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>
54 #include <linux/sched/cond_resched.h>
58 #include "autogroup.h"
61 * The initial- and re-scaling of tunables is configurable
65 * SCHED_TUNABLESCALING_NONE - unscaled, always *1
66 * SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus)
67 * SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus
69 * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus))
71 unsigned int sysctl_sched_tunable_scaling = SCHED_TUNABLESCALING_LOG;
74 * Minimal preemption granularity for CPU-bound tasks:
76 * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds)
78 unsigned int sysctl_sched_base_slice = 750000ULL;
79 static unsigned int normalized_sysctl_sched_base_slice = 750000ULL;
82 * After fork, child runs first. If set to 0 (default) then
83 * parent will (try to) run first.
85 unsigned int sysctl_sched_child_runs_first __read_mostly;
87 const_debug unsigned int sysctl_sched_migration_cost = 500000UL;
89 int sched_thermal_decay_shift;
90 static int __init setup_sched_thermal_decay_shift(char *str)
94 if (kstrtoint(str, 0, &_shift))
95 pr_warn("Unable to set scheduler thermal pressure decay shift parameter\n");
97 sched_thermal_decay_shift = clamp(_shift, 0, 10);
100 __setup("sched_thermal_decay_shift=", setup_sched_thermal_decay_shift);
104 * For asym packing, by default the lower numbered CPU has higher priority.
106 int __weak arch_asym_cpu_priority(int cpu)
112 * The margin used when comparing utilization with CPU capacity.
116 #define fits_capacity(cap, max) ((cap) * 1280 < (max) * 1024)
119 * The margin used when comparing CPU capacities.
120 * is 'cap1' noticeably greater than 'cap2'
124 #define capacity_greater(cap1, cap2) ((cap1) * 1024 > (cap2) * 1078)
127 #ifdef CONFIG_CFS_BANDWIDTH
129 * Amount of runtime to allocate from global (tg) to local (per-cfs_rq) pool
130 * each time a cfs_rq requests quota.
132 * Note: in the case that the slice exceeds the runtime remaining (either due
133 * to consumption or the quota being specified to be smaller than the slice)
134 * we will always only issue the remaining available time.
136 * (default: 5 msec, units: microseconds)
138 static unsigned int sysctl_sched_cfs_bandwidth_slice = 5000UL;
141 #ifdef CONFIG_NUMA_BALANCING
142 /* Restrict the NUMA promotion throughput (MB/s) for each target node. */
143 static unsigned int sysctl_numa_balancing_promote_rate_limit = 65536;
147 static struct ctl_table sched_fair_sysctls[] = {
149 .procname = "sched_child_runs_first",
150 .data = &sysctl_sched_child_runs_first,
151 .maxlen = sizeof(unsigned int),
153 .proc_handler = proc_dointvec,
155 #ifdef CONFIG_CFS_BANDWIDTH
157 .procname = "sched_cfs_bandwidth_slice_us",
158 .data = &sysctl_sched_cfs_bandwidth_slice,
159 .maxlen = sizeof(unsigned int),
161 .proc_handler = proc_dointvec_minmax,
162 .extra1 = SYSCTL_ONE,
165 #ifdef CONFIG_NUMA_BALANCING
167 .procname = "numa_balancing_promote_rate_limit_MBps",
168 .data = &sysctl_numa_balancing_promote_rate_limit,
169 .maxlen = sizeof(unsigned int),
171 .proc_handler = proc_dointvec_minmax,
172 .extra1 = SYSCTL_ZERO,
174 #endif /* CONFIG_NUMA_BALANCING */
178 static int __init sched_fair_sysctl_init(void)
180 register_sysctl_init("kernel", sched_fair_sysctls);
183 late_initcall(sched_fair_sysctl_init);
186 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
192 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
198 static inline void update_load_set(struct load_weight *lw, unsigned long w)
205 * Increase the granularity value when there are more CPUs,
206 * because with more CPUs the 'effective latency' as visible
207 * to users decreases. But the relationship is not linear,
208 * so pick a second-best guess by going with the log2 of the
211 * This idea comes from the SD scheduler of Con Kolivas:
213 static unsigned int get_update_sysctl_factor(void)
215 unsigned int cpus = min_t(unsigned int, num_online_cpus(), 8);
218 switch (sysctl_sched_tunable_scaling) {
219 case SCHED_TUNABLESCALING_NONE:
222 case SCHED_TUNABLESCALING_LINEAR:
225 case SCHED_TUNABLESCALING_LOG:
227 factor = 1 + ilog2(cpus);
234 static void update_sysctl(void)
236 unsigned int factor = get_update_sysctl_factor();
238 #define SET_SYSCTL(name) \
239 (sysctl_##name = (factor) * normalized_sysctl_##name)
240 SET_SYSCTL(sched_base_slice);
244 void __init sched_init_granularity(void)
249 #define WMULT_CONST (~0U)
250 #define WMULT_SHIFT 32
252 static void __update_inv_weight(struct load_weight *lw)
256 if (likely(lw->inv_weight))
259 w = scale_load_down(lw->weight);
261 if (BITS_PER_LONG > 32 && unlikely(w >= WMULT_CONST))
263 else if (unlikely(!w))
264 lw->inv_weight = WMULT_CONST;
266 lw->inv_weight = WMULT_CONST / w;
270 * delta_exec * weight / lw.weight
272 * (delta_exec * (weight * lw->inv_weight)) >> WMULT_SHIFT
274 * Either weight := NICE_0_LOAD and lw \e sched_prio_to_wmult[], in which case
275 * we're guaranteed shift stays positive because inv_weight is guaranteed to
276 * fit 32 bits, and NICE_0_LOAD gives another 10 bits; therefore shift >= 22.
278 * Or, weight =< lw.weight (because lw.weight is the runqueue weight), thus
279 * weight/lw.weight <= 1, and therefore our shift will also be positive.
281 static u64 __calc_delta(u64 delta_exec, unsigned long weight, struct load_weight *lw)
283 u64 fact = scale_load_down(weight);
284 u32 fact_hi = (u32)(fact >> 32);
285 int shift = WMULT_SHIFT;
288 __update_inv_weight(lw);
290 if (unlikely(fact_hi)) {
296 fact = mul_u32_u32(fact, lw->inv_weight);
298 fact_hi = (u32)(fact >> 32);
305 return mul_u64_u32_shr(delta_exec, fact, shift);
311 static inline u64 calc_delta_fair(u64 delta, struct sched_entity *se)
313 if (unlikely(se->load.weight != NICE_0_LOAD))
314 delta = __calc_delta(delta, NICE_0_LOAD, &se->load);
319 const struct sched_class fair_sched_class;
321 /**************************************************************
322 * CFS operations on generic schedulable entities:
325 #ifdef CONFIG_FAIR_GROUP_SCHED
327 /* Walk up scheduling entities hierarchy */
328 #define for_each_sched_entity(se) \
329 for (; se; se = se->parent)
331 static inline bool list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
333 struct rq *rq = rq_of(cfs_rq);
334 int cpu = cpu_of(rq);
337 return rq->tmp_alone_branch == &rq->leaf_cfs_rq_list;
342 * Ensure we either appear before our parent (if already
343 * enqueued) or force our parent to appear after us when it is
344 * enqueued. The fact that we always enqueue bottom-up
345 * reduces this to two cases and a special case for the root
346 * cfs_rq. Furthermore, it also means that we will always reset
347 * tmp_alone_branch either when the branch is connected
348 * to a tree or when we reach the top of the tree
350 if (cfs_rq->tg->parent &&
351 cfs_rq->tg->parent->cfs_rq[cpu]->on_list) {
353 * If parent is already on the list, we add the child
354 * just before. Thanks to circular linked property of
355 * the list, this means to put the child at the tail
356 * of the list that starts by parent.
358 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
359 &(cfs_rq->tg->parent->cfs_rq[cpu]->leaf_cfs_rq_list));
361 * The branch is now connected to its tree so we can
362 * reset tmp_alone_branch to the beginning of the
365 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
369 if (!cfs_rq->tg->parent) {
371 * cfs rq without parent should be put
372 * at the tail of the list.
374 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
375 &rq->leaf_cfs_rq_list);
377 * We have reach the top of a tree so we can reset
378 * tmp_alone_branch to the beginning of the list.
380 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
385 * The parent has not already been added so we want to
386 * make sure that it will be put after us.
387 * tmp_alone_branch points to the begin of the branch
388 * where we will add parent.
390 list_add_rcu(&cfs_rq->leaf_cfs_rq_list, rq->tmp_alone_branch);
392 * update tmp_alone_branch to points to the new begin
395 rq->tmp_alone_branch = &cfs_rq->leaf_cfs_rq_list;
399 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
401 if (cfs_rq->on_list) {
402 struct rq *rq = rq_of(cfs_rq);
405 * With cfs_rq being unthrottled/throttled during an enqueue,
406 * it can happen the tmp_alone_branch points the a leaf that
407 * we finally want to del. In this case, tmp_alone_branch moves
408 * to the prev element but it will point to rq->leaf_cfs_rq_list
409 * at the end of the enqueue.
411 if (rq->tmp_alone_branch == &cfs_rq->leaf_cfs_rq_list)
412 rq->tmp_alone_branch = cfs_rq->leaf_cfs_rq_list.prev;
414 list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
419 static inline void assert_list_leaf_cfs_rq(struct rq *rq)
421 SCHED_WARN_ON(rq->tmp_alone_branch != &rq->leaf_cfs_rq_list);
424 /* Iterate thr' all leaf cfs_rq's on a runqueue */
425 #define for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) \
426 list_for_each_entry_safe(cfs_rq, pos, &rq->leaf_cfs_rq_list, \
429 /* Do the two (enqueued) entities belong to the same group ? */
430 static inline struct cfs_rq *
431 is_same_group(struct sched_entity *se, struct sched_entity *pse)
433 if (se->cfs_rq == pse->cfs_rq)
439 static inline struct sched_entity *parent_entity(const struct sched_entity *se)
445 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
447 int se_depth, pse_depth;
450 * preemption test can be made between sibling entities who are in the
451 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
452 * both tasks until we find their ancestors who are siblings of common
456 /* First walk up until both entities are at same depth */
457 se_depth = (*se)->depth;
458 pse_depth = (*pse)->depth;
460 while (se_depth > pse_depth) {
462 *se = parent_entity(*se);
465 while (pse_depth > se_depth) {
467 *pse = parent_entity(*pse);
470 while (!is_same_group(*se, *pse)) {
471 *se = parent_entity(*se);
472 *pse = parent_entity(*pse);
476 static int tg_is_idle(struct task_group *tg)
481 static int cfs_rq_is_idle(struct cfs_rq *cfs_rq)
483 return cfs_rq->idle > 0;
486 static int se_is_idle(struct sched_entity *se)
488 if (entity_is_task(se))
489 return task_has_idle_policy(task_of(se));
490 return cfs_rq_is_idle(group_cfs_rq(se));
493 #else /* !CONFIG_FAIR_GROUP_SCHED */
495 #define for_each_sched_entity(se) \
496 for (; se; se = NULL)
498 static inline bool list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
503 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
507 static inline void assert_list_leaf_cfs_rq(struct rq *rq)
511 #define for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) \
512 for (cfs_rq = &rq->cfs, pos = NULL; cfs_rq; cfs_rq = pos)
514 static inline struct sched_entity *parent_entity(struct sched_entity *se)
520 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
524 static inline int tg_is_idle(struct task_group *tg)
529 static int cfs_rq_is_idle(struct cfs_rq *cfs_rq)
534 static int se_is_idle(struct sched_entity *se)
539 #endif /* CONFIG_FAIR_GROUP_SCHED */
541 static __always_inline
542 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec);
544 /**************************************************************
545 * Scheduling class tree data structure manipulation methods:
548 static inline u64 max_vruntime(u64 max_vruntime, u64 vruntime)
550 s64 delta = (s64)(vruntime - max_vruntime);
552 max_vruntime = vruntime;
557 static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime)
559 s64 delta = (s64)(vruntime - min_vruntime);
561 min_vruntime = vruntime;
566 static inline bool entity_before(const struct sched_entity *a,
567 const struct sched_entity *b)
569 return (s64)(a->vruntime - b->vruntime) < 0;
572 static inline s64 entity_key(struct cfs_rq *cfs_rq, struct sched_entity *se)
574 return (s64)(se->vruntime - cfs_rq->min_vruntime);
577 #define __node_2_se(node) \
578 rb_entry((node), struct sched_entity, run_node)
581 * Compute virtual time from the per-task service numbers:
583 * Fair schedulers conserve lag:
587 * Where lag_i is given by:
589 * lag_i = S - s_i = w_i * (V - v_i)
591 * Where S is the ideal service time and V is it's virtual time counterpart.
595 * \Sum w_i * (V - v_i) = 0
596 * \Sum w_i * V - w_i * v_i = 0
598 * From which we can solve an expression for V in v_i (which we have in
601 * \Sum v_i * w_i \Sum v_i * w_i
602 * V = -------------- = --------------
605 * Specifically, this is the weighted average of all entity virtual runtimes.
607 * [[ NOTE: this is only equal to the ideal scheduler under the condition
608 * that join/leave operations happen at lag_i = 0, otherwise the
609 * virtual time has non-continguous motion equivalent to:
613 * Also see the comment in place_entity() that deals with this. ]]
615 * However, since v_i is u64, and the multiplcation could easily overflow
616 * transform it into a relative form that uses smaller quantities:
618 * Substitute: v_i == (v_i - v0) + v0
620 * \Sum ((v_i - v0) + v0) * w_i \Sum (v_i - v0) * w_i
621 * V = ---------------------------- = --------------------- + v0
624 * Which we track using:
626 * v0 := cfs_rq->min_vruntime
627 * \Sum (v_i - v0) * w_i := cfs_rq->avg_vruntime
628 * \Sum w_i := cfs_rq->avg_load
630 * Since min_vruntime is a monotonic increasing variable that closely tracks
631 * the per-task service, these deltas: (v_i - v), will be in the order of the
632 * maximal (virtual) lag induced in the system due to quantisation.
634 * Also, we use scale_load_down() to reduce the size.
636 * As measured, the max (key * weight) value was ~44 bits for a kernel build.
639 avg_vruntime_add(struct cfs_rq *cfs_rq, struct sched_entity *se)
641 unsigned long weight = scale_load_down(se->load.weight);
642 s64 key = entity_key(cfs_rq, se);
644 cfs_rq->avg_vruntime += key * weight;
645 cfs_rq->avg_load += weight;
649 avg_vruntime_sub(struct cfs_rq *cfs_rq, struct sched_entity *se)
651 unsigned long weight = scale_load_down(se->load.weight);
652 s64 key = entity_key(cfs_rq, se);
654 cfs_rq->avg_vruntime -= key * weight;
655 cfs_rq->avg_load -= weight;
659 void avg_vruntime_update(struct cfs_rq *cfs_rq, s64 delta)
662 * v' = v + d ==> avg_vruntime' = avg_runtime - d*avg_load
664 cfs_rq->avg_vruntime -= cfs_rq->avg_load * delta;
668 * Specifically: avg_runtime() + 0 must result in entity_eligible() := true
669 * For this to be so, the result of this function must have a left bias.
671 u64 avg_vruntime(struct cfs_rq *cfs_rq)
673 struct sched_entity *curr = cfs_rq->curr;
674 s64 avg = cfs_rq->avg_vruntime;
675 long load = cfs_rq->avg_load;
677 if (curr && curr->on_rq) {
678 unsigned long weight = scale_load_down(curr->load.weight);
680 avg += entity_key(cfs_rq, curr) * weight;
685 /* sign flips effective floor / ceil */
688 avg = div_s64(avg, load);
691 return cfs_rq->min_vruntime + avg;
695 * lag_i = S - s_i = w_i * (V - v_i)
697 * However, since V is approximated by the weighted average of all entities it
698 * is possible -- by addition/removal/reweight to the tree -- to move V around
699 * and end up with a larger lag than we started with.
701 * Limit this to either double the slice length with a minimum of TICK_NSEC
702 * since that is the timing granularity.
704 * EEVDF gives the following limit for a steady state system:
706 * -r_max < lag < max(r_max, q)
708 * XXX could add max_slice to the augmented data to track this.
710 static void update_entity_lag(struct cfs_rq *cfs_rq, struct sched_entity *se)
714 SCHED_WARN_ON(!se->on_rq);
715 lag = avg_vruntime(cfs_rq) - se->vruntime;
717 limit = calc_delta_fair(max_t(u64, 2*se->slice, TICK_NSEC), se);
718 se->vlag = clamp(lag, -limit, limit);
722 * Entity is eligible once it received less service than it ought to have,
725 * lag_i = S - s_i = w_i*(V - v_i)
727 * lag_i >= 0 -> V >= v_i
730 * V = ------------------ + v
733 * lag_i >= 0 -> \Sum (v_i - v)*w_i >= (v_i - v)*(\Sum w_i)
735 * Note: using 'avg_vruntime() > se->vruntime' is inacurate due
736 * to the loss in precision caused by the division.
738 int entity_eligible(struct cfs_rq *cfs_rq, struct sched_entity *se)
740 struct sched_entity *curr = cfs_rq->curr;
741 s64 avg = cfs_rq->avg_vruntime;
742 long load = cfs_rq->avg_load;
744 if (curr && curr->on_rq) {
745 unsigned long weight = scale_load_down(curr->load.weight);
747 avg += entity_key(cfs_rq, curr) * weight;
751 return avg >= entity_key(cfs_rq, se) * load;
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_first_entity(cfs_rq);
771 struct sched_entity *curr = cfs_rq->curr;
773 u64 vruntime = cfs_rq->min_vruntime;
777 vruntime = curr->vruntime;
784 vruntime = se->vruntime;
786 vruntime = min_vruntime(vruntime, se->vruntime);
789 /* ensure we never gain time by being placed backwards. */
790 u64_u32_store(cfs_rq->min_vruntime,
791 __update_min_vruntime(cfs_rq, vruntime));
794 static inline bool __entity_less(struct rb_node *a, const struct rb_node *b)
796 return entity_before(__node_2_se(a), __node_2_se(b));
799 #define deadline_gt(field, lse, rse) ({ (s64)((lse)->field - (rse)->field) > 0; })
801 static inline void __update_min_deadline(struct sched_entity *se, struct rb_node *node)
804 struct sched_entity *rse = __node_2_se(node);
805 if (deadline_gt(min_deadline, se, rse))
806 se->min_deadline = rse->min_deadline;
811 * se->min_deadline = min(se->deadline, left->min_deadline, right->min_deadline)
813 static inline bool min_deadline_update(struct sched_entity *se, bool exit)
815 u64 old_min_deadline = se->min_deadline;
816 struct rb_node *node = &se->run_node;
818 se->min_deadline = se->deadline;
819 __update_min_deadline(se, node->rb_right);
820 __update_min_deadline(se, node->rb_left);
822 return se->min_deadline == old_min_deadline;
825 RB_DECLARE_CALLBACKS(static, min_deadline_cb, struct sched_entity,
826 run_node, min_deadline, min_deadline_update);
829 * Enqueue an entity into the rb-tree:
831 static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
833 avg_vruntime_add(cfs_rq, se);
834 se->min_deadline = se->deadline;
835 rb_add_augmented_cached(&se->run_node, &cfs_rq->tasks_timeline,
836 __entity_less, &min_deadline_cb);
839 static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
841 rb_erase_augmented_cached(&se->run_node, &cfs_rq->tasks_timeline,
843 avg_vruntime_sub(cfs_rq, se);
846 struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq)
848 struct rb_node *left = rb_first_cached(&cfs_rq->tasks_timeline);
853 return __node_2_se(left);
857 * Earliest Eligible Virtual Deadline First
859 * In order to provide latency guarantees for different request sizes
860 * EEVDF selects the best runnable task from two criteria:
862 * 1) the task must be eligible (must be owed service)
864 * 2) from those tasks that meet 1), we select the one
865 * with the earliest virtual deadline.
867 * We can do this in O(log n) time due to an augmented RB-tree. The
868 * tree keeps the entries sorted on service, but also functions as a
869 * heap based on the deadline by keeping:
871 * se->min_deadline = min(se->deadline, se->{left,right}->min_deadline)
873 * Which allows an EDF like search on (sub)trees.
875 static struct sched_entity *__pick_eevdf(struct cfs_rq *cfs_rq)
877 struct rb_node *node = cfs_rq->tasks_timeline.rb_root.rb_node;
878 struct sched_entity *curr = cfs_rq->curr;
879 struct sched_entity *best = NULL;
880 struct sched_entity *best_left = NULL;
882 if (curr && (!curr->on_rq || !entity_eligible(cfs_rq, curr)))
887 * Once selected, run a task until it either becomes non-eligible or
888 * until it gets a new slice. See the HACK in set_next_entity().
890 if (sched_feat(RUN_TO_PARITY) && curr && curr->vlag == curr->deadline)
894 struct sched_entity *se = __node_2_se(node);
897 * If this entity is not eligible, try the left subtree.
899 if (!entity_eligible(cfs_rq, se)) {
900 node = node->rb_left;
905 * Now we heap search eligible trees for the best (min_)deadline
907 if (!best || deadline_gt(deadline, best, se))
911 * Every se in a left branch is eligible, keep track of the
912 * branch with the best min_deadline
915 struct sched_entity *left = __node_2_se(node->rb_left);
917 if (!best_left || deadline_gt(min_deadline, best_left, left))
921 * min_deadline is in the left branch. rb_left and all
922 * descendants are eligible, so immediately switch to the second
925 if (left->min_deadline == se->min_deadline)
929 /* min_deadline is at this node, no need to look right */
930 if (se->deadline == se->min_deadline)
933 /* else min_deadline is in the right branch. */
934 node = node->rb_right;
938 * We ran into an eligible node which is itself the best.
939 * (Or nr_running == 0 and both are NULL)
941 if (!best_left || (s64)(best_left->min_deadline - best->deadline) > 0)
945 * Now best_left and all of its children are eligible, and we are just
946 * looking for deadline == min_deadline
948 node = &best_left->run_node;
950 struct sched_entity *se = __node_2_se(node);
952 /* min_deadline is the current node */
953 if (se->deadline == se->min_deadline)
956 /* min_deadline is in the left branch */
958 __node_2_se(node->rb_left)->min_deadline == se->min_deadline) {
959 node = node->rb_left;
963 /* else min_deadline is in the right branch */
964 node = node->rb_right;
969 static struct sched_entity *pick_eevdf(struct cfs_rq *cfs_rq)
971 struct sched_entity *se = __pick_eevdf(cfs_rq);
974 struct sched_entity *left = __pick_first_entity(cfs_rq);
976 pr_err("EEVDF scheduling fail, picking leftmost\n");
984 #ifdef CONFIG_SCHED_DEBUG
985 struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
987 struct rb_node *last = rb_last(&cfs_rq->tasks_timeline.rb_root);
992 return __node_2_se(last);
995 /**************************************************************
996 * Scheduling class statistics methods:
999 int sched_update_scaling(void)
1001 unsigned int factor = get_update_sysctl_factor();
1003 #define WRT_SYSCTL(name) \
1004 (normalized_sysctl_##name = sysctl_##name / (factor))
1005 WRT_SYSCTL(sched_base_slice);
1013 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se);
1016 * XXX: strictly: vd_i += N*r_i/w_i such that: vd_i > ve_i
1017 * this is probably good enough.
1019 static void update_deadline(struct cfs_rq *cfs_rq, struct sched_entity *se)
1021 if ((s64)(se->vruntime - se->deadline) < 0)
1025 * For EEVDF the virtual time slope is determined by w_i (iow.
1026 * nice) while the request time r_i is determined by
1027 * sysctl_sched_base_slice.
1029 se->slice = sysctl_sched_base_slice;
1032 * EEVDF: vd_i = ve_i + r_i / w_i
1034 se->deadline = se->vruntime + calc_delta_fair(se->slice, se);
1037 * The task has consumed its request, reschedule.
1039 if (cfs_rq->nr_running > 1) {
1040 resched_curr(rq_of(cfs_rq));
1041 clear_buddies(cfs_rq, se);
1048 static int select_idle_sibling(struct task_struct *p, int prev_cpu, int cpu);
1049 static unsigned long task_h_load(struct task_struct *p);
1050 static unsigned long capacity_of(int cpu);
1052 /* Give new sched_entity start runnable values to heavy its load in infant time */
1053 void init_entity_runnable_average(struct sched_entity *se)
1055 struct sched_avg *sa = &se->avg;
1057 memset(sa, 0, sizeof(*sa));
1060 * Tasks are initialized with full load to be seen as heavy tasks until
1061 * they get a chance to stabilize to their real load level.
1062 * Group entities are initialized with zero load to reflect the fact that
1063 * nothing has been attached to the task group yet.
1065 if (entity_is_task(se))
1066 sa->load_avg = scale_load_down(se->load.weight);
1068 /* when this task enqueue'ed, it will contribute to its cfs_rq's load_avg */
1072 * With new tasks being created, their initial util_avgs are extrapolated
1073 * based on the cfs_rq's current util_avg:
1075 * util_avg = cfs_rq->util_avg / (cfs_rq->load_avg + 1) * se.load.weight
1077 * However, in many cases, the above util_avg does not give a desired
1078 * value. Moreover, the sum of the util_avgs may be divergent, such
1079 * as when the series is a harmonic series.
1081 * To solve this problem, we also cap the util_avg of successive tasks to
1082 * only 1/2 of the left utilization budget:
1084 * util_avg_cap = (cpu_scale - cfs_rq->avg.util_avg) / 2^n
1086 * where n denotes the nth task and cpu_scale the CPU capacity.
1088 * For example, for a CPU with 1024 of capacity, a simplest series from
1089 * the beginning would be like:
1091 * task util_avg: 512, 256, 128, 64, 32, 16, 8, ...
1092 * cfs_rq util_avg: 512, 768, 896, 960, 992, 1008, 1016, ...
1094 * Finally, that extrapolated util_avg is clamped to the cap (util_avg_cap)
1095 * if util_avg > util_avg_cap.
1097 void post_init_entity_util_avg(struct task_struct *p)
1099 struct sched_entity *se = &p->se;
1100 struct cfs_rq *cfs_rq = cfs_rq_of(se);
1101 struct sched_avg *sa = &se->avg;
1102 long cpu_scale = arch_scale_cpu_capacity(cpu_of(rq_of(cfs_rq)));
1103 long cap = (long)(cpu_scale - cfs_rq->avg.util_avg) / 2;
1105 if (p->sched_class != &fair_sched_class) {
1107 * For !fair tasks do:
1109 update_cfs_rq_load_avg(now, cfs_rq);
1110 attach_entity_load_avg(cfs_rq, se);
1111 switched_from_fair(rq, p);
1113 * such that the next switched_to_fair() has the
1116 se->avg.last_update_time = cfs_rq_clock_pelt(cfs_rq);
1121 if (cfs_rq->avg.util_avg != 0) {
1122 sa->util_avg = cfs_rq->avg.util_avg * se->load.weight;
1123 sa->util_avg /= (cfs_rq->avg.load_avg + 1);
1125 if (sa->util_avg > cap)
1132 sa->runnable_avg = sa->util_avg;
1135 #else /* !CONFIG_SMP */
1136 void init_entity_runnable_average(struct sched_entity *se)
1139 void post_init_entity_util_avg(struct task_struct *p)
1142 static void update_tg_load_avg(struct cfs_rq *cfs_rq)
1145 #endif /* CONFIG_SMP */
1148 * Update the current task's runtime statistics.
1150 static void update_curr(struct cfs_rq *cfs_rq)
1152 struct sched_entity *curr = cfs_rq->curr;
1153 u64 now = rq_clock_task(rq_of(cfs_rq));
1156 if (unlikely(!curr))
1159 delta_exec = now - curr->exec_start;
1160 if (unlikely((s64)delta_exec <= 0))
1163 curr->exec_start = now;
1165 if (schedstat_enabled()) {
1166 struct sched_statistics *stats;
1168 stats = __schedstats_from_se(curr);
1169 __schedstat_set(stats->exec_max,
1170 max(delta_exec, stats->exec_max));
1173 curr->sum_exec_runtime += delta_exec;
1174 schedstat_add(cfs_rq->exec_clock, delta_exec);
1176 curr->vruntime += calc_delta_fair(delta_exec, curr);
1177 update_deadline(cfs_rq, curr);
1178 update_min_vruntime(cfs_rq);
1180 if (entity_is_task(curr)) {
1181 struct task_struct *curtask = task_of(curr);
1183 trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime);
1184 cgroup_account_cputime(curtask, delta_exec);
1185 account_group_exec_runtime(curtask, delta_exec);
1188 account_cfs_rq_runtime(cfs_rq, delta_exec);
1191 static void update_curr_fair(struct rq *rq)
1193 update_curr(cfs_rq_of(&rq->curr->se));
1197 update_stats_wait_start_fair(struct cfs_rq *cfs_rq, struct sched_entity *se)
1199 struct sched_statistics *stats;
1200 struct task_struct *p = NULL;
1202 if (!schedstat_enabled())
1205 stats = __schedstats_from_se(se);
1207 if (entity_is_task(se))
1210 __update_stats_wait_start(rq_of(cfs_rq), p, stats);
1214 update_stats_wait_end_fair(struct cfs_rq *cfs_rq, struct sched_entity *se)
1216 struct sched_statistics *stats;
1217 struct task_struct *p = NULL;
1219 if (!schedstat_enabled())
1222 stats = __schedstats_from_se(se);
1225 * When the sched_schedstat changes from 0 to 1, some sched se
1226 * maybe already in the runqueue, the se->statistics.wait_start
1227 * will be 0.So it will let the delta wrong. We need to avoid this
1230 if (unlikely(!schedstat_val(stats->wait_start)))
1233 if (entity_is_task(se))
1236 __update_stats_wait_end(rq_of(cfs_rq), p, stats);
1240 update_stats_enqueue_sleeper_fair(struct cfs_rq *cfs_rq, struct sched_entity *se)
1242 struct sched_statistics *stats;
1243 struct task_struct *tsk = NULL;
1245 if (!schedstat_enabled())
1248 stats = __schedstats_from_se(se);
1250 if (entity_is_task(se))
1253 __update_stats_enqueue_sleeper(rq_of(cfs_rq), tsk, stats);
1257 * Task is being enqueued - update stats:
1260 update_stats_enqueue_fair(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
1262 if (!schedstat_enabled())
1266 * Are we enqueueing a waiting task? (for current tasks
1267 * a dequeue/enqueue event is a NOP)
1269 if (se != cfs_rq->curr)
1270 update_stats_wait_start_fair(cfs_rq, se);
1272 if (flags & ENQUEUE_WAKEUP)
1273 update_stats_enqueue_sleeper_fair(cfs_rq, se);
1277 update_stats_dequeue_fair(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
1280 if (!schedstat_enabled())
1284 * Mark the end of the wait period if dequeueing a
1287 if (se != cfs_rq->curr)
1288 update_stats_wait_end_fair(cfs_rq, se);
1290 if ((flags & DEQUEUE_SLEEP) && entity_is_task(se)) {
1291 struct task_struct *tsk = task_of(se);
1294 /* XXX racy against TTWU */
1295 state = READ_ONCE(tsk->__state);
1296 if (state & TASK_INTERRUPTIBLE)
1297 __schedstat_set(tsk->stats.sleep_start,
1298 rq_clock(rq_of(cfs_rq)));
1299 if (state & TASK_UNINTERRUPTIBLE)
1300 __schedstat_set(tsk->stats.block_start,
1301 rq_clock(rq_of(cfs_rq)));
1306 * We are picking a new current task - update its stats:
1309 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
1312 * We are starting a new run period:
1314 se->exec_start = rq_clock_task(rq_of(cfs_rq));
1317 /**************************************************
1318 * Scheduling class queueing methods:
1321 static inline bool is_core_idle(int cpu)
1323 #ifdef CONFIG_SCHED_SMT
1326 for_each_cpu(sibling, cpu_smt_mask(cpu)) {
1330 if (!idle_cpu(sibling))
1339 #define NUMA_IMBALANCE_MIN 2
1342 adjust_numa_imbalance(int imbalance, int dst_running, int imb_numa_nr)
1345 * Allow a NUMA imbalance if busy CPUs is less than the maximum
1346 * threshold. Above this threshold, individual tasks may be contending
1347 * for both memory bandwidth and any shared HT resources. This is an
1348 * approximation as the number of running tasks may not be related to
1349 * the number of busy CPUs due to sched_setaffinity.
1351 if (dst_running > imb_numa_nr)
1355 * Allow a small imbalance based on a simple pair of communicating
1356 * tasks that remain local when the destination is lightly loaded.
1358 if (imbalance <= NUMA_IMBALANCE_MIN)
1363 #endif /* CONFIG_NUMA */
1365 #ifdef CONFIG_NUMA_BALANCING
1367 * Approximate time to scan a full NUMA task in ms. The task scan period is
1368 * calculated based on the tasks virtual memory size and
1369 * numa_balancing_scan_size.
1371 unsigned int sysctl_numa_balancing_scan_period_min = 1000;
1372 unsigned int sysctl_numa_balancing_scan_period_max = 60000;
1374 /* Portion of address space to scan in MB */
1375 unsigned int sysctl_numa_balancing_scan_size = 256;
1377 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
1378 unsigned int sysctl_numa_balancing_scan_delay = 1000;
1380 /* The page with hint page fault latency < threshold in ms is considered hot */
1381 unsigned int sysctl_numa_balancing_hot_threshold = MSEC_PER_SEC;
1384 refcount_t refcount;
1386 spinlock_t lock; /* nr_tasks, tasks */
1391 struct rcu_head rcu;
1392 unsigned long total_faults;
1393 unsigned long max_faults_cpu;
1395 * faults[] array is split into two regions: faults_mem and faults_cpu.
1397 * Faults_cpu is used to decide whether memory should move
1398 * towards the CPU. As a consequence, these stats are weighted
1399 * more by CPU use than by memory faults.
1401 unsigned long faults[];
1405 * For functions that can be called in multiple contexts that permit reading
1406 * ->numa_group (see struct task_struct for locking rules).
1408 static struct numa_group *deref_task_numa_group(struct task_struct *p)
1410 return rcu_dereference_check(p->numa_group, p == current ||
1411 (lockdep_is_held(__rq_lockp(task_rq(p))) && !READ_ONCE(p->on_cpu)));
1414 static struct numa_group *deref_curr_numa_group(struct task_struct *p)
1416 return rcu_dereference_protected(p->numa_group, p == current);
1419 static inline unsigned long group_faults_priv(struct numa_group *ng);
1420 static inline unsigned long group_faults_shared(struct numa_group *ng);
1422 static unsigned int task_nr_scan_windows(struct task_struct *p)
1424 unsigned long rss = 0;
1425 unsigned long nr_scan_pages;
1428 * Calculations based on RSS as non-present and empty pages are skipped
1429 * by the PTE scanner and NUMA hinting faults should be trapped based
1432 nr_scan_pages = sysctl_numa_balancing_scan_size << (20 - PAGE_SHIFT);
1433 rss = get_mm_rss(p->mm);
1435 rss = nr_scan_pages;
1437 rss = round_up(rss, nr_scan_pages);
1438 return rss / nr_scan_pages;
1441 /* For sanity's sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */
1442 #define MAX_SCAN_WINDOW 2560
1444 static unsigned int task_scan_min(struct task_struct *p)
1446 unsigned int scan_size = READ_ONCE(sysctl_numa_balancing_scan_size);
1447 unsigned int scan, floor;
1448 unsigned int windows = 1;
1450 if (scan_size < MAX_SCAN_WINDOW)
1451 windows = MAX_SCAN_WINDOW / scan_size;
1452 floor = 1000 / windows;
1454 scan = sysctl_numa_balancing_scan_period_min / task_nr_scan_windows(p);
1455 return max_t(unsigned int, floor, scan);
1458 static unsigned int task_scan_start(struct task_struct *p)
1460 unsigned long smin = task_scan_min(p);
1461 unsigned long period = smin;
1462 struct numa_group *ng;
1464 /* Scale the maximum scan period with the amount of shared memory. */
1466 ng = rcu_dereference(p->numa_group);
1468 unsigned long shared = group_faults_shared(ng);
1469 unsigned long private = group_faults_priv(ng);
1471 period *= refcount_read(&ng->refcount);
1472 period *= shared + 1;
1473 period /= private + shared + 1;
1477 return max(smin, period);
1480 static unsigned int task_scan_max(struct task_struct *p)
1482 unsigned long smin = task_scan_min(p);
1484 struct numa_group *ng;
1486 /* Watch for min being lower than max due to floor calculations */
1487 smax = sysctl_numa_balancing_scan_period_max / task_nr_scan_windows(p);
1489 /* Scale the maximum scan period with the amount of shared memory. */
1490 ng = deref_curr_numa_group(p);
1492 unsigned long shared = group_faults_shared(ng);
1493 unsigned long private = group_faults_priv(ng);
1494 unsigned long period = smax;
1496 period *= refcount_read(&ng->refcount);
1497 period *= shared + 1;
1498 period /= private + shared + 1;
1500 smax = max(smax, period);
1503 return max(smin, smax);
1506 static void account_numa_enqueue(struct rq *rq, struct task_struct *p)
1508 rq->nr_numa_running += (p->numa_preferred_nid != NUMA_NO_NODE);
1509 rq->nr_preferred_running += (p->numa_preferred_nid == task_node(p));
1512 static void account_numa_dequeue(struct rq *rq, struct task_struct *p)
1514 rq->nr_numa_running -= (p->numa_preferred_nid != NUMA_NO_NODE);
1515 rq->nr_preferred_running -= (p->numa_preferred_nid == task_node(p));
1518 /* Shared or private faults. */
1519 #define NR_NUMA_HINT_FAULT_TYPES 2
1521 /* Memory and CPU locality */
1522 #define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2)
1524 /* Averaged statistics, and temporary buffers. */
1525 #define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2)
1527 pid_t task_numa_group_id(struct task_struct *p)
1529 struct numa_group *ng;
1533 ng = rcu_dereference(p->numa_group);
1542 * The averaged statistics, shared & private, memory & CPU,
1543 * occupy the first half of the array. The second half of the
1544 * array is for current counters, which are averaged into the
1545 * first set by task_numa_placement.
1547 static inline int task_faults_idx(enum numa_faults_stats s, int nid, int priv)
1549 return NR_NUMA_HINT_FAULT_TYPES * (s * nr_node_ids + nid) + priv;
1552 static inline unsigned long task_faults(struct task_struct *p, int nid)
1554 if (!p->numa_faults)
1557 return p->numa_faults[task_faults_idx(NUMA_MEM, nid, 0)] +
1558 p->numa_faults[task_faults_idx(NUMA_MEM, nid, 1)];
1561 static inline unsigned long group_faults(struct task_struct *p, int nid)
1563 struct numa_group *ng = deref_task_numa_group(p);
1568 return ng->faults[task_faults_idx(NUMA_MEM, nid, 0)] +
1569 ng->faults[task_faults_idx(NUMA_MEM, nid, 1)];
1572 static inline unsigned long group_faults_cpu(struct numa_group *group, int nid)
1574 return group->faults[task_faults_idx(NUMA_CPU, nid, 0)] +
1575 group->faults[task_faults_idx(NUMA_CPU, nid, 1)];
1578 static inline unsigned long group_faults_priv(struct numa_group *ng)
1580 unsigned long faults = 0;
1583 for_each_online_node(node) {
1584 faults += ng->faults[task_faults_idx(NUMA_MEM, node, 1)];
1590 static inline unsigned long group_faults_shared(struct numa_group *ng)
1592 unsigned long faults = 0;
1595 for_each_online_node(node) {
1596 faults += ng->faults[task_faults_idx(NUMA_MEM, node, 0)];
1603 * A node triggering more than 1/3 as many NUMA faults as the maximum is
1604 * considered part of a numa group's pseudo-interleaving set. Migrations
1605 * between these nodes are slowed down, to allow things to settle down.
1607 #define ACTIVE_NODE_FRACTION 3
1609 static bool numa_is_active_node(int nid, struct numa_group *ng)
1611 return group_faults_cpu(ng, nid) * ACTIVE_NODE_FRACTION > ng->max_faults_cpu;
1614 /* Handle placement on systems where not all nodes are directly connected. */
1615 static unsigned long score_nearby_nodes(struct task_struct *p, int nid,
1616 int lim_dist, bool task)
1618 unsigned long score = 0;
1622 * All nodes are directly connected, and the same distance
1623 * from each other. No need for fancy placement algorithms.
1625 if (sched_numa_topology_type == NUMA_DIRECT)
1628 /* sched_max_numa_distance may be changed in parallel. */
1629 max_dist = READ_ONCE(sched_max_numa_distance);
1631 * This code is called for each node, introducing N^2 complexity,
1632 * which should be ok given the number of nodes rarely exceeds 8.
1634 for_each_online_node(node) {
1635 unsigned long faults;
1636 int dist = node_distance(nid, node);
1639 * The furthest away nodes in the system are not interesting
1640 * for placement; nid was already counted.
1642 if (dist >= max_dist || node == nid)
1646 * On systems with a backplane NUMA topology, compare groups
1647 * of nodes, and move tasks towards the group with the most
1648 * memory accesses. When comparing two nodes at distance
1649 * "hoplimit", only nodes closer by than "hoplimit" are part
1650 * of each group. Skip other nodes.
1652 if (sched_numa_topology_type == NUMA_BACKPLANE && dist >= lim_dist)
1655 /* Add up the faults from nearby nodes. */
1657 faults = task_faults(p, node);
1659 faults = group_faults(p, node);
1662 * On systems with a glueless mesh NUMA topology, there are
1663 * no fixed "groups of nodes". Instead, nodes that are not
1664 * directly connected bounce traffic through intermediate
1665 * nodes; a numa_group can occupy any set of nodes.
1666 * The further away a node is, the less the faults count.
1667 * This seems to result in good task placement.
1669 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
1670 faults *= (max_dist - dist);
1671 faults /= (max_dist - LOCAL_DISTANCE);
1681 * These return the fraction of accesses done by a particular task, or
1682 * task group, on a particular numa node. The group weight is given a
1683 * larger multiplier, in order to group tasks together that are almost
1684 * evenly spread out between numa nodes.
1686 static inline unsigned long task_weight(struct task_struct *p, int nid,
1689 unsigned long faults, total_faults;
1691 if (!p->numa_faults)
1694 total_faults = p->total_numa_faults;
1699 faults = task_faults(p, nid);
1700 faults += score_nearby_nodes(p, nid, dist, true);
1702 return 1000 * faults / total_faults;
1705 static inline unsigned long group_weight(struct task_struct *p, int nid,
1708 struct numa_group *ng = deref_task_numa_group(p);
1709 unsigned long faults, total_faults;
1714 total_faults = ng->total_faults;
1719 faults = group_faults(p, nid);
1720 faults += score_nearby_nodes(p, nid, dist, false);
1722 return 1000 * faults / total_faults;
1726 * If memory tiering mode is enabled, cpupid of slow memory page is
1727 * used to record scan time instead of CPU and PID. When tiering mode
1728 * is disabled at run time, the scan time (in cpupid) will be
1729 * interpreted as CPU and PID. So CPU needs to be checked to avoid to
1730 * access out of array bound.
1732 static inline bool cpupid_valid(int cpupid)
1734 return cpupid_to_cpu(cpupid) < nr_cpu_ids;
1738 * For memory tiering mode, if there are enough free pages (more than
1739 * enough watermark defined here) in fast memory node, to take full
1740 * advantage of fast memory capacity, all recently accessed slow
1741 * memory pages will be migrated to fast memory node without
1742 * considering hot threshold.
1744 static bool pgdat_free_space_enough(struct pglist_data *pgdat)
1747 unsigned long enough_wmark;
1749 enough_wmark = max(1UL * 1024 * 1024 * 1024 >> PAGE_SHIFT,
1750 pgdat->node_present_pages >> 4);
1751 for (z = pgdat->nr_zones - 1; z >= 0; z--) {
1752 struct zone *zone = pgdat->node_zones + z;
1754 if (!populated_zone(zone))
1757 if (zone_watermark_ok(zone, 0,
1758 wmark_pages(zone, WMARK_PROMO) + enough_wmark,
1766 * For memory tiering mode, when page tables are scanned, the scan
1767 * time will be recorded in struct page in addition to make page
1768 * PROT_NONE for slow memory page. So when the page is accessed, in
1769 * hint page fault handler, the hint page fault latency is calculated
1772 * hint page fault latency = hint page fault time - scan time
1774 * The smaller the hint page fault latency, the higher the possibility
1775 * for the page to be hot.
1777 static int numa_hint_fault_latency(struct page *page)
1779 int last_time, time;
1781 time = jiffies_to_msecs(jiffies);
1782 last_time = xchg_page_access_time(page, time);
1784 return (time - last_time) & PAGE_ACCESS_TIME_MASK;
1788 * For memory tiering mode, too high promotion/demotion throughput may
1789 * hurt application latency. So we provide a mechanism to rate limit
1790 * the number of pages that are tried to be promoted.
1792 static bool numa_promotion_rate_limit(struct pglist_data *pgdat,
1793 unsigned long rate_limit, int nr)
1795 unsigned long nr_cand;
1796 unsigned int now, start;
1798 now = jiffies_to_msecs(jiffies);
1799 mod_node_page_state(pgdat, PGPROMOTE_CANDIDATE, nr);
1800 nr_cand = node_page_state(pgdat, PGPROMOTE_CANDIDATE);
1801 start = pgdat->nbp_rl_start;
1802 if (now - start > MSEC_PER_SEC &&
1803 cmpxchg(&pgdat->nbp_rl_start, start, now) == start)
1804 pgdat->nbp_rl_nr_cand = nr_cand;
1805 if (nr_cand - pgdat->nbp_rl_nr_cand >= rate_limit)
1810 #define NUMA_MIGRATION_ADJUST_STEPS 16
1812 static void numa_promotion_adjust_threshold(struct pglist_data *pgdat,
1813 unsigned long rate_limit,
1814 unsigned int ref_th)
1816 unsigned int now, start, th_period, unit_th, th;
1817 unsigned long nr_cand, ref_cand, diff_cand;
1819 now = jiffies_to_msecs(jiffies);
1820 th_period = sysctl_numa_balancing_scan_period_max;
1821 start = pgdat->nbp_th_start;
1822 if (now - start > th_period &&
1823 cmpxchg(&pgdat->nbp_th_start, start, now) == start) {
1824 ref_cand = rate_limit *
1825 sysctl_numa_balancing_scan_period_max / MSEC_PER_SEC;
1826 nr_cand = node_page_state(pgdat, PGPROMOTE_CANDIDATE);
1827 diff_cand = nr_cand - pgdat->nbp_th_nr_cand;
1828 unit_th = ref_th * 2 / NUMA_MIGRATION_ADJUST_STEPS;
1829 th = pgdat->nbp_threshold ? : ref_th;
1830 if (diff_cand > ref_cand * 11 / 10)
1831 th = max(th - unit_th, unit_th);
1832 else if (diff_cand < ref_cand * 9 / 10)
1833 th = min(th + unit_th, ref_th * 2);
1834 pgdat->nbp_th_nr_cand = nr_cand;
1835 pgdat->nbp_threshold = th;
1839 bool should_numa_migrate_memory(struct task_struct *p, struct page * page,
1840 int src_nid, int dst_cpu)
1842 struct numa_group *ng = deref_curr_numa_group(p);
1843 int dst_nid = cpu_to_node(dst_cpu);
1844 int last_cpupid, this_cpupid;
1847 * The pages in slow memory node should be migrated according
1848 * to hot/cold instead of private/shared.
1850 if (sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING &&
1851 !node_is_toptier(src_nid)) {
1852 struct pglist_data *pgdat;
1853 unsigned long rate_limit;
1854 unsigned int latency, th, def_th;
1856 pgdat = NODE_DATA(dst_nid);
1857 if (pgdat_free_space_enough(pgdat)) {
1858 /* workload changed, reset hot threshold */
1859 pgdat->nbp_threshold = 0;
1863 def_th = sysctl_numa_balancing_hot_threshold;
1864 rate_limit = sysctl_numa_balancing_promote_rate_limit << \
1866 numa_promotion_adjust_threshold(pgdat, rate_limit, def_th);
1868 th = pgdat->nbp_threshold ? : def_th;
1869 latency = numa_hint_fault_latency(page);
1873 return !numa_promotion_rate_limit(pgdat, rate_limit,
1874 thp_nr_pages(page));
1877 this_cpupid = cpu_pid_to_cpupid(dst_cpu, current->pid);
1878 last_cpupid = page_cpupid_xchg_last(page, this_cpupid);
1880 if (!(sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING) &&
1881 !node_is_toptier(src_nid) && !cpupid_valid(last_cpupid))
1885 * Allow first faults or private faults to migrate immediately early in
1886 * the lifetime of a task. The magic number 4 is based on waiting for
1887 * two full passes of the "multi-stage node selection" test that is
1890 if ((p->numa_preferred_nid == NUMA_NO_NODE || p->numa_scan_seq <= 4) &&
1891 (cpupid_pid_unset(last_cpupid) || cpupid_match_pid(p, last_cpupid)))
1895 * Multi-stage node selection is used in conjunction with a periodic
1896 * migration fault to build a temporal task<->page relation. By using
1897 * a two-stage filter we remove short/unlikely relations.
1899 * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate
1900 * a task's usage of a particular page (n_p) per total usage of this
1901 * page (n_t) (in a given time-span) to a probability.
1903 * Our periodic faults will sample this probability and getting the
1904 * same result twice in a row, given these samples are fully
1905 * independent, is then given by P(n)^2, provided our sample period
1906 * is sufficiently short compared to the usage pattern.
1908 * This quadric squishes small probabilities, making it less likely we
1909 * act on an unlikely task<->page relation.
1911 if (!cpupid_pid_unset(last_cpupid) &&
1912 cpupid_to_nid(last_cpupid) != dst_nid)
1915 /* Always allow migrate on private faults */
1916 if (cpupid_match_pid(p, last_cpupid))
1919 /* A shared fault, but p->numa_group has not been set up yet. */
1924 * Destination node is much more heavily used than the source
1925 * node? Allow migration.
1927 if (group_faults_cpu(ng, dst_nid) > group_faults_cpu(ng, src_nid) *
1928 ACTIVE_NODE_FRACTION)
1932 * Distribute memory according to CPU & memory use on each node,
1933 * with 3/4 hysteresis to avoid unnecessary memory migrations:
1935 * faults_cpu(dst) 3 faults_cpu(src)
1936 * --------------- * - > ---------------
1937 * faults_mem(dst) 4 faults_mem(src)
1939 return group_faults_cpu(ng, dst_nid) * group_faults(p, src_nid) * 3 >
1940 group_faults_cpu(ng, src_nid) * group_faults(p, dst_nid) * 4;
1944 * 'numa_type' describes the node at the moment of load balancing.
1947 /* The node has spare capacity that can be used to run more tasks. */
1950 * The node is fully used and the tasks don't compete for more CPU
1951 * cycles. Nevertheless, some tasks might wait before running.
1955 * The node is overloaded and can't provide expected CPU cycles to all
1961 /* Cached statistics for all CPUs within a node */
1964 unsigned long runnable;
1966 /* Total compute capacity of CPUs on a node */
1967 unsigned long compute_capacity;
1968 unsigned int nr_running;
1969 unsigned int weight;
1970 enum numa_type node_type;
1974 struct task_numa_env {
1975 struct task_struct *p;
1977 int src_cpu, src_nid;
1978 int dst_cpu, dst_nid;
1981 struct numa_stats src_stats, dst_stats;
1986 struct task_struct *best_task;
1991 static unsigned long cpu_load(struct rq *rq);
1992 static unsigned long cpu_runnable(struct rq *rq);
1995 numa_type numa_classify(unsigned int imbalance_pct,
1996 struct numa_stats *ns)
1998 if ((ns->nr_running > ns->weight) &&
1999 (((ns->compute_capacity * 100) < (ns->util * imbalance_pct)) ||
2000 ((ns->compute_capacity * imbalance_pct) < (ns->runnable * 100))))
2001 return node_overloaded;
2003 if ((ns->nr_running < ns->weight) ||
2004 (((ns->compute_capacity * 100) > (ns->util * imbalance_pct)) &&
2005 ((ns->compute_capacity * imbalance_pct) > (ns->runnable * 100))))
2006 return node_has_spare;
2008 return node_fully_busy;
2011 #ifdef CONFIG_SCHED_SMT
2012 /* Forward declarations of select_idle_sibling helpers */
2013 static inline bool test_idle_cores(int cpu);
2014 static inline int numa_idle_core(int idle_core, int cpu)
2016 if (!static_branch_likely(&sched_smt_present) ||
2017 idle_core >= 0 || !test_idle_cores(cpu))
2021 * Prefer cores instead of packing HT siblings
2022 * and triggering future load balancing.
2024 if (is_core_idle(cpu))
2030 static inline int numa_idle_core(int idle_core, int cpu)
2037 * Gather all necessary information to make NUMA balancing placement
2038 * decisions that are compatible with standard load balancer. This
2039 * borrows code and logic from update_sg_lb_stats but sharing a
2040 * common implementation is impractical.
2042 static void update_numa_stats(struct task_numa_env *env,
2043 struct numa_stats *ns, int nid,
2046 int cpu, idle_core = -1;
2048 memset(ns, 0, sizeof(*ns));
2052 for_each_cpu(cpu, cpumask_of_node(nid)) {
2053 struct rq *rq = cpu_rq(cpu);
2055 ns->load += cpu_load(rq);
2056 ns->runnable += cpu_runnable(rq);
2057 ns->util += cpu_util_cfs(cpu);
2058 ns->nr_running += rq->cfs.h_nr_running;
2059 ns->compute_capacity += capacity_of(cpu);
2061 if (find_idle && idle_core < 0 && !rq->nr_running && idle_cpu(cpu)) {
2062 if (READ_ONCE(rq->numa_migrate_on) ||
2063 !cpumask_test_cpu(cpu, env->p->cpus_ptr))
2066 if (ns->idle_cpu == -1)
2069 idle_core = numa_idle_core(idle_core, cpu);
2074 ns->weight = cpumask_weight(cpumask_of_node(nid));
2076 ns->node_type = numa_classify(env->imbalance_pct, ns);
2079 ns->idle_cpu = idle_core;
2082 static void task_numa_assign(struct task_numa_env *env,
2083 struct task_struct *p, long imp)
2085 struct rq *rq = cpu_rq(env->dst_cpu);
2087 /* Check if run-queue part of active NUMA balance. */
2088 if (env->best_cpu != env->dst_cpu && xchg(&rq->numa_migrate_on, 1)) {
2090 int start = env->dst_cpu;
2092 /* Find alternative idle CPU. */
2093 for_each_cpu_wrap(cpu, cpumask_of_node(env->dst_nid), start + 1) {
2094 if (cpu == env->best_cpu || !idle_cpu(cpu) ||
2095 !cpumask_test_cpu(cpu, env->p->cpus_ptr)) {
2100 rq = cpu_rq(env->dst_cpu);
2101 if (!xchg(&rq->numa_migrate_on, 1))
2105 /* Failed to find an alternative idle CPU */
2111 * Clear previous best_cpu/rq numa-migrate flag, since task now
2112 * found a better CPU to move/swap.
2114 if (env->best_cpu != -1 && env->best_cpu != env->dst_cpu) {
2115 rq = cpu_rq(env->best_cpu);
2116 WRITE_ONCE(rq->numa_migrate_on, 0);
2120 put_task_struct(env->best_task);
2125 env->best_imp = imp;
2126 env->best_cpu = env->dst_cpu;
2129 static bool load_too_imbalanced(long src_load, long dst_load,
2130 struct task_numa_env *env)
2133 long orig_src_load, orig_dst_load;
2134 long src_capacity, dst_capacity;
2137 * The load is corrected for the CPU capacity available on each node.
2140 * ------------ vs ---------
2141 * src_capacity dst_capacity
2143 src_capacity = env->src_stats.compute_capacity;
2144 dst_capacity = env->dst_stats.compute_capacity;
2146 imb = abs(dst_load * src_capacity - src_load * dst_capacity);
2148 orig_src_load = env->src_stats.load;
2149 orig_dst_load = env->dst_stats.load;
2151 old_imb = abs(orig_dst_load * src_capacity - orig_src_load * dst_capacity);
2153 /* Would this change make things worse? */
2154 return (imb > old_imb);
2158 * Maximum NUMA importance can be 1998 (2*999);
2159 * SMALLIMP @ 30 would be close to 1998/64.
2160 * Used to deter task migration.
2165 * This checks if the overall compute and NUMA accesses of the system would
2166 * be improved if the source tasks was migrated to the target dst_cpu taking
2167 * into account that it might be best if task running on the dst_cpu should
2168 * be exchanged with the source task
2170 static bool task_numa_compare(struct task_numa_env *env,
2171 long taskimp, long groupimp, bool maymove)
2173 struct numa_group *cur_ng, *p_ng = deref_curr_numa_group(env->p);
2174 struct rq *dst_rq = cpu_rq(env->dst_cpu);
2175 long imp = p_ng ? groupimp : taskimp;
2176 struct task_struct *cur;
2177 long src_load, dst_load;
2178 int dist = env->dist;
2181 bool stopsearch = false;
2183 if (READ_ONCE(dst_rq->numa_migrate_on))
2187 cur = rcu_dereference(dst_rq->curr);
2188 if (cur && ((cur->flags & PF_EXITING) || is_idle_task(cur)))
2192 * Because we have preemption enabled we can get migrated around and
2193 * end try selecting ourselves (current == env->p) as a swap candidate.
2195 if (cur == env->p) {
2201 if (maymove && moveimp >= env->best_imp)
2207 /* Skip this swap candidate if cannot move to the source cpu. */
2208 if (!cpumask_test_cpu(env->src_cpu, cur->cpus_ptr))
2212 * Skip this swap candidate if it is not moving to its preferred
2213 * node and the best task is.
2215 if (env->best_task &&
2216 env->best_task->numa_preferred_nid == env->src_nid &&
2217 cur->numa_preferred_nid != env->src_nid) {
2222 * "imp" is the fault differential for the source task between the
2223 * source and destination node. Calculate the total differential for
2224 * the source task and potential destination task. The more negative
2225 * the value is, the more remote accesses that would be expected to
2226 * be incurred if the tasks were swapped.
2228 * If dst and source tasks are in the same NUMA group, or not
2229 * in any group then look only at task weights.
2231 cur_ng = rcu_dereference(cur->numa_group);
2232 if (cur_ng == p_ng) {
2234 * Do not swap within a group or between tasks that have
2235 * no group if there is spare capacity. Swapping does
2236 * not address the load imbalance and helps one task at
2237 * the cost of punishing another.
2239 if (env->dst_stats.node_type == node_has_spare)
2242 imp = taskimp + task_weight(cur, env->src_nid, dist) -
2243 task_weight(cur, env->dst_nid, dist);
2245 * Add some hysteresis to prevent swapping the
2246 * tasks within a group over tiny differences.
2252 * Compare the group weights. If a task is all by itself
2253 * (not part of a group), use the task weight instead.
2256 imp += group_weight(cur, env->src_nid, dist) -
2257 group_weight(cur, env->dst_nid, dist);
2259 imp += task_weight(cur, env->src_nid, dist) -
2260 task_weight(cur, env->dst_nid, dist);
2263 /* Discourage picking a task already on its preferred node */
2264 if (cur->numa_preferred_nid == env->dst_nid)
2268 * Encourage picking a task that moves to its preferred node.
2269 * This potentially makes imp larger than it's maximum of
2270 * 1998 (see SMALLIMP and task_weight for why) but in this
2271 * case, it does not matter.
2273 if (cur->numa_preferred_nid == env->src_nid)
2276 if (maymove && moveimp > imp && moveimp > env->best_imp) {
2283 * Prefer swapping with a task moving to its preferred node over a
2286 if (env->best_task && cur->numa_preferred_nid == env->src_nid &&
2287 env->best_task->numa_preferred_nid != env->src_nid) {
2292 * If the NUMA importance is less than SMALLIMP,
2293 * task migration might only result in ping pong
2294 * of tasks and also hurt performance due to cache
2297 if (imp < SMALLIMP || imp <= env->best_imp + SMALLIMP / 2)
2301 * In the overloaded case, try and keep the load balanced.
2303 load = task_h_load(env->p) - task_h_load(cur);
2307 dst_load = env->dst_stats.load + load;
2308 src_load = env->src_stats.load - load;
2310 if (load_too_imbalanced(src_load, dst_load, env))
2314 /* Evaluate an idle CPU for a task numa move. */
2316 int cpu = env->dst_stats.idle_cpu;
2318 /* Nothing cached so current CPU went idle since the search. */
2323 * If the CPU is no longer truly idle and the previous best CPU
2324 * is, keep using it.
2326 if (!idle_cpu(cpu) && env->best_cpu >= 0 &&
2327 idle_cpu(env->best_cpu)) {
2328 cpu = env->best_cpu;
2334 task_numa_assign(env, cur, imp);
2337 * If a move to idle is allowed because there is capacity or load
2338 * balance improves then stop the search. While a better swap
2339 * candidate may exist, a search is not free.
2341 if (maymove && !cur && env->best_cpu >= 0 && idle_cpu(env->best_cpu))
2345 * If a swap candidate must be identified and the current best task
2346 * moves its preferred node then stop the search.
2348 if (!maymove && env->best_task &&
2349 env->best_task->numa_preferred_nid == env->src_nid) {
2358 static void task_numa_find_cpu(struct task_numa_env *env,
2359 long taskimp, long groupimp)
2361 bool maymove = false;
2365 * If dst node has spare capacity, then check if there is an
2366 * imbalance that would be overruled by the load balancer.
2368 if (env->dst_stats.node_type == node_has_spare) {
2369 unsigned int imbalance;
2370 int src_running, dst_running;
2373 * Would movement cause an imbalance? Note that if src has
2374 * more running tasks that the imbalance is ignored as the
2375 * move improves the imbalance from the perspective of the
2376 * CPU load balancer.
2378 src_running = env->src_stats.nr_running - 1;
2379 dst_running = env->dst_stats.nr_running + 1;
2380 imbalance = max(0, dst_running - src_running);
2381 imbalance = adjust_numa_imbalance(imbalance, dst_running,
2384 /* Use idle CPU if there is no imbalance */
2387 if (env->dst_stats.idle_cpu >= 0) {
2388 env->dst_cpu = env->dst_stats.idle_cpu;
2389 task_numa_assign(env, NULL, 0);
2394 long src_load, dst_load, load;
2396 * If the improvement from just moving env->p direction is better
2397 * than swapping tasks around, check if a move is possible.
2399 load = task_h_load(env->p);
2400 dst_load = env->dst_stats.load + load;
2401 src_load = env->src_stats.load - load;
2402 maymove = !load_too_imbalanced(src_load, dst_load, env);
2405 for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) {
2406 /* Skip this CPU if the source task cannot migrate */
2407 if (!cpumask_test_cpu(cpu, env->p->cpus_ptr))
2411 if (task_numa_compare(env, taskimp, groupimp, maymove))
2416 static int task_numa_migrate(struct task_struct *p)
2418 struct task_numa_env env = {
2421 .src_cpu = task_cpu(p),
2422 .src_nid = task_node(p),
2424 .imbalance_pct = 112,
2430 unsigned long taskweight, groupweight;
2431 struct sched_domain *sd;
2432 long taskimp, groupimp;
2433 struct numa_group *ng;
2438 * Pick the lowest SD_NUMA domain, as that would have the smallest
2439 * imbalance and would be the first to start moving tasks about.
2441 * And we want to avoid any moving of tasks about, as that would create
2442 * random movement of tasks -- counter the numa conditions we're trying
2446 sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu));
2448 env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2;
2449 env.imb_numa_nr = sd->imb_numa_nr;
2454 * Cpusets can break the scheduler domain tree into smaller
2455 * balance domains, some of which do not cross NUMA boundaries.
2456 * Tasks that are "trapped" in such domains cannot be migrated
2457 * elsewhere, so there is no point in (re)trying.
2459 if (unlikely(!sd)) {
2460 sched_setnuma(p, task_node(p));
2464 env.dst_nid = p->numa_preferred_nid;
2465 dist = env.dist = node_distance(env.src_nid, env.dst_nid);
2466 taskweight = task_weight(p, env.src_nid, dist);
2467 groupweight = group_weight(p, env.src_nid, dist);
2468 update_numa_stats(&env, &env.src_stats, env.src_nid, false);
2469 taskimp = task_weight(p, env.dst_nid, dist) - taskweight;
2470 groupimp = group_weight(p, env.dst_nid, dist) - groupweight;
2471 update_numa_stats(&env, &env.dst_stats, env.dst_nid, true);
2473 /* Try to find a spot on the preferred nid. */
2474 task_numa_find_cpu(&env, taskimp, groupimp);
2477 * Look at other nodes in these cases:
2478 * - there is no space available on the preferred_nid
2479 * - the task is part of a numa_group that is interleaved across
2480 * multiple NUMA nodes; in order to better consolidate the group,
2481 * we need to check other locations.
2483 ng = deref_curr_numa_group(p);
2484 if (env.best_cpu == -1 || (ng && ng->active_nodes > 1)) {
2485 for_each_node_state(nid, N_CPU) {
2486 if (nid == env.src_nid || nid == p->numa_preferred_nid)
2489 dist = node_distance(env.src_nid, env.dst_nid);
2490 if (sched_numa_topology_type == NUMA_BACKPLANE &&
2492 taskweight = task_weight(p, env.src_nid, dist);
2493 groupweight = group_weight(p, env.src_nid, dist);
2496 /* Only consider nodes where both task and groups benefit */
2497 taskimp = task_weight(p, nid, dist) - taskweight;
2498 groupimp = group_weight(p, nid, dist) - groupweight;
2499 if (taskimp < 0 && groupimp < 0)
2504 update_numa_stats(&env, &env.dst_stats, env.dst_nid, true);
2505 task_numa_find_cpu(&env, taskimp, groupimp);
2510 * If the task is part of a workload that spans multiple NUMA nodes,
2511 * and is migrating into one of the workload's active nodes, remember
2512 * this node as the task's preferred numa node, so the workload can
2514 * A task that migrated to a second choice node will be better off
2515 * trying for a better one later. Do not set the preferred node here.
2518 if (env.best_cpu == -1)
2521 nid = cpu_to_node(env.best_cpu);
2523 if (nid != p->numa_preferred_nid)
2524 sched_setnuma(p, nid);
2527 /* No better CPU than the current one was found. */
2528 if (env.best_cpu == -1) {
2529 trace_sched_stick_numa(p, env.src_cpu, NULL, -1);
2533 best_rq = cpu_rq(env.best_cpu);
2534 if (env.best_task == NULL) {
2535 ret = migrate_task_to(p, env.best_cpu);
2536 WRITE_ONCE(best_rq->numa_migrate_on, 0);
2538 trace_sched_stick_numa(p, env.src_cpu, NULL, env.best_cpu);
2542 ret = migrate_swap(p, env.best_task, env.best_cpu, env.src_cpu);
2543 WRITE_ONCE(best_rq->numa_migrate_on, 0);
2546 trace_sched_stick_numa(p, env.src_cpu, env.best_task, env.best_cpu);
2547 put_task_struct(env.best_task);
2551 /* Attempt to migrate a task to a CPU on the preferred node. */
2552 static void numa_migrate_preferred(struct task_struct *p)
2554 unsigned long interval = HZ;
2556 /* This task has no NUMA fault statistics yet */
2557 if (unlikely(p->numa_preferred_nid == NUMA_NO_NODE || !p->numa_faults))
2560 /* Periodically retry migrating the task to the preferred node */
2561 interval = min(interval, msecs_to_jiffies(p->numa_scan_period) / 16);
2562 p->numa_migrate_retry = jiffies + interval;
2564 /* Success if task is already running on preferred CPU */
2565 if (task_node(p) == p->numa_preferred_nid)
2568 /* Otherwise, try migrate to a CPU on the preferred node */
2569 task_numa_migrate(p);
2573 * Find out how many nodes the workload is actively running on. Do this by
2574 * tracking the nodes from which NUMA hinting faults are triggered. This can
2575 * be different from the set of nodes where the workload's memory is currently
2578 static void numa_group_count_active_nodes(struct numa_group *numa_group)
2580 unsigned long faults, max_faults = 0;
2581 int nid, active_nodes = 0;
2583 for_each_node_state(nid, N_CPU) {
2584 faults = group_faults_cpu(numa_group, nid);
2585 if (faults > max_faults)
2586 max_faults = faults;
2589 for_each_node_state(nid, N_CPU) {
2590 faults = group_faults_cpu(numa_group, nid);
2591 if (faults * ACTIVE_NODE_FRACTION > max_faults)
2595 numa_group->max_faults_cpu = max_faults;
2596 numa_group->active_nodes = active_nodes;
2600 * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
2601 * increments. The more local the fault statistics are, the higher the scan
2602 * period will be for the next scan window. If local/(local+remote) ratio is
2603 * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS)
2604 * the scan period will decrease. Aim for 70% local accesses.
2606 #define NUMA_PERIOD_SLOTS 10
2607 #define NUMA_PERIOD_THRESHOLD 7
2610 * Increase the scan period (slow down scanning) if the majority of
2611 * our memory is already on our local node, or if the majority of
2612 * the page accesses are shared with other processes.
2613 * Otherwise, decrease the scan period.
2615 static void update_task_scan_period(struct task_struct *p,
2616 unsigned long shared, unsigned long private)
2618 unsigned int period_slot;
2619 int lr_ratio, ps_ratio;
2622 unsigned long remote = p->numa_faults_locality[0];
2623 unsigned long local = p->numa_faults_locality[1];
2626 * If there were no record hinting faults then either the task is
2627 * completely idle or all activity is in areas that are not of interest
2628 * to automatic numa balancing. Related to that, if there were failed
2629 * migration then it implies we are migrating too quickly or the local
2630 * node is overloaded. In either case, scan slower
2632 if (local + shared == 0 || p->numa_faults_locality[2]) {
2633 p->numa_scan_period = min(p->numa_scan_period_max,
2634 p->numa_scan_period << 1);
2636 p->mm->numa_next_scan = jiffies +
2637 msecs_to_jiffies(p->numa_scan_period);
2643 * Prepare to scale scan period relative to the current period.
2644 * == NUMA_PERIOD_THRESHOLD scan period stays the same
2645 * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
2646 * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
2648 period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS);
2649 lr_ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote);
2650 ps_ratio = (private * NUMA_PERIOD_SLOTS) / (private + shared);
2652 if (ps_ratio >= NUMA_PERIOD_THRESHOLD) {
2654 * Most memory accesses are local. There is no need to
2655 * do fast NUMA scanning, since memory is already local.
2657 int slot = ps_ratio - NUMA_PERIOD_THRESHOLD;
2660 diff = slot * period_slot;
2661 } else if (lr_ratio >= NUMA_PERIOD_THRESHOLD) {
2663 * Most memory accesses are shared with other tasks.
2664 * There is no point in continuing fast NUMA scanning,
2665 * since other tasks may just move the memory elsewhere.
2667 int slot = lr_ratio - NUMA_PERIOD_THRESHOLD;
2670 diff = slot * period_slot;
2673 * Private memory faults exceed (SLOTS-THRESHOLD)/SLOTS,
2674 * yet they are not on the local NUMA node. Speed up
2675 * NUMA scanning to get the memory moved over.
2677 int ratio = max(lr_ratio, ps_ratio);
2678 diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot;
2681 p->numa_scan_period = clamp(p->numa_scan_period + diff,
2682 task_scan_min(p), task_scan_max(p));
2683 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
2687 * Get the fraction of time the task has been running since the last
2688 * NUMA placement cycle. The scheduler keeps similar statistics, but
2689 * decays those on a 32ms period, which is orders of magnitude off
2690 * from the dozens-of-seconds NUMA balancing period. Use the scheduler
2691 * stats only if the task is so new there are no NUMA statistics yet.
2693 static u64 numa_get_avg_runtime(struct task_struct *p, u64 *period)
2695 u64 runtime, delta, now;
2696 /* Use the start of this time slice to avoid calculations. */
2697 now = p->se.exec_start;
2698 runtime = p->se.sum_exec_runtime;
2700 if (p->last_task_numa_placement) {
2701 delta = runtime - p->last_sum_exec_runtime;
2702 *period = now - p->last_task_numa_placement;
2704 /* Avoid time going backwards, prevent potential divide error: */
2705 if (unlikely((s64)*period < 0))
2708 delta = p->se.avg.load_sum;
2709 *period = LOAD_AVG_MAX;
2712 p->last_sum_exec_runtime = runtime;
2713 p->last_task_numa_placement = now;
2719 * Determine the preferred nid for a task in a numa_group. This needs to
2720 * be done in a way that produces consistent results with group_weight,
2721 * otherwise workloads might not converge.
2723 static int preferred_group_nid(struct task_struct *p, int nid)
2728 /* Direct connections between all NUMA nodes. */
2729 if (sched_numa_topology_type == NUMA_DIRECT)
2733 * On a system with glueless mesh NUMA topology, group_weight
2734 * scores nodes according to the number of NUMA hinting faults on
2735 * both the node itself, and on nearby nodes.
2737 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
2738 unsigned long score, max_score = 0;
2739 int node, max_node = nid;
2741 dist = sched_max_numa_distance;
2743 for_each_node_state(node, N_CPU) {
2744 score = group_weight(p, node, dist);
2745 if (score > max_score) {
2754 * Finding the preferred nid in a system with NUMA backplane
2755 * interconnect topology is more involved. The goal is to locate
2756 * tasks from numa_groups near each other in the system, and
2757 * untangle workloads from different sides of the system. This requires
2758 * searching down the hierarchy of node groups, recursively searching
2759 * inside the highest scoring group of nodes. The nodemask tricks
2760 * keep the complexity of the search down.
2762 nodes = node_states[N_CPU];
2763 for (dist = sched_max_numa_distance; dist > LOCAL_DISTANCE; dist--) {
2764 unsigned long max_faults = 0;
2765 nodemask_t max_group = NODE_MASK_NONE;
2768 /* Are there nodes at this distance from each other? */
2769 if (!find_numa_distance(dist))
2772 for_each_node_mask(a, nodes) {
2773 unsigned long faults = 0;
2774 nodemask_t this_group;
2775 nodes_clear(this_group);
2777 /* Sum group's NUMA faults; includes a==b case. */
2778 for_each_node_mask(b, nodes) {
2779 if (node_distance(a, b) < dist) {
2780 faults += group_faults(p, b);
2781 node_set(b, this_group);
2782 node_clear(b, nodes);
2786 /* Remember the top group. */
2787 if (faults > max_faults) {
2788 max_faults = faults;
2789 max_group = this_group;
2791 * subtle: at the smallest distance there is
2792 * just one node left in each "group", the
2793 * winner is the preferred nid.
2798 /* Next round, evaluate the nodes within max_group. */
2806 static void task_numa_placement(struct task_struct *p)
2808 int seq, nid, max_nid = NUMA_NO_NODE;
2809 unsigned long max_faults = 0;
2810 unsigned long fault_types[2] = { 0, 0 };
2811 unsigned long total_faults;
2812 u64 runtime, period;
2813 spinlock_t *group_lock = NULL;
2814 struct numa_group *ng;
2817 * The p->mm->numa_scan_seq field gets updated without
2818 * exclusive access. Use READ_ONCE() here to ensure
2819 * that the field is read in a single access:
2821 seq = READ_ONCE(p->mm->numa_scan_seq);
2822 if (p->numa_scan_seq == seq)
2824 p->numa_scan_seq = seq;
2825 p->numa_scan_period_max = task_scan_max(p);
2827 total_faults = p->numa_faults_locality[0] +
2828 p->numa_faults_locality[1];
2829 runtime = numa_get_avg_runtime(p, &period);
2831 /* If the task is part of a group prevent parallel updates to group stats */
2832 ng = deref_curr_numa_group(p);
2834 group_lock = &ng->lock;
2835 spin_lock_irq(group_lock);
2838 /* Find the node with the highest number of faults */
2839 for_each_online_node(nid) {
2840 /* Keep track of the offsets in numa_faults array */
2841 int mem_idx, membuf_idx, cpu_idx, cpubuf_idx;
2842 unsigned long faults = 0, group_faults = 0;
2845 for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) {
2846 long diff, f_diff, f_weight;
2848 mem_idx = task_faults_idx(NUMA_MEM, nid, priv);
2849 membuf_idx = task_faults_idx(NUMA_MEMBUF, nid, priv);
2850 cpu_idx = task_faults_idx(NUMA_CPU, nid, priv);
2851 cpubuf_idx = task_faults_idx(NUMA_CPUBUF, nid, priv);
2853 /* Decay existing window, copy faults since last scan */
2854 diff = p->numa_faults[membuf_idx] - p->numa_faults[mem_idx] / 2;
2855 fault_types[priv] += p->numa_faults[membuf_idx];
2856 p->numa_faults[membuf_idx] = 0;
2859 * Normalize the faults_from, so all tasks in a group
2860 * count according to CPU use, instead of by the raw
2861 * number of faults. Tasks with little runtime have
2862 * little over-all impact on throughput, and thus their
2863 * faults are less important.
2865 f_weight = div64_u64(runtime << 16, period + 1);
2866 f_weight = (f_weight * p->numa_faults[cpubuf_idx]) /
2868 f_diff = f_weight - p->numa_faults[cpu_idx] / 2;
2869 p->numa_faults[cpubuf_idx] = 0;
2871 p->numa_faults[mem_idx] += diff;
2872 p->numa_faults[cpu_idx] += f_diff;
2873 faults += p->numa_faults[mem_idx];
2874 p->total_numa_faults += diff;
2877 * safe because we can only change our own group
2879 * mem_idx represents the offset for a given
2880 * nid and priv in a specific region because it
2881 * is at the beginning of the numa_faults array.
2883 ng->faults[mem_idx] += diff;
2884 ng->faults[cpu_idx] += f_diff;
2885 ng->total_faults += diff;
2886 group_faults += ng->faults[mem_idx];
2891 if (faults > max_faults) {
2892 max_faults = faults;
2895 } else if (group_faults > max_faults) {
2896 max_faults = group_faults;
2901 /* Cannot migrate task to CPU-less node */
2902 if (max_nid != NUMA_NO_NODE && !node_state(max_nid, N_CPU)) {
2903 int near_nid = max_nid;
2904 int distance, near_distance = INT_MAX;
2906 for_each_node_state(nid, N_CPU) {
2907 distance = node_distance(max_nid, nid);
2908 if (distance < near_distance) {
2910 near_distance = distance;
2917 numa_group_count_active_nodes(ng);
2918 spin_unlock_irq(group_lock);
2919 max_nid = preferred_group_nid(p, max_nid);
2923 /* Set the new preferred node */
2924 if (max_nid != p->numa_preferred_nid)
2925 sched_setnuma(p, max_nid);
2928 update_task_scan_period(p, fault_types[0], fault_types[1]);
2931 static inline int get_numa_group(struct numa_group *grp)
2933 return refcount_inc_not_zero(&grp->refcount);
2936 static inline void put_numa_group(struct numa_group *grp)
2938 if (refcount_dec_and_test(&grp->refcount))
2939 kfree_rcu(grp, rcu);
2942 static void task_numa_group(struct task_struct *p, int cpupid, int flags,
2945 struct numa_group *grp, *my_grp;
2946 struct task_struct *tsk;
2948 int cpu = cpupid_to_cpu(cpupid);
2951 if (unlikely(!deref_curr_numa_group(p))) {
2952 unsigned int size = sizeof(struct numa_group) +
2953 NR_NUMA_HINT_FAULT_STATS *
2954 nr_node_ids * sizeof(unsigned long);
2956 grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
2960 refcount_set(&grp->refcount, 1);
2961 grp->active_nodes = 1;
2962 grp->max_faults_cpu = 0;
2963 spin_lock_init(&grp->lock);
2966 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2967 grp->faults[i] = p->numa_faults[i];
2969 grp->total_faults = p->total_numa_faults;
2972 rcu_assign_pointer(p->numa_group, grp);
2976 tsk = READ_ONCE(cpu_rq(cpu)->curr);
2978 if (!cpupid_match_pid(tsk, cpupid))
2981 grp = rcu_dereference(tsk->numa_group);
2985 my_grp = deref_curr_numa_group(p);
2990 * Only join the other group if its bigger; if we're the bigger group,
2991 * the other task will join us.
2993 if (my_grp->nr_tasks > grp->nr_tasks)
2997 * Tie-break on the grp address.
2999 if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp)
3002 /* Always join threads in the same process. */
3003 if (tsk->mm == current->mm)
3006 /* Simple filter to avoid false positives due to PID collisions */
3007 if (flags & TNF_SHARED)
3010 /* Update priv based on whether false sharing was detected */
3013 if (join && !get_numa_group(grp))
3021 WARN_ON_ONCE(irqs_disabled());
3022 double_lock_irq(&my_grp->lock, &grp->lock);
3024 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) {
3025 my_grp->faults[i] -= p->numa_faults[i];
3026 grp->faults[i] += p->numa_faults[i];
3028 my_grp->total_faults -= p->total_numa_faults;
3029 grp->total_faults += p->total_numa_faults;
3034 spin_unlock(&my_grp->lock);
3035 spin_unlock_irq(&grp->lock);
3037 rcu_assign_pointer(p->numa_group, grp);
3039 put_numa_group(my_grp);
3048 * Get rid of NUMA statistics associated with a task (either current or dead).
3049 * If @final is set, the task is dead and has reached refcount zero, so we can
3050 * safely free all relevant data structures. Otherwise, there might be
3051 * concurrent reads from places like load balancing and procfs, and we should
3052 * reset the data back to default state without freeing ->numa_faults.
3054 void task_numa_free(struct task_struct *p, bool final)
3056 /* safe: p either is current or is being freed by current */
3057 struct numa_group *grp = rcu_dereference_raw(p->numa_group);
3058 unsigned long *numa_faults = p->numa_faults;
3059 unsigned long flags;
3066 spin_lock_irqsave(&grp->lock, flags);
3067 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
3068 grp->faults[i] -= p->numa_faults[i];
3069 grp->total_faults -= p->total_numa_faults;
3072 spin_unlock_irqrestore(&grp->lock, flags);
3073 RCU_INIT_POINTER(p->numa_group, NULL);
3074 put_numa_group(grp);
3078 p->numa_faults = NULL;
3081 p->total_numa_faults = 0;
3082 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
3088 * Got a PROT_NONE fault for a page on @node.
3090 void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags)
3092 struct task_struct *p = current;
3093 bool migrated = flags & TNF_MIGRATED;
3094 int cpu_node = task_node(current);
3095 int local = !!(flags & TNF_FAULT_LOCAL);
3096 struct numa_group *ng;
3099 if (!static_branch_likely(&sched_numa_balancing))
3102 /* for example, ksmd faulting in a user's mm */
3107 * NUMA faults statistics are unnecessary for the slow memory
3108 * node for memory tiering mode.
3110 if (!node_is_toptier(mem_node) &&
3111 (sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING ||
3112 !cpupid_valid(last_cpupid)))
3115 /* Allocate buffer to track faults on a per-node basis */
3116 if (unlikely(!p->numa_faults)) {
3117 int size = sizeof(*p->numa_faults) *
3118 NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids;
3120 p->numa_faults = kzalloc(size, GFP_KERNEL|__GFP_NOWARN);
3121 if (!p->numa_faults)
3124 p->total_numa_faults = 0;
3125 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
3129 * First accesses are treated as private, otherwise consider accesses
3130 * to be private if the accessing pid has not changed
3132 if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) {
3135 priv = cpupid_match_pid(p, last_cpupid);
3136 if (!priv && !(flags & TNF_NO_GROUP))
3137 task_numa_group(p, last_cpupid, flags, &priv);
3141 * If a workload spans multiple NUMA nodes, a shared fault that
3142 * occurs wholly within the set of nodes that the workload is
3143 * actively using should be counted as local. This allows the
3144 * scan rate to slow down when a workload has settled down.
3146 ng = deref_curr_numa_group(p);
3147 if (!priv && !local && ng && ng->active_nodes > 1 &&
3148 numa_is_active_node(cpu_node, ng) &&
3149 numa_is_active_node(mem_node, ng))
3153 * Retry to migrate task to preferred node periodically, in case it
3154 * previously failed, or the scheduler moved us.
3156 if (time_after(jiffies, p->numa_migrate_retry)) {
3157 task_numa_placement(p);
3158 numa_migrate_preferred(p);
3162 p->numa_pages_migrated += pages;
3163 if (flags & TNF_MIGRATE_FAIL)
3164 p->numa_faults_locality[2] += pages;
3166 p->numa_faults[task_faults_idx(NUMA_MEMBUF, mem_node, priv)] += pages;
3167 p->numa_faults[task_faults_idx(NUMA_CPUBUF, cpu_node, priv)] += pages;
3168 p->numa_faults_locality[local] += pages;
3171 static void reset_ptenuma_scan(struct task_struct *p)
3174 * We only did a read acquisition of the mmap sem, so
3175 * p->mm->numa_scan_seq is written to without exclusive access
3176 * and the update is not guaranteed to be atomic. That's not
3177 * much of an issue though, since this is just used for
3178 * statistical sampling. Use READ_ONCE/WRITE_ONCE, which are not
3179 * expensive, to avoid any form of compiler optimizations:
3181 WRITE_ONCE(p->mm->numa_scan_seq, READ_ONCE(p->mm->numa_scan_seq) + 1);
3182 p->mm->numa_scan_offset = 0;
3185 static bool vma_is_accessed(struct vm_area_struct *vma)
3189 * Allow unconditional access first two times, so that all the (pages)
3190 * of VMAs get prot_none fault introduced irrespective of accesses.
3191 * This is also done to avoid any side effect of task scanning
3192 * amplifying the unfairness of disjoint set of VMAs' access.
3194 if (READ_ONCE(current->mm->numa_scan_seq) < 2)
3197 pids = vma->numab_state->access_pids[0] | vma->numab_state->access_pids[1];
3198 return test_bit(hash_32(current->pid, ilog2(BITS_PER_LONG)), &pids);
3201 #define VMA_PID_RESET_PERIOD (4 * sysctl_numa_balancing_scan_delay)
3204 * The expensive part of numa migration is done from task_work context.
3205 * Triggered from task_tick_numa().
3207 static void task_numa_work(struct callback_head *work)
3209 unsigned long migrate, next_scan, now = jiffies;
3210 struct task_struct *p = current;
3211 struct mm_struct *mm = p->mm;
3212 u64 runtime = p->se.sum_exec_runtime;
3213 struct vm_area_struct *vma;
3214 unsigned long start, end;
3215 unsigned long nr_pte_updates = 0;
3216 long pages, virtpages;
3217 struct vma_iterator vmi;
3219 SCHED_WARN_ON(p != container_of(work, struct task_struct, numa_work));
3223 * Who cares about NUMA placement when they're dying.
3225 * NOTE: make sure not to dereference p->mm before this check,
3226 * exit_task_work() happens _after_ exit_mm() so we could be called
3227 * without p->mm even though we still had it when we enqueued this
3230 if (p->flags & PF_EXITING)
3233 if (!mm->numa_next_scan) {
3234 mm->numa_next_scan = now +
3235 msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
3239 * Enforce maximal scan/migration frequency..
3241 migrate = mm->numa_next_scan;
3242 if (time_before(now, migrate))
3245 if (p->numa_scan_period == 0) {
3246 p->numa_scan_period_max = task_scan_max(p);
3247 p->numa_scan_period = task_scan_start(p);
3250 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
3251 if (!try_cmpxchg(&mm->numa_next_scan, &migrate, next_scan))
3255 * Delay this task enough that another task of this mm will likely win
3256 * the next time around.
3258 p->node_stamp += 2 * TICK_NSEC;
3260 start = mm->numa_scan_offset;
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))
3270 vma_iter_init(&vmi, mm, start);
3271 vma = vma_next(&vmi);
3273 reset_ptenuma_scan(p);
3275 vma_iter_set(&vmi, start);
3276 vma = vma_next(&vmi);
3280 if (!vma_migratable(vma) || !vma_policy_mof(vma) ||
3281 is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_MIXEDMAP)) {
3286 * Shared library pages mapped by multiple processes are not
3287 * migrated as it is expected they are cache replicated. Avoid
3288 * hinting faults in read-only file-backed mappings or the vdso
3289 * as migrating the pages will be of marginal benefit.
3292 (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ)))
3296 * Skip inaccessible VMAs to avoid any confusion between
3297 * PROT_NONE and NUMA hinting ptes
3299 if (!vma_is_accessible(vma))
3302 /* Initialise new per-VMA NUMAB state. */
3303 if (!vma->numab_state) {
3304 vma->numab_state = kzalloc(sizeof(struct vma_numab_state),
3306 if (!vma->numab_state)
3309 vma->numab_state->next_scan = now +
3310 msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
3312 /* Reset happens after 4 times scan delay of scan start */
3313 vma->numab_state->next_pid_reset = vma->numab_state->next_scan +
3314 msecs_to_jiffies(VMA_PID_RESET_PERIOD);
3318 * Scanning the VMA's of short lived tasks add more overhead. So
3319 * delay the scan for new VMAs.
3321 if (mm->numa_scan_seq && time_before(jiffies,
3322 vma->numab_state->next_scan))
3325 /* Do not scan the VMA if task has not accessed */
3326 if (!vma_is_accessed(vma))
3330 * RESET access PIDs regularly for old VMAs. Resetting after checking
3331 * vma for recent access to avoid clearing PID info before access..
3333 if (mm->numa_scan_seq &&
3334 time_after(jiffies, vma->numab_state->next_pid_reset)) {
3335 vma->numab_state->next_pid_reset = vma->numab_state->next_pid_reset +
3336 msecs_to_jiffies(VMA_PID_RESET_PERIOD);
3337 vma->numab_state->access_pids[0] = READ_ONCE(vma->numab_state->access_pids[1]);
3338 vma->numab_state->access_pids[1] = 0;
3342 start = max(start, vma->vm_start);
3343 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
3344 end = min(end, vma->vm_end);
3345 nr_pte_updates = change_prot_numa(vma, start, end);
3348 * Try to scan sysctl_numa_balancing_size worth of
3349 * hpages that have at least one present PTE that
3350 * is not already pte-numa. If the VMA contains
3351 * areas that are unused or already full of prot_numa
3352 * PTEs, scan up to virtpages, to skip through those
3356 pages -= (end - start) >> PAGE_SHIFT;
3357 virtpages -= (end - start) >> PAGE_SHIFT;
3360 if (pages <= 0 || virtpages <= 0)
3364 } while (end != vma->vm_end);
3365 } for_each_vma(vmi, vma);
3369 * It is possible to reach the end of the VMA list but the last few
3370 * VMAs are not guaranteed to the vma_migratable. If they are not, we
3371 * would find the !migratable VMA on the next scan but not reset the
3372 * scanner to the start so check it now.
3375 mm->numa_scan_offset = start;
3377 reset_ptenuma_scan(p);
3378 mmap_read_unlock(mm);
3381 * Make sure tasks use at least 32x as much time to run other code
3382 * than they used here, to limit NUMA PTE scanning overhead to 3% max.
3383 * Usually update_task_scan_period slows down scanning enough; on an
3384 * overloaded system we need to limit overhead on a per task basis.
3386 if (unlikely(p->se.sum_exec_runtime != runtime)) {
3387 u64 diff = p->se.sum_exec_runtime - runtime;
3388 p->node_stamp += 32 * diff;
3392 void init_numa_balancing(unsigned long clone_flags, struct task_struct *p)
3395 struct mm_struct *mm = p->mm;
3398 mm_users = atomic_read(&mm->mm_users);
3399 if (mm_users == 1) {
3400 mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
3401 mm->numa_scan_seq = 0;
3405 p->numa_scan_seq = mm ? mm->numa_scan_seq : 0;
3406 p->numa_scan_period = sysctl_numa_balancing_scan_delay;
3407 p->numa_migrate_retry = 0;
3408 /* Protect against double add, see task_tick_numa and task_numa_work */
3409 p->numa_work.next = &p->numa_work;
3410 p->numa_faults = NULL;
3411 p->numa_pages_migrated = 0;
3412 p->total_numa_faults = 0;
3413 RCU_INIT_POINTER(p->numa_group, NULL);
3414 p->last_task_numa_placement = 0;
3415 p->last_sum_exec_runtime = 0;
3417 init_task_work(&p->numa_work, task_numa_work);
3419 /* New address space, reset the preferred nid */
3420 if (!(clone_flags & CLONE_VM)) {
3421 p->numa_preferred_nid = NUMA_NO_NODE;
3426 * New thread, keep existing numa_preferred_nid which should be copied
3427 * already by arch_dup_task_struct but stagger when scans start.
3432 delay = min_t(unsigned int, task_scan_max(current),
3433 current->numa_scan_period * mm_users * NSEC_PER_MSEC);
3434 delay += 2 * TICK_NSEC;
3435 p->node_stamp = delay;
3440 * Drive the periodic memory faults..
3442 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
3444 struct callback_head *work = &curr->numa_work;
3448 * We don't care about NUMA placement if we don't have memory.
3450 if (!curr->mm || (curr->flags & (PF_EXITING | PF_KTHREAD)) || work->next != work)
3454 * Using runtime rather than walltime has the dual advantage that
3455 * we (mostly) drive the selection from busy threads and that the
3456 * task needs to have done some actual work before we bother with
3459 now = curr->se.sum_exec_runtime;
3460 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
3462 if (now > curr->node_stamp + period) {
3463 if (!curr->node_stamp)
3464 curr->numa_scan_period = task_scan_start(curr);
3465 curr->node_stamp += period;
3467 if (!time_before(jiffies, curr->mm->numa_next_scan))
3468 task_work_add(curr, work, TWA_RESUME);
3472 static void update_scan_period(struct task_struct *p, int new_cpu)
3474 int src_nid = cpu_to_node(task_cpu(p));
3475 int dst_nid = cpu_to_node(new_cpu);
3477 if (!static_branch_likely(&sched_numa_balancing))
3480 if (!p->mm || !p->numa_faults || (p->flags & PF_EXITING))
3483 if (src_nid == dst_nid)
3487 * Allow resets if faults have been trapped before one scan
3488 * has completed. This is most likely due to a new task that
3489 * is pulled cross-node due to wakeups or load balancing.
3491 if (p->numa_scan_seq) {
3493 * Avoid scan adjustments if moving to the preferred
3494 * node or if the task was not previously running on
3495 * the preferred node.
3497 if (dst_nid == p->numa_preferred_nid ||
3498 (p->numa_preferred_nid != NUMA_NO_NODE &&
3499 src_nid != p->numa_preferred_nid))
3503 p->numa_scan_period = task_scan_start(p);
3507 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
3511 static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p)
3515 static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p)
3519 static inline void update_scan_period(struct task_struct *p, int new_cpu)
3523 #endif /* CONFIG_NUMA_BALANCING */
3526 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
3528 update_load_add(&cfs_rq->load, se->load.weight);
3530 if (entity_is_task(se)) {
3531 struct rq *rq = rq_of(cfs_rq);
3533 account_numa_enqueue(rq, task_of(se));
3534 list_add(&se->group_node, &rq->cfs_tasks);
3537 cfs_rq->nr_running++;
3539 cfs_rq->idle_nr_running++;
3543 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
3545 update_load_sub(&cfs_rq->load, se->load.weight);
3547 if (entity_is_task(se)) {
3548 account_numa_dequeue(rq_of(cfs_rq), task_of(se));
3549 list_del_init(&se->group_node);
3552 cfs_rq->nr_running--;
3554 cfs_rq->idle_nr_running--;
3558 * Signed add and clamp on underflow.
3560 * Explicitly do a load-store to ensure the intermediate value never hits
3561 * memory. This allows lockless observations without ever seeing the negative
3564 #define add_positive(_ptr, _val) do { \
3565 typeof(_ptr) ptr = (_ptr); \
3566 typeof(_val) val = (_val); \
3567 typeof(*ptr) res, var = READ_ONCE(*ptr); \
3571 if (val < 0 && res > var) \
3574 WRITE_ONCE(*ptr, res); \
3578 * Unsigned subtract and clamp on underflow.
3580 * Explicitly do a load-store to ensure the intermediate value never hits
3581 * memory. This allows lockless observations without ever seeing the negative
3584 #define sub_positive(_ptr, _val) do { \
3585 typeof(_ptr) ptr = (_ptr); \
3586 typeof(*ptr) val = (_val); \
3587 typeof(*ptr) res, var = READ_ONCE(*ptr); \
3591 WRITE_ONCE(*ptr, res); \
3595 * Remove and clamp on negative, from a local variable.
3597 * A variant of sub_positive(), which does not use explicit load-store
3598 * and is thus optimized for local variable updates.
3600 #define lsub_positive(_ptr, _val) do { \
3601 typeof(_ptr) ptr = (_ptr); \
3602 *ptr -= min_t(typeof(*ptr), *ptr, _val); \
3607 enqueue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3609 cfs_rq->avg.load_avg += se->avg.load_avg;
3610 cfs_rq->avg.load_sum += se_weight(se) * se->avg.load_sum;
3614 dequeue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3616 sub_positive(&cfs_rq->avg.load_avg, se->avg.load_avg);
3617 sub_positive(&cfs_rq->avg.load_sum, se_weight(se) * se->avg.load_sum);
3618 /* See update_cfs_rq_load_avg() */
3619 cfs_rq->avg.load_sum = max_t(u32, cfs_rq->avg.load_sum,
3620 cfs_rq->avg.load_avg * PELT_MIN_DIVIDER);
3624 enqueue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
3626 dequeue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
3629 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
3630 unsigned long weight)
3632 unsigned long old_weight = se->load.weight;
3635 /* commit outstanding execution time */
3636 if (cfs_rq->curr == se)
3637 update_curr(cfs_rq);
3639 avg_vruntime_sub(cfs_rq, se);
3640 update_load_sub(&cfs_rq->load, se->load.weight);
3642 dequeue_load_avg(cfs_rq, se);
3644 update_load_set(&se->load, weight);
3648 * Because we keep se->vlag = V - v_i, while: lag_i = w_i*(V - v_i),
3649 * we need to scale se->vlag when w_i changes.
3651 se->vlag = div_s64(se->vlag * old_weight, weight);
3653 s64 deadline = se->deadline - se->vruntime;
3655 * When the weight changes, the virtual time slope changes and
3656 * we should adjust the relative virtual deadline accordingly.
3658 deadline = div_s64(deadline * old_weight, weight);
3659 se->deadline = se->vruntime + deadline;
3660 min_deadline_cb_propagate(&se->run_node, NULL);
3665 u32 divider = get_pelt_divider(&se->avg);
3667 se->avg.load_avg = div_u64(se_weight(se) * se->avg.load_sum, divider);
3671 enqueue_load_avg(cfs_rq, se);
3673 update_load_add(&cfs_rq->load, se->load.weight);
3674 if (cfs_rq->curr != se)
3675 avg_vruntime_add(cfs_rq, se);
3679 void reweight_task(struct task_struct *p, int prio)
3681 struct sched_entity *se = &p->se;
3682 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3683 struct load_weight *load = &se->load;
3684 unsigned long weight = scale_load(sched_prio_to_weight[prio]);
3686 reweight_entity(cfs_rq, se, weight);
3687 load->inv_weight = sched_prio_to_wmult[prio];
3690 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
3692 #ifdef CONFIG_FAIR_GROUP_SCHED
3695 * All this does is approximate the hierarchical proportion which includes that
3696 * global sum we all love to hate.
3698 * That is, the weight of a group entity, is the proportional share of the
3699 * group weight based on the group runqueue weights. That is:
3701 * tg->weight * grq->load.weight
3702 * ge->load.weight = ----------------------------- (1)
3703 * \Sum grq->load.weight
3705 * Now, because computing that sum is prohibitively expensive to compute (been
3706 * there, done that) we approximate it with this average stuff. The average
3707 * moves slower and therefore the approximation is cheaper and more stable.
3709 * So instead of the above, we substitute:
3711 * grq->load.weight -> grq->avg.load_avg (2)
3713 * which yields the following:
3715 * tg->weight * grq->avg.load_avg
3716 * ge->load.weight = ------------------------------ (3)
3719 * Where: tg->load_avg ~= \Sum grq->avg.load_avg
3721 * That is shares_avg, and it is right (given the approximation (2)).
3723 * The problem with it is that because the average is slow -- it was designed
3724 * to be exactly that of course -- this leads to transients in boundary
3725 * conditions. In specific, the case where the group was idle and we start the
3726 * one task. It takes time for our CPU's grq->avg.load_avg to build up,
3727 * yielding bad latency etc..
3729 * Now, in that special case (1) reduces to:
3731 * tg->weight * grq->load.weight
3732 * ge->load.weight = ----------------------------- = tg->weight (4)
3735 * That is, the sum collapses because all other CPUs are idle; the UP scenario.
3737 * So what we do is modify our approximation (3) to approach (4) in the (near)
3742 * tg->weight * grq->load.weight
3743 * --------------------------------------------------- (5)
3744 * tg->load_avg - grq->avg.load_avg + grq->load.weight
3746 * But because grq->load.weight can drop to 0, resulting in a divide by zero,
3747 * we need to use grq->avg.load_avg as its lower bound, which then gives:
3750 * tg->weight * grq->load.weight
3751 * ge->load.weight = ----------------------------- (6)
3756 * tg_load_avg' = tg->load_avg - grq->avg.load_avg +
3757 * max(grq->load.weight, grq->avg.load_avg)
3759 * And that is shares_weight and is icky. In the (near) UP case it approaches
3760 * (4) while in the normal case it approaches (3). It consistently
3761 * overestimates the ge->load.weight and therefore:
3763 * \Sum ge->load.weight >= tg->weight
3767 static long calc_group_shares(struct cfs_rq *cfs_rq)
3769 long tg_weight, tg_shares, load, shares;
3770 struct task_group *tg = cfs_rq->tg;
3772 tg_shares = READ_ONCE(tg->shares);
3774 load = max(scale_load_down(cfs_rq->load.weight), cfs_rq->avg.load_avg);
3776 tg_weight = atomic_long_read(&tg->load_avg);
3778 /* Ensure tg_weight >= load */
3779 tg_weight -= cfs_rq->tg_load_avg_contrib;
3782 shares = (tg_shares * load);
3784 shares /= tg_weight;
3787 * MIN_SHARES has to be unscaled here to support per-CPU partitioning
3788 * of a group with small tg->shares value. It is a floor value which is
3789 * assigned as a minimum load.weight to the sched_entity representing
3790 * the group on a CPU.
3792 * E.g. on 64-bit for a group with tg->shares of scale_load(15)=15*1024
3793 * on an 8-core system with 8 tasks each runnable on one CPU shares has
3794 * to be 15*1024*1/8=1920 instead of scale_load(MIN_SHARES)=2*1024. In
3795 * case no task is runnable on a CPU MIN_SHARES=2 should be returned
3798 return clamp_t(long, shares, MIN_SHARES, tg_shares);
3800 #endif /* CONFIG_SMP */
3803 * Recomputes the group entity based on the current state of its group
3806 static void update_cfs_group(struct sched_entity *se)
3808 struct cfs_rq *gcfs_rq = group_cfs_rq(se);
3814 if (throttled_hierarchy(gcfs_rq))
3818 shares = READ_ONCE(gcfs_rq->tg->shares);
3820 if (likely(se->load.weight == shares))
3823 shares = calc_group_shares(gcfs_rq);
3826 reweight_entity(cfs_rq_of(se), se, shares);
3829 #else /* CONFIG_FAIR_GROUP_SCHED */
3830 static inline void update_cfs_group(struct sched_entity *se)
3833 #endif /* CONFIG_FAIR_GROUP_SCHED */
3835 static inline void cfs_rq_util_change(struct cfs_rq *cfs_rq, int flags)
3837 struct rq *rq = rq_of(cfs_rq);
3839 if (&rq->cfs == cfs_rq) {
3841 * There are a few boundary cases this might miss but it should
3842 * get called often enough that that should (hopefully) not be
3845 * It will not get called when we go idle, because the idle
3846 * thread is a different class (!fair), nor will the utilization
3847 * number include things like RT tasks.
3849 * As is, the util number is not freq-invariant (we'd have to
3850 * implement arch_scale_freq_capacity() for that).
3852 * See cpu_util_cfs().
3854 cpufreq_update_util(rq, flags);
3859 static inline bool load_avg_is_decayed(struct sched_avg *sa)
3867 if (sa->runnable_sum)
3871 * _avg must be null when _sum are null because _avg = _sum / divider
3872 * Make sure that rounding and/or propagation of PELT values never
3875 SCHED_WARN_ON(sa->load_avg ||
3882 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
3884 return u64_u32_load_copy(cfs_rq->avg.last_update_time,
3885 cfs_rq->last_update_time_copy);
3887 #ifdef CONFIG_FAIR_GROUP_SCHED
3889 * Because list_add_leaf_cfs_rq always places a child cfs_rq on the list
3890 * immediately before a parent cfs_rq, and cfs_rqs are removed from the list
3891 * bottom-up, we only have to test whether the cfs_rq before us on the list
3893 * If cfs_rq is not on the list, test whether a child needs its to be added to
3894 * connect a branch to the tree * (see list_add_leaf_cfs_rq() for details).
3896 static inline bool child_cfs_rq_on_list(struct cfs_rq *cfs_rq)
3898 struct cfs_rq *prev_cfs_rq;
3899 struct list_head *prev;
3901 if (cfs_rq->on_list) {
3902 prev = cfs_rq->leaf_cfs_rq_list.prev;
3904 struct rq *rq = rq_of(cfs_rq);
3906 prev = rq->tmp_alone_branch;
3909 prev_cfs_rq = container_of(prev, struct cfs_rq, leaf_cfs_rq_list);
3911 return (prev_cfs_rq->tg->parent == cfs_rq->tg);
3914 static inline bool cfs_rq_is_decayed(struct cfs_rq *cfs_rq)
3916 if (cfs_rq->load.weight)
3919 if (!load_avg_is_decayed(&cfs_rq->avg))
3922 if (child_cfs_rq_on_list(cfs_rq))
3929 * update_tg_load_avg - update the tg's load avg
3930 * @cfs_rq: the cfs_rq whose avg changed
3932 * This function 'ensures': tg->load_avg := \Sum tg->cfs_rq[]->avg.load.
3933 * However, because tg->load_avg is a global value there are performance
3936 * In order to avoid having to look at the other cfs_rq's, we use a
3937 * differential update where we store the last value we propagated. This in
3938 * turn allows skipping updates if the differential is 'small'.
3940 * Updating tg's load_avg is necessary before update_cfs_share().
3942 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq)
3944 long delta = cfs_rq->avg.load_avg - cfs_rq->tg_load_avg_contrib;
3947 * No need to update load_avg for root_task_group as it is not used.
3949 if (cfs_rq->tg == &root_task_group)
3952 if (abs(delta) > cfs_rq->tg_load_avg_contrib / 64) {
3953 atomic_long_add(delta, &cfs_rq->tg->load_avg);
3954 cfs_rq->tg_load_avg_contrib = cfs_rq->avg.load_avg;
3959 * Called within set_task_rq() right before setting a task's CPU. The
3960 * caller only guarantees p->pi_lock is held; no other assumptions,
3961 * including the state of rq->lock, should be made.
3963 void set_task_rq_fair(struct sched_entity *se,
3964 struct cfs_rq *prev, struct cfs_rq *next)
3966 u64 p_last_update_time;
3967 u64 n_last_update_time;
3969 if (!sched_feat(ATTACH_AGE_LOAD))
3973 * We are supposed to update the task to "current" time, then its up to
3974 * date and ready to go to new CPU/cfs_rq. But we have difficulty in
3975 * getting what current time is, so simply throw away the out-of-date
3976 * time. This will result in the wakee task is less decayed, but giving
3977 * the wakee more load sounds not bad.
3979 if (!(se->avg.last_update_time && prev))
3982 p_last_update_time = cfs_rq_last_update_time(prev);
3983 n_last_update_time = cfs_rq_last_update_time(next);
3985 __update_load_avg_blocked_se(p_last_update_time, se);
3986 se->avg.last_update_time = n_last_update_time;
3990 * When on migration a sched_entity joins/leaves the PELT hierarchy, we need to
3991 * propagate its contribution. The key to this propagation is the invariant
3992 * that for each group:
3994 * ge->avg == grq->avg (1)
3996 * _IFF_ we look at the pure running and runnable sums. Because they
3997 * represent the very same entity, just at different points in the hierarchy.
3999 * Per the above update_tg_cfs_util() and update_tg_cfs_runnable() are trivial
4000 * and simply copies the running/runnable sum over (but still wrong, because
4001 * the group entity and group rq do not have their PELT windows aligned).
4003 * However, update_tg_cfs_load() is more complex. So we have:
4005 * ge->avg.load_avg = ge->load.weight * ge->avg.runnable_avg (2)
4007 * And since, like util, the runnable part should be directly transferable,
4008 * the following would _appear_ to be the straight forward approach:
4010 * grq->avg.load_avg = grq->load.weight * grq->avg.runnable_avg (3)
4012 * And per (1) we have:
4014 * ge->avg.runnable_avg == grq->avg.runnable_avg
4018 * ge->load.weight * grq->avg.load_avg
4019 * ge->avg.load_avg = ----------------------------------- (4)
4022 * Except that is wrong!
4024 * Because while for entities historical weight is not important and we
4025 * really only care about our future and therefore can consider a pure
4026 * runnable sum, runqueues can NOT do this.
4028 * We specifically want runqueues to have a load_avg that includes
4029 * historical weights. Those represent the blocked load, the load we expect
4030 * to (shortly) return to us. This only works by keeping the weights as
4031 * integral part of the sum. We therefore cannot decompose as per (3).
4033 * Another reason this doesn't work is that runnable isn't a 0-sum entity.
4034 * Imagine a rq with 2 tasks that each are runnable 2/3 of the time. Then the
4035 * rq itself is runnable anywhere between 2/3 and 1 depending on how the
4036 * runnable section of these tasks overlap (or not). If they were to perfectly
4037 * align the rq as a whole would be runnable 2/3 of the time. If however we
4038 * always have at least 1 runnable task, the rq as a whole is always runnable.
4040 * So we'll have to approximate.. :/
4042 * Given the constraint:
4044 * ge->avg.running_sum <= ge->avg.runnable_sum <= LOAD_AVG_MAX
4046 * We can construct a rule that adds runnable to a rq by assuming minimal
4049 * On removal, we'll assume each task is equally runnable; which yields:
4051 * grq->avg.runnable_sum = grq->avg.load_sum / grq->load.weight
4053 * XXX: only do this for the part of runnable > running ?
4057 update_tg_cfs_util(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq)
4059 long delta_sum, delta_avg = gcfs_rq->avg.util_avg - se->avg.util_avg;
4060 u32 new_sum, divider;
4062 /* Nothing to update */
4067 * cfs_rq->avg.period_contrib can be used for both cfs_rq and se.
4068 * See ___update_load_avg() for details.
4070 divider = get_pelt_divider(&cfs_rq->avg);
4073 /* Set new sched_entity's utilization */
4074 se->avg.util_avg = gcfs_rq->avg.util_avg;
4075 new_sum = se->avg.util_avg * divider;
4076 delta_sum = (long)new_sum - (long)se->avg.util_sum;
4077 se->avg.util_sum = new_sum;
4079 /* Update parent cfs_rq utilization */
4080 add_positive(&cfs_rq->avg.util_avg, delta_avg);
4081 add_positive(&cfs_rq->avg.util_sum, delta_sum);
4083 /* See update_cfs_rq_load_avg() */
4084 cfs_rq->avg.util_sum = max_t(u32, cfs_rq->avg.util_sum,
4085 cfs_rq->avg.util_avg * PELT_MIN_DIVIDER);
4089 update_tg_cfs_runnable(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq)
4091 long delta_sum, delta_avg = gcfs_rq->avg.runnable_avg - se->avg.runnable_avg;
4092 u32 new_sum, divider;
4094 /* Nothing to update */
4099 * cfs_rq->avg.period_contrib can be used for both cfs_rq and se.
4100 * See ___update_load_avg() for details.
4102 divider = get_pelt_divider(&cfs_rq->avg);
4104 /* Set new sched_entity's runnable */
4105 se->avg.runnable_avg = gcfs_rq->avg.runnable_avg;
4106 new_sum = se->avg.runnable_avg * divider;
4107 delta_sum = (long)new_sum - (long)se->avg.runnable_sum;
4108 se->avg.runnable_sum = new_sum;
4110 /* Update parent cfs_rq runnable */
4111 add_positive(&cfs_rq->avg.runnable_avg, delta_avg);
4112 add_positive(&cfs_rq->avg.runnable_sum, delta_sum);
4113 /* See update_cfs_rq_load_avg() */
4114 cfs_rq->avg.runnable_sum = max_t(u32, cfs_rq->avg.runnable_sum,
4115 cfs_rq->avg.runnable_avg * PELT_MIN_DIVIDER);
4119 update_tg_cfs_load(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq)
4121 long delta_avg, running_sum, runnable_sum = gcfs_rq->prop_runnable_sum;
4122 unsigned long load_avg;
4130 gcfs_rq->prop_runnable_sum = 0;
4133 * cfs_rq->avg.period_contrib can be used for both cfs_rq and se.
4134 * See ___update_load_avg() for details.
4136 divider = get_pelt_divider(&cfs_rq->avg);
4138 if (runnable_sum >= 0) {
4140 * Add runnable; clip at LOAD_AVG_MAX. Reflects that until
4141 * the CPU is saturated running == runnable.
4143 runnable_sum += se->avg.load_sum;
4144 runnable_sum = min_t(long, runnable_sum, divider);
4147 * Estimate the new unweighted runnable_sum of the gcfs_rq by
4148 * assuming all tasks are equally runnable.
4150 if (scale_load_down(gcfs_rq->load.weight)) {
4151 load_sum = div_u64(gcfs_rq->avg.load_sum,
4152 scale_load_down(gcfs_rq->load.weight));
4155 /* But make sure to not inflate se's runnable */
4156 runnable_sum = min(se->avg.load_sum, load_sum);
4160 * runnable_sum can't be lower than running_sum
4161 * Rescale running sum to be in the same range as runnable sum
4162 * running_sum is in [0 : LOAD_AVG_MAX << SCHED_CAPACITY_SHIFT]
4163 * runnable_sum is in [0 : LOAD_AVG_MAX]
4165 running_sum = se->avg.util_sum >> SCHED_CAPACITY_SHIFT;
4166 runnable_sum = max(runnable_sum, running_sum);
4168 load_sum = se_weight(se) * runnable_sum;
4169 load_avg = div_u64(load_sum, divider);
4171 delta_avg = load_avg - se->avg.load_avg;
4175 delta_sum = load_sum - (s64)se_weight(se) * se->avg.load_sum;
4177 se->avg.load_sum = runnable_sum;
4178 se->avg.load_avg = load_avg;
4179 add_positive(&cfs_rq->avg.load_avg, delta_avg);
4180 add_positive(&cfs_rq->avg.load_sum, delta_sum);
4181 /* See update_cfs_rq_load_avg() */
4182 cfs_rq->avg.load_sum = max_t(u32, cfs_rq->avg.load_sum,
4183 cfs_rq->avg.load_avg * PELT_MIN_DIVIDER);
4186 static inline void add_tg_cfs_propagate(struct cfs_rq *cfs_rq, long runnable_sum)
4188 cfs_rq->propagate = 1;
4189 cfs_rq->prop_runnable_sum += runnable_sum;
4192 /* Update task and its cfs_rq load average */
4193 static inline int propagate_entity_load_avg(struct sched_entity *se)
4195 struct cfs_rq *cfs_rq, *gcfs_rq;
4197 if (entity_is_task(se))
4200 gcfs_rq = group_cfs_rq(se);
4201 if (!gcfs_rq->propagate)
4204 gcfs_rq->propagate = 0;
4206 cfs_rq = cfs_rq_of(se);
4208 add_tg_cfs_propagate(cfs_rq, gcfs_rq->prop_runnable_sum);
4210 update_tg_cfs_util(cfs_rq, se, gcfs_rq);
4211 update_tg_cfs_runnable(cfs_rq, se, gcfs_rq);
4212 update_tg_cfs_load(cfs_rq, se, gcfs_rq);
4214 trace_pelt_cfs_tp(cfs_rq);
4215 trace_pelt_se_tp(se);
4221 * Check if we need to update the load and the utilization of a blocked
4224 static inline bool skip_blocked_update(struct sched_entity *se)
4226 struct cfs_rq *gcfs_rq = group_cfs_rq(se);
4229 * If sched_entity still have not zero load or utilization, we have to
4232 if (se->avg.load_avg || se->avg.util_avg)
4236 * If there is a pending propagation, we have to update the load and
4237 * the utilization of the sched_entity:
4239 if (gcfs_rq->propagate)
4243 * Otherwise, the load and the utilization of the sched_entity is
4244 * already zero and there is no pending propagation, so it will be a
4245 * waste of time to try to decay it:
4250 #else /* CONFIG_FAIR_GROUP_SCHED */
4252 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq) {}
4254 static inline int propagate_entity_load_avg(struct sched_entity *se)
4259 static inline void add_tg_cfs_propagate(struct cfs_rq *cfs_rq, long runnable_sum) {}
4261 #endif /* CONFIG_FAIR_GROUP_SCHED */
4263 #ifdef CONFIG_NO_HZ_COMMON
4264 static inline void migrate_se_pelt_lag(struct sched_entity *se)
4266 u64 throttled = 0, now, lut;
4267 struct cfs_rq *cfs_rq;
4271 if (load_avg_is_decayed(&se->avg))
4274 cfs_rq = cfs_rq_of(se);
4278 is_idle = is_idle_task(rcu_dereference(rq->curr));
4282 * The lag estimation comes with a cost we don't want to pay all the
4283 * time. Hence, limiting to the case where the source CPU is idle and
4284 * we know we are at the greatest risk to have an outdated clock.
4290 * Estimated "now" is: last_update_time + cfs_idle_lag + rq_idle_lag, where:
4292 * last_update_time (the cfs_rq's last_update_time)
4293 * = cfs_rq_clock_pelt()@cfs_rq_idle
4294 * = rq_clock_pelt()@cfs_rq_idle
4295 * - cfs->throttled_clock_pelt_time@cfs_rq_idle
4297 * cfs_idle_lag (delta between rq's update and cfs_rq's update)
4298 * = rq_clock_pelt()@rq_idle - rq_clock_pelt()@cfs_rq_idle
4300 * rq_idle_lag (delta between now and rq's update)
4301 * = sched_clock_cpu() - rq_clock()@rq_idle
4303 * We can then write:
4305 * now = rq_clock_pelt()@rq_idle - cfs->throttled_clock_pelt_time +
4306 * sched_clock_cpu() - rq_clock()@rq_idle
4308 * rq_clock_pelt()@rq_idle is rq->clock_pelt_idle
4309 * rq_clock()@rq_idle is rq->clock_idle
4310 * cfs->throttled_clock_pelt_time@cfs_rq_idle
4311 * is cfs_rq->throttled_pelt_idle
4314 #ifdef CONFIG_CFS_BANDWIDTH
4315 throttled = u64_u32_load(cfs_rq->throttled_pelt_idle);
4316 /* The clock has been stopped for throttling */
4317 if (throttled == U64_MAX)
4320 now = u64_u32_load(rq->clock_pelt_idle);
4322 * Paired with _update_idle_rq_clock_pelt(). It ensures at the worst case
4323 * is observed the old clock_pelt_idle value and the new clock_idle,
4324 * which lead to an underestimation. The opposite would lead to an
4328 lut = cfs_rq_last_update_time(cfs_rq);
4333 * cfs_rq->avg.last_update_time is more recent than our
4334 * estimation, let's use it.
4338 now += sched_clock_cpu(cpu_of(rq)) - u64_u32_load(rq->clock_idle);
4340 __update_load_avg_blocked_se(now, se);
4343 static void migrate_se_pelt_lag(struct sched_entity *se) {}
4347 * update_cfs_rq_load_avg - update the cfs_rq's load/util averages
4348 * @now: current time, as per cfs_rq_clock_pelt()
4349 * @cfs_rq: cfs_rq to update
4351 * The cfs_rq avg is the direct sum of all its entities (blocked and runnable)
4352 * avg. The immediate corollary is that all (fair) tasks must be attached.
4354 * cfs_rq->avg is used for task_h_load() and update_cfs_share() for example.
4356 * Return: true if the load decayed or we removed load.
4358 * Since both these conditions indicate a changed cfs_rq->avg.load we should
4359 * call update_tg_load_avg() when this function returns true.
4362 update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq)
4364 unsigned long removed_load = 0, removed_util = 0, removed_runnable = 0;
4365 struct sched_avg *sa = &cfs_rq->avg;
4368 if (cfs_rq->removed.nr) {
4370 u32 divider = get_pelt_divider(&cfs_rq->avg);
4372 raw_spin_lock(&cfs_rq->removed.lock);
4373 swap(cfs_rq->removed.util_avg, removed_util);
4374 swap(cfs_rq->removed.load_avg, removed_load);
4375 swap(cfs_rq->removed.runnable_avg, removed_runnable);
4376 cfs_rq->removed.nr = 0;
4377 raw_spin_unlock(&cfs_rq->removed.lock);
4380 sub_positive(&sa->load_avg, r);
4381 sub_positive(&sa->load_sum, r * divider);
4382 /* See sa->util_sum below */
4383 sa->load_sum = max_t(u32, sa->load_sum, sa->load_avg * PELT_MIN_DIVIDER);
4386 sub_positive(&sa->util_avg, r);
4387 sub_positive(&sa->util_sum, r * divider);
4389 * Because of rounding, se->util_sum might ends up being +1 more than
4390 * cfs->util_sum. Although this is not a problem by itself, detaching
4391 * a lot of tasks with the rounding problem between 2 updates of
4392 * util_avg (~1ms) can make cfs->util_sum becoming null whereas
4393 * cfs_util_avg is not.
4394 * Check that util_sum is still above its lower bound for the new
4395 * util_avg. Given that period_contrib might have moved since the last
4396 * sync, we are only sure that util_sum must be above or equal to
4397 * util_avg * minimum possible divider
4399 sa->util_sum = max_t(u32, sa->util_sum, sa->util_avg * PELT_MIN_DIVIDER);
4401 r = removed_runnable;
4402 sub_positive(&sa->runnable_avg, r);
4403 sub_positive(&sa->runnable_sum, r * divider);
4404 /* See sa->util_sum above */
4405 sa->runnable_sum = max_t(u32, sa->runnable_sum,
4406 sa->runnable_avg * PELT_MIN_DIVIDER);
4409 * removed_runnable is the unweighted version of removed_load so we
4410 * can use it to estimate removed_load_sum.
4412 add_tg_cfs_propagate(cfs_rq,
4413 -(long)(removed_runnable * divider) >> SCHED_CAPACITY_SHIFT);
4418 decayed |= __update_load_avg_cfs_rq(now, cfs_rq);
4419 u64_u32_store_copy(sa->last_update_time,
4420 cfs_rq->last_update_time_copy,
4421 sa->last_update_time);
4426 * attach_entity_load_avg - attach this entity to its cfs_rq load avg
4427 * @cfs_rq: cfs_rq to attach to
4428 * @se: sched_entity to attach
4430 * Must call update_cfs_rq_load_avg() before this, since we rely on
4431 * cfs_rq->avg.last_update_time being current.
4433 static void attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
4436 * cfs_rq->avg.period_contrib can be used for both cfs_rq and se.
4437 * See ___update_load_avg() for details.
4439 u32 divider = get_pelt_divider(&cfs_rq->avg);
4442 * When we attach the @se to the @cfs_rq, we must align the decay
4443 * window because without that, really weird and wonderful things can
4448 se->avg.last_update_time = cfs_rq->avg.last_update_time;
4449 se->avg.period_contrib = cfs_rq->avg.period_contrib;
4452 * Hell(o) Nasty stuff.. we need to recompute _sum based on the new
4453 * period_contrib. This isn't strictly correct, but since we're
4454 * entirely outside of the PELT hierarchy, nobody cares if we truncate
4457 se->avg.util_sum = se->avg.util_avg * divider;
4459 se->avg.runnable_sum = se->avg.runnable_avg * divider;
4461 se->avg.load_sum = se->avg.load_avg * divider;
4462 if (se_weight(se) < se->avg.load_sum)
4463 se->avg.load_sum = div_u64(se->avg.load_sum, se_weight(se));
4465 se->avg.load_sum = 1;
4467 enqueue_load_avg(cfs_rq, se);
4468 cfs_rq->avg.util_avg += se->avg.util_avg;
4469 cfs_rq->avg.util_sum += se->avg.util_sum;
4470 cfs_rq->avg.runnable_avg += se->avg.runnable_avg;
4471 cfs_rq->avg.runnable_sum += se->avg.runnable_sum;
4473 add_tg_cfs_propagate(cfs_rq, se->avg.load_sum);
4475 cfs_rq_util_change(cfs_rq, 0);
4477 trace_pelt_cfs_tp(cfs_rq);
4481 * detach_entity_load_avg - detach this entity from its cfs_rq load avg
4482 * @cfs_rq: cfs_rq to detach from
4483 * @se: sched_entity to detach
4485 * Must call update_cfs_rq_load_avg() before this, since we rely on
4486 * cfs_rq->avg.last_update_time being current.
4488 static void detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
4490 dequeue_load_avg(cfs_rq, se);
4491 sub_positive(&cfs_rq->avg.util_avg, se->avg.util_avg);
4492 sub_positive(&cfs_rq->avg.util_sum, se->avg.util_sum);
4493 /* See update_cfs_rq_load_avg() */
4494 cfs_rq->avg.util_sum = max_t(u32, cfs_rq->avg.util_sum,
4495 cfs_rq->avg.util_avg * PELT_MIN_DIVIDER);
4497 sub_positive(&cfs_rq->avg.runnable_avg, se->avg.runnable_avg);
4498 sub_positive(&cfs_rq->avg.runnable_sum, se->avg.runnable_sum);
4499 /* See update_cfs_rq_load_avg() */
4500 cfs_rq->avg.runnable_sum = max_t(u32, cfs_rq->avg.runnable_sum,
4501 cfs_rq->avg.runnable_avg * PELT_MIN_DIVIDER);
4503 add_tg_cfs_propagate(cfs_rq, -se->avg.load_sum);
4505 cfs_rq_util_change(cfs_rq, 0);
4507 trace_pelt_cfs_tp(cfs_rq);
4511 * Optional action to be done while updating the load average
4513 #define UPDATE_TG 0x1
4514 #define SKIP_AGE_LOAD 0x2
4515 #define DO_ATTACH 0x4
4516 #define DO_DETACH 0x8
4518 /* Update task and its cfs_rq load average */
4519 static inline void update_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
4521 u64 now = cfs_rq_clock_pelt(cfs_rq);
4525 * Track task load average for carrying it to new CPU after migrated, and
4526 * track group sched_entity load average for task_h_load calc in migration
4528 if (se->avg.last_update_time && !(flags & SKIP_AGE_LOAD))
4529 __update_load_avg_se(now, cfs_rq, se);
4531 decayed = update_cfs_rq_load_avg(now, cfs_rq);
4532 decayed |= propagate_entity_load_avg(se);
4534 if (!se->avg.last_update_time && (flags & DO_ATTACH)) {
4537 * DO_ATTACH means we're here from enqueue_entity().
4538 * !last_update_time means we've passed through
4539 * migrate_task_rq_fair() indicating we migrated.
4541 * IOW we're enqueueing a task on a new CPU.
4543 attach_entity_load_avg(cfs_rq, se);
4544 update_tg_load_avg(cfs_rq);
4546 } else if (flags & DO_DETACH) {
4548 * DO_DETACH means we're here from dequeue_entity()
4549 * and we are migrating task out of the CPU.
4551 detach_entity_load_avg(cfs_rq, se);
4552 update_tg_load_avg(cfs_rq);
4553 } else if (decayed) {
4554 cfs_rq_util_change(cfs_rq, 0);
4556 if (flags & UPDATE_TG)
4557 update_tg_load_avg(cfs_rq);
4562 * Synchronize entity load avg of dequeued entity without locking
4565 static void sync_entity_load_avg(struct sched_entity *se)
4567 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4568 u64 last_update_time;
4570 last_update_time = cfs_rq_last_update_time(cfs_rq);
4571 __update_load_avg_blocked_se(last_update_time, se);
4575 * Task first catches up with cfs_rq, and then subtract
4576 * itself from the cfs_rq (task must be off the queue now).
4578 static void remove_entity_load_avg(struct sched_entity *se)
4580 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4581 unsigned long flags;
4584 * tasks cannot exit without having gone through wake_up_new_task() ->
4585 * enqueue_task_fair() which will have added things to the cfs_rq,
4586 * so we can remove unconditionally.
4589 sync_entity_load_avg(se);
4591 raw_spin_lock_irqsave(&cfs_rq->removed.lock, flags);
4592 ++cfs_rq->removed.nr;
4593 cfs_rq->removed.util_avg += se->avg.util_avg;
4594 cfs_rq->removed.load_avg += se->avg.load_avg;
4595 cfs_rq->removed.runnable_avg += se->avg.runnable_avg;
4596 raw_spin_unlock_irqrestore(&cfs_rq->removed.lock, flags);
4599 static inline unsigned long cfs_rq_runnable_avg(struct cfs_rq *cfs_rq)
4601 return cfs_rq->avg.runnable_avg;
4604 static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq)
4606 return cfs_rq->avg.load_avg;
4609 static int newidle_balance(struct rq *this_rq, struct rq_flags *rf);
4611 static inline unsigned long task_util(struct task_struct *p)
4613 return READ_ONCE(p->se.avg.util_avg);
4616 static inline unsigned long _task_util_est(struct task_struct *p)
4618 struct util_est ue = READ_ONCE(p->se.avg.util_est);
4620 return max(ue.ewma, (ue.enqueued & ~UTIL_AVG_UNCHANGED));
4623 static inline unsigned long task_util_est(struct task_struct *p)
4625 return max(task_util(p), _task_util_est(p));
4628 #ifdef CONFIG_UCLAMP_TASK
4629 static inline unsigned long uclamp_task_util(struct task_struct *p,
4630 unsigned long uclamp_min,
4631 unsigned long uclamp_max)
4633 return clamp(task_util_est(p), uclamp_min, uclamp_max);
4636 static inline unsigned long uclamp_task_util(struct task_struct *p,
4637 unsigned long uclamp_min,
4638 unsigned long uclamp_max)
4640 return task_util_est(p);
4644 static inline void util_est_enqueue(struct cfs_rq *cfs_rq,
4645 struct task_struct *p)
4647 unsigned int enqueued;
4649 if (!sched_feat(UTIL_EST))
4652 /* Update root cfs_rq's estimated utilization */
4653 enqueued = cfs_rq->avg.util_est.enqueued;
4654 enqueued += _task_util_est(p);
4655 WRITE_ONCE(cfs_rq->avg.util_est.enqueued, enqueued);
4657 trace_sched_util_est_cfs_tp(cfs_rq);
4660 static inline void util_est_dequeue(struct cfs_rq *cfs_rq,
4661 struct task_struct *p)
4663 unsigned int enqueued;
4665 if (!sched_feat(UTIL_EST))
4668 /* Update root cfs_rq's estimated utilization */
4669 enqueued = cfs_rq->avg.util_est.enqueued;
4670 enqueued -= min_t(unsigned int, enqueued, _task_util_est(p));
4671 WRITE_ONCE(cfs_rq->avg.util_est.enqueued, enqueued);
4673 trace_sched_util_est_cfs_tp(cfs_rq);
4676 #define UTIL_EST_MARGIN (SCHED_CAPACITY_SCALE / 100)
4679 * Check if a (signed) value is within a specified (unsigned) margin,
4680 * based on the observation that:
4682 * abs(x) < y := (unsigned)(x + y - 1) < (2 * y - 1)
4684 * NOTE: this only works when value + margin < INT_MAX.
4686 static inline bool within_margin(int value, int margin)
4688 return ((unsigned int)(value + margin - 1) < (2 * margin - 1));
4691 static inline void util_est_update(struct cfs_rq *cfs_rq,
4692 struct task_struct *p,
4695 long last_ewma_diff, last_enqueued_diff;
4698 if (!sched_feat(UTIL_EST))
4702 * Skip update of task's estimated utilization when the task has not
4703 * yet completed an activation, e.g. being migrated.
4709 * If the PELT values haven't changed since enqueue time,
4710 * skip the util_est update.
4712 ue = p->se.avg.util_est;
4713 if (ue.enqueued & UTIL_AVG_UNCHANGED)
4716 last_enqueued_diff = ue.enqueued;
4719 * Reset EWMA on utilization increases, the moving average is used only
4720 * to smooth utilization decreases.
4722 ue.enqueued = task_util(p);
4723 if (sched_feat(UTIL_EST_FASTUP)) {
4724 if (ue.ewma < ue.enqueued) {
4725 ue.ewma = ue.enqueued;
4731 * Skip update of task's estimated utilization when its members are
4732 * already ~1% close to its last activation value.
4734 last_ewma_diff = ue.enqueued - ue.ewma;
4735 last_enqueued_diff -= ue.enqueued;
4736 if (within_margin(last_ewma_diff, UTIL_EST_MARGIN)) {
4737 if (!within_margin(last_enqueued_diff, UTIL_EST_MARGIN))
4744 * To avoid overestimation of actual task utilization, skip updates if
4745 * we cannot grant there is idle time in this CPU.
4747 if (task_util(p) > capacity_orig_of(cpu_of(rq_of(cfs_rq))))
4751 * Update Task's estimated utilization
4753 * When *p completes an activation we can consolidate another sample
4754 * of the task size. This is done by storing the current PELT value
4755 * as ue.enqueued and by using this value to update the Exponential
4756 * Weighted Moving Average (EWMA):
4758 * ewma(t) = w * task_util(p) + (1-w) * ewma(t-1)
4759 * = w * task_util(p) + ewma(t-1) - w * ewma(t-1)
4760 * = w * (task_util(p) - ewma(t-1)) + ewma(t-1)
4761 * = w * ( last_ewma_diff ) + ewma(t-1)
4762 * = w * (last_ewma_diff + ewma(t-1) / w)
4764 * Where 'w' is the weight of new samples, which is configured to be
4765 * 0.25, thus making w=1/4 ( >>= UTIL_EST_WEIGHT_SHIFT)
4767 ue.ewma <<= UTIL_EST_WEIGHT_SHIFT;
4768 ue.ewma += last_ewma_diff;
4769 ue.ewma >>= UTIL_EST_WEIGHT_SHIFT;
4771 ue.enqueued |= UTIL_AVG_UNCHANGED;
4772 WRITE_ONCE(p->se.avg.util_est, ue);
4774 trace_sched_util_est_se_tp(&p->se);
4777 static inline int util_fits_cpu(unsigned long util,
4778 unsigned long uclamp_min,
4779 unsigned long uclamp_max,
4782 unsigned long capacity_orig, capacity_orig_thermal;
4783 unsigned long capacity = capacity_of(cpu);
4784 bool fits, uclamp_max_fits;
4787 * Check if the real util fits without any uclamp boost/cap applied.
4789 fits = fits_capacity(util, capacity);
4791 if (!uclamp_is_used())
4795 * We must use capacity_orig_of() for comparing against uclamp_min and
4796 * uclamp_max. We only care about capacity pressure (by using
4797 * capacity_of()) for comparing against the real util.
4799 * If a task is boosted to 1024 for example, we don't want a tiny
4800 * pressure to skew the check whether it fits a CPU or not.
4802 * Similarly if a task is capped to capacity_orig_of(little_cpu), it
4803 * should fit a little cpu even if there's some pressure.
4805 * Only exception is for thermal pressure since it has a direct impact
4806 * on available OPP of the system.
4808 * We honour it for uclamp_min only as a drop in performance level
4809 * could result in not getting the requested minimum performance level.
4811 * For uclamp_max, we can tolerate a drop in performance level as the
4812 * goal is to cap the task. So it's okay if it's getting less.
4814 capacity_orig = capacity_orig_of(cpu);
4815 capacity_orig_thermal = capacity_orig - arch_scale_thermal_pressure(cpu);
4818 * We want to force a task to fit a cpu as implied by uclamp_max.
4819 * But we do have some corner cases to cater for..
4825 * |_ _ _ _ _ _ _ _ _ ___ _ _ _ | _ | _ _ _ _ _ uclamp_max
4828 * | | | | | | | (util somewhere in this region)
4831 * +----------------------------------------
4834 * In the above example if a task is capped to a specific performance
4835 * point, y, then when:
4837 * * util = 80% of x then it does not fit on cpu0 and should migrate
4839 * * util = 80% of y then it is forced to fit on cpu1 to honour
4840 * uclamp_max request.
4842 * which is what we're enforcing here. A task always fits if
4843 * uclamp_max <= capacity_orig. But when uclamp_max > capacity_orig,
4844 * the normal upmigration rules should withhold still.
4846 * Only exception is when we are on max capacity, then we need to be
4847 * careful not to block overutilized state. This is so because:
4849 * 1. There's no concept of capping at max_capacity! We can't go
4850 * beyond this performance level anyway.
4851 * 2. The system is being saturated when we're operating near
4852 * max capacity, it doesn't make sense to block overutilized.
4854 uclamp_max_fits = (capacity_orig == SCHED_CAPACITY_SCALE) && (uclamp_max == SCHED_CAPACITY_SCALE);
4855 uclamp_max_fits = !uclamp_max_fits && (uclamp_max <= capacity_orig);
4856 fits = fits || uclamp_max_fits;
4861 * | ___ (region a, capped, util >= uclamp_max)
4863 * |_ _ _ _ _ _ _ _ _ ___ _ _ _ | _ | _ _ _ _ _ uclamp_max
4865 * | ___ | | | | (region b, uclamp_min <= util <= uclamp_max)
4866 * |_ _ _|_ _|_ _ _ _| _ | _ _ _| _ | _ _ _ _ _ uclamp_min
4868 * | | | | | | | (region c, boosted, util < uclamp_min)
4869 * +----------------------------------------
4872 * a) If util > uclamp_max, then we're capped, we don't care about
4873 * actual fitness value here. We only care if uclamp_max fits
4874 * capacity without taking margin/pressure into account.
4875 * See comment above.
4877 * b) If uclamp_min <= util <= uclamp_max, then the normal
4878 * fits_capacity() rules apply. Except we need to ensure that we
4879 * enforce we remain within uclamp_max, see comment above.
4881 * c) If util < uclamp_min, then we are boosted. Same as (b) but we
4882 * need to take into account the boosted value fits the CPU without
4883 * taking margin/pressure into account.
4885 * Cases (a) and (b) are handled in the 'fits' variable already. We
4886 * just need to consider an extra check for case (c) after ensuring we
4887 * handle the case uclamp_min > uclamp_max.
4889 uclamp_min = min(uclamp_min, uclamp_max);
4890 if (fits && (util < uclamp_min) && (uclamp_min > capacity_orig_thermal))
4896 static inline int task_fits_cpu(struct task_struct *p, int cpu)
4898 unsigned long uclamp_min = uclamp_eff_value(p, UCLAMP_MIN);
4899 unsigned long uclamp_max = uclamp_eff_value(p, UCLAMP_MAX);
4900 unsigned long util = task_util_est(p);
4902 * Return true only if the cpu fully fits the task requirements, which
4903 * include the utilization but also the performance hints.
4905 return (util_fits_cpu(util, uclamp_min, uclamp_max, cpu) > 0);
4908 static inline void update_misfit_status(struct task_struct *p, struct rq *rq)
4910 if (!sched_asym_cpucap_active())
4913 if (!p || p->nr_cpus_allowed == 1) {
4914 rq->misfit_task_load = 0;
4918 if (task_fits_cpu(p, cpu_of(rq))) {
4919 rq->misfit_task_load = 0;
4924 * Make sure that misfit_task_load will not be null even if
4925 * task_h_load() returns 0.
4927 rq->misfit_task_load = max_t(unsigned long, task_h_load(p), 1);
4930 #else /* CONFIG_SMP */
4932 static inline bool cfs_rq_is_decayed(struct cfs_rq *cfs_rq)
4937 #define UPDATE_TG 0x0
4938 #define SKIP_AGE_LOAD 0x0
4939 #define DO_ATTACH 0x0
4940 #define DO_DETACH 0x0
4942 static inline void update_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int not_used1)
4944 cfs_rq_util_change(cfs_rq, 0);
4947 static inline void remove_entity_load_avg(struct sched_entity *se) {}
4950 attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
4952 detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
4954 static inline int newidle_balance(struct rq *rq, struct rq_flags *rf)
4960 util_est_enqueue(struct cfs_rq *cfs_rq, struct task_struct *p) {}
4963 util_est_dequeue(struct cfs_rq *cfs_rq, struct task_struct *p) {}
4966 util_est_update(struct cfs_rq *cfs_rq, struct task_struct *p,
4968 static inline void update_misfit_status(struct task_struct *p, struct rq *rq) {}
4970 #endif /* CONFIG_SMP */
4973 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
4975 u64 vslice, vruntime = avg_vruntime(cfs_rq);
4978 se->slice = sysctl_sched_base_slice;
4979 vslice = calc_delta_fair(se->slice, se);
4982 * Due to how V is constructed as the weighted average of entities,
4983 * adding tasks with positive lag, or removing tasks with negative lag
4984 * will move 'time' backwards, this can screw around with the lag of
4987 * EEVDF: placement strategy #1 / #2
4989 if (sched_feat(PLACE_LAG) && cfs_rq->nr_running) {
4990 struct sched_entity *curr = cfs_rq->curr;
4996 * If we want to place a task and preserve lag, we have to
4997 * consider the effect of the new entity on the weighted
4998 * average and compensate for this, otherwise lag can quickly
5001 * Lag is defined as:
5003 * lag_i = S - s_i = w_i * (V - v_i)
5005 * To avoid the 'w_i' term all over the place, we only track
5008 * vl_i = V - v_i <=> v_i = V - vl_i
5010 * And we take V to be the weighted average of all v:
5012 * V = (\Sum w_j*v_j) / W
5014 * Where W is: \Sum w_j
5016 * Then, the weighted average after adding an entity with lag
5019 * V' = (\Sum w_j*v_j + w_i*v_i) / (W + w_i)
5020 * = (W*V + w_i*(V - vl_i)) / (W + w_i)
5021 * = (W*V + w_i*V - w_i*vl_i) / (W + w_i)
5022 * = (V*(W + w_i) - w_i*l) / (W + w_i)
5023 * = V - w_i*vl_i / (W + w_i)
5025 * And the actual lag after adding an entity with vl_i is:
5028 * = V - w_i*vl_i / (W + w_i) - (V - vl_i)
5029 * = vl_i - w_i*vl_i / (W + w_i)
5031 * Which is strictly less than vl_i. So in order to preserve lag
5032 * we should inflate the lag before placement such that the
5033 * effective lag after placement comes out right.
5035 * As such, invert the above relation for vl'_i to get the vl_i
5036 * we need to use such that the lag after placement is the lag
5037 * we computed before dequeue.
5039 * vl'_i = vl_i - w_i*vl_i / (W + w_i)
5040 * = ((W + w_i)*vl_i - w_i*vl_i) / (W + w_i)
5042 * (W + w_i)*vl'_i = (W + w_i)*vl_i - w_i*vl_i
5045 * vl_i = (W + w_i)*vl'_i / W
5047 load = cfs_rq->avg_load;
5048 if (curr && curr->on_rq)
5049 load += scale_load_down(curr->load.weight);
5051 lag *= load + scale_load_down(se->load.weight);
5052 if (WARN_ON_ONCE(!load))
5054 lag = div_s64(lag, load);
5057 se->vruntime = vruntime - lag;
5060 * When joining the competition; the exisiting tasks will be,
5061 * on average, halfway through their slice, as such start tasks
5062 * off with half a slice to ease into the competition.
5064 if (sched_feat(PLACE_DEADLINE_INITIAL) && (flags & ENQUEUE_INITIAL))
5068 * EEVDF: vd_i = ve_i + r_i/w_i
5070 se->deadline = se->vruntime + vslice;
5073 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
5074 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq);
5076 static inline bool cfs_bandwidth_used(void);
5079 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
5081 bool curr = cfs_rq->curr == se;
5084 * If we're the current task, we must renormalise before calling
5088 place_entity(cfs_rq, se, flags);
5090 update_curr(cfs_rq);
5093 * When enqueuing a sched_entity, we must:
5094 * - Update loads to have both entity and cfs_rq synced with now.
5095 * - For group_entity, update its runnable_weight to reflect the new
5096 * h_nr_running of its group cfs_rq.
5097 * - For group_entity, update its weight to reflect the new share of
5099 * - Add its new weight to cfs_rq->load.weight
5101 update_load_avg(cfs_rq, se, UPDATE_TG | DO_ATTACH);
5102 se_update_runnable(se);
5104 * XXX update_load_avg() above will have attached us to the pelt sum;
5105 * but update_cfs_group() here will re-adjust the weight and have to
5106 * undo/redo all that. Seems wasteful.
5108 update_cfs_group(se);
5111 * XXX now that the entity has been re-weighted, and it's lag adjusted,
5112 * we can place the entity.
5115 place_entity(cfs_rq, se, flags);
5117 account_entity_enqueue(cfs_rq, se);
5119 /* Entity has migrated, no longer consider this task hot */
5120 if (flags & ENQUEUE_MIGRATED)
5123 check_schedstat_required();
5124 update_stats_enqueue_fair(cfs_rq, se, flags);
5126 __enqueue_entity(cfs_rq, se);
5129 if (cfs_rq->nr_running == 1) {
5130 check_enqueue_throttle(cfs_rq);
5131 if (!throttled_hierarchy(cfs_rq)) {
5132 list_add_leaf_cfs_rq(cfs_rq);
5134 #ifdef CONFIG_CFS_BANDWIDTH
5135 struct rq *rq = rq_of(cfs_rq);
5137 if (cfs_rq_throttled(cfs_rq) && !cfs_rq->throttled_clock)
5138 cfs_rq->throttled_clock = rq_clock(rq);
5139 if (!cfs_rq->throttled_clock_self)
5140 cfs_rq->throttled_clock_self = rq_clock(rq);
5146 static void __clear_buddies_next(struct sched_entity *se)
5148 for_each_sched_entity(se) {
5149 struct cfs_rq *cfs_rq = cfs_rq_of(se);
5150 if (cfs_rq->next != se)
5153 cfs_rq->next = NULL;
5157 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
5159 if (cfs_rq->next == se)
5160 __clear_buddies_next(se);
5163 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
5166 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
5168 int action = UPDATE_TG;
5170 if (entity_is_task(se) && task_on_rq_migrating(task_of(se)))
5171 action |= DO_DETACH;
5174 * Update run-time statistics of the 'current'.
5176 update_curr(cfs_rq);
5179 * When dequeuing a sched_entity, we must:
5180 * - Update loads to have both entity and cfs_rq synced with now.
5181 * - For group_entity, update its runnable_weight to reflect the new
5182 * h_nr_running of its group cfs_rq.
5183 * - Subtract its previous weight from cfs_rq->load.weight.
5184 * - For group entity, update its weight to reflect the new share
5185 * of its group cfs_rq.
5187 update_load_avg(cfs_rq, se, action);
5188 se_update_runnable(se);
5190 update_stats_dequeue_fair(cfs_rq, se, flags);
5192 clear_buddies(cfs_rq, se);
5194 update_entity_lag(cfs_rq, se);
5195 if (se != cfs_rq->curr)
5196 __dequeue_entity(cfs_rq, se);
5198 account_entity_dequeue(cfs_rq, se);
5200 /* return excess runtime on last dequeue */
5201 return_cfs_rq_runtime(cfs_rq);
5203 update_cfs_group(se);
5206 * Now advance min_vruntime if @se was the entity holding it back,
5207 * except when: DEQUEUE_SAVE && !DEQUEUE_MOVE, in this case we'll be
5208 * put back on, and if we advance min_vruntime, we'll be placed back
5209 * further than we started -- ie. we'll be penalized.
5211 if ((flags & (DEQUEUE_SAVE | DEQUEUE_MOVE)) != DEQUEUE_SAVE)
5212 update_min_vruntime(cfs_rq);
5214 if (cfs_rq->nr_running == 0)
5215 update_idle_cfs_rq_clock_pelt(cfs_rq);
5219 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
5221 clear_buddies(cfs_rq, se);
5223 /* 'current' is not kept within the tree. */
5226 * Any task has to be enqueued before it get to execute on
5227 * a CPU. So account for the time it spent waiting on the
5230 update_stats_wait_end_fair(cfs_rq, se);
5231 __dequeue_entity(cfs_rq, se);
5232 update_load_avg(cfs_rq, se, UPDATE_TG);
5234 * HACK, stash a copy of deadline at the point of pick in vlag,
5235 * which isn't used until dequeue.
5237 se->vlag = se->deadline;
5240 update_stats_curr_start(cfs_rq, se);
5244 * Track our maximum slice length, if the CPU's load is at
5245 * least twice that of our own weight (i.e. dont track it
5246 * when there are only lesser-weight tasks around):
5248 if (schedstat_enabled() &&
5249 rq_of(cfs_rq)->cfs.load.weight >= 2*se->load.weight) {
5250 struct sched_statistics *stats;
5252 stats = __schedstats_from_se(se);
5253 __schedstat_set(stats->slice_max,
5254 max((u64)stats->slice_max,
5255 se->sum_exec_runtime - se->prev_sum_exec_runtime));
5258 se->prev_sum_exec_runtime = se->sum_exec_runtime;
5262 * Pick the next process, keeping these things in mind, in this order:
5263 * 1) keep things fair between processes/task groups
5264 * 2) pick the "next" process, since someone really wants that to run
5265 * 3) pick the "last" process, for cache locality
5266 * 4) do not run the "skip" process, if something else is available
5268 static struct sched_entity *
5269 pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr)
5272 * Enabling NEXT_BUDDY will affect latency but not fairness.
5274 if (sched_feat(NEXT_BUDDY) &&
5275 cfs_rq->next && entity_eligible(cfs_rq, cfs_rq->next))
5276 return cfs_rq->next;
5278 return pick_eevdf(cfs_rq);
5281 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
5283 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
5286 * If still on the runqueue then deactivate_task()
5287 * was not called and update_curr() has to be done:
5290 update_curr(cfs_rq);
5292 /* throttle cfs_rqs exceeding runtime */
5293 check_cfs_rq_runtime(cfs_rq);
5296 update_stats_wait_start_fair(cfs_rq, prev);
5297 /* Put 'current' back into the tree. */
5298 __enqueue_entity(cfs_rq, prev);
5299 /* in !on_rq case, update occurred at dequeue */
5300 update_load_avg(cfs_rq, prev, 0);
5302 cfs_rq->curr = NULL;
5306 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
5309 * Update run-time statistics of the 'current'.
5311 update_curr(cfs_rq);
5314 * Ensure that runnable average is periodically updated.
5316 update_load_avg(cfs_rq, curr, UPDATE_TG);
5317 update_cfs_group(curr);
5319 #ifdef CONFIG_SCHED_HRTICK
5321 * queued ticks are scheduled to match the slice, so don't bother
5322 * validating it and just reschedule.
5325 resched_curr(rq_of(cfs_rq));
5329 * don't let the period tick interfere with the hrtick preemption
5331 if (!sched_feat(DOUBLE_TICK) &&
5332 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
5338 /**************************************************
5339 * CFS bandwidth control machinery
5342 #ifdef CONFIG_CFS_BANDWIDTH
5344 #ifdef CONFIG_JUMP_LABEL
5345 static struct static_key __cfs_bandwidth_used;
5347 static inline bool cfs_bandwidth_used(void)
5349 return static_key_false(&__cfs_bandwidth_used);
5352 void cfs_bandwidth_usage_inc(void)
5354 static_key_slow_inc_cpuslocked(&__cfs_bandwidth_used);
5357 void cfs_bandwidth_usage_dec(void)
5359 static_key_slow_dec_cpuslocked(&__cfs_bandwidth_used);
5361 #else /* CONFIG_JUMP_LABEL */
5362 static bool cfs_bandwidth_used(void)
5367 void cfs_bandwidth_usage_inc(void) {}
5368 void cfs_bandwidth_usage_dec(void) {}
5369 #endif /* CONFIG_JUMP_LABEL */
5372 * default period for cfs group bandwidth.
5373 * default: 0.1s, units: nanoseconds
5375 static inline u64 default_cfs_period(void)
5377 return 100000000ULL;
5380 static inline u64 sched_cfs_bandwidth_slice(void)
5382 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
5386 * Replenish runtime according to assigned quota. We use sched_clock_cpu
5387 * directly instead of rq->clock to avoid adding additional synchronization
5390 * requires cfs_b->lock
5392 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
5396 if (unlikely(cfs_b->quota == RUNTIME_INF))
5399 cfs_b->runtime += cfs_b->quota;
5400 runtime = cfs_b->runtime_snap - cfs_b->runtime;
5402 cfs_b->burst_time += runtime;
5406 cfs_b->runtime = min(cfs_b->runtime, cfs_b->quota + cfs_b->burst);
5407 cfs_b->runtime_snap = cfs_b->runtime;
5410 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
5412 return &tg->cfs_bandwidth;
5415 /* returns 0 on failure to allocate runtime */
5416 static int __assign_cfs_rq_runtime(struct cfs_bandwidth *cfs_b,
5417 struct cfs_rq *cfs_rq, u64 target_runtime)
5419 u64 min_amount, amount = 0;
5421 lockdep_assert_held(&cfs_b->lock);
5423 /* note: this is a positive sum as runtime_remaining <= 0 */
5424 min_amount = target_runtime - cfs_rq->runtime_remaining;
5426 if (cfs_b->quota == RUNTIME_INF)
5427 amount = min_amount;
5429 start_cfs_bandwidth(cfs_b);
5431 if (cfs_b->runtime > 0) {
5432 amount = min(cfs_b->runtime, min_amount);
5433 cfs_b->runtime -= amount;
5438 cfs_rq->runtime_remaining += amount;
5440 return cfs_rq->runtime_remaining > 0;
5443 /* returns 0 on failure to allocate runtime */
5444 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
5446 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
5449 raw_spin_lock(&cfs_b->lock);
5450 ret = __assign_cfs_rq_runtime(cfs_b, cfs_rq, sched_cfs_bandwidth_slice());
5451 raw_spin_unlock(&cfs_b->lock);
5456 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
5458 /* dock delta_exec before expiring quota (as it could span periods) */
5459 cfs_rq->runtime_remaining -= delta_exec;
5461 if (likely(cfs_rq->runtime_remaining > 0))
5464 if (cfs_rq->throttled)
5467 * if we're unable to extend our runtime we resched so that the active
5468 * hierarchy can be throttled
5470 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
5471 resched_curr(rq_of(cfs_rq));
5474 static __always_inline
5475 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
5477 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
5480 __account_cfs_rq_runtime(cfs_rq, delta_exec);
5483 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
5485 return cfs_bandwidth_used() && cfs_rq->throttled;
5488 /* check whether cfs_rq, or any parent, is throttled */
5489 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
5491 return cfs_bandwidth_used() && cfs_rq->throttle_count;
5495 * Ensure that neither of the group entities corresponding to src_cpu or
5496 * dest_cpu are members of a throttled hierarchy when performing group
5497 * load-balance operations.
5499 static inline int throttled_lb_pair(struct task_group *tg,
5500 int src_cpu, int dest_cpu)
5502 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
5504 src_cfs_rq = tg->cfs_rq[src_cpu];
5505 dest_cfs_rq = tg->cfs_rq[dest_cpu];
5507 return throttled_hierarchy(src_cfs_rq) ||
5508 throttled_hierarchy(dest_cfs_rq);
5511 static int tg_unthrottle_up(struct task_group *tg, void *data)
5513 struct rq *rq = data;
5514 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
5516 cfs_rq->throttle_count--;
5517 if (!cfs_rq->throttle_count) {
5518 cfs_rq->throttled_clock_pelt_time += rq_clock_pelt(rq) -
5519 cfs_rq->throttled_clock_pelt;
5521 /* Add cfs_rq with load or one or more already running entities to the list */
5522 if (!cfs_rq_is_decayed(cfs_rq))
5523 list_add_leaf_cfs_rq(cfs_rq);
5525 if (cfs_rq->throttled_clock_self) {
5526 u64 delta = rq_clock(rq) - cfs_rq->throttled_clock_self;
5528 cfs_rq->throttled_clock_self = 0;
5530 if (SCHED_WARN_ON((s64)delta < 0))
5533 cfs_rq->throttled_clock_self_time += delta;
5540 static int tg_throttle_down(struct task_group *tg, void *data)
5542 struct rq *rq = data;
5543 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
5545 /* group is entering throttled state, stop time */
5546 if (!cfs_rq->throttle_count) {
5547 cfs_rq->throttled_clock_pelt = rq_clock_pelt(rq);
5548 list_del_leaf_cfs_rq(cfs_rq);
5550 SCHED_WARN_ON(cfs_rq->throttled_clock_self);
5551 if (cfs_rq->nr_running)
5552 cfs_rq->throttled_clock_self = rq_clock(rq);
5554 cfs_rq->throttle_count++;
5559 static bool throttle_cfs_rq(struct cfs_rq *cfs_rq)
5561 struct rq *rq = rq_of(cfs_rq);
5562 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
5563 struct sched_entity *se;
5564 long task_delta, idle_task_delta, dequeue = 1;
5566 raw_spin_lock(&cfs_b->lock);
5567 /* This will start the period timer if necessary */
5568 if (__assign_cfs_rq_runtime(cfs_b, cfs_rq, 1)) {
5570 * We have raced with bandwidth becoming available, and if we
5571 * actually throttled the timer might not unthrottle us for an
5572 * entire period. We additionally needed to make sure that any
5573 * subsequent check_cfs_rq_runtime calls agree not to throttle
5574 * us, as we may commit to do cfs put_prev+pick_next, so we ask
5575 * for 1ns of runtime rather than just check cfs_b.
5579 list_add_tail_rcu(&cfs_rq->throttled_list,
5580 &cfs_b->throttled_cfs_rq);
5582 raw_spin_unlock(&cfs_b->lock);
5585 return false; /* Throttle no longer required. */
5587 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
5589 /* freeze hierarchy runnable averages while throttled */
5591 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
5594 task_delta = cfs_rq->h_nr_running;
5595 idle_task_delta = cfs_rq->idle_h_nr_running;
5596 for_each_sched_entity(se) {
5597 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
5598 /* throttled entity or throttle-on-deactivate */
5602 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
5604 if (cfs_rq_is_idle(group_cfs_rq(se)))
5605 idle_task_delta = cfs_rq->h_nr_running;
5607 qcfs_rq->h_nr_running -= task_delta;
5608 qcfs_rq->idle_h_nr_running -= idle_task_delta;
5610 if (qcfs_rq->load.weight) {
5611 /* Avoid re-evaluating load for this entity: */
5612 se = parent_entity(se);
5617 for_each_sched_entity(se) {
5618 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
5619 /* throttled entity or throttle-on-deactivate */
5623 update_load_avg(qcfs_rq, se, 0);
5624 se_update_runnable(se);
5626 if (cfs_rq_is_idle(group_cfs_rq(se)))
5627 idle_task_delta = cfs_rq->h_nr_running;
5629 qcfs_rq->h_nr_running -= task_delta;
5630 qcfs_rq->idle_h_nr_running -= idle_task_delta;
5633 /* At this point se is NULL and we are at root level*/
5634 sub_nr_running(rq, task_delta);
5638 * Note: distribution will already see us throttled via the
5639 * throttled-list. rq->lock protects completion.
5641 cfs_rq->throttled = 1;
5642 SCHED_WARN_ON(cfs_rq->throttled_clock);
5643 if (cfs_rq->nr_running)
5644 cfs_rq->throttled_clock = rq_clock(rq);
5648 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
5650 struct rq *rq = rq_of(cfs_rq);
5651 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
5652 struct sched_entity *se;
5653 long task_delta, idle_task_delta;
5655 se = cfs_rq->tg->se[cpu_of(rq)];
5657 cfs_rq->throttled = 0;
5659 update_rq_clock(rq);
5661 raw_spin_lock(&cfs_b->lock);
5662 if (cfs_rq->throttled_clock) {
5663 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
5664 cfs_rq->throttled_clock = 0;
5666 list_del_rcu(&cfs_rq->throttled_list);
5667 raw_spin_unlock(&cfs_b->lock);
5669 /* update hierarchical throttle state */
5670 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
5672 if (!cfs_rq->load.weight) {
5673 if (!cfs_rq->on_list)
5676 * Nothing to run but something to decay (on_list)?
5677 * Complete the branch.
5679 for_each_sched_entity(se) {
5680 if (list_add_leaf_cfs_rq(cfs_rq_of(se)))
5683 goto unthrottle_throttle;
5686 task_delta = cfs_rq->h_nr_running;
5687 idle_task_delta = cfs_rq->idle_h_nr_running;
5688 for_each_sched_entity(se) {
5689 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
5693 enqueue_entity(qcfs_rq, se, ENQUEUE_WAKEUP);
5695 if (cfs_rq_is_idle(group_cfs_rq(se)))
5696 idle_task_delta = cfs_rq->h_nr_running;
5698 qcfs_rq->h_nr_running += task_delta;
5699 qcfs_rq->idle_h_nr_running += idle_task_delta;
5701 /* end evaluation on encountering a throttled cfs_rq */
5702 if (cfs_rq_throttled(qcfs_rq))
5703 goto unthrottle_throttle;
5706 for_each_sched_entity(se) {
5707 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
5709 update_load_avg(qcfs_rq, se, UPDATE_TG);
5710 se_update_runnable(se);
5712 if (cfs_rq_is_idle(group_cfs_rq(se)))
5713 idle_task_delta = cfs_rq->h_nr_running;
5715 qcfs_rq->h_nr_running += task_delta;
5716 qcfs_rq->idle_h_nr_running += idle_task_delta;
5718 /* end evaluation on encountering a throttled cfs_rq */
5719 if (cfs_rq_throttled(qcfs_rq))
5720 goto unthrottle_throttle;
5723 /* At this point se is NULL and we are at root level*/
5724 add_nr_running(rq, task_delta);
5726 unthrottle_throttle:
5727 assert_list_leaf_cfs_rq(rq);
5729 /* Determine whether we need to wake up potentially idle CPU: */
5730 if (rq->curr == rq->idle && rq->cfs.nr_running)
5735 static void __cfsb_csd_unthrottle(void *arg)
5737 struct cfs_rq *cursor, *tmp;
5738 struct rq *rq = arg;
5744 * Iterating over the list can trigger several call to
5745 * update_rq_clock() in unthrottle_cfs_rq().
5746 * Do it once and skip the potential next ones.
5748 update_rq_clock(rq);
5749 rq_clock_start_loop_update(rq);
5752 * Since we hold rq lock we're safe from concurrent manipulation of
5753 * the CSD list. However, this RCU critical section annotates the
5754 * fact that we pair with sched_free_group_rcu(), so that we cannot
5755 * race with group being freed in the window between removing it
5756 * from the list and advancing to the next entry in the list.
5760 list_for_each_entry_safe(cursor, tmp, &rq->cfsb_csd_list,
5761 throttled_csd_list) {
5762 list_del_init(&cursor->throttled_csd_list);
5764 if (cfs_rq_throttled(cursor))
5765 unthrottle_cfs_rq(cursor);
5770 rq_clock_stop_loop_update(rq);
5774 static inline void __unthrottle_cfs_rq_async(struct cfs_rq *cfs_rq)
5776 struct rq *rq = rq_of(cfs_rq);
5779 if (rq == this_rq()) {
5780 unthrottle_cfs_rq(cfs_rq);
5784 /* Already enqueued */
5785 if (SCHED_WARN_ON(!list_empty(&cfs_rq->throttled_csd_list)))
5788 first = list_empty(&rq->cfsb_csd_list);
5789 list_add_tail(&cfs_rq->throttled_csd_list, &rq->cfsb_csd_list);
5791 smp_call_function_single_async(cpu_of(rq), &rq->cfsb_csd);
5794 static inline void __unthrottle_cfs_rq_async(struct cfs_rq *cfs_rq)
5796 unthrottle_cfs_rq(cfs_rq);
5800 static void unthrottle_cfs_rq_async(struct cfs_rq *cfs_rq)
5802 lockdep_assert_rq_held(rq_of(cfs_rq));
5804 if (SCHED_WARN_ON(!cfs_rq_throttled(cfs_rq) ||
5805 cfs_rq->runtime_remaining <= 0))
5808 __unthrottle_cfs_rq_async(cfs_rq);
5811 static bool distribute_cfs_runtime(struct cfs_bandwidth *cfs_b)
5813 struct cfs_rq *local_unthrottle = NULL;
5814 int this_cpu = smp_processor_id();
5815 u64 runtime, remaining = 1;
5816 bool throttled = false;
5817 struct cfs_rq *cfs_rq;
5822 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
5831 rq_lock_irqsave(rq, &rf);
5832 if (!cfs_rq_throttled(cfs_rq))
5836 /* Already queued for async unthrottle */
5837 if (!list_empty(&cfs_rq->throttled_csd_list))
5841 /* By the above checks, this should never be true */
5842 SCHED_WARN_ON(cfs_rq->runtime_remaining > 0);
5844 raw_spin_lock(&cfs_b->lock);
5845 runtime = -cfs_rq->runtime_remaining + 1;
5846 if (runtime > cfs_b->runtime)
5847 runtime = cfs_b->runtime;
5848 cfs_b->runtime -= runtime;
5849 remaining = cfs_b->runtime;
5850 raw_spin_unlock(&cfs_b->lock);
5852 cfs_rq->runtime_remaining += runtime;
5854 /* we check whether we're throttled above */
5855 if (cfs_rq->runtime_remaining > 0) {
5856 if (cpu_of(rq) != this_cpu ||
5857 SCHED_WARN_ON(local_unthrottle))
5858 unthrottle_cfs_rq_async(cfs_rq);
5860 local_unthrottle = cfs_rq;
5866 rq_unlock_irqrestore(rq, &rf);
5870 if (local_unthrottle) {
5871 rq = cpu_rq(this_cpu);
5872 rq_lock_irqsave(rq, &rf);
5873 if (cfs_rq_throttled(local_unthrottle))
5874 unthrottle_cfs_rq(local_unthrottle);
5875 rq_unlock_irqrestore(rq, &rf);
5882 * Responsible for refilling a task_group's bandwidth and unthrottling its
5883 * cfs_rqs as appropriate. If there has been no activity within the last
5884 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
5885 * used to track this state.
5887 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun, unsigned long flags)
5891 /* no need to continue the timer with no bandwidth constraint */
5892 if (cfs_b->quota == RUNTIME_INF)
5893 goto out_deactivate;
5895 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
5896 cfs_b->nr_periods += overrun;
5898 /* Refill extra burst quota even if cfs_b->idle */
5899 __refill_cfs_bandwidth_runtime(cfs_b);
5902 * idle depends on !throttled (for the case of a large deficit), and if
5903 * we're going inactive then everything else can be deferred
5905 if (cfs_b->idle && !throttled)
5906 goto out_deactivate;
5909 /* mark as potentially idle for the upcoming period */
5914 /* account preceding periods in which throttling occurred */
5915 cfs_b->nr_throttled += overrun;
5918 * This check is repeated as we release cfs_b->lock while we unthrottle.
5920 while (throttled && cfs_b->runtime > 0) {
5921 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
5922 /* we can't nest cfs_b->lock while distributing bandwidth */
5923 throttled = distribute_cfs_runtime(cfs_b);
5924 raw_spin_lock_irqsave(&cfs_b->lock, flags);
5928 * While we are ensured activity in the period following an
5929 * unthrottle, this also covers the case in which the new bandwidth is
5930 * insufficient to cover the existing bandwidth deficit. (Forcing the
5931 * timer to remain active while there are any throttled entities.)
5941 /* a cfs_rq won't donate quota below this amount */
5942 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
5943 /* minimum remaining period time to redistribute slack quota */
5944 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
5945 /* how long we wait to gather additional slack before distributing */
5946 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
5949 * Are we near the end of the current quota period?
5951 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
5952 * hrtimer base being cleared by hrtimer_start. In the case of
5953 * migrate_hrtimers, base is never cleared, so we are fine.
5955 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
5957 struct hrtimer *refresh_timer = &cfs_b->period_timer;
5960 /* if the call-back is running a quota refresh is already occurring */
5961 if (hrtimer_callback_running(refresh_timer))
5964 /* is a quota refresh about to occur? */
5965 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
5966 if (remaining < (s64)min_expire)
5972 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
5974 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
5976 /* if there's a quota refresh soon don't bother with slack */
5977 if (runtime_refresh_within(cfs_b, min_left))
5980 /* don't push forwards an existing deferred unthrottle */
5981 if (cfs_b->slack_started)
5983 cfs_b->slack_started = true;
5985 hrtimer_start(&cfs_b->slack_timer,
5986 ns_to_ktime(cfs_bandwidth_slack_period),
5990 /* we know any runtime found here is valid as update_curr() precedes return */
5991 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
5993 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
5994 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
5996 if (slack_runtime <= 0)
5999 raw_spin_lock(&cfs_b->lock);
6000 if (cfs_b->quota != RUNTIME_INF) {
6001 cfs_b->runtime += slack_runtime;
6003 /* we are under rq->lock, defer unthrottling using a timer */
6004 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
6005 !list_empty(&cfs_b->throttled_cfs_rq))
6006 start_cfs_slack_bandwidth(cfs_b);
6008 raw_spin_unlock(&cfs_b->lock);
6010 /* even if it's not valid for return we don't want to try again */
6011 cfs_rq->runtime_remaining -= slack_runtime;
6014 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
6016 if (!cfs_bandwidth_used())
6019 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
6022 __return_cfs_rq_runtime(cfs_rq);
6026 * This is done with a timer (instead of inline with bandwidth return) since
6027 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
6029 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
6031 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
6032 unsigned long flags;
6034 /* confirm we're still not at a refresh boundary */
6035 raw_spin_lock_irqsave(&cfs_b->lock, flags);
6036 cfs_b->slack_started = false;
6038 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
6039 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
6043 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice)
6044 runtime = cfs_b->runtime;
6046 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
6051 distribute_cfs_runtime(cfs_b);
6055 * When a group wakes up we want to make sure that its quota is not already
6056 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
6057 * runtime as update_curr() throttling can not trigger until it's on-rq.
6059 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
6061 if (!cfs_bandwidth_used())
6064 /* an active group must be handled by the update_curr()->put() path */
6065 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
6068 /* ensure the group is not already throttled */
6069 if (cfs_rq_throttled(cfs_rq))
6072 /* update runtime allocation */
6073 account_cfs_rq_runtime(cfs_rq, 0);
6074 if (cfs_rq->runtime_remaining <= 0)
6075 throttle_cfs_rq(cfs_rq);
6078 static void sync_throttle(struct task_group *tg, int cpu)
6080 struct cfs_rq *pcfs_rq, *cfs_rq;
6082 if (!cfs_bandwidth_used())
6088 cfs_rq = tg->cfs_rq[cpu];
6089 pcfs_rq = tg->parent->cfs_rq[cpu];
6091 cfs_rq->throttle_count = pcfs_rq->throttle_count;
6092 cfs_rq->throttled_clock_pelt = rq_clock_pelt(cpu_rq(cpu));
6095 /* conditionally throttle active cfs_rq's from put_prev_entity() */
6096 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
6098 if (!cfs_bandwidth_used())
6101 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
6105 * it's possible for a throttled entity to be forced into a running
6106 * state (e.g. set_curr_task), in this case we're finished.
6108 if (cfs_rq_throttled(cfs_rq))
6111 return throttle_cfs_rq(cfs_rq);
6114 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
6116 struct cfs_bandwidth *cfs_b =
6117 container_of(timer, struct cfs_bandwidth, slack_timer);
6119 do_sched_cfs_slack_timer(cfs_b);
6121 return HRTIMER_NORESTART;
6124 extern const u64 max_cfs_quota_period;
6126 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
6128 struct cfs_bandwidth *cfs_b =
6129 container_of(timer, struct cfs_bandwidth, period_timer);
6130 unsigned long flags;
6135 raw_spin_lock_irqsave(&cfs_b->lock, flags);
6137 overrun = hrtimer_forward_now(timer, cfs_b->period);
6141 idle = do_sched_cfs_period_timer(cfs_b, overrun, flags);
6144 u64 new, old = ktime_to_ns(cfs_b->period);
6147 * Grow period by a factor of 2 to avoid losing precision.
6148 * Precision loss in the quota/period ratio can cause __cfs_schedulable
6152 if (new < max_cfs_quota_period) {
6153 cfs_b->period = ns_to_ktime(new);
6157 pr_warn_ratelimited(
6158 "cfs_period_timer[cpu%d]: period too short, scaling up (new cfs_period_us = %lld, cfs_quota_us = %lld)\n",
6160 div_u64(new, NSEC_PER_USEC),
6161 div_u64(cfs_b->quota, NSEC_PER_USEC));
6163 pr_warn_ratelimited(
6164 "cfs_period_timer[cpu%d]: period too short, but cannot scale up without losing precision (cfs_period_us = %lld, cfs_quota_us = %lld)\n",
6166 div_u64(old, NSEC_PER_USEC),
6167 div_u64(cfs_b->quota, NSEC_PER_USEC));
6170 /* reset count so we don't come right back in here */
6175 cfs_b->period_active = 0;
6176 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
6178 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
6181 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b, struct cfs_bandwidth *parent)
6183 raw_spin_lock_init(&cfs_b->lock);
6185 cfs_b->quota = RUNTIME_INF;
6186 cfs_b->period = ns_to_ktime(default_cfs_period());
6188 cfs_b->hierarchical_quota = parent ? parent->hierarchical_quota : RUNTIME_INF;
6190 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
6191 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED);
6192 cfs_b->period_timer.function = sched_cfs_period_timer;
6194 /* Add a random offset so that timers interleave */
6195 hrtimer_set_expires(&cfs_b->period_timer,
6196 get_random_u32_below(cfs_b->period));
6197 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
6198 cfs_b->slack_timer.function = sched_cfs_slack_timer;
6199 cfs_b->slack_started = false;
6202 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
6204 cfs_rq->runtime_enabled = 0;
6205 INIT_LIST_HEAD(&cfs_rq->throttled_list);
6207 INIT_LIST_HEAD(&cfs_rq->throttled_csd_list);
6211 void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
6213 lockdep_assert_held(&cfs_b->lock);
6215 if (cfs_b->period_active)
6218 cfs_b->period_active = 1;
6219 hrtimer_forward_now(&cfs_b->period_timer, cfs_b->period);
6220 hrtimer_start_expires(&cfs_b->period_timer, HRTIMER_MODE_ABS_PINNED);
6223 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
6225 int __maybe_unused i;
6227 /* init_cfs_bandwidth() was not called */
6228 if (!cfs_b->throttled_cfs_rq.next)
6231 hrtimer_cancel(&cfs_b->period_timer);
6232 hrtimer_cancel(&cfs_b->slack_timer);
6235 * It is possible that we still have some cfs_rq's pending on a CSD
6236 * list, though this race is very rare. In order for this to occur, we
6237 * must have raced with the last task leaving the group while there
6238 * exist throttled cfs_rq(s), and the period_timer must have queued the
6239 * CSD item but the remote cpu has not yet processed it. To handle this,
6240 * we can simply flush all pending CSD work inline here. We're
6241 * guaranteed at this point that no additional cfs_rq of this group can
6245 for_each_possible_cpu(i) {
6246 struct rq *rq = cpu_rq(i);
6247 unsigned long flags;
6249 if (list_empty(&rq->cfsb_csd_list))
6252 local_irq_save(flags);
6253 __cfsb_csd_unthrottle(rq);
6254 local_irq_restore(flags);
6260 * Both these CPU hotplug callbacks race against unregister_fair_sched_group()
6262 * The race is harmless, since modifying bandwidth settings of unhooked group
6263 * bits doesn't do much.
6266 /* cpu online callback */
6267 static void __maybe_unused update_runtime_enabled(struct rq *rq)
6269 struct task_group *tg;
6271 lockdep_assert_rq_held(rq);
6274 list_for_each_entry_rcu(tg, &task_groups, list) {
6275 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
6276 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
6278 raw_spin_lock(&cfs_b->lock);
6279 cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF;
6280 raw_spin_unlock(&cfs_b->lock);
6285 /* cpu offline callback */
6286 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
6288 struct task_group *tg;
6290 lockdep_assert_rq_held(rq);
6293 * The rq clock has already been updated in the
6294 * set_rq_offline(), so we should skip updating
6295 * the rq clock again in unthrottle_cfs_rq().
6297 rq_clock_start_loop_update(rq);
6300 list_for_each_entry_rcu(tg, &task_groups, list) {
6301 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
6303 if (!cfs_rq->runtime_enabled)
6307 * clock_task is not advancing so we just need to make sure
6308 * there's some valid quota amount
6310 cfs_rq->runtime_remaining = 1;
6312 * Offline rq is schedulable till CPU is completely disabled
6313 * in take_cpu_down(), so we prevent new cfs throttling here.
6315 cfs_rq->runtime_enabled = 0;
6317 if (cfs_rq_throttled(cfs_rq))
6318 unthrottle_cfs_rq(cfs_rq);
6322 rq_clock_stop_loop_update(rq);
6325 bool cfs_task_bw_constrained(struct task_struct *p)
6327 struct cfs_rq *cfs_rq = task_cfs_rq(p);
6329 if (!cfs_bandwidth_used())
6332 if (cfs_rq->runtime_enabled ||
6333 tg_cfs_bandwidth(cfs_rq->tg)->hierarchical_quota != RUNTIME_INF)
6339 #ifdef CONFIG_NO_HZ_FULL
6340 /* called from pick_next_task_fair() */
6341 static void sched_fair_update_stop_tick(struct rq *rq, struct task_struct *p)
6343 int cpu = cpu_of(rq);
6345 if (!sched_feat(HZ_BW) || !cfs_bandwidth_used())
6348 if (!tick_nohz_full_cpu(cpu))
6351 if (rq->nr_running != 1)
6355 * We know there is only one task runnable and we've just picked it. The
6356 * normal enqueue path will have cleared TICK_DEP_BIT_SCHED if we will
6357 * be otherwise able to stop the tick. Just need to check if we are using
6358 * bandwidth control.
6360 if (cfs_task_bw_constrained(p))
6361 tick_nohz_dep_set_cpu(cpu, TICK_DEP_BIT_SCHED);
6365 #else /* CONFIG_CFS_BANDWIDTH */
6367 static inline bool cfs_bandwidth_used(void)
6372 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {}
6373 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; }
6374 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
6375 static inline void sync_throttle(struct task_group *tg, int cpu) {}
6376 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
6378 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
6383 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
6388 static inline int throttled_lb_pair(struct task_group *tg,
6389 int src_cpu, int dest_cpu)
6394 #ifdef CONFIG_FAIR_GROUP_SCHED
6395 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b, struct cfs_bandwidth *parent) {}
6396 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
6399 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
6403 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
6404 static inline void update_runtime_enabled(struct rq *rq) {}
6405 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
6406 #ifdef CONFIG_CGROUP_SCHED
6407 bool cfs_task_bw_constrained(struct task_struct *p)
6412 #endif /* CONFIG_CFS_BANDWIDTH */
6414 #if !defined(CONFIG_CFS_BANDWIDTH) || !defined(CONFIG_NO_HZ_FULL)
6415 static inline void sched_fair_update_stop_tick(struct rq *rq, struct task_struct *p) {}
6418 /**************************************************
6419 * CFS operations on tasks:
6422 #ifdef CONFIG_SCHED_HRTICK
6423 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
6425 struct sched_entity *se = &p->se;
6427 SCHED_WARN_ON(task_rq(p) != rq);
6429 if (rq->cfs.h_nr_running > 1) {
6430 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
6431 u64 slice = se->slice;
6432 s64 delta = slice - ran;
6435 if (task_current(rq, p))
6439 hrtick_start(rq, delta);
6444 * called from enqueue/dequeue and updates the hrtick when the
6445 * current task is from our class and nr_running is low enough
6448 static void hrtick_update(struct rq *rq)
6450 struct task_struct *curr = rq->curr;
6452 if (!hrtick_enabled_fair(rq) || curr->sched_class != &fair_sched_class)
6455 hrtick_start_fair(rq, curr);
6457 #else /* !CONFIG_SCHED_HRTICK */
6459 hrtick_start_fair(struct rq *rq, struct task_struct *p)
6463 static inline void hrtick_update(struct rq *rq)
6469 static inline bool cpu_overutilized(int cpu)
6471 unsigned long rq_util_min = uclamp_rq_get(cpu_rq(cpu), UCLAMP_MIN);
6472 unsigned long rq_util_max = uclamp_rq_get(cpu_rq(cpu), UCLAMP_MAX);
6474 /* Return true only if the utilization doesn't fit CPU's capacity */
6475 return !util_fits_cpu(cpu_util_cfs(cpu), rq_util_min, rq_util_max, cpu);
6478 static inline void update_overutilized_status(struct rq *rq)
6480 if (!READ_ONCE(rq->rd->overutilized) && cpu_overutilized(rq->cpu)) {
6481 WRITE_ONCE(rq->rd->overutilized, SG_OVERUTILIZED);
6482 trace_sched_overutilized_tp(rq->rd, SG_OVERUTILIZED);
6486 static inline void update_overutilized_status(struct rq *rq) { }
6489 /* Runqueue only has SCHED_IDLE tasks enqueued */
6490 static int sched_idle_rq(struct rq *rq)
6492 return unlikely(rq->nr_running == rq->cfs.idle_h_nr_running &&
6497 static int sched_idle_cpu(int cpu)
6499 return sched_idle_rq(cpu_rq(cpu));
6504 * The enqueue_task method is called before nr_running is
6505 * increased. Here we update the fair scheduling stats and
6506 * then put the task into the rbtree:
6509 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
6511 struct cfs_rq *cfs_rq;
6512 struct sched_entity *se = &p->se;
6513 int idle_h_nr_running = task_has_idle_policy(p);
6514 int task_new = !(flags & ENQUEUE_WAKEUP);
6517 * The code below (indirectly) updates schedutil which looks at
6518 * the cfs_rq utilization to select a frequency.
6519 * Let's add the task's estimated utilization to the cfs_rq's
6520 * estimated utilization, before we update schedutil.
6522 util_est_enqueue(&rq->cfs, p);
6525 * If in_iowait is set, the code below may not trigger any cpufreq
6526 * utilization updates, so do it here explicitly with the IOWAIT flag
6530 cpufreq_update_util(rq, SCHED_CPUFREQ_IOWAIT);
6532 for_each_sched_entity(se) {
6535 cfs_rq = cfs_rq_of(se);
6536 enqueue_entity(cfs_rq, se, flags);
6538 cfs_rq->h_nr_running++;
6539 cfs_rq->idle_h_nr_running += idle_h_nr_running;
6541 if (cfs_rq_is_idle(cfs_rq))
6542 idle_h_nr_running = 1;
6544 /* end evaluation on encountering a throttled cfs_rq */
6545 if (cfs_rq_throttled(cfs_rq))
6546 goto enqueue_throttle;
6548 flags = ENQUEUE_WAKEUP;
6551 for_each_sched_entity(se) {
6552 cfs_rq = cfs_rq_of(se);
6554 update_load_avg(cfs_rq, se, UPDATE_TG);
6555 se_update_runnable(se);
6556 update_cfs_group(se);
6558 cfs_rq->h_nr_running++;
6559 cfs_rq->idle_h_nr_running += idle_h_nr_running;
6561 if (cfs_rq_is_idle(cfs_rq))
6562 idle_h_nr_running = 1;
6564 /* end evaluation on encountering a throttled cfs_rq */
6565 if (cfs_rq_throttled(cfs_rq))
6566 goto enqueue_throttle;
6569 /* At this point se is NULL and we are at root level*/
6570 add_nr_running(rq, 1);
6573 * Since new tasks are assigned an initial util_avg equal to
6574 * half of the spare capacity of their CPU, tiny tasks have the
6575 * ability to cross the overutilized threshold, which will
6576 * result in the load balancer ruining all the task placement
6577 * done by EAS. As a way to mitigate that effect, do not account
6578 * for the first enqueue operation of new tasks during the
6579 * overutilized flag detection.
6581 * A better way of solving this problem would be to wait for
6582 * the PELT signals of tasks to converge before taking them
6583 * into account, but that is not straightforward to implement,
6584 * and the following generally works well enough in practice.
6587 update_overutilized_status(rq);
6590 assert_list_leaf_cfs_rq(rq);
6595 static void set_next_buddy(struct sched_entity *se);
6598 * The dequeue_task method is called before nr_running is
6599 * decreased. We remove the task from the rbtree and
6600 * update the fair scheduling stats:
6602 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
6604 struct cfs_rq *cfs_rq;
6605 struct sched_entity *se = &p->se;
6606 int task_sleep = flags & DEQUEUE_SLEEP;
6607 int idle_h_nr_running = task_has_idle_policy(p);
6608 bool was_sched_idle = sched_idle_rq(rq);
6610 util_est_dequeue(&rq->cfs, p);
6612 for_each_sched_entity(se) {
6613 cfs_rq = cfs_rq_of(se);
6614 dequeue_entity(cfs_rq, se, flags);
6616 cfs_rq->h_nr_running--;
6617 cfs_rq->idle_h_nr_running -= idle_h_nr_running;
6619 if (cfs_rq_is_idle(cfs_rq))
6620 idle_h_nr_running = 1;
6622 /* end evaluation on encountering a throttled cfs_rq */
6623 if (cfs_rq_throttled(cfs_rq))
6624 goto dequeue_throttle;
6626 /* Don't dequeue parent if it has other entities besides us */
6627 if (cfs_rq->load.weight) {
6628 /* Avoid re-evaluating load for this entity: */
6629 se = parent_entity(se);
6631 * Bias pick_next to pick a task from this cfs_rq, as
6632 * p is sleeping when it is within its sched_slice.
6634 if (task_sleep && se && !throttled_hierarchy(cfs_rq))
6638 flags |= DEQUEUE_SLEEP;
6641 for_each_sched_entity(se) {
6642 cfs_rq = cfs_rq_of(se);
6644 update_load_avg(cfs_rq, se, UPDATE_TG);
6645 se_update_runnable(se);
6646 update_cfs_group(se);
6648 cfs_rq->h_nr_running--;
6649 cfs_rq->idle_h_nr_running -= idle_h_nr_running;
6651 if (cfs_rq_is_idle(cfs_rq))
6652 idle_h_nr_running = 1;
6654 /* end evaluation on encountering a throttled cfs_rq */
6655 if (cfs_rq_throttled(cfs_rq))
6656 goto dequeue_throttle;
6660 /* At this point se is NULL and we are at root level*/
6661 sub_nr_running(rq, 1);
6663 /* balance early to pull high priority tasks */
6664 if (unlikely(!was_sched_idle && sched_idle_rq(rq)))
6665 rq->next_balance = jiffies;
6668 util_est_update(&rq->cfs, p, task_sleep);
6674 /* Working cpumask for: load_balance, load_balance_newidle. */
6675 static DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
6676 static DEFINE_PER_CPU(cpumask_var_t, select_rq_mask);
6677 static DEFINE_PER_CPU(cpumask_var_t, should_we_balance_tmpmask);
6679 #ifdef CONFIG_NO_HZ_COMMON
6682 cpumask_var_t idle_cpus_mask;
6684 int has_blocked; /* Idle CPUS has blocked load */
6685 int needs_update; /* Newly idle CPUs need their next_balance collated */
6686 unsigned long next_balance; /* in jiffy units */
6687 unsigned long next_blocked; /* Next update of blocked load in jiffies */
6688 } nohz ____cacheline_aligned;
6690 #endif /* CONFIG_NO_HZ_COMMON */
6692 static unsigned long cpu_load(struct rq *rq)
6694 return cfs_rq_load_avg(&rq->cfs);
6698 * cpu_load_without - compute CPU load without any contributions from *p
6699 * @cpu: the CPU which load is requested
6700 * @p: the task which load should be discounted
6702 * The load of a CPU is defined by the load of tasks currently enqueued on that
6703 * CPU as well as tasks which are currently sleeping after an execution on that
6706 * This method returns the load of the specified CPU by discounting the load of
6707 * the specified task, whenever the task is currently contributing to the CPU
6710 static unsigned long cpu_load_without(struct rq *rq, struct task_struct *p)
6712 struct cfs_rq *cfs_rq;
6715 /* Task has no contribution or is new */
6716 if (cpu_of(rq) != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time))
6717 return cpu_load(rq);
6720 load = READ_ONCE(cfs_rq->avg.load_avg);
6722 /* Discount task's util from CPU's util */
6723 lsub_positive(&load, task_h_load(p));
6728 static unsigned long cpu_runnable(struct rq *rq)
6730 return cfs_rq_runnable_avg(&rq->cfs);
6733 static unsigned long cpu_runnable_without(struct rq *rq, struct task_struct *p)
6735 struct cfs_rq *cfs_rq;
6736 unsigned int runnable;
6738 /* Task has no contribution or is new */
6739 if (cpu_of(rq) != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time))
6740 return cpu_runnable(rq);
6743 runnable = READ_ONCE(cfs_rq->avg.runnable_avg);
6745 /* Discount task's runnable from CPU's runnable */
6746 lsub_positive(&runnable, p->se.avg.runnable_avg);
6751 static unsigned long capacity_of(int cpu)
6753 return cpu_rq(cpu)->cpu_capacity;
6756 static void record_wakee(struct task_struct *p)
6759 * Only decay a single time; tasks that have less then 1 wakeup per
6760 * jiffy will not have built up many flips.
6762 if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) {
6763 current->wakee_flips >>= 1;
6764 current->wakee_flip_decay_ts = jiffies;
6767 if (current->last_wakee != p) {
6768 current->last_wakee = p;
6769 current->wakee_flips++;
6774 * Detect M:N waker/wakee relationships via a switching-frequency heuristic.
6776 * A waker of many should wake a different task than the one last awakened
6777 * at a frequency roughly N times higher than one of its wakees.
6779 * In order to determine whether we should let the load spread vs consolidating
6780 * to shared cache, we look for a minimum 'flip' frequency of llc_size in one
6781 * partner, and a factor of lls_size higher frequency in the other.
6783 * With both conditions met, we can be relatively sure that the relationship is
6784 * non-monogamous, with partner count exceeding socket size.
6786 * Waker/wakee being client/server, worker/dispatcher, interrupt source or
6787 * whatever is irrelevant, spread criteria is apparent partner count exceeds
6790 static int wake_wide(struct task_struct *p)
6792 unsigned int master = current->wakee_flips;
6793 unsigned int slave = p->wakee_flips;
6794 int factor = __this_cpu_read(sd_llc_size);
6797 swap(master, slave);
6798 if (slave < factor || master < slave * factor)
6804 * The purpose of wake_affine() is to quickly determine on which CPU we can run
6805 * soonest. For the purpose of speed we only consider the waking and previous
6808 * wake_affine_idle() - only considers 'now', it check if the waking CPU is
6809 * cache-affine and is (or will be) idle.
6811 * wake_affine_weight() - considers the weight to reflect the average
6812 * scheduling latency of the CPUs. This seems to work
6813 * for the overloaded case.
6816 wake_affine_idle(int this_cpu, int prev_cpu, int sync)
6819 * If this_cpu is idle, it implies the wakeup is from interrupt
6820 * context. Only allow the move if cache is shared. Otherwise an
6821 * interrupt intensive workload could force all tasks onto one
6822 * node depending on the IO topology or IRQ affinity settings.
6824 * If the prev_cpu is idle and cache affine then avoid a migration.
6825 * There is no guarantee that the cache hot data from an interrupt
6826 * is more important than cache hot data on the prev_cpu and from
6827 * a cpufreq perspective, it's better to have higher utilisation
6830 if (available_idle_cpu(this_cpu) && cpus_share_cache(this_cpu, prev_cpu))
6831 return available_idle_cpu(prev_cpu) ? prev_cpu : this_cpu;
6833 if (sync && cpu_rq(this_cpu)->nr_running == 1)
6836 if (available_idle_cpu(prev_cpu))
6839 return nr_cpumask_bits;
6843 wake_affine_weight(struct sched_domain *sd, struct task_struct *p,
6844 int this_cpu, int prev_cpu, int sync)
6846 s64 this_eff_load, prev_eff_load;
6847 unsigned long task_load;
6849 this_eff_load = cpu_load(cpu_rq(this_cpu));
6852 unsigned long current_load = task_h_load(current);
6854 if (current_load > this_eff_load)
6857 this_eff_load -= current_load;
6860 task_load = task_h_load(p);
6862 this_eff_load += task_load;
6863 if (sched_feat(WA_BIAS))
6864 this_eff_load *= 100;
6865 this_eff_load *= capacity_of(prev_cpu);
6867 prev_eff_load = cpu_load(cpu_rq(prev_cpu));
6868 prev_eff_load -= task_load;
6869 if (sched_feat(WA_BIAS))
6870 prev_eff_load *= 100 + (sd->imbalance_pct - 100) / 2;
6871 prev_eff_load *= capacity_of(this_cpu);
6874 * If sync, adjust the weight of prev_eff_load such that if
6875 * prev_eff == this_eff that select_idle_sibling() will consider
6876 * stacking the wakee on top of the waker if no other CPU is
6882 return this_eff_load < prev_eff_load ? this_cpu : nr_cpumask_bits;
6885 static int wake_affine(struct sched_domain *sd, struct task_struct *p,
6886 int this_cpu, int prev_cpu, int sync)
6888 int target = nr_cpumask_bits;
6890 if (sched_feat(WA_IDLE))
6891 target = wake_affine_idle(this_cpu, prev_cpu, sync);
6893 if (sched_feat(WA_WEIGHT) && target == nr_cpumask_bits)
6894 target = wake_affine_weight(sd, p, this_cpu, prev_cpu, sync);
6896 schedstat_inc(p->stats.nr_wakeups_affine_attempts);
6897 if (target != this_cpu)
6900 schedstat_inc(sd->ttwu_move_affine);
6901 schedstat_inc(p->stats.nr_wakeups_affine);
6905 static struct sched_group *
6906 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu);
6909 * find_idlest_group_cpu - find the idlest CPU among the CPUs in the group.
6912 find_idlest_group_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
6914 unsigned long load, min_load = ULONG_MAX;
6915 unsigned int min_exit_latency = UINT_MAX;
6916 u64 latest_idle_timestamp = 0;
6917 int least_loaded_cpu = this_cpu;
6918 int shallowest_idle_cpu = -1;
6921 /* Check if we have any choice: */
6922 if (group->group_weight == 1)
6923 return cpumask_first(sched_group_span(group));
6925 /* Traverse only the allowed CPUs */
6926 for_each_cpu_and(i, sched_group_span(group), p->cpus_ptr) {
6927 struct rq *rq = cpu_rq(i);
6929 if (!sched_core_cookie_match(rq, p))
6932 if (sched_idle_cpu(i))
6935 if (available_idle_cpu(i)) {
6936 struct cpuidle_state *idle = idle_get_state(rq);
6937 if (idle && idle->exit_latency < min_exit_latency) {
6939 * We give priority to a CPU whose idle state
6940 * has the smallest exit latency irrespective
6941 * of any idle timestamp.
6943 min_exit_latency = idle->exit_latency;
6944 latest_idle_timestamp = rq->idle_stamp;
6945 shallowest_idle_cpu = i;
6946 } else if ((!idle || idle->exit_latency == min_exit_latency) &&
6947 rq->idle_stamp > latest_idle_timestamp) {
6949 * If equal or no active idle state, then
6950 * the most recently idled CPU might have
6953 latest_idle_timestamp = rq->idle_stamp;
6954 shallowest_idle_cpu = i;
6956 } else if (shallowest_idle_cpu == -1) {
6957 load = cpu_load(cpu_rq(i));
6958 if (load < min_load) {
6960 least_loaded_cpu = i;
6965 return shallowest_idle_cpu != -1 ? shallowest_idle_cpu : least_loaded_cpu;
6968 static inline int find_idlest_cpu(struct sched_domain *sd, struct task_struct *p,
6969 int cpu, int prev_cpu, int sd_flag)
6973 if (!cpumask_intersects(sched_domain_span(sd), p->cpus_ptr))
6977 * We need task's util for cpu_util_without, sync it up to
6978 * prev_cpu's last_update_time.
6980 if (!(sd_flag & SD_BALANCE_FORK))
6981 sync_entity_load_avg(&p->se);
6984 struct sched_group *group;
6985 struct sched_domain *tmp;
6988 if (!(sd->flags & sd_flag)) {
6993 group = find_idlest_group(sd, p, cpu);
6999 new_cpu = find_idlest_group_cpu(group, p, cpu);
7000 if (new_cpu == cpu) {
7001 /* Now try balancing at a lower domain level of 'cpu': */
7006 /* Now try balancing at a lower domain level of 'new_cpu': */
7008 weight = sd->span_weight;
7010 for_each_domain(cpu, tmp) {
7011 if (weight <= tmp->span_weight)
7013 if (tmp->flags & sd_flag)
7021 static inline int __select_idle_cpu(int cpu, struct task_struct *p)
7023 if ((available_idle_cpu(cpu) || sched_idle_cpu(cpu)) &&
7024 sched_cpu_cookie_match(cpu_rq(cpu), p))
7030 #ifdef CONFIG_SCHED_SMT
7031 DEFINE_STATIC_KEY_FALSE(sched_smt_present);
7032 EXPORT_SYMBOL_GPL(sched_smt_present);
7034 static inline void set_idle_cores(int cpu, int val)
7036 struct sched_domain_shared *sds;
7038 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
7040 WRITE_ONCE(sds->has_idle_cores, val);
7043 static inline bool test_idle_cores(int cpu)
7045 struct sched_domain_shared *sds;
7047 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
7049 return READ_ONCE(sds->has_idle_cores);
7055 * Scans the local SMT mask to see if the entire core is idle, and records this
7056 * information in sd_llc_shared->has_idle_cores.
7058 * Since SMT siblings share all cache levels, inspecting this limited remote
7059 * state should be fairly cheap.
7061 void __update_idle_core(struct rq *rq)
7063 int core = cpu_of(rq);
7067 if (test_idle_cores(core))
7070 for_each_cpu(cpu, cpu_smt_mask(core)) {
7074 if (!available_idle_cpu(cpu))
7078 set_idle_cores(core, 1);
7084 * Scan the entire LLC domain for idle cores; this dynamically switches off if
7085 * there are no idle cores left in the system; tracked through
7086 * sd_llc->shared->has_idle_cores and enabled through update_idle_core() above.
7088 static int select_idle_core(struct task_struct *p, int core, struct cpumask *cpus, int *idle_cpu)
7093 for_each_cpu(cpu, cpu_smt_mask(core)) {
7094 if (!available_idle_cpu(cpu)) {
7096 if (*idle_cpu == -1) {
7097 if (sched_idle_cpu(cpu) && cpumask_test_cpu(cpu, p->cpus_ptr)) {
7105 if (*idle_cpu == -1 && cpumask_test_cpu(cpu, p->cpus_ptr))
7112 cpumask_andnot(cpus, cpus, cpu_smt_mask(core));
7117 * Scan the local SMT mask for idle CPUs.
7119 static int select_idle_smt(struct task_struct *p, int target)
7123 for_each_cpu_and(cpu, cpu_smt_mask(target), p->cpus_ptr) {
7126 if (available_idle_cpu(cpu) || sched_idle_cpu(cpu))
7133 #else /* CONFIG_SCHED_SMT */
7135 static inline void set_idle_cores(int cpu, int val)
7139 static inline bool test_idle_cores(int cpu)
7144 static inline int select_idle_core(struct task_struct *p, int core, struct cpumask *cpus, int *idle_cpu)
7146 return __select_idle_cpu(core, p);
7149 static inline int select_idle_smt(struct task_struct *p, int target)
7154 #endif /* CONFIG_SCHED_SMT */
7157 * Scan the LLC domain for idle CPUs; this is dynamically regulated by
7158 * comparing the average scan cost (tracked in sd->avg_scan_cost) against the
7159 * average idle time for this rq (as found in rq->avg_idle).
7161 static int select_idle_cpu(struct task_struct *p, struct sched_domain *sd, bool has_idle_core, int target)
7163 struct cpumask *cpus = this_cpu_cpumask_var_ptr(select_rq_mask);
7164 int i, cpu, idle_cpu = -1, nr = INT_MAX;
7165 struct sched_domain_shared *sd_share;
7166 struct rq *this_rq = this_rq();
7167 int this = smp_processor_id();
7168 struct sched_domain *this_sd = NULL;
7171 cpumask_and(cpus, sched_domain_span(sd), p->cpus_ptr);
7173 if (sched_feat(SIS_PROP) && !has_idle_core) {
7174 u64 avg_cost, avg_idle, span_avg;
7175 unsigned long now = jiffies;
7177 this_sd = rcu_dereference(*this_cpu_ptr(&sd_llc));
7182 * If we're busy, the assumption that the last idle period
7183 * predicts the future is flawed; age away the remaining
7184 * predicted idle time.
7186 if (unlikely(this_rq->wake_stamp < now)) {
7187 while (this_rq->wake_stamp < now && this_rq->wake_avg_idle) {
7188 this_rq->wake_stamp++;
7189 this_rq->wake_avg_idle >>= 1;
7193 avg_idle = this_rq->wake_avg_idle;
7194 avg_cost = this_sd->avg_scan_cost + 1;
7196 span_avg = sd->span_weight * avg_idle;
7197 if (span_avg > 4*avg_cost)
7198 nr = div_u64(span_avg, avg_cost);
7202 time = cpu_clock(this);
7205 if (sched_feat(SIS_UTIL)) {
7206 sd_share = rcu_dereference(per_cpu(sd_llc_shared, target));
7208 /* because !--nr is the condition to stop scan */
7209 nr = READ_ONCE(sd_share->nr_idle_scan) + 1;
7210 /* overloaded LLC is unlikely to have idle cpu/core */
7216 for_each_cpu_wrap(cpu, cpus, target + 1) {
7217 if (has_idle_core) {
7218 i = select_idle_core(p, cpu, cpus, &idle_cpu);
7219 if ((unsigned int)i < nr_cpumask_bits)
7225 idle_cpu = __select_idle_cpu(cpu, p);
7226 if ((unsigned int)idle_cpu < nr_cpumask_bits)
7232 set_idle_cores(target, false);
7234 if (sched_feat(SIS_PROP) && this_sd && !has_idle_core) {
7235 time = cpu_clock(this) - time;
7238 * Account for the scan cost of wakeups against the average
7241 this_rq->wake_avg_idle -= min(this_rq->wake_avg_idle, time);
7243 update_avg(&this_sd->avg_scan_cost, time);
7250 * Scan the asym_capacity domain for idle CPUs; pick the first idle one on which
7251 * the task fits. If no CPU is big enough, but there are idle ones, try to
7252 * maximize capacity.
7255 select_idle_capacity(struct task_struct *p, struct sched_domain *sd, int target)
7257 unsigned long task_util, util_min, util_max, best_cap = 0;
7258 int fits, best_fits = 0;
7259 int cpu, best_cpu = -1;
7260 struct cpumask *cpus;
7262 cpus = this_cpu_cpumask_var_ptr(select_rq_mask);
7263 cpumask_and(cpus, sched_domain_span(sd), p->cpus_ptr);
7265 task_util = task_util_est(p);
7266 util_min = uclamp_eff_value(p, UCLAMP_MIN);
7267 util_max = uclamp_eff_value(p, UCLAMP_MAX);
7269 for_each_cpu_wrap(cpu, cpus, target) {
7270 unsigned long cpu_cap = capacity_of(cpu);
7272 if (!available_idle_cpu(cpu) && !sched_idle_cpu(cpu))
7275 fits = util_fits_cpu(task_util, util_min, util_max, cpu);
7277 /* This CPU fits with all requirements */
7281 * Only the min performance hint (i.e. uclamp_min) doesn't fit.
7282 * Look for the CPU with best capacity.
7285 cpu_cap = capacity_orig_of(cpu) - thermal_load_avg(cpu_rq(cpu));
7288 * First, select CPU which fits better (-1 being better than 0).
7289 * Then, select the one with best capacity at same level.
7291 if ((fits < best_fits) ||
7292 ((fits == best_fits) && (cpu_cap > best_cap))) {
7302 static inline bool asym_fits_cpu(unsigned long util,
7303 unsigned long util_min,
7304 unsigned long util_max,
7307 if (sched_asym_cpucap_active())
7309 * Return true only if the cpu fully fits the task requirements
7310 * which include the utilization and the performance hints.
7312 return (util_fits_cpu(util, util_min, util_max, cpu) > 0);
7318 * Try and locate an idle core/thread in the LLC cache domain.
7320 static int select_idle_sibling(struct task_struct *p, int prev, int target)
7322 bool has_idle_core = false;
7323 struct sched_domain *sd;
7324 unsigned long task_util, util_min, util_max;
7325 int i, recent_used_cpu;
7328 * On asymmetric system, update task utilization because we will check
7329 * that the task fits with cpu's capacity.
7331 if (sched_asym_cpucap_active()) {
7332 sync_entity_load_avg(&p->se);
7333 task_util = task_util_est(p);
7334 util_min = uclamp_eff_value(p, UCLAMP_MIN);
7335 util_max = uclamp_eff_value(p, UCLAMP_MAX);
7339 * per-cpu select_rq_mask usage
7341 lockdep_assert_irqs_disabled();
7343 if ((available_idle_cpu(target) || sched_idle_cpu(target)) &&
7344 asym_fits_cpu(task_util, util_min, util_max, target))
7348 * If the previous CPU is cache affine and idle, don't be stupid:
7350 if (prev != target && cpus_share_cache(prev, target) &&
7351 (available_idle_cpu(prev) || sched_idle_cpu(prev)) &&
7352 asym_fits_cpu(task_util, util_min, util_max, prev))
7356 * Allow a per-cpu kthread to stack with the wakee if the
7357 * kworker thread and the tasks previous CPUs are the same.
7358 * The assumption is that the wakee queued work for the
7359 * per-cpu kthread that is now complete and the wakeup is
7360 * essentially a sync wakeup. An obvious example of this
7361 * pattern is IO completions.
7363 if (is_per_cpu_kthread(current) &&
7365 prev == smp_processor_id() &&
7366 this_rq()->nr_running <= 1 &&
7367 asym_fits_cpu(task_util, util_min, util_max, prev)) {
7371 /* Check a recently used CPU as a potential idle candidate: */
7372 recent_used_cpu = p->recent_used_cpu;
7373 p->recent_used_cpu = prev;
7374 if (recent_used_cpu != prev &&
7375 recent_used_cpu != target &&
7376 cpus_share_cache(recent_used_cpu, target) &&
7377 (available_idle_cpu(recent_used_cpu) || sched_idle_cpu(recent_used_cpu)) &&
7378 cpumask_test_cpu(recent_used_cpu, p->cpus_ptr) &&
7379 asym_fits_cpu(task_util, util_min, util_max, recent_used_cpu)) {
7380 return recent_used_cpu;
7384 * For asymmetric CPU capacity systems, our domain of interest is
7385 * sd_asym_cpucapacity rather than sd_llc.
7387 if (sched_asym_cpucap_active()) {
7388 sd = rcu_dereference(per_cpu(sd_asym_cpucapacity, target));
7390 * On an asymmetric CPU capacity system where an exclusive
7391 * cpuset defines a symmetric island (i.e. one unique
7392 * capacity_orig value through the cpuset), the key will be set
7393 * but the CPUs within that cpuset will not have a domain with
7394 * SD_ASYM_CPUCAPACITY. These should follow the usual symmetric
7398 i = select_idle_capacity(p, sd, target);
7399 return ((unsigned)i < nr_cpumask_bits) ? i : target;
7403 sd = rcu_dereference(per_cpu(sd_llc, target));
7407 if (sched_smt_active()) {
7408 has_idle_core = test_idle_cores(target);
7410 if (!has_idle_core && cpus_share_cache(prev, target)) {
7411 i = select_idle_smt(p, prev);
7412 if ((unsigned int)i < nr_cpumask_bits)
7417 i = select_idle_cpu(p, sd, has_idle_core, target);
7418 if ((unsigned)i < nr_cpumask_bits)
7425 * cpu_util() - Estimates the amount of CPU capacity used by CFS tasks.
7426 * @cpu: the CPU to get the utilization for
7427 * @p: task for which the CPU utilization should be predicted or NULL
7428 * @dst_cpu: CPU @p migrates to, -1 if @p moves from @cpu or @p == NULL
7429 * @boost: 1 to enable boosting, otherwise 0
7431 * The unit of the return value must be the same as the one of CPU capacity
7432 * so that CPU utilization can be compared with CPU capacity.
7434 * CPU utilization is the sum of running time of runnable tasks plus the
7435 * recent utilization of currently non-runnable tasks on that CPU.
7436 * It represents the amount of CPU capacity currently used by CFS tasks in
7437 * the range [0..max CPU capacity] with max CPU capacity being the CPU
7438 * capacity at f_max.
7440 * The estimated CPU utilization is defined as the maximum between CPU
7441 * utilization and sum of the estimated utilization of the currently
7442 * runnable tasks on that CPU. It preserves a utilization "snapshot" of
7443 * previously-executed tasks, which helps better deduce how busy a CPU will
7444 * be when a long-sleeping task wakes up. The contribution to CPU utilization
7445 * of such a task would be significantly decayed at this point of time.
7447 * Boosted CPU utilization is defined as max(CPU runnable, CPU utilization).
7448 * CPU contention for CFS tasks can be detected by CPU runnable > CPU
7449 * utilization. Boosting is implemented in cpu_util() so that internal
7450 * users (e.g. EAS) can use it next to external users (e.g. schedutil),
7451 * latter via cpu_util_cfs_boost().
7453 * CPU utilization can be higher than the current CPU capacity
7454 * (f_curr/f_max * max CPU capacity) or even the max CPU capacity because
7455 * of rounding errors as well as task migrations or wakeups of new tasks.
7456 * CPU utilization has to be capped to fit into the [0..max CPU capacity]
7457 * range. Otherwise a group of CPUs (CPU0 util = 121% + CPU1 util = 80%)
7458 * could be seen as over-utilized even though CPU1 has 20% of spare CPU
7459 * capacity. CPU utilization is allowed to overshoot current CPU capacity
7460 * though since this is useful for predicting the CPU capacity required
7461 * after task migrations (scheduler-driven DVFS).
7463 * Return: (Boosted) (estimated) utilization for the specified CPU.
7465 static unsigned long
7466 cpu_util(int cpu, struct task_struct *p, int dst_cpu, int boost)
7468 struct cfs_rq *cfs_rq = &cpu_rq(cpu)->cfs;
7469 unsigned long util = READ_ONCE(cfs_rq->avg.util_avg);
7470 unsigned long runnable;
7473 runnable = READ_ONCE(cfs_rq->avg.runnable_avg);
7474 util = max(util, runnable);
7478 * If @dst_cpu is -1 or @p migrates from @cpu to @dst_cpu remove its
7479 * contribution. If @p migrates from another CPU to @cpu add its
7480 * contribution. In all the other cases @cpu is not impacted by the
7481 * migration so its util_avg is already correct.
7483 if (p && task_cpu(p) == cpu && dst_cpu != cpu)
7484 lsub_positive(&util, task_util(p));
7485 else if (p && task_cpu(p) != cpu && dst_cpu == cpu)
7486 util += task_util(p);
7488 if (sched_feat(UTIL_EST)) {
7489 unsigned long util_est;
7491 util_est = READ_ONCE(cfs_rq->avg.util_est.enqueued);
7494 * During wake-up @p isn't enqueued yet and doesn't contribute
7495 * to any cpu_rq(cpu)->cfs.avg.util_est.enqueued.
7496 * If @dst_cpu == @cpu add it to "simulate" cpu_util after @p
7497 * has been enqueued.
7499 * During exec (@dst_cpu = -1) @p is enqueued and does
7500 * contribute to cpu_rq(cpu)->cfs.util_est.enqueued.
7501 * Remove it to "simulate" cpu_util without @p's contribution.
7503 * Despite the task_on_rq_queued(@p) check there is still a
7504 * small window for a possible race when an exec
7505 * select_task_rq_fair() races with LB's detach_task().
7509 * p->on_rq = TASK_ON_RQ_MIGRATING;
7510 * -------------------------------- A
7512 * dequeue_task_fair() + Race Time
7513 * util_est_dequeue() /
7514 * -------------------------------- B
7516 * The additional check "current == p" is required to further
7517 * reduce the race window.
7520 util_est += _task_util_est(p);
7521 else if (p && unlikely(task_on_rq_queued(p) || current == p))
7522 lsub_positive(&util_est, _task_util_est(p));
7524 util = max(util, util_est);
7527 return min(util, capacity_orig_of(cpu));
7530 unsigned long cpu_util_cfs(int cpu)
7532 return cpu_util(cpu, NULL, -1, 0);
7535 unsigned long cpu_util_cfs_boost(int cpu)
7537 return cpu_util(cpu, NULL, -1, 1);
7541 * cpu_util_without: compute cpu utilization without any contributions from *p
7542 * @cpu: the CPU which utilization is requested
7543 * @p: the task which utilization should be discounted
7545 * The utilization of a CPU is defined by the utilization of tasks currently
7546 * enqueued on that CPU as well as tasks which are currently sleeping after an
7547 * execution on that CPU.
7549 * This method returns the utilization of the specified CPU by discounting the
7550 * utilization of the specified task, whenever the task is currently
7551 * contributing to the CPU utilization.
7553 static unsigned long cpu_util_without(int cpu, struct task_struct *p)
7555 /* Task has no contribution or is new */
7556 if (cpu != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time))
7559 return cpu_util(cpu, p, -1, 0);
7563 * energy_env - Utilization landscape for energy estimation.
7564 * @task_busy_time: Utilization contribution by the task for which we test the
7565 * placement. Given by eenv_task_busy_time().
7566 * @pd_busy_time: Utilization of the whole perf domain without the task
7567 * contribution. Given by eenv_pd_busy_time().
7568 * @cpu_cap: Maximum CPU capacity for the perf domain.
7569 * @pd_cap: Entire perf domain capacity. (pd->nr_cpus * cpu_cap).
7572 unsigned long task_busy_time;
7573 unsigned long pd_busy_time;
7574 unsigned long cpu_cap;
7575 unsigned long pd_cap;
7579 * Compute the task busy time for compute_energy(). This time cannot be
7580 * injected directly into effective_cpu_util() because of the IRQ scaling.
7581 * The latter only makes sense with the most recent CPUs where the task has
7584 static inline void eenv_task_busy_time(struct energy_env *eenv,
7585 struct task_struct *p, int prev_cpu)
7587 unsigned long busy_time, max_cap = arch_scale_cpu_capacity(prev_cpu);
7588 unsigned long irq = cpu_util_irq(cpu_rq(prev_cpu));
7590 if (unlikely(irq >= max_cap))
7591 busy_time = max_cap;
7593 busy_time = scale_irq_capacity(task_util_est(p), irq, max_cap);
7595 eenv->task_busy_time = busy_time;
7599 * Compute the perf_domain (PD) busy time for compute_energy(). Based on the
7600 * utilization for each @pd_cpus, it however doesn't take into account
7601 * clamping since the ratio (utilization / cpu_capacity) is already enough to
7602 * scale the EM reported power consumption at the (eventually clamped)
7605 * The contribution of the task @p for which we want to estimate the
7606 * energy cost is removed (by cpu_util()) and must be calculated
7607 * separately (see eenv_task_busy_time). This ensures:
7609 * - A stable PD utilization, no matter which CPU of that PD we want to place
7612 * - A fair comparison between CPUs as the task contribution (task_util())
7613 * will always be the same no matter which CPU utilization we rely on
7614 * (util_avg or util_est).
7616 * Set @eenv busy time for the PD that spans @pd_cpus. This busy time can't
7617 * exceed @eenv->pd_cap.
7619 static inline void eenv_pd_busy_time(struct energy_env *eenv,
7620 struct cpumask *pd_cpus,
7621 struct task_struct *p)
7623 unsigned long busy_time = 0;
7626 for_each_cpu(cpu, pd_cpus) {
7627 unsigned long util = cpu_util(cpu, p, -1, 0);
7629 busy_time += effective_cpu_util(cpu, util, ENERGY_UTIL, NULL);
7632 eenv->pd_busy_time = min(eenv->pd_cap, busy_time);
7636 * Compute the maximum utilization for compute_energy() when the task @p
7637 * is placed on the cpu @dst_cpu.
7639 * Returns the maximum utilization among @eenv->cpus. This utilization can't
7640 * exceed @eenv->cpu_cap.
7642 static inline unsigned long
7643 eenv_pd_max_util(struct energy_env *eenv, struct cpumask *pd_cpus,
7644 struct task_struct *p, int dst_cpu)
7646 unsigned long max_util = 0;
7649 for_each_cpu(cpu, pd_cpus) {
7650 struct task_struct *tsk = (cpu == dst_cpu) ? p : NULL;
7651 unsigned long util = cpu_util(cpu, p, dst_cpu, 1);
7652 unsigned long eff_util;
7655 * Performance domain frequency: utilization clamping
7656 * must be considered since it affects the selection
7657 * of the performance domain frequency.
7658 * NOTE: in case RT tasks are running, by default the
7659 * FREQUENCY_UTIL's utilization can be max OPP.
7661 eff_util = effective_cpu_util(cpu, util, FREQUENCY_UTIL, tsk);
7662 max_util = max(max_util, eff_util);
7665 return min(max_util, eenv->cpu_cap);
7669 * compute_energy(): Use the Energy Model to estimate the energy that @pd would
7670 * consume for a given utilization landscape @eenv. When @dst_cpu < 0, the task
7671 * contribution is ignored.
7673 static inline unsigned long
7674 compute_energy(struct energy_env *eenv, struct perf_domain *pd,
7675 struct cpumask *pd_cpus, struct task_struct *p, int dst_cpu)
7677 unsigned long max_util = eenv_pd_max_util(eenv, pd_cpus, p, dst_cpu);
7678 unsigned long busy_time = eenv->pd_busy_time;
7681 busy_time = min(eenv->pd_cap, busy_time + eenv->task_busy_time);
7683 return em_cpu_energy(pd->em_pd, max_util, busy_time, eenv->cpu_cap);
7687 * find_energy_efficient_cpu(): Find most energy-efficient target CPU for the
7688 * waking task. find_energy_efficient_cpu() looks for the CPU with maximum
7689 * spare capacity in each performance domain and uses it as a potential
7690 * candidate to execute the task. Then, it uses the Energy Model to figure
7691 * out which of the CPU candidates is the most energy-efficient.
7693 * The rationale for this heuristic is as follows. In a performance domain,
7694 * all the most energy efficient CPU candidates (according to the Energy
7695 * Model) are those for which we'll request a low frequency. When there are
7696 * several CPUs for which the frequency request will be the same, we don't
7697 * have enough data to break the tie between them, because the Energy Model
7698 * only includes active power costs. With this model, if we assume that
7699 * frequency requests follow utilization (e.g. using schedutil), the CPU with
7700 * the maximum spare capacity in a performance domain is guaranteed to be among
7701 * the best candidates of the performance domain.
7703 * In practice, it could be preferable from an energy standpoint to pack
7704 * small tasks on a CPU in order to let other CPUs go in deeper idle states,
7705 * but that could also hurt our chances to go cluster idle, and we have no
7706 * ways to tell with the current Energy Model if this is actually a good
7707 * idea or not. So, find_energy_efficient_cpu() basically favors
7708 * cluster-packing, and spreading inside a cluster. That should at least be
7709 * a good thing for latency, and this is consistent with the idea that most
7710 * of the energy savings of EAS come from the asymmetry of the system, and
7711 * not so much from breaking the tie between identical CPUs. That's also the
7712 * reason why EAS is enabled in the topology code only for systems where
7713 * SD_ASYM_CPUCAPACITY is set.
7715 * NOTE: Forkees are not accepted in the energy-aware wake-up path because
7716 * they don't have any useful utilization data yet and it's not possible to
7717 * forecast their impact on energy consumption. Consequently, they will be
7718 * placed by find_idlest_cpu() on the least loaded CPU, which might turn out
7719 * to be energy-inefficient in some use-cases. The alternative would be to
7720 * bias new tasks towards specific types of CPUs first, or to try to infer
7721 * their util_avg from the parent task, but those heuristics could hurt
7722 * other use-cases too. So, until someone finds a better way to solve this,
7723 * let's keep things simple by re-using the existing slow path.
7725 static int find_energy_efficient_cpu(struct task_struct *p, int prev_cpu)
7727 struct cpumask *cpus = this_cpu_cpumask_var_ptr(select_rq_mask);
7728 unsigned long prev_delta = ULONG_MAX, best_delta = ULONG_MAX;
7729 unsigned long p_util_min = uclamp_is_used() ? uclamp_eff_value(p, UCLAMP_MIN) : 0;
7730 unsigned long p_util_max = uclamp_is_used() ? uclamp_eff_value(p, UCLAMP_MAX) : 1024;
7731 struct root_domain *rd = this_rq()->rd;
7732 int cpu, best_energy_cpu, target = -1;
7733 int prev_fits = -1, best_fits = -1;
7734 unsigned long best_thermal_cap = 0;
7735 unsigned long prev_thermal_cap = 0;
7736 struct sched_domain *sd;
7737 struct perf_domain *pd;
7738 struct energy_env eenv;
7741 pd = rcu_dereference(rd->pd);
7742 if (!pd || READ_ONCE(rd->overutilized))
7746 * Energy-aware wake-up happens on the lowest sched_domain starting
7747 * from sd_asym_cpucapacity spanning over this_cpu and prev_cpu.
7749 sd = rcu_dereference(*this_cpu_ptr(&sd_asym_cpucapacity));
7750 while (sd && !cpumask_test_cpu(prev_cpu, sched_domain_span(sd)))
7757 sync_entity_load_avg(&p->se);
7758 if (!uclamp_task_util(p, p_util_min, p_util_max))
7761 eenv_task_busy_time(&eenv, p, prev_cpu);
7763 for (; pd; pd = pd->next) {
7764 unsigned long util_min = p_util_min, util_max = p_util_max;
7765 unsigned long cpu_cap, cpu_thermal_cap, util;
7766 unsigned long cur_delta, max_spare_cap = 0;
7767 unsigned long rq_util_min, rq_util_max;
7768 unsigned long prev_spare_cap = 0;
7769 int max_spare_cap_cpu = -1;
7770 unsigned long base_energy;
7771 int fits, max_fits = -1;
7773 cpumask_and(cpus, perf_domain_span(pd), cpu_online_mask);
7775 if (cpumask_empty(cpus))
7778 /* Account thermal pressure for the energy estimation */
7779 cpu = cpumask_first(cpus);
7780 cpu_thermal_cap = arch_scale_cpu_capacity(cpu);
7781 cpu_thermal_cap -= arch_scale_thermal_pressure(cpu);
7783 eenv.cpu_cap = cpu_thermal_cap;
7786 for_each_cpu(cpu, cpus) {
7787 struct rq *rq = cpu_rq(cpu);
7789 eenv.pd_cap += cpu_thermal_cap;
7791 if (!cpumask_test_cpu(cpu, sched_domain_span(sd)))
7794 if (!cpumask_test_cpu(cpu, p->cpus_ptr))
7797 util = cpu_util(cpu, p, cpu, 0);
7798 cpu_cap = capacity_of(cpu);
7801 * Skip CPUs that cannot satisfy the capacity request.
7802 * IOW, placing the task there would make the CPU
7803 * overutilized. Take uclamp into account to see how
7804 * much capacity we can get out of the CPU; this is
7805 * aligned with sched_cpu_util().
7807 if (uclamp_is_used() && !uclamp_rq_is_idle(rq)) {
7809 * Open code uclamp_rq_util_with() except for
7810 * the clamp() part. Ie: apply max aggregation
7811 * only. util_fits_cpu() logic requires to
7812 * operate on non clamped util but must use the
7813 * max-aggregated uclamp_{min, max}.
7815 rq_util_min = uclamp_rq_get(rq, UCLAMP_MIN);
7816 rq_util_max = uclamp_rq_get(rq, UCLAMP_MAX);
7818 util_min = max(rq_util_min, p_util_min);
7819 util_max = max(rq_util_max, p_util_max);
7822 fits = util_fits_cpu(util, util_min, util_max, cpu);
7826 lsub_positive(&cpu_cap, util);
7828 if (cpu == prev_cpu) {
7829 /* Always use prev_cpu as a candidate. */
7830 prev_spare_cap = cpu_cap;
7832 } else if ((fits > max_fits) ||
7833 ((fits == max_fits) && (cpu_cap > max_spare_cap))) {
7835 * Find the CPU with the maximum spare capacity
7836 * among the remaining CPUs in the performance
7839 max_spare_cap = cpu_cap;
7840 max_spare_cap_cpu = cpu;
7845 if (max_spare_cap_cpu < 0 && prev_spare_cap == 0)
7848 eenv_pd_busy_time(&eenv, cpus, p);
7849 /* Compute the 'base' energy of the pd, without @p */
7850 base_energy = compute_energy(&eenv, pd, cpus, p, -1);
7852 /* Evaluate the energy impact of using prev_cpu. */
7853 if (prev_spare_cap > 0) {
7854 prev_delta = compute_energy(&eenv, pd, cpus, p,
7856 /* CPU utilization has changed */
7857 if (prev_delta < base_energy)
7859 prev_delta -= base_energy;
7860 prev_thermal_cap = cpu_thermal_cap;
7861 best_delta = min(best_delta, prev_delta);
7864 /* Evaluate the energy impact of using max_spare_cap_cpu. */
7865 if (max_spare_cap_cpu >= 0 && max_spare_cap > prev_spare_cap) {
7866 /* Current best energy cpu fits better */
7867 if (max_fits < best_fits)
7871 * Both don't fit performance hint (i.e. uclamp_min)
7872 * but best energy cpu has better capacity.
7874 if ((max_fits < 0) &&
7875 (cpu_thermal_cap <= best_thermal_cap))
7878 cur_delta = compute_energy(&eenv, pd, cpus, p,
7880 /* CPU utilization has changed */
7881 if (cur_delta < base_energy)
7883 cur_delta -= base_energy;
7886 * Both fit for the task but best energy cpu has lower
7889 if ((max_fits > 0) && (best_fits > 0) &&
7890 (cur_delta >= best_delta))
7893 best_delta = cur_delta;
7894 best_energy_cpu = max_spare_cap_cpu;
7895 best_fits = max_fits;
7896 best_thermal_cap = cpu_thermal_cap;
7901 if ((best_fits > prev_fits) ||
7902 ((best_fits > 0) && (best_delta < prev_delta)) ||
7903 ((best_fits < 0) && (best_thermal_cap > prev_thermal_cap)))
7904 target = best_energy_cpu;
7915 * select_task_rq_fair: Select target runqueue for the waking task in domains
7916 * that have the relevant SD flag set. In practice, this is SD_BALANCE_WAKE,
7917 * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
7919 * Balances load by selecting the idlest CPU in the idlest group, or under
7920 * certain conditions an idle sibling CPU if the domain has SD_WAKE_AFFINE set.
7922 * Returns the target CPU number.
7925 select_task_rq_fair(struct task_struct *p, int prev_cpu, int wake_flags)
7927 int sync = (wake_flags & WF_SYNC) && !(current->flags & PF_EXITING);
7928 struct sched_domain *tmp, *sd = NULL;
7929 int cpu = smp_processor_id();
7930 int new_cpu = prev_cpu;
7931 int want_affine = 0;
7932 /* SD_flags and WF_flags share the first nibble */
7933 int sd_flag = wake_flags & 0xF;
7936 * required for stable ->cpus_allowed
7938 lockdep_assert_held(&p->pi_lock);
7939 if (wake_flags & WF_TTWU) {
7942 if ((wake_flags & WF_CURRENT_CPU) &&
7943 cpumask_test_cpu(cpu, p->cpus_ptr))
7946 if (sched_energy_enabled()) {
7947 new_cpu = find_energy_efficient_cpu(p, prev_cpu);
7953 want_affine = !wake_wide(p) && cpumask_test_cpu(cpu, p->cpus_ptr);
7957 for_each_domain(cpu, tmp) {
7959 * If both 'cpu' and 'prev_cpu' are part of this domain,
7960 * cpu is a valid SD_WAKE_AFFINE target.
7962 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
7963 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
7964 if (cpu != prev_cpu)
7965 new_cpu = wake_affine(tmp, p, cpu, prev_cpu, sync);
7967 sd = NULL; /* Prefer wake_affine over balance flags */
7972 * Usually only true for WF_EXEC and WF_FORK, as sched_domains
7973 * usually do not have SD_BALANCE_WAKE set. That means wakeup
7974 * will usually go to the fast path.
7976 if (tmp->flags & sd_flag)
7978 else if (!want_affine)
7984 new_cpu = find_idlest_cpu(sd, p, cpu, prev_cpu, sd_flag);
7985 } else if (wake_flags & WF_TTWU) { /* XXX always ? */
7987 new_cpu = select_idle_sibling(p, prev_cpu, new_cpu);
7995 * Called immediately before a task is migrated to a new CPU; task_cpu(p) and
7996 * cfs_rq_of(p) references at time of call are still valid and identify the
7997 * previous CPU. The caller guarantees p->pi_lock or task_rq(p)->lock is held.
7999 static void migrate_task_rq_fair(struct task_struct *p, int new_cpu)
8001 struct sched_entity *se = &p->se;
8003 if (!task_on_rq_migrating(p)) {
8004 remove_entity_load_avg(se);
8007 * Here, the task's PELT values have been updated according to
8008 * the current rq's clock. But if that clock hasn't been
8009 * updated in a while, a substantial idle time will be missed,
8010 * leading to an inflation after wake-up on the new rq.
8012 * Estimate the missing time from the cfs_rq last_update_time
8013 * and update sched_avg to improve the PELT continuity after
8016 migrate_se_pelt_lag(se);
8019 /* Tell new CPU we are migrated */
8020 se->avg.last_update_time = 0;
8022 update_scan_period(p, new_cpu);
8025 static void task_dead_fair(struct task_struct *p)
8027 remove_entity_load_avg(&p->se);
8031 balance_fair(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
8036 return newidle_balance(rq, rf) != 0;
8038 #endif /* CONFIG_SMP */
8040 static void set_next_buddy(struct sched_entity *se)
8042 for_each_sched_entity(se) {
8043 if (SCHED_WARN_ON(!se->on_rq))
8047 cfs_rq_of(se)->next = se;
8052 * Preempt the current task with a newly woken task if needed:
8054 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
8056 struct task_struct *curr = rq->curr;
8057 struct sched_entity *se = &curr->se, *pse = &p->se;
8058 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
8059 int next_buddy_marked = 0;
8060 int cse_is_idle, pse_is_idle;
8062 if (unlikely(se == pse))
8066 * This is possible from callers such as attach_tasks(), in which we
8067 * unconditionally check_preempt_curr() after an enqueue (which may have
8068 * lead to a throttle). This both saves work and prevents false
8069 * next-buddy nomination below.
8071 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
8074 if (sched_feat(NEXT_BUDDY) && !(wake_flags & WF_FORK)) {
8075 set_next_buddy(pse);
8076 next_buddy_marked = 1;
8080 * We can come here with TIF_NEED_RESCHED already set from new task
8083 * Note: this also catches the edge-case of curr being in a throttled
8084 * group (e.g. via set_curr_task), since update_curr() (in the
8085 * enqueue of curr) will have resulted in resched being set. This
8086 * prevents us from potentially nominating it as a false LAST_BUDDY
8089 if (test_tsk_need_resched(curr))
8092 /* Idle tasks are by definition preempted by non-idle tasks. */
8093 if (unlikely(task_has_idle_policy(curr)) &&
8094 likely(!task_has_idle_policy(p)))
8098 * Batch and idle tasks do not preempt non-idle tasks (their preemption
8099 * is driven by the tick):
8101 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
8104 find_matching_se(&se, &pse);
8107 cse_is_idle = se_is_idle(se);
8108 pse_is_idle = se_is_idle(pse);
8111 * Preempt an idle group in favor of a non-idle group (and don't preempt
8112 * in the inverse case).
8114 if (cse_is_idle && !pse_is_idle)
8116 if (cse_is_idle != pse_is_idle)
8119 cfs_rq = cfs_rq_of(se);
8120 update_curr(cfs_rq);
8123 * XXX pick_eevdf(cfs_rq) != se ?
8125 if (pick_eevdf(cfs_rq) == pse)
8135 static struct task_struct *pick_task_fair(struct rq *rq)
8137 struct sched_entity *se;
8138 struct cfs_rq *cfs_rq;
8142 if (!cfs_rq->nr_running)
8146 struct sched_entity *curr = cfs_rq->curr;
8148 /* When we pick for a remote RQ, we'll not have done put_prev_entity() */
8151 update_curr(cfs_rq);
8155 if (unlikely(check_cfs_rq_runtime(cfs_rq)))
8159 se = pick_next_entity(cfs_rq, curr);
8160 cfs_rq = group_cfs_rq(se);
8167 struct task_struct *
8168 pick_next_task_fair(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
8170 struct cfs_rq *cfs_rq = &rq->cfs;
8171 struct sched_entity *se;
8172 struct task_struct *p;
8176 if (!sched_fair_runnable(rq))
8179 #ifdef CONFIG_FAIR_GROUP_SCHED
8180 if (!prev || prev->sched_class != &fair_sched_class)
8184 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
8185 * likely that a next task is from the same cgroup as the current.
8187 * Therefore attempt to avoid putting and setting the entire cgroup
8188 * hierarchy, only change the part that actually changes.
8192 struct sched_entity *curr = cfs_rq->curr;
8195 * Since we got here without doing put_prev_entity() we also
8196 * have to consider cfs_rq->curr. If it is still a runnable
8197 * entity, update_curr() will update its vruntime, otherwise
8198 * forget we've ever seen it.
8202 update_curr(cfs_rq);
8207 * This call to check_cfs_rq_runtime() will do the
8208 * throttle and dequeue its entity in the parent(s).
8209 * Therefore the nr_running test will indeed
8212 if (unlikely(check_cfs_rq_runtime(cfs_rq))) {
8215 if (!cfs_rq->nr_running)
8222 se = pick_next_entity(cfs_rq, curr);
8223 cfs_rq = group_cfs_rq(se);
8229 * Since we haven't yet done put_prev_entity and if the selected task
8230 * is a different task than we started out with, try and touch the
8231 * least amount of cfs_rqs.
8234 struct sched_entity *pse = &prev->se;
8236 while (!(cfs_rq = is_same_group(se, pse))) {
8237 int se_depth = se->depth;
8238 int pse_depth = pse->depth;
8240 if (se_depth <= pse_depth) {
8241 put_prev_entity(cfs_rq_of(pse), pse);
8242 pse = parent_entity(pse);
8244 if (se_depth >= pse_depth) {
8245 set_next_entity(cfs_rq_of(se), se);
8246 se = parent_entity(se);
8250 put_prev_entity(cfs_rq, pse);
8251 set_next_entity(cfs_rq, se);
8258 put_prev_task(rq, prev);
8261 se = pick_next_entity(cfs_rq, NULL);
8262 set_next_entity(cfs_rq, se);
8263 cfs_rq = group_cfs_rq(se);
8268 done: __maybe_unused;
8271 * Move the next running task to the front of
8272 * the list, so our cfs_tasks list becomes MRU
8275 list_move(&p->se.group_node, &rq->cfs_tasks);
8278 if (hrtick_enabled_fair(rq))
8279 hrtick_start_fair(rq, p);
8281 update_misfit_status(p, rq);
8282 sched_fair_update_stop_tick(rq, p);
8290 new_tasks = newidle_balance(rq, rf);
8293 * Because newidle_balance() releases (and re-acquires) rq->lock, it is
8294 * possible for any higher priority task to appear. In that case we
8295 * must re-start the pick_next_entity() loop.
8304 * rq is about to be idle, check if we need to update the
8305 * lost_idle_time of clock_pelt
8307 update_idle_rq_clock_pelt(rq);
8312 static struct task_struct *__pick_next_task_fair(struct rq *rq)
8314 return pick_next_task_fair(rq, NULL, NULL);
8318 * Account for a descheduled task:
8320 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
8322 struct sched_entity *se = &prev->se;
8323 struct cfs_rq *cfs_rq;
8325 for_each_sched_entity(se) {
8326 cfs_rq = cfs_rq_of(se);
8327 put_prev_entity(cfs_rq, se);
8332 * sched_yield() is very simple
8334 static void yield_task_fair(struct rq *rq)
8336 struct task_struct *curr = rq->curr;
8337 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
8338 struct sched_entity *se = &curr->se;
8341 * Are we the only task in the tree?
8343 if (unlikely(rq->nr_running == 1))
8346 clear_buddies(cfs_rq, se);
8348 update_rq_clock(rq);
8350 * Update run-time statistics of the 'current'.
8352 update_curr(cfs_rq);
8354 * Tell update_rq_clock() that we've just updated,
8355 * so we don't do microscopic update in schedule()
8356 * and double the fastpath cost.
8358 rq_clock_skip_update(rq);
8360 se->deadline += calc_delta_fair(se->slice, se);
8363 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p)
8365 struct sched_entity *se = &p->se;
8367 /* throttled hierarchies are not runnable */
8368 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
8371 /* Tell the scheduler that we'd really like pse to run next. */
8374 yield_task_fair(rq);
8380 /**************************************************
8381 * Fair scheduling class load-balancing methods.
8385 * The purpose of load-balancing is to achieve the same basic fairness the
8386 * per-CPU scheduler provides, namely provide a proportional amount of compute
8387 * time to each task. This is expressed in the following equation:
8389 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
8391 * Where W_i,n is the n-th weight average for CPU i. The instantaneous weight
8392 * W_i,0 is defined as:
8394 * W_i,0 = \Sum_j w_i,j (2)
8396 * Where w_i,j is the weight of the j-th runnable task on CPU i. This weight
8397 * is derived from the nice value as per sched_prio_to_weight[].
8399 * The weight average is an exponential decay average of the instantaneous
8402 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
8404 * C_i is the compute capacity of CPU i, typically it is the
8405 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
8406 * can also include other factors [XXX].
8408 * To achieve this balance we define a measure of imbalance which follows
8409 * directly from (1):
8411 * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4)
8413 * We them move tasks around to minimize the imbalance. In the continuous
8414 * function space it is obvious this converges, in the discrete case we get
8415 * a few fun cases generally called infeasible weight scenarios.
8418 * - infeasible weights;
8419 * - local vs global optima in the discrete case. ]
8424 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
8425 * for all i,j solution, we create a tree of CPUs that follows the hardware
8426 * topology where each level pairs two lower groups (or better). This results
8427 * in O(log n) layers. Furthermore we reduce the number of CPUs going up the
8428 * tree to only the first of the previous level and we decrease the frequency
8429 * of load-balance at each level inv. proportional to the number of CPUs in
8435 * \Sum { --- * --- * 2^i } = O(n) (5)
8437 * `- size of each group
8438 * | | `- number of CPUs doing load-balance
8440 * `- sum over all levels
8442 * Coupled with a limit on how many tasks we can migrate every balance pass,
8443 * this makes (5) the runtime complexity of the balancer.
8445 * An important property here is that each CPU is still (indirectly) connected
8446 * to every other CPU in at most O(log n) steps:
8448 * The adjacency matrix of the resulting graph is given by:
8451 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
8454 * And you'll find that:
8456 * A^(log_2 n)_i,j != 0 for all i,j (7)
8458 * Showing there's indeed a path between every CPU in at most O(log n) steps.
8459 * The task movement gives a factor of O(m), giving a convergence complexity
8462 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
8467 * In order to avoid CPUs going idle while there's still work to do, new idle
8468 * balancing is more aggressive and has the newly idle CPU iterate up the domain
8469 * tree itself instead of relying on other CPUs to bring it work.
8471 * This adds some complexity to both (5) and (8) but it reduces the total idle
8479 * Cgroups make a horror show out of (2), instead of a simple sum we get:
8482 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
8487 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
8489 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on CPU i.
8491 * The big problem is S_k, its a global sum needed to compute a local (W_i)
8494 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
8495 * rewrite all of this once again.]
8498 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
8500 enum fbq_type { regular, remote, all };
8503 * 'group_type' describes the group of CPUs at the moment of load balancing.
8505 * The enum is ordered by pulling priority, with the group with lowest priority
8506 * first so the group_type can simply be compared when selecting the busiest
8507 * group. See update_sd_pick_busiest().
8510 /* The group has spare capacity that can be used to run more tasks. */
8511 group_has_spare = 0,
8513 * The group is fully used and the tasks don't compete for more CPU
8514 * cycles. Nevertheless, some tasks might wait before running.
8518 * One task doesn't fit with CPU's capacity and must be migrated to a
8519 * more powerful CPU.
8523 * Balance SMT group that's fully busy. Can benefit from migration
8524 * a task on SMT with busy sibling to another CPU on idle core.
8528 * SD_ASYM_PACKING only: One local CPU with higher capacity is available,
8529 * and the task should be migrated to it instead of running on the
8534 * The tasks' affinity constraints previously prevented the scheduler
8535 * from balancing the load across the system.
8539 * The CPU is overloaded and can't provide expected CPU cycles to all
8545 enum migration_type {
8552 #define LBF_ALL_PINNED 0x01
8553 #define LBF_NEED_BREAK 0x02
8554 #define LBF_DST_PINNED 0x04
8555 #define LBF_SOME_PINNED 0x08
8556 #define LBF_ACTIVE_LB 0x10
8559 struct sched_domain *sd;
8567 struct cpumask *dst_grpmask;
8569 enum cpu_idle_type idle;
8571 /* The set of CPUs under consideration for load-balancing */
8572 struct cpumask *cpus;
8577 unsigned int loop_break;
8578 unsigned int loop_max;
8580 enum fbq_type fbq_type;
8581 enum migration_type migration_type;
8582 struct list_head tasks;
8586 * Is this task likely cache-hot:
8588 static int task_hot(struct task_struct *p, struct lb_env *env)
8592 lockdep_assert_rq_held(env->src_rq);
8594 if (p->sched_class != &fair_sched_class)
8597 if (unlikely(task_has_idle_policy(p)))
8600 /* SMT siblings share cache */
8601 if (env->sd->flags & SD_SHARE_CPUCAPACITY)
8605 * Buddy candidates are cache hot:
8607 if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running &&
8608 (&p->se == cfs_rq_of(&p->se)->next))
8611 if (sysctl_sched_migration_cost == -1)
8615 * Don't migrate task if the task's cookie does not match
8616 * with the destination CPU's core cookie.
8618 if (!sched_core_cookie_match(cpu_rq(env->dst_cpu), p))
8621 if (sysctl_sched_migration_cost == 0)
8624 delta = rq_clock_task(env->src_rq) - p->se.exec_start;
8626 return delta < (s64)sysctl_sched_migration_cost;
8629 #ifdef CONFIG_NUMA_BALANCING
8631 * Returns 1, if task migration degrades locality
8632 * Returns 0, if task migration improves locality i.e migration preferred.
8633 * Returns -1, if task migration is not affected by locality.
8635 static int migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
8637 struct numa_group *numa_group = rcu_dereference(p->numa_group);
8638 unsigned long src_weight, dst_weight;
8639 int src_nid, dst_nid, dist;
8641 if (!static_branch_likely(&sched_numa_balancing))
8644 if (!p->numa_faults || !(env->sd->flags & SD_NUMA))
8647 src_nid = cpu_to_node(env->src_cpu);
8648 dst_nid = cpu_to_node(env->dst_cpu);
8650 if (src_nid == dst_nid)
8653 /* Migrating away from the preferred node is always bad. */
8654 if (src_nid == p->numa_preferred_nid) {
8655 if (env->src_rq->nr_running > env->src_rq->nr_preferred_running)
8661 /* Encourage migration to the preferred node. */
8662 if (dst_nid == p->numa_preferred_nid)
8665 /* Leaving a core idle is often worse than degrading locality. */
8666 if (env->idle == CPU_IDLE)
8669 dist = node_distance(src_nid, dst_nid);
8671 src_weight = group_weight(p, src_nid, dist);
8672 dst_weight = group_weight(p, dst_nid, dist);
8674 src_weight = task_weight(p, src_nid, dist);
8675 dst_weight = task_weight(p, dst_nid, dist);
8678 return dst_weight < src_weight;
8682 static inline int migrate_degrades_locality(struct task_struct *p,
8690 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
8693 int can_migrate_task(struct task_struct *p, struct lb_env *env)
8697 lockdep_assert_rq_held(env->src_rq);
8700 * We do not migrate tasks that are:
8701 * 1) throttled_lb_pair, or
8702 * 2) cannot be migrated to this CPU due to cpus_ptr, or
8703 * 3) running (obviously), or
8704 * 4) are cache-hot on their current CPU.
8706 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
8709 /* Disregard pcpu kthreads; they are where they need to be. */
8710 if (kthread_is_per_cpu(p))
8713 if (!cpumask_test_cpu(env->dst_cpu, p->cpus_ptr)) {
8716 schedstat_inc(p->stats.nr_failed_migrations_affine);
8718 env->flags |= LBF_SOME_PINNED;
8721 * Remember if this task can be migrated to any other CPU in
8722 * our sched_group. We may want to revisit it if we couldn't
8723 * meet load balance goals by pulling other tasks on src_cpu.
8725 * Avoid computing new_dst_cpu
8727 * - if we have already computed one in current iteration
8728 * - if it's an active balance
8730 if (env->idle == CPU_NEWLY_IDLE ||
8731 env->flags & (LBF_DST_PINNED | LBF_ACTIVE_LB))
8734 /* Prevent to re-select dst_cpu via env's CPUs: */
8735 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
8736 if (cpumask_test_cpu(cpu, p->cpus_ptr)) {
8737 env->flags |= LBF_DST_PINNED;
8738 env->new_dst_cpu = cpu;
8746 /* Record that we found at least one task that could run on dst_cpu */
8747 env->flags &= ~LBF_ALL_PINNED;
8749 if (task_on_cpu(env->src_rq, p)) {
8750 schedstat_inc(p->stats.nr_failed_migrations_running);
8755 * Aggressive migration if:
8757 * 2) destination numa is preferred
8758 * 3) task is cache cold, or
8759 * 4) too many balance attempts have failed.
8761 if (env->flags & LBF_ACTIVE_LB)
8764 tsk_cache_hot = migrate_degrades_locality(p, env);
8765 if (tsk_cache_hot == -1)
8766 tsk_cache_hot = task_hot(p, env);
8768 if (tsk_cache_hot <= 0 ||
8769 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
8770 if (tsk_cache_hot == 1) {
8771 schedstat_inc(env->sd->lb_hot_gained[env->idle]);
8772 schedstat_inc(p->stats.nr_forced_migrations);
8777 schedstat_inc(p->stats.nr_failed_migrations_hot);
8782 * detach_task() -- detach the task for the migration specified in env
8784 static void detach_task(struct task_struct *p, struct lb_env *env)
8786 lockdep_assert_rq_held(env->src_rq);
8788 deactivate_task(env->src_rq, p, DEQUEUE_NOCLOCK);
8789 set_task_cpu(p, env->dst_cpu);
8793 * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
8794 * part of active balancing operations within "domain".
8796 * Returns a task if successful and NULL otherwise.
8798 static struct task_struct *detach_one_task(struct lb_env *env)
8800 struct task_struct *p;
8802 lockdep_assert_rq_held(env->src_rq);
8804 list_for_each_entry_reverse(p,
8805 &env->src_rq->cfs_tasks, se.group_node) {
8806 if (!can_migrate_task(p, env))
8809 detach_task(p, env);
8812 * Right now, this is only the second place where
8813 * lb_gained[env->idle] is updated (other is detach_tasks)
8814 * so we can safely collect stats here rather than
8815 * inside detach_tasks().
8817 schedstat_inc(env->sd->lb_gained[env->idle]);
8824 * detach_tasks() -- tries to detach up to imbalance load/util/tasks from
8825 * busiest_rq, as part of a balancing operation within domain "sd".
8827 * Returns number of detached tasks if successful and 0 otherwise.
8829 static int detach_tasks(struct lb_env *env)
8831 struct list_head *tasks = &env->src_rq->cfs_tasks;
8832 unsigned long util, load;
8833 struct task_struct *p;
8836 lockdep_assert_rq_held(env->src_rq);
8839 * Source run queue has been emptied by another CPU, clear
8840 * LBF_ALL_PINNED flag as we will not test any task.
8842 if (env->src_rq->nr_running <= 1) {
8843 env->flags &= ~LBF_ALL_PINNED;
8847 if (env->imbalance <= 0)
8850 while (!list_empty(tasks)) {
8852 * We don't want to steal all, otherwise we may be treated likewise,
8853 * which could at worst lead to a livelock crash.
8855 if (env->idle != CPU_NOT_IDLE && env->src_rq->nr_running <= 1)
8860 * We've more or less seen every task there is, call it quits
8861 * unless we haven't found any movable task yet.
8863 if (env->loop > env->loop_max &&
8864 !(env->flags & LBF_ALL_PINNED))
8867 /* take a breather every nr_migrate tasks */
8868 if (env->loop > env->loop_break) {
8869 env->loop_break += SCHED_NR_MIGRATE_BREAK;
8870 env->flags |= LBF_NEED_BREAK;
8874 p = list_last_entry(tasks, struct task_struct, se.group_node);
8876 if (!can_migrate_task(p, env))
8879 switch (env->migration_type) {
8882 * Depending of the number of CPUs and tasks and the
8883 * cgroup hierarchy, task_h_load() can return a null
8884 * value. Make sure that env->imbalance decreases
8885 * otherwise detach_tasks() will stop only after
8886 * detaching up to loop_max tasks.
8888 load = max_t(unsigned long, task_h_load(p), 1);
8890 if (sched_feat(LB_MIN) &&
8891 load < 16 && !env->sd->nr_balance_failed)
8895 * Make sure that we don't migrate too much load.
8896 * Nevertheless, let relax the constraint if
8897 * scheduler fails to find a good waiting task to
8900 if (shr_bound(load, env->sd->nr_balance_failed) > env->imbalance)
8903 env->imbalance -= load;
8907 util = task_util_est(p);
8909 if (util > env->imbalance)
8912 env->imbalance -= util;
8919 case migrate_misfit:
8920 /* This is not a misfit task */
8921 if (task_fits_cpu(p, env->src_cpu))
8928 detach_task(p, env);
8929 list_add(&p->se.group_node, &env->tasks);
8933 #ifdef CONFIG_PREEMPTION
8935 * NEWIDLE balancing is a source of latency, so preemptible
8936 * kernels will stop after the first task is detached to minimize
8937 * the critical section.
8939 if (env->idle == CPU_NEWLY_IDLE)
8944 * We only want to steal up to the prescribed amount of
8947 if (env->imbalance <= 0)
8952 list_move(&p->se.group_node, tasks);
8956 * Right now, this is one of only two places we collect this stat
8957 * so we can safely collect detach_one_task() stats here rather
8958 * than inside detach_one_task().
8960 schedstat_add(env->sd->lb_gained[env->idle], detached);
8966 * attach_task() -- attach the task detached by detach_task() to its new rq.
8968 static void attach_task(struct rq *rq, struct task_struct *p)
8970 lockdep_assert_rq_held(rq);
8972 WARN_ON_ONCE(task_rq(p) != rq);
8973 activate_task(rq, p, ENQUEUE_NOCLOCK);
8974 check_preempt_curr(rq, p, 0);
8978 * attach_one_task() -- attaches the task returned from detach_one_task() to
8981 static void attach_one_task(struct rq *rq, struct task_struct *p)
8986 update_rq_clock(rq);
8992 * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
8995 static void attach_tasks(struct lb_env *env)
8997 struct list_head *tasks = &env->tasks;
8998 struct task_struct *p;
9001 rq_lock(env->dst_rq, &rf);
9002 update_rq_clock(env->dst_rq);
9004 while (!list_empty(tasks)) {
9005 p = list_first_entry(tasks, struct task_struct, se.group_node);
9006 list_del_init(&p->se.group_node);
9008 attach_task(env->dst_rq, p);
9011 rq_unlock(env->dst_rq, &rf);
9014 #ifdef CONFIG_NO_HZ_COMMON
9015 static inline bool cfs_rq_has_blocked(struct cfs_rq *cfs_rq)
9017 if (cfs_rq->avg.load_avg)
9020 if (cfs_rq->avg.util_avg)
9026 static inline bool others_have_blocked(struct rq *rq)
9028 if (READ_ONCE(rq->avg_rt.util_avg))
9031 if (READ_ONCE(rq->avg_dl.util_avg))
9034 if (thermal_load_avg(rq))
9037 #ifdef CONFIG_HAVE_SCHED_AVG_IRQ
9038 if (READ_ONCE(rq->avg_irq.util_avg))
9045 static inline void update_blocked_load_tick(struct rq *rq)
9047 WRITE_ONCE(rq->last_blocked_load_update_tick, jiffies);
9050 static inline void update_blocked_load_status(struct rq *rq, bool has_blocked)
9053 rq->has_blocked_load = 0;
9056 static inline bool cfs_rq_has_blocked(struct cfs_rq *cfs_rq) { return false; }
9057 static inline bool others_have_blocked(struct rq *rq) { return false; }
9058 static inline void update_blocked_load_tick(struct rq *rq) {}
9059 static inline void update_blocked_load_status(struct rq *rq, bool has_blocked) {}
9062 static bool __update_blocked_others(struct rq *rq, bool *done)
9064 const struct sched_class *curr_class;
9065 u64 now = rq_clock_pelt(rq);
9066 unsigned long thermal_pressure;
9070 * update_load_avg() can call cpufreq_update_util(). Make sure that RT,
9071 * DL and IRQ signals have been updated before updating CFS.
9073 curr_class = rq->curr->sched_class;
9075 thermal_pressure = arch_scale_thermal_pressure(cpu_of(rq));
9077 decayed = update_rt_rq_load_avg(now, rq, curr_class == &rt_sched_class) |
9078 update_dl_rq_load_avg(now, rq, curr_class == &dl_sched_class) |
9079 update_thermal_load_avg(rq_clock_thermal(rq), rq, thermal_pressure) |
9080 update_irq_load_avg(rq, 0);
9082 if (others_have_blocked(rq))
9088 #ifdef CONFIG_FAIR_GROUP_SCHED
9090 static bool __update_blocked_fair(struct rq *rq, bool *done)
9092 struct cfs_rq *cfs_rq, *pos;
9093 bool decayed = false;
9094 int cpu = cpu_of(rq);
9097 * Iterates the task_group tree in a bottom up fashion, see
9098 * list_add_leaf_cfs_rq() for details.
9100 for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) {
9101 struct sched_entity *se;
9103 if (update_cfs_rq_load_avg(cfs_rq_clock_pelt(cfs_rq), cfs_rq)) {
9104 update_tg_load_avg(cfs_rq);
9106 if (cfs_rq->nr_running == 0)
9107 update_idle_cfs_rq_clock_pelt(cfs_rq);
9109 if (cfs_rq == &rq->cfs)
9113 /* Propagate pending load changes to the parent, if any: */
9114 se = cfs_rq->tg->se[cpu];
9115 if (se && !skip_blocked_update(se))
9116 update_load_avg(cfs_rq_of(se), se, UPDATE_TG);
9119 * There can be a lot of idle CPU cgroups. Don't let fully
9120 * decayed cfs_rqs linger on the list.
9122 if (cfs_rq_is_decayed(cfs_rq))
9123 list_del_leaf_cfs_rq(cfs_rq);
9125 /* Don't need periodic decay once load/util_avg are null */
9126 if (cfs_rq_has_blocked(cfs_rq))
9134 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
9135 * This needs to be done in a top-down fashion because the load of a child
9136 * group is a fraction of its parents load.
9138 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
9140 struct rq *rq = rq_of(cfs_rq);
9141 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
9142 unsigned long now = jiffies;
9145 if (cfs_rq->last_h_load_update == now)
9148 WRITE_ONCE(cfs_rq->h_load_next, NULL);
9149 for_each_sched_entity(se) {
9150 cfs_rq = cfs_rq_of(se);
9151 WRITE_ONCE(cfs_rq->h_load_next, se);
9152 if (cfs_rq->last_h_load_update == now)
9157 cfs_rq->h_load = cfs_rq_load_avg(cfs_rq);
9158 cfs_rq->last_h_load_update = now;
9161 while ((se = READ_ONCE(cfs_rq->h_load_next)) != NULL) {
9162 load = cfs_rq->h_load;
9163 load = div64_ul(load * se->avg.load_avg,
9164 cfs_rq_load_avg(cfs_rq) + 1);
9165 cfs_rq = group_cfs_rq(se);
9166 cfs_rq->h_load = load;
9167 cfs_rq->last_h_load_update = now;
9171 static unsigned long task_h_load(struct task_struct *p)
9173 struct cfs_rq *cfs_rq = task_cfs_rq(p);
9175 update_cfs_rq_h_load(cfs_rq);
9176 return div64_ul(p->se.avg.load_avg * cfs_rq->h_load,
9177 cfs_rq_load_avg(cfs_rq) + 1);
9180 static bool __update_blocked_fair(struct rq *rq, bool *done)
9182 struct cfs_rq *cfs_rq = &rq->cfs;
9185 decayed = update_cfs_rq_load_avg(cfs_rq_clock_pelt(cfs_rq), cfs_rq);
9186 if (cfs_rq_has_blocked(cfs_rq))
9192 static unsigned long task_h_load(struct task_struct *p)
9194 return p->se.avg.load_avg;
9198 static void update_blocked_averages(int cpu)
9200 bool decayed = false, done = true;
9201 struct rq *rq = cpu_rq(cpu);
9204 rq_lock_irqsave(rq, &rf);
9205 update_blocked_load_tick(rq);
9206 update_rq_clock(rq);
9208 decayed |= __update_blocked_others(rq, &done);
9209 decayed |= __update_blocked_fair(rq, &done);
9211 update_blocked_load_status(rq, !done);
9213 cpufreq_update_util(rq, 0);
9214 rq_unlock_irqrestore(rq, &rf);
9217 /********** Helpers for find_busiest_group ************************/
9220 * sg_lb_stats - stats of a sched_group required for load_balancing
9222 struct sg_lb_stats {
9223 unsigned long avg_load; /*Avg load across the CPUs of the group */
9224 unsigned long group_load; /* Total load over the CPUs of the group */
9225 unsigned long group_capacity;
9226 unsigned long group_util; /* Total utilization over the CPUs of the group */
9227 unsigned long group_runnable; /* Total runnable time over the CPUs of the group */
9228 unsigned int sum_nr_running; /* Nr of tasks running in the group */
9229 unsigned int sum_h_nr_running; /* Nr of CFS tasks running in the group */
9230 unsigned int idle_cpus;
9231 unsigned int group_weight;
9232 enum group_type group_type;
9233 unsigned int group_asym_packing; /* Tasks should be moved to preferred CPU */
9234 unsigned int group_smt_balance; /* Task on busy SMT be moved */
9235 unsigned long group_misfit_task_load; /* A CPU has a task too big for its capacity */
9236 #ifdef CONFIG_NUMA_BALANCING
9237 unsigned int nr_numa_running;
9238 unsigned int nr_preferred_running;
9243 * sd_lb_stats - Structure to store the statistics of a sched_domain
9244 * during load balancing.
9246 struct sd_lb_stats {
9247 struct sched_group *busiest; /* Busiest group in this sd */
9248 struct sched_group *local; /* Local group in this sd */
9249 unsigned long total_load; /* Total load of all groups in sd */
9250 unsigned long total_capacity; /* Total capacity of all groups in sd */
9251 unsigned long avg_load; /* Average load across all groups in sd */
9252 unsigned int prefer_sibling; /* tasks should go to sibling first */
9254 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
9255 struct sg_lb_stats local_stat; /* Statistics of the local group */
9258 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
9261 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
9262 * local_stat because update_sg_lb_stats() does a full clear/assignment.
9263 * We must however set busiest_stat::group_type and
9264 * busiest_stat::idle_cpus to the worst busiest group because
9265 * update_sd_pick_busiest() reads these before assignment.
9267 *sds = (struct sd_lb_stats){
9271 .total_capacity = 0UL,
9273 .idle_cpus = UINT_MAX,
9274 .group_type = group_has_spare,
9279 static unsigned long scale_rt_capacity(int cpu)
9281 struct rq *rq = cpu_rq(cpu);
9282 unsigned long max = arch_scale_cpu_capacity(cpu);
9283 unsigned long used, free;
9286 irq = cpu_util_irq(rq);
9288 if (unlikely(irq >= max))
9292 * avg_rt.util_avg and avg_dl.util_avg track binary signals
9293 * (running and not running) with weights 0 and 1024 respectively.
9294 * avg_thermal.load_avg tracks thermal pressure and the weighted
9295 * average uses the actual delta max capacity(load).
9297 used = READ_ONCE(rq->avg_rt.util_avg);
9298 used += READ_ONCE(rq->avg_dl.util_avg);
9299 used += thermal_load_avg(rq);
9301 if (unlikely(used >= max))
9306 return scale_irq_capacity(free, irq, max);
9309 static void update_cpu_capacity(struct sched_domain *sd, int cpu)
9311 unsigned long capacity = scale_rt_capacity(cpu);
9312 struct sched_group *sdg = sd->groups;
9314 cpu_rq(cpu)->cpu_capacity_orig = arch_scale_cpu_capacity(cpu);
9319 cpu_rq(cpu)->cpu_capacity = capacity;
9320 trace_sched_cpu_capacity_tp(cpu_rq(cpu));
9322 sdg->sgc->capacity = capacity;
9323 sdg->sgc->min_capacity = capacity;
9324 sdg->sgc->max_capacity = capacity;
9327 void update_group_capacity(struct sched_domain *sd, int cpu)
9329 struct sched_domain *child = sd->child;
9330 struct sched_group *group, *sdg = sd->groups;
9331 unsigned long capacity, min_capacity, max_capacity;
9332 unsigned long interval;
9334 interval = msecs_to_jiffies(sd->balance_interval);
9335 interval = clamp(interval, 1UL, max_load_balance_interval);
9336 sdg->sgc->next_update = jiffies + interval;
9339 update_cpu_capacity(sd, cpu);
9344 min_capacity = ULONG_MAX;
9347 if (child->flags & SD_OVERLAP) {
9349 * SD_OVERLAP domains cannot assume that child groups
9350 * span the current group.
9353 for_each_cpu(cpu, sched_group_span(sdg)) {
9354 unsigned long cpu_cap = capacity_of(cpu);
9356 capacity += cpu_cap;
9357 min_capacity = min(cpu_cap, min_capacity);
9358 max_capacity = max(cpu_cap, max_capacity);
9362 * !SD_OVERLAP domains can assume that child groups
9363 * span the current group.
9366 group = child->groups;
9368 struct sched_group_capacity *sgc = group->sgc;
9370 capacity += sgc->capacity;
9371 min_capacity = min(sgc->min_capacity, min_capacity);
9372 max_capacity = max(sgc->max_capacity, max_capacity);
9373 group = group->next;
9374 } while (group != child->groups);
9377 sdg->sgc->capacity = capacity;
9378 sdg->sgc->min_capacity = min_capacity;
9379 sdg->sgc->max_capacity = max_capacity;
9383 * Check whether the capacity of the rq has been noticeably reduced by side
9384 * activity. The imbalance_pct is used for the threshold.
9385 * Return true is the capacity is reduced
9388 check_cpu_capacity(struct rq *rq, struct sched_domain *sd)
9390 return ((rq->cpu_capacity * sd->imbalance_pct) <
9391 (rq->cpu_capacity_orig * 100));
9395 * Check whether a rq has a misfit task and if it looks like we can actually
9396 * help that task: we can migrate the task to a CPU of higher capacity, or
9397 * the task's current CPU is heavily pressured.
9399 static inline int check_misfit_status(struct rq *rq, struct sched_domain *sd)
9401 return rq->misfit_task_load &&
9402 (rq->cpu_capacity_orig < rq->rd->max_cpu_capacity ||
9403 check_cpu_capacity(rq, sd));
9407 * Group imbalance indicates (and tries to solve) the problem where balancing
9408 * groups is inadequate due to ->cpus_ptr constraints.
9410 * Imagine a situation of two groups of 4 CPUs each and 4 tasks each with a
9411 * cpumask covering 1 CPU of the first group and 3 CPUs of the second group.
9414 * { 0 1 2 3 } { 4 5 6 7 }
9417 * If we were to balance group-wise we'd place two tasks in the first group and
9418 * two tasks in the second group. Clearly this is undesired as it will overload
9419 * cpu 3 and leave one of the CPUs in the second group unused.
9421 * The current solution to this issue is detecting the skew in the first group
9422 * by noticing the lower domain failed to reach balance and had difficulty
9423 * moving tasks due to affinity constraints.
9425 * When this is so detected; this group becomes a candidate for busiest; see
9426 * update_sd_pick_busiest(). And calculate_imbalance() and
9427 * find_busiest_group() avoid some of the usual balance conditions to allow it
9428 * to create an effective group imbalance.
9430 * This is a somewhat tricky proposition since the next run might not find the
9431 * group imbalance and decide the groups need to be balanced again. A most
9432 * subtle and fragile situation.
9435 static inline int sg_imbalanced(struct sched_group *group)
9437 return group->sgc->imbalance;
9441 * group_has_capacity returns true if the group has spare capacity that could
9442 * be used by some tasks.
9443 * We consider that a group has spare capacity if the number of task is
9444 * smaller than the number of CPUs or if the utilization is lower than the
9445 * available capacity for CFS tasks.
9446 * For the latter, we use a threshold to stabilize the state, to take into
9447 * account the variance of the tasks' load and to return true if the available
9448 * capacity in meaningful for the load balancer.
9449 * As an example, an available capacity of 1% can appear but it doesn't make
9450 * any benefit for the load balance.
9453 group_has_capacity(unsigned int imbalance_pct, struct sg_lb_stats *sgs)
9455 if (sgs->sum_nr_running < sgs->group_weight)
9458 if ((sgs->group_capacity * imbalance_pct) <
9459 (sgs->group_runnable * 100))
9462 if ((sgs->group_capacity * 100) >
9463 (sgs->group_util * imbalance_pct))
9470 * group_is_overloaded returns true if the group has more tasks than it can
9472 * group_is_overloaded is not equals to !group_has_capacity because a group
9473 * with the exact right number of tasks, has no more spare capacity but is not
9474 * overloaded so both group_has_capacity and group_is_overloaded return
9478 group_is_overloaded(unsigned int imbalance_pct, struct sg_lb_stats *sgs)
9480 if (sgs->sum_nr_running <= sgs->group_weight)
9483 if ((sgs->group_capacity * 100) <
9484 (sgs->group_util * imbalance_pct))
9487 if ((sgs->group_capacity * imbalance_pct) <
9488 (sgs->group_runnable * 100))
9495 group_type group_classify(unsigned int imbalance_pct,
9496 struct sched_group *group,
9497 struct sg_lb_stats *sgs)
9499 if (group_is_overloaded(imbalance_pct, sgs))
9500 return group_overloaded;
9502 if (sg_imbalanced(group))
9503 return group_imbalanced;
9505 if (sgs->group_asym_packing)
9506 return group_asym_packing;
9508 if (sgs->group_smt_balance)
9509 return group_smt_balance;
9511 if (sgs->group_misfit_task_load)
9512 return group_misfit_task;
9514 if (!group_has_capacity(imbalance_pct, sgs))
9515 return group_fully_busy;
9517 return group_has_spare;
9521 * sched_use_asym_prio - Check whether asym_packing priority must be used
9522 * @sd: The scheduling domain of the load balancing
9525 * Always use CPU priority when balancing load between SMT siblings. When
9526 * balancing load between cores, it is not sufficient that @cpu is idle. Only
9527 * use CPU priority if the whole core is idle.
9529 * Returns: True if the priority of @cpu must be followed. False otherwise.
9531 static bool sched_use_asym_prio(struct sched_domain *sd, int cpu)
9533 if (!sched_smt_active())
9536 return sd->flags & SD_SHARE_CPUCAPACITY || is_core_idle(cpu);
9540 * sched_asym - Check if the destination CPU can do asym_packing load balance
9541 * @env: The load balancing environment
9542 * @sds: Load-balancing data with statistics of the local group
9543 * @sgs: Load-balancing statistics of the candidate busiest group
9544 * @group: The candidate busiest group
9546 * @env::dst_cpu can do asym_packing if it has higher priority than the
9547 * preferred CPU of @group.
9549 * SMT is a special case. If we are balancing load between cores, @env::dst_cpu
9550 * can do asym_packing balance only if all its SMT siblings are idle. Also, it
9551 * can only do it if @group is an SMT group and has exactly on busy CPU. Larger
9552 * imbalances in the number of CPUS are dealt with in find_busiest_group().
9554 * If we are balancing load within an SMT core, or at DIE domain level, always
9557 * Return: true if @env::dst_cpu can do with asym_packing load balance. False
9561 sched_asym(struct lb_env *env, struct sd_lb_stats *sds, struct sg_lb_stats *sgs,
9562 struct sched_group *group)
9564 /* Ensure that the whole local core is idle, if applicable. */
9565 if (!sched_use_asym_prio(env->sd, env->dst_cpu))
9569 * CPU priorities does not make sense for SMT cores with more than one
9572 if (group->flags & SD_SHARE_CPUCAPACITY) {
9573 if (sgs->group_weight - sgs->idle_cpus != 1)
9577 return sched_asym_prefer(env->dst_cpu, group->asym_prefer_cpu);
9580 /* One group has more than one SMT CPU while the other group does not */
9581 static inline bool smt_vs_nonsmt_groups(struct sched_group *sg1,
9582 struct sched_group *sg2)
9587 return (sg1->flags & SD_SHARE_CPUCAPACITY) !=
9588 (sg2->flags & SD_SHARE_CPUCAPACITY);
9591 static inline bool smt_balance(struct lb_env *env, struct sg_lb_stats *sgs,
9592 struct sched_group *group)
9594 if (env->idle == CPU_NOT_IDLE)
9598 * For SMT source group, it is better to move a task
9599 * to a CPU that doesn't have multiple tasks sharing its CPU capacity.
9600 * Note that if a group has a single SMT, SD_SHARE_CPUCAPACITY
9603 if (group->flags & SD_SHARE_CPUCAPACITY &&
9604 sgs->sum_h_nr_running > 1)
9610 static inline long sibling_imbalance(struct lb_env *env,
9611 struct sd_lb_stats *sds,
9612 struct sg_lb_stats *busiest,
9613 struct sg_lb_stats *local)
9615 int ncores_busiest, ncores_local;
9618 if (env->idle == CPU_NOT_IDLE || !busiest->sum_nr_running)
9621 ncores_busiest = sds->busiest->cores;
9622 ncores_local = sds->local->cores;
9624 if (ncores_busiest == ncores_local) {
9625 imbalance = busiest->sum_nr_running;
9626 lsub_positive(&imbalance, local->sum_nr_running);
9630 /* Balance such that nr_running/ncores ratio are same on both groups */
9631 imbalance = ncores_local * busiest->sum_nr_running;
9632 lsub_positive(&imbalance, ncores_busiest * local->sum_nr_running);
9633 /* Normalize imbalance and do rounding on normalization */
9634 imbalance = 2 * imbalance + ncores_local + ncores_busiest;
9635 imbalance /= ncores_local + ncores_busiest;
9637 /* Take advantage of resource in an empty sched group */
9638 if (imbalance <= 1 && local->sum_nr_running == 0 &&
9639 busiest->sum_nr_running > 1)
9646 sched_reduced_capacity(struct rq *rq, struct sched_domain *sd)
9649 * When there is more than 1 task, the group_overloaded case already
9650 * takes care of cpu with reduced capacity
9652 if (rq->cfs.h_nr_running != 1)
9655 return check_cpu_capacity(rq, sd);
9659 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
9660 * @env: The load balancing environment.
9661 * @sds: Load-balancing data with statistics of the local group.
9662 * @group: sched_group whose statistics are to be updated.
9663 * @sgs: variable to hold the statistics for this group.
9664 * @sg_status: Holds flag indicating the status of the sched_group
9666 static inline void update_sg_lb_stats(struct lb_env *env,
9667 struct sd_lb_stats *sds,
9668 struct sched_group *group,
9669 struct sg_lb_stats *sgs,
9672 int i, nr_running, local_group;
9674 memset(sgs, 0, sizeof(*sgs));
9676 local_group = group == sds->local;
9678 for_each_cpu_and(i, sched_group_span(group), env->cpus) {
9679 struct rq *rq = cpu_rq(i);
9680 unsigned long load = cpu_load(rq);
9682 sgs->group_load += load;
9683 sgs->group_util += cpu_util_cfs(i);
9684 sgs->group_runnable += cpu_runnable(rq);
9685 sgs->sum_h_nr_running += rq->cfs.h_nr_running;
9687 nr_running = rq->nr_running;
9688 sgs->sum_nr_running += nr_running;
9691 *sg_status |= SG_OVERLOAD;
9693 if (cpu_overutilized(i))
9694 *sg_status |= SG_OVERUTILIZED;
9696 #ifdef CONFIG_NUMA_BALANCING
9697 sgs->nr_numa_running += rq->nr_numa_running;
9698 sgs->nr_preferred_running += rq->nr_preferred_running;
9701 * No need to call idle_cpu() if nr_running is not 0
9703 if (!nr_running && idle_cpu(i)) {
9705 /* Idle cpu can't have misfit task */
9712 if (env->sd->flags & SD_ASYM_CPUCAPACITY) {
9713 /* Check for a misfit task on the cpu */
9714 if (sgs->group_misfit_task_load < rq->misfit_task_load) {
9715 sgs->group_misfit_task_load = rq->misfit_task_load;
9716 *sg_status |= SG_OVERLOAD;
9718 } else if ((env->idle != CPU_NOT_IDLE) &&
9719 sched_reduced_capacity(rq, env->sd)) {
9720 /* Check for a task running on a CPU with reduced capacity */
9721 if (sgs->group_misfit_task_load < load)
9722 sgs->group_misfit_task_load = load;
9726 sgs->group_capacity = group->sgc->capacity;
9728 sgs->group_weight = group->group_weight;
9730 /* Check if dst CPU is idle and preferred to this group */
9731 if (!local_group && env->sd->flags & SD_ASYM_PACKING &&
9732 env->idle != CPU_NOT_IDLE && sgs->sum_h_nr_running &&
9733 sched_asym(env, sds, sgs, group)) {
9734 sgs->group_asym_packing = 1;
9737 /* Check for loaded SMT group to be balanced to dst CPU */
9738 if (!local_group && smt_balance(env, sgs, group))
9739 sgs->group_smt_balance = 1;
9741 sgs->group_type = group_classify(env->sd->imbalance_pct, group, sgs);
9743 /* Computing avg_load makes sense only when group is overloaded */
9744 if (sgs->group_type == group_overloaded)
9745 sgs->avg_load = (sgs->group_load * SCHED_CAPACITY_SCALE) /
9746 sgs->group_capacity;
9750 * update_sd_pick_busiest - return 1 on busiest group
9751 * @env: The load balancing environment.
9752 * @sds: sched_domain statistics
9753 * @sg: sched_group candidate to be checked for being the busiest
9754 * @sgs: sched_group statistics
9756 * Determine if @sg is a busier group than the previously selected
9759 * Return: %true if @sg is a busier group than the previously selected
9760 * busiest group. %false otherwise.
9762 static bool update_sd_pick_busiest(struct lb_env *env,
9763 struct sd_lb_stats *sds,
9764 struct sched_group *sg,
9765 struct sg_lb_stats *sgs)
9767 struct sg_lb_stats *busiest = &sds->busiest_stat;
9769 /* Make sure that there is at least one task to pull */
9770 if (!sgs->sum_h_nr_running)
9774 * Don't try to pull misfit tasks we can't help.
9775 * We can use max_capacity here as reduction in capacity on some
9776 * CPUs in the group should either be possible to resolve
9777 * internally or be covered by avg_load imbalance (eventually).
9779 if ((env->sd->flags & SD_ASYM_CPUCAPACITY) &&
9780 (sgs->group_type == group_misfit_task) &&
9781 (!capacity_greater(capacity_of(env->dst_cpu), sg->sgc->max_capacity) ||
9782 sds->local_stat.group_type != group_has_spare))
9785 if (sgs->group_type > busiest->group_type)
9788 if (sgs->group_type < busiest->group_type)
9792 * The candidate and the current busiest group are the same type of
9793 * group. Let check which one is the busiest according to the type.
9796 switch (sgs->group_type) {
9797 case group_overloaded:
9798 /* Select the overloaded group with highest avg_load. */
9799 if (sgs->avg_load <= busiest->avg_load)
9803 case group_imbalanced:
9805 * Select the 1st imbalanced group as we don't have any way to
9806 * choose one more than another.
9810 case group_asym_packing:
9811 /* Prefer to move from lowest priority CPU's work */
9812 if (sched_asym_prefer(sg->asym_prefer_cpu, sds->busiest->asym_prefer_cpu))
9816 case group_misfit_task:
9818 * If we have more than one misfit sg go with the biggest
9821 if (sgs->group_misfit_task_load < busiest->group_misfit_task_load)
9825 case group_smt_balance:
9827 * Check if we have spare CPUs on either SMT group to
9828 * choose has spare or fully busy handling.
9830 if (sgs->idle_cpus != 0 || busiest->idle_cpus != 0)
9835 case group_fully_busy:
9837 * Select the fully busy group with highest avg_load. In
9838 * theory, there is no need to pull task from such kind of
9839 * group because tasks have all compute capacity that they need
9840 * but we can still improve the overall throughput by reducing
9841 * contention when accessing shared HW resources.
9843 * XXX for now avg_load is not computed and always 0 so we
9844 * select the 1st one, except if @sg is composed of SMT
9848 if (sgs->avg_load < busiest->avg_load)
9851 if (sgs->avg_load == busiest->avg_load) {
9853 * SMT sched groups need more help than non-SMT groups.
9854 * If @sg happens to also be SMT, either choice is good.
9856 if (sds->busiest->flags & SD_SHARE_CPUCAPACITY)
9862 case group_has_spare:
9864 * Do not pick sg with SMT CPUs over sg with pure CPUs,
9865 * as we do not want to pull task off SMT core with one task
9866 * and make the core idle.
9868 if (smt_vs_nonsmt_groups(sds->busiest, sg)) {
9869 if (sg->flags & SD_SHARE_CPUCAPACITY && sgs->sum_h_nr_running <= 1)
9877 * Select not overloaded group with lowest number of idle cpus
9878 * and highest number of running tasks. We could also compare
9879 * the spare capacity which is more stable but it can end up
9880 * that the group has less spare capacity but finally more idle
9881 * CPUs which means less opportunity to pull tasks.
9883 if (sgs->idle_cpus > busiest->idle_cpus)
9885 else if ((sgs->idle_cpus == busiest->idle_cpus) &&
9886 (sgs->sum_nr_running <= busiest->sum_nr_running))
9893 * Candidate sg has no more than one task per CPU and has higher
9894 * per-CPU capacity. Migrating tasks to less capable CPUs may harm
9895 * throughput. Maximize throughput, power/energy consequences are not
9898 if ((env->sd->flags & SD_ASYM_CPUCAPACITY) &&
9899 (sgs->group_type <= group_fully_busy) &&
9900 (capacity_greater(sg->sgc->min_capacity, capacity_of(env->dst_cpu))))
9906 #ifdef CONFIG_NUMA_BALANCING
9907 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
9909 if (sgs->sum_h_nr_running > sgs->nr_numa_running)
9911 if (sgs->sum_h_nr_running > sgs->nr_preferred_running)
9916 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
9918 if (rq->nr_running > rq->nr_numa_running)
9920 if (rq->nr_running > rq->nr_preferred_running)
9925 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
9930 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
9934 #endif /* CONFIG_NUMA_BALANCING */
9940 * task_running_on_cpu - return 1 if @p is running on @cpu.
9943 static unsigned int task_running_on_cpu(int cpu, struct task_struct *p)
9945 /* Task has no contribution or is new */
9946 if (cpu != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time))
9949 if (task_on_rq_queued(p))
9956 * idle_cpu_without - would a given CPU be idle without p ?
9957 * @cpu: the processor on which idleness is tested.
9958 * @p: task which should be ignored.
9960 * Return: 1 if the CPU would be idle. 0 otherwise.
9962 static int idle_cpu_without(int cpu, struct task_struct *p)
9964 struct rq *rq = cpu_rq(cpu);
9966 if (rq->curr != rq->idle && rq->curr != p)
9970 * rq->nr_running can't be used but an updated version without the
9971 * impact of p on cpu must be used instead. The updated nr_running
9972 * be computed and tested before calling idle_cpu_without().
9976 if (rq->ttwu_pending)
9984 * update_sg_wakeup_stats - Update sched_group's statistics for wakeup.
9985 * @sd: The sched_domain level to look for idlest group.
9986 * @group: sched_group whose statistics are to be updated.
9987 * @sgs: variable to hold the statistics for this group.
9988 * @p: The task for which we look for the idlest group/CPU.
9990 static inline void update_sg_wakeup_stats(struct sched_domain *sd,
9991 struct sched_group *group,
9992 struct sg_lb_stats *sgs,
9993 struct task_struct *p)
9997 memset(sgs, 0, sizeof(*sgs));
9999 /* Assume that task can't fit any CPU of the group */
10000 if (sd->flags & SD_ASYM_CPUCAPACITY)
10001 sgs->group_misfit_task_load = 1;
10003 for_each_cpu(i, sched_group_span(group)) {
10004 struct rq *rq = cpu_rq(i);
10005 unsigned int local;
10007 sgs->group_load += cpu_load_without(rq, p);
10008 sgs->group_util += cpu_util_without(i, p);
10009 sgs->group_runnable += cpu_runnable_without(rq, p);
10010 local = task_running_on_cpu(i, p);
10011 sgs->sum_h_nr_running += rq->cfs.h_nr_running - local;
10013 nr_running = rq->nr_running - local;
10014 sgs->sum_nr_running += nr_running;
10017 * No need to call idle_cpu_without() if nr_running is not 0
10019 if (!nr_running && idle_cpu_without(i, p))
10022 /* Check if task fits in the CPU */
10023 if (sd->flags & SD_ASYM_CPUCAPACITY &&
10024 sgs->group_misfit_task_load &&
10025 task_fits_cpu(p, i))
10026 sgs->group_misfit_task_load = 0;
10030 sgs->group_capacity = group->sgc->capacity;
10032 sgs->group_weight = group->group_weight;
10034 sgs->group_type = group_classify(sd->imbalance_pct, group, sgs);
10037 * Computing avg_load makes sense only when group is fully busy or
10040 if (sgs->group_type == group_fully_busy ||
10041 sgs->group_type == group_overloaded)
10042 sgs->avg_load = (sgs->group_load * SCHED_CAPACITY_SCALE) /
10043 sgs->group_capacity;
10046 static bool update_pick_idlest(struct sched_group *idlest,
10047 struct sg_lb_stats *idlest_sgs,
10048 struct sched_group *group,
10049 struct sg_lb_stats *sgs)
10051 if (sgs->group_type < idlest_sgs->group_type)
10054 if (sgs->group_type > idlest_sgs->group_type)
10058 * The candidate and the current idlest group are the same type of
10059 * group. Let check which one is the idlest according to the type.
10062 switch (sgs->group_type) {
10063 case group_overloaded:
10064 case group_fully_busy:
10065 /* Select the group with lowest avg_load. */
10066 if (idlest_sgs->avg_load <= sgs->avg_load)
10070 case group_imbalanced:
10071 case group_asym_packing:
10072 case group_smt_balance:
10073 /* Those types are not used in the slow wakeup path */
10076 case group_misfit_task:
10077 /* Select group with the highest max capacity */
10078 if (idlest->sgc->max_capacity >= group->sgc->max_capacity)
10082 case group_has_spare:
10083 /* Select group with most idle CPUs */
10084 if (idlest_sgs->idle_cpus > sgs->idle_cpus)
10087 /* Select group with lowest group_util */
10088 if (idlest_sgs->idle_cpus == sgs->idle_cpus &&
10089 idlest_sgs->group_util <= sgs->group_util)
10099 * find_idlest_group() finds and returns the least busy CPU group within the
10102 * Assumes p is allowed on at least one CPU in sd.
10104 static struct sched_group *
10105 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
10107 struct sched_group *idlest = NULL, *local = NULL, *group = sd->groups;
10108 struct sg_lb_stats local_sgs, tmp_sgs;
10109 struct sg_lb_stats *sgs;
10110 unsigned long imbalance;
10111 struct sg_lb_stats idlest_sgs = {
10112 .avg_load = UINT_MAX,
10113 .group_type = group_overloaded,
10119 /* Skip over this group if it has no CPUs allowed */
10120 if (!cpumask_intersects(sched_group_span(group),
10124 /* Skip over this group if no cookie matched */
10125 if (!sched_group_cookie_match(cpu_rq(this_cpu), p, group))
10128 local_group = cpumask_test_cpu(this_cpu,
10129 sched_group_span(group));
10138 update_sg_wakeup_stats(sd, group, sgs, p);
10140 if (!local_group && update_pick_idlest(idlest, &idlest_sgs, group, sgs)) {
10145 } while (group = group->next, group != sd->groups);
10148 /* There is no idlest group to push tasks to */
10152 /* The local group has been skipped because of CPU affinity */
10157 * If the local group is idler than the selected idlest group
10158 * don't try and push the task.
10160 if (local_sgs.group_type < idlest_sgs.group_type)
10164 * If the local group is busier than the selected idlest group
10165 * try and push the task.
10167 if (local_sgs.group_type > idlest_sgs.group_type)
10170 switch (local_sgs.group_type) {
10171 case group_overloaded:
10172 case group_fully_busy:
10174 /* Calculate allowed imbalance based on load */
10175 imbalance = scale_load_down(NICE_0_LOAD) *
10176 (sd->imbalance_pct-100) / 100;
10179 * When comparing groups across NUMA domains, it's possible for
10180 * the local domain to be very lightly loaded relative to the
10181 * remote domains but "imbalance" skews the comparison making
10182 * remote CPUs look much more favourable. When considering
10183 * cross-domain, add imbalance to the load on the remote node
10184 * and consider staying local.
10187 if ((sd->flags & SD_NUMA) &&
10188 ((idlest_sgs.avg_load + imbalance) >= local_sgs.avg_load))
10192 * If the local group is less loaded than the selected
10193 * idlest group don't try and push any tasks.
10195 if (idlest_sgs.avg_load >= (local_sgs.avg_load + imbalance))
10198 if (100 * local_sgs.avg_load <= sd->imbalance_pct * idlest_sgs.avg_load)
10202 case group_imbalanced:
10203 case group_asym_packing:
10204 case group_smt_balance:
10205 /* Those type are not used in the slow wakeup path */
10208 case group_misfit_task:
10209 /* Select group with the highest max capacity */
10210 if (local->sgc->max_capacity >= idlest->sgc->max_capacity)
10214 case group_has_spare:
10216 if (sd->flags & SD_NUMA) {
10217 int imb_numa_nr = sd->imb_numa_nr;
10218 #ifdef CONFIG_NUMA_BALANCING
10221 * If there is spare capacity at NUMA, try to select
10222 * the preferred node
10224 if (cpu_to_node(this_cpu) == p->numa_preferred_nid)
10227 idlest_cpu = cpumask_first(sched_group_span(idlest));
10228 if (cpu_to_node(idlest_cpu) == p->numa_preferred_nid)
10230 #endif /* CONFIG_NUMA_BALANCING */
10232 * Otherwise, keep the task close to the wakeup source
10233 * and improve locality if the number of running tasks
10234 * would remain below threshold where an imbalance is
10235 * allowed while accounting for the possibility the
10236 * task is pinned to a subset of CPUs. If there is a
10237 * real need of migration, periodic load balance will
10240 if (p->nr_cpus_allowed != NR_CPUS) {
10241 struct cpumask *cpus = this_cpu_cpumask_var_ptr(select_rq_mask);
10243 cpumask_and(cpus, sched_group_span(local), p->cpus_ptr);
10244 imb_numa_nr = min(cpumask_weight(cpus), sd->imb_numa_nr);
10247 imbalance = abs(local_sgs.idle_cpus - idlest_sgs.idle_cpus);
10248 if (!adjust_numa_imbalance(imbalance,
10249 local_sgs.sum_nr_running + 1,
10254 #endif /* CONFIG_NUMA */
10257 * Select group with highest number of idle CPUs. We could also
10258 * compare the utilization which is more stable but it can end
10259 * up that the group has less spare capacity but finally more
10260 * idle CPUs which means more opportunity to run task.
10262 if (local_sgs.idle_cpus >= idlest_sgs.idle_cpus)
10270 static void update_idle_cpu_scan(struct lb_env *env,
10271 unsigned long sum_util)
10273 struct sched_domain_shared *sd_share;
10274 int llc_weight, pct;
10277 * Update the number of CPUs to scan in LLC domain, which could
10278 * be used as a hint in select_idle_cpu(). The update of sd_share
10279 * could be expensive because it is within a shared cache line.
10280 * So the write of this hint only occurs during periodic load
10281 * balancing, rather than CPU_NEWLY_IDLE, because the latter
10282 * can fire way more frequently than the former.
10284 if (!sched_feat(SIS_UTIL) || env->idle == CPU_NEWLY_IDLE)
10287 llc_weight = per_cpu(sd_llc_size, env->dst_cpu);
10288 if (env->sd->span_weight != llc_weight)
10291 sd_share = rcu_dereference(per_cpu(sd_llc_shared, env->dst_cpu));
10296 * The number of CPUs to search drops as sum_util increases, when
10297 * sum_util hits 85% or above, the scan stops.
10298 * The reason to choose 85% as the threshold is because this is the
10299 * imbalance_pct(117) when a LLC sched group is overloaded.
10301 * let y = SCHED_CAPACITY_SCALE - p * x^2 [1]
10302 * and y'= y / SCHED_CAPACITY_SCALE
10304 * x is the ratio of sum_util compared to the CPU capacity:
10305 * x = sum_util / (llc_weight * SCHED_CAPACITY_SCALE)
10306 * y' is the ratio of CPUs to be scanned in the LLC domain,
10307 * and the number of CPUs to scan is calculated by:
10309 * nr_scan = llc_weight * y' [2]
10311 * When x hits the threshold of overloaded, AKA, when
10312 * x = 100 / pct, y drops to 0. According to [1],
10313 * p should be SCHED_CAPACITY_SCALE * pct^2 / 10000
10315 * Scale x by SCHED_CAPACITY_SCALE:
10316 * x' = sum_util / llc_weight; [3]
10318 * and finally [1] becomes:
10319 * y = SCHED_CAPACITY_SCALE -
10320 * x'^2 * pct^2 / (10000 * SCHED_CAPACITY_SCALE) [4]
10325 do_div(x, llc_weight);
10328 pct = env->sd->imbalance_pct;
10329 tmp = x * x * pct * pct;
10330 do_div(tmp, 10000 * SCHED_CAPACITY_SCALE);
10331 tmp = min_t(long, tmp, SCHED_CAPACITY_SCALE);
10332 y = SCHED_CAPACITY_SCALE - tmp;
10336 do_div(y, SCHED_CAPACITY_SCALE);
10337 if ((int)y != sd_share->nr_idle_scan)
10338 WRITE_ONCE(sd_share->nr_idle_scan, (int)y);
10342 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
10343 * @env: The load balancing environment.
10344 * @sds: variable to hold the statistics for this sched_domain.
10347 static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
10349 struct sched_group *sg = env->sd->groups;
10350 struct sg_lb_stats *local = &sds->local_stat;
10351 struct sg_lb_stats tmp_sgs;
10352 unsigned long sum_util = 0;
10356 struct sg_lb_stats *sgs = &tmp_sgs;
10359 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_span(sg));
10364 if (env->idle != CPU_NEWLY_IDLE ||
10365 time_after_eq(jiffies, sg->sgc->next_update))
10366 update_group_capacity(env->sd, env->dst_cpu);
10369 update_sg_lb_stats(env, sds, sg, sgs, &sg_status);
10375 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
10377 sds->busiest_stat = *sgs;
10381 /* Now, start updating sd_lb_stats */
10382 sds->total_load += sgs->group_load;
10383 sds->total_capacity += sgs->group_capacity;
10385 sum_util += sgs->group_util;
10387 } while (sg != env->sd->groups);
10390 * Indicate that the child domain of the busiest group prefers tasks
10391 * go to a child's sibling domains first. NB the flags of a sched group
10392 * are those of the child domain.
10395 sds->prefer_sibling = !!(sds->busiest->flags & SD_PREFER_SIBLING);
10398 if (env->sd->flags & SD_NUMA)
10399 env->fbq_type = fbq_classify_group(&sds->busiest_stat);
10401 if (!env->sd->parent) {
10402 struct root_domain *rd = env->dst_rq->rd;
10404 /* update overload indicator if we are at root domain */
10405 WRITE_ONCE(rd->overload, sg_status & SG_OVERLOAD);
10407 /* Update over-utilization (tipping point, U >= 0) indicator */
10408 WRITE_ONCE(rd->overutilized, sg_status & SG_OVERUTILIZED);
10409 trace_sched_overutilized_tp(rd, sg_status & SG_OVERUTILIZED);
10410 } else if (sg_status & SG_OVERUTILIZED) {
10411 struct root_domain *rd = env->dst_rq->rd;
10413 WRITE_ONCE(rd->overutilized, SG_OVERUTILIZED);
10414 trace_sched_overutilized_tp(rd, SG_OVERUTILIZED);
10417 update_idle_cpu_scan(env, sum_util);
10421 * calculate_imbalance - Calculate the amount of imbalance present within the
10422 * groups of a given sched_domain during load balance.
10423 * @env: load balance environment
10424 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
10426 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
10428 struct sg_lb_stats *local, *busiest;
10430 local = &sds->local_stat;
10431 busiest = &sds->busiest_stat;
10433 if (busiest->group_type == group_misfit_task) {
10434 if (env->sd->flags & SD_ASYM_CPUCAPACITY) {
10435 /* Set imbalance to allow misfit tasks to be balanced. */
10436 env->migration_type = migrate_misfit;
10437 env->imbalance = 1;
10440 * Set load imbalance to allow moving task from cpu
10441 * with reduced capacity.
10443 env->migration_type = migrate_load;
10444 env->imbalance = busiest->group_misfit_task_load;
10449 if (busiest->group_type == group_asym_packing) {
10451 * In case of asym capacity, we will try to migrate all load to
10452 * the preferred CPU.
10454 env->migration_type = migrate_task;
10455 env->imbalance = busiest->sum_h_nr_running;
10459 if (busiest->group_type == group_smt_balance) {
10460 /* Reduce number of tasks sharing CPU capacity */
10461 env->migration_type = migrate_task;
10462 env->imbalance = 1;
10466 if (busiest->group_type == group_imbalanced) {
10468 * In the group_imb case we cannot rely on group-wide averages
10469 * to ensure CPU-load equilibrium, try to move any task to fix
10470 * the imbalance. The next load balance will take care of
10471 * balancing back the system.
10473 env->migration_type = migrate_task;
10474 env->imbalance = 1;
10479 * Try to use spare capacity of local group without overloading it or
10480 * emptying busiest.
10482 if (local->group_type == group_has_spare) {
10483 if ((busiest->group_type > group_fully_busy) &&
10484 !(env->sd->flags & SD_SHARE_PKG_RESOURCES)) {
10486 * If busiest is overloaded, try to fill spare
10487 * capacity. This might end up creating spare capacity
10488 * in busiest or busiest still being overloaded but
10489 * there is no simple way to directly compute the
10490 * amount of load to migrate in order to balance the
10493 env->migration_type = migrate_util;
10494 env->imbalance = max(local->group_capacity, local->group_util) -
10498 * In some cases, the group's utilization is max or even
10499 * higher than capacity because of migrations but the
10500 * local CPU is (newly) idle. There is at least one
10501 * waiting task in this overloaded busiest group. Let's
10504 if (env->idle != CPU_NOT_IDLE && env->imbalance == 0) {
10505 env->migration_type = migrate_task;
10506 env->imbalance = 1;
10512 if (busiest->group_weight == 1 || sds->prefer_sibling) {
10514 * When prefer sibling, evenly spread running tasks on
10517 env->migration_type = migrate_task;
10518 env->imbalance = sibling_imbalance(env, sds, busiest, local);
10522 * If there is no overload, we just want to even the number of
10525 env->migration_type = migrate_task;
10526 env->imbalance = max_t(long, 0,
10527 (local->idle_cpus - busiest->idle_cpus));
10531 /* Consider allowing a small imbalance between NUMA groups */
10532 if (env->sd->flags & SD_NUMA) {
10533 env->imbalance = adjust_numa_imbalance(env->imbalance,
10534 local->sum_nr_running + 1,
10535 env->sd->imb_numa_nr);
10539 /* Number of tasks to move to restore balance */
10540 env->imbalance >>= 1;
10546 * Local is fully busy but has to take more load to relieve the
10549 if (local->group_type < group_overloaded) {
10551 * Local will become overloaded so the avg_load metrics are
10555 local->avg_load = (local->group_load * SCHED_CAPACITY_SCALE) /
10556 local->group_capacity;
10559 * If the local group is more loaded than the selected
10560 * busiest group don't try to pull any tasks.
10562 if (local->avg_load >= busiest->avg_load) {
10563 env->imbalance = 0;
10567 sds->avg_load = (sds->total_load * SCHED_CAPACITY_SCALE) /
10568 sds->total_capacity;
10571 * If the local group is more loaded than the average system
10572 * load, don't try to pull any tasks.
10574 if (local->avg_load >= sds->avg_load) {
10575 env->imbalance = 0;
10582 * Both group are or will become overloaded and we're trying to get all
10583 * the CPUs to the average_load, so we don't want to push ourselves
10584 * above the average load, nor do we wish to reduce the max loaded CPU
10585 * below the average load. At the same time, we also don't want to
10586 * reduce the group load below the group capacity. Thus we look for
10587 * the minimum possible imbalance.
10589 env->migration_type = migrate_load;
10590 env->imbalance = min(
10591 (busiest->avg_load - sds->avg_load) * busiest->group_capacity,
10592 (sds->avg_load - local->avg_load) * local->group_capacity
10593 ) / SCHED_CAPACITY_SCALE;
10596 /******* find_busiest_group() helpers end here *********************/
10599 * Decision matrix according to the local and busiest group type:
10601 * busiest \ local has_spare fully_busy misfit asym imbalanced overloaded
10602 * has_spare nr_idle balanced N/A N/A balanced balanced
10603 * fully_busy nr_idle nr_idle N/A N/A balanced balanced
10604 * misfit_task force N/A N/A N/A N/A N/A
10605 * asym_packing force force N/A N/A force force
10606 * imbalanced force force N/A N/A force force
10607 * overloaded force force N/A N/A force avg_load
10609 * N/A : Not Applicable because already filtered while updating
10611 * balanced : The system is balanced for these 2 groups.
10612 * force : Calculate the imbalance as load migration is probably needed.
10613 * avg_load : Only if imbalance is significant enough.
10614 * nr_idle : dst_cpu is not busy and the number of idle CPUs is quite
10615 * different in groups.
10619 * find_busiest_group - Returns the busiest group within the sched_domain
10620 * if there is an imbalance.
10621 * @env: The load balancing environment.
10623 * Also calculates the amount of runnable load which should be moved
10624 * to restore balance.
10626 * Return: - The busiest group if imbalance exists.
10628 static struct sched_group *find_busiest_group(struct lb_env *env)
10630 struct sg_lb_stats *local, *busiest;
10631 struct sd_lb_stats sds;
10633 init_sd_lb_stats(&sds);
10636 * Compute the various statistics relevant for load balancing at
10639 update_sd_lb_stats(env, &sds);
10641 /* There is no busy sibling group to pull tasks from */
10645 busiest = &sds.busiest_stat;
10647 /* Misfit tasks should be dealt with regardless of the avg load */
10648 if (busiest->group_type == group_misfit_task)
10649 goto force_balance;
10651 if (sched_energy_enabled()) {
10652 struct root_domain *rd = env->dst_rq->rd;
10654 if (rcu_dereference(rd->pd) && !READ_ONCE(rd->overutilized))
10658 /* ASYM feature bypasses nice load balance check */
10659 if (busiest->group_type == group_asym_packing)
10660 goto force_balance;
10663 * If the busiest group is imbalanced the below checks don't
10664 * work because they assume all things are equal, which typically
10665 * isn't true due to cpus_ptr constraints and the like.
10667 if (busiest->group_type == group_imbalanced)
10668 goto force_balance;
10670 local = &sds.local_stat;
10672 * If the local group is busier than the selected busiest group
10673 * don't try and pull any tasks.
10675 if (local->group_type > busiest->group_type)
10679 * When groups are overloaded, use the avg_load to ensure fairness
10682 if (local->group_type == group_overloaded) {
10684 * If the local group is more loaded than the selected
10685 * busiest group don't try to pull any tasks.
10687 if (local->avg_load >= busiest->avg_load)
10690 /* XXX broken for overlapping NUMA groups */
10691 sds.avg_load = (sds.total_load * SCHED_CAPACITY_SCALE) /
10692 sds.total_capacity;
10695 * Don't pull any tasks if this group is already above the
10696 * domain average load.
10698 if (local->avg_load >= sds.avg_load)
10702 * If the busiest group is more loaded, use imbalance_pct to be
10705 if (100 * busiest->avg_load <=
10706 env->sd->imbalance_pct * local->avg_load)
10711 * Try to move all excess tasks to a sibling domain of the busiest
10712 * group's child domain.
10714 if (sds.prefer_sibling && local->group_type == group_has_spare &&
10715 sibling_imbalance(env, &sds, busiest, local) > 1)
10716 goto force_balance;
10718 if (busiest->group_type != group_overloaded) {
10719 if (env->idle == CPU_NOT_IDLE) {
10721 * If the busiest group is not overloaded (and as a
10722 * result the local one too) but this CPU is already
10723 * busy, let another idle CPU try to pull task.
10728 if (busiest->group_type == group_smt_balance &&
10729 smt_vs_nonsmt_groups(sds.local, sds.busiest)) {
10730 /* Let non SMT CPU pull from SMT CPU sharing with sibling */
10731 goto force_balance;
10734 if (busiest->group_weight > 1 &&
10735 local->idle_cpus <= (busiest->idle_cpus + 1)) {
10737 * If the busiest group is not overloaded
10738 * and there is no imbalance between this and busiest
10739 * group wrt idle CPUs, it is balanced. The imbalance
10740 * becomes significant if the diff is greater than 1
10741 * otherwise we might end up to just move the imbalance
10742 * on another group. Of course this applies only if
10743 * there is more than 1 CPU per group.
10748 if (busiest->sum_h_nr_running == 1) {
10750 * busiest doesn't have any tasks waiting to run
10757 /* Looks like there is an imbalance. Compute it */
10758 calculate_imbalance(env, &sds);
10759 return env->imbalance ? sds.busiest : NULL;
10762 env->imbalance = 0;
10767 * find_busiest_queue - find the busiest runqueue among the CPUs in the group.
10769 static struct rq *find_busiest_queue(struct lb_env *env,
10770 struct sched_group *group)
10772 struct rq *busiest = NULL, *rq;
10773 unsigned long busiest_util = 0, busiest_load = 0, busiest_capacity = 1;
10774 unsigned int busiest_nr = 0;
10777 for_each_cpu_and(i, sched_group_span(group), env->cpus) {
10778 unsigned long capacity, load, util;
10779 unsigned int nr_running;
10783 rt = fbq_classify_rq(rq);
10786 * We classify groups/runqueues into three groups:
10787 * - regular: there are !numa tasks
10788 * - remote: there are numa tasks that run on the 'wrong' node
10789 * - all: there is no distinction
10791 * In order to avoid migrating ideally placed numa tasks,
10792 * ignore those when there's better options.
10794 * If we ignore the actual busiest queue to migrate another
10795 * task, the next balance pass can still reduce the busiest
10796 * queue by moving tasks around inside the node.
10798 * If we cannot move enough load due to this classification
10799 * the next pass will adjust the group classification and
10800 * allow migration of more tasks.
10802 * Both cases only affect the total convergence complexity.
10804 if (rt > env->fbq_type)
10807 nr_running = rq->cfs.h_nr_running;
10811 capacity = capacity_of(i);
10814 * For ASYM_CPUCAPACITY domains, don't pick a CPU that could
10815 * eventually lead to active_balancing high->low capacity.
10816 * Higher per-CPU capacity is considered better than balancing
10819 if (env->sd->flags & SD_ASYM_CPUCAPACITY &&
10820 !capacity_greater(capacity_of(env->dst_cpu), capacity) &&
10825 * Make sure we only pull tasks from a CPU of lower priority
10826 * when balancing between SMT siblings.
10828 * If balancing between cores, let lower priority CPUs help
10829 * SMT cores with more than one busy sibling.
10831 if ((env->sd->flags & SD_ASYM_PACKING) &&
10832 sched_use_asym_prio(env->sd, i) &&
10833 sched_asym_prefer(i, env->dst_cpu) &&
10837 switch (env->migration_type) {
10840 * When comparing with load imbalance, use cpu_load()
10841 * which is not scaled with the CPU capacity.
10843 load = cpu_load(rq);
10845 if (nr_running == 1 && load > env->imbalance &&
10846 !check_cpu_capacity(rq, env->sd))
10850 * For the load comparisons with the other CPUs,
10851 * consider the cpu_load() scaled with the CPU
10852 * capacity, so that the load can be moved away
10853 * from the CPU that is potentially running at a
10856 * Thus we're looking for max(load_i / capacity_i),
10857 * crosswise multiplication to rid ourselves of the
10858 * division works out to:
10859 * load_i * capacity_j > load_j * capacity_i;
10860 * where j is our previous maximum.
10862 if (load * busiest_capacity > busiest_load * capacity) {
10863 busiest_load = load;
10864 busiest_capacity = capacity;
10870 util = cpu_util_cfs_boost(i);
10873 * Don't try to pull utilization from a CPU with one
10874 * running task. Whatever its utilization, we will fail
10877 if (nr_running <= 1)
10880 if (busiest_util < util) {
10881 busiest_util = util;
10887 if (busiest_nr < nr_running) {
10888 busiest_nr = nr_running;
10893 case migrate_misfit:
10895 * For ASYM_CPUCAPACITY domains with misfit tasks we
10896 * simply seek the "biggest" misfit task.
10898 if (rq->misfit_task_load > busiest_load) {
10899 busiest_load = rq->misfit_task_load;
10912 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
10913 * so long as it is large enough.
10915 #define MAX_PINNED_INTERVAL 512
10918 asym_active_balance(struct lb_env *env)
10921 * ASYM_PACKING needs to force migrate tasks from busy but lower
10922 * priority CPUs in order to pack all tasks in the highest priority
10923 * CPUs. When done between cores, do it only if the whole core if the
10924 * whole core is idle.
10926 * If @env::src_cpu is an SMT core with busy siblings, let
10927 * the lower priority @env::dst_cpu help it. Do not follow
10930 return env->idle != CPU_NOT_IDLE && (env->sd->flags & SD_ASYM_PACKING) &&
10931 sched_use_asym_prio(env->sd, env->dst_cpu) &&
10932 (sched_asym_prefer(env->dst_cpu, env->src_cpu) ||
10933 !sched_use_asym_prio(env->sd, env->src_cpu));
10937 imbalanced_active_balance(struct lb_env *env)
10939 struct sched_domain *sd = env->sd;
10942 * The imbalanced case includes the case of pinned tasks preventing a fair
10943 * distribution of the load on the system but also the even distribution of the
10944 * threads on a system with spare capacity
10946 if ((env->migration_type == migrate_task) &&
10947 (sd->nr_balance_failed > sd->cache_nice_tries+2))
10953 static int need_active_balance(struct lb_env *env)
10955 struct sched_domain *sd = env->sd;
10957 if (asym_active_balance(env))
10960 if (imbalanced_active_balance(env))
10964 * The dst_cpu is idle and the src_cpu CPU has only 1 CFS task.
10965 * It's worth migrating the task if the src_cpu's capacity is reduced
10966 * because of other sched_class or IRQs if more capacity stays
10967 * available on dst_cpu.
10969 if ((env->idle != CPU_NOT_IDLE) &&
10970 (env->src_rq->cfs.h_nr_running == 1)) {
10971 if ((check_cpu_capacity(env->src_rq, sd)) &&
10972 (capacity_of(env->src_cpu)*sd->imbalance_pct < capacity_of(env->dst_cpu)*100))
10976 if (env->migration_type == migrate_misfit)
10982 static int active_load_balance_cpu_stop(void *data);
10984 static int should_we_balance(struct lb_env *env)
10986 struct cpumask *swb_cpus = this_cpu_cpumask_var_ptr(should_we_balance_tmpmask);
10987 struct sched_group *sg = env->sd->groups;
10988 int cpu, idle_smt = -1;
10991 * Ensure the balancing environment is consistent; can happen
10992 * when the softirq triggers 'during' hotplug.
10994 if (!cpumask_test_cpu(env->dst_cpu, env->cpus))
10998 * In the newly idle case, we will allow all the CPUs
10999 * to do the newly idle load balance.
11001 * However, we bail out if we already have tasks or a wakeup pending,
11002 * to optimize wakeup latency.
11004 if (env->idle == CPU_NEWLY_IDLE) {
11005 if (env->dst_rq->nr_running > 0 || env->dst_rq->ttwu_pending)
11010 cpumask_copy(swb_cpus, group_balance_mask(sg));
11011 /* Try to find first idle CPU */
11012 for_each_cpu_and(cpu, swb_cpus, env->cpus) {
11013 if (!idle_cpu(cpu))
11017 * Don't balance to idle SMT in busy core right away when
11018 * balancing cores, but remember the first idle SMT CPU for
11019 * later consideration. Find CPU on an idle core first.
11021 if (!(env->sd->flags & SD_SHARE_CPUCAPACITY) && !is_core_idle(cpu)) {
11022 if (idle_smt == -1)
11025 * If the core is not idle, and first SMT sibling which is
11026 * idle has been found, then its not needed to check other
11027 * SMT siblings for idleness:
11029 #ifdef CONFIG_SCHED_SMT
11030 cpumask_andnot(swb_cpus, swb_cpus, cpu_smt_mask(cpu));
11035 /* Are we the first idle CPU? */
11036 return cpu == env->dst_cpu;
11039 if (idle_smt == env->dst_cpu)
11042 /* Are we the first CPU of this group ? */
11043 return group_balance_cpu(sg) == env->dst_cpu;
11047 * Check this_cpu to ensure it is balanced within domain. Attempt to move
11048 * tasks if there is an imbalance.
11050 static int load_balance(int this_cpu, struct rq *this_rq,
11051 struct sched_domain *sd, enum cpu_idle_type idle,
11052 int *continue_balancing)
11054 int ld_moved, cur_ld_moved, active_balance = 0;
11055 struct sched_domain *sd_parent = sd->parent;
11056 struct sched_group *group;
11057 struct rq *busiest;
11058 struct rq_flags rf;
11059 struct cpumask *cpus = this_cpu_cpumask_var_ptr(load_balance_mask);
11060 struct lb_env env = {
11062 .dst_cpu = this_cpu,
11064 .dst_grpmask = group_balance_mask(sd->groups),
11066 .loop_break = SCHED_NR_MIGRATE_BREAK,
11069 .tasks = LIST_HEAD_INIT(env.tasks),
11072 cpumask_and(cpus, sched_domain_span(sd), cpu_active_mask);
11074 schedstat_inc(sd->lb_count[idle]);
11077 if (!should_we_balance(&env)) {
11078 *continue_balancing = 0;
11082 group = find_busiest_group(&env);
11084 schedstat_inc(sd->lb_nobusyg[idle]);
11088 busiest = find_busiest_queue(&env, group);
11090 schedstat_inc(sd->lb_nobusyq[idle]);
11094 WARN_ON_ONCE(busiest == env.dst_rq);
11096 schedstat_add(sd->lb_imbalance[idle], env.imbalance);
11098 env.src_cpu = busiest->cpu;
11099 env.src_rq = busiest;
11102 /* Clear this flag as soon as we find a pullable task */
11103 env.flags |= LBF_ALL_PINNED;
11104 if (busiest->nr_running > 1) {
11106 * Attempt to move tasks. If find_busiest_group has found
11107 * an imbalance but busiest->nr_running <= 1, the group is
11108 * still unbalanced. ld_moved simply stays zero, so it is
11109 * correctly treated as an imbalance.
11111 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
11114 rq_lock_irqsave(busiest, &rf);
11115 update_rq_clock(busiest);
11118 * cur_ld_moved - load moved in current iteration
11119 * ld_moved - cumulative load moved across iterations
11121 cur_ld_moved = detach_tasks(&env);
11124 * We've detached some tasks from busiest_rq. Every
11125 * task is masked "TASK_ON_RQ_MIGRATING", so we can safely
11126 * unlock busiest->lock, and we are able to be sure
11127 * that nobody can manipulate the tasks in parallel.
11128 * See task_rq_lock() family for the details.
11131 rq_unlock(busiest, &rf);
11133 if (cur_ld_moved) {
11134 attach_tasks(&env);
11135 ld_moved += cur_ld_moved;
11138 local_irq_restore(rf.flags);
11140 if (env.flags & LBF_NEED_BREAK) {
11141 env.flags &= ~LBF_NEED_BREAK;
11142 /* Stop if we tried all running tasks */
11143 if (env.loop < busiest->nr_running)
11148 * Revisit (affine) tasks on src_cpu that couldn't be moved to
11149 * us and move them to an alternate dst_cpu in our sched_group
11150 * where they can run. The upper limit on how many times we
11151 * iterate on same src_cpu is dependent on number of CPUs in our
11154 * This changes load balance semantics a bit on who can move
11155 * load to a given_cpu. In addition to the given_cpu itself
11156 * (or a ilb_cpu acting on its behalf where given_cpu is
11157 * nohz-idle), we now have balance_cpu in a position to move
11158 * load to given_cpu. In rare situations, this may cause
11159 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
11160 * _independently_ and at _same_ time to move some load to
11161 * given_cpu) causing excess load to be moved to given_cpu.
11162 * This however should not happen so much in practice and
11163 * moreover subsequent load balance cycles should correct the
11164 * excess load moved.
11166 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
11168 /* Prevent to re-select dst_cpu via env's CPUs */
11169 __cpumask_clear_cpu(env.dst_cpu, env.cpus);
11171 env.dst_rq = cpu_rq(env.new_dst_cpu);
11172 env.dst_cpu = env.new_dst_cpu;
11173 env.flags &= ~LBF_DST_PINNED;
11175 env.loop_break = SCHED_NR_MIGRATE_BREAK;
11178 * Go back to "more_balance" rather than "redo" since we
11179 * need to continue with same src_cpu.
11185 * We failed to reach balance because of affinity.
11188 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
11190 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0)
11191 *group_imbalance = 1;
11194 /* All tasks on this runqueue were pinned by CPU affinity */
11195 if (unlikely(env.flags & LBF_ALL_PINNED)) {
11196 __cpumask_clear_cpu(cpu_of(busiest), cpus);
11198 * Attempting to continue load balancing at the current
11199 * sched_domain level only makes sense if there are
11200 * active CPUs remaining as possible busiest CPUs to
11201 * pull load from which are not contained within the
11202 * destination group that is receiving any migrated
11205 if (!cpumask_subset(cpus, env.dst_grpmask)) {
11207 env.loop_break = SCHED_NR_MIGRATE_BREAK;
11210 goto out_all_pinned;
11215 schedstat_inc(sd->lb_failed[idle]);
11217 * Increment the failure counter only on periodic balance.
11218 * We do not want newidle balance, which can be very
11219 * frequent, pollute the failure counter causing
11220 * excessive cache_hot migrations and active balances.
11222 if (idle != CPU_NEWLY_IDLE)
11223 sd->nr_balance_failed++;
11225 if (need_active_balance(&env)) {
11226 unsigned long flags;
11228 raw_spin_rq_lock_irqsave(busiest, flags);
11231 * Don't kick the active_load_balance_cpu_stop,
11232 * if the curr task on busiest CPU can't be
11233 * moved to this_cpu:
11235 if (!cpumask_test_cpu(this_cpu, busiest->curr->cpus_ptr)) {
11236 raw_spin_rq_unlock_irqrestore(busiest, flags);
11237 goto out_one_pinned;
11240 /* Record that we found at least one task that could run on this_cpu */
11241 env.flags &= ~LBF_ALL_PINNED;
11244 * ->active_balance synchronizes accesses to
11245 * ->active_balance_work. Once set, it's cleared
11246 * only after active load balance is finished.
11248 if (!busiest->active_balance) {
11249 busiest->active_balance = 1;
11250 busiest->push_cpu = this_cpu;
11251 active_balance = 1;
11253 raw_spin_rq_unlock_irqrestore(busiest, flags);
11255 if (active_balance) {
11256 stop_one_cpu_nowait(cpu_of(busiest),
11257 active_load_balance_cpu_stop, busiest,
11258 &busiest->active_balance_work);
11262 sd->nr_balance_failed = 0;
11265 if (likely(!active_balance) || need_active_balance(&env)) {
11266 /* We were unbalanced, so reset the balancing interval */
11267 sd->balance_interval = sd->min_interval;
11274 * We reach balance although we may have faced some affinity
11275 * constraints. Clear the imbalance flag only if other tasks got
11276 * a chance to move and fix the imbalance.
11278 if (sd_parent && !(env.flags & LBF_ALL_PINNED)) {
11279 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
11281 if (*group_imbalance)
11282 *group_imbalance = 0;
11287 * We reach balance because all tasks are pinned at this level so
11288 * we can't migrate them. Let the imbalance flag set so parent level
11289 * can try to migrate them.
11291 schedstat_inc(sd->lb_balanced[idle]);
11293 sd->nr_balance_failed = 0;
11299 * newidle_balance() disregards balance intervals, so we could
11300 * repeatedly reach this code, which would lead to balance_interval
11301 * skyrocketing in a short amount of time. Skip the balance_interval
11302 * increase logic to avoid that.
11304 if (env.idle == CPU_NEWLY_IDLE)
11307 /* tune up the balancing interval */
11308 if ((env.flags & LBF_ALL_PINNED &&
11309 sd->balance_interval < MAX_PINNED_INTERVAL) ||
11310 sd->balance_interval < sd->max_interval)
11311 sd->balance_interval *= 2;
11316 static inline unsigned long
11317 get_sd_balance_interval(struct sched_domain *sd, int cpu_busy)
11319 unsigned long interval = sd->balance_interval;
11322 interval *= sd->busy_factor;
11324 /* scale ms to jiffies */
11325 interval = msecs_to_jiffies(interval);
11328 * Reduce likelihood of busy balancing at higher domains racing with
11329 * balancing at lower domains by preventing their balancing periods
11330 * from being multiples of each other.
11335 interval = clamp(interval, 1UL, max_load_balance_interval);
11341 update_next_balance(struct sched_domain *sd, unsigned long *next_balance)
11343 unsigned long interval, next;
11345 /* used by idle balance, so cpu_busy = 0 */
11346 interval = get_sd_balance_interval(sd, 0);
11347 next = sd->last_balance + interval;
11349 if (time_after(*next_balance, next))
11350 *next_balance = next;
11354 * active_load_balance_cpu_stop is run by the CPU stopper. It pushes
11355 * running tasks off the busiest CPU onto idle CPUs. It requires at
11356 * least 1 task to be running on each physical CPU where possible, and
11357 * avoids physical / logical imbalances.
11359 static int active_load_balance_cpu_stop(void *data)
11361 struct rq *busiest_rq = data;
11362 int busiest_cpu = cpu_of(busiest_rq);
11363 int target_cpu = busiest_rq->push_cpu;
11364 struct rq *target_rq = cpu_rq(target_cpu);
11365 struct sched_domain *sd;
11366 struct task_struct *p = NULL;
11367 struct rq_flags rf;
11369 rq_lock_irq(busiest_rq, &rf);
11371 * Between queueing the stop-work and running it is a hole in which
11372 * CPUs can become inactive. We should not move tasks from or to
11375 if (!cpu_active(busiest_cpu) || !cpu_active(target_cpu))
11378 /* Make sure the requested CPU hasn't gone down in the meantime: */
11379 if (unlikely(busiest_cpu != smp_processor_id() ||
11380 !busiest_rq->active_balance))
11383 /* Is there any task to move? */
11384 if (busiest_rq->nr_running <= 1)
11388 * This condition is "impossible", if it occurs
11389 * we need to fix it. Originally reported by
11390 * Bjorn Helgaas on a 128-CPU setup.
11392 WARN_ON_ONCE(busiest_rq == target_rq);
11394 /* Search for an sd spanning us and the target CPU. */
11396 for_each_domain(target_cpu, sd) {
11397 if (cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
11402 struct lb_env env = {
11404 .dst_cpu = target_cpu,
11405 .dst_rq = target_rq,
11406 .src_cpu = busiest_rq->cpu,
11407 .src_rq = busiest_rq,
11409 .flags = LBF_ACTIVE_LB,
11412 schedstat_inc(sd->alb_count);
11413 update_rq_clock(busiest_rq);
11415 p = detach_one_task(&env);
11417 schedstat_inc(sd->alb_pushed);
11418 /* Active balancing done, reset the failure counter. */
11419 sd->nr_balance_failed = 0;
11421 schedstat_inc(sd->alb_failed);
11426 busiest_rq->active_balance = 0;
11427 rq_unlock(busiest_rq, &rf);
11430 attach_one_task(target_rq, p);
11432 local_irq_enable();
11437 static DEFINE_SPINLOCK(balancing);
11440 * Scale the max load_balance interval with the number of CPUs in the system.
11441 * This trades load-balance latency on larger machines for less cross talk.
11443 void update_max_interval(void)
11445 max_load_balance_interval = HZ*num_online_cpus()/10;
11448 static inline bool update_newidle_cost(struct sched_domain *sd, u64 cost)
11450 if (cost > sd->max_newidle_lb_cost) {
11452 * Track max cost of a domain to make sure to not delay the
11453 * next wakeup on the CPU.
11455 sd->max_newidle_lb_cost = cost;
11456 sd->last_decay_max_lb_cost = jiffies;
11457 } else if (time_after(jiffies, sd->last_decay_max_lb_cost + HZ)) {
11459 * Decay the newidle max times by ~1% per second to ensure that
11460 * it is not outdated and the current max cost is actually
11463 sd->max_newidle_lb_cost = (sd->max_newidle_lb_cost * 253) / 256;
11464 sd->last_decay_max_lb_cost = jiffies;
11473 * It checks each scheduling domain to see if it is due to be balanced,
11474 * and initiates a balancing operation if so.
11476 * Balancing parameters are set up in init_sched_domains.
11478 static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle)
11480 int continue_balancing = 1;
11482 int busy = idle != CPU_IDLE && !sched_idle_cpu(cpu);
11483 unsigned long interval;
11484 struct sched_domain *sd;
11485 /* Earliest time when we have to do rebalance again */
11486 unsigned long next_balance = jiffies + 60*HZ;
11487 int update_next_balance = 0;
11488 int need_serialize, need_decay = 0;
11492 for_each_domain(cpu, sd) {
11494 * Decay the newidle max times here because this is a regular
11495 * visit to all the domains.
11497 need_decay = update_newidle_cost(sd, 0);
11498 max_cost += sd->max_newidle_lb_cost;
11501 * Stop the load balance at this level. There is another
11502 * CPU in our sched group which is doing load balancing more
11505 if (!continue_balancing) {
11511 interval = get_sd_balance_interval(sd, busy);
11513 need_serialize = sd->flags & SD_SERIALIZE;
11514 if (need_serialize) {
11515 if (!spin_trylock(&balancing))
11519 if (time_after_eq(jiffies, sd->last_balance + interval)) {
11520 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
11522 * The LBF_DST_PINNED logic could have changed
11523 * env->dst_cpu, so we can't know our idle
11524 * state even if we migrated tasks. Update it.
11526 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
11527 busy = idle != CPU_IDLE && !sched_idle_cpu(cpu);
11529 sd->last_balance = jiffies;
11530 interval = get_sd_balance_interval(sd, busy);
11532 if (need_serialize)
11533 spin_unlock(&balancing);
11535 if (time_after(next_balance, sd->last_balance + interval)) {
11536 next_balance = sd->last_balance + interval;
11537 update_next_balance = 1;
11542 * Ensure the rq-wide value also decays but keep it at a
11543 * reasonable floor to avoid funnies with rq->avg_idle.
11545 rq->max_idle_balance_cost =
11546 max((u64)sysctl_sched_migration_cost, max_cost);
11551 * next_balance will be updated only when there is a need.
11552 * When the cpu is attached to null domain for ex, it will not be
11555 if (likely(update_next_balance))
11556 rq->next_balance = next_balance;
11560 static inline int on_null_domain(struct rq *rq)
11562 return unlikely(!rcu_dereference_sched(rq->sd));
11565 #ifdef CONFIG_NO_HZ_COMMON
11567 * idle load balancing details
11568 * - When one of the busy CPUs notice that there may be an idle rebalancing
11569 * needed, they will kick the idle load balancer, which then does idle
11570 * load balancing for all the idle CPUs.
11571 * - HK_TYPE_MISC CPUs are used for this task, because HK_TYPE_SCHED not set
11575 static inline int find_new_ilb(void)
11578 const struct cpumask *hk_mask;
11580 hk_mask = housekeeping_cpumask(HK_TYPE_MISC);
11582 for_each_cpu_and(ilb, nohz.idle_cpus_mask, hk_mask) {
11584 if (ilb == smp_processor_id())
11595 * Kick a CPU to do the nohz balancing, if it is time for it. We pick any
11596 * idle CPU in the HK_TYPE_MISC housekeeping set (if there is one).
11598 static void kick_ilb(unsigned int flags)
11603 * Increase nohz.next_balance only when if full ilb is triggered but
11604 * not if we only update stats.
11606 if (flags & NOHZ_BALANCE_KICK)
11607 nohz.next_balance = jiffies+1;
11609 ilb_cpu = find_new_ilb();
11611 if (ilb_cpu >= nr_cpu_ids)
11615 * Access to rq::nohz_csd is serialized by NOHZ_KICK_MASK; he who sets
11616 * the first flag owns it; cleared by nohz_csd_func().
11618 flags = atomic_fetch_or(flags, nohz_flags(ilb_cpu));
11619 if (flags & NOHZ_KICK_MASK)
11623 * This way we generate an IPI on the target CPU which
11624 * is idle. And the softirq performing nohz idle load balance
11625 * will be run before returning from the IPI.
11627 smp_call_function_single_async(ilb_cpu, &cpu_rq(ilb_cpu)->nohz_csd);
11631 * Current decision point for kicking the idle load balancer in the presence
11632 * of idle CPUs in the system.
11634 static void nohz_balancer_kick(struct rq *rq)
11636 unsigned long now = jiffies;
11637 struct sched_domain_shared *sds;
11638 struct sched_domain *sd;
11639 int nr_busy, i, cpu = rq->cpu;
11640 unsigned int flags = 0;
11642 if (unlikely(rq->idle_balance))
11646 * We may be recently in ticked or tickless idle mode. At the first
11647 * busy tick after returning from idle, we will update the busy stats.
11649 nohz_balance_exit_idle(rq);
11652 * None are in tickless mode and hence no need for NOHZ idle load
11655 if (likely(!atomic_read(&nohz.nr_cpus)))
11658 if (READ_ONCE(nohz.has_blocked) &&
11659 time_after(now, READ_ONCE(nohz.next_blocked)))
11660 flags = NOHZ_STATS_KICK;
11662 if (time_before(now, nohz.next_balance))
11665 if (rq->nr_running >= 2) {
11666 flags = NOHZ_STATS_KICK | NOHZ_BALANCE_KICK;
11672 sd = rcu_dereference(rq->sd);
11675 * If there's a CFS task and the current CPU has reduced
11676 * capacity; kick the ILB to see if there's a better CPU to run
11679 if (rq->cfs.h_nr_running >= 1 && check_cpu_capacity(rq, sd)) {
11680 flags = NOHZ_STATS_KICK | NOHZ_BALANCE_KICK;
11685 sd = rcu_dereference(per_cpu(sd_asym_packing, cpu));
11688 * When ASYM_PACKING; see if there's a more preferred CPU
11689 * currently idle; in which case, kick the ILB to move tasks
11692 * When balancing betwen cores, all the SMT siblings of the
11693 * preferred CPU must be idle.
11695 for_each_cpu_and(i, sched_domain_span(sd), nohz.idle_cpus_mask) {
11696 if (sched_use_asym_prio(sd, i) &&
11697 sched_asym_prefer(i, cpu)) {
11698 flags = NOHZ_STATS_KICK | NOHZ_BALANCE_KICK;
11704 sd = rcu_dereference(per_cpu(sd_asym_cpucapacity, cpu));
11707 * When ASYM_CPUCAPACITY; see if there's a higher capacity CPU
11708 * to run the misfit task on.
11710 if (check_misfit_status(rq, sd)) {
11711 flags = NOHZ_STATS_KICK | NOHZ_BALANCE_KICK;
11716 * For asymmetric systems, we do not want to nicely balance
11717 * cache use, instead we want to embrace asymmetry and only
11718 * ensure tasks have enough CPU capacity.
11720 * Skip the LLC logic because it's not relevant in that case.
11725 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
11728 * If there is an imbalance between LLC domains (IOW we could
11729 * increase the overall cache use), we need some less-loaded LLC
11730 * domain to pull some load. Likewise, we may need to spread
11731 * load within the current LLC domain (e.g. packed SMT cores but
11732 * other CPUs are idle). We can't really know from here how busy
11733 * the others are - so just get a nohz balance going if it looks
11734 * like this LLC domain has tasks we could move.
11736 nr_busy = atomic_read(&sds->nr_busy_cpus);
11738 flags = NOHZ_STATS_KICK | NOHZ_BALANCE_KICK;
11745 if (READ_ONCE(nohz.needs_update))
11746 flags |= NOHZ_NEXT_KICK;
11752 static void set_cpu_sd_state_busy(int cpu)
11754 struct sched_domain *sd;
11757 sd = rcu_dereference(per_cpu(sd_llc, cpu));
11759 if (!sd || !sd->nohz_idle)
11763 atomic_inc(&sd->shared->nr_busy_cpus);
11768 void nohz_balance_exit_idle(struct rq *rq)
11770 SCHED_WARN_ON(rq != this_rq());
11772 if (likely(!rq->nohz_tick_stopped))
11775 rq->nohz_tick_stopped = 0;
11776 cpumask_clear_cpu(rq->cpu, nohz.idle_cpus_mask);
11777 atomic_dec(&nohz.nr_cpus);
11779 set_cpu_sd_state_busy(rq->cpu);
11782 static void set_cpu_sd_state_idle(int cpu)
11784 struct sched_domain *sd;
11787 sd = rcu_dereference(per_cpu(sd_llc, cpu));
11789 if (!sd || sd->nohz_idle)
11793 atomic_dec(&sd->shared->nr_busy_cpus);
11799 * This routine will record that the CPU is going idle with tick stopped.
11800 * This info will be used in performing idle load balancing in the future.
11802 void nohz_balance_enter_idle(int cpu)
11804 struct rq *rq = cpu_rq(cpu);
11806 SCHED_WARN_ON(cpu != smp_processor_id());
11808 /* If this CPU is going down, then nothing needs to be done: */
11809 if (!cpu_active(cpu))
11812 /* Spare idle load balancing on CPUs that don't want to be disturbed: */
11813 if (!housekeeping_cpu(cpu, HK_TYPE_SCHED))
11817 * Can be set safely without rq->lock held
11818 * If a clear happens, it will have evaluated last additions because
11819 * rq->lock is held during the check and the clear
11821 rq->has_blocked_load = 1;
11824 * The tick is still stopped but load could have been added in the
11825 * meantime. We set the nohz.has_blocked flag to trig a check of the
11826 * *_avg. The CPU is already part of nohz.idle_cpus_mask so the clear
11827 * of nohz.has_blocked can only happen after checking the new load
11829 if (rq->nohz_tick_stopped)
11832 /* If we're a completely isolated CPU, we don't play: */
11833 if (on_null_domain(rq))
11836 rq->nohz_tick_stopped = 1;
11838 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
11839 atomic_inc(&nohz.nr_cpus);
11842 * Ensures that if nohz_idle_balance() fails to observe our
11843 * @idle_cpus_mask store, it must observe the @has_blocked
11844 * and @needs_update stores.
11846 smp_mb__after_atomic();
11848 set_cpu_sd_state_idle(cpu);
11850 WRITE_ONCE(nohz.needs_update, 1);
11853 * Each time a cpu enter idle, we assume that it has blocked load and
11854 * enable the periodic update of the load of idle cpus
11856 WRITE_ONCE(nohz.has_blocked, 1);
11859 static bool update_nohz_stats(struct rq *rq)
11861 unsigned int cpu = rq->cpu;
11863 if (!rq->has_blocked_load)
11866 if (!cpumask_test_cpu(cpu, nohz.idle_cpus_mask))
11869 if (!time_after(jiffies, READ_ONCE(rq->last_blocked_load_update_tick)))
11872 update_blocked_averages(cpu);
11874 return rq->has_blocked_load;
11878 * Internal function that runs load balance for all idle cpus. The load balance
11879 * can be a simple update of blocked load or a complete load balance with
11880 * tasks movement depending of flags.
11882 static void _nohz_idle_balance(struct rq *this_rq, unsigned int flags)
11884 /* Earliest time when we have to do rebalance again */
11885 unsigned long now = jiffies;
11886 unsigned long next_balance = now + 60*HZ;
11887 bool has_blocked_load = false;
11888 int update_next_balance = 0;
11889 int this_cpu = this_rq->cpu;
11893 SCHED_WARN_ON((flags & NOHZ_KICK_MASK) == NOHZ_BALANCE_KICK);
11896 * We assume there will be no idle load after this update and clear
11897 * the has_blocked flag. If a cpu enters idle in the mean time, it will
11898 * set the has_blocked flag and trigger another update of idle load.
11899 * Because a cpu that becomes idle, is added to idle_cpus_mask before
11900 * setting the flag, we are sure to not clear the state and not
11901 * check the load of an idle cpu.
11903 * Same applies to idle_cpus_mask vs needs_update.
11905 if (flags & NOHZ_STATS_KICK)
11906 WRITE_ONCE(nohz.has_blocked, 0);
11907 if (flags & NOHZ_NEXT_KICK)
11908 WRITE_ONCE(nohz.needs_update, 0);
11911 * Ensures that if we miss the CPU, we must see the has_blocked
11912 * store from nohz_balance_enter_idle().
11917 * Start with the next CPU after this_cpu so we will end with this_cpu and let a
11918 * chance for other idle cpu to pull load.
11920 for_each_cpu_wrap(balance_cpu, nohz.idle_cpus_mask, this_cpu+1) {
11921 if (!idle_cpu(balance_cpu))
11925 * If this CPU gets work to do, stop the load balancing
11926 * work being done for other CPUs. Next load
11927 * balancing owner will pick it up.
11929 if (need_resched()) {
11930 if (flags & NOHZ_STATS_KICK)
11931 has_blocked_load = true;
11932 if (flags & NOHZ_NEXT_KICK)
11933 WRITE_ONCE(nohz.needs_update, 1);
11937 rq = cpu_rq(balance_cpu);
11939 if (flags & NOHZ_STATS_KICK)
11940 has_blocked_load |= update_nohz_stats(rq);
11943 * If time for next balance is due,
11946 if (time_after_eq(jiffies, rq->next_balance)) {
11947 struct rq_flags rf;
11949 rq_lock_irqsave(rq, &rf);
11950 update_rq_clock(rq);
11951 rq_unlock_irqrestore(rq, &rf);
11953 if (flags & NOHZ_BALANCE_KICK)
11954 rebalance_domains(rq, CPU_IDLE);
11957 if (time_after(next_balance, rq->next_balance)) {
11958 next_balance = rq->next_balance;
11959 update_next_balance = 1;
11964 * next_balance will be updated only when there is a need.
11965 * When the CPU is attached to null domain for ex, it will not be
11968 if (likely(update_next_balance))
11969 nohz.next_balance = next_balance;
11971 if (flags & NOHZ_STATS_KICK)
11972 WRITE_ONCE(nohz.next_blocked,
11973 now + msecs_to_jiffies(LOAD_AVG_PERIOD));
11976 /* There is still blocked load, enable periodic update */
11977 if (has_blocked_load)
11978 WRITE_ONCE(nohz.has_blocked, 1);
11982 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
11983 * rebalancing for all the cpus for whom scheduler ticks are stopped.
11985 static bool nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
11987 unsigned int flags = this_rq->nohz_idle_balance;
11992 this_rq->nohz_idle_balance = 0;
11994 if (idle != CPU_IDLE)
11997 _nohz_idle_balance(this_rq, flags);
12003 * Check if we need to run the ILB for updating blocked load before entering
12006 void nohz_run_idle_balance(int cpu)
12008 unsigned int flags;
12010 flags = atomic_fetch_andnot(NOHZ_NEWILB_KICK, nohz_flags(cpu));
12013 * Update the blocked load only if no SCHED_SOFTIRQ is about to happen
12014 * (ie NOHZ_STATS_KICK set) and will do the same.
12016 if ((flags == NOHZ_NEWILB_KICK) && !need_resched())
12017 _nohz_idle_balance(cpu_rq(cpu), NOHZ_STATS_KICK);
12020 static void nohz_newidle_balance(struct rq *this_rq)
12022 int this_cpu = this_rq->cpu;
12025 * This CPU doesn't want to be disturbed by scheduler
12028 if (!housekeeping_cpu(this_cpu, HK_TYPE_SCHED))
12031 /* Will wake up very soon. No time for doing anything else*/
12032 if (this_rq->avg_idle < sysctl_sched_migration_cost)
12035 /* Don't need to update blocked load of idle CPUs*/
12036 if (!READ_ONCE(nohz.has_blocked) ||
12037 time_before(jiffies, READ_ONCE(nohz.next_blocked)))
12041 * Set the need to trigger ILB in order to update blocked load
12042 * before entering idle state.
12044 atomic_or(NOHZ_NEWILB_KICK, nohz_flags(this_cpu));
12047 #else /* !CONFIG_NO_HZ_COMMON */
12048 static inline void nohz_balancer_kick(struct rq *rq) { }
12050 static inline bool nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
12055 static inline void nohz_newidle_balance(struct rq *this_rq) { }
12056 #endif /* CONFIG_NO_HZ_COMMON */
12059 * newidle_balance is called by schedule() if this_cpu is about to become
12060 * idle. Attempts to pull tasks from other CPUs.
12063 * < 0 - we released the lock and there are !fair tasks present
12064 * 0 - failed, no new tasks
12065 * > 0 - success, new (fair) tasks present
12067 static int newidle_balance(struct rq *this_rq, struct rq_flags *rf)
12069 unsigned long next_balance = jiffies + HZ;
12070 int this_cpu = this_rq->cpu;
12071 u64 t0, t1, curr_cost = 0;
12072 struct sched_domain *sd;
12073 int pulled_task = 0;
12075 update_misfit_status(NULL, this_rq);
12078 * There is a task waiting to run. No need to search for one.
12079 * Return 0; the task will be enqueued when switching to idle.
12081 if (this_rq->ttwu_pending)
12085 * We must set idle_stamp _before_ calling idle_balance(), such that we
12086 * measure the duration of idle_balance() as idle time.
12088 this_rq->idle_stamp = rq_clock(this_rq);
12091 * Do not pull tasks towards !active CPUs...
12093 if (!cpu_active(this_cpu))
12097 * This is OK, because current is on_cpu, which avoids it being picked
12098 * for load-balance and preemption/IRQs are still disabled avoiding
12099 * further scheduler activity on it and we're being very careful to
12100 * re-start the picking loop.
12102 rq_unpin_lock(this_rq, rf);
12105 sd = rcu_dereference_check_sched_domain(this_rq->sd);
12107 if (!READ_ONCE(this_rq->rd->overload) ||
12108 (sd && this_rq->avg_idle < sd->max_newidle_lb_cost)) {
12111 update_next_balance(sd, &next_balance);
12118 raw_spin_rq_unlock(this_rq);
12120 t0 = sched_clock_cpu(this_cpu);
12121 update_blocked_averages(this_cpu);
12124 for_each_domain(this_cpu, sd) {
12125 int continue_balancing = 1;
12128 update_next_balance(sd, &next_balance);
12130 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost)
12133 if (sd->flags & SD_BALANCE_NEWIDLE) {
12135 pulled_task = load_balance(this_cpu, this_rq,
12136 sd, CPU_NEWLY_IDLE,
12137 &continue_balancing);
12139 t1 = sched_clock_cpu(this_cpu);
12140 domain_cost = t1 - t0;
12141 update_newidle_cost(sd, domain_cost);
12143 curr_cost += domain_cost;
12148 * Stop searching for tasks to pull if there are
12149 * now runnable tasks on this rq.
12151 if (pulled_task || this_rq->nr_running > 0 ||
12152 this_rq->ttwu_pending)
12157 raw_spin_rq_lock(this_rq);
12159 if (curr_cost > this_rq->max_idle_balance_cost)
12160 this_rq->max_idle_balance_cost = curr_cost;
12163 * While browsing the domains, we released the rq lock, a task could
12164 * have been enqueued in the meantime. Since we're not going idle,
12165 * pretend we pulled a task.
12167 if (this_rq->cfs.h_nr_running && !pulled_task)
12170 /* Is there a task of a high priority class? */
12171 if (this_rq->nr_running != this_rq->cfs.h_nr_running)
12175 /* Move the next balance forward */
12176 if (time_after(this_rq->next_balance, next_balance))
12177 this_rq->next_balance = next_balance;
12180 this_rq->idle_stamp = 0;
12182 nohz_newidle_balance(this_rq);
12184 rq_repin_lock(this_rq, rf);
12186 return pulled_task;
12190 * run_rebalance_domains is triggered when needed from the scheduler tick.
12191 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
12193 static __latent_entropy void run_rebalance_domains(struct softirq_action *h)
12195 struct rq *this_rq = this_rq();
12196 enum cpu_idle_type idle = this_rq->idle_balance ?
12197 CPU_IDLE : CPU_NOT_IDLE;
12200 * If this CPU has a pending nohz_balance_kick, then do the
12201 * balancing on behalf of the other idle CPUs whose ticks are
12202 * stopped. Do nohz_idle_balance *before* rebalance_domains to
12203 * give the idle CPUs a chance to load balance. Else we may
12204 * load balance only within the local sched_domain hierarchy
12205 * and abort nohz_idle_balance altogether if we pull some load.
12207 if (nohz_idle_balance(this_rq, idle))
12210 /* normal load balance */
12211 update_blocked_averages(this_rq->cpu);
12212 rebalance_domains(this_rq, idle);
12216 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
12218 void trigger_load_balance(struct rq *rq)
12221 * Don't need to rebalance while attached to NULL domain or
12222 * runqueue CPU is not active
12224 if (unlikely(on_null_domain(rq) || !cpu_active(cpu_of(rq))))
12227 if (time_after_eq(jiffies, rq->next_balance))
12228 raise_softirq(SCHED_SOFTIRQ);
12230 nohz_balancer_kick(rq);
12233 static void rq_online_fair(struct rq *rq)
12237 update_runtime_enabled(rq);
12240 static void rq_offline_fair(struct rq *rq)
12244 /* Ensure any throttled groups are reachable by pick_next_task */
12245 unthrottle_offline_cfs_rqs(rq);
12248 #endif /* CONFIG_SMP */
12250 #ifdef CONFIG_SCHED_CORE
12252 __entity_slice_used(struct sched_entity *se, int min_nr_tasks)
12254 u64 rtime = se->sum_exec_runtime - se->prev_sum_exec_runtime;
12255 u64 slice = se->slice;
12257 return (rtime * min_nr_tasks > slice);
12260 #define MIN_NR_TASKS_DURING_FORCEIDLE 2
12261 static inline void task_tick_core(struct rq *rq, struct task_struct *curr)
12263 if (!sched_core_enabled(rq))
12267 * If runqueue has only one task which used up its slice and
12268 * if the sibling is forced idle, then trigger schedule to
12269 * give forced idle task a chance.
12271 * sched_slice() considers only this active rq and it gets the
12272 * whole slice. But during force idle, we have siblings acting
12273 * like a single runqueue and hence we need to consider runnable
12274 * tasks on this CPU and the forced idle CPU. Ideally, we should
12275 * go through the forced idle rq, but that would be a perf hit.
12276 * We can assume that the forced idle CPU has at least
12277 * MIN_NR_TASKS_DURING_FORCEIDLE - 1 tasks and use that to check
12278 * if we need to give up the CPU.
12280 if (rq->core->core_forceidle_count && rq->cfs.nr_running == 1 &&
12281 __entity_slice_used(&curr->se, MIN_NR_TASKS_DURING_FORCEIDLE))
12286 * se_fi_update - Update the cfs_rq->min_vruntime_fi in a CFS hierarchy if needed.
12288 static void se_fi_update(const struct sched_entity *se, unsigned int fi_seq,
12291 for_each_sched_entity(se) {
12292 struct cfs_rq *cfs_rq = cfs_rq_of(se);
12295 if (cfs_rq->forceidle_seq == fi_seq)
12297 cfs_rq->forceidle_seq = fi_seq;
12300 cfs_rq->min_vruntime_fi = cfs_rq->min_vruntime;
12304 void task_vruntime_update(struct rq *rq, struct task_struct *p, bool in_fi)
12306 struct sched_entity *se = &p->se;
12308 if (p->sched_class != &fair_sched_class)
12311 se_fi_update(se, rq->core->core_forceidle_seq, in_fi);
12314 bool cfs_prio_less(const struct task_struct *a, const struct task_struct *b,
12317 struct rq *rq = task_rq(a);
12318 const struct sched_entity *sea = &a->se;
12319 const struct sched_entity *seb = &b->se;
12320 struct cfs_rq *cfs_rqa;
12321 struct cfs_rq *cfs_rqb;
12324 SCHED_WARN_ON(task_rq(b)->core != rq->core);
12326 #ifdef CONFIG_FAIR_GROUP_SCHED
12328 * Find an se in the hierarchy for tasks a and b, such that the se's
12329 * are immediate siblings.
12331 while (sea->cfs_rq->tg != seb->cfs_rq->tg) {
12332 int sea_depth = sea->depth;
12333 int seb_depth = seb->depth;
12335 if (sea_depth >= seb_depth)
12336 sea = parent_entity(sea);
12337 if (sea_depth <= seb_depth)
12338 seb = parent_entity(seb);
12341 se_fi_update(sea, rq->core->core_forceidle_seq, in_fi);
12342 se_fi_update(seb, rq->core->core_forceidle_seq, in_fi);
12344 cfs_rqa = sea->cfs_rq;
12345 cfs_rqb = seb->cfs_rq;
12347 cfs_rqa = &task_rq(a)->cfs;
12348 cfs_rqb = &task_rq(b)->cfs;
12352 * Find delta after normalizing se's vruntime with its cfs_rq's
12353 * min_vruntime_fi, which would have been updated in prior calls
12354 * to se_fi_update().
12356 delta = (s64)(sea->vruntime - seb->vruntime) +
12357 (s64)(cfs_rqb->min_vruntime_fi - cfs_rqa->min_vruntime_fi);
12362 static int task_is_throttled_fair(struct task_struct *p, int cpu)
12364 struct cfs_rq *cfs_rq;
12366 #ifdef CONFIG_FAIR_GROUP_SCHED
12367 cfs_rq = task_group(p)->cfs_rq[cpu];
12369 cfs_rq = &cpu_rq(cpu)->cfs;
12371 return throttled_hierarchy(cfs_rq);
12374 static inline void task_tick_core(struct rq *rq, struct task_struct *curr) {}
12378 * scheduler tick hitting a task of our scheduling class.
12380 * NOTE: This function can be called remotely by the tick offload that
12381 * goes along full dynticks. Therefore no local assumption can be made
12382 * and everything must be accessed through the @rq and @curr passed in
12385 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
12387 struct cfs_rq *cfs_rq;
12388 struct sched_entity *se = &curr->se;
12390 for_each_sched_entity(se) {
12391 cfs_rq = cfs_rq_of(se);
12392 entity_tick(cfs_rq, se, queued);
12395 if (static_branch_unlikely(&sched_numa_balancing))
12396 task_tick_numa(rq, curr);
12398 update_misfit_status(curr, rq);
12399 update_overutilized_status(task_rq(curr));
12401 task_tick_core(rq, curr);
12405 * called on fork with the child task as argument from the parent's context
12406 * - child not yet on the tasklist
12407 * - preemption disabled
12409 static void task_fork_fair(struct task_struct *p)
12411 struct sched_entity *se = &p->se, *curr;
12412 struct cfs_rq *cfs_rq;
12413 struct rq *rq = this_rq();
12414 struct rq_flags rf;
12417 update_rq_clock(rq);
12419 cfs_rq = task_cfs_rq(current);
12420 curr = cfs_rq->curr;
12422 update_curr(cfs_rq);
12423 place_entity(cfs_rq, se, ENQUEUE_INITIAL);
12424 rq_unlock(rq, &rf);
12428 * Priority of the task has changed. Check to see if we preempt
12429 * the current task.
12432 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
12434 if (!task_on_rq_queued(p))
12437 if (rq->cfs.nr_running == 1)
12441 * Reschedule if we are currently running on this runqueue and
12442 * our priority decreased, or if we are not currently running on
12443 * this runqueue and our priority is higher than the current's
12445 if (task_current(rq, p)) {
12446 if (p->prio > oldprio)
12449 check_preempt_curr(rq, p, 0);
12452 #ifdef CONFIG_FAIR_GROUP_SCHED
12454 * Propagate the changes of the sched_entity across the tg tree to make it
12455 * visible to the root
12457 static void propagate_entity_cfs_rq(struct sched_entity *se)
12459 struct cfs_rq *cfs_rq = cfs_rq_of(se);
12461 if (cfs_rq_throttled(cfs_rq))
12464 if (!throttled_hierarchy(cfs_rq))
12465 list_add_leaf_cfs_rq(cfs_rq);
12467 /* Start to propagate at parent */
12470 for_each_sched_entity(se) {
12471 cfs_rq = cfs_rq_of(se);
12473 update_load_avg(cfs_rq, se, UPDATE_TG);
12475 if (cfs_rq_throttled(cfs_rq))
12478 if (!throttled_hierarchy(cfs_rq))
12479 list_add_leaf_cfs_rq(cfs_rq);
12483 static void propagate_entity_cfs_rq(struct sched_entity *se) { }
12486 static void detach_entity_cfs_rq(struct sched_entity *se)
12488 struct cfs_rq *cfs_rq = cfs_rq_of(se);
12492 * In case the task sched_avg hasn't been attached:
12493 * - A forked task which hasn't been woken up by wake_up_new_task().
12494 * - A task which has been woken up by try_to_wake_up() but is
12495 * waiting for actually being woken up by sched_ttwu_pending().
12497 if (!se->avg.last_update_time)
12501 /* Catch up with the cfs_rq and remove our load when we leave */
12502 update_load_avg(cfs_rq, se, 0);
12503 detach_entity_load_avg(cfs_rq, se);
12504 update_tg_load_avg(cfs_rq);
12505 propagate_entity_cfs_rq(se);
12508 static void attach_entity_cfs_rq(struct sched_entity *se)
12510 struct cfs_rq *cfs_rq = cfs_rq_of(se);
12512 /* Synchronize entity with its cfs_rq */
12513 update_load_avg(cfs_rq, se, sched_feat(ATTACH_AGE_LOAD) ? 0 : SKIP_AGE_LOAD);
12514 attach_entity_load_avg(cfs_rq, se);
12515 update_tg_load_avg(cfs_rq);
12516 propagate_entity_cfs_rq(se);
12519 static void detach_task_cfs_rq(struct task_struct *p)
12521 struct sched_entity *se = &p->se;
12523 detach_entity_cfs_rq(se);
12526 static void attach_task_cfs_rq(struct task_struct *p)
12528 struct sched_entity *se = &p->se;
12530 attach_entity_cfs_rq(se);
12533 static void switched_from_fair(struct rq *rq, struct task_struct *p)
12535 detach_task_cfs_rq(p);
12538 static void switched_to_fair(struct rq *rq, struct task_struct *p)
12540 attach_task_cfs_rq(p);
12542 if (task_on_rq_queued(p)) {
12544 * We were most likely switched from sched_rt, so
12545 * kick off the schedule if running, otherwise just see
12546 * if we can still preempt the current task.
12548 if (task_current(rq, p))
12551 check_preempt_curr(rq, p, 0);
12555 /* Account for a task changing its policy or group.
12557 * This routine is mostly called to set cfs_rq->curr field when a task
12558 * migrates between groups/classes.
12560 static void set_next_task_fair(struct rq *rq, struct task_struct *p, bool first)
12562 struct sched_entity *se = &p->se;
12565 if (task_on_rq_queued(p)) {
12567 * Move the next running task to the front of the list, so our
12568 * cfs_tasks list becomes MRU one.
12570 list_move(&se->group_node, &rq->cfs_tasks);
12574 for_each_sched_entity(se) {
12575 struct cfs_rq *cfs_rq = cfs_rq_of(se);
12577 set_next_entity(cfs_rq, se);
12578 /* ensure bandwidth has been allocated on our new cfs_rq */
12579 account_cfs_rq_runtime(cfs_rq, 0);
12583 void init_cfs_rq(struct cfs_rq *cfs_rq)
12585 cfs_rq->tasks_timeline = RB_ROOT_CACHED;
12586 u64_u32_store(cfs_rq->min_vruntime, (u64)(-(1LL << 20)));
12588 raw_spin_lock_init(&cfs_rq->removed.lock);
12592 #ifdef CONFIG_FAIR_GROUP_SCHED
12593 static void task_change_group_fair(struct task_struct *p)
12596 * We couldn't detach or attach a forked task which
12597 * hasn't been woken up by wake_up_new_task().
12599 if (READ_ONCE(p->__state) == TASK_NEW)
12602 detach_task_cfs_rq(p);
12605 /* Tell se's cfs_rq has been changed -- migrated */
12606 p->se.avg.last_update_time = 0;
12608 set_task_rq(p, task_cpu(p));
12609 attach_task_cfs_rq(p);
12612 void free_fair_sched_group(struct task_group *tg)
12616 for_each_possible_cpu(i) {
12618 kfree(tg->cfs_rq[i]);
12627 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
12629 struct sched_entity *se;
12630 struct cfs_rq *cfs_rq;
12633 tg->cfs_rq = kcalloc(nr_cpu_ids, sizeof(cfs_rq), GFP_KERNEL);
12636 tg->se = kcalloc(nr_cpu_ids, sizeof(se), GFP_KERNEL);
12640 tg->shares = NICE_0_LOAD;
12642 init_cfs_bandwidth(tg_cfs_bandwidth(tg), tg_cfs_bandwidth(parent));
12644 for_each_possible_cpu(i) {
12645 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
12646 GFP_KERNEL, cpu_to_node(i));
12650 se = kzalloc_node(sizeof(struct sched_entity_stats),
12651 GFP_KERNEL, cpu_to_node(i));
12655 init_cfs_rq(cfs_rq);
12656 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
12657 init_entity_runnable_average(se);
12668 void online_fair_sched_group(struct task_group *tg)
12670 struct sched_entity *se;
12671 struct rq_flags rf;
12675 for_each_possible_cpu(i) {
12678 rq_lock_irq(rq, &rf);
12679 update_rq_clock(rq);
12680 attach_entity_cfs_rq(se);
12681 sync_throttle(tg, i);
12682 rq_unlock_irq(rq, &rf);
12686 void unregister_fair_sched_group(struct task_group *tg)
12688 unsigned long flags;
12692 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
12694 for_each_possible_cpu(cpu) {
12696 remove_entity_load_avg(tg->se[cpu]);
12699 * Only empty task groups can be destroyed; so we can speculatively
12700 * check on_list without danger of it being re-added.
12702 if (!tg->cfs_rq[cpu]->on_list)
12707 raw_spin_rq_lock_irqsave(rq, flags);
12708 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
12709 raw_spin_rq_unlock_irqrestore(rq, flags);
12713 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
12714 struct sched_entity *se, int cpu,
12715 struct sched_entity *parent)
12717 struct rq *rq = cpu_rq(cpu);
12721 init_cfs_rq_runtime(cfs_rq);
12723 tg->cfs_rq[cpu] = cfs_rq;
12726 /* se could be NULL for root_task_group */
12731 se->cfs_rq = &rq->cfs;
12734 se->cfs_rq = parent->my_q;
12735 se->depth = parent->depth + 1;
12739 /* guarantee group entities always have weight */
12740 update_load_set(&se->load, NICE_0_LOAD);
12741 se->parent = parent;
12744 static DEFINE_MUTEX(shares_mutex);
12746 static int __sched_group_set_shares(struct task_group *tg, unsigned long shares)
12750 lockdep_assert_held(&shares_mutex);
12753 * We can't change the weight of the root cgroup.
12758 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
12760 if (tg->shares == shares)
12763 tg->shares = shares;
12764 for_each_possible_cpu(i) {
12765 struct rq *rq = cpu_rq(i);
12766 struct sched_entity *se = tg->se[i];
12767 struct rq_flags rf;
12769 /* Propagate contribution to hierarchy */
12770 rq_lock_irqsave(rq, &rf);
12771 update_rq_clock(rq);
12772 for_each_sched_entity(se) {
12773 update_load_avg(cfs_rq_of(se), se, UPDATE_TG);
12774 update_cfs_group(se);
12776 rq_unlock_irqrestore(rq, &rf);
12782 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
12786 mutex_lock(&shares_mutex);
12787 if (tg_is_idle(tg))
12790 ret = __sched_group_set_shares(tg, shares);
12791 mutex_unlock(&shares_mutex);
12796 int sched_group_set_idle(struct task_group *tg, long idle)
12800 if (tg == &root_task_group)
12803 if (idle < 0 || idle > 1)
12806 mutex_lock(&shares_mutex);
12808 if (tg->idle == idle) {
12809 mutex_unlock(&shares_mutex);
12815 for_each_possible_cpu(i) {
12816 struct rq *rq = cpu_rq(i);
12817 struct sched_entity *se = tg->se[i];
12818 struct cfs_rq *parent_cfs_rq, *grp_cfs_rq = tg->cfs_rq[i];
12819 bool was_idle = cfs_rq_is_idle(grp_cfs_rq);
12820 long idle_task_delta;
12821 struct rq_flags rf;
12823 rq_lock_irqsave(rq, &rf);
12825 grp_cfs_rq->idle = idle;
12826 if (WARN_ON_ONCE(was_idle == cfs_rq_is_idle(grp_cfs_rq)))
12830 parent_cfs_rq = cfs_rq_of(se);
12831 if (cfs_rq_is_idle(grp_cfs_rq))
12832 parent_cfs_rq->idle_nr_running++;
12834 parent_cfs_rq->idle_nr_running--;
12837 idle_task_delta = grp_cfs_rq->h_nr_running -
12838 grp_cfs_rq->idle_h_nr_running;
12839 if (!cfs_rq_is_idle(grp_cfs_rq))
12840 idle_task_delta *= -1;
12842 for_each_sched_entity(se) {
12843 struct cfs_rq *cfs_rq = cfs_rq_of(se);
12848 cfs_rq->idle_h_nr_running += idle_task_delta;
12850 /* Already accounted at parent level and above. */
12851 if (cfs_rq_is_idle(cfs_rq))
12856 rq_unlock_irqrestore(rq, &rf);
12859 /* Idle groups have minimum weight. */
12860 if (tg_is_idle(tg))
12861 __sched_group_set_shares(tg, scale_load(WEIGHT_IDLEPRIO));
12863 __sched_group_set_shares(tg, NICE_0_LOAD);
12865 mutex_unlock(&shares_mutex);
12869 #else /* CONFIG_FAIR_GROUP_SCHED */
12871 void free_fair_sched_group(struct task_group *tg) { }
12873 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
12878 void online_fair_sched_group(struct task_group *tg) { }
12880 void unregister_fair_sched_group(struct task_group *tg) { }
12882 #endif /* CONFIG_FAIR_GROUP_SCHED */
12885 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
12887 struct sched_entity *se = &task->se;
12888 unsigned int rr_interval = 0;
12891 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
12894 if (rq->cfs.load.weight)
12895 rr_interval = NS_TO_JIFFIES(se->slice);
12897 return rr_interval;
12901 * All the scheduling class methods:
12903 DEFINE_SCHED_CLASS(fair) = {
12905 .enqueue_task = enqueue_task_fair,
12906 .dequeue_task = dequeue_task_fair,
12907 .yield_task = yield_task_fair,
12908 .yield_to_task = yield_to_task_fair,
12910 .check_preempt_curr = check_preempt_wakeup,
12912 .pick_next_task = __pick_next_task_fair,
12913 .put_prev_task = put_prev_task_fair,
12914 .set_next_task = set_next_task_fair,
12917 .balance = balance_fair,
12918 .pick_task = pick_task_fair,
12919 .select_task_rq = select_task_rq_fair,
12920 .migrate_task_rq = migrate_task_rq_fair,
12922 .rq_online = rq_online_fair,
12923 .rq_offline = rq_offline_fair,
12925 .task_dead = task_dead_fair,
12926 .set_cpus_allowed = set_cpus_allowed_common,
12929 .task_tick = task_tick_fair,
12930 .task_fork = task_fork_fair,
12932 .prio_changed = prio_changed_fair,
12933 .switched_from = switched_from_fair,
12934 .switched_to = switched_to_fair,
12936 .get_rr_interval = get_rr_interval_fair,
12938 .update_curr = update_curr_fair,
12940 #ifdef CONFIG_FAIR_GROUP_SCHED
12941 .task_change_group = task_change_group_fair,
12944 #ifdef CONFIG_SCHED_CORE
12945 .task_is_throttled = task_is_throttled_fair,
12948 #ifdef CONFIG_UCLAMP_TASK
12949 .uclamp_enabled = 1,
12953 #ifdef CONFIG_SCHED_DEBUG
12954 void print_cfs_stats(struct seq_file *m, int cpu)
12956 struct cfs_rq *cfs_rq, *pos;
12959 for_each_leaf_cfs_rq_safe(cpu_rq(cpu), cfs_rq, pos)
12960 print_cfs_rq(m, cpu, cfs_rq);
12964 #ifdef CONFIG_NUMA_BALANCING
12965 void show_numa_stats(struct task_struct *p, struct seq_file *m)
12968 unsigned long tsf = 0, tpf = 0, gsf = 0, gpf = 0;
12969 struct numa_group *ng;
12972 ng = rcu_dereference(p->numa_group);
12973 for_each_online_node(node) {
12974 if (p->numa_faults) {
12975 tsf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 0)];
12976 tpf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 1)];
12979 gsf = ng->faults[task_faults_idx(NUMA_MEM, node, 0)],
12980 gpf = ng->faults[task_faults_idx(NUMA_MEM, node, 1)];
12982 print_numa_stats(m, node, tsf, tpf, gsf, gpf);
12986 #endif /* CONFIG_NUMA_BALANCING */
12987 #endif /* CONFIG_SCHED_DEBUG */
12989 __init void init_sched_fair_class(void)
12994 for_each_possible_cpu(i) {
12995 zalloc_cpumask_var_node(&per_cpu(load_balance_mask, i), GFP_KERNEL, cpu_to_node(i));
12996 zalloc_cpumask_var_node(&per_cpu(select_rq_mask, i), GFP_KERNEL, cpu_to_node(i));
12997 zalloc_cpumask_var_node(&per_cpu(should_we_balance_tmpmask, i),
12998 GFP_KERNEL, cpu_to_node(i));
13000 #ifdef CONFIG_CFS_BANDWIDTH
13001 INIT_CSD(&cpu_rq(i)->cfsb_csd, __cfsb_csd_unthrottle, cpu_rq(i));
13002 INIT_LIST_HEAD(&cpu_rq(i)->cfsb_csd_list);
13006 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
13008 #ifdef CONFIG_NO_HZ_COMMON
13009 nohz.next_balance = jiffies;
13010 nohz.next_blocked = jiffies;
13011 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);