 /* end evaluation on encountering a throttled cfs_rq */
6772 if (cfs_rq_throttled(cfs_rq))
6773 goto enqueue_throttle;
6776 /* At this point se is NULL and we are at root level*/
6777 add_nr_running(rq, 1);
6780 * Since new tasks are assigned an initial util_avg equal to
6781 * half of the spare capacity of their CPU, tiny tasks have the
6782 * ability to cross the overutilized threshold, which will
6783 * result in the load balancer ruining all the task placement
6784 * done by EAS. As a way to mitigate that effect, do not account
6785 * for the first enqueue operation of new tasks during the
6786 * overutilized flag detection.
6788 * A better way of solving this problem would be to wait for
6789 * the PELT signals of tasks to converge before taking them
6790 * into account, but that is not straightforward to implement,
6791 * and the following generally works well enough in practice.
6794 update_overutilized_status(rq);
6797 assert_list_leaf_cfs_rq(rq);
6802 static void set_next_buddy(struct sched_entity *se);
6805 * The dequeue_task method is called before nr_running is
6806 * decreased. We remove the task from the rbtree and
6807 * update the fair scheduling stats:
6809 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
6811 struct cfs_rq *cfs_rq;
6812 struct sched_entity *se = &p->se;
6813 int task_sleep = flags & DEQUEUE_SLEEP;
6814 int idle_h_nr_running = task_has_idle_policy(p);
6815 bool was_sched_idle = sched_idle_rq(rq);
6817 util_est_dequeue(&rq->cfs, p);
6819 for_each_sched_entity(se) {
6820 cfs_rq = cfs_rq_of(se);
6821 dequeue_entity(cfs_rq, se, flags);
6823 cfs_rq->h_nr_running--;
6824 cfs_rq->idle_h_nr_running -= idle_h_nr_running;
6826 if (cfs_rq_is_idle(cfs_rq))
6827 idle_h_nr_running = 1;
6829 /* end evaluation on encountering a throttled cfs_rq */
6830 if (cfs_rq_throttled(cfs_rq))
6831 goto dequeue_throttle;
6833 /* Don't dequeue parent if it has other entities besides us */
6834 if (cfs_rq->load.weight) {
6835 /* Avoid re-evaluating load for this entity: */
6836 se = parent_entity(se);
6838 * Bias pick_next to pick a task from this cfs_rq, as
6839 * p is sleeping when it is within its sched_slice.
6841 if (task_sleep && se && !throttled_hierarchy(cfs_rq))
6845 flags |= DEQUEUE_SLEEP;
6848 for_each_sched_entity(se) {
6849 cfs_rq = cfs_rq_of(se);
6851 update_load_avg(cfs_rq, se, UPDATE_TG);
6852 se_update_runnable(se);
6853 update_cfs_group(se);
6855 cfs_rq->h_nr_running--;
6856 cfs_rq->idle_h_nr_running -= idle_h_nr_running;
6858 if (cfs_rq_is_idle(cfs_rq))
6859 idle_h_nr_running = 1;
6861 /* end evaluation on encountering a throttled cfs_rq */
6862 if (cfs_rq_throttled(cfs_rq))
6863 goto dequeue_throttle;
6867 /* At this point se is NULL and we are at root level*/
6868 sub_nr_running(rq, 1);
6870 /* balance early to pull high priority tasks */
6871 if (unlikely(!was_sched_idle && sched_idle_rq(rq)))
6872 rq->next_balance = jiffies;
6875 util_est_update(&rq->cfs, p, task_sleep);
6881 /* Working cpumask for: load_balance, load_balance_newidle. */
6882 static DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
6883 static DEFINE_PER_CPU(cpumask_var_t, select_rq_mask);
6884 static DEFINE_PER_CPU(cpumask_var_t, should_we_balance_tmpmask);
6886 #ifdef CONFIG_NO_HZ_COMMON
6889 cpumask_var_t idle_cpus_mask;
6891 int has_blocked; /* Idle CPUS has blocked load */
6892 int needs_update; /* Newly idle CPUs need their next_balance collated */
6893 unsigned long next_balance; /* in jiffy units */
6894 unsigned long next_blocked; /* Next update of blocked load in jiffies */
6895 } nohz ____cacheline_aligned;
6897 #endif /* CONFIG_NO_HZ_COMMON */
6899 static unsigned long cpu_load(struct rq *rq)
6901 return cfs_rq_load_avg(&rq->cfs);
6905 * cpu_load_without - compute CPU load without any contributions from *p
6906 * @cpu: the CPU which load is requested
6907 * @p: the task which load should be discounted
6909 * The load of a CPU is defined by the load of tasks currently enqueued on that
6910 * CPU as well as tasks which are currently sleeping after an execution on that
6913 * This method returns the load of the specified CPU by discounting the load of
6914 * the specified task, whenever the task is currently contributing to the CPU
6917 static unsigned long cpu_load_without(struct rq *rq, struct task_struct *p)
6919 struct cfs_rq *cfs_rq;
6922 /* Task has no contribution or is new */
6923 if (cpu_of(rq) != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time))
6924 return cpu_load(rq);
6927 load = READ_ONCE(cfs_rq->avg.load_avg);
6929 /* Discount task's util from CPU's util */
6930 lsub_positive(&load, task_h_load(p));
6935 static unsigned long cpu_runnable(struct rq *rq)
6937 return cfs_rq_runnable_avg(&rq->cfs);
6940 static unsigned long cpu_runnable_without(struct rq *rq, struct task_struct *p)
6942 struct cfs_rq *cfs_rq;
6943 unsigned int runnable;
6945 /* Task has no contribution or is new */
6946 if (cpu_of(rq) != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time))
6947 return cpu_runnable(rq);
6950 runnable = READ_ONCE(cfs_rq->avg.runnable_avg);
6952 /* Discount task's runnable from CPU's runnable */
6953 lsub_positive(&runnable, p->se.avg.runnable_avg);
6958 static unsigned long capacity_of(int cpu)
6960 return cpu_rq(cpu)->cpu_capacity;
6963 static void record_wakee(struct task_struct *p)
6966 * Only decay a single time; tasks that have less then 1 wakeup per
6967 * jiffy will not have built up many flips.
6969 if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) {
6970 current->wakee_flips >>= 1;
6971 current->wakee_flip_decay_ts = jiffies;
6974 if (current->last_wakee != p) {
6975 current->last_wakee = p;
6976 current->wakee_flips++;
6981 * Detect M:N waker/wakee relationships via a switching-frequency heuristic.
6983 * A waker of many should wake a different task than the one last awakened
6984 * at a frequency roughly N times higher than one of its wakees.
6986 * In order to determine whether we should let the load spread vs consolidating
6987 * to shared cache, we look for a minimum 'flip' frequency of llc_size in one
6988 * partner, and a factor of lls_size higher frequency in the other.
6990 * With both conditions met, we can be relatively sure that the relationship is
6991 * non-monogamous, with partner count exceeding socket size.
6993 * Waker/wakee being client/server, worker/dispatcher, interrupt source or
6994 * whatever is irrelevant, spread criteria is apparent partner count exceeds
6997 static int wake_wide(struct task_struct *p)
6999 unsigned int master = current->wakee_flips;
7000 unsigned int slave = p->wakee_flips;
7001 int factor = __this_cpu_read(sd_llc_size);
7004 swap(master, slave);
7005 if (slave < factor || master < slave * factor)
7011 * The purpose of wake_affine() is to quickly determine on which CPU we can run
7012 * soonest. For the purpose of speed we only consider the waking and previous
7015 * wake_affine_idle() - only considers 'now', it check if the waking CPU is
7016 * cache-affine and is (or will be) idle.
7018 * wake_affine_weight() - considers the weight to reflect the average
7019 * scheduling latency of the CPUs. This seems to work
7020 * for the overloaded case.
7023 wake_affine_idle(int this_cpu, int prev_cpu, int sync)
7026 * If this_cpu is idle, it implies the wakeup is from interrupt
7027 * context. Only allow the move if cache is shared. Otherwise an
7028 * interrupt intensive workload could force all tasks onto one
7029 * node depending on the IO topology or IRQ affinity settings.
7031 * If the prev_cpu is idle and cache affine then avoid a migration.
7032 * There is no guarantee that the cache hot data from an interrupt
7033 * is more important than cache hot data on the prev_cpu and from
7034 * a cpufreq perspective, it's better to have higher utilisation
7037 if (available_idle_cpu(this_cpu) && cpus_share_cache(this_cpu, prev_cpu))
7038 return available_idle_cpu(prev_cpu) ? prev_cpu : this_cpu;
7040 if (sync && cpu_rq(this_cpu)->nr_running == 1)
7043 if (available_idle_cpu(prev_cpu))
7046 return nr_cpumask_bits;
7050 wake_affine_weight(struct sched_domain *sd, struct task_struct *p,
7051 int this_cpu, int prev_cpu, int sync)
7053 s64 this_eff_load, prev_eff_load;
7054 unsigned long task_load;
7056 this_eff_load = cpu_load(cpu_rq(this_cpu));
7059 unsigned long current_load = task_h_load(current);
7061 if (current_load > this_eff_load)
7064 this_eff_load -= current_load;
7067 task_load = task_h_load(p);
7069 this_eff_load += task_load;
7070 if (sched_feat(WA_BIAS))
7071 this_eff_load *= 100;
7072 this_eff_load *= capacity_of(prev_cpu);
7074 prev_eff_load = cpu_load(cpu_rq(prev_cpu));
7075 prev_eff_load -= task_load;
7076 if (sched_feat(WA_BIAS))
7077 prev_eff_load *= 100 + (sd->imbalance_pct - 100) / 2;
7078 prev_eff_load *= capacity_of(this_cpu);
7081 * If sync, adjust the weight of prev_eff_load such that if
7082 * prev_eff == this_eff that select_idle_sibling() will consider
7083 * stacking the wakee on top of the waker if no other CPU is
7089 return this_eff_load < prev_eff_load ? this_cpu : nr_cpumask_bits;
7092 static int wake_affine(struct sched_domain *sd, struct task_struct *p,
7093 int this_cpu, int prev_cpu, int sync)
7095 int target = nr_cpumask_bits;
7097 if (sched_feat(WA_IDLE))
7098 target = wake_affine_idle(this_cpu, prev_cpu, sync);
7100 if (sched_feat(WA_WEIGHT) && target == nr_cpumask_bits)
7101 target = wake_affine_weight(sd, p, this_cpu, prev_cpu, sync);
7103 schedstat_inc(p->stats.nr_wakeups_affine_attempts);
7104 if (target != this_cpu)
7107 schedstat_inc(sd->ttwu_move_affine);
7108 schedstat_inc(p->stats.nr_wakeups_affine);
7112 static struct sched_group *
7113 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu);
7116 * find_idlest_group_cpu - find the idlest CPU among the CPUs in the group.
7119 find_idlest_group_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
7121 unsigned long load, min_load = ULONG_MAX;
7122 unsigned int min_exit_latency = UINT_MAX;
7123 u64 latest_idle_timestamp = 0;
7124 int least_loaded_cpu = this_cpu;
7125 int shallowest_idle_cpu = -1;
7128 /* Check if we have any choice: */
7129 if (group->group_weight == 1)
7130 return cpumask_first(sched_group_span(group));
7132 /* Traverse only the allowed CPUs */
7133 for_each_cpu_and(i, sched_group_span(group), p->cpus_ptr) {
7134 struct rq *rq = cpu_rq(i);
7136 if (!sched_core_cookie_match(rq, p))
7139 if (sched_idle_cpu(i))
7142 if (available_idle_cpu(i)) {
7143 struct cpuidle_state *idle = idle_get_state(rq);
7144 if (idle && idle->exit_latency < min_exit_latency) {
7146 * We give priority to a CPU whose idle state
7147 * has the smallest exit latency irrespective
7148 * of any idle timestamp.
7150 min_exit_latency = idle->exit_latency;
7151 latest_idle_timestamp = rq->idle_stamp;
7152 shallowest_idle_cpu = i;
7153 } else if ((!idle || idle->exit_latency == min_exit_latency) &&
7154 rq->idle_stamp > latest_idle_timestamp) {
7156 * If equal or no active idle state, then
7157 * the most recently idled CPU might have
7160 latest_idle_timestamp = rq->idle_stamp;
7161 shallowest_idle_cpu = i;
7163 } else if (shallowest_idle_cpu == -1) {
7164 load = cpu_load(cpu_rq(i));
7165 if (load < min_load) {
7167 least_loaded_cpu = i;
7172 return shallowest_idle_cpu != -1 ? shallowest_idle_cpu : least_loaded_cpu;
7175 static inline int find_idlest_cpu(struct sched_domain *sd, struct task_struct *p,
7176 int cpu, int prev_cpu, int sd_flag)
7180 if (!cpumask_intersects(sched_domain_span(sd), p->cpus_ptr))
7184 * We need task's util for cpu_util_without, sync it up to
7185 * prev_cpu's last_update_time.
7187 if (!(sd_flag & SD_BALANCE_FORK))
7188 sync_entity_load_avg(&p->se);
7191 struct sched_group *group;
7192 struct sched_domain *tmp;
7195 if (!(sd->flags & sd_flag)) {
7200 group = find_idlest_group(sd, p, cpu);
7206 new_cpu = find_idlest_group_cpu(group, p, cpu);
7207 if (new_cpu == cpu) {
7208 /* Now try balancing at a lower domain level of 'cpu': */
7213 /* Now try balancing at a lower domain level of 'new_cpu': */
7215 weight = sd->span_weight;
7217 for_each_domain(cpu, tmp) {
7218 if (weight <= tmp->span_weight)
7220 if (tmp->flags & sd_flag)
7228 static inline int __select_idle_cpu(int cpu, struct task_struct *p)
7230 if ((available_idle_cpu(cpu) || sched_idle_cpu(cpu)) &&
7231 sched_cpu_cookie_match(cpu_rq(cpu), p))
7237 #ifdef CONFIG_SCHED_SMT
7238 DEFINE_STATIC_KEY_FALSE(sched_smt_present);
7239 EXPORT_SYMBOL_GPL(sched_smt_present);
7241 static inline void set_idle_cores(int cpu, int val)
7243 struct sched_domain_shared *sds;
7245 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
7247 WRITE_ONCE(sds->has_idle_cores, val);
7250 static inline bool test_idle_cores(int cpu)
7252 struct sched_domain_shared *sds;
7254 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
7256 return READ_ONCE(sds->has_idle_cores);
7262 * Scans the local SMT mask to see if the entire core is idle, and records this
7263 * information in sd_llc_shared->has_idle_cores.
7265 * Since SMT siblings share all cache levels, inspecting this limited remote
7266 * state should be fairly cheap.
7268 void __update_idle_core(struct rq *rq)
7270 int core = cpu_of(rq);
7274 if (test_idle_cores(core))
7277 for_each_cpu(cpu, cpu_smt_mask(core)) {
7281 if (!available_idle_cpu(cpu))
7285 set_idle_cores(core, 1);
7291 * Scan the entire LLC domain for idle cores; this dynamically switches off if
7292 * there are no idle cores left in the system; tracked through
7293 * sd_llc->shared->has_idle_cores and enabled through update_idle_core() above.
7295 static int select_idle_core(struct task_struct *p, int core, struct cpumask *cpus, int *idle_cpu)
7300 for_each_cpu(cpu, cpu_smt_mask(core)) {
7301 if (!available_idle_cpu(cpu)) {
7303 if (*idle_cpu == -1) {
7304 if (sched_idle_cpu(cpu) && cpumask_test_cpu(cpu, cpus)) {
7312 if (*idle_cpu == -1 && cpumask_test_cpu(cpu, cpus))
7319 cpumask_andnot(cpus, cpus, cpu_smt_mask(core));
7324 * Scan the local SMT mask for idle CPUs.
7326 static int select_idle_smt(struct task_struct *p, struct sched_domain *sd, int target)
7330 for_each_cpu_and(cpu, cpu_smt_mask(target), p->cpus_ptr) {
7334 * Check if the CPU is in the LLC scheduling domain of @target.
7335 * Due to isolcpus, there is no guarantee that all the siblings are in the domain.
7337 if (!cpumask_test_cpu(cpu, sched_domain_span(sd)))
7339 if (available_idle_cpu(cpu) || sched_idle_cpu(cpu))
7346 #else /* CONFIG_SCHED_SMT */
7348 static inline void set_idle_cores(int cpu, int val)
7352 static inline bool test_idle_cores(int cpu)
7357 static inline int select_idle_core(struct task_struct *p, int core, struct cpumask *cpus, int *idle_cpu)
7359 return __select_idle_cpu(core, p);
7362 static inline int select_idle_smt(struct task_struct *p, struct sched_domain *sd, int target)
7367 #endif /* CONFIG_SCHED_SMT */
7370 * Scan the LLC domain for idle CPUs; this is dynamically regulated by
7371 * comparing the average scan cost (tracked in sd->avg_scan_cost) against the
7372 * average idle time for this rq (as found in rq->avg_idle).
7374 static int select_idle_cpu(struct task_struct *p, struct sched_domain *sd, bool has_idle_core, int target)
7376 struct cpumask *cpus = this_cpu_cpumask_var_ptr(select_rq_mask);
7377 int i, cpu, idle_cpu = -1, nr = INT_MAX;
7378 struct sched_domain_shared *sd_share;
7380 cpumask_and(cpus, sched_domain_span(sd), p->cpus_ptr);
7382 if (sched_feat(SIS_UTIL)) {
7383 sd_share = rcu_dereference(per_cpu(sd_llc_shared, target));
7385 /* because !--nr is the condition to stop scan */
7386 nr = READ_ONCE(sd_share->nr_idle_scan) + 1;
7387 /* overloaded LLC is unlikely to have idle cpu/core */
7393 if (static_branch_unlikely(&sched_cluster_active)) {
7394 struct sched_group *sg = sd->groups;
7396 if (sg->flags & SD_CLUSTER) {
7397 for_each_cpu_wrap(cpu, sched_group_span(sg), target + 1) {
7398 if (!cpumask_test_cpu(cpu, cpus))
7401 if (has_idle_core) {
7402 i = select_idle_core(p, cpu, cpus, &idle_cpu);
7403 if ((unsigned int)i < nr_cpumask_bits)
7408 idle_cpu = __select_idle_cpu(cpu, p);
7409 if ((unsigned int)idle_cpu < nr_cpumask_bits)
7413 cpumask_andnot(cpus, cpus, sched_group_span(sg));
7417 for_each_cpu_wrap(cpu, cpus, target + 1) {
7418 if (has_idle_core) {
7419 i = select_idle_core(p, cpu, cpus, &idle_cpu);
7420 if ((unsigned int)i < nr_cpumask_bits)
7426 idle_cpu = __select_idle_cpu(cpu, p);
7427 if ((unsigned int)idle_cpu < nr_cpumask_bits)
7433 set_idle_cores(target, false);
7439 * Scan the asym_capacity domain for idle CPUs; pick the first idle one on which
7440 * the task fits. If no CPU is big enough, but there are idle ones, try to
7441 * maximize capacity.
7444 select_idle_capacity(struct task_struct *p, struct sched_domain *sd, int target)
7446 unsigned long task_util, util_min, util_max, best_cap = 0;
7447 int fits, best_fits = 0;
7448 int cpu, best_cpu = -1;
7449 struct cpumask *cpus;
7451 cpus = this_cpu_cpumask_var_ptr(select_rq_mask);
7452 cpumask_and(cpus, sched_domain_span(sd), p->cpus_ptr);
7454 task_util = task_util_est(p);
7455 util_min = uclamp_eff_value(p, UCLAMP_MIN);
7456 util_max = uclamp_eff_value(p, UCLAMP_MAX);
7458 for_each_cpu_wrap(cpu, cpus, target) {
7459 unsigned long cpu_cap = capacity_of(cpu);
7461 if (!available_idle_cpu(cpu) && !sched_idle_cpu(cpu))
7464 fits = util_fits_cpu(task_util, util_min, util_max, cpu);
7466 /* This CPU fits with all requirements */
7470 * Only the min performance hint (i.e. uclamp_min) doesn't fit.
7471 * Look for the CPU with best capacity.
7474 cpu_cap = arch_scale_cpu_capacity(cpu) - thermal_load_avg(cpu_rq(cpu));
7477 * First, select CPU which fits better (-1 being better than 0).
7478 * Then, select the one with best capacity at same level.
7480 if ((fits < best_fits) ||
7481 ((fits == best_fits) && (cpu_cap > best_cap))) {
7491 static inline bool asym_fits_cpu(unsigned long util,
7492 unsigned long util_min,
7493 unsigned long util_max,
7496 if (sched_asym_cpucap_active())
7498 * Return true only if the cpu fully fits the task requirements
7499 * which include the utilization and the performance hints.
7501 return (util_fits_cpu(util, util_min, util_max, cpu) > 0);
7507 * Try and locate an idle core/thread in the LLC cache domain.
7509 static int select_idle_sibling(struct task_struct *p, int prev, int target)
7511 bool has_idle_core = false;
7512 struct sched_domain *sd;
7513 unsigned long task_util, util_min, util_max;
7514 int i, recent_used_cpu, prev_aff = -1;
7517 * On asymmetric system, update task utilization because we will check
7518 * that the task fits with cpu's capacity.
7520 if (sched_asym_cpucap_active()) {
7521 sync_entity_load_avg(&p->se);
7522 task_util = task_util_est(p);
7523 util_min = uclamp_eff_value(p, UCLAMP_MIN);
7524 util_max = uclamp_eff_value(p, UCLAMP_MAX);
7528 * per-cpu select_rq_mask usage
7530 lockdep_assert_irqs_disabled();
7532 if ((available_idle_cpu(target) || sched_idle_cpu(target)) &&
7533 asym_fits_cpu(task_util, util_min, util_max, target))
7537 * If the previous CPU is cache affine and idle, don't be stupid:
7539 if (prev != target && cpus_share_cache(prev, target) &&
7540 (available_idle_cpu(prev) || sched_idle_cpu(prev)) &&
7541 asym_fits_cpu(task_util, util_min, util_max, prev)) {
7543 if (!static_branch_unlikely(&sched_cluster_active) ||
7544 cpus_share_resources(prev, target))
7551 * Allow a per-cpu kthread to stack with the wakee if the
7552 * kworker thread and the tasks previous CPUs are the same.
7553 * The assumption is that the wakee queued work for the
7554 * per-cpu kthread that is now complete and the wakeup is
7555 * essentially a sync wakeup. An obvious example of this
7556 * pattern is IO completions.
7558 if (is_per_cpu_kthread(current) &&
7560 prev == smp_processor_id() &&
7561 this_rq()->nr_running <= 1 &&
7562 asym_fits_cpu(task_util, util_min, util_max, prev)) {
7566 /* Check a recently used CPU as a potential idle candidate: */
7567 recent_used_cpu = p->recent_used_cpu;
7568 p->recent_used_cpu = prev;
7569 if (recent_used_cpu != prev &&
7570 recent_used_cpu != target &&
7571 cpus_share_cache(recent_used_cpu, target) &&
7572 (available_idle_cpu(recent_used_cpu) || sched_idle_cpu(recent_used_cpu)) &&
7573 cpumask_test_cpu(recent_used_cpu, p->cpus_ptr) &&
7574 asym_fits_cpu(task_util, util_min, util_max, recent_used_cpu)) {
7576 if (!static_branch_unlikely(&sched_cluster_active) ||
7577 cpus_share_resources(recent_used_cpu, target))
7578 return recent_used_cpu;
7581 recent_used_cpu = -1;
7585 * For asymmetric CPU capacity systems, our domain of interest is
7586 * sd_asym_cpucapacity rather than sd_llc.
7588 if (sched_asym_cpucap_active()) {
7589 sd = rcu_dereference(per_cpu(sd_asym_cpucapacity, target));
7591 * On an asymmetric CPU capacity system where an exclusive
7592 * cpuset defines a symmetric island (i.e. one unique
7593 * capacity_orig value through the cpuset), the key will be set
7594 * but the CPUs within that cpuset will not have a domain with
7595 * SD_ASYM_CPUCAPACITY. These should follow the usual symmetric
7599 i = select_idle_capacity(p, sd, target);
7600 return ((unsigned)i < nr_cpumask_bits) ? i : target;
7604 sd = rcu_dereference(per_cpu(sd_llc, target));
7608 if (sched_smt_active()) {
7609 has_idle_core = test_idle_cores(target);
7611 if (!has_idle_core && cpus_share_cache(prev, target)) {
7612 i = select_idle_smt(p, sd, prev);
7613 if ((unsigned int)i < nr_cpumask_bits)
7618 i = select_idle_cpu(p, sd, has_idle_core, target);
7619 if ((unsigned)i < nr_cpumask_bits)
7623 * For cluster machines which have lower sharing cache like L2 or
7624 * LLC Tag, we tend to find an idle CPU in the target's cluster
7625 * first. But prev_cpu or recent_used_cpu may also be a good candidate,
7626 * use them if possible when no idle CPU found in select_idle_cpu().
7628 if ((unsigned int)prev_aff < nr_cpumask_bits)
7630 if ((unsigned int)recent_used_cpu < nr_cpumask_bits)
7631 return recent_used_cpu;
7637 * cpu_util() - Estimates the amount of CPU capacity used by CFS tasks.
7638 * @cpu: the CPU to get the utilization for
7639 * @p: task for which the CPU utilization should be predicted or NULL
7640 * @dst_cpu: CPU @p migrates to, -1 if @p moves from @cpu or @p == NULL
7641 * @boost: 1 to enable boosting, otherwise 0
7643 * The unit of the return value must be the same as the one of CPU capacity
7644 * so that CPU utilization can be compared with CPU capacity.
7646 * CPU utilization is the sum of running time of runnable tasks plus the
7647 * recent utilization of currently non-runnable tasks on that CPU.
7648 * It represents the amount of CPU capacity currently used by CFS tasks in
7649 * the range [0..max CPU capacity] with max CPU capacity being the CPU
7650 * capacity at f_max.
7652 * The estimated CPU utilization is defined as the maximum between CPU
7653 * utilization and sum of the estimated utilization of the currently
7654 * runnable tasks on that CPU. It preserves a utilization "snapshot" of
7655 * previously-executed tasks, which helps better deduce how busy a CPU will
7656 * be when a long-sleeping task wakes up. The contribution to CPU utilization
7657 * of such a task would be significantly decayed at this point of time.
7659 * Boosted CPU utilization is defined as max(CPU runnable, CPU utilization).
7660 * CPU contention for CFS tasks can be detected by CPU runnable > CPU
7661 * utilization. Boosting is implemented in cpu_util() so that internal
7662 * users (e.g. EAS) can use it next to external users (e.g. schedutil),
7663 * latter via cpu_util_cfs_boost().
7665 * CPU utilization can be higher than the current CPU capacity
7666 * (f_curr/f_max * max CPU capacity) or even the max CPU capacity because
7667 * of rounding errors as well as task migrations or wakeups of new tasks.
7668 * CPU utilization has to be capped to fit into the [0..max CPU capacity]
7669 * range. Otherwise a group of CPUs (CPU0 util = 121% + CPU1 util = 80%)
7670 * could be seen as over-utilized even though CPU1 has 20% of spare CPU
7671 * capacity. CPU utilization is allowed to overshoot current CPU capacity
7672 * though since this is useful for predicting the CPU capacity required
7673 * after task migrations (scheduler-driven DVFS).
7675 * Return: (Boosted) (estimated) utilization for the specified CPU.
7677 static unsigned long
7678 cpu_util(int cpu, struct task_struct *p, int dst_cpu, int boost)
7680 struct cfs_rq *cfs_rq = &cpu_rq(cpu)->cfs;
7681 unsigned long util = READ_ONCE(cfs_rq->avg.util_avg);
7682 unsigned long runnable;
7685 runnable = READ_ONCE(cfs_rq->avg.runnable_avg);
7686 util = max(util, runnable);
7690 * If @dst_cpu is -1 or @p migrates from @cpu to @dst_cpu remove its
7691 * contribution. If @p migrates from another CPU to @cpu add its
7692 * contribution. In all the other cases @cpu is not impacted by the
7693 * migration so its util_avg is already correct.
7695 if (p && task_cpu(p) == cpu && dst_cpu != cpu)
7696 lsub_positive(&util, task_util(p));
7697 else if (p && task_cpu(p) != cpu && dst_cpu == cpu)
7698 util += task_util(p);
7700 if (sched_feat(UTIL_EST)) {
7701 unsigned long util_est;
7703 util_est = READ_ONCE(cfs_rq->avg.util_est);
7706 * During wake-up @p isn't enqueued yet and doesn't contribute
7707 * to any cpu_rq(cpu)->cfs.avg.util_est.
7708 * If @dst_cpu == @cpu add it to "simulate" cpu_util after @p
7709 * has been enqueued.
7711 * During exec (@dst_cpu = -1) @p is enqueued and does
7712 * contribute to cpu_rq(cpu)->cfs.util_est.
7713 * Remove it to "simulate" cpu_util without @p's contribution.
7715 * Despite the task_on_rq_queued(@p) check there is still a
7716 * small window for a possible race when an exec
7717 * select_task_rq_fair() races with LB's detach_task().
7721 * p->on_rq = TASK_ON_RQ_MIGRATING;
7722 * -------------------------------- A
7724 * dequeue_task_fair() + Race Time
7725 * util_est_dequeue() /
7726 * -------------------------------- B
7728 * The additional check "current == p" is required to further
7729 * reduce the race window.
7732 util_est += _task_util_est(p);
7733 else if (p && unlikely(task_on_rq_queued(p) || current == p))
7734 lsub_positive(&util_est, _task_util_est(p));
7736 util = max(util, util_est);
7739 return min(util, arch_scale_cpu_capacity(cpu));
7742 unsigned long cpu_util_cfs(int cpu)
7744 return cpu_util(cpu, NULL, -1, 0);
7747 unsigned long cpu_util_cfs_boost(int cpu)
7749 return cpu_util(cpu, NULL, -1, 1);
7753 * cpu_util_without: compute cpu utilization without any contributions from *p
7754 * @cpu: the CPU which utilization is requested
7755 * @p: the task which utilization should be discounted
7757 * The utilization of a CPU is defined by the utilization of tasks currently
7758 * enqueued on that CPU as well as tasks which are currently sleeping after an
7759 * execution on that CPU.
7761 * This method returns the utilization of the specified CPU by discounting the
7762 * utilization of the specified task, whenever the task is currently
7763 * contributing to the CPU utilization.
7765 static unsigned long cpu_util_without(int cpu, struct task_struct *p)
7767 /* Task has no contribution or is new */
7768 if (cpu != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time))
7771 return cpu_util(cpu, p, -1, 0);
7775 * energy_env - Utilization landscape for energy estimation.
7776 * @task_busy_time: Utilization contribution by the task for which we test the
7777 * placement. Given by eenv_task_busy_time().
7778 * @pd_busy_time: Utilization of the whole perf domain without the task
7779 * contribution. Given by eenv_pd_busy_time().
7780 * @cpu_cap: Maximum CPU capacity for the perf domain.
7781 * @pd_cap: Entire perf domain capacity. (pd->nr_cpus * cpu_cap).
7784 unsigned long task_busy_time;
7785 unsigned long pd_busy_time;
7786 unsigned long cpu_cap;
7787 unsigned long pd_cap;
7791 * Compute the task busy time for compute_energy(). This time cannot be
7792 * injected directly into effective_cpu_util() because of the IRQ scaling.
7793 * The latter only makes sense with the most recent CPUs where the task has
7796 static inline void eenv_task_busy_time(struct energy_env *eenv,
7797 struct task_struct *p, int prev_cpu)
7799 unsigned long busy_time, max_cap = arch_scale_cpu_capacity(prev_cpu);
7800 unsigned long irq = cpu_util_irq(cpu_rq(prev_cpu));
7802 if (unlikely(irq >= max_cap))
7803 busy_time = max_cap;
7805 busy_time = scale_irq_capacity(task_util_est(p), irq, max_cap);
7807 eenv->task_busy_time = busy_time;
7811 * Compute the perf_domain (PD) busy time for compute_energy(). Based on the
7812 * utilization for each @pd_cpus, it however doesn't take into account
7813 * clamping since the ratio (utilization / cpu_capacity) is already enough to
7814 * scale the EM reported power consumption at the (eventually clamped)
7817 * The contribution of the task @p for which we want to estimate the
7818 * energy cost is removed (by cpu_util()) and must be calculated
7819 * separately (see eenv_task_busy_time). This ensures:
7821 * - A stable PD utilization, no matter which CPU of that PD we want to place
7824 * - A fair comparison between CPUs as the task contribution (task_util())
7825 * will always be the same no matter which CPU utilization we rely on
7826 * (util_avg or util_est).
7828 * Set @eenv busy time for the PD that spans @pd_cpus. This busy time can't
7829 * exceed @eenv->pd_cap.
7831 static inline void eenv_pd_busy_time(struct energy_env *eenv,
7832 struct cpumask *pd_cpus,
7833 struct task_struct *p)
7835 unsigned long busy_time = 0;
7838 for_each_cpu(cpu, pd_cpus) {
7839 unsigned long util = cpu_util(cpu, p, -1, 0);
7841 busy_time += effective_cpu_util(cpu, util, NULL, NULL);
7844 eenv->pd_busy_time = min(eenv->pd_cap, busy_time);
7848 * Compute the maximum utilization for compute_energy() when the task @p
7849 * is placed on the cpu @dst_cpu.
7851 * Returns the maximum utilization among @eenv->cpus. This utilization can't
7852 * exceed @eenv->cpu_cap.
7854 static inline unsigned long
7855 eenv_pd_max_util(struct energy_env *eenv, struct cpumask *pd_cpus,
7856 struct task_struct *p, int dst_cpu)
7858 unsigned long max_util = 0;
7861 for_each_cpu(cpu, pd_cpus) {
7862 struct task_struct *tsk = (cpu == dst_cpu) ? p : NULL;
7863 unsigned long util = cpu_util(cpu, p, dst_cpu, 1);
7864 unsigned long eff_util, min, max;
7867 * Performance domain frequency: utilization clamping
7868 * must be considered since it affects the selection
7869 * of the performance domain frequency.
7870 * NOTE: in case RT tasks are running, by default the
7871 * FREQUENCY_UTIL's utilization can be max OPP.
7873 eff_util = effective_cpu_util(cpu, util, &min, &max);
7875 /* Task's uclamp can modify min and max value */
7876 if (tsk && uclamp_is_used()) {
7877 min = max(min, uclamp_eff_value(p, UCLAMP_MIN));
7880 * If there is no active max uclamp constraint,
7881 * directly use task's one, otherwise keep max.
7883 if (uclamp_rq_is_idle(cpu_rq(cpu)))
7884 max = uclamp_eff_value(p, UCLAMP_MAX);
7886 max = max(max, uclamp_eff_value(p, UCLAMP_MAX));
7889 eff_util = sugov_effective_cpu_perf(cpu, eff_util, min, max);
7890 max_util = max(max_util, eff_util);
7893 return min(max_util, eenv->cpu_cap);
7897 * compute_energy(): Use the Energy Model to estimate the energy that @pd would
7898 * consume for a given utilization landscape @eenv. When @dst_cpu < 0, the task
7899 * contribution is ignored.
7901 static inline unsigned long
7902 compute_energy(struct energy_env *eenv, struct perf_domain *pd,
7903 struct cpumask *pd_cpus, struct task_struct *p, int dst_cpu)
7905 unsigned long max_util = eenv_pd_max_util(eenv, pd_cpus, p, dst_cpu);
7906 unsigned long busy_time = eenv->pd_busy_time;
7907 unsigned long energy;
7910 busy_time = min(eenv->pd_cap, busy_time + eenv->task_busy_time);
7912 energy = em_cpu_energy(pd->em_pd, max_util, busy_time, eenv->cpu_cap);
7914 trace_sched_compute_energy_tp(p, dst_cpu, energy, max_util, busy_time);
7920 * find_energy_efficient_cpu(): Find most energy-efficient target CPU for the
7921 * waking task. find_energy_efficient_cpu() looks for the CPU with maximum
7922 * spare capacity in each performance domain and uses it as a potential
7923 * candidate to execute the task. Then, it uses the Energy Model to figure
7924 * out which of the CPU candidates is the most energy-efficient.
7926 * The rationale for this heuristic is as follows. In a performance domain,
7927 * all the most energy efficient CPU candidates (according to the Energy
7928 * Model) are those for which we'll request a low frequency. When there are
7929 * several CPUs for which the frequency request will be the same, we don't
7930 * have enough data to break the tie between them, because the Energy Model
7931 * only includes active power costs. With this model, if we assume that
7932 * frequency requests follow utilization (e.g. using schedutil), the CPU with
7933 * the maximum spare capacity in a performance domain is guaranteed to be among
7934 * the best candidates of the performance domain.
7936 * In practice, it could be preferable from an energy standpoint to pack
7937 * small tasks on a CPU in order to let other CPUs go in deeper idle states,
7938 * but that could also hurt our chances to go cluster idle, and we have no
7939 * ways to tell with the current Energy Model if this is actually a good
7940 * idea or not. So, find_energy_efficient_cpu() basically favors
7941 * cluster-packing, and spreading inside a cluster. That should at least be
7942 * a good thing for latency, and this is consistent with the idea that most
7943 * of the energy savings of EAS come from the asymmetry of the system, and
7944 * not so much from breaking the tie between identical CPUs. That's also the
7945 * reason why EAS is enabled in the topology code only for systems where
7946 * SD_ASYM_CPUCAPACITY is set.
7948 * NOTE: Forkees are not accepted in the energy-aware wake-up path because
7949 * they don't have any useful utilization data yet and it's not possible to
7950 * forecast their impact on energy consumption. Consequently, they will be
7951 * placed by find_idlest_cpu() on the least loaded CPU, which might turn out
7952 * to be energy-inefficient in some use-cases. The alternative would be to
7953 * bias new tasks towards specific types of CPUs first, or to try to infer
7954 * their util_avg from the parent task, but those heuristics could hurt
7955 * other use-cases too. So, until someone finds a better way to solve this,
7956 * let's keep things simple by re-using the existing slow path.
7958 static int find_energy_efficient_cpu(struct task_struct *p, int prev_cpu)
7960 struct cpumask *cpus = this_cpu_cpumask_var_ptr(select_rq_mask);
7961 unsigned long prev_delta = ULONG_MAX, best_delta = ULONG_MAX;
7962 unsigned long p_util_min = uclamp_is_used() ? uclamp_eff_value(p, UCLAMP_MIN) : 0;
7963 unsigned long p_util_max = uclamp_is_used() ? uclamp_eff_value(p, UCLAMP_MAX) : 1024;
7964 struct root_domain *rd = this_rq()->rd;
7965 int cpu, best_energy_cpu, target = -1;
7966 int prev_fits = -1, best_fits = -1;
7967 unsigned long best_thermal_cap = 0;
7968 unsigned long prev_thermal_cap = 0;
7969 struct sched_domain *sd;
7970 struct perf_domain *pd;
7971 struct energy_env eenv;
7974 pd = rcu_dereference(rd->pd);
7975 if (!pd || READ_ONCE(rd->overutilized))
7979 * Energy-aware wake-up happens on the lowest sched_domain starting
7980 * from sd_asym_cpucapacity spanning over this_cpu and prev_cpu.
7982 sd = rcu_dereference(*this_cpu_ptr(&sd_asym_cpucapacity));
7983 while (sd && !cpumask_test_cpu(prev_cpu, sched_domain_span(sd)))
7990 sync_entity_load_avg(&p->se);
7991 if (!task_util_est(p) && p_util_min == 0)
7994 eenv_task_busy_time(&eenv, p, prev_cpu);
7996 for (; pd; pd = pd->next) {
7997 unsigned long util_min = p_util_min, util_max = p_util_max;
7998 unsigned long cpu_cap, cpu_thermal_cap, util;
7999 long prev_spare_cap = -1, max_spare_cap = -1;
8000 unsigned long rq_util_min, rq_util_max;
8001 unsigned long cur_delta, base_energy;
8002 int max_spare_cap_cpu = -1;
8003 int fits, max_fits = -1;
8005 cpumask_and(cpus, perf_domain_span(pd), cpu_online_mask);
8007 if (cpumask_empty(cpus))
8010 /* Account thermal pressure for the energy estimation */
8011 cpu = cpumask_first(cpus);
8012 cpu_thermal_cap = arch_scale_cpu_capacity(cpu);
8013 cpu_thermal_cap -= arch_scale_thermal_pressure(cpu);
8015 eenv.cpu_cap = cpu_thermal_cap;
8018 for_each_cpu(cpu, cpus) {
8019 struct rq *rq = cpu_rq(cpu);
8021 eenv.pd_cap += cpu_thermal_cap;
8023 if (!cpumask_test_cpu(cpu, sched_domain_span(sd)))
8026 if (!cpumask_test_cpu(cpu, p->cpus_ptr))
8029 util = cpu_util(cpu, p, cpu, 0);
8030 cpu_cap = capacity_of(cpu);
8033 * Skip CPUs that cannot satisfy the capacity request.
8034 * IOW, placing the task there would make the CPU
8035 * overutilized. Take uclamp into account to see how
8036 * much capacity we can get out of the CPU; this is
8037 * aligned with sched_cpu_util().
8039 if (uclamp_is_used() && !uclamp_rq_is_idle(rq)) {
8041 * Open code uclamp_rq_util_with() except for
8042 * the clamp() part. Ie: apply max aggregation
8043 * only. util_fits_cpu() logic requires to
8044 * operate on non clamped util but must use the
8045 * max-aggregated uclamp_{min, max}.
8047 rq_util_min = uclamp_rq_get(rq, UCLAMP_MIN);
8048 rq_util_max = uclamp_rq_get(rq, UCLAMP_MAX);
8050 util_min = max(rq_util_min, p_util_min);
8051 util_max = max(rq_util_max, p_util_max);
8054 fits = util_fits_cpu(util, util_min, util_max, cpu);
8058 lsub_positive(&cpu_cap, util);
8060 if (cpu == prev_cpu) {
8061 /* Always use prev_cpu as a candidate. */
8062 prev_spare_cap = cpu_cap;
8064 } else if ((fits > max_fits) ||
8065 ((fits == max_fits) && ((long)cpu_cap > max_spare_cap))) {
8067 * Find the CPU with the maximum spare capacity
8068 * among the remaining CPUs in the performance
8071 max_spare_cap = cpu_cap;
8072 max_spare_cap_cpu = cpu;
8077 if (max_spare_cap_cpu < 0 && prev_spare_cap < 0)
8080 eenv_pd_busy_time(&eenv, cpus, p);
8081 /* Compute the 'base' energy of the pd, without @p */
8082 base_energy = compute_energy(&eenv, pd, cpus, p, -1);
8084 /* Evaluate the energy impact of using prev_cpu. */
8085 if (prev_spare_cap > -1) {
8086 prev_delta = compute_energy(&eenv, pd, cpus, p,
8088 /* CPU utilization has changed */
8089 if (prev_delta < base_energy)
8091 prev_delta -= base_energy;
8092 prev_thermal_cap = cpu_thermal_cap;
8093 best_delta = min(best_delta, prev_delta);
8096 /* Evaluate the energy impact of using max_spare_cap_cpu. */
8097 if (max_spare_cap_cpu >= 0 && max_spare_cap > prev_spare_cap) {
8098 /* Current best energy cpu fits better */
8099 if (max_fits < best_fits)
8103 * Both don't fit performance hint (i.e. uclamp_min)
8104 * but best energy cpu has better capacity.
8106 if ((max_fits < 0) &&
8107 (cpu_thermal_cap <= best_thermal_cap))
8110 cur_delta = compute_energy(&eenv, pd, cpus, p,
8112 /* CPU utilization has changed */
8113 if (cur_delta < base_energy)
8115 cur_delta -= base_energy;
8118 * Both fit for the task but best energy cpu has lower
8121 if ((max_fits > 0) && (best_fits > 0) &&
8122 (cur_delta >= best_delta))
8125 best_delta = cur_delta;
8126 best_energy_cpu = max_spare_cap_cpu;
8127 best_fits = max_fits;
8128 best_thermal_cap = cpu_thermal_cap;
8133 if ((best_fits > prev_fits) ||
8134 ((best_fits > 0) && (best_delta < prev_delta)) ||
8135 ((best_fits < 0) && (best_thermal_cap > prev_thermal_cap)))
8136 target = best_energy_cpu;
8147 * select_task_rq_fair: Select target runqueue for the waking task in domains
8148 * that have the relevant SD flag set. In practice, this is SD_BALANCE_WAKE,
8149 * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
8151 * Balances load by selecting the idlest CPU in the idlest group, or under
8152 * certain conditions an idle sibling CPU if the domain has SD_WAKE_AFFINE set.
8154 * Returns the target CPU number.
8157 select_task_rq_fair(struct task_struct *p, int prev_cpu, int wake_flags)
8159 int sync = (wake_flags & WF_SYNC) && !(current->flags & PF_EXITING);
8160 struct sched_domain *tmp, *sd = NULL;
8161 int cpu = smp_processor_id();
8162 int new_cpu = prev_cpu;
8163 int want_affine = 0;
8164 /* SD_flags and WF_flags share the first nibble */
8165 int sd_flag = wake_flags & 0xF;
8168 * required for stable ->cpus_allowed
8170 lockdep_assert_held(&p->pi_lock);
8171 if (wake_flags & WF_TTWU) {
8174 if ((wake_flags & WF_CURRENT_CPU) &&
8175 cpumask_test_cpu(cpu, p->cpus_ptr))
8178 if (sched_energy_enabled()) {
8179 new_cpu = find_energy_efficient_cpu(p, prev_cpu);
8185 want_affine = !wake_wide(p) && cpumask_test_cpu(cpu, p->cpus_ptr);
8189 for_each_domain(cpu, tmp) {
8191 * If both 'cpu' and 'prev_cpu' are part of this domain,
8192 * cpu is a valid SD_WAKE_AFFINE target.
8194 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
8195 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
8196 if (cpu != prev_cpu)
8197 new_cpu = wake_affine(tmp, p, cpu, prev_cpu, sync);
8199 sd = NULL; /* Prefer wake_affine over balance flags */
8204 * Usually only true for WF_EXEC and WF_FORK, as sched_domains
8205 * usually do not have SD_BALANCE_WAKE set. That means wakeup
8206 * will usually go to the fast path.
8208 if (tmp->flags & sd_flag)
8210 else if (!want_affine)
8216 new_cpu = find_idlest_cpu(sd, p, cpu, prev_cpu, sd_flag);
8217 } else if (wake_flags & WF_TTWU) { /* XXX always ? */
8219 new_cpu = select_idle_sibling(p, prev_cpu, new_cpu);
8227 * Called immediately before a task is migrated to a new CPU; task_cpu(p) and
8228 * cfs_rq_of(p) references at time of call are still valid and identify the
8229 * previous CPU. The caller guarantees p->pi_lock or task_rq(p)->lock is held.
8231 static void migrate_task_rq_fair(struct task_struct *p, int new_cpu)
8233 struct sched_entity *se = &p->se;
8235 if (!task_on_rq_migrating(p)) {
8236 remove_entity_load_avg(se);
8239 * Here, the task's PELT values have been updated according to
8240 * the current rq's clock. But if that clock hasn't been
8241 * updated in a while, a substantial idle time will be missed,
8242 * leading to an inflation after wake-up on the new rq.
8244 * Estimate the missing time from the cfs_rq last_update_time
8245 * and update sched_avg to improve the PELT continuity after
8248 migrate_se_pelt_lag(se);
8251 /* Tell new CPU we are migrated */
8252 se->avg.last_update_time = 0;
8254 update_scan_period(p, new_cpu);
8257 static void task_dead_fair(struct task_struct *p)
8259 remove_entity_load_avg(&p->se);
8263 balance_fair(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
8268 return newidle_balance(rq, rf) != 0;
8270 #endif /* CONFIG_SMP */
8272 static void set_next_buddy(struct sched_entity *se)
8274 for_each_sched_entity(se) {
8275 if (SCHED_WARN_ON(!se->on_rq))
8279 cfs_rq_of(se)->next = se;
8284 * Preempt the current task with a newly woken task if needed:
8286 static void check_preempt_wakeup_fair(struct rq *rq, struct task_struct *p, int wake_flags)
8288 struct task_struct *curr = rq->curr;
8289 struct sched_entity *se = &curr->se, *pse = &p->se;
8290 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
8291 int cse_is_idle, pse_is_idle;
8293 if (unlikely(se == pse))
8297 * This is possible from callers such as attach_tasks(), in which we
8298 * unconditionally wakeup_preempt() after an enqueue (which may have
8299 * lead to a throttle). This both saves work and prevents false
8300 * next-buddy nomination below.
8302 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
8305 if (sched_feat(NEXT_BUDDY) && !(wake_flags & WF_FORK)) {
8306 set_next_buddy(pse);
8310 * We can come here with TIF_NEED_RESCHED already set from new task
8313 * Note: this also catches the edge-case of curr being in a throttled
8314 * group (e.g. via set_curr_task), since update_curr() (in the
8315 * enqueue of curr) will have resulted in resched being set. This
8316 * prevents us from potentially nominating it as a false LAST_BUDDY
8319 if (test_tsk_need_resched(curr))
8322 /* Idle tasks are by definition preempted by non-idle tasks. */
8323 if (unlikely(task_has_idle_policy(curr)) &&
8324 likely(!task_has_idle_policy(p)))
8328 * Batch and idle tasks do not preempt non-idle tasks (their preemption
8329 * is driven by the tick):
8331 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
8334 find_matching_se(&se, &pse);
8337 cse_is_idle = se_is_idle(se);
8338 pse_is_idle = se_is_idle(pse);
8341 * Preempt an idle group in favor of a non-idle group (and don't preempt
8342 * in the inverse case).
8344 if (cse_is_idle && !pse_is_idle)
8346 if (cse_is_idle != pse_is_idle)
8349 cfs_rq = cfs_rq_of(se);
8350 update_curr(cfs_rq);
8353 * XXX pick_eevdf(cfs_rq) != se ?
8355 if (pick_eevdf(cfs_rq) == pse)
8365 static struct task_struct *pick_task_fair(struct rq *rq)
8367 struct sched_entity *se;
8368 struct cfs_rq *cfs_rq;
8372 if (!cfs_rq->nr_running)
8376 struct sched_entity *curr = cfs_rq->curr;
8378 /* When we pick for a remote RQ, we'll not have done put_prev_entity() */
8381 update_curr(cfs_rq);
8385 if (unlikely(check_cfs_rq_runtime(cfs_rq)))
8389 se = pick_next_entity(cfs_rq);
8390 cfs_rq = group_cfs_rq(se);
8397 struct task_struct *
8398 pick_next_task_fair(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
8400 struct cfs_rq *cfs_rq = &rq->cfs;
8401 struct sched_entity *se;
8402 struct task_struct *p;
8406 if (!sched_fair_runnable(rq))
8409 #ifdef CONFIG_FAIR_GROUP_SCHED
8410 if (!prev || prev->sched_class != &fair_sched_class)
8414 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
8415 * likely that a next task is from the same cgroup as the current.
8417 * Therefore attempt to avoid putting and setting the entire cgroup
8418 * hierarchy, only change the part that actually changes.
8422 struct sched_entity *curr = cfs_rq->curr;
8425 * Since we got here without doing put_prev_entity() we also
8426 * have to consider cfs_rq->curr. If it is still a runnable
8427 * entity, update_curr() will update its vruntime, otherwise
8428 * forget we've ever seen it.
8432 update_curr(cfs_rq);
8437 * This call to check_cfs_rq_runtime() will do the
8438 * throttle and dequeue its entity in the parent(s).
8439 * Therefore the nr_running test will indeed
8442 if (unlikely(check_cfs_rq_runtime(cfs_rq))) {
8445 if (!cfs_rq->nr_running)
8452 se = pick_next_entity(cfs_rq);
8453 cfs_rq = group_cfs_rq(se);
8459 * Since we haven't yet done put_prev_entity and if the selected task
8460 * is a different task than we started out with, try and touch the
8461 * least amount of cfs_rqs.
8464 struct sched_entity *pse = &prev->se;
8466 while (!(cfs_rq = is_same_group(se, pse))) {
8467 int se_depth = se->depth;
8468 int pse_depth = pse->depth;
8470 if (se_depth <= pse_depth) {
8471 put_prev_entity(cfs_rq_of(pse), pse);
8472 pse = parent_entity(pse);
8474 if (se_depth >= pse_depth) {
8475 set_next_entity(cfs_rq_of(se), se);
8476 se = parent_entity(se);
8480 put_prev_entity(cfs_rq, pse);
8481 set_next_entity(cfs_rq, se);
8488 put_prev_task(rq, prev);
8491 se = pick_next_entity(cfs_rq);
8492 set_next_entity(cfs_rq, se);
8493 cfs_rq = group_cfs_rq(se);
8498 done: __maybe_unused;
8501 * Move the next running task to the front of
8502 * the list, so our cfs_tasks list becomes MRU
8505 list_move(&p->se.group_node, &rq->cfs_tasks);
8508 if (hrtick_enabled_fair(rq))
8509 hrtick_start_fair(rq, p);
8511 update_misfit_status(p, rq);
8512 sched_fair_update_stop_tick(rq, p);
8520 new_tasks = newidle_balance(rq, rf);
8523 * Because newidle_balance() releases (and re-acquires) rq->lock, it is
8524 * possible for any higher priority task to appear. In that case we
8525 * must re-start the pick_next_entity() loop.
8534 * rq is about to be idle, check if we need to update the
8535 * lost_idle_time of clock_pelt
8537 update_idle_rq_clock_pelt(rq);
8542 static struct task_struct *__pick_next_task_fair(struct rq *rq)
8544 return pick_next_task_fair(rq, NULL, NULL);
8548 * Account for a descheduled task:
8550 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
8552 struct sched_entity *se = &prev->se;
8553 struct cfs_rq *cfs_rq;
8555 for_each_sched_entity(se) {
8556 cfs_rq = cfs_rq_of(se);
8557 put_prev_entity(cfs_rq, se);
8562 * sched_yield() is very simple
8564 static void yield_task_fair(struct rq *rq)
8566 struct task_struct *curr = rq->curr;
8567 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
8568 struct sched_entity *se = &curr->se;
8571 * Are we the only task in the tree?
8573 if (unlikely(rq->nr_running == 1))
8576 clear_buddies(cfs_rq, se);
8578 update_rq_clock(rq);
8580 * Update run-time statistics of the 'current'.
8582 update_curr(cfs_rq);
8584 * Tell update_rq_clock() that we've just updated,
8585 * so we don't do microscopic update in schedule()
8586 * and double the fastpath cost.
8588 rq_clock_skip_update(rq);
8590 se->deadline += calc_delta_fair(se->slice, se);
8593 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p)
8595 struct sched_entity *se = &p->se;
8597 /* throttled hierarchies are not runnable */
8598 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
8601 /* Tell the scheduler that we'd really like pse to run next. */
8604 yield_task_fair(rq);
8610 /**************************************************
8611 * Fair scheduling class load-balancing methods.
8615 * The purpose of load-balancing is to achieve the same basic fairness the
8616 * per-CPU scheduler provides, namely provide a proportional amount of compute
8617 * time to each task. This is expressed in the following equation:
8619 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
8621 * Where W_i,n is the n-th weight average for CPU i. The instantaneous weight
8622 * W_i,0 is defined as:
8624 * W_i,0 = \Sum_j w_i,j (2)
8626 * Where w_i,j is the weight of the j-th runnable task on CPU i. This weight
8627 * is derived from the nice value as per sched_prio_to_weight[].
8629 * The weight average is an exponential decay average of the instantaneous
8632 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
8634 * C_i is the compute capacity of CPU i, typically it is the
8635 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
8636 * can also include other factors [XXX].
8638 * To achieve this balance we define a measure of imbalance which follows
8639 * directly from (1):
8641 * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4)
8643 * We them move tasks around to minimize the imbalance. In the continuous
8644 * function space it is obvious this converges, in the discrete case we get
8645 * a few fun cases generally called infeasible weight scenarios.
8648 * - infeasible weights;
8649 * - local vs global optima in the discrete case. ]
8654 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
8655 * for all i,j solution, we create a tree of CPUs that follows the hardware
8656 * topology where each level pairs two lower groups (or better). This results
8657 * in O(log n) layers. Furthermore we reduce the number of CPUs going up the
8658 * tree to only the first of the previous level and we decrease the frequency
8659 * of load-balance at each level inv. proportional to the number of CPUs in
8665 * \Sum { --- * --- * 2^i } = O(n) (5)
8667 * `- size of each group
8668 * | | `- number of CPUs doing load-balance
8670 * `- sum over all levels
8672 * Coupled with a limit on how many tasks we can migrate every balance pass,
8673 * this makes (5) the runtime complexity of the balancer.
8675 * An important property here is that each CPU is still (indirectly) connected
8676 * to every other CPU in at most O(log n) steps:
8678 * The adjacency matrix of the resulting graph is given by:
8681 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
8684 * And you'll find that:
8686 * A^(log_2 n)_i,j != 0 for all i,j (7)
8688 * Showing there's indeed a path between every CPU in at most O(log n) steps.
8689 * The task movement gives a factor of O(m), giving a convergence complexity
8692 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
8697 * In order to avoid CPUs going idle while there's still work to do, new idle
8698 * balancing is more aggressive and has the newly idle CPU iterate up the domain
8699 * tree itself instead of relying on other CPUs to bring it work.
8701 * This adds some complexity to both (5) and (8) but it reduces the total idle
8709 * Cgroups make a horror show out of (2), instead of a simple sum we get:
8712 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
8717 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
8719 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on CPU i.
8721 * The big problem is S_k, its a global sum needed to compute a local (W_i)
8724 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
8725 * rewrite all of this once again.]
8728 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
8730 enum fbq_type { regular, remote, all };
8733 * 'group_type' describes the group of CPUs at the moment of load balancing.
8735 * The enum is ordered by pulling priority, with the group with lowest priority
8736 * first so the group_type can simply be compared when selecting the busiest
8737 * group. See update_sd_pick_busiest().
8740 /* The group has spare capacity that can be used to run more tasks. */
8741 group_has_spare = 0,
8743 * The group is fully used and the tasks don't compete for more CPU
8744 * cycles. Nevertheless, some tasks might wait before running.
8748 * One task doesn't fit with CPU's capacity and must be migrated to a
8749 * more powerful CPU.
8753 * Balance SMT group that's fully busy. Can benefit from migration
8754 * a task on SMT with busy sibling to another CPU on idle core.
8758 * SD_ASYM_PACKING only: One local CPU with higher capacity is available,
8759 * and the task should be migrated to it instead of running on the
8764 * The tasks' affinity constraints previously prevented the scheduler
8765 * from balancing the load across the system.
8769 * The CPU is overloaded and can't provide expected CPU cycles to all
8775 enum migration_type {
8782 #define LBF_ALL_PINNED 0x01
8783 #define LBF_NEED_BREAK 0x02
8784 #define LBF_DST_PINNED 0x04
8785 #define LBF_SOME_PINNED 0x08
8786 #define LBF_ACTIVE_LB 0x10
8789 struct sched_domain *sd;
8797 struct cpumask *dst_grpmask;
8799 enum cpu_idle_type idle;
8801 /* The set of CPUs under consideration for load-balancing */
8802 struct cpumask *cpus;
8807 unsigned int loop_break;
8808 unsigned int loop_max;
8810 enum fbq_type fbq_type;
8811 enum migration_type migration_type;
8812 struct list_head tasks;
8816 * Is this task likely cache-hot:
8818 static int task_hot(struct task_struct *p, struct lb_env *env)
8822 lockdep_assert_rq_held(env->src_rq);
8824 if (p->sched_class != &fair_sched_class)
8827 if (unlikely(task_has_idle_policy(p)))
8830 /* SMT siblings share cache */
8831 if (env->sd->flags & SD_SHARE_CPUCAPACITY)
8835 * Buddy candidates are cache hot:
8837 if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running &&
8838 (&p->se == cfs_rq_of(&p->se)->next))
8841 if (sysctl_sched_migration_cost == -1)
8845 * Don't migrate task if the task's cookie does not match
8846 * with the destination CPU's core cookie.
8848 if (!sched_core_cookie_match(cpu_rq(env->dst_cpu), p))
8851 if (sysctl_sched_migration_cost == 0)
8854 delta = rq_clock_task(env->src_rq) - p->se.exec_start;
8856 return delta < (s64)sysctl_sched_migration_cost;
8859 #ifdef CONFIG_NUMA_BALANCING
8861 * Returns 1, if task migration degrades locality
8862 * Returns 0, if task migration improves locality i.e migration preferred.
8863 * Returns -1, if task migration is not affected by locality.
8865 static int migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
8867 struct numa_group *numa_group = rcu_dereference(p->numa_group);
8868 unsigned long src_weight, dst_weight;
8869 int src_nid, dst_nid, dist;
8871 if (!static_branch_likely(&sched_numa_balancing))
8874 if (!p->numa_faults || !(env->sd->flags & SD_NUMA))
8877 src_nid = cpu_to_node(env->src_cpu);
8878 dst_nid = cpu_to_node(env->dst_cpu);
8880 if (src_nid == dst_nid)
8883 /* Migrating away from the preferred node is always bad. */
8884 if (src_nid == p->numa_preferred_nid) {
8885 if (env->src_rq->nr_running > env->src_rq->nr_preferred_running)
8891 /* Encourage migration to the preferred node. */
8892 if (dst_nid == p->numa_preferred_nid)
8895 /* Leaving a core idle is often worse than degrading locality. */
8896 if (env->idle == CPU_IDLE)
8899 dist = node_distance(src_nid, dst_nid);
8901 src_weight = group_weight(p, src_nid, dist);
8902 dst_weight = group_weight(p, dst_nid, dist);
8904 src_weight = task_weight(p, src_nid, dist);
8905 dst_weight = task_weight(p, dst_nid, dist);
8908 return dst_weight < src_weight;
8912 static inline int migrate_degrades_locality(struct task_struct *p,
8920 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
8923 int can_migrate_task(struct task_struct *p, struct lb_env *env)
8927 lockdep_assert_rq_held(env->src_rq);
8930 * We do not migrate tasks that are:
8931 * 1) throttled_lb_pair, or
8932 * 2) cannot be migrated to this CPU due to cpus_ptr, or
8933 * 3) running (obviously), or
8934 * 4) are cache-hot on their current CPU.
8936 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
8939 /* Disregard pcpu kthreads; they are where they need to be. */
8940 if (kthread_is_per_cpu(p))
8943 if (!cpumask_test_cpu(env->dst_cpu, p->cpus_ptr)) {
8946 schedstat_inc(p->stats.nr_failed_migrations_affine);
8948 env->flags |= LBF_SOME_PINNED;
8951 * Remember if this task can be migrated to any other CPU in
8952 * our sched_group. We may want to revisit it if we couldn't
8953 * meet load balance goals by pulling other tasks on src_cpu.
8955 * Avoid computing new_dst_cpu
8957 * - if we have already computed one in current iteration
8958 * - if it's an active balance
8960 if (env->idle == CPU_NEWLY_IDLE ||
8961 env->flags & (LBF_DST_PINNED | LBF_ACTIVE_LB))
8964 /* Prevent to re-select dst_cpu via env's CPUs: */
8965 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
8966 if (cpumask_test_cpu(cpu, p->cpus_ptr)) {
8967 env->flags |= LBF_DST_PINNED;
8968 env->new_dst_cpu = cpu;
8976 /* Record that we found at least one task that could run on dst_cpu */
8977 env->flags &= ~LBF_ALL_PINNED;
8979 if (task_on_cpu(env->src_rq, p)) {
8980 schedstat_inc(p->stats.nr_failed_migrations_running);
8985 * Aggressive migration if:
8987 * 2) destination numa is preferred
8988 * 3) task is cache cold, or
8989 * 4) too many balance attempts have failed.
8991 if (env->flags & LBF_ACTIVE_LB)
8994 tsk_cache_hot = migrate_degrades_locality(p, env);
8995 if (tsk_cache_hot == -1)
8996 tsk_cache_hot = task_hot(p, env);
8998 if (tsk_cache_hot <= 0 ||
8999 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
9000 if (tsk_cache_hot == 1) {
9001 schedstat_inc(env->sd->lb_hot_gained[env->idle]);
9002 schedstat_inc(p->stats.nr_forced_migrations);
9007 schedstat_inc(p->stats.nr_failed_migrations_hot);
9012 * detach_task() -- detach the task for the migration specified in env
9014 static void detach_task(struct task_struct *p, struct lb_env *env)
9016 lockdep_assert_rq_held(env->src_rq);
9018 deactivate_task(env->src_rq, p, DEQUEUE_NOCLOCK);
9019 set_task_cpu(p, env->dst_cpu);
9023 * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
9024 * part of active balancing operations within "domain".
9026 * Returns a task if successful and NULL otherwise.
9028 static struct task_struct *detach_one_task(struct lb_env *env)
9030 struct task_struct *p;
9032 lockdep_assert_rq_held(env->src_rq);
9034 list_for_each_entry_reverse(p,
9035 &env->src_rq->cfs_tasks, se.group_node) {
9036 if (!can_migrate_task(p, env))
9039 detach_task(p, env);
9042 * Right now, this is only the second place where
9043 * lb_gained[env->idle] is updated (other is detach_tasks)
9044 * so we can safely collect stats here rather than
9045 * inside detach_tasks().
9047 schedstat_inc(env->sd->lb_gained[env->idle]);
9054 * detach_tasks() -- tries to detach up to imbalance load/util/tasks from
9055 * busiest_rq, as part of a balancing operation within domain "sd".
9057 * Returns number of detached tasks if successful and 0 otherwise.
9059 static int detach_tasks(struct lb_env *env)
9061 struct list_head *tasks = &env->src_rq->cfs_tasks;
9062 unsigned long util, load;
9063 struct task_struct *p;
9066 lockdep_assert_rq_held(env->src_rq);
9069 * Source run queue has been emptied by another CPU, clear
9070 * LBF_ALL_PINNED flag as we will not test any task.
9072 if (env->src_rq->nr_running <= 1) {
9073 env->flags &= ~LBF_ALL_PINNED;
9077 if (env->imbalance <= 0)
9080 while (!list_empty(tasks)) {
9082 * We don't want to steal all, otherwise we may be treated likewise,
9083 * which could at worst lead to a livelock crash.
9085 if (env->idle != CPU_NOT_IDLE && env->src_rq->nr_running <= 1)
9090 * We've more or less seen every task there is, call it quits
9091 * unless we haven't found any movable task yet.
9093 if (env->loop > env->loop_max &&
9094 !(env->flags & LBF_ALL_PINNED))
9097 /* take a breather every nr_migrate tasks */
9098 if (env->loop > env->loop_break) {
9099 env->loop_break += SCHED_NR_MIGRATE_BREAK;
9100 env->flags |= LBF_NEED_BREAK;
9104 p = list_last_entry(tasks, struct task_struct, se.group_node);
9106 if (!can_migrate_task(p, env))
9109 switch (env->migration_type) {
9112 * Depending of the number of CPUs and tasks and the
9113 * cgroup hierarchy, task_h_load() can return a null
9114 * value. Make sure that env->imbalance decreases
9115 * otherwise detach_tasks() will stop only after
9116 * detaching up to loop_max tasks.
9118 load = max_t(unsigned long, task_h_load(p), 1);
9120 if (sched_feat(LB_MIN) &&
9121 load < 16 && !env->sd->nr_balance_failed)
9125 * Make sure that we don't migrate too much load.
9126 * Nevertheless, let relax the constraint if
9127 * scheduler fails to find a good waiting task to
9130 if (shr_bound(load, env->sd->nr_balance_failed) > env->imbalance)
9133 env->imbalance -= load;
9137 util = task_util_est(p);
9139 if (shr_bound(util, env->sd->nr_balance_failed) > env->imbalance)
9142 env->imbalance -= util;
9149 case migrate_misfit:
9150 /* This is not a misfit task */
9151 if (task_fits_cpu(p, env->src_cpu))
9158 detach_task(p, env);
9159 list_add(&p->se.group_node, &env->tasks);
9163 #ifdef CONFIG_PREEMPTION
9165 * NEWIDLE balancing is a source of latency, so preemptible
9166 * kernels will stop after the first task is detached to minimize
9167 * the critical section.
9169 if (env->idle == CPU_NEWLY_IDLE)
9174 * We only want to steal up to the prescribed amount of
9177 if (env->imbalance <= 0)
9182 list_move(&p->se.group_node, tasks);
9186 * Right now, this is one of only two places we collect this stat
9187 * so we can safely collect detach_one_task() stats here rather
9188 * than inside detach_one_task().
9190 schedstat_add(env->sd->lb_gained[env->idle], detached);
9196 * attach_task() -- attach the task detached by detach_task() to its new rq.
9198 static void attach_task(struct rq *rq, struct task_struct *p)
9200 lockdep_assert_rq_held(rq);
9202 WARN_ON_ONCE(task_rq(p) != rq);
9203 activate_task(rq, p, ENQUEUE_NOCLOCK);
9204 wakeup_preempt(rq, p, 0);
9208 * attach_one_task() -- attaches the task returned from detach_one_task() to
9211 static void attach_one_task(struct rq *rq, struct task_struct *p)
9216 update_rq_clock(rq);
9222 * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
9225 static void attach_tasks(struct lb_env *env)
9227 struct list_head *tasks = &env->tasks;
9228 struct task_struct *p;
9231 rq_lock(env->dst_rq, &rf);
9232 update_rq_clock(env->dst_rq);
9234 while (!list_empty(tasks)) {
9235 p = list_first_entry(tasks, struct task_struct, se.group_node);
9236 list_del_init(&p->se.group_node);
9238 attach_task(env->dst_rq, p);
9241 rq_unlock(env->dst_rq, &rf);
9244 #ifdef CONFIG_NO_HZ_COMMON
9245 static inline bool cfs_rq_has_blocked(struct cfs_rq *cfs_rq)
9247 if (cfs_rq->avg.load_avg)
9250 if (cfs_rq->avg.util_avg)
9256 static inline bool others_have_blocked(struct rq *rq)
9258 if (cpu_util_rt(rq))
9261 if (cpu_util_dl(rq))
9264 if (thermal_load_avg(rq))
9267 if (cpu_util_irq(rq))
9273 static inline void update_blocked_load_tick(struct rq *rq)
9275 WRITE_ONCE(rq->last_blocked_load_update_tick, jiffies);
9278 static inline void update_blocked_load_status(struct rq *rq, bool has_blocked)
9281 rq->has_blocked_load = 0;
9284 static inline bool cfs_rq_has_blocked(struct cfs_rq *cfs_rq) { return false; }
9285 static inline bool others_have_blocked(struct rq *rq) { return false; }
9286 static inline void update_blocked_load_tick(struct rq *rq) {}
9287 static inline void update_blocked_load_status(struct rq *rq, bool has_blocked) {}
9290 static bool __update_blocked_others(struct rq *rq, bool *done)
9292 const struct sched_class *curr_class;
9293 u64 now = rq_clock_pelt(rq);
9294 unsigned long thermal_pressure;
9298 * update_load_avg() can call cpufreq_update_util(). Make sure that RT,
9299 * DL and IRQ signals have been updated before updating CFS.
9301 curr_class = rq->curr->sched_class;
9303 thermal_pressure = arch_scale_thermal_pressure(cpu_of(rq));
9305 decayed = update_rt_rq_load_avg(now, rq, curr_class == &rt_sched_class) |
9306 update_dl_rq_load_avg(now, rq, curr_class == &dl_sched_class) |
9307 update_thermal_load_avg(rq_clock_thermal(rq), rq, thermal_pressure) |
9308 update_irq_load_avg(rq, 0);
9310 if (others_have_blocked(rq))
9316 #ifdef CONFIG_FAIR_GROUP_SCHED
9318 static bool __update_blocked_fair(struct rq *rq, bool *done)
9320 struct cfs_rq *cfs_rq, *pos;
9321 bool decayed = false;
9322 int cpu = cpu_of(rq);
9325 * Iterates the task_group tree in a bottom up fashion, see
9326 * list_add_leaf_cfs_rq() for details.
9328 for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) {
9329 struct sched_entity *se;
9331 if (update_cfs_rq_load_avg(cfs_rq_clock_pelt(cfs_rq), cfs_rq)) {
9332 update_tg_load_avg(cfs_rq);
9334 if (cfs_rq->nr_running == 0)
9335 update_idle_cfs_rq_clock_pelt(cfs_rq);
9337 if (cfs_rq == &rq->cfs)
9341 /* Propagate pending load changes to the parent, if any: */
9342 se = cfs_rq->tg->se[cpu];
9343 if (se && !skip_blocked_update(se))
9344 update_load_avg(cfs_rq_of(se), se, UPDATE_TG);
9347 * There can be a lot of idle CPU cgroups. Don't let fully
9348 * decayed cfs_rqs linger on the list.
9350 if (cfs_rq_is_decayed(cfs_rq))
9351 list_del_leaf_cfs_rq(cfs_rq);
9353 /* Don't need periodic decay once load/util_avg are null */
9354 if (cfs_rq_has_blocked(cfs_rq))
9362 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
9363 * This needs to be done in a top-down fashion because the load of a child
9364 * group is a fraction of its parents load.
9366 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
9368 struct rq *rq = rq_of(cfs_rq);
9369 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
9370 unsigned long now = jiffies;
9373 if (cfs_rq->last_h_load_update == now)
9376 WRITE_ONCE(cfs_rq->h_load_next, NULL);
9377 for_each_sched_entity(se) {
9378 cfs_rq = cfs_rq_of(se);
9379 WRITE_ONCE(cfs_rq->h_load_next, se);
9380 if (cfs_rq->last_h_load_update == now)
9385 cfs_rq->h_load = cfs_rq_load_avg(cfs_rq);
9386 cfs_rq->last_h_load_update = now;
9389 while ((se = READ_ONCE(cfs_rq->h_load_next)) != NULL) {
9390 load = cfs_rq->h_load;
9391 load = div64_ul(load * se->avg.load_avg,
9392 cfs_rq_load_avg(cfs_rq) + 1);
9393 cfs_rq = group_cfs_rq(se);
9394 cfs_rq->h_load = load;
9395 cfs_rq->last_h_load_update = now;
9399 static unsigned long task_h_load(struct task_struct *p)
9401 struct cfs_rq *cfs_rq = task_cfs_rq(p);
9403 update_cfs_rq_h_load(cfs_rq);
9404 return div64_ul(p->se.avg.load_avg * cfs_rq->h_load,
9405 cfs_rq_load_avg(cfs_rq) + 1);
9408 static bool __update_blocked_fair(struct rq *rq, bool *done)
9410 struct cfs_rq *cfs_rq = &rq->cfs;
9413 decayed = update_cfs_rq_load_avg(cfs_rq_clock_pelt(cfs_rq), cfs_rq);
9414 if (cfs_rq_has_blocked(cfs_rq))
9420 static unsigned long task_h_load(struct task_struct *p)
9422 return p->se.avg.load_avg;
9426 static void update_blocked_averages(int cpu)
9428 bool decayed = false, done = true;
9429 struct rq *rq = cpu_rq(cpu);
9432 rq_lock_irqsave(rq, &rf);
9433 update_blocked_load_tick(rq);
9434 update_rq_clock(rq);
9436 decayed |= __update_blocked_others(rq, &done);
9437 decayed |= __update_blocked_fair(rq, &done);
9439 update_blocked_load_status(rq, !done);
9441 cpufreq_update_util(rq, 0);
9442 rq_unlock_irqrestore(rq, &rf);
9445 /********** Helpers for find_busiest_group ************************/
9448 * sg_lb_stats - stats of a sched_group required for load_balancing
9450 struct sg_lb_stats {
9451 unsigned long avg_load; /*Avg load across the CPUs of the group */
9452 unsigned long group_load; /* Total load over the CPUs of the group */
9453 unsigned long group_capacity;
9454 unsigned long group_util; /* Total utilization over the CPUs of the group */
9455 unsigned long group_runnable; /* Total runnable time over the CPUs of the group */
9456 unsigned int sum_nr_running; /* Nr of tasks running in the group */
9457 unsigned int sum_h_nr_running; /* Nr of CFS tasks running in the group */
9458 unsigned int idle_cpus;
9459 unsigned int group_weight;
9460 enum group_type group_type;
9461 unsigned int group_asym_packing; /* Tasks should be moved to preferred CPU */
9462 unsigned int group_smt_balance; /* Task on busy SMT be moved */
9463 unsigned long group_misfit_task_load; /* A CPU has a task too big for its capacity */
9464 #ifdef CONFIG_NUMA_BALANCING
9465 unsigned int nr_numa_running;
9466 unsigned int nr_preferred_running;
9471 * sd_lb_stats - Structure to store the statistics of a sched_domain
9472 * during load balancing.
9474 struct sd_lb_stats {
9475 struct sched_group *busiest; /* Busiest group in this sd */
9476 struct sched_group *local; /* Local group in this sd */
9477 unsigned long total_load; /* Total load of all groups in sd */
9478 unsigned long total_capacity; /* Total capacity of all groups in sd */
9479 unsigned long avg_load; /* Average load across all groups in sd */
9480 unsigned int prefer_sibling; /* tasks should go to sibling first */
9482 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
9483 struct sg_lb_stats local_stat; /* Statistics of the local group */
9486 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
9489 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
9490 * local_stat because update_sg_lb_stats() does a full clear/assignment.
9491 * We must however set busiest_stat::group_type and
9492 * busiest_stat::idle_cpus to the worst busiest group because
9493 * update_sd_pick_busiest() reads these before assignment.
9495 *sds = (struct sd_lb_stats){
9499 .total_capacity = 0UL,
9501 .idle_cpus = UINT_MAX,
9502 .group_type = group_has_spare,
9507 static unsigned long scale_rt_capacity(int cpu)
9509 struct rq *rq = cpu_rq(cpu);
9510 unsigned long max = arch_scale_cpu_capacity(cpu);
9511 unsigned long used, free;
9514 irq = cpu_util_irq(rq);
9516 if (unlikely(irq >= max))
9520 * avg_rt.util_avg and avg_dl.util_avg track binary signals
9521 * (running and not running) with weights 0 and 1024 respectively.
9522 * avg_thermal.load_avg tracks thermal pressure and the weighted
9523 * average uses the actual delta max capacity(load).
9525 used = cpu_util_rt(rq);
9526 used += cpu_util_dl(rq);
9527 used += thermal_load_avg(rq);
9529 if (unlikely(used >= max))
9534 return scale_irq_capacity(free, irq, max);
9537 static void update_cpu_capacity(struct sched_domain *sd, int cpu)
9539 unsigned long capacity = scale_rt_capacity(cpu);
9540 struct sched_group *sdg = sd->groups;
9545 cpu_rq(cpu)->cpu_capacity = capacity;
9546 trace_sched_cpu_capacity_tp(cpu_rq(cpu));
9548 sdg->sgc->capacity = capacity;
9549 sdg->sgc->min_capacity = capacity;
9550 sdg->sgc->max_capacity = capacity;
9553 void update_group_capacity(struct sched_domain *sd, int cpu)
9555 struct sched_domain *child = sd->child;
9556 struct sched_group *group, *sdg = sd->groups;
9557 unsigned long capacity, min_capacity, max_capacity;
9558 unsigned long interval;
9560 interval = msecs_to_jiffies(sd->balance_interval);
9561 interval = clamp(interval, 1UL, max_load_balance_interval);
9562 sdg->sgc->next_update = jiffies + interval;
9565 update_cpu_capacity(sd, cpu);
9570 min_capacity = ULONG_MAX;
9573 if (child->flags & SD_OVERLAP) {
9575 * SD_OVERLAP domains cannot assume that child groups
9576 * span the current group.
9579 for_each_cpu(cpu, sched_group_span(sdg)) {
9580 unsigned long cpu_cap = capacity_of(cpu);
9582 capacity += cpu_cap;
9583 min_capacity = min(cpu_cap, min_capacity);
9584 max_capacity = max(cpu_cap, max_capacity);
9588 * !SD_OVERLAP domains can assume that child groups
9589 * span the current group.
9592 group = child->groups;
9594 struct sched_group_capacity *sgc = group->sgc;
9596 capacity += sgc->capacity;
9597 min_capacity = min(sgc->min_capacity, min_capacity);
9598 max_capacity = max(sgc->max_capacity, max_capacity);
9599 group = group->next;
9600 } while (group != child->groups);
9603 sdg->sgc->capacity = capacity;
9604 sdg->sgc->min_capacity = min_capacity;
9605 sdg->sgc->max_capacity = max_capacity;
9609 * Check whether the capacity of the rq has been noticeably reduced by side
9610 * activity. The imbalance_pct is used for the threshold.
9611 * Return true is the capacity is reduced
9614 check_cpu_capacity(struct rq *rq, struct sched_domain *sd)
9616 return ((rq->cpu_capacity * sd->imbalance_pct) <
9617 (arch_scale_cpu_capacity(cpu_of(rq)) * 100));
9621 * Check whether a rq has a misfit task and if it looks like we can actually
9622 * help that task: we can migrate the task to a CPU of higher capacity, or
9623 * the task's current CPU is heavily pressured.
9625 static inline int check_misfit_status(struct rq *rq, struct sched_domain *sd)
9627 return rq->misfit_task_load &&
9628 (arch_scale_cpu_capacity(rq->cpu) < rq->rd->max_cpu_capacity ||
9629 check_cpu_capacity(rq, sd));
9633 * Group imbalance indicates (and tries to solve) the problem where balancing
9634 * groups is inadequate due to ->cpus_ptr constraints.
9636 * Imagine a situation of two groups of 4 CPUs each and 4 tasks each with a
9637 * cpumask covering 1 CPU of the first group and 3 CPUs of the second group.
9640 * { 0 1 2 3 } { 4 5 6 7 }
9643 * If we were to balance group-wise we'd place two tasks in the first group and
9644 * two tasks in the second group. Clearly this is undesired as it will overload
9645 * cpu 3 and leave one of the CPUs in the second group unused.
9647 * The current solution to this issue is detecting the skew in the first group
9648 * by noticing the lower domain failed to reach balance and had difficulty
9649 * moving tasks due to affinity constraints.
9651 * When this is so detected; this group becomes a candidate for busiest; see
9652 * update_sd_pick_busiest(). And calculate_imbalance() and
9653 * find_busiest_group() avoid some of the usual balance conditions to allow it
9654 * to create an effective group imbalance.
9656 * This is a somewhat tricky proposition since the next run might not find the
9657 * group imbalance and decide the groups need to be balanced again. A most
9658 * subtle and fragile situation.
9661 static inline int sg_imbalanced(struct sched_group *group)
9663 return group->sgc->imbalance;
9667 * group_has_capacity returns true if the group has spare capacity that could
9668 * be used by some tasks.
9669 * We consider that a group has spare capacity if the number of task is
9670 * smaller than the number of CPUs or if the utilization is lower than the
9671 * available capacity for CFS tasks.
9672 * For the latter, we use a threshold to stabilize the state, to take into
9673 * account the variance of the tasks' load and to return true if the available
9674 * capacity in meaningful for the load balancer.
9675 * As an example, an available capacity of 1% can appear but it doesn't make
9676 * any benefit for the load balance.
9679 group_has_capacity(unsigned int imbalance_pct, struct sg_lb_stats *sgs)
9681 if (sgs->sum_nr_running < sgs->group_weight)
9684 if ((sgs->group_capacity * imbalance_pct) <
9685 (sgs->group_runnable * 100))
9688 if ((sgs->group_capacity * 100) >
9689 (sgs->group_util * imbalance_pct))
9696 * group_is_overloaded returns true if the group has more tasks than it can
9698 * group_is_overloaded is not equals to !group_has_capacity because a group
9699 * with the exact right number of tasks, has no more spare capacity but is not
9700 * overloaded so both group_has_capacity and group_is_overloaded return
9704 group_is_overloaded(unsigned int imbalance_pct, struct sg_lb_stats *sgs)
9706 if (sgs->sum_nr_running <= sgs->group_weight)
9709 if ((sgs->group_capacity * 100) <
9710 (sgs->group_util * imbalance_pct))
9713 if ((sgs->group_capacity * imbalance_pct) <
9714 (sgs->group_runnable * 100))
9721 group_type group_classify(unsigned int imbalance_pct,
9722 struct sched_group *group,
9723 struct sg_lb_stats *sgs)
9725 if (group_is_overloaded(imbalance_pct, sgs))
9726 return group_overloaded;
9728 if (sg_imbalanced(group))
9729 return group_imbalanced;
9731 if (sgs->group_asym_packing)
9732 return group_asym_packing;
9734 if (sgs->group_smt_balance)
9735 return group_smt_balance;
9737 if (sgs->group_misfit_task_load)
9738 return group_misfit_task;
9740 if (!group_has_capacity(imbalance_pct, sgs))
9741 return group_fully_busy;
9743 return group_has_spare;
9747 * sched_use_asym_prio - Check whether asym_packing priority must be used
9748 * @sd: The scheduling domain of the load balancing
9751 * Always use CPU priority when balancing load between SMT siblings. When
9752 * balancing load between cores, it is not sufficient that @cpu is idle. Only
9753 * use CPU priority if the whole core is idle.
9755 * Returns: True if the priority of @cpu must be followed. False otherwise.
9757 static bool sched_use_asym_prio(struct sched_domain *sd, int cpu)
9759 if (!(sd->flags & SD_ASYM_PACKING))
9762 if (!sched_smt_active())
9765 return sd->flags & SD_SHARE_CPUCAPACITY || is_core_idle(cpu);
9768 static inline bool sched_asym(struct sched_domain *sd, int dst_cpu, int src_cpu)
9771 * First check if @dst_cpu can do asym_packing load balance. Only do it
9772 * if it has higher priority than @src_cpu.
9774 return sched_use_asym_prio(sd, dst_cpu) &&
9775 sched_asym_prefer(dst_cpu, src_cpu);
9779 * sched_group_asym - Check if the destination CPU can do asym_packing balance
9780 * @env: The load balancing environment
9781 * @sgs: Load-balancing statistics of the candidate busiest group
9782 * @group: The candidate busiest group
9784 * @env::dst_cpu can do asym_packing if it has higher priority than the
9785 * preferred CPU of @group.
9787 * Return: true if @env::dst_cpu can do with asym_packing load balance. False
9791 sched_group_asym(struct lb_env *env, struct sg_lb_stats *sgs, struct sched_group *group)
9794 * CPU priorities do not make sense for SMT cores with more than one
9797 if ((group->flags & SD_SHARE_CPUCAPACITY) &&
9798 (sgs->group_weight - sgs->idle_cpus != 1))
9801 return sched_asym(env->sd, env->dst_cpu, group->asym_prefer_cpu);
9804 /* One group has more than one SMT CPU while the other group does not */
9805 static inline bool smt_vs_nonsmt_groups(struct sched_group *sg1,
9806 struct sched_group *sg2)
9811 return (sg1->flags & SD_SHARE_CPUCAPACITY) !=
9812 (sg2->flags & SD_SHARE_CPUCAPACITY);
9815 static inline bool smt_balance(struct lb_env *env, struct sg_lb_stats *sgs,
9816 struct sched_group *group)
9818 if (env->idle == CPU_NOT_IDLE)
9822 * For SMT source group, it is better to move a task
9823 * to a CPU that doesn't have multiple tasks sharing its CPU capacity.
9824 * Note that if a group has a single SMT, SD_SHARE_CPUCAPACITY
9827 if (group->flags & SD_SHARE_CPUCAPACITY &&
9828 sgs->sum_h_nr_running > 1)
9834 static inline long sibling_imbalance(struct lb_env *env,
9835 struct sd_lb_stats *sds,
9836 struct sg_lb_stats *busiest,
9837 struct sg_lb_stats *local)
9839 int ncores_busiest, ncores_local;
9842 if (env->idle == CPU_NOT_IDLE || !busiest->sum_nr_running)
9845 ncores_busiest = sds->busiest->cores;
9846 ncores_local = sds->local->cores;
9848 if (ncores_busiest == ncores_local) {
9849 imbalance = busiest->sum_nr_running;
9850 lsub_positive(&imbalance, local->sum_nr_running);
9854 /* Balance such that nr_running/ncores ratio are same on both groups */
9855 imbalance = ncores_local * busiest->sum_nr_running;
9856 lsub_positive(&imbalance, ncores_busiest * local->sum_nr_running);
9857 /* Normalize imbalance and do rounding on normalization */
9858 imbalance = 2 * imbalance + ncores_local + ncores_busiest;
9859 imbalance /= ncores_local + ncores_busiest;
9861 /* Take advantage of resource in an empty sched group */
9862 if (imbalance <= 1 && local->sum_nr_running == 0 &&
9863 busiest->sum_nr_running > 1)
9870 sched_reduced_capacity(struct rq *rq, struct sched_domain *sd)
9873 * When there is more than 1 task, the group_overloaded case already
9874 * takes care of cpu with reduced capacity
9876 if (rq->cfs.h_nr_running != 1)
9879 return check_cpu_capacity(rq, sd);
9883 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
9884 * @env: The load balancing environment.
9885 * @sds: Load-balancing data with statistics of the local group.
9886 * @group: sched_group whose statistics are to be updated.
9887 * @sgs: variable to hold the statistics for this group.
9888 * @sg_status: Holds flag indicating the status of the sched_group
9890 static inline void update_sg_lb_stats(struct lb_env *env,
9891 struct sd_lb_stats *sds,
9892 struct sched_group *group,
9893 struct sg_lb_stats *sgs,
9896 int i, nr_running, local_group;
9898 memset(sgs, 0, sizeof(*sgs));
9900 local_group = group == sds->local;
9902 for_each_cpu_and(i, sched_group_span(group), env->cpus) {
9903 struct rq *rq = cpu_rq(i);
9904 unsigned long load = cpu_load(rq);
9906 sgs->group_load += load;
9907 sgs->group_util += cpu_util_cfs(i);
9908 sgs->group_runnable += cpu_runnable(rq);
9909 sgs->sum_h_nr_running += rq->cfs.h_nr_running;
9911 nr_running = rq->nr_running;
9912 sgs->sum_nr_running += nr_running;
9915 *sg_status |= SG_OVERLOAD;
9917 if (cpu_overutilized(i))
9918 *sg_status |= SG_OVERUTILIZED;
9920 #ifdef CONFIG_NUMA_BALANCING
9921 sgs->nr_numa_running += rq->nr_numa_running;
9922 sgs->nr_preferred_running += rq->nr_preferred_running;
9925 * No need to call idle_cpu() if nr_running is not 0
9927 if (!nr_running && idle_cpu(i)) {
9929 /* Idle cpu can't have misfit task */
9936 if (env->sd->flags & SD_ASYM_CPUCAPACITY) {
9937 /* Check for a misfit task on the cpu */
9938 if (sgs->group_misfit_task_load < rq->misfit_task_load) {
9939 sgs->group_misfit_task_load = rq->misfit_task_load;
9940 *sg_status |= SG_OVERLOAD;
9942 } else if ((env->idle != CPU_NOT_IDLE) &&
9943 sched_reduced_capacity(rq, env->sd)) {
9944 /* Check for a task running on a CPU with reduced capacity */
9945 if (sgs->group_misfit_task_load < load)
9946 sgs->group_misfit_task_load = load;
9950 sgs->group_capacity = group->sgc->capacity;
9952 sgs->group_weight = group->group_weight;
9954 /* Check if dst CPU is idle and preferred to this group */
9955 if (!local_group && env->idle != CPU_NOT_IDLE && sgs->sum_h_nr_running &&
9956 sched_group_asym(env, sgs, group))
9957 sgs->group_asym_packing = 1;
9959 /* Check for loaded SMT group to be balanced to dst CPU */
9960 if (!local_group && smt_balance(env, sgs, group))
9961 sgs->group_smt_balance = 1;
9963 sgs->group_type = group_classify(env->sd->imbalance_pct, group, sgs);
9965 /* Computing avg_load makes sense only when group is overloaded */
9966 if (sgs->group_type == group_overloaded)
9967 sgs->avg_load = (sgs->group_load * SCHED_CAPACITY_SCALE) /
9968 sgs->group_capacity;
9972 * update_sd_pick_busiest - return 1 on busiest group
9973 * @env: The load balancing environment.
9974 * @sds: sched_domain statistics
9975 * @sg: sched_group candidate to be checked for being the busiest
9976 * @sgs: sched_group statistics
9978 * Determine if @sg is a busier group than the previously selected
9981 * Return: %true if @sg is a busier group than the previously selected
9982 * busiest group. %false otherwise.
9984 static bool update_sd_pick_busiest(struct lb_env *env,
9985 struct sd_lb_stats *sds,
9986 struct sched_group *sg,
9987 struct sg_lb_stats *sgs)
9989 struct sg_lb_stats *busiest = &sds->busiest_stat;
9991 /* Make sure that there is at least one task to pull */
9992 if (!sgs->sum_h_nr_running)
9996 * Don't try to pull misfit tasks we can't help.
9997 * We can use max_capacity here as reduction in capacity on some
9998 * CPUs in the group should either be possible to resolve
9999 * internally or be covered by avg_load imbalance (eventually).
10001 if ((env->sd->flags & SD_ASYM_CPUCAPACITY) &&
10002 (sgs->group_type == group_misfit_task) &&
10003 (!capacity_greater(capacity_of(env->dst_cpu), sg->sgc->max_capacity) ||
10004 sds->local_stat.group_type != group_has_spare))
10007 if (sgs->group_type > busiest->group_type)
10010 if (sgs->group_type < busiest->group_type)
10014 * The candidate and the current busiest group are the same type of
10015 * group. Let check which one is the busiest according to the type.
10018 switch (sgs->group_type) {
10019 case group_overloaded:
10020 /* Select the overloaded group with highest avg_load. */
10021 return sgs->avg_load > busiest->avg_load;
10023 case group_imbalanced:
10025 * Select the 1st imbalanced group as we don't have any way to
10026 * choose one more than another.
10030 case group_asym_packing:
10031 /* Prefer to move from lowest priority CPU's work */
10032 return sched_asym_prefer(sds->busiest->asym_prefer_cpu, sg->asym_prefer_cpu);
10034 case group_misfit_task:
10036 * If we have more than one misfit sg go with the biggest
10039 return sgs->group_misfit_task_load > busiest->group_misfit_task_load;
10041 case group_smt_balance:
10043 * Check if we have spare CPUs on either SMT group to
10044 * choose has spare or fully busy handling.
10046 if (sgs->idle_cpus != 0 || busiest->idle_cpus != 0)
10051 case group_fully_busy:
10053 * Select the fully busy group with highest avg_load. In
10054 * theory, there is no need to pull task from such kind of
10055 * group because tasks have all compute capacity that they need
10056 * but we can still improve the overall throughput by reducing
10057 * contention when accessing shared HW resources.
10059 * XXX for now avg_load is not computed and always 0 so we
10060 * select the 1st one, except if @sg is composed of SMT
10064 if (sgs->avg_load < busiest->avg_load)
10067 if (sgs->avg_load == busiest->avg_load) {
10069 * SMT sched groups need more help than non-SMT groups.
10070 * If @sg happens to also be SMT, either choice is good.
10072 if (sds->busiest->flags & SD_SHARE_CPUCAPACITY)
10078 case group_has_spare:
10080 * Do not pick sg with SMT CPUs over sg with pure CPUs,
10081 * as we do not want to pull task off SMT core with one task
10082 * and make the core idle.
10084 if (smt_vs_nonsmt_groups(sds->busiest, sg)) {
10085 if (sg->flags & SD_SHARE_CPUCAPACITY && sgs->sum_h_nr_running <= 1)
10093 * Select not overloaded group with lowest number of idle cpus
10094 * and highest number of running tasks. We could also compare
10095 * the spare capacity which is more stable but it can end up
10096 * that the group has less spare capacity but finally more idle
10097 * CPUs which means less opportunity to pull tasks.
10099 if (sgs->idle_cpus > busiest->idle_cpus)
10101 else if ((sgs->idle_cpus == busiest->idle_cpus) &&
10102 (sgs->sum_nr_running <= busiest->sum_nr_running))
10109 * Candidate sg has no more than one task per CPU and has higher
10110 * per-CPU capacity. Migrating tasks to less capable CPUs may harm
10111 * throughput. Maximize throughput, power/energy consequences are not
10114 if ((env->sd->flags & SD_ASYM_CPUCAPACITY) &&
10115 (sgs->group_type <= group_fully_busy) &&
10116 (capacity_greater(sg->sgc->min_capacity, capacity_of(env->dst_cpu))))
10122 #ifdef CONFIG_NUMA_BALANCING
10123 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
10125 if (sgs->sum_h_nr_running > sgs->nr_numa_running)
10127 if (sgs->sum_h_nr_running > sgs->nr_preferred_running)
10132 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
10134 if (rq->nr_running > rq->nr_numa_running)
10136 if (rq->nr_running > rq->nr_preferred_running)
10141 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
10146 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
10150 #endif /* CONFIG_NUMA_BALANCING */
10153 struct sg_lb_stats;
10156 * task_running_on_cpu - return 1 if @p is running on @cpu.
10159 static unsigned int task_running_on_cpu(int cpu, struct task_struct *p)
10161 /* Task has no contribution or is new */
10162 if (cpu != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time))
10165 if (task_on_rq_queued(p))
10172 * idle_cpu_without - would a given CPU be idle without p ?
10173 * @cpu: the processor on which idleness is tested.
10174 * @p: task which should be ignored.
10176 * Return: 1 if the CPU would be idle. 0 otherwise.
10178 static int idle_cpu_without(int cpu, struct task_struct *p)
10180 struct rq *rq = cpu_rq(cpu);
10182 if (rq->curr != rq->idle && rq->curr != p)
10186 * rq->nr_running can't be used but an updated version without the
10187 * impact of p on cpu must be used instead. The updated nr_running
10188 * be computed and tested before calling idle_cpu_without().
10191 if (rq->ttwu_pending)
10198 * update_sg_wakeup_stats - Update sched_group's statistics for wakeup.
10199 * @sd: The sched_domain level to look for idlest group.
10200 * @group: sched_group whose statistics are to be updated.
10201 * @sgs: variable to hold the statistics for this group.
10202 * @p: The task for which we look for the idlest group/CPU.
10204 static inline void update_sg_wakeup_stats(struct sched_domain *sd,
10205 struct sched_group *group,
10206 struct sg_lb_stats *sgs,
10207 struct task_struct *p)
10211 memset(sgs, 0, sizeof(*sgs));
10213 /* Assume that task can't fit any CPU of the group */
10214 if (sd->flags & SD_ASYM_CPUCAPACITY)
10215 sgs->group_misfit_task_load = 1;
10217 for_each_cpu(i, sched_group_span(group)) {
10218 struct rq *rq = cpu_rq(i);
10219 unsigned int local;
10221 sgs->group_load += cpu_load_without(rq, p);
10222 sgs->group_util += cpu_util_without(i, p);
10223 sgs->group_runnable += cpu_runnable_without(rq, p);
10224 local = task_running_on_cpu(i, p);
10225 sgs->sum_h_nr_running += rq->cfs.h_nr_running - local;
10227 nr_running = rq->nr_running - local;
10228 sgs->sum_nr_running += nr_running;
10231 * No need to call idle_cpu_without() if nr_running is not 0
10233 if (!nr_running && idle_cpu_without(i, p))
10236 /* Check if task fits in the CPU */
10237 if (sd->flags & SD_ASYM_CPUCAPACITY &&
10238 sgs->group_misfit_task_load &&
10239 task_fits_cpu(p, i))
10240 sgs->group_misfit_task_load = 0;
10244 sgs->group_capacity = group->sgc->capacity;
10246 sgs->group_weight = group->group_weight;
10248 sgs->group_type = group_classify(sd->imbalance_pct, group, sgs);
10251 * Computing avg_load makes sense only when group is fully busy or
10254 if (sgs->group_type == group_fully_busy ||
10255 sgs->group_type == group_overloaded)
10256 sgs->avg_load = (sgs->group_load * SCHED_CAPACITY_SCALE) /
10257 sgs->group_capacity;
10260 static bool update_pick_idlest(struct sched_group *idlest,
10261 struct sg_lb_stats *idlest_sgs,
10262 struct sched_group *group,
10263 struct sg_lb_stats *sgs)
10265 if (sgs->group_type < idlest_sgs->group_type)
10268 if (sgs->group_type > idlest_sgs->group_type)
10272 * The candidate and the current idlest group are the same type of
10273 * group. Let check which one is the idlest according to the type.
10276 switch (sgs->group_type) {
10277 case group_overloaded:
10278 case group_fully_busy:
10279 /* Select the group with lowest avg_load. */
10280 if (idlest_sgs->avg_load <= sgs->avg_load)
10284 case group_imbalanced:
10285 case group_asym_packing:
10286 case group_smt_balance:
10287 /* Those types are not used in the slow wakeup path */
10290 case group_misfit_task:
10291 /* Select group with the highest max capacity */
10292 if (idlest->sgc->max_capacity >= group->sgc->max_capacity)
10296 case group_has_spare:
10297 /* Select group with most idle CPUs */
10298 if (idlest_sgs->idle_cpus > sgs->idle_cpus)
10301 /* Select group with lowest group_util */
10302 if (idlest_sgs->idle_cpus == sgs->idle_cpus &&
10303 idlest_sgs->group_util <= sgs->group_util)
10313 * find_idlest_group() finds and returns the least busy CPU group within the
10316 * Assumes p is allowed on at least one CPU in sd.
10318 static struct sched_group *
10319 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
10321 struct sched_group *idlest = NULL, *local = NULL, *group = sd->groups;
10322 struct sg_lb_stats local_sgs, tmp_sgs;
10323 struct sg_lb_stats *sgs;
10324 unsigned long imbalance;
10325 struct sg_lb_stats idlest_sgs = {
10326 .avg_load = UINT_MAX,
10327 .group_type = group_overloaded,
10333 /* Skip over this group if it has no CPUs allowed */
10334 if (!cpumask_intersects(sched_group_span(group),
10338 /* Skip over this group if no cookie matched */
10339 if (!sched_group_cookie_match(cpu_rq(this_cpu), p, group))
10342 local_group = cpumask_test_cpu(this_cpu,
10343 sched_group_span(group));
10352 update_sg_wakeup_stats(sd, group, sgs, p);
10354 if (!local_group && update_pick_idlest(idlest, &idlest_sgs, group, sgs)) {
10359 } while (group = group->next, group != sd->groups);
10362 /* There is no idlest group to push tasks to */
10366 /* The local group has been skipped because of CPU affinity */
10371 * If the local group is idler than the selected idlest group
10372 * don't try and push the task.
10374 if (local_sgs.group_type < idlest_sgs.group_type)
10378 * If the local group is busier than the selected idlest group
10379 * try and push the task.
10381 if (local_sgs.group_type > idlest_sgs.group_type)
10384 switch (local_sgs.group_type) {
10385 case group_overloaded:
10386 case group_fully_busy:
10388 /* Calculate allowed imbalance based on load */
10389 imbalance = scale_load_down(NICE_0_LOAD) *
10390 (sd->imbalance_pct-100) / 100;
10393 * When comparing groups across NUMA domains, it's possible for
10394 * the local domain to be very lightly loaded relative to the
10395 * remote domains but "imbalance" skews the comparison making
10396 * remote CPUs look much more favourable. When considering
10397 * cross-domain, add imbalance to the load on the remote node
10398 * and consider staying local.
10401 if ((sd->flags & SD_NUMA) &&
10402 ((idlest_sgs.avg_load + imbalance) >= local_sgs.avg_load))
10406 * If the local group is less loaded than the selected
10407 * idlest group don't try and push any tasks.
10409 if (idlest_sgs.avg_load >= (local_sgs.avg_load + imbalance))
10412 if (100 * local_sgs.avg_load <= sd->imbalance_pct * idlest_sgs.avg_load)
10416 case group_imbalanced:
10417 case group_asym_packing:
10418 case group_smt_balance:
10419 /* Those type are not used in the slow wakeup path */
10422 case group_misfit_task:
10423 /* Select group with the highest max capacity */
10424 if (local->sgc->max_capacity >= idlest->sgc->max_capacity)
10428 case group_has_spare:
10430 if (sd->flags & SD_NUMA) {
10431 int imb_numa_nr = sd->imb_numa_nr;
10432 #ifdef CONFIG_NUMA_BALANCING
10435 * If there is spare capacity at NUMA, try to select
10436 * the preferred node
10438 if (cpu_to_node(this_cpu) == p->numa_preferred_nid)
10441 idlest_cpu = cpumask_first(sched_group_span(idlest));
10442 if (cpu_to_node(idlest_cpu) == p->numa_preferred_nid)
10444 #endif /* CONFIG_NUMA_BALANCING */
10446 * Otherwise, keep the task close to the wakeup source
10447 * and improve locality if the number of running tasks
10448 * would remain below threshold where an imbalance is
10449 * allowed while accounting for the possibility the
10450 * task is pinned to a subset of CPUs. If there is a
10451 * real need of migration, periodic load balance will
10454 if (p->nr_cpus_allowed != NR_CPUS) {
10455 struct cpumask *cpus = this_cpu_cpumask_var_ptr(select_rq_mask);
10457 cpumask_and(cpus, sched_group_span(local), p->cpus_ptr);
10458 imb_numa_nr = min(cpumask_weight(cpus), sd->imb_numa_nr);
10461 imbalance = abs(local_sgs.idle_cpus - idlest_sgs.idle_cpus);
10462 if (!adjust_numa_imbalance(imbalance,
10463 local_sgs.sum_nr_running + 1,
10468 #endif /* CONFIG_NUMA */
10471 * Select group with highest number of idle CPUs. We could also
10472 * compare the utilization which is more stable but it can end
10473 * up that the group has less spare capacity but finally more
10474 * idle CPUs which means more opportunity to run task.
10476 if (local_sgs.idle_cpus >= idlest_sgs.idle_cpus)
10484 static void update_idle_cpu_scan(struct lb_env *env,
10485 unsigned long sum_util)
10487 struct sched_domain_shared *sd_share;
10488 int llc_weight, pct;
10491 * Update the number of CPUs to scan in LLC domain, which could
10492 * be used as a hint in select_idle_cpu(). The update of sd_share
10493 * could be expensive because it is within a shared cache line.
10494 * So the write of this hint only occurs during periodic load
10495 * balancing, rather than CPU_NEWLY_IDLE, because the latter
10496 * can fire way more frequently than the former.
10498 if (!sched_feat(SIS_UTIL) || env->idle == CPU_NEWLY_IDLE)
10501 llc_weight = per_cpu(sd_llc_size, env->dst_cpu);
10502 if (env->sd->span_weight != llc_weight)
10505 sd_share = rcu_dereference(per_cpu(sd_llc_shared, env->dst_cpu));
10510 * The number of CPUs to search drops as sum_util increases, when
10511 * sum_util hits 85% or above, the scan stops.
10512 * The reason to choose 85% as the threshold is because this is the
10513 * imbalance_pct(117) when a LLC sched group is overloaded.
10515 * let y = SCHED_CAPACITY_SCALE - p * x^2 [1]
10516 * and y'= y / SCHED_CAPACITY_SCALE
10518 * x is the ratio of sum_util compared to the CPU capacity:
10519 * x = sum_util / (llc_weight * SCHED_CAPACITY_SCALE)
10520 * y' is the ratio of CPUs to be scanned in the LLC domain,
10521 * and the number of CPUs to scan is calculated by:
10523 * nr_scan = llc_weight * y' [2]
10525 * When x hits the threshold of overloaded, AKA, when
10526 * x = 100 / pct, y drops to 0. According to [1],
10527 * p should be SCHED_CAPACITY_SCALE * pct^2 / 10000
10529 * Scale x by SCHED_CAPACITY_SCALE:
10530 * x' = sum_util / llc_weight; [3]
10532 * and finally [1] becomes:
10533 * y = SCHED_CAPACITY_SCALE -
10534 * x'^2 * pct^2 / (10000 * SCHED_CAPACITY_SCALE) [4]
10539 do_div(x, llc_weight);
10542 pct = env->sd->imbalance_pct;
10543 tmp = x * x * pct * pct;
10544 do_div(tmp, 10000 * SCHED_CAPACITY_SCALE);
10545 tmp = min_t(long, tmp, SCHED_CAPACITY_SCALE);
10546 y = SCHED_CAPACITY_SCALE - tmp;
10550 do_div(y, SCHED_CAPACITY_SCALE);
10551 if ((int)y != sd_share->nr_idle_scan)
10552 WRITE_ONCE(sd_share->nr_idle_scan, (int)y);
10556 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
10557 * @env: The load balancing environment.
10558 * @sds: variable to hold the statistics for this sched_domain.
10561 static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
10563 struct sched_group *sg = env->sd->groups;
10564 struct sg_lb_stats *local = &sds->local_stat;
10565 struct sg_lb_stats tmp_sgs;
10566 unsigned long sum_util = 0;
10570 struct sg_lb_stats *sgs = &tmp_sgs;
10573 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_span(sg));
10578 if (env->idle != CPU_NEWLY_IDLE ||
10579 time_after_eq(jiffies, sg->sgc->next_update))
10580 update_group_capacity(env->sd, env->dst_cpu);
10583 update_sg_lb_stats(env, sds, sg, sgs, &sg_status);
10585 if (!local_group && update_sd_pick_busiest(env, sds, sg, sgs)) {
10587 sds->busiest_stat = *sgs;
10590 /* Now, start updating sd_lb_stats */
10591 sds->total_load += sgs->group_load;
10592 sds->total_capacity += sgs->group_capacity;
10594 sum_util += sgs->group_util;
10596 } while (sg != env->sd->groups);
10599 * Indicate that the child domain of the busiest group prefers tasks
10600 * go to a child's sibling domains first. NB the flags of a sched group
10601 * are those of the child domain.
10604 sds->prefer_sibling = !!(sds->busiest->flags & SD_PREFER_SIBLING);
10607 if (env->sd->flags & SD_NUMA)
10608 env->fbq_type = fbq_classify_group(&sds->busiest_stat);
10610 if (!env->sd->parent) {
10611 struct root_domain *rd = env->dst_rq->rd;
10613 /* update overload indicator if we are at root domain */
10614 WRITE_ONCE(rd->overload, sg_status & SG_OVERLOAD);
10616 /* Update over-utilization (tipping point, U >= 0) indicator */
10617 WRITE_ONCE(rd->overutilized, sg_status & SG_OVERUTILIZED);
10618 trace_sched_overutilized_tp(rd, sg_status & SG_OVERUTILIZED);
10619 } else if (sg_status & SG_OVERUTILIZED) {
10620 struct root_domain *rd = env->dst_rq->rd;
10622 WRITE_ONCE(rd->overutilized, SG_OVERUTILIZED);
10623 trace_sched_overutilized_tp(rd, SG_OVERUTILIZED);
10626 update_idle_cpu_scan(env, sum_util);
10630 * calculate_imbalance - Calculate the amount of imbalance present within the
10631 * groups of a given sched_domain during load balance.
10632 * @env: load balance environment
10633 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
10635 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
10637 struct sg_lb_stats *local, *busiest;
10639 local = &sds->local_stat;
10640 busiest = &sds->busiest_stat;
10642 if (busiest->group_type == group_misfit_task) {
10643 if (env->sd->flags & SD_ASYM_CPUCAPACITY) {
10644 /* Set imbalance to allow misfit tasks to be balanced. */
10645 env->migration_type = migrate_misfit;
10646 env->imbalance = 1;
10649 * Set load imbalance to allow moving task from cpu
10650 * with reduced capacity.
10652 env->migration_type = migrate_load;
10653 env->imbalance = busiest->group_misfit_task_load;
10658 if (busiest->group_type == group_asym_packing) {
10660 * In case of asym capacity, we will try to migrate all load to
10661 * the preferred CPU.
10663 env->migration_type = migrate_task;
10664 env->imbalance = busiest->sum_h_nr_running;
10668 if (busiest->group_type == group_smt_balance) {
10669 /* Reduce number of tasks sharing CPU capacity */
10670 env->migration_type = migrate_task;
10671 env->imbalance = 1;
10675 if (busiest->group_type == group_imbalanced) {
10677 * In the group_imb case we cannot rely on group-wide averages
10678 * to ensure CPU-load equilibrium, try to move any task to fix
10679 * the imbalance. The next load balance will take care of
10680 * balancing back the system.
10682 env->migration_type = migrate_task;
10683 env->imbalance = 1;
10688 * Try to use spare capacity of local group without overloading it or
10689 * emptying busiest.
10691 if (local->group_type == group_has_spare) {
10692 if ((busiest->group_type > group_fully_busy) &&
10693 !(env->sd->flags & SD_SHARE_LLC)) {
10695 * If busiest is overloaded, try to fill spare
10696 * capacity. This might end up creating spare capacity
10697 * in busiest or busiest still being overloaded but
10698 * there is no simple way to directly compute the
10699 * amount of load to migrate in order to balance the
10702 env->migration_type = migrate_util;
10703 env->imbalance = max(local->group_capacity, local->group_util) -
10707 * In some cases, the group's utilization is max or even
10708 * higher than capacity because of migrations but the
10709 * local CPU is (newly) idle. There is at least one
10710 * waiting task in this overloaded busiest group. Let's
10713 if (env->idle != CPU_NOT_IDLE && env->imbalance == 0) {
10714 env->migration_type = migrate_task;
10715 env->imbalance = 1;
10721 if (busiest->group_weight == 1 || sds->prefer_sibling) {
10723 * When prefer sibling, evenly spread running tasks on
10726 env->migration_type = migrate_task;
10727 env->imbalance = sibling_imbalance(env, sds, busiest, local);
10731 * If there is no overload, we just want to even the number of
10734 env->migration_type = migrate_task;
10735 env->imbalance = max_t(long, 0,
10736 (local->idle_cpus - busiest->idle_cpus));
10740 /* Consider allowing a small imbalance between NUMA groups */
10741 if (env->sd->flags & SD_NUMA) {
10742 env->imbalance = adjust_numa_imbalance(env->imbalance,
10743 local->sum_nr_running + 1,
10744 env->sd->imb_numa_nr);
10748 /* Number of tasks to move to restore balance */
10749 env->imbalance >>= 1;
10755 * Local is fully busy but has to take more load to relieve the
10758 if (local->group_type < group_overloaded) {
10760 * Local will become overloaded so the avg_load metrics are
10764 local->avg_load = (local->group_load * SCHED_CAPACITY_SCALE) /
10765 local->group_capacity;
10768 * If the local group is more loaded than the selected
10769 * busiest group don't try to pull any tasks.
10771 if (local->avg_load >= busiest->avg_load) {
10772 env->imbalance = 0;
10776 sds->avg_load = (sds->total_load * SCHED_CAPACITY_SCALE) /
10777 sds->total_capacity;
10780 * If the local group is more loaded than the average system
10781 * load, don't try to pull any tasks.
10783 if (local->avg_load >= sds->avg_load) {
10784 env->imbalance = 0;
10791 * Both group are or will become overloaded and we're trying to get all
10792 * the CPUs to the average_load, so we don't want to push ourselves
10793 * above the average load, nor do we wish to reduce the max loaded CPU
10794 * below the average load. At the same time, we also don't want to
10795 * reduce the group load below the group capacity. Thus we look for
10796 * the minimum possible imbalance.
10798 env->migration_type = migrate_load;
10799 env->imbalance = min(
10800 (busiest->avg_load - sds->avg_load) * busiest->group_capacity,
10801 (sds->avg_load - local->avg_load) * local->group_capacity
10802 ) / SCHED_CAPACITY_SCALE;
10805 /******* find_busiest_group() helpers end here *********************/
10808 * Decision matrix according to the local and busiest group type:
10810 * busiest \ local has_spare fully_busy misfit asym imbalanced overloaded
10811 * has_spare nr_idle balanced N/A N/A balanced balanced
10812 * fully_busy nr_idle nr_idle N/A N/A balanced balanced
10813 * misfit_task force N/A N/A N/A N/A N/A
10814 * asym_packing force force N/A N/A force force
10815 * imbalanced force force N/A N/A force force
10816 * overloaded force force N/A N/A force avg_load
10818 * N/A : Not Applicable because already filtered while updating
10820 * balanced : The system is balanced for these 2 groups.
10821 * force : Calculate the imbalance as load migration is probably needed.
10822 * avg_load : Only if imbalance is significant enough.
10823 * nr_idle : dst_cpu is not busy and the number of idle CPUs is quite
10824 * different in groups.
10828 * find_busiest_group - Returns the busiest group within the sched_domain
10829 * if there is an imbalance.
10830 * @env: The load balancing environment.
10832 * Also calculates the amount of runnable load which should be moved
10833 * to restore balance.
10835 * Return: - The busiest group if imbalance exists.
10837 static struct sched_group *find_busiest_group(struct lb_env *env)
10839 struct sg_lb_stats *local, *busiest;
10840 struct sd_lb_stats sds;
10842 init_sd_lb_stats(&sds);
10845 * Compute the various statistics relevant for load balancing at
10848 update_sd_lb_stats(env, &sds);
10850 /* There is no busy sibling group to pull tasks from */
10854 busiest = &sds.busiest_stat;
10856 /* Misfit tasks should be dealt with regardless of the avg load */
10857 if (busiest->group_type == group_misfit_task)
10858 goto force_balance;
10860 if (sched_energy_enabled()) {
10861 struct root_domain *rd = env->dst_rq->rd;
10863 if (rcu_dereference(rd->pd) && !READ_ONCE(rd->overutilized))
10867 /* ASYM feature bypasses nice load balance check */
10868 if (busiest->group_type == group_asym_packing)
10869 goto force_balance;
10872 * If the busiest group is imbalanced the below checks don't
10873 * work because they assume all things are equal, which typically
10874 * isn't true due to cpus_ptr constraints and the like.
10876 if (busiest->group_type == group_imbalanced)
10877 goto force_balance;
10879 local = &sds.local_stat;
10881 * If the local group is busier than the selected busiest group
10882 * don't try and pull any tasks.
10884 if (local->group_type > busiest->group_type)
10888 * When groups are overloaded, use the avg_load to ensure fairness
10891 if (local->group_type == group_overloaded) {
10893 * If the local group is more loaded than the selected
10894 * busiest group don't try to pull any tasks.
10896 if (local->avg_load >= busiest->avg_load)
10899 /* XXX broken for overlapping NUMA groups */
10900 sds.avg_load = (sds.total_load * SCHED_CAPACITY_SCALE) /
10901 sds.total_capacity;
10904 * Don't pull any tasks if this group is already above the
10905 * domain average load.
10907 if (local->avg_load >= sds.avg_load)
10911 * If the busiest group is more loaded, use imbalance_pct to be
10914 if (100 * busiest->avg_load <=
10915 env->sd->imbalance_pct * local->avg_load)
10920 * Try to move all excess tasks to a sibling domain of the busiest
10921 * group's child domain.
10923 if (sds.prefer_sibling && local->group_type == group_has_spare &&
10924 sibling_imbalance(env, &sds, busiest, local) > 1)
10925 goto force_balance;
10927 if (busiest->group_type != group_overloaded) {
10928 if (env->idle == CPU_NOT_IDLE) {
10930 * If the busiest group is not overloaded (and as a
10931 * result the local one too) but this CPU is already
10932 * busy, let another idle CPU try to pull task.
10937 if (busiest->group_type == group_smt_balance &&
10938 smt_vs_nonsmt_groups(sds.local, sds.busiest)) {
10939 /* Let non SMT CPU pull from SMT CPU sharing with sibling */
10940 goto force_balance;
10943 if (busiest->group_weight > 1 &&
10944 local->idle_cpus <= (busiest->idle_cpus + 1)) {
10946 * If the busiest group is not overloaded
10947 * and there is no imbalance between this and busiest
10948 * group wrt idle CPUs, it is balanced. The imbalance
10949 * becomes significant if the diff is greater than 1
10950 * otherwise we might end up to just move the imbalance
10951 * on another group. Of course this applies only if
10952 * there is more than 1 CPU per group.
10957 if (busiest->sum_h_nr_running == 1) {
10959 * busiest doesn't have any tasks waiting to run
10966 /* Looks like there is an imbalance. Compute it */
10967 calculate_imbalance(env, &sds);
10968 return env->imbalance ? sds.busiest : NULL;
10971 env->imbalance = 0;
10976 * find_busiest_queue - find the busiest runqueue among the CPUs in the group.
10978 static struct rq *find_busiest_queue(struct lb_env *env,
10979 struct sched_group *group)
10981 struct rq *busiest = NULL, *rq;
10982 unsigned long busiest_util = 0, busiest_load = 0, busiest_capacity = 1;
10983 unsigned int busiest_nr = 0;
10986 for_each_cpu_and(i, sched_group_span(group), env->cpus) {
10987 unsigned long capacity, load, util;
10988 unsigned int nr_running;
10992 rt = fbq_classify_rq(rq);
10995 * We classify groups/runqueues into three groups:
10996 * - regular: there are !numa tasks
10997 * - remote: there are numa tasks that run on the 'wrong' node
10998 * - all: there is no distinction
11000 * In order to avoid migrating ideally placed numa tasks,
11001 * ignore those when there's better options.
11003 * If we ignore the actual busiest queue to migrate another
11004 * task, the next balance pass can still reduce the busiest
11005 * queue by moving tasks around inside the node.
11007 * If we cannot move enough load due to this classification
11008 * the next pass will adjust the group classification and
11009 * allow migration of more tasks.
11011 * Both cases only affect the total convergence complexity.
11013 if (rt > env->fbq_type)
11016 nr_running = rq->cfs.h_nr_running;
11020 capacity = capacity_of(i);
11023 * For ASYM_CPUCAPACITY domains, don't pick a CPU that could
11024 * eventually lead to active_balancing high->low capacity.
11025 * Higher per-CPU capacity is considered better than balancing
11028 if (env->sd->flags & SD_ASYM_CPUCAPACITY &&
11029 !capacity_greater(capacity_of(env->dst_cpu), capacity) &&
11034 * Make sure we only pull tasks from a CPU of lower priority
11035 * when balancing between SMT siblings.
11037 * If balancing between cores, let lower priority CPUs help
11038 * SMT cores with more than one busy sibling.
11040 if (sched_asym(env->sd, i, env->dst_cpu) && nr_running == 1)
11043 switch (env->migration_type) {
11046 * When comparing with load imbalance, use cpu_load()
11047 * which is not scaled with the CPU capacity.
11049 load = cpu_load(rq);
11051 if (nr_running == 1 && load > env->imbalance &&
11052 !check_cpu_capacity(rq, env->sd))
11056 * For the load comparisons with the other CPUs,
11057 * consider the cpu_load() scaled with the CPU
11058 * capacity, so that the load can be moved away
11059 * from the CPU that is potentially running at a
11062 * Thus we're looking for max(load_i / capacity_i),
11063 * crosswise multiplication to rid ourselves of the
11064 * division works out to:
11065 * load_i * capacity_j > load_j * capacity_i;
11066 * where j is our previous maximum.
11068 if (load * busiest_capacity > busiest_load * capacity) {
11069 busiest_load = load;
11070 busiest_capacity = capacity;
11076 util = cpu_util_cfs_boost(i);
11079 * Don't try to pull utilization from a CPU with one
11080 * running task. Whatever its utilization, we will fail
11083 if (nr_running <= 1)
11086 if (busiest_util < util) {
11087 busiest_util = util;
11093 if (busiest_nr < nr_running) {
11094 busiest_nr = nr_running;
11099 case migrate_misfit:
11101 * For ASYM_CPUCAPACITY domains with misfit tasks we
11102 * simply seek the "biggest" misfit task.
11104 if (rq->misfit_task_load > busiest_load) {
11105 busiest_load = rq->misfit_task_load;
11118 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
11119 * so long as it is large enough.
11121 #define MAX_PINNED_INTERVAL 512
11124 asym_active_balance(struct lb_env *env)
11127 * ASYM_PACKING needs to force migrate tasks from busy but lower
11128 * priority CPUs in order to pack all tasks in the highest priority
11129 * CPUs. When done between cores, do it only if the whole core if the
11130 * whole core is idle.
11132 * If @env::src_cpu is an SMT core with busy siblings, let
11133 * the lower priority @env::dst_cpu help it. Do not follow
11136 return env->idle != CPU_NOT_IDLE && sched_use_asym_prio(env->sd, env->dst_cpu) &&
11137 (sched_asym_prefer(env->dst_cpu, env->src_cpu) ||
11138 !sched_use_asym_prio(env->sd, env->src_cpu));
11142 imbalanced_active_balance(struct lb_env *env)
11144 struct sched_domain *sd = env->sd;
11147 * The imbalanced case includes the case of pinned tasks preventing a fair
11148 * distribution of the load on the system but also the even distribution of the
11149 * threads on a system with spare capacity
11151 if ((env->migration_type == migrate_task) &&
11152 (sd->nr_balance_failed > sd->cache_nice_tries+2))
11158 static int need_active_balance(struct lb_env *env)
11160 struct sched_domain *sd = env->sd;
11162 if (asym_active_balance(env))
11165 if (imbalanced_active_balance(env))
11169 * The dst_cpu is idle and the src_cpu CPU has only 1 CFS task.
11170 * It's worth migrating the task if the src_cpu's capacity is reduced
11171 * because of other sched_class or IRQs if more capacity stays
11172 * available on dst_cpu.
11174 if ((env->idle != CPU_NOT_IDLE) &&
11175 (env->src_rq->cfs.h_nr_running == 1)) {
11176 if ((check_cpu_capacity(env->src_rq, sd)) &&
11177 (capacity_of(env->src_cpu)*sd->imbalance_pct < capacity_of(env->dst_cpu)*100))
11181 if (env->migration_type == migrate_misfit)
11187 static int active_load_balance_cpu_stop(void *data);
11189 static int should_we_balance(struct lb_env *env)
11191 struct cpumask *swb_cpus = this_cpu_cpumask_var_ptr(should_we_balance_tmpmask);
11192 struct sched_group *sg = env->sd->groups;
11193 int cpu, idle_smt = -1;
11196 * Ensure the balancing environment is consistent; can happen
11197 * when the softirq triggers 'during' hotplug.
11199 if (!cpumask_test_cpu(env->dst_cpu, env->cpus))
11203 * In the newly idle case, we will allow all the CPUs
11204 * to do the newly idle load balance.
11206 * However, we bail out if we already have tasks or a wakeup pending,
11207 * to optimize wakeup latency.
11209 if (env->idle == CPU_NEWLY_IDLE) {
11210 if (env->dst_rq->nr_running > 0 || env->dst_rq->ttwu_pending)
11215 cpumask_copy(swb_cpus, group_balance_mask(sg));
11216 /* Try to find first idle CPU */
11217 for_each_cpu_and(cpu, swb_cpus, env->cpus) {
11218 if (!idle_cpu(cpu))
11222 * Don't balance to idle SMT in busy core right away when
11223 * balancing cores, but remember the first idle SMT CPU for
11224 * later consideration. Find CPU on an idle core first.
11226 if (!(env->sd->flags & SD_SHARE_CPUCAPACITY) && !is_core_idle(cpu)) {
11227 if (idle_smt == -1)
11230 * If the core is not idle, and first SMT sibling which is
11231 * idle has been found, then its not needed to check other
11232 * SMT siblings for idleness:
11234 #ifdef CONFIG_SCHED_SMT
11235 cpumask_andnot(swb_cpus, swb_cpus, cpu_smt_mask(cpu));
11241 * Are we the first idle core in a non-SMT domain or higher,
11242 * or the first idle CPU in a SMT domain?
11244 return cpu == env->dst_cpu;
11247 /* Are we the first idle CPU with busy siblings? */
11248 if (idle_smt != -1)
11249 return idle_smt == env->dst_cpu;
11251 /* Are we the first CPU of this group ? */
11252 return group_balance_cpu(sg) == env->dst_cpu;
11256 * Check this_cpu to ensure it is balanced within domain. Attempt to move
11257 * tasks if there is an imbalance.
11259 static int load_balance(int this_cpu, struct rq *this_rq,
11260 struct sched_domain *sd, enum cpu_idle_type idle,
11261 int *continue_balancing)
11263 int ld_moved, cur_ld_moved, active_balance = 0;
11264 struct sched_domain *sd_parent = sd->parent;
11265 struct sched_group *group;
11266 struct rq *busiest;
11267 struct rq_flags rf;
11268 struct cpumask *cpus = this_cpu_cpumask_var_ptr(load_balance_mask);
11269 struct lb_env env = {
11271 .dst_cpu = this_cpu,
11273 .dst_grpmask = group_balance_mask(sd->groups),
11275 .loop_break = SCHED_NR_MIGRATE_BREAK,
11278 .tasks = LIST_HEAD_INIT(env.tasks),
11281 cpumask_and(cpus, sched_domain_span(sd), cpu_active_mask);
11283 schedstat_inc(sd->lb_count[idle]);
11286 if (!should_we_balance(&env)) {
11287 *continue_balancing = 0;
11291 group = find_busiest_group(&env);
11293 schedstat_inc(sd->lb_nobusyg[idle]);
11297 busiest = find_busiest_queue(&env, group);
11299 schedstat_inc(sd->lb_nobusyq[idle]);
11303 WARN_ON_ONCE(busiest == env.dst_rq);
11305 schedstat_add(sd->lb_imbalance[idle], env.imbalance);
11307 env.src_cpu = busiest->cpu;
11308 env.src_rq = busiest;
11311 /* Clear this flag as soon as we find a pullable task */
11312 env.flags |= LBF_ALL_PINNED;
11313 if (busiest->nr_running > 1) {
11315 * Attempt to move tasks. If find_busiest_group has found
11316 * an imbalance but busiest->nr_running <= 1, the group is
11317 * still unbalanced. ld_moved simply stays zero, so it is
11318 * correctly treated as an imbalance.
11320 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
11323 rq_lock_irqsave(busiest, &rf);
11324 update_rq_clock(busiest);
11327 * cur_ld_moved - load moved in current iteration
11328 * ld_moved - cumulative load moved across iterations
11330 cur_ld_moved = detach_tasks(&env);
11333 * We've detached some tasks from busiest_rq. Every
11334 * task is masked "TASK_ON_RQ_MIGRATING", so we can safely
11335 * unlock busiest->lock, and we are able to be sure
11336 * that nobody can manipulate the tasks in parallel.
11337 * See task_rq_lock() family for the details.
11340 rq_unlock(busiest, &rf);
11342 if (cur_ld_moved) {
11343 attach_tasks(&env);
11344 ld_moved += cur_ld_moved;
11347 local_irq_restore(rf.flags);
11349 if (env.flags & LBF_NEED_BREAK) {
11350 env.flags &= ~LBF_NEED_BREAK;
11351 /* Stop if we tried all running tasks */
11352 if (env.loop < busiest->nr_running)
11357 * Revisit (affine) tasks on src_cpu that couldn't be moved to
11358 * us and move them to an alternate dst_cpu in our sched_group
11359 * where they can run. The upper limit on how many times we
11360 * iterate on same src_cpu is dependent on number of CPUs in our
11363 * This changes load balance semantics a bit on who can move
11364 * load to a given_cpu. In addition to the given_cpu itself
11365 * (or a ilb_cpu acting on its behalf where given_cpu is
11366 * nohz-idle), we now have balance_cpu in a position to move
11367 * load to given_cpu. In rare situations, this may cause
11368 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
11369 * _independently_ and at _same_ time to move some load to
11370 * given_cpu) causing excess load to be moved to given_cpu.
11371 * This however should not happen so much in practice and
11372 * moreover subsequent load balance cycles should correct the
11373 * excess load moved.
11375 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
11377 /* Prevent to re-select dst_cpu via env's CPUs */
11378 __cpumask_clear_cpu(env.dst_cpu, env.cpus);
11380 env.dst_rq = cpu_rq(env.new_dst_cpu);
11381 env.dst_cpu = env.new_dst_cpu;
11382 env.flags &= ~LBF_DST_PINNED;
11384 env.loop_break = SCHED_NR_MIGRATE_BREAK;
11387 * Go back to "more_balance" rather than "redo" since we
11388 * need to continue with same src_cpu.
11394 * We failed to reach balance because of affinity.
11397 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
11399 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0)
11400 *group_imbalance = 1;
11403 /* All tasks on this runqueue were pinned by CPU affinity */
11404 if (unlikely(env.flags & LBF_ALL_PINNED)) {
11405 __cpumask_clear_cpu(cpu_of(busiest), cpus);
11407 * Attempting to continue load balancing at the current
11408 * sched_domain level only makes sense if there are
11409 * active CPUs remaining as possible busiest CPUs to
11410 * pull load from which are not contained within the
11411 * destination group that is receiving any migrated
11414 if (!cpumask_subset(cpus, env.dst_grpmask)) {
11416 env.loop_break = SCHED_NR_MIGRATE_BREAK;
11419 goto out_all_pinned;
11424 schedstat_inc(sd->lb_failed[idle]);
11426 * Increment the failure counter only on periodic balance.
11427 * We do not want newidle balance, which can be very
11428 * frequent, pollute the failure counter causing
11429 * excessive cache_hot migrations and active balances.
11431 if (idle != CPU_NEWLY_IDLE)
11432 sd->nr_balance_failed++;
11434 if (need_active_balance(&env)) {
11435 unsigned long flags;
11437 raw_spin_rq_lock_irqsave(busiest, flags);
11440 * Don't kick the active_load_balance_cpu_stop,
11441 * if the curr task on busiest CPU can't be
11442 * moved to this_cpu:
11444 if (!cpumask_test_cpu(this_cpu, busiest->curr->cpus_ptr)) {
11445 raw_spin_rq_unlock_irqrestore(busiest, flags);
11446 goto out_one_pinned;
11449 /* Record that we found at least one task that could run on this_cpu */
11450 env.flags &= ~LBF_ALL_PINNED;
11453 * ->active_balance synchronizes accesses to
11454 * ->active_balance_work. Once set, it's cleared
11455 * only after active load balance is finished.
11457 if (!busiest->active_balance) {
11458 busiest->active_balance = 1;
11459 busiest->push_cpu = this_cpu;
11460 active_balance = 1;
11464 raw_spin_rq_unlock_irqrestore(busiest, flags);
11465 if (active_balance) {
11466 stop_one_cpu_nowait(cpu_of(busiest),
11467 active_load_balance_cpu_stop, busiest,
11468 &busiest->active_balance_work);
11473 sd->nr_balance_failed = 0;
11476 if (likely(!active_balance) || need_active_balance(&env)) {
11477 /* We were unbalanced, so reset the balancing interval */
11478 sd->balance_interval = sd->min_interval;
11485 * We reach balance although we may have faced some affinity
11486 * constraints. Clear the imbalance flag only if other tasks got
11487 * a chance to move and fix the imbalance.
11489 if (sd_parent && !(env.flags & LBF_ALL_PINNED)) {
11490 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
11492 if (*group_imbalance)
11493 *group_imbalance = 0;
11498 * We reach balance because all tasks are pinned at this level so
11499 * we can't migrate them. Let the imbalance flag set so parent level
11500 * can try to migrate them.
11502 schedstat_inc(sd->lb_balanced[idle]);
11504 sd->nr_balance_failed = 0;
11510 * newidle_balance() disregards balance intervals, so we could
11511 * repeatedly reach this code, which would lead to balance_interval
11512 * skyrocketing in a short amount of time. Skip the balance_interval
11513 * increase logic to avoid that.
11515 if (env.idle == CPU_NEWLY_IDLE)
11518 /* tune up the balancing interval */
11519 if ((env.flags & LBF_ALL_PINNED &&
11520 sd->balance_interval < MAX_PINNED_INTERVAL) ||
11521 sd->balance_interval < sd->max_interval)
11522 sd->balance_interval *= 2;
11527 static inline unsigned long
11528 get_sd_balance_interval(struct sched_domain *sd, int cpu_busy)
11530 unsigned long interval = sd->balance_interval;
11533 interval *= sd->busy_factor;
11535 /* scale ms to jiffies */
11536 interval = msecs_to_jiffies(interval);
11539 * Reduce likelihood of busy balancing at higher domains racing with
11540 * balancing at lower domains by preventing their balancing periods
11541 * from being multiples of each other.
11546 interval = clamp(interval, 1UL, max_load_balance_interval);
11552 update_next_balance(struct sched_domain *sd, unsigned long *next_balance)
11554 unsigned long interval, next;
11556 /* used by idle balance, so cpu_busy = 0 */
11557 interval = get_sd_balance_interval(sd, 0);
11558 next = sd->last_balance + interval;
11560 if (time_after(*next_balance, next))
11561 *next_balance = next;
11565 * active_load_balance_cpu_stop is run by the CPU stopper. It pushes
11566 * running tasks off the busiest CPU onto idle CPUs. It requires at
11567 * least 1 task to be running on each physical CPU where possible, and
11568 * avoids physical / logical imbalances.
11570 static int active_load_balance_cpu_stop(void *data)
11572 struct rq *busiest_rq = data;
11573 int busiest_cpu = cpu_of(busiest_rq);
11574 int target_cpu = busiest_rq->push_cpu;
11575 struct rq *target_rq = cpu_rq(target_cpu);
11576 struct sched_domain *sd;
11577 struct task_struct *p = NULL;
11578 struct rq_flags rf;
11580 rq_lock_irq(busiest_rq, &rf);
11582 * Between queueing the stop-work and running it is a hole in which
11583 * CPUs can become inactive. We should not move tasks from or to
11586 if (!cpu_active(busiest_cpu) || !cpu_active(target_cpu))
11589 /* Make sure the requested CPU hasn't gone down in the meantime: */
11590 if (unlikely(busiest_cpu != smp_processor_id() ||
11591 !busiest_rq->active_balance))
11594 /* Is there any task to move? */
11595 if (busiest_rq->nr_running <= 1)
11599 * This condition is "impossible", if it occurs
11600 * we need to fix it. Originally reported by
11601 * Bjorn Helgaas on a 128-CPU setup.
11603 WARN_ON_ONCE(busiest_rq == target_rq);
11605 /* Search for an sd spanning us and the target CPU. */
11607 for_each_domain(target_cpu, sd) {
11608 if (cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
11613 struct lb_env env = {
11615 .dst_cpu = target_cpu,
11616 .dst_rq = target_rq,
11617 .src_cpu = busiest_rq->cpu,
11618 .src_rq = busiest_rq,
11620 .flags = LBF_ACTIVE_LB,
11623 schedstat_inc(sd->alb_count);
11624 update_rq_clock(busiest_rq);
11626 p = detach_one_task(&env);
11628 schedstat_inc(sd->alb_pushed);
11629 /* Active balancing done, reset the failure counter. */
11630 sd->nr_balance_failed = 0;
11632 schedstat_inc(sd->alb_failed);
11637 busiest_rq->active_balance = 0;
11638 rq_unlock(busiest_rq, &rf);
11641 attach_one_task(target_rq, p);
11643 local_irq_enable();
11648 static DEFINE_SPINLOCK(balancing);
11651 * Scale the max load_balance interval with the number of CPUs in the system.
11652 * This trades load-balance latency on larger machines for less cross talk.
11654 void update_max_interval(void)
11656 max_load_balance_interval = HZ*num_online_cpus()/10;
11659 static inline bool update_newidle_cost(struct sched_domain *sd, u64 cost)
11661 if (cost > sd->max_newidle_lb_cost) {
11663 * Track max cost of a domain to make sure to not delay the
11664 * next wakeup on the CPU.
11666 sd->max_newidle_lb_cost = cost;
11667 sd->last_decay_max_lb_cost = jiffies;
11668 } else if (time_after(jiffies, sd->last_decay_max_lb_cost + HZ)) {
11670 * Decay the newidle max times by ~1% per second to ensure that
11671 * it is not outdated and the current max cost is actually
11674 sd->max_newidle_lb_cost = (sd->max_newidle_lb_cost * 253) / 256;
11675 sd->last_decay_max_lb_cost = jiffies;
11684 * It checks each scheduling domain to see if it is due to be balanced,
11685 * and initiates a balancing operation if so.
11687 * Balancing parameters are set up in init_sched_domains.
11689 static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle)
11691 int continue_balancing = 1;
11693 int busy = idle != CPU_IDLE && !sched_idle_cpu(cpu);
11694 unsigned long interval;
11695 struct sched_domain *sd;
11696 /* Earliest time when we have to do rebalance again */
11697 unsigned long next_balance = jiffies + 60*HZ;
11698 int update_next_balance = 0;
11699 int need_serialize, need_decay = 0;
11703 for_each_domain(cpu, sd) {
11705 * Decay the newidle max times here because this is a regular
11706 * visit to all the domains.
11708 need_decay = update_newidle_cost(sd, 0);
11709 max_cost += sd->max_newidle_lb_cost;
11712 * Stop the load balance at this level. There is another
11713 * CPU in our sched group which is doing load balancing more
11716 if (!continue_balancing) {
11722 interval = get_sd_balance_interval(sd, busy);
11724 need_serialize = sd->flags & SD_SERIALIZE;
11725 if (need_serialize) {
11726 if (!spin_trylock(&balancing))
11730 if (time_after_eq(jiffies, sd->last_balance + interval)) {
11731 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
11733 * The LBF_DST_PINNED logic could have changed
11734 * env->dst_cpu, so we can't know our idle
11735 * state even if we migrated tasks. Update it.
11737 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
11738 busy = idle != CPU_IDLE && !sched_idle_cpu(cpu);
11740 sd->last_balance = jiffies;
11741 interval = get_sd_balance_interval(sd, busy);
11743 if (need_serialize)
11744 spin_unlock(&balancing);
11746 if (time_after(next_balance, sd->last_balance + interval)) {
11747 next_balance = sd->last_balance + interval;
11748 update_next_balance = 1;
11753 * Ensure the rq-wide value also decays but keep it at a
11754 * reasonable floor to avoid funnies with rq->avg_idle.
11756 rq->max_idle_balance_cost =
11757 max((u64)sysctl_sched_migration_cost, max_cost);
11762 * next_balance will be updated only when there is a need.
11763 * When the cpu is attached to null domain for ex, it will not be
11766 if (likely(update_next_balance))
11767 rq->next_balance = next_balance;
11771 static inline int on_null_domain(struct rq *rq)
11773 return unlikely(!rcu_dereference_sched(rq->sd));
11776 #ifdef CONFIG_NO_HZ_COMMON
11778 * NOHZ idle load balancing (ILB) details:
11780 * - When one of the busy CPUs notices that there may be an idle rebalancing
11781 * needed, they will kick the idle load balancer, which then does idle
11782 * load balancing for all the idle CPUs.
11784 * - HK_TYPE_MISC CPUs are used for this task, because HK_TYPE_SCHED is not set
11787 static inline int find_new_ilb(void)
11789 const struct cpumask *hk_mask;
11792 hk_mask = housekeeping_cpumask(HK_TYPE_MISC);
11794 for_each_cpu_and(ilb_cpu, nohz.idle_cpus_mask, hk_mask) {
11796 if (ilb_cpu == smp_processor_id())
11799 if (idle_cpu(ilb_cpu))
11807 * Kick a CPU to do the NOHZ balancing, if it is time for it, via a cross-CPU
11808 * SMP function call (IPI).
11810 * We pick the first idle CPU in the HK_TYPE_MISC housekeeping set (if there is one).
11812 static void kick_ilb(unsigned int flags)
11817 * Increase nohz.next_balance only when if full ilb is triggered but
11818 * not if we only update stats.
11820 if (flags & NOHZ_BALANCE_KICK)
11821 nohz.next_balance = jiffies+1;
11823 ilb_cpu = find_new_ilb();
11828 * Access to rq::nohz_csd is serialized by NOHZ_KICK_MASK; he who sets
11829 * the first flag owns it; cleared by nohz_csd_func().
11831 flags = atomic_fetch_or(flags, nohz_flags(ilb_cpu));
11832 if (flags & NOHZ_KICK_MASK)
11836 * This way we generate an IPI on the target CPU which
11837 * is idle, and the softirq performing NOHZ idle load balancing
11838 * will be run before returning from the IPI.
11840 smp_call_function_single_async(ilb_cpu, &cpu_rq(ilb_cpu)->nohz_csd);
11844 * Current decision point for kicking the idle load balancer in the presence
11845 * of idle CPUs in the system.
11847 static void nohz_balancer_kick(struct rq *rq)
11849 unsigned long now = jiffies;
11850 struct sched_domain_shared *sds;
11851 struct sched_domain *sd;
11852 int nr_busy, i, cpu = rq->cpu;
11853 unsigned int flags = 0;
11855 if (unlikely(rq->idle_balance))
11859 * We may be recently in ticked or tickless idle mode. At the first
11860 * busy tick after returning from idle, we will update the busy stats.
11862 nohz_balance_exit_idle(rq);
11865 * None are in tickless mode and hence no need for NOHZ idle load
11868 if (likely(!atomic_read(&nohz.nr_cpus)))
11871 if (READ_ONCE(nohz.has_blocked) &&
11872 time_after(now, READ_ONCE(nohz.next_blocked)))
11873 flags = NOHZ_STATS_KICK;
11875 if (time_before(now, nohz.next_balance))
11878 if (rq->nr_running >= 2) {
11879 flags = NOHZ_STATS_KICK | NOHZ_BALANCE_KICK;
11885 sd = rcu_dereference(rq->sd);
11888 * If there's a runnable CFS task and the current CPU has reduced
11889 * capacity, kick the ILB to see if there's a better CPU to run on:
11891 if (rq->cfs.h_nr_running >= 1 && check_cpu_capacity(rq, sd)) {
11892 flags = NOHZ_STATS_KICK | NOHZ_BALANCE_KICK;
11897 sd = rcu_dereference(per_cpu(sd_asym_packing, cpu));
11900 * When ASYM_PACKING; see if there's a more preferred CPU
11901 * currently idle; in which case, kick the ILB to move tasks
11904 * When balancing betwen cores, all the SMT siblings of the
11905 * preferred CPU must be idle.
11907 for_each_cpu_and(i, sched_domain_span(sd), nohz.idle_cpus_mask) {
11908 if (sched_asym(sd, i, cpu)) {
11909 flags = NOHZ_STATS_KICK | NOHZ_BALANCE_KICK;
11915 sd = rcu_dereference(per_cpu(sd_asym_cpucapacity, cpu));
11918 * When ASYM_CPUCAPACITY; see if there's a higher capacity CPU
11919 * to run the misfit task on.
11921 if (check_misfit_status(rq, sd)) {
11922 flags = NOHZ_STATS_KICK | NOHZ_BALANCE_KICK;
11927 * For asymmetric systems, we do not want to nicely balance
11928 * cache use, instead we want to embrace asymmetry and only
11929 * ensure tasks have enough CPU capacity.
11931 * Skip the LLC logic because it's not relevant in that case.
11936 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
11939 * If there is an imbalance between LLC domains (IOW we could
11940 * increase the overall cache utilization), we need a less-loaded LLC
11941 * domain to pull some load from. Likewise, we may need to spread
11942 * load within the current LLC domain (e.g. packed SMT cores but
11943 * other CPUs are idle). We can't really know from here how busy
11944 * the others are - so just get a NOHZ balance going if it looks
11945 * like this LLC domain has tasks we could move.
11947 nr_busy = atomic_read(&sds->nr_busy_cpus);
11949 flags = NOHZ_STATS_KICK | NOHZ_BALANCE_KICK;
11956 if (READ_ONCE(nohz.needs_update))
11957 flags |= NOHZ_NEXT_KICK;
11963 static void set_cpu_sd_state_busy(int cpu)
11965 struct sched_domain *sd;
11968 sd = rcu_dereference(per_cpu(sd_llc, cpu));
11970 if (!sd || !sd->nohz_idle)
11974 atomic_inc(&sd->shared->nr_busy_cpus);
11979 void nohz_balance_exit_idle(struct rq *rq)
11981 SCHED_WARN_ON(rq != this_rq());
11983 if (likely(!rq->nohz_tick_stopped))
11986 rq->nohz_tick_stopped = 0;
11987 cpumask_clear_cpu(rq->cpu, nohz.idle_cpus_mask);
11988 atomic_dec(&nohz.nr_cpus);
11990 set_cpu_sd_state_busy(rq->cpu);
11993 static void set_cpu_sd_state_idle(int cpu)
11995 struct sched_domain *sd;
11998 sd = rcu_dereference(per_cpu(sd_llc, cpu));
12000 if (!sd || sd->nohz_idle)
12004 atomic_dec(&sd->shared->nr_busy_cpus);
12010 * This routine will record that the CPU is going idle with tick stopped.
12011 * This info will be used in performing idle load balancing in the future.
12013 void nohz_balance_enter_idle(int cpu)
12015 struct rq *rq = cpu_rq(cpu);
12017 SCHED_WARN_ON(cpu != smp_processor_id());
12019 /* If this CPU is going down, then nothing needs to be done: */
12020 if (!cpu_active(cpu))
12023 /* Spare idle load balancing on CPUs that don't want to be disturbed: */
12024 if (!housekeeping_cpu(cpu, HK_TYPE_SCHED))
12028 * Can be set safely without rq->lock held
12029 * If a clear happens, it will have evaluated last additions because
12030 * rq->lock is held during the check and the clear
12032 rq->has_blocked_load = 1;
12035 * The tick is still stopped but load could have been added in the
12036 * meantime. We set the nohz.has_blocked flag to trig a check of the
12037 * *_avg. The CPU is already part of nohz.idle_cpus_mask so the clear
12038 * of nohz.has_blocked can only happen after checking the new load
12040 if (rq->nohz_tick_stopped)
12043 /* If we're a completely isolated CPU, we don't play: */
12044 if (on_null_domain(rq))
12047 rq->nohz_tick_stopped = 1;
12049 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
12050 atomic_inc(&nohz.nr_cpus);
12053 * Ensures that if nohz_idle_balance() fails to observe our
12054 * @idle_cpus_mask store, it must observe the @has_blocked
12055 * and @needs_update stores.
12057 smp_mb__after_atomic();
12059 set_cpu_sd_state_idle(cpu);
12061 WRITE_ONCE(nohz.needs_update, 1);
12064 * Each time a cpu enter idle, we assume that it has blocked load and
12065 * enable the periodic update of the load of idle cpus
12067 WRITE_ONCE(nohz.has_blocked, 1);
12070 static bool update_nohz_stats(struct rq *rq)
12072 unsigned int cpu = rq->cpu;
12074 if (!rq->has_blocked_load)
12077 if (!cpumask_test_cpu(cpu, nohz.idle_cpus_mask))
12080 if (!time_after(jiffies, READ_ONCE(rq->last_blocked_load_update_tick)))
12083 update_blocked_averages(cpu);
12085 return rq->has_blocked_load;
12089 * Internal function that runs load balance for all idle cpus. The load balance
12090 * can be a simple update of blocked load or a complete load balance with
12091 * tasks movement depending of flags.
12093 static void _nohz_idle_balance(struct rq *this_rq, unsigned int flags)
12095 /* Earliest time when we have to do rebalance again */
12096 unsigned long now = jiffies;
12097 unsigned long next_balance = now + 60*HZ;
12098 bool has_blocked_load = false;
12099 int update_next_balance = 0;
12100 int this_cpu = this_rq->cpu;
12104 SCHED_WARN_ON((flags & NOHZ_KICK_MASK) == NOHZ_BALANCE_KICK);
12107 * We assume there will be no idle load after this update and clear
12108 * the has_blocked flag. If a cpu enters idle in the mean time, it will
12109 * set the has_blocked flag and trigger another update of idle load.
12110 * Because a cpu that becomes idle, is added to idle_cpus_mask before
12111 * setting the flag, we are sure to not clear the state and not
12112 * check the load of an idle cpu.
12114 * Same applies to idle_cpus_mask vs needs_update.
12116 if (flags & NOHZ_STATS_KICK)
12117 WRITE_ONCE(nohz.has_blocked, 0);
12118 if (flags & NOHZ_NEXT_KICK)
12119 WRITE_ONCE(nohz.needs_update, 0);
12122 * Ensures that if we miss the CPU, we must see the has_blocked
12123 * store from nohz_balance_enter_idle().
12128 * Start with the next CPU after this_cpu so we will end with this_cpu and let a
12129 * chance for other idle cpu to pull load.
12131 for_each_cpu_wrap(balance_cpu, nohz.idle_cpus_mask, this_cpu+1) {
12132 if (!idle_cpu(balance_cpu))
12136 * If this CPU gets work to do, stop the load balancing
12137 * work being done for other CPUs. Next load
12138 * balancing owner will pick it up.
12140 if (need_resched()) {
12141 if (flags & NOHZ_STATS_KICK)
12142 has_blocked_load = true;
12143 if (flags & NOHZ_NEXT_KICK)
12144 WRITE_ONCE(nohz.needs_update, 1);
12148 rq = cpu_rq(balance_cpu);
12150 if (flags & NOHZ_STATS_KICK)
12151 has_blocked_load |= update_nohz_stats(rq);
12154 * If time for next balance is due,
12157 if (time_after_eq(jiffies, rq->next_balance)) {
12158 struct rq_flags rf;
12160 rq_lock_irqsave(rq, &rf);
12161 update_rq_clock(rq);
12162 rq_unlock_irqrestore(rq, &rf);
12164 if (flags & NOHZ_BALANCE_KICK)
12165 rebalance_domains(rq, CPU_IDLE);
12168 if (time_after(next_balance, rq->next_balance)) {
12169 next_balance = rq->next_balance;
12170 update_next_balance = 1;
12175 * next_balance will be updated only when there is a need.
12176 * When the CPU is attached to null domain for ex, it will not be
12179 if (likely(update_next_balance))
12180 nohz.next_balance = next_balance;
12182 if (flags & NOHZ_STATS_KICK)
12183 WRITE_ONCE(nohz.next_blocked,
12184 now + msecs_to_jiffies(LOAD_AVG_PERIOD));
12187 /* There is still blocked load, enable periodic update */
12188 if (has_blocked_load)
12189 WRITE_ONCE(nohz.has_blocked, 1);
12193 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
12194 * rebalancing for all the cpus for whom scheduler ticks are stopped.
12196 static bool nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
12198 unsigned int flags = this_rq->nohz_idle_balance;
12203 this_rq->nohz_idle_balance = 0;
12205 if (idle != CPU_IDLE)
12208 _nohz_idle_balance(this_rq, flags);
12214 * Check if we need to directly run the ILB for updating blocked load before
12215 * entering idle state. Here we run ILB directly without issuing IPIs.
12217 * Note that when this function is called, the tick may not yet be stopped on
12218 * this CPU yet. nohz.idle_cpus_mask is updated only when tick is stopped and
12219 * cleared on the next busy tick. In other words, nohz.idle_cpus_mask updates
12220 * don't align with CPUs enter/exit idle to avoid bottlenecks due to high idle
12221 * entry/exit rate (usec). So it is possible that _nohz_idle_balance() is
12222 * called from this function on (this) CPU that's not yet in the mask. That's
12223 * OK because the goal of nohz_run_idle_balance() is to run ILB only for
12224 * updating the blocked load of already idle CPUs without waking up one of
12225 * those idle CPUs and outside the preempt disable / irq off phase of the local
12226 * cpu about to enter idle, because it can take a long time.
12228 void nohz_run_idle_balance(int cpu)
12230 unsigned int flags;
12232 flags = atomic_fetch_andnot(NOHZ_NEWILB_KICK, nohz_flags(cpu));
12235 * Update the blocked load only if no SCHED_SOFTIRQ is about to happen
12236 * (ie NOHZ_STATS_KICK set) and will do the same.
12238 if ((flags == NOHZ_NEWILB_KICK) && !need_resched())
12239 _nohz_idle_balance(cpu_rq(cpu), NOHZ_STATS_KICK);
12242 static void nohz_newidle_balance(struct rq *this_rq)
12244 int this_cpu = this_rq->cpu;
12247 * This CPU doesn't want to be disturbed by scheduler
12250 if (!housekeeping_cpu(this_cpu, HK_TYPE_SCHED))
12253 /* Will wake up very soon. No time for doing anything else*/
12254 if (this_rq->avg_idle < sysctl_sched_migration_cost)
12257 /* Don't need to update blocked load of idle CPUs*/
12258 if (!READ_ONCE(nohz.has_blocked) ||
12259 time_before(jiffies, READ_ONCE(nohz.next_blocked)))
12263 * Set the need to trigger ILB in order to update blocked load
12264 * before entering idle state.
12266 atomic_or(NOHZ_NEWILB_KICK, nohz_flags(this_cpu));
12269 #else /* !CONFIG_NO_HZ_COMMON */
12270 static inline void nohz_balancer_kick(struct rq *rq) { }
12272 static inline bool nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
12277 static inline void nohz_newidle_balance(struct rq *this_rq) { }
12278 #endif /* CONFIG_NO_HZ_COMMON */
12281 * newidle_balance is called by schedule() if this_cpu is about to become
12282 * idle. Attempts to pull tasks from other CPUs.
12285 * < 0 - we released the lock and there are !fair tasks present
12286 * 0 - failed, no new tasks
12287 * > 0 - success, new (fair) tasks present
12289 static int newidle_balance(struct rq *this_rq, struct rq_flags *rf)
12291 unsigned long next_balance = jiffies + HZ;
12292 int this_cpu = this_rq->cpu;
12293 u64 t0, t1, curr_cost = 0;
12294 struct sched_domain *sd;
12295 int pulled_task = 0;
12297 update_misfit_status(NULL, this_rq);
12300 * There is a task waiting to run. No need to search for one.
12301 * Return 0; the task will be enqueued when switching to idle.
12303 if (this_rq->ttwu_pending)
12307 * We must set idle_stamp _before_ calling idle_balance(), such that we
12308 * measure the duration of idle_balance() as idle time.
12310 this_rq->idle_stamp = rq_clock(this_rq);
12313 * Do not pull tasks towards !active CPUs...
12315 if (!cpu_active(this_cpu))
12319 * This is OK, because current is on_cpu, which avoids it being picked
12320 * for load-balance and preemption/IRQs are still disabled avoiding
12321 * further scheduler activity on it and we're being very careful to
12322 * re-start the picking loop.
12324 rq_unpin_lock(this_rq, rf);
12327 sd = rcu_dereference_check_sched_domain(this_rq->sd);
12329 if (!READ_ONCE(this_rq->rd->overload) ||
12330 (sd && this_rq->avg_idle < sd->max_newidle_lb_cost)) {
12333 update_next_balance(sd, &next_balance);
12340 raw_spin_rq_unlock(this_rq);
12342 t0 = sched_clock_cpu(this_cpu);
12343 update_blocked_averages(this_cpu);
12346 for_each_domain(this_cpu, sd) {
12347 int continue_balancing = 1;
12350 update_next_balance(sd, &next_balance);
12352 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost)
12355 if (sd->flags & SD_BALANCE_NEWIDLE) {
12357 pulled_task = load_balance(this_cpu, this_rq,
12358 sd, CPU_NEWLY_IDLE,
12359 &continue_balancing);
12361 t1 = sched_clock_cpu(this_cpu);
12362 domain_cost = t1 - t0;
12363 update_newidle_cost(sd, domain_cost);
12365 curr_cost += domain_cost;
12370 * Stop searching for tasks to pull if there are
12371 * now runnable tasks on this rq.
12373 if (pulled_task || this_rq->nr_running > 0 ||
12374 this_rq->ttwu_pending)
12379 raw_spin_rq_lock(this_rq);
12381 if (curr_cost > this_rq->max_idle_balance_cost)
12382 this_rq->max_idle_balance_cost = curr_cost;
12385 * While browsing the domains, we released the rq lock, a task could
12386 * have been enqueued in the meantime. Since we're not going idle,
12387 * pretend we pulled a task.
12389 if (this_rq->cfs.h_nr_running && !pulled_task)
12392 /* Is there a task of a high priority class? */
12393 if (this_rq->nr_running != this_rq->cfs.h_nr_running)
12397 /* Move the next balance forward */
12398 if (time_after(this_rq->next_balance, next_balance))
12399 this_rq->next_balance = next_balance;
12402 this_rq->idle_stamp = 0;
12404 nohz_newidle_balance(this_rq);
12406 rq_repin_lock(this_rq, rf);
12408 return pulled_task;
12412 * run_rebalance_domains is triggered when needed from the scheduler tick.
12413 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
12415 static __latent_entropy void run_rebalance_domains(struct softirq_action *h)
12417 struct rq *this_rq = this_rq();
12418 enum cpu_idle_type idle = this_rq->idle_balance ?
12419 CPU_IDLE : CPU_NOT_IDLE;
12422 * If this CPU has a pending nohz_balance_kick, then do the
12423 * balancing on behalf of the other idle CPUs whose ticks are
12424 * stopped. Do nohz_idle_balance *before* rebalance_domains to
12425 * give the idle CPUs a chance to load balance. Else we may
12426 * load balance only within the local sched_domain hierarchy
12427 * and abort nohz_idle_balance altogether if we pull some load.
12429 if (nohz_idle_balance(this_rq, idle))
12432 /* normal load balance */
12433 update_blocked_averages(this_rq->cpu);
12434 rebalance_domains(this_rq, idle);
12438 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
12440 void trigger_load_balance(struct rq *rq)
12443 * Don't need to rebalance while attached to NULL domain or
12444 * runqueue CPU is not active
12446 if (unlikely(on_null_domain(rq) || !cpu_active(cpu_of(rq))))
12449 if (time_after_eq(jiffies, rq->next_balance))
12450 raise_softirq(SCHED_SOFTIRQ);
12452 nohz_balancer_kick(rq);
12455 static void rq_online_fair(struct rq *rq)
12459 update_runtime_enabled(rq);
12462 static void rq_offline_fair(struct rq *rq)
12466 /* Ensure any throttled groups are reachable by pick_next_task */
12467 unthrottle_offline_cfs_rqs(rq);
12469 /* Ensure that we remove rq contribution to group share: */
12470 clear_tg_offline_cfs_rqs(rq);
12473 #endif /* CONFIG_SMP */
12475 #ifdef CONFIG_SCHED_CORE
12477 __entity_slice_used(struct sched_entity *se, int min_nr_tasks)
12479 u64 rtime = se->sum_exec_runtime - se->prev_sum_exec_runtime;
12480 u64 slice = se->slice;
12482 return (rtime * min_nr_tasks > slice);
12485 #define MIN_NR_TASKS_DURING_FORCEIDLE 2
12486 static inline void task_tick_core(struct rq *rq, struct task_struct *curr)
12488 if (!sched_core_enabled(rq))
12492 * If runqueue has only one task which used up its slice and
12493 * if the sibling is forced idle, then trigger schedule to
12494 * give forced idle task a chance.
12496 * sched_slice() considers only this active rq and it gets the
12497 * whole slice. But during force idle, we have siblings acting
12498 * like a single runqueue and hence we need to consider runnable
12499 * tasks on this CPU and the forced idle CPU. Ideally, we should
12500 * go through the forced idle rq, but that would be a perf hit.
12501 * We can assume that the forced idle CPU has at least
12502 * MIN_NR_TASKS_DURING_FORCEIDLE - 1 tasks and use that to check
12503 * if we need to give up the CPU.
12505 if (rq->core->core_forceidle_count && rq->cfs.nr_running == 1 &&
12506 __entity_slice_used(&curr->se, MIN_NR_TASKS_DURING_FORCEIDLE))
12511 * se_fi_update - Update the cfs_rq->min_vruntime_fi in a CFS hierarchy if needed.
12513 static void se_fi_update(const struct sched_entity *se, unsigned int fi_seq,
12516 for_each_sched_entity(se) {
12517 struct cfs_rq *cfs_rq = cfs_rq_of(se);
12520 if (cfs_rq->forceidle_seq == fi_seq)
12522 cfs_rq->forceidle_seq = fi_seq;
12525 cfs_rq->min_vruntime_fi = cfs_rq->min_vruntime;
12529 void task_vruntime_update(struct rq *rq, struct task_struct *p, bool in_fi)
12531 struct sched_entity *se = &p->se;
12533 if (p->sched_class != &fair_sched_class)
12536 se_fi_update(se, rq->core->core_forceidle_seq, in_fi);
12539 bool cfs_prio_less(const struct task_struct *a, const struct task_struct *b,
12542 struct rq *rq = task_rq(a);
12543 const struct sched_entity *sea = &a->se;
12544 const struct sched_entity *seb = &b->se;
12545 struct cfs_rq *cfs_rqa;
12546 struct cfs_rq *cfs_rqb;
12549 SCHED_WARN_ON(task_rq(b)->core != rq->core);
12551 #ifdef CONFIG_FAIR_GROUP_SCHED
12553 * Find an se in the hierarchy for tasks a and b, such that the se's
12554 * are immediate siblings.
12556 while (sea->cfs_rq->tg != seb->cfs_rq->tg) {
12557 int sea_depth = sea->depth;
12558 int seb_depth = seb->depth;
12560 if (sea_depth >= seb_depth)
12561 sea = parent_entity(sea);
12562 if (sea_depth <= seb_depth)
12563 seb = parent_entity(seb);
12566 se_fi_update(sea, rq->core->core_forceidle_seq, in_fi);
12567 se_fi_update(seb, rq->core->core_forceidle_seq, in_fi);
12569 cfs_rqa = sea->cfs_rq;
12570 cfs_rqb = seb->cfs_rq;
12572 cfs_rqa = &task_rq(a)->cfs;
12573 cfs_rqb = &task_rq(b)->cfs;
12577 * Find delta after normalizing se's vruntime with its cfs_rq's
12578 * min_vruntime_fi, which would have been updated in prior calls
12579 * to se_fi_update().
12581 delta = (s64)(sea->vruntime - seb->vruntime) +
12582 (s64)(cfs_rqb->min_vruntime_fi - cfs_rqa->min_vruntime_fi);
12587 static int task_is_throttled_fair(struct task_struct *p, int cpu)
12589 struct cfs_rq *cfs_rq;
12591 #ifdef CONFIG_FAIR_GROUP_SCHED
12592 cfs_rq = task_group(p)->cfs_rq[cpu];
12594 cfs_rq = &cpu_rq(cpu)->cfs;
12596 return throttled_hierarchy(cfs_rq);
12599 static inline void task_tick_core(struct rq *rq, struct task_struct *curr) {}
12603 * scheduler tick hitting a task of our scheduling class.
12605 * NOTE: This function can be called remotely by the tick offload that
12606 * goes along full dynticks. Therefore no local assumption can be made
12607 * and everything must be accessed through the @rq and @curr passed in
12610 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
12612 struct cfs_rq *cfs_rq;
12613 struct sched_entity *se = &curr->se;
12615 for_each_sched_entity(se) {
12616 cfs_rq = cfs_rq_of(se);
12617 entity_tick(cfs_rq, se, queued);
12620 if (static_branch_unlikely(&sched_numa_balancing))
12621 task_tick_numa(rq, curr);
12623 update_misfit_status(curr, rq);
12624 update_overutilized_status(task_rq(curr));
12626 task_tick_core(rq, curr);
12630 * called on fork with the child task as argument from the parent's context
12631 * - child not yet on the tasklist
12632 * - preemption disabled
12634 static void task_fork_fair(struct task_struct *p)
12636 struct sched_entity *se = &p->se, *curr;
12637 struct cfs_rq *cfs_rq;
12638 struct rq *rq = this_rq();
12639 struct rq_flags rf;
12642 update_rq_clock(rq);
12644 cfs_rq = task_cfs_rq(current);
12645 curr = cfs_rq->curr;
12647 update_curr(cfs_rq);
12648 place_entity(cfs_rq, se, ENQUEUE_INITIAL);
12649 rq_unlock(rq, &rf);
12653 * Priority of the task has changed. Check to see if we preempt
12654 * the current task.
12657 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
12659 if (!task_on_rq_queued(p))
12662 if (rq->cfs.nr_running == 1)
12666 * Reschedule if we are currently running on this runqueue and
12667 * our priority decreased, or if we are not currently running on
12668 * this runqueue and our priority is higher than the current's
12670 if (task_current(rq, p)) {
12671 if (p->prio > oldprio)
12674 wakeup_preempt(rq, p, 0);
12677 #ifdef CONFIG_FAIR_GROUP_SCHED
12679 * Propagate the changes of the sched_entity across the tg tree to make it
12680 * visible to the root
12682 static void propagate_entity_cfs_rq(struct sched_entity *se)
12684 struct cfs_rq *cfs_rq = cfs_rq_of(se);
12686 if (cfs_rq_throttled(cfs_rq))
12689 if (!throttled_hierarchy(cfs_rq))
12690 list_add_leaf_cfs_rq(cfs_rq);
12692 /* Start to propagate at parent */
12695 for_each_sched_entity(se) {
12696 cfs_rq = cfs_rq_of(se);
12698 update_load_avg(cfs_rq, se, UPDATE_TG);
12700 if (cfs_rq_throttled(cfs_rq))
12703 if (!throttled_hierarchy(cfs_rq))
12704 list_add_leaf_cfs_rq(cfs_rq);
12708 static void propagate_entity_cfs_rq(struct sched_entity *se) { }
12711 static void detach_entity_cfs_rq(struct sched_entity *se)
12713 struct cfs_rq *cfs_rq = cfs_rq_of(se);
12717 * In case the task sched_avg hasn't been attached:
12718 * - A forked task which hasn't been woken up by wake_up_new_task().
12719 * - A task which has been woken up by try_to_wake_up() but is
12720 * waiting for actually being woken up by sched_ttwu_pending().
12722 if (!se->avg.last_update_time)
12726 /* Catch up with the cfs_rq and remove our load when we leave */
12727 update_load_avg(cfs_rq, se, 0);
12728 detach_entity_load_avg(cfs_rq, se);
12729 update_tg_load_avg(cfs_rq);
12730 propagate_entity_cfs_rq(se);
12733 static void attach_entity_cfs_rq(struct sched_entity *se)
12735 struct cfs_rq *cfs_rq = cfs_rq_of(se);
12737 /* Synchronize entity with its cfs_rq */
12738 update_load_avg(cfs_rq, se, sched_feat(ATTACH_AGE_LOAD) ? 0 : SKIP_AGE_LOAD);
12739 attach_entity_load_avg(cfs_rq, se);
12740 update_tg_load_avg(cfs_rq);
12741 propagate_entity_cfs_rq(se);
12744 static void detach_task_cfs_rq(struct task_struct *p)
12746 struct sched_entity *se = &p->se;
12748 detach_entity_cfs_rq(se);
12751 static void attach_task_cfs_rq(struct task_struct *p)
12753 struct sched_entity *se = &p->se;
12755 attach_entity_cfs_rq(se);
12758 static void switched_from_fair(struct rq *rq, struct task_struct *p)
12760 detach_task_cfs_rq(p);
12763 static void switched_to_fair(struct rq *rq, struct task_struct *p)
12765 attach_task_cfs_rq(p);
12767 if (task_on_rq_queued(p)) {
12769 * We were most likely switched from sched_rt, so
12770 * kick off the schedule if running, otherwise just see
12771 * if we can still preempt the current task.
12773 if (task_current(rq, p))
12776 wakeup_preempt(rq, p, 0);
12780 /* Account for a task changing its policy or group.
12782 * This routine is mostly called to set cfs_rq->curr field when a task
12783 * migrates between groups/classes.
12785 static void set_next_task_fair(struct rq *rq, struct task_struct *p, bool first)
12787 struct sched_entity *se = &p->se;
12790 if (task_on_rq_queued(p)) {
12792 * Move the next running task to the front of the list, so our
12793 * cfs_tasks list becomes MRU one.
12795 list_move(&se->group_node, &rq->cfs_tasks);
12799 for_each_sched_entity(se) {
12800 struct cfs_rq *cfs_rq = cfs_rq_of(se);
12802 set_next_entity(cfs_rq, se);
12803 /* ensure bandwidth has been allocated on our new cfs_rq */
12804 account_cfs_rq_runtime(cfs_rq, 0);
12808 void init_cfs_rq(struct cfs_rq *cfs_rq)
12810 cfs_rq->tasks_timeline = RB_ROOT_CACHED;
12811 u64_u32_store(cfs_rq->min_vruntime, (u64)(-(1LL << 20)));
12813 raw_spin_lock_init(&cfs_rq->removed.lock);
12817 #ifdef CONFIG_FAIR_GROUP_SCHED
12818 static void task_change_group_fair(struct task_struct *p)
12821 * We couldn't detach or attach a forked task which
12822 * hasn't been woken up by wake_up_new_task().
12824 if (READ_ONCE(p->__state) == TASK_NEW)
12827 detach_task_cfs_rq(p);
12830 /* Tell se's cfs_rq has been changed -- migrated */
12831 p->se.avg.last_update_time = 0;
12833 set_task_rq(p, task_cpu(p));
12834 attach_task_cfs_rq(p);
12837 void free_fair_sched_group(struct task_group *tg)
12841 for_each_possible_cpu(i) {
12843 kfree(tg->cfs_rq[i]);
12852 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
12854 struct sched_entity *se;
12855 struct cfs_rq *cfs_rq;
12858 tg->cfs_rq = kcalloc(nr_cpu_ids, sizeof(cfs_rq), GFP_KERNEL);
12861 tg->se = kcalloc(nr_cpu_ids, sizeof(se), GFP_KERNEL);
12865 tg->shares = NICE_0_LOAD;
12867 init_cfs_bandwidth(tg_cfs_bandwidth(tg), tg_cfs_bandwidth(parent));
12869 for_each_possible_cpu(i) {
12870 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
12871 GFP_KERNEL, cpu_to_node(i));
12875 se = kzalloc_node(sizeof(struct sched_entity_stats),
12876 GFP_KERNEL, cpu_to_node(i));
12880 init_cfs_rq(cfs_rq);
12881 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
12882 init_entity_runnable_average(se);
12893 void online_fair_sched_group(struct task_group *tg)
12895 struct sched_entity *se;
12896 struct rq_flags rf;
12900 for_each_possible_cpu(i) {
12903 rq_lock_irq(rq, &rf);
12904 update_rq_clock(rq);
12905 attach_entity_cfs_rq(se);
12906 sync_throttle(tg, i);
12907 rq_unlock_irq(rq, &rf);
12911 void unregister_fair_sched_group(struct task_group *tg)
12913 unsigned long flags;
12917 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
12919 for_each_possible_cpu(cpu) {
12921 remove_entity_load_avg(tg->se[cpu]);
12924 * Only empty task groups can be destroyed; so we can speculatively
12925 * check on_list without danger of it being re-added.
12927 if (!tg->cfs_rq[cpu]->on_list)
12932 raw_spin_rq_lock_irqsave(rq, flags);
12933 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
12934 raw_spin_rq_unlock_irqrestore(rq, flags);
12938 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
12939 struct sched_entity *se, int cpu,
12940 struct sched_entity *parent)
12942 struct rq *rq = cpu_rq(cpu);
12946 init_cfs_rq_runtime(cfs_rq);
12948 tg->cfs_rq[cpu] = cfs_rq;
12951 /* se could be NULL for root_task_group */
12956 se->cfs_rq = &rq->cfs;
12959 se->cfs_rq = parent->my_q;
12960 se->depth = parent->depth + 1;
12964 /* guarantee group entities always have weight */
12965 update_load_set(&se->load, NICE_0_LOAD);
12966 se->parent = parent;
12969 static DEFINE_MUTEX(shares_mutex);
12971 static int __sched_group_set_shares(struct task_group *tg, unsigned long shares)
12975 lockdep_assert_held(&shares_mutex);
12978 * We can't change the weight of the root cgroup.
12983 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
12985 if (tg->shares == shares)
12988 tg->shares = shares;
12989 for_each_possible_cpu(i) {
12990 struct rq *rq = cpu_rq(i);
12991 struct sched_entity *se = tg->se[i];
12992 struct rq_flags rf;
12994 /* Propagate contribution to hierarchy */
12995 rq_lock_irqsave(rq, &rf);
12996 update_rq_clock(rq);
12997 for_each_sched_entity(se) {
12998 update_load_avg(cfs_rq_of(se), se, UPDATE_TG);
12999 update_cfs_group(se);
13001 rq_unlock_irqrestore(rq, &rf);
13007 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
13011 mutex_lock(&shares_mutex);
13012 if (tg_is_idle(tg))
13015 ret = __sched_group_set_shares(tg, shares);
13016 mutex_unlock(&shares_mutex);
13021 int sched_group_set_idle(struct task_group *tg, long idle)
13025 if (tg == &root_task_group)
13028 if (idle < 0 || idle > 1)
13031 mutex_lock(&shares_mutex);
13033 if (tg->idle == idle) {
13034 mutex_unlock(&shares_mutex);
13040 for_each_possible_cpu(i) {
13041 struct rq *rq = cpu_rq(i);
13042 struct sched_entity *se = tg->se[i];
13043 struct cfs_rq *parent_cfs_rq, *grp_cfs_rq = tg->cfs_rq[i];
13044 bool was_idle = cfs_rq_is_idle(grp_cfs_rq);
13045 long idle_task_delta;
13046 struct rq_flags rf;
13048 rq_lock_irqsave(rq, &rf);
13050 grp_cfs_rq->idle = idle;
13051 if (WARN_ON_ONCE(was_idle == cfs_rq_is_idle(grp_cfs_rq)))
13055 parent_cfs_rq = cfs_rq_of(se);
13056 if (cfs_rq_is_idle(grp_cfs_rq))
13057 parent_cfs_rq->idle_nr_running++;
13059 parent_cfs_rq->idle_nr_running--;
13062 idle_task_delta = grp_cfs_rq->h_nr_running -
13063 grp_cfs_rq->idle_h_nr_running;
13064 if (!cfs_rq_is_idle(grp_cfs_rq))
13065 idle_task_delta *= -1;
13067 for_each_sched_entity(se) {
13068 struct cfs_rq *cfs_rq = cfs_rq_of(se);
13073 cfs_rq->idle_h_nr_running += idle_task_delta;
13075 /* Already accounted at parent level and above. */
13076 if (cfs_rq_is_idle(cfs_rq))
13081 rq_unlock_irqrestore(rq, &rf);
13084 /* Idle groups have minimum weight. */
13085 if (tg_is_idle(tg))
13086 __sched_group_set_shares(tg, scale_load(WEIGHT_IDLEPRIO));
13088 __sched_group_set_shares(tg, NICE_0_LOAD);
13090 mutex_unlock(&shares_mutex);
13094 #endif /* CONFIG_FAIR_GROUP_SCHED */
13097 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
13099 struct sched_entity *se = &task->se;
13100 unsigned int rr_interval = 0;
13103 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
13106 if (rq->cfs.load.weight)
13107 rr_interval = NS_TO_JIFFIES(se->slice);
13109 return rr_interval;
13113 * All the scheduling class methods:
13115 DEFINE_SCHED_CLASS(fair) = {
13117 .enqueue_task = enqueue_task_fair,
13118 .dequeue_task = dequeue_task_fair,
13119 .yield_task = yield_task_fair,
13120 .yield_to_task = yield_to_task_fair,
13122 .wakeup_preempt = check_preempt_wakeup_fair,
13124 .pick_next_task = __pick_next_task_fair,
13125 .put_prev_task = put_prev_task_fair,
13126 .set_next_task = set_next_task_fair,
13129 .balance = balance_fair,
13130 .pick_task = pick_task_fair,
13131 .select_task_rq = select_task_rq_fair,
13132 .migrate_task_rq = migrate_task_rq_fair,
13134 .rq_online = rq_online_fair,
13135 .rq_offline = rq_offline_fair,
13137 .task_dead = task_dead_fair,
13138 .set_cpus_allowed = set_cpus_allowed_common,
13141 .task_tick = task_tick_fair,
13142 .task_fork = task_fork_fair,
13144 .prio_changed = prio_changed_fair,
13145 .switched_from = switched_from_fair,
13146 .switched_to = switched_to_fair,
13148 .get_rr_interval = get_rr_interval_fair,
13150 .update_curr = update_curr_fair,
13152 #ifdef CONFIG_FAIR_GROUP_SCHED
13153 .task_change_group = task_change_group_fair,
13156 #ifdef CONFIG_SCHED_CORE
13157 .task_is_throttled = task_is_throttled_fair,
13160 #ifdef CONFIG_UCLAMP_TASK
13161 .uclamp_enabled = 1,
13165 #ifdef CONFIG_SCHED_DEBUG
13166 void print_cfs_stats(struct seq_file *m, int cpu)
13168 struct cfs_rq *cfs_rq, *pos;
13171 for_each_leaf_cfs_rq_safe(cpu_rq(cpu), cfs_rq, pos)
13172 print_cfs_rq(m, cpu, cfs_rq);
13176 #ifdef CONFIG_NUMA_BALANCING
13177 void show_numa_stats(struct task_struct *p, struct seq_file *m)
13180 unsigned long tsf = 0, tpf = 0, gsf = 0, gpf = 0;
13181 struct numa_group *ng;
13184 ng = rcu_dereference(p->numa_group);
13185 for_each_online_node(node) {
13186 if (p->numa_faults) {
13187 tsf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 0)];
13188 tpf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 1)];
13191 gsf = ng->faults[task_faults_idx(NUMA_MEM, node, 0)],
13192 gpf = ng->faults[task_faults_idx(NUMA_MEM, node, 1)];
13194 print_numa_stats(m, node, tsf, tpf, gsf, gpf);
13198 #endif /* CONFIG_NUMA_BALANCING */
13199 #endif /* CONFIG_SCHED_DEBUG */
13201 __init void init_sched_fair_class(void)
13206 for_each_possible_cpu(i) {
13207 zalloc_cpumask_var_node(&per_cpu(load_balance_mask, i), GFP_KERNEL, cpu_to_node(i));
13208 zalloc_cpumask_var_node(&per_cpu(select_rq_mask, i), GFP_KERNEL, cpu_to_node(i));
13209 zalloc_cpumask_var_node(&per_cpu(should_we_balance_tmpmask, i),
13210 GFP_KERNEL, cpu_to_node(i));
13212 #ifdef CONFIG_CFS_BANDWIDTH
13213 INIT_CSD(&cpu_rq(i)->cfsb_csd, __cfsb_csd_unthrottle, cpu_rq(i));
13214 INIT_LIST_HEAD(&cpu_rq(i)->cfsb_csd_list);
13218 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
13220 #ifdef CONFIG_NO_HZ_COMMON
13221 nohz.next_balance = jiffies;
13222 nohz.next_blocked = jiffies;
13223 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);