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_eevdf(struct cfs_rq *cfs_rq, struct sched_entity *se,
3630 unsigned long weight)
3632 unsigned long old_weight = se->load.weight;
3633 u64 avruntime = avg_vruntime(cfs_rq);
3640 * COROLLARY #1: The virtual runtime of the entity needs to be
3641 * adjusted if re-weight at !0-lag point.
3643 * Proof: For contradiction assume this is not true, so we can
3644 * re-weight without changing vruntime at !0-lag point.
3646 * Weight VRuntime Avg-VRuntime
3650 * Since lag needs to be preserved through re-weight:
3652 * lag = (V - v)*w = (V'- v')*w', where v = v'
3653 * ==> V' = (V - v)*w/w' + v (1)
3655 * Let W be the total weight of the entities before reweight,
3656 * since V' is the new weighted average of entities:
3658 * V' = (WV + w'v - wv) / (W + w' - w) (2)
3660 * by using (1) & (2) we obtain:
3662 * (WV + w'v - wv) / (W + w' - w) = (V - v)*w/w' + v
3663 * ==> (WV-Wv+Wv+w'v-wv)/(W+w'-w) = (V - v)*w/w' + v
3664 * ==> (WV - Wv)/(W + w' - w) + v = (V - v)*w/w' + v
3665 * ==> (V - v)*W/(W + w' - w) = (V - v)*w/w' (3)
3667 * Since we are doing at !0-lag point which means V != v, we
3670 * ==> W / (W + w' - w) = w / w'
3671 * ==> Ww' = Ww + ww' - ww
3672 * ==> W * (w' - w) = w * (w' - w)
3673 * ==> W = w (re-weight indicates w' != w)
3675 * So the cfs_rq contains only one entity, hence vruntime of
3676 * the entity @v should always equal to the cfs_rq's weighted
3677 * average vruntime @V, which means we will always re-weight
3678 * at 0-lag point, thus breach assumption. Proof completed.
3681 * COROLLARY #2: Re-weight does NOT affect weighted average
3682 * vruntime of all the entities.
3684 * Proof: According to corollary #1, Eq. (1) should be:
3686 * (V - v)*w = (V' - v')*w'
3687 * ==> v' = V' - (V - v)*w/w' (4)
3689 * According to the weighted average formula, we have:
3691 * V' = (WV - wv + w'v') / (W - w + w')
3692 * = (WV - wv + w'(V' - (V - v)w/w')) / (W - w + w')
3693 * = (WV - wv + w'V' - Vw + wv) / (W - w + w')
3694 * = (WV + w'V' - Vw) / (W - w + w')
3696 * ==> V'*(W - w + w') = WV + w'V' - Vw
3697 * ==> V' * (W - w) = (W - w) * V (5)
3699 * If the entity is the only one in the cfs_rq, then reweight
3700 * always occurs at 0-lag point, so V won't change. Or else
3701 * there are other entities, hence W != w, then Eq. (5) turns
3702 * into V' = V. So V won't change in either case, proof done.
3705 * So according to corollary #1 & #2, the effect of re-weight
3706 * on vruntime should be:
3708 * v' = V' - (V - v) * w / w' (4)
3709 * = V - (V - v) * w / w'
3713 if (avruntime != se->vruntime) {
3714 vlag = (s64)(avruntime - se->vruntime);
3715 vlag = div_s64(vlag * old_weight, weight);
3716 se->vruntime = avruntime - vlag;
3723 * When the weight changes, the virtual time slope changes and
3724 * we should adjust the relative virtual deadline accordingly.
3726 * d' = v' + (d - v)*w/w'
3727 * = V' - (V - v)*w/w' + (d - v)*w/w'
3728 * = V - (V - v)*w/w' + (d - v)*w/w'
3729 * = V + (d - V)*w/w'
3731 vslice = (s64)(se->deadline - avruntime);
3732 vslice = div_s64(vslice * old_weight, weight);
3733 se->deadline = avruntime + vslice;
3736 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
3737 unsigned long weight)
3739 bool curr = cfs_rq->curr == se;
3742 /* commit outstanding execution time */
3744 update_curr(cfs_rq);
3746 __dequeue_entity(cfs_rq, se);
3747 update_load_sub(&cfs_rq->load, se->load.weight);
3749 dequeue_load_avg(cfs_rq, se);
3753 * Because we keep se->vlag = V - v_i, while: lag_i = w_i*(V - v_i),
3754 * we need to scale se->vlag when w_i changes.
3756 se->vlag = div_s64(se->vlag * se->load.weight, weight);
3758 reweight_eevdf(cfs_rq, se, weight);
3761 update_load_set(&se->load, weight);
3765 u32 divider = get_pelt_divider(&se->avg);
3767 se->avg.load_avg = div_u64(se_weight(se) * se->avg.load_sum, divider);
3771 enqueue_load_avg(cfs_rq, se);
3773 update_load_add(&cfs_rq->load, se->load.weight);
3776 * The entity's vruntime has been adjusted, so let's check
3777 * whether the rq-wide min_vruntime needs updated too. Since
3778 * the calculations above require stable min_vruntime rather
3779 * than up-to-date one, we do the update at the end of the
3782 __enqueue_entity(cfs_rq, se);
3783 update_min_vruntime(cfs_rq);
3788 void reweight_task(struct task_struct *p, int prio)
3790 struct sched_entity *se = &p->se;
3791 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3792 struct load_weight *load = &se->load;
3793 unsigned long weight = scale_load(sched_prio_to_weight[prio]);
3795 reweight_entity(cfs_rq, se, weight);
3796 load->inv_weight = sched_prio_to_wmult[prio];
3799 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
3801 #ifdef CONFIG_FAIR_GROUP_SCHED
3804 * All this does is approximate the hierarchical proportion which includes that
3805 * global sum we all love to hate.
3807 * That is, the weight of a group entity, is the proportional share of the
3808 * group weight based on the group runqueue weights. That is:
3810 * tg->weight * grq->load.weight
3811 * ge->load.weight = ----------------------------- (1)
3812 * \Sum grq->load.weight
3814 * Now, because computing that sum is prohibitively expensive to compute (been
3815 * there, done that) we approximate it with this average stuff. The average
3816 * moves slower and therefore the approximation is cheaper and more stable.
3818 * So instead of the above, we substitute:
3820 * grq->load.weight -> grq->avg.load_avg (2)
3822 * which yields the following:
3824 * tg->weight * grq->avg.load_avg
3825 * ge->load.weight = ------------------------------ (3)
3828 * Where: tg->load_avg ~= \Sum grq->avg.load_avg
3830 * That is shares_avg, and it is right (given the approximation (2)).
3832 * The problem with it is that because the average is slow -- it was designed
3833 * to be exactly that of course -- this leads to transients in boundary
3834 * conditions. In specific, the case where the group was idle and we start the
3835 * one task. It takes time for our CPU's grq->avg.load_avg to build up,
3836 * yielding bad latency etc..
3838 * Now, in that special case (1) reduces to:
3840 * tg->weight * grq->load.weight
3841 * ge->load.weight = ----------------------------- = tg->weight (4)
3844 * That is, the sum collapses because all other CPUs are idle; the UP scenario.
3846 * So what we do is modify our approximation (3) to approach (4) in the (near)
3851 * tg->weight * grq->load.weight
3852 * --------------------------------------------------- (5)
3853 * tg->load_avg - grq->avg.load_avg + grq->load.weight
3855 * But because grq->load.weight can drop to 0, resulting in a divide by zero,
3856 * we need to use grq->avg.load_avg as its lower bound, which then gives:
3859 * tg->weight * grq->load.weight
3860 * ge->load.weight = ----------------------------- (6)
3865 * tg_load_avg' = tg->load_avg - grq->avg.load_avg +
3866 * max(grq->load.weight, grq->avg.load_avg)
3868 * And that is shares_weight and is icky. In the (near) UP case it approaches
3869 * (4) while in the normal case it approaches (3). It consistently
3870 * overestimates the ge->load.weight and therefore:
3872 * \Sum ge->load.weight >= tg->weight
3876 static long calc_group_shares(struct cfs_rq *cfs_rq)
3878 long tg_weight, tg_shares, load, shares;
3879 struct task_group *tg = cfs_rq->tg;
3881 tg_shares = READ_ONCE(tg->shares);
3883 load = max(scale_load_down(cfs_rq->load.weight), cfs_rq->avg.load_avg);
3885 tg_weight = atomic_long_read(&tg->load_avg);
3887 /* Ensure tg_weight >= load */
3888 tg_weight -= cfs_rq->tg_load_avg_contrib;
3891 shares = (tg_shares * load);
3893 shares /= tg_weight;
3896 * MIN_SHARES has to be unscaled here to support per-CPU partitioning
3897 * of a group with small tg->shares value. It is a floor value which is
3898 * assigned as a minimum load.weight to the sched_entity representing
3899 * the group on a CPU.
3901 * E.g. on 64-bit for a group with tg->shares of scale_load(15)=15*1024
3902 * on an 8-core system with 8 tasks each runnable on one CPU shares has
3903 * to be 15*1024*1/8=1920 instead of scale_load(MIN_SHARES)=2*1024. In
3904 * case no task is runnable on a CPU MIN_SHARES=2 should be returned
3907 return clamp_t(long, shares, MIN_SHARES, tg_shares);
3909 #endif /* CONFIG_SMP */
3912 * Recomputes the group entity based on the current state of its group
3915 static void update_cfs_group(struct sched_entity *se)
3917 struct cfs_rq *gcfs_rq = group_cfs_rq(se);
3923 if (throttled_hierarchy(gcfs_rq))
3927 shares = READ_ONCE(gcfs_rq->tg->shares);
3929 shares = calc_group_shares(gcfs_rq);
3931 if (unlikely(se->load.weight != shares))
3932 reweight_entity(cfs_rq_of(se), se, shares);
3935 #else /* CONFIG_FAIR_GROUP_SCHED */
3936 static inline void update_cfs_group(struct sched_entity *se)
3939 #endif /* CONFIG_FAIR_GROUP_SCHED */
3941 static inline void cfs_rq_util_change(struct cfs_rq *cfs_rq, int flags)
3943 struct rq *rq = rq_of(cfs_rq);
3945 if (&rq->cfs == cfs_rq) {
3947 * There are a few boundary cases this might miss but it should
3948 * get called often enough that that should (hopefully) not be
3951 * It will not get called when we go idle, because the idle
3952 * thread is a different class (!fair), nor will the utilization
3953 * number include things like RT tasks.
3955 * As is, the util number is not freq-invariant (we'd have to
3956 * implement arch_scale_freq_capacity() for that).
3958 * See cpu_util_cfs().
3960 cpufreq_update_util(rq, flags);
3965 static inline bool load_avg_is_decayed(struct sched_avg *sa)
3973 if (sa->runnable_sum)
3977 * _avg must be null when _sum are null because _avg = _sum / divider
3978 * Make sure that rounding and/or propagation of PELT values never
3981 SCHED_WARN_ON(sa->load_avg ||
3988 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
3990 return u64_u32_load_copy(cfs_rq->avg.last_update_time,
3991 cfs_rq->last_update_time_copy);
3993 #ifdef CONFIG_FAIR_GROUP_SCHED
3995 * Because list_add_leaf_cfs_rq always places a child cfs_rq on the list
3996 * immediately before a parent cfs_rq, and cfs_rqs are removed from the list
3997 * bottom-up, we only have to test whether the cfs_rq before us on the list
3999 * If cfs_rq is not on the list, test whether a child needs its to be added to
4000 * connect a branch to the tree * (see list_add_leaf_cfs_rq() for details).
4002 static inline bool child_cfs_rq_on_list(struct cfs_rq *cfs_rq)
4004 struct cfs_rq *prev_cfs_rq;
4005 struct list_head *prev;
4007 if (cfs_rq->on_list) {
4008 prev = cfs_rq->leaf_cfs_rq_list.prev;
4010 struct rq *rq = rq_of(cfs_rq);
4012 prev = rq->tmp_alone_branch;
4015 prev_cfs_rq = container_of(prev, struct cfs_rq, leaf_cfs_rq_list);
4017 return (prev_cfs_rq->tg->parent == cfs_rq->tg);
4020 static inline bool cfs_rq_is_decayed(struct cfs_rq *cfs_rq)
4022 if (cfs_rq->load.weight)
4025 if (!load_avg_is_decayed(&cfs_rq->avg))
4028 if (child_cfs_rq_on_list(cfs_rq))
4035 * update_tg_load_avg - update the tg's load avg
4036 * @cfs_rq: the cfs_rq whose avg changed
4038 * This function 'ensures': tg->load_avg := \Sum tg->cfs_rq[]->avg.load.
4039 * However, because tg->load_avg is a global value there are performance
4042 * In order to avoid having to look at the other cfs_rq's, we use a
4043 * differential update where we store the last value we propagated. This in
4044 * turn allows skipping updates if the differential is 'small'.
4046 * Updating tg's load_avg is necessary before update_cfs_share().
4048 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq)
4050 long delta = cfs_rq->avg.load_avg - cfs_rq->tg_load_avg_contrib;
4053 * No need to update load_avg for root_task_group as it is not used.
4055 if (cfs_rq->tg == &root_task_group)
4058 if (abs(delta) > cfs_rq->tg_load_avg_contrib / 64) {
4059 atomic_long_add(delta, &cfs_rq->tg->load_avg);
4060 cfs_rq->tg_load_avg_contrib = cfs_rq->avg.load_avg;
4065 * Called within set_task_rq() right before setting a task's CPU. The
4066 * caller only guarantees p->pi_lock is held; no other assumptions,
4067 * including the state of rq->lock, should be made.
4069 void set_task_rq_fair(struct sched_entity *se,
4070 struct cfs_rq *prev, struct cfs_rq *next)
4072 u64 p_last_update_time;
4073 u64 n_last_update_time;
4075 if (!sched_feat(ATTACH_AGE_LOAD))
4079 * We are supposed to update the task to "current" time, then its up to
4080 * date and ready to go to new CPU/cfs_rq. But we have difficulty in
4081 * getting what current time is, so simply throw away the out-of-date
4082 * time. This will result in the wakee task is less decayed, but giving
4083 * the wakee more load sounds not bad.
4085 if (!(se->avg.last_update_time && prev))
4088 p_last_update_time = cfs_rq_last_update_time(prev);
4089 n_last_update_time = cfs_rq_last_update_time(next);
4091 __update_load_avg_blocked_se(p_last_update_time, se);
4092 se->avg.last_update_time = n_last_update_time;
4096 * When on migration a sched_entity joins/leaves the PELT hierarchy, we need to
4097 * propagate its contribution. The key to this propagation is the invariant
4098 * that for each group:
4100 * ge->avg == grq->avg (1)
4102 * _IFF_ we look at the pure running and runnable sums. Because they
4103 * represent the very same entity, just at different points in the hierarchy.
4105 * Per the above update_tg_cfs_util() and update_tg_cfs_runnable() are trivial
4106 * and simply copies the running/runnable sum over (but still wrong, because
4107 * the group entity and group rq do not have their PELT windows aligned).
4109 * However, update_tg_cfs_load() is more complex. So we have:
4111 * ge->avg.load_avg = ge->load.weight * ge->avg.runnable_avg (2)
4113 * And since, like util, the runnable part should be directly transferable,
4114 * the following would _appear_ to be the straight forward approach:
4116 * grq->avg.load_avg = grq->load.weight * grq->avg.runnable_avg (3)
4118 * And per (1) we have:
4120 * ge->avg.runnable_avg == grq->avg.runnable_avg
4124 * ge->load.weight * grq->avg.load_avg
4125 * ge->avg.load_avg = ----------------------------------- (4)
4128 * Except that is wrong!
4130 * Because while for entities historical weight is not important and we
4131 * really only care about our future and therefore can consider a pure
4132 * runnable sum, runqueues can NOT do this.
4134 * We specifically want runqueues to have a load_avg that includes
4135 * historical weights. Those represent the blocked load, the load we expect
4136 * to (shortly) return to us. This only works by keeping the weights as
4137 * integral part of the sum. We therefore cannot decompose as per (3).
4139 * Another reason this doesn't work is that runnable isn't a 0-sum entity.
4140 * Imagine a rq with 2 tasks that each are runnable 2/3 of the time. Then the
4141 * rq itself is runnable anywhere between 2/3 and 1 depending on how the
4142 * runnable section of these tasks overlap (or not). If they were to perfectly
4143 * align the rq as a whole would be runnable 2/3 of the time. If however we
4144 * always have at least 1 runnable task, the rq as a whole is always runnable.
4146 * So we'll have to approximate.. :/
4148 * Given the constraint:
4150 * ge->avg.running_sum <= ge->avg.runnable_sum <= LOAD_AVG_MAX
4152 * We can construct a rule that adds runnable to a rq by assuming minimal
4155 * On removal, we'll assume each task is equally runnable; which yields:
4157 * grq->avg.runnable_sum = grq->avg.load_sum / grq->load.weight
4159 * XXX: only do this for the part of runnable > running ?
4163 update_tg_cfs_util(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq)
4165 long delta_sum, delta_avg = gcfs_rq->avg.util_avg - se->avg.util_avg;
4166 u32 new_sum, divider;
4168 /* Nothing to update */
4173 * cfs_rq->avg.period_contrib can be used for both cfs_rq and se.
4174 * See ___update_load_avg() for details.
4176 divider = get_pelt_divider(&cfs_rq->avg);
4179 /* Set new sched_entity's utilization */
4180 se->avg.util_avg = gcfs_rq->avg.util_avg;
4181 new_sum = se->avg.util_avg * divider;
4182 delta_sum = (long)new_sum - (long)se->avg.util_sum;
4183 se->avg.util_sum = new_sum;
4185 /* Update parent cfs_rq utilization */
4186 add_positive(&cfs_rq->avg.util_avg, delta_avg);
4187 add_positive(&cfs_rq->avg.util_sum, delta_sum);
4189 /* See update_cfs_rq_load_avg() */
4190 cfs_rq->avg.util_sum = max_t(u32, cfs_rq->avg.util_sum,
4191 cfs_rq->avg.util_avg * PELT_MIN_DIVIDER);
4195 update_tg_cfs_runnable(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq)
4197 long delta_sum, delta_avg = gcfs_rq->avg.runnable_avg - se->avg.runnable_avg;
4198 u32 new_sum, divider;
4200 /* Nothing to update */
4205 * cfs_rq->avg.period_contrib can be used for both cfs_rq and se.
4206 * See ___update_load_avg() for details.
4208 divider = get_pelt_divider(&cfs_rq->avg);
4210 /* Set new sched_entity's runnable */
4211 se->avg.runnable_avg = gcfs_rq->avg.runnable_avg;
4212 new_sum = se->avg.runnable_avg * divider;
4213 delta_sum = (long)new_sum - (long)se->avg.runnable_sum;
4214 se->avg.runnable_sum = new_sum;
4216 /* Update parent cfs_rq runnable */
4217 add_positive(&cfs_rq->avg.runnable_avg, delta_avg);
4218 add_positive(&cfs_rq->avg.runnable_sum, delta_sum);
4219 /* See update_cfs_rq_load_avg() */
4220 cfs_rq->avg.runnable_sum = max_t(u32, cfs_rq->avg.runnable_sum,
4221 cfs_rq->avg.runnable_avg * PELT_MIN_DIVIDER);
4225 update_tg_cfs_load(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq)
4227 long delta_avg, running_sum, runnable_sum = gcfs_rq->prop_runnable_sum;
4228 unsigned long load_avg;
4236 gcfs_rq->prop_runnable_sum = 0;
4239 * cfs_rq->avg.period_contrib can be used for both cfs_rq and se.
4240 * See ___update_load_avg() for details.
4242 divider = get_pelt_divider(&cfs_rq->avg);
4244 if (runnable_sum >= 0) {
4246 * Add runnable; clip at LOAD_AVG_MAX. Reflects that until
4247 * the CPU is saturated running == runnable.
4249 runnable_sum += se->avg.load_sum;
4250 runnable_sum = min_t(long, runnable_sum, divider);
4253 * Estimate the new unweighted runnable_sum of the gcfs_rq by
4254 * assuming all tasks are equally runnable.
4256 if (scale_load_down(gcfs_rq->load.weight)) {
4257 load_sum = div_u64(gcfs_rq->avg.load_sum,
4258 scale_load_down(gcfs_rq->load.weight));
4261 /* But make sure to not inflate se's runnable */
4262 runnable_sum = min(se->avg.load_sum, load_sum);
4266 * runnable_sum can't be lower than running_sum
4267 * Rescale running sum to be in the same range as runnable sum
4268 * running_sum is in [0 : LOAD_AVG_MAX << SCHED_CAPACITY_SHIFT]
4269 * runnable_sum is in [0 : LOAD_AVG_MAX]
4271 running_sum = se->avg.util_sum >> SCHED_CAPACITY_SHIFT;
4272 runnable_sum = max(runnable_sum, running_sum);
4274 load_sum = se_weight(se) * runnable_sum;
4275 load_avg = div_u64(load_sum, divider);
4277 delta_avg = load_avg - se->avg.load_avg;
4281 delta_sum = load_sum - (s64)se_weight(se) * se->avg.load_sum;
4283 se->avg.load_sum = runnable_sum;
4284 se->avg.load_avg = load_avg;
4285 add_positive(&cfs_rq->avg.load_avg, delta_avg);
4286 add_positive(&cfs_rq->avg.load_sum, delta_sum);
4287 /* See update_cfs_rq_load_avg() */
4288 cfs_rq->avg.load_sum = max_t(u32, cfs_rq->avg.load_sum,
4289 cfs_rq->avg.load_avg * PELT_MIN_DIVIDER);
4292 static inline void add_tg_cfs_propagate(struct cfs_rq *cfs_rq, long runnable_sum)
4294 cfs_rq->propagate = 1;
4295 cfs_rq->prop_runnable_sum += runnable_sum;
4298 /* Update task and its cfs_rq load average */
4299 static inline int propagate_entity_load_avg(struct sched_entity *se)
4301 struct cfs_rq *cfs_rq, *gcfs_rq;
4303 if (entity_is_task(se))
4306 gcfs_rq = group_cfs_rq(se);
4307 if (!gcfs_rq->propagate)
4310 gcfs_rq->propagate = 0;
4312 cfs_rq = cfs_rq_of(se);
4314 add_tg_cfs_propagate(cfs_rq, gcfs_rq->prop_runnable_sum);
4316 update_tg_cfs_util(cfs_rq, se, gcfs_rq);
4317 update_tg_cfs_runnable(cfs_rq, se, gcfs_rq);
4318 update_tg_cfs_load(cfs_rq, se, gcfs_rq);
4320 trace_pelt_cfs_tp(cfs_rq);
4321 trace_pelt_se_tp(se);
4327 * Check if we need to update the load and the utilization of a blocked
4330 static inline bool skip_blocked_update(struct sched_entity *se)
4332 struct cfs_rq *gcfs_rq = group_cfs_rq(se);
4335 * If sched_entity still have not zero load or utilization, we have to
4338 if (se->avg.load_avg || se->avg.util_avg)
4342 * If there is a pending propagation, we have to update the load and
4343 * the utilization of the sched_entity:
4345 if (gcfs_rq->propagate)
4349 * Otherwise, the load and the utilization of the sched_entity is
4350 * already zero and there is no pending propagation, so it will be a
4351 * waste of time to try to decay it:
4356 #else /* CONFIG_FAIR_GROUP_SCHED */
4358 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq) {}
4360 static inline int propagate_entity_load_avg(struct sched_entity *se)
4365 static inline void add_tg_cfs_propagate(struct cfs_rq *cfs_rq, long runnable_sum) {}
4367 #endif /* CONFIG_FAIR_GROUP_SCHED */
4369 #ifdef CONFIG_NO_HZ_COMMON
4370 static inline void migrate_se_pelt_lag(struct sched_entity *se)
4372 u64 throttled = 0, now, lut;
4373 struct cfs_rq *cfs_rq;
4377 if (load_avg_is_decayed(&se->avg))
4380 cfs_rq = cfs_rq_of(se);
4384 is_idle = is_idle_task(rcu_dereference(rq->curr));
4388 * The lag estimation comes with a cost we don't want to pay all the
4389 * time. Hence, limiting to the case where the source CPU is idle and
4390 * we know we are at the greatest risk to have an outdated clock.
4396 * Estimated "now" is: last_update_time + cfs_idle_lag + rq_idle_lag, where:
4398 * last_update_time (the cfs_rq's last_update_time)
4399 * = cfs_rq_clock_pelt()@cfs_rq_idle
4400 * = rq_clock_pelt()@cfs_rq_idle
4401 * - cfs->throttled_clock_pelt_time@cfs_rq_idle
4403 * cfs_idle_lag (delta between rq's update and cfs_rq's update)
4404 * = rq_clock_pelt()@rq_idle - rq_clock_pelt()@cfs_rq_idle
4406 * rq_idle_lag (delta between now and rq's update)
4407 * = sched_clock_cpu() - rq_clock()@rq_idle
4409 * We can then write:
4411 * now = rq_clock_pelt()@rq_idle - cfs->throttled_clock_pelt_time +
4412 * sched_clock_cpu() - rq_clock()@rq_idle
4414 * rq_clock_pelt()@rq_idle is rq->clock_pelt_idle
4415 * rq_clock()@rq_idle is rq->clock_idle
4416 * cfs->throttled_clock_pelt_time@cfs_rq_idle
4417 * is cfs_rq->throttled_pelt_idle
4420 #ifdef CONFIG_CFS_BANDWIDTH
4421 throttled = u64_u32_load(cfs_rq->throttled_pelt_idle);
4422 /* The clock has been stopped for throttling */
4423 if (throttled == U64_MAX)
4426 now = u64_u32_load(rq->clock_pelt_idle);
4428 * Paired with _update_idle_rq_clock_pelt(). It ensures at the worst case
4429 * is observed the old clock_pelt_idle value and the new clock_idle,
4430 * which lead to an underestimation. The opposite would lead to an
4434 lut = cfs_rq_last_update_time(cfs_rq);
4439 * cfs_rq->avg.last_update_time is more recent than our
4440 * estimation, let's use it.
4444 now += sched_clock_cpu(cpu_of(rq)) - u64_u32_load(rq->clock_idle);
4446 __update_load_avg_blocked_se(now, se);
4449 static void migrate_se_pelt_lag(struct sched_entity *se) {}
4453 * update_cfs_rq_load_avg - update the cfs_rq's load/util averages
4454 * @now: current time, as per cfs_rq_clock_pelt()
4455 * @cfs_rq: cfs_rq to update
4457 * The cfs_rq avg is the direct sum of all its entities (blocked and runnable)
4458 * avg. The immediate corollary is that all (fair) tasks must be attached.
4460 * cfs_rq->avg is used for task_h_load() and update_cfs_share() for example.
4462 * Return: true if the load decayed or we removed load.
4464 * Since both these conditions indicate a changed cfs_rq->avg.load we should
4465 * call update_tg_load_avg() when this function returns true.
4468 update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq)
4470 unsigned long removed_load = 0, removed_util = 0, removed_runnable = 0;
4471 struct sched_avg *sa = &cfs_rq->avg;
4474 if (cfs_rq->removed.nr) {
4476 u32 divider = get_pelt_divider(&cfs_rq->avg);
4478 raw_spin_lock(&cfs_rq->removed.lock);
4479 swap(cfs_rq->removed.util_avg, removed_util);
4480 swap(cfs_rq->removed.load_avg, removed_load);
4481 swap(cfs_rq->removed.runnable_avg, removed_runnable);
4482 cfs_rq->removed.nr = 0;
4483 raw_spin_unlock(&cfs_rq->removed.lock);
4486 sub_positive(&sa->load_avg, r);
4487 sub_positive(&sa->load_sum, r * divider);
4488 /* See sa->util_sum below */
4489 sa->load_sum = max_t(u32, sa->load_sum, sa->load_avg * PELT_MIN_DIVIDER);
4492 sub_positive(&sa->util_avg, r);
4493 sub_positive(&sa->util_sum, r * divider);
4495 * Because of rounding, se->util_sum might ends up being +1 more than
4496 * cfs->util_sum. Although this is not a problem by itself, detaching
4497 * a lot of tasks with the rounding problem between 2 updates of
4498 * util_avg (~1ms) can make cfs->util_sum becoming null whereas
4499 * cfs_util_avg is not.
4500 * Check that util_sum is still above its lower bound for the new
4501 * util_avg. Given that period_contrib might have moved since the last
4502 * sync, we are only sure that util_sum must be above or equal to
4503 * util_avg * minimum possible divider
4505 sa->util_sum = max_t(u32, sa->util_sum, sa->util_avg * PELT_MIN_DIVIDER);
4507 r = removed_runnable;
4508 sub_positive(&sa->runnable_avg, r);
4509 sub_positive(&sa->runnable_sum, r * divider);
4510 /* See sa->util_sum above */
4511 sa->runnable_sum = max_t(u32, sa->runnable_sum,
4512 sa->runnable_avg * PELT_MIN_DIVIDER);
4515 * removed_runnable is the unweighted version of removed_load so we
4516 * can use it to estimate removed_load_sum.
4518 add_tg_cfs_propagate(cfs_rq,
4519 -(long)(removed_runnable * divider) >> SCHED_CAPACITY_SHIFT);
4524 decayed |= __update_load_avg_cfs_rq(now, cfs_rq);
4525 u64_u32_store_copy(sa->last_update_time,
4526 cfs_rq->last_update_time_copy,
4527 sa->last_update_time);
4532 * attach_entity_load_avg - attach this entity to its cfs_rq load avg
4533 * @cfs_rq: cfs_rq to attach to
4534 * @se: sched_entity to attach
4536 * Must call update_cfs_rq_load_avg() before this, since we rely on
4537 * cfs_rq->avg.last_update_time being current.
4539 static void attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
4542 * cfs_rq->avg.period_contrib can be used for both cfs_rq and se.
4543 * See ___update_load_avg() for details.
4545 u32 divider = get_pelt_divider(&cfs_rq->avg);
4548 * When we attach the @se to the @cfs_rq, we must align the decay
4549 * window because without that, really weird and wonderful things can
4554 se->avg.last_update_time = cfs_rq->avg.last_update_time;
4555 se->avg.period_contrib = cfs_rq->avg.period_contrib;
4558 * Hell(o) Nasty stuff.. we need to recompute _sum based on the new
4559 * period_contrib. This isn't strictly correct, but since we're
4560 * entirely outside of the PELT hierarchy, nobody cares if we truncate
4563 se->avg.util_sum = se->avg.util_avg * divider;
4565 se->avg.runnable_sum = se->avg.runnable_avg * divider;
4567 se->avg.load_sum = se->avg.load_avg * divider;
4568 if (se_weight(se) < se->avg.load_sum)
4569 se->avg.load_sum = div_u64(se->avg.load_sum, se_weight(se));
4571 se->avg.load_sum = 1;
4573 enqueue_load_avg(cfs_rq, se);
4574 cfs_rq->avg.util_avg += se->avg.util_avg;
4575 cfs_rq->avg.util_sum += se->avg.util_sum;
4576 cfs_rq->avg.runnable_avg += se->avg.runnable_avg;
4577 cfs_rq->avg.runnable_sum += se->avg.runnable_sum;
4579 add_tg_cfs_propagate(cfs_rq, se->avg.load_sum);
4581 cfs_rq_util_change(cfs_rq, 0);
4583 trace_pelt_cfs_tp(cfs_rq);
4587 * detach_entity_load_avg - detach this entity from its cfs_rq load avg
4588 * @cfs_rq: cfs_rq to detach from
4589 * @se: sched_entity to detach
4591 * Must call update_cfs_rq_load_avg() before this, since we rely on
4592 * cfs_rq->avg.last_update_time being current.
4594 static void detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
4596 dequeue_load_avg(cfs_rq, se);
4597 sub_positive(&cfs_rq->avg.util_avg, se->avg.util_avg);
4598 sub_positive(&cfs_rq->avg.util_sum, se->avg.util_sum);
4599 /* See update_cfs_rq_load_avg() */
4600 cfs_rq->avg.util_sum = max_t(u32, cfs_rq->avg.util_sum,
4601 cfs_rq->avg.util_avg * PELT_MIN_DIVIDER);
4603 sub_positive(&cfs_rq->avg.runnable_avg, se->avg.runnable_avg);
4604 sub_positive(&cfs_rq->avg.runnable_sum, se->avg.runnable_sum);
4605 /* See update_cfs_rq_load_avg() */
4606 cfs_rq->avg.runnable_sum = max_t(u32, cfs_rq->avg.runnable_sum,
4607 cfs_rq->avg.runnable_avg * PELT_MIN_DIVIDER);
4609 add_tg_cfs_propagate(cfs_rq, -se->avg.load_sum);
4611 cfs_rq_util_change(cfs_rq, 0);
4613 trace_pelt_cfs_tp(cfs_rq);
4617 * Optional action to be done while updating the load average
4619 #define UPDATE_TG 0x1
4620 #define SKIP_AGE_LOAD 0x2
4621 #define DO_ATTACH 0x4
4622 #define DO_DETACH 0x8
4624 /* Update task and its cfs_rq load average */
4625 static inline void update_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
4627 u64 now = cfs_rq_clock_pelt(cfs_rq);
4631 * Track task load average for carrying it to new CPU after migrated, and
4632 * track group sched_entity load average for task_h_load calc in migration
4634 if (se->avg.last_update_time && !(flags & SKIP_AGE_LOAD))
4635 __update_load_avg_se(now, cfs_rq, se);
4637 decayed = update_cfs_rq_load_avg(now, cfs_rq);
4638 decayed |= propagate_entity_load_avg(se);
4640 if (!se->avg.last_update_time && (flags & DO_ATTACH)) {
4643 * DO_ATTACH means we're here from enqueue_entity().
4644 * !last_update_time means we've passed through
4645 * migrate_task_rq_fair() indicating we migrated.
4647 * IOW we're enqueueing a task on a new CPU.
4649 attach_entity_load_avg(cfs_rq, se);
4650 update_tg_load_avg(cfs_rq);
4652 } else if (flags & DO_DETACH) {
4654 * DO_DETACH means we're here from dequeue_entity()
4655 * and we are migrating task out of the CPU.
4657 detach_entity_load_avg(cfs_rq, se);
4658 update_tg_load_avg(cfs_rq);
4659 } else if (decayed) {
4660 cfs_rq_util_change(cfs_rq, 0);
4662 if (flags & UPDATE_TG)
4663 update_tg_load_avg(cfs_rq);
4668 * Synchronize entity load avg of dequeued entity without locking
4671 static void sync_entity_load_avg(struct sched_entity *se)
4673 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4674 u64 last_update_time;
4676 last_update_time = cfs_rq_last_update_time(cfs_rq);
4677 __update_load_avg_blocked_se(last_update_time, se);
4681 * Task first catches up with cfs_rq, and then subtract
4682 * itself from the cfs_rq (task must be off the queue now).
4684 static void remove_entity_load_avg(struct sched_entity *se)
4686 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4687 unsigned long flags;
4690 * tasks cannot exit without having gone through wake_up_new_task() ->
4691 * enqueue_task_fair() which will have added things to the cfs_rq,
4692 * so we can remove unconditionally.
4695 sync_entity_load_avg(se);
4697 raw_spin_lock_irqsave(&cfs_rq->removed.lock, flags);
4698 ++cfs_rq->removed.nr;
4699 cfs_rq->removed.util_avg += se->avg.util_avg;
4700 cfs_rq->removed.load_avg += se->avg.load_avg;
4701 cfs_rq->removed.runnable_avg += se->avg.runnable_avg;
4702 raw_spin_unlock_irqrestore(&cfs_rq->removed.lock, flags);
4705 static inline unsigned long cfs_rq_runnable_avg(struct cfs_rq *cfs_rq)
4707 return cfs_rq->avg.runnable_avg;
4710 static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq)
4712 return cfs_rq->avg.load_avg;
4715 static int newidle_balance(struct rq *this_rq, struct rq_flags *rf);
4717 static inline unsigned long task_util(struct task_struct *p)
4719 return READ_ONCE(p->se.avg.util_avg);
4722 static inline unsigned long _task_util_est(struct task_struct *p)
4724 struct util_est ue = READ_ONCE(p->se.avg.util_est);
4726 return max(ue.ewma, (ue.enqueued & ~UTIL_AVG_UNCHANGED));
4729 static inline unsigned long task_util_est(struct task_struct *p)
4731 return max(task_util(p), _task_util_est(p));
4734 static inline void util_est_enqueue(struct cfs_rq *cfs_rq,
4735 struct task_struct *p)
4737 unsigned int enqueued;
4739 if (!sched_feat(UTIL_EST))
4742 /* Update root cfs_rq's estimated utilization */
4743 enqueued = cfs_rq->avg.util_est.enqueued;
4744 enqueued += _task_util_est(p);
4745 WRITE_ONCE(cfs_rq->avg.util_est.enqueued, enqueued);
4747 trace_sched_util_est_cfs_tp(cfs_rq);
4750 static inline void util_est_dequeue(struct cfs_rq *cfs_rq,
4751 struct task_struct *p)
4753 unsigned int enqueued;
4755 if (!sched_feat(UTIL_EST))
4758 /* Update root cfs_rq's estimated utilization */
4759 enqueued = cfs_rq->avg.util_est.enqueued;
4760 enqueued -= min_t(unsigned int, enqueued, _task_util_est(p));
4761 WRITE_ONCE(cfs_rq->avg.util_est.enqueued, enqueued);
4763 trace_sched_util_est_cfs_tp(cfs_rq);
4766 #define UTIL_EST_MARGIN (SCHED_CAPACITY_SCALE / 100)
4769 * Check if a (signed) value is within a specified (unsigned) margin,
4770 * based on the observation that:
4772 * abs(x) < y := (unsigned)(x + y - 1) < (2 * y - 1)
4774 * NOTE: this only works when value + margin < INT_MAX.
4776 static inline bool within_margin(int value, int margin)
4778 return ((unsigned int)(value + margin - 1) < (2 * margin - 1));
4781 static inline void util_est_update(struct cfs_rq *cfs_rq,
4782 struct task_struct *p,
4785 long last_ewma_diff, last_enqueued_diff;
4788 if (!sched_feat(UTIL_EST))
4792 * Skip update of task's estimated utilization when the task has not
4793 * yet completed an activation, e.g. being migrated.
4799 * If the PELT values haven't changed since enqueue time,
4800 * skip the util_est update.
4802 ue = p->se.avg.util_est;
4803 if (ue.enqueued & UTIL_AVG_UNCHANGED)
4806 last_enqueued_diff = ue.enqueued;
4809 * Reset EWMA on utilization increases, the moving average is used only
4810 * to smooth utilization decreases.
4812 ue.enqueued = task_util(p);
4813 if (sched_feat(UTIL_EST_FASTUP)) {
4814 if (ue.ewma < ue.enqueued) {
4815 ue.ewma = ue.enqueued;
4821 * Skip update of task's estimated utilization when its members are
4822 * already ~1% close to its last activation value.
4824 last_ewma_diff = ue.enqueued - ue.ewma;
4825 last_enqueued_diff -= ue.enqueued;
4826 if (within_margin(last_ewma_diff, UTIL_EST_MARGIN)) {
4827 if (!within_margin(last_enqueued_diff, UTIL_EST_MARGIN))
4834 * To avoid overestimation of actual task utilization, skip updates if
4835 * we cannot grant there is idle time in this CPU.
4837 if (task_util(p) > capacity_orig_of(cpu_of(rq_of(cfs_rq))))
4841 * Update Task's estimated utilization
4843 * When *p completes an activation we can consolidate another sample
4844 * of the task size. This is done by storing the current PELT value
4845 * as ue.enqueued and by using this value to update the Exponential
4846 * Weighted Moving Average (EWMA):
4848 * ewma(t) = w * task_util(p) + (1-w) * ewma(t-1)
4849 * = w * task_util(p) + ewma(t-1) - w * ewma(t-1)
4850 * = w * (task_util(p) - ewma(t-1)) + ewma(t-1)
4851 * = w * ( last_ewma_diff ) + ewma(t-1)
4852 * = w * (last_ewma_diff + ewma(t-1) / w)
4854 * Where 'w' is the weight of new samples, which is configured to be
4855 * 0.25, thus making w=1/4 ( >>= UTIL_EST_WEIGHT_SHIFT)
4857 ue.ewma <<= UTIL_EST_WEIGHT_SHIFT;
4858 ue.ewma += last_ewma_diff;
4859 ue.ewma >>= UTIL_EST_WEIGHT_SHIFT;
4861 ue.enqueued |= UTIL_AVG_UNCHANGED;
4862 WRITE_ONCE(p->se.avg.util_est, ue);
4864 trace_sched_util_est_se_tp(&p->se);
4867 static inline int util_fits_cpu(unsigned long util,
4868 unsigned long uclamp_min,
4869 unsigned long uclamp_max,
4872 unsigned long capacity_orig, capacity_orig_thermal;
4873 unsigned long capacity = capacity_of(cpu);
4874 bool fits, uclamp_max_fits;
4877 * Check if the real util fits without any uclamp boost/cap applied.
4879 fits = fits_capacity(util, capacity);
4881 if (!uclamp_is_used())
4885 * We must use capacity_orig_of() for comparing against uclamp_min and
4886 * uclamp_max. We only care about capacity pressure (by using
4887 * capacity_of()) for comparing against the real util.
4889 * If a task is boosted to 1024 for example, we don't want a tiny
4890 * pressure to skew the check whether it fits a CPU or not.
4892 * Similarly if a task is capped to capacity_orig_of(little_cpu), it
4893 * should fit a little cpu even if there's some pressure.
4895 * Only exception is for thermal pressure since it has a direct impact
4896 * on available OPP of the system.
4898 * We honour it for uclamp_min only as a drop in performance level
4899 * could result in not getting the requested minimum performance level.
4901 * For uclamp_max, we can tolerate a drop in performance level as the
4902 * goal is to cap the task. So it's okay if it's getting less.
4904 capacity_orig = capacity_orig_of(cpu);
4905 capacity_orig_thermal = capacity_orig - arch_scale_thermal_pressure(cpu);
4908 * We want to force a task to fit a cpu as implied by uclamp_max.
4909 * But we do have some corner cases to cater for..
4915 * |_ _ _ _ _ _ _ _ _ ___ _ _ _ | _ | _ _ _ _ _ uclamp_max
4918 * | | | | | | | (util somewhere in this region)
4921 * +----------------------------------------
4924 * In the above example if a task is capped to a specific performance
4925 * point, y, then when:
4927 * * util = 80% of x then it does not fit on cpu0 and should migrate
4929 * * util = 80% of y then it is forced to fit on cpu1 to honour
4930 * uclamp_max request.
4932 * which is what we're enforcing here. A task always fits if
4933 * uclamp_max <= capacity_orig. But when uclamp_max > capacity_orig,
4934 * the normal upmigration rules should withhold still.
4936 * Only exception is when we are on max capacity, then we need to be
4937 * careful not to block overutilized state. This is so because:
4939 * 1. There's no concept of capping at max_capacity! We can't go
4940 * beyond this performance level anyway.
4941 * 2. The system is being saturated when we're operating near
4942 * max capacity, it doesn't make sense to block overutilized.
4944 uclamp_max_fits = (capacity_orig == SCHED_CAPACITY_SCALE) && (uclamp_max == SCHED_CAPACITY_SCALE);
4945 uclamp_max_fits = !uclamp_max_fits && (uclamp_max <= capacity_orig);
4946 fits = fits || uclamp_max_fits;
4951 * | ___ (region a, capped, util >= uclamp_max)
4953 * |_ _ _ _ _ _ _ _ _ ___ _ _ _ | _ | _ _ _ _ _ uclamp_max
4955 * | ___ | | | | (region b, uclamp_min <= util <= uclamp_max)
4956 * |_ _ _|_ _|_ _ _ _| _ | _ _ _| _ | _ _ _ _ _ uclamp_min
4958 * | | | | | | | (region c, boosted, util < uclamp_min)
4959 * +----------------------------------------
4962 * a) If util > uclamp_max, then we're capped, we don't care about
4963 * actual fitness value here. We only care if uclamp_max fits
4964 * capacity without taking margin/pressure into account.
4965 * See comment above.
4967 * b) If uclamp_min <= util <= uclamp_max, then the normal
4968 * fits_capacity() rules apply. Except we need to ensure that we
4969 * enforce we remain within uclamp_max, see comment above.
4971 * c) If util < uclamp_min, then we are boosted. Same as (b) but we
4972 * need to take into account the boosted value fits the CPU without
4973 * taking margin/pressure into account.
4975 * Cases (a) and (b) are handled in the 'fits' variable already. We
4976 * just need to consider an extra check for case (c) after ensuring we
4977 * handle the case uclamp_min > uclamp_max.
4979 uclamp_min = min(uclamp_min, uclamp_max);
4980 if (fits && (util < uclamp_min) && (uclamp_min > capacity_orig_thermal))
4986 static inline int task_fits_cpu(struct task_struct *p, int cpu)
4988 unsigned long uclamp_min = uclamp_eff_value(p, UCLAMP_MIN);
4989 unsigned long uclamp_max = uclamp_eff_value(p, UCLAMP_MAX);
4990 unsigned long util = task_util_est(p);
4992 * Return true only if the cpu fully fits the task requirements, which
4993 * include the utilization but also the performance hints.
4995 return (util_fits_cpu(util, uclamp_min, uclamp_max, cpu) > 0);
4998 static inline void update_misfit_status(struct task_struct *p, struct rq *rq)
5000 if (!sched_asym_cpucap_active())
5003 if (!p || p->nr_cpus_allowed == 1) {
5004 rq->misfit_task_load = 0;
5008 if (task_fits_cpu(p, cpu_of(rq))) {
5009 rq->misfit_task_load = 0;
5014 * Make sure that misfit_task_load will not be null even if
5015 * task_h_load() returns 0.
5017 rq->misfit_task_load = max_t(unsigned long, task_h_load(p), 1);
5020 #else /* CONFIG_SMP */
5022 static inline bool cfs_rq_is_decayed(struct cfs_rq *cfs_rq)
5024 return !cfs_rq->nr_running;
5027 #define UPDATE_TG 0x0
5028 #define SKIP_AGE_LOAD 0x0
5029 #define DO_ATTACH 0x0
5030 #define DO_DETACH 0x0
5032 static inline void update_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int not_used1)
5034 cfs_rq_util_change(cfs_rq, 0);
5037 static inline void remove_entity_load_avg(struct sched_entity *se) {}
5040 attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
5042 detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
5044 static inline int newidle_balance(struct rq *rq, struct rq_flags *rf)
5050 util_est_enqueue(struct cfs_rq *cfs_rq, struct task_struct *p) {}
5053 util_est_dequeue(struct cfs_rq *cfs_rq, struct task_struct *p) {}
5056 util_est_update(struct cfs_rq *cfs_rq, struct task_struct *p,
5058 static inline void update_misfit_status(struct task_struct *p, struct rq *rq) {}
5060 #endif /* CONFIG_SMP */
5063 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
5065 u64 vslice, vruntime = avg_vruntime(cfs_rq);
5068 se->slice = sysctl_sched_base_slice;
5069 vslice = calc_delta_fair(se->slice, se);
5072 * Due to how V is constructed as the weighted average of entities,
5073 * adding tasks with positive lag, or removing tasks with negative lag
5074 * will move 'time' backwards, this can screw around with the lag of
5077 * EEVDF: placement strategy #1 / #2
5079 if (sched_feat(PLACE_LAG) && cfs_rq->nr_running) {
5080 struct sched_entity *curr = cfs_rq->curr;
5086 * If we want to place a task and preserve lag, we have to
5087 * consider the effect of the new entity on the weighted
5088 * average and compensate for this, otherwise lag can quickly
5091 * Lag is defined as:
5093 * lag_i = S - s_i = w_i * (V - v_i)
5095 * To avoid the 'w_i' term all over the place, we only track
5098 * vl_i = V - v_i <=> v_i = V - vl_i
5100 * And we take V to be the weighted average of all v:
5102 * V = (\Sum w_j*v_j) / W
5104 * Where W is: \Sum w_j
5106 * Then, the weighted average after adding an entity with lag
5109 * V' = (\Sum w_j*v_j + w_i*v_i) / (W + w_i)
5110 * = (W*V + w_i*(V - vl_i)) / (W + w_i)
5111 * = (W*V + w_i*V - w_i*vl_i) / (W + w_i)
5112 * = (V*(W + w_i) - w_i*l) / (W + w_i)
5113 * = V - w_i*vl_i / (W + w_i)
5115 * And the actual lag after adding an entity with vl_i is:
5118 * = V - w_i*vl_i / (W + w_i) - (V - vl_i)
5119 * = vl_i - w_i*vl_i / (W + w_i)
5121 * Which is strictly less than vl_i. So in order to preserve lag
5122 * we should inflate the lag before placement such that the
5123 * effective lag after placement comes out right.
5125 * As such, invert the above relation for vl'_i to get the vl_i
5126 * we need to use such that the lag after placement is the lag
5127 * we computed before dequeue.
5129 * vl'_i = vl_i - w_i*vl_i / (W + w_i)
5130 * = ((W + w_i)*vl_i - w_i*vl_i) / (W + w_i)
5132 * (W + w_i)*vl'_i = (W + w_i)*vl_i - w_i*vl_i
5135 * vl_i = (W + w_i)*vl'_i / W
5137 load = cfs_rq->avg_load;
5138 if (curr && curr->on_rq)
5139 load += scale_load_down(curr->load.weight);
5141 lag *= load + scale_load_down(se->load.weight);
5142 if (WARN_ON_ONCE(!load))
5144 lag = div_s64(lag, load);
5147 se->vruntime = vruntime - lag;
5150 * When joining the competition; the exisiting tasks will be,
5151 * on average, halfway through their slice, as such start tasks
5152 * off with half a slice to ease into the competition.
5154 if (sched_feat(PLACE_DEADLINE_INITIAL) && (flags & ENQUEUE_INITIAL))
5158 * EEVDF: vd_i = ve_i + r_i/w_i
5160 se->deadline = se->vruntime + vslice;
5163 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
5164 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq);
5166 static inline bool cfs_bandwidth_used(void);
5169 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
5171 bool curr = cfs_rq->curr == se;
5174 * If we're the current task, we must renormalise before calling
5178 place_entity(cfs_rq, se, flags);
5180 update_curr(cfs_rq);
5183 * When enqueuing a sched_entity, we must:
5184 * - Update loads to have both entity and cfs_rq synced with now.
5185 * - For group_entity, update its runnable_weight to reflect the new
5186 * h_nr_running of its group cfs_rq.
5187 * - For group_entity, update its weight to reflect the new share of
5189 * - Add its new weight to cfs_rq->load.weight
5191 update_load_avg(cfs_rq, se, UPDATE_TG | DO_ATTACH);
5192 se_update_runnable(se);
5194 * XXX update_load_avg() above will have attached us to the pelt sum;
5195 * but update_cfs_group() here will re-adjust the weight and have to
5196 * undo/redo all that. Seems wasteful.
5198 update_cfs_group(se);
5201 * XXX now that the entity has been re-weighted, and it's lag adjusted,
5202 * we can place the entity.
5205 place_entity(cfs_rq, se, flags);
5207 account_entity_enqueue(cfs_rq, se);
5209 /* Entity has migrated, no longer consider this task hot */
5210 if (flags & ENQUEUE_MIGRATED)
5213 check_schedstat_required();
5214 update_stats_enqueue_fair(cfs_rq, se, flags);
5216 __enqueue_entity(cfs_rq, se);
5219 if (cfs_rq->nr_running == 1) {
5220 check_enqueue_throttle(cfs_rq);
5221 if (!throttled_hierarchy(cfs_rq)) {
5222 list_add_leaf_cfs_rq(cfs_rq);
5224 #ifdef CONFIG_CFS_BANDWIDTH
5225 struct rq *rq = rq_of(cfs_rq);
5227 if (cfs_rq_throttled(cfs_rq) && !cfs_rq->throttled_clock)
5228 cfs_rq->throttled_clock = rq_clock(rq);
5229 if (!cfs_rq->throttled_clock_self)
5230 cfs_rq->throttled_clock_self = rq_clock(rq);
5236 static void __clear_buddies_next(struct sched_entity *se)
5238 for_each_sched_entity(se) {
5239 struct cfs_rq *cfs_rq = cfs_rq_of(se);
5240 if (cfs_rq->next != se)
5243 cfs_rq->next = NULL;
5247 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
5249 if (cfs_rq->next == se)
5250 __clear_buddies_next(se);
5253 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
5256 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
5258 int action = UPDATE_TG;
5260 if (entity_is_task(se) && task_on_rq_migrating(task_of(se)))
5261 action |= DO_DETACH;
5264 * Update run-time statistics of the 'current'.
5266 update_curr(cfs_rq);
5269 * When dequeuing a sched_entity, we must:
5270 * - Update loads to have both entity and cfs_rq synced with now.
5271 * - For group_entity, update its runnable_weight to reflect the new
5272 * h_nr_running of its group cfs_rq.
5273 * - Subtract its previous weight from cfs_rq->load.weight.
5274 * - For group entity, update its weight to reflect the new share
5275 * of its group cfs_rq.
5277 update_load_avg(cfs_rq, se, action);
5278 se_update_runnable(se);
5280 update_stats_dequeue_fair(cfs_rq, se, flags);
5282 clear_buddies(cfs_rq, se);
5284 update_entity_lag(cfs_rq, se);
5285 if (se != cfs_rq->curr)
5286 __dequeue_entity(cfs_rq, se);
5288 account_entity_dequeue(cfs_rq, se);
5290 /* return excess runtime on last dequeue */
5291 return_cfs_rq_runtime(cfs_rq);
5293 update_cfs_group(se);
5296 * Now advance min_vruntime if @se was the entity holding it back,
5297 * except when: DEQUEUE_SAVE && !DEQUEUE_MOVE, in this case we'll be
5298 * put back on, and if we advance min_vruntime, we'll be placed back
5299 * further than we started -- ie. we'll be penalized.
5301 if ((flags & (DEQUEUE_SAVE | DEQUEUE_MOVE)) != DEQUEUE_SAVE)
5302 update_min_vruntime(cfs_rq);
5304 if (cfs_rq->nr_running == 0)
5305 update_idle_cfs_rq_clock_pelt(cfs_rq);
5309 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
5311 clear_buddies(cfs_rq, se);
5313 /* 'current' is not kept within the tree. */
5316 * Any task has to be enqueued before it get to execute on
5317 * a CPU. So account for the time it spent waiting on the
5320 update_stats_wait_end_fair(cfs_rq, se);
5321 __dequeue_entity(cfs_rq, se);
5322 update_load_avg(cfs_rq, se, UPDATE_TG);
5324 * HACK, stash a copy of deadline at the point of pick in vlag,
5325 * which isn't used until dequeue.
5327 se->vlag = se->deadline;
5330 update_stats_curr_start(cfs_rq, se);
5334 * Track our maximum slice length, if the CPU's load is at
5335 * least twice that of our own weight (i.e. dont track it
5336 * when there are only lesser-weight tasks around):
5338 if (schedstat_enabled() &&
5339 rq_of(cfs_rq)->cfs.load.weight >= 2*se->load.weight) {
5340 struct sched_statistics *stats;
5342 stats = __schedstats_from_se(se);
5343 __schedstat_set(stats->slice_max,
5344 max((u64)stats->slice_max,
5345 se->sum_exec_runtime - se->prev_sum_exec_runtime));
5348 se->prev_sum_exec_runtime = se->sum_exec_runtime;
5352 * Pick the next process, keeping these things in mind, in this order:
5353 * 1) keep things fair between processes/task groups
5354 * 2) pick the "next" process, since someone really wants that to run
5355 * 3) pick the "last" process, for cache locality
5356 * 4) do not run the "skip" process, if something else is available
5358 static struct sched_entity *
5359 pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr)
5362 * Enabling NEXT_BUDDY will affect latency but not fairness.
5364 if (sched_feat(NEXT_BUDDY) &&
5365 cfs_rq->next && entity_eligible(cfs_rq, cfs_rq->next))
5366 return cfs_rq->next;
5368 return pick_eevdf(cfs_rq);
5371 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
5373 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
5376 * If still on the runqueue then deactivate_task()
5377 * was not called and update_curr() has to be done:
5380 update_curr(cfs_rq);
5382 /* throttle cfs_rqs exceeding runtime */
5383 check_cfs_rq_runtime(cfs_rq);
5386 update_stats_wait_start_fair(cfs_rq, prev);
5387 /* Put 'current' back into the tree. */
5388 __enqueue_entity(cfs_rq, prev);
5389 /* in !on_rq case, update occurred at dequeue */
5390 update_load_avg(cfs_rq, prev, 0);
5392 cfs_rq->curr = NULL;
5396 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
5399 * Update run-time statistics of the 'current'.
5401 update_curr(cfs_rq);
5404 * Ensure that runnable average is periodically updated.
5406 update_load_avg(cfs_rq, curr, UPDATE_TG);
5407 update_cfs_group(curr);
5409 #ifdef CONFIG_SCHED_HRTICK
5411 * queued ticks are scheduled to match the slice, so don't bother
5412 * validating it and just reschedule.
5415 resched_curr(rq_of(cfs_rq));
5419 * don't let the period tick interfere with the hrtick preemption
5421 if (!sched_feat(DOUBLE_TICK) &&
5422 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
5428 /**************************************************
5429 * CFS bandwidth control machinery
5432 #ifdef CONFIG_CFS_BANDWIDTH
5434 #ifdef CONFIG_JUMP_LABEL
5435 static struct static_key __cfs_bandwidth_used;
5437 static inline bool cfs_bandwidth_used(void)
5439 return static_key_false(&__cfs_bandwidth_used);
5442 void cfs_bandwidth_usage_inc(void)
5444 static_key_slow_inc_cpuslocked(&__cfs_bandwidth_used);
5447 void cfs_bandwidth_usage_dec(void)
5449 static_key_slow_dec_cpuslocked(&__cfs_bandwidth_used);
5451 #else /* CONFIG_JUMP_LABEL */
5452 static bool cfs_bandwidth_used(void)
5457 void cfs_bandwidth_usage_inc(void) {}
5458 void cfs_bandwidth_usage_dec(void) {}
5459 #endif /* CONFIG_JUMP_LABEL */
5462 * default period for cfs group bandwidth.
5463 * default: 0.1s, units: nanoseconds
5465 static inline u64 default_cfs_period(void)
5467 return 100000000ULL;
5470 static inline u64 sched_cfs_bandwidth_slice(void)
5472 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
5476 * Replenish runtime according to assigned quota. We use sched_clock_cpu
5477 * directly instead of rq->clock to avoid adding additional synchronization
5480 * requires cfs_b->lock
5482 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
5486 if (unlikely(cfs_b->quota == RUNTIME_INF))
5489 cfs_b->runtime += cfs_b->quota;
5490 runtime = cfs_b->runtime_snap - cfs_b->runtime;
5492 cfs_b->burst_time += runtime;
5496 cfs_b->runtime = min(cfs_b->runtime, cfs_b->quota + cfs_b->burst);
5497 cfs_b->runtime_snap = cfs_b->runtime;
5500 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
5502 return &tg->cfs_bandwidth;
5505 /* returns 0 on failure to allocate runtime */
5506 static int __assign_cfs_rq_runtime(struct cfs_bandwidth *cfs_b,
5507 struct cfs_rq *cfs_rq, u64 target_runtime)
5509 u64 min_amount, amount = 0;
5511 lockdep_assert_held(&cfs_b->lock);
5513 /* note: this is a positive sum as runtime_remaining <= 0 */
5514 min_amount = target_runtime - cfs_rq->runtime_remaining;
5516 if (cfs_b->quota == RUNTIME_INF)
5517 amount = min_amount;
5519 start_cfs_bandwidth(cfs_b);
5521 if (cfs_b->runtime > 0) {
5522 amount = min(cfs_b->runtime, min_amount);
5523 cfs_b->runtime -= amount;
5528 cfs_rq->runtime_remaining += amount;
5530 return cfs_rq->runtime_remaining > 0;
5533 /* returns 0 on failure to allocate runtime */
5534 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
5536 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
5539 raw_spin_lock(&cfs_b->lock);
5540 ret = __assign_cfs_rq_runtime(cfs_b, cfs_rq, sched_cfs_bandwidth_slice());
5541 raw_spin_unlock(&cfs_b->lock);
5546 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
5548 /* dock delta_exec before expiring quota (as it could span periods) */
5549 cfs_rq->runtime_remaining -= delta_exec;
5551 if (likely(cfs_rq->runtime_remaining > 0))
5554 if (cfs_rq->throttled)
5557 * if we're unable to extend our runtime we resched so that the active
5558 * hierarchy can be throttled
5560 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
5561 resched_curr(rq_of(cfs_rq));
5564 static __always_inline
5565 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
5567 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
5570 __account_cfs_rq_runtime(cfs_rq, delta_exec);
5573 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
5575 return cfs_bandwidth_used() && cfs_rq->throttled;
5578 /* check whether cfs_rq, or any parent, is throttled */
5579 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
5581 return cfs_bandwidth_used() && cfs_rq->throttle_count;
5585 * Ensure that neither of the group entities corresponding to src_cpu or
5586 * dest_cpu are members of a throttled hierarchy when performing group
5587 * load-balance operations.
5589 static inline int throttled_lb_pair(struct task_group *tg,
5590 int src_cpu, int dest_cpu)
5592 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
5594 src_cfs_rq = tg->cfs_rq[src_cpu];
5595 dest_cfs_rq = tg->cfs_rq[dest_cpu];
5597 return throttled_hierarchy(src_cfs_rq) ||
5598 throttled_hierarchy(dest_cfs_rq);
5601 static int tg_unthrottle_up(struct task_group *tg, void *data)
5603 struct rq *rq = data;
5604 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
5606 cfs_rq->throttle_count--;
5607 if (!cfs_rq->throttle_count) {
5608 cfs_rq->throttled_clock_pelt_time += rq_clock_pelt(rq) -
5609 cfs_rq->throttled_clock_pelt;
5611 /* Add cfs_rq with load or one or more already running entities to the list */
5612 if (!cfs_rq_is_decayed(cfs_rq))
5613 list_add_leaf_cfs_rq(cfs_rq);
5615 if (cfs_rq->throttled_clock_self) {
5616 u64 delta = rq_clock(rq) - cfs_rq->throttled_clock_self;
5618 cfs_rq->throttled_clock_self = 0;
5620 if (SCHED_WARN_ON((s64)delta < 0))
5623 cfs_rq->throttled_clock_self_time += delta;
5630 static int tg_throttle_down(struct task_group *tg, void *data)
5632 struct rq *rq = data;
5633 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
5635 /* group is entering throttled state, stop time */
5636 if (!cfs_rq->throttle_count) {
5637 cfs_rq->throttled_clock_pelt = rq_clock_pelt(rq);
5638 list_del_leaf_cfs_rq(cfs_rq);
5640 SCHED_WARN_ON(cfs_rq->throttled_clock_self);
5641 if (cfs_rq->nr_running)
5642 cfs_rq->throttled_clock_self = rq_clock(rq);
5644 cfs_rq->throttle_count++;
5649 static bool throttle_cfs_rq(struct cfs_rq *cfs_rq)
5651 struct rq *rq = rq_of(cfs_rq);
5652 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
5653 struct sched_entity *se;
5654 long task_delta, idle_task_delta, dequeue = 1;
5656 raw_spin_lock(&cfs_b->lock);
5657 /* This will start the period timer if necessary */
5658 if (__assign_cfs_rq_runtime(cfs_b, cfs_rq, 1)) {
5660 * We have raced with bandwidth becoming available, and if we
5661 * actually throttled the timer might not unthrottle us for an
5662 * entire period. We additionally needed to make sure that any
5663 * subsequent check_cfs_rq_runtime calls agree not to throttle
5664 * us, as we may commit to do cfs put_prev+pick_next, so we ask
5665 * for 1ns of runtime rather than just check cfs_b.
5669 list_add_tail_rcu(&cfs_rq->throttled_list,
5670 &cfs_b->throttled_cfs_rq);
5672 raw_spin_unlock(&cfs_b->lock);
5675 return false; /* Throttle no longer required. */
5677 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
5679 /* freeze hierarchy runnable averages while throttled */
5681 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
5684 task_delta = cfs_rq->h_nr_running;
5685 idle_task_delta = cfs_rq->idle_h_nr_running;
5686 for_each_sched_entity(se) {
5687 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
5688 /* throttled entity or throttle-on-deactivate */
5692 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
5694 if (cfs_rq_is_idle(group_cfs_rq(se)))
5695 idle_task_delta = cfs_rq->h_nr_running;
5697 qcfs_rq->h_nr_running -= task_delta;
5698 qcfs_rq->idle_h_nr_running -= idle_task_delta;
5700 if (qcfs_rq->load.weight) {
5701 /* Avoid re-evaluating load for this entity: */
5702 se = parent_entity(se);
5707 for_each_sched_entity(se) {
5708 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
5709 /* throttled entity or throttle-on-deactivate */
5713 update_load_avg(qcfs_rq, se, 0);
5714 se_update_runnable(se);
5716 if (cfs_rq_is_idle(group_cfs_rq(se)))
5717 idle_task_delta = cfs_rq->h_nr_running;
5719 qcfs_rq->h_nr_running -= task_delta;
5720 qcfs_rq->idle_h_nr_running -= idle_task_delta;
5723 /* At this point se is NULL and we are at root level*/
5724 sub_nr_running(rq, task_delta);
5728 * Note: distribution will already see us throttled via the
5729 * throttled-list. rq->lock protects completion.
5731 cfs_rq->throttled = 1;
5732 SCHED_WARN_ON(cfs_rq->throttled_clock);
5733 if (cfs_rq->nr_running)
5734 cfs_rq->throttled_clock = rq_clock(rq);
5738 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
5740 struct rq *rq = rq_of(cfs_rq);
5741 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
5742 struct sched_entity *se;
5743 long task_delta, idle_task_delta;
5745 se = cfs_rq->tg->se[cpu_of(rq)];
5747 cfs_rq->throttled = 0;
5749 update_rq_clock(rq);
5751 raw_spin_lock(&cfs_b->lock);
5752 if (cfs_rq->throttled_clock) {
5753 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
5754 cfs_rq->throttled_clock = 0;
5756 list_del_rcu(&cfs_rq->throttled_list);
5757 raw_spin_unlock(&cfs_b->lock);
5759 /* update hierarchical throttle state */
5760 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
5762 if (!cfs_rq->load.weight) {
5763 if (!cfs_rq->on_list)
5766 * Nothing to run but something to decay (on_list)?
5767 * Complete the branch.
5769 for_each_sched_entity(se) {
5770 if (list_add_leaf_cfs_rq(cfs_rq_of(se)))
5773 goto unthrottle_throttle;
5776 task_delta = cfs_rq->h_nr_running;
5777 idle_task_delta = cfs_rq->idle_h_nr_running;
5778 for_each_sched_entity(se) {
5779 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
5783 enqueue_entity(qcfs_rq, se, ENQUEUE_WAKEUP);
5785 if (cfs_rq_is_idle(group_cfs_rq(se)))
5786 idle_task_delta = cfs_rq->h_nr_running;
5788 qcfs_rq->h_nr_running += task_delta;
5789 qcfs_rq->idle_h_nr_running += idle_task_delta;
5791 /* end evaluation on encountering a throttled cfs_rq */
5792 if (cfs_rq_throttled(qcfs_rq))
5793 goto unthrottle_throttle;
5796 for_each_sched_entity(se) {
5797 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
5799 update_load_avg(qcfs_rq, se, UPDATE_TG);
5800 se_update_runnable(se);
5802 if (cfs_rq_is_idle(group_cfs_rq(se)))
5803 idle_task_delta = cfs_rq->h_nr_running;
5805 qcfs_rq->h_nr_running += task_delta;
5806 qcfs_rq->idle_h_nr_running += idle_task_delta;
5808 /* end evaluation on encountering a throttled cfs_rq */
5809 if (cfs_rq_throttled(qcfs_rq))
5810 goto unthrottle_throttle;
5813 /* At this point se is NULL and we are at root level*/
5814 add_nr_running(rq, task_delta);
5816 unthrottle_throttle:
5817 assert_list_leaf_cfs_rq(rq);
5819 /* Determine whether we need to wake up potentially idle CPU: */
5820 if (rq->curr == rq->idle && rq->cfs.nr_running)
5825 static void __cfsb_csd_unthrottle(void *arg)
5827 struct cfs_rq *cursor, *tmp;
5828 struct rq *rq = arg;
5834 * Iterating over the list can trigger several call to
5835 * update_rq_clock() in unthrottle_cfs_rq().
5836 * Do it once and skip the potential next ones.
5838 update_rq_clock(rq);
5839 rq_clock_start_loop_update(rq);
5842 * Since we hold rq lock we're safe from concurrent manipulation of
5843 * the CSD list. However, this RCU critical section annotates the
5844 * fact that we pair with sched_free_group_rcu(), so that we cannot
5845 * race with group being freed in the window between removing it
5846 * from the list and advancing to the next entry in the list.
5850 list_for_each_entry_safe(cursor, tmp, &rq->cfsb_csd_list,
5851 throttled_csd_list) {
5852 list_del_init(&cursor->throttled_csd_list);
5854 if (cfs_rq_throttled(cursor))
5855 unthrottle_cfs_rq(cursor);
5860 rq_clock_stop_loop_update(rq);
5864 static inline void __unthrottle_cfs_rq_async(struct cfs_rq *cfs_rq)
5866 struct rq *rq = rq_of(cfs_rq);
5869 if (rq == this_rq()) {
5870 unthrottle_cfs_rq(cfs_rq);
5874 /* Already enqueued */
5875 if (SCHED_WARN_ON(!list_empty(&cfs_rq->throttled_csd_list)))
5878 first = list_empty(&rq->cfsb_csd_list);
5879 list_add_tail(&cfs_rq->throttled_csd_list, &rq->cfsb_csd_list);
5881 smp_call_function_single_async(cpu_of(rq), &rq->cfsb_csd);
5884 static inline void __unthrottle_cfs_rq_async(struct cfs_rq *cfs_rq)
5886 unthrottle_cfs_rq(cfs_rq);
5890 static void unthrottle_cfs_rq_async(struct cfs_rq *cfs_rq)
5892 lockdep_assert_rq_held(rq_of(cfs_rq));
5894 if (SCHED_WARN_ON(!cfs_rq_throttled(cfs_rq) ||
5895 cfs_rq->runtime_remaining <= 0))
5898 __unthrottle_cfs_rq_async(cfs_rq);
5901 static bool distribute_cfs_runtime(struct cfs_bandwidth *cfs_b)
5903 struct cfs_rq *local_unthrottle = NULL;
5904 int this_cpu = smp_processor_id();
5905 u64 runtime, remaining = 1;
5906 bool throttled = false;
5907 struct cfs_rq *cfs_rq;
5912 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
5921 rq_lock_irqsave(rq, &rf);
5922 if (!cfs_rq_throttled(cfs_rq))
5926 /* Already queued for async unthrottle */
5927 if (!list_empty(&cfs_rq->throttled_csd_list))
5931 /* By the above checks, this should never be true */
5932 SCHED_WARN_ON(cfs_rq->runtime_remaining > 0);
5934 raw_spin_lock(&cfs_b->lock);
5935 runtime = -cfs_rq->runtime_remaining + 1;
5936 if (runtime > cfs_b->runtime)
5937 runtime = cfs_b->runtime;
5938 cfs_b->runtime -= runtime;
5939 remaining = cfs_b->runtime;
5940 raw_spin_unlock(&cfs_b->lock);
5942 cfs_rq->runtime_remaining += runtime;
5944 /* we check whether we're throttled above */
5945 if (cfs_rq->runtime_remaining > 0) {
5946 if (cpu_of(rq) != this_cpu ||
5947 SCHED_WARN_ON(local_unthrottle))
5948 unthrottle_cfs_rq_async(cfs_rq);
5950 local_unthrottle = cfs_rq;
5956 rq_unlock_irqrestore(rq, &rf);
5960 if (local_unthrottle) {
5961 rq = cpu_rq(this_cpu);
5962 rq_lock_irqsave(rq, &rf);
5963 if (cfs_rq_throttled(local_unthrottle))
5964 unthrottle_cfs_rq(local_unthrottle);
5965 rq_unlock_irqrestore(rq, &rf);
5972 * Responsible for refilling a task_group's bandwidth and unthrottling its
5973 * cfs_rqs as appropriate. If there has been no activity within the last
5974 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
5975 * used to track this state.
5977 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun, unsigned long flags)
5981 /* no need to continue the timer with no bandwidth constraint */
5982 if (cfs_b->quota == RUNTIME_INF)
5983 goto out_deactivate;
5985 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
5986 cfs_b->nr_periods += overrun;
5988 /* Refill extra burst quota even if cfs_b->idle */
5989 __refill_cfs_bandwidth_runtime(cfs_b);
5992 * idle depends on !throttled (for the case of a large deficit), and if
5993 * we're going inactive then everything else can be deferred
5995 if (cfs_b->idle && !throttled)
5996 goto out_deactivate;
5999 /* mark as potentially idle for the upcoming period */
6004 /* account preceding periods in which throttling occurred */
6005 cfs_b->nr_throttled += overrun;
6008 * This check is repeated as we release cfs_b->lock while we unthrottle.
6010 while (throttled && cfs_b->runtime > 0) {
6011 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
6012 /* we can't nest cfs_b->lock while distributing bandwidth */
6013 throttled = distribute_cfs_runtime(cfs_b);
6014 raw_spin_lock_irqsave(&cfs_b->lock, flags);
6018 * While we are ensured activity in the period following an
6019 * unthrottle, this also covers the case in which the new bandwidth is
6020 * insufficient to cover the existing bandwidth deficit. (Forcing the
6021 * timer to remain active while there are any throttled entities.)
6031 /* a cfs_rq won't donate quota below this amount */
6032 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
6033 /* minimum remaining period time to redistribute slack quota */
6034 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
6035 /* how long we wait to gather additional slack before distributing */
6036 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
6039 * Are we near the end of the current quota period?
6041 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
6042 * hrtimer base being cleared by hrtimer_start. In the case of
6043 * migrate_hrtimers, base is never cleared, so we are fine.
6045 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
6047 struct hrtimer *refresh_timer = &cfs_b->period_timer;
6050 /* if the call-back is running a quota refresh is already occurring */
6051 if (hrtimer_callback_running(refresh_timer))
6054 /* is a quota refresh about to occur? */
6055 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
6056 if (remaining < (s64)min_expire)
6062 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
6064 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
6066 /* if there's a quota refresh soon don't bother with slack */
6067 if (runtime_refresh_within(cfs_b, min_left))
6070 /* don't push forwards an existing deferred unthrottle */
6071 if (cfs_b->slack_started)
6073 cfs_b->slack_started = true;
6075 hrtimer_start(&cfs_b->slack_timer,
6076 ns_to_ktime(cfs_bandwidth_slack_period),
6080 /* we know any runtime found here is valid as update_curr() precedes return */
6081 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
6083 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
6084 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
6086 if (slack_runtime <= 0)
6089 raw_spin_lock(&cfs_b->lock);
6090 if (cfs_b->quota != RUNTIME_INF) {
6091 cfs_b->runtime += slack_runtime;
6093 /* we are under rq->lock, defer unthrottling using a timer */
6094 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
6095 !list_empty(&cfs_b->throttled_cfs_rq))
6096 start_cfs_slack_bandwidth(cfs_b);
6098 raw_spin_unlock(&cfs_b->lock);
6100 /* even if it's not valid for return we don't want to try again */
6101 cfs_rq->runtime_remaining -= slack_runtime;
6104 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
6106 if (!cfs_bandwidth_used())
6109 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
6112 __return_cfs_rq_runtime(cfs_rq);
6116 * This is done with a timer (instead of inline with bandwidth return) since
6117 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
6119 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
6121 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
6122 unsigned long flags;
6124 /* confirm we're still not at a refresh boundary */
6125 raw_spin_lock_irqsave(&cfs_b->lock, flags);
6126 cfs_b->slack_started = false;
6128 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
6129 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
6133 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice)
6134 runtime = cfs_b->runtime;
6136 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
6141 distribute_cfs_runtime(cfs_b);
6145 * When a group wakes up we want to make sure that its quota is not already
6146 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
6147 * runtime as update_curr() throttling can not trigger until it's on-rq.
6149 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
6151 if (!cfs_bandwidth_used())
6154 /* an active group must be handled by the update_curr()->put() path */
6155 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
6158 /* ensure the group is not already throttled */
6159 if (cfs_rq_throttled(cfs_rq))
6162 /* update runtime allocation */
6163 account_cfs_rq_runtime(cfs_rq, 0);
6164 if (cfs_rq->runtime_remaining <= 0)
6165 throttle_cfs_rq(cfs_rq);
6168 static void sync_throttle(struct task_group *tg, int cpu)
6170 struct cfs_rq *pcfs_rq, *cfs_rq;
6172 if (!cfs_bandwidth_used())
6178 cfs_rq = tg->cfs_rq[cpu];
6179 pcfs_rq = tg->parent->cfs_rq[cpu];
6181 cfs_rq->throttle_count = pcfs_rq->throttle_count;
6182 cfs_rq->throttled_clock_pelt = rq_clock_pelt(cpu_rq(cpu));
6185 /* conditionally throttle active cfs_rq's from put_prev_entity() */
6186 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
6188 if (!cfs_bandwidth_used())
6191 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
6195 * it's possible for a throttled entity to be forced into a running
6196 * state (e.g. set_curr_task), in this case we're finished.
6198 if (cfs_rq_throttled(cfs_rq))
6201 return throttle_cfs_rq(cfs_rq);
6204 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
6206 struct cfs_bandwidth *cfs_b =
6207 container_of(timer, struct cfs_bandwidth, slack_timer);
6209 do_sched_cfs_slack_timer(cfs_b);
6211 return HRTIMER_NORESTART;
6214 extern const u64 max_cfs_quota_period;
6216 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
6218 struct cfs_bandwidth *cfs_b =
6219 container_of(timer, struct cfs_bandwidth, period_timer);
6220 unsigned long flags;
6225 raw_spin_lock_irqsave(&cfs_b->lock, flags);
6227 overrun = hrtimer_forward_now(timer, cfs_b->period);
6231 idle = do_sched_cfs_period_timer(cfs_b, overrun, flags);
6234 u64 new, old = ktime_to_ns(cfs_b->period);
6237 * Grow period by a factor of 2 to avoid losing precision.
6238 * Precision loss in the quota/period ratio can cause __cfs_schedulable
6242 if (new < max_cfs_quota_period) {
6243 cfs_b->period = ns_to_ktime(new);
6247 pr_warn_ratelimited(
6248 "cfs_period_timer[cpu%d]: period too short, scaling up (new cfs_period_us = %lld, cfs_quota_us = %lld)\n",
6250 div_u64(new, NSEC_PER_USEC),
6251 div_u64(cfs_b->quota, NSEC_PER_USEC));
6253 pr_warn_ratelimited(
6254 "cfs_period_timer[cpu%d]: period too short, but cannot scale up without losing precision (cfs_period_us = %lld, cfs_quota_us = %lld)\n",
6256 div_u64(old, NSEC_PER_USEC),
6257 div_u64(cfs_b->quota, NSEC_PER_USEC));
6260 /* reset count so we don't come right back in here */
6265 cfs_b->period_active = 0;
6266 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
6268 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
6271 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b, struct cfs_bandwidth *parent)
6273 raw_spin_lock_init(&cfs_b->lock);
6275 cfs_b->quota = RUNTIME_INF;
6276 cfs_b->period = ns_to_ktime(default_cfs_period());
6278 cfs_b->hierarchical_quota = parent ? parent->hierarchical_quota : RUNTIME_INF;
6280 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
6281 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED);
6282 cfs_b->period_timer.function = sched_cfs_period_timer;
6284 /* Add a random offset so that timers interleave */
6285 hrtimer_set_expires(&cfs_b->period_timer,
6286 get_random_u32_below(cfs_b->period));
6287 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
6288 cfs_b->slack_timer.function = sched_cfs_slack_timer;
6289 cfs_b->slack_started = false;
6292 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
6294 cfs_rq->runtime_enabled = 0;
6295 INIT_LIST_HEAD(&cfs_rq->throttled_list);
6297 INIT_LIST_HEAD(&cfs_rq->throttled_csd_list);
6301 void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
6303 lockdep_assert_held(&cfs_b->lock);
6305 if (cfs_b->period_active)
6308 cfs_b->period_active = 1;
6309 hrtimer_forward_now(&cfs_b->period_timer, cfs_b->period);
6310 hrtimer_start_expires(&cfs_b->period_timer, HRTIMER_MODE_ABS_PINNED);
6313 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
6315 int __maybe_unused i;
6317 /* init_cfs_bandwidth() was not called */
6318 if (!cfs_b->throttled_cfs_rq.next)
6321 hrtimer_cancel(&cfs_b->period_timer);
6322 hrtimer_cancel(&cfs_b->slack_timer);
6325 * It is possible that we still have some cfs_rq's pending on a CSD
6326 * list, though this race is very rare. In order for this to occur, we
6327 * must have raced with the last task leaving the group while there
6328 * exist throttled cfs_rq(s), and the period_timer must have queued the
6329 * CSD item but the remote cpu has not yet processed it. To handle this,
6330 * we can simply flush all pending CSD work inline here. We're
6331 * guaranteed at this point that no additional cfs_rq of this group can
6335 for_each_possible_cpu(i) {
6336 struct rq *rq = cpu_rq(i);
6337 unsigned long flags;
6339 if (list_empty(&rq->cfsb_csd_list))
6342 local_irq_save(flags);
6343 __cfsb_csd_unthrottle(rq);
6344 local_irq_restore(flags);
6350 * Both these CPU hotplug callbacks race against unregister_fair_sched_group()
6352 * The race is harmless, since modifying bandwidth settings of unhooked group
6353 * bits doesn't do much.
6356 /* cpu online callback */
6357 static void __maybe_unused update_runtime_enabled(struct rq *rq)
6359 struct task_group *tg;
6361 lockdep_assert_rq_held(rq);
6364 list_for_each_entry_rcu(tg, &task_groups, list) {
6365 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
6366 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
6368 raw_spin_lock(&cfs_b->lock);
6369 cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF;
6370 raw_spin_unlock(&cfs_b->lock);
6375 /* cpu offline callback */
6376 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
6378 struct task_group *tg;
6380 lockdep_assert_rq_held(rq);
6383 * The rq clock has already been updated in the
6384 * set_rq_offline(), so we should skip updating
6385 * the rq clock again in unthrottle_cfs_rq().
6387 rq_clock_start_loop_update(rq);
6390 list_for_each_entry_rcu(tg, &task_groups, list) {
6391 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
6393 if (!cfs_rq->runtime_enabled)
6397 * clock_task is not advancing so we just need to make sure
6398 * there's some valid quota amount
6400 cfs_rq->runtime_remaining = 1;
6402 * Offline rq is schedulable till CPU is completely disabled
6403 * in take_cpu_down(), so we prevent new cfs throttling here.
6405 cfs_rq->runtime_enabled = 0;
6407 if (cfs_rq_throttled(cfs_rq))
6408 unthrottle_cfs_rq(cfs_rq);
6412 rq_clock_stop_loop_update(rq);
6415 bool cfs_task_bw_constrained(struct task_struct *p)
6417 struct cfs_rq *cfs_rq = task_cfs_rq(p);
6419 if (!cfs_bandwidth_used())
6422 if (cfs_rq->runtime_enabled ||
6423 tg_cfs_bandwidth(cfs_rq->tg)->hierarchical_quota != RUNTIME_INF)
6429 #ifdef CONFIG_NO_HZ_FULL
6430 /* called from pick_next_task_fair() */
6431 static void sched_fair_update_stop_tick(struct rq *rq, struct task_struct *p)
6433 int cpu = cpu_of(rq);
6435 if (!sched_feat(HZ_BW) || !cfs_bandwidth_used())
6438 if (!tick_nohz_full_cpu(cpu))
6441 if (rq->nr_running != 1)
6445 * We know there is only one task runnable and we've just picked it. The
6446 * normal enqueue path will have cleared TICK_DEP_BIT_SCHED if we will
6447 * be otherwise able to stop the tick. Just need to check if we are using
6448 * bandwidth control.
6450 if (cfs_task_bw_constrained(p))
6451 tick_nohz_dep_set_cpu(cpu, TICK_DEP_BIT_SCHED);
6455 #else /* CONFIG_CFS_BANDWIDTH */
6457 static inline bool cfs_bandwidth_used(void)
6462 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {}
6463 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; }
6464 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
6465 static inline void sync_throttle(struct task_group *tg, int cpu) {}
6466 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
6468 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
6473 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
6478 static inline int throttled_lb_pair(struct task_group *tg,
6479 int src_cpu, int dest_cpu)
6484 #ifdef CONFIG_FAIR_GROUP_SCHED
6485 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b, struct cfs_bandwidth *parent) {}
6486 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
6489 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
6493 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
6494 static inline void update_runtime_enabled(struct rq *rq) {}
6495 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
6496 #ifdef CONFIG_CGROUP_SCHED
6497 bool cfs_task_bw_constrained(struct task_struct *p)
6502 #endif /* CONFIG_CFS_BANDWIDTH */
6504 #if !defined(CONFIG_CFS_BANDWIDTH) || !defined(CONFIG_NO_HZ_FULL)
6505 static inline void sched_fair_update_stop_tick(struct rq *rq, struct task_struct *p) {}
6508 /**************************************************
6509 * CFS operations on tasks:
6512 #ifdef CONFIG_SCHED_HRTICK
6513 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
6515 struct sched_entity *se = &p->se;
6517 SCHED_WARN_ON(task_rq(p) != rq);
6519 if (rq->cfs.h_nr_running > 1) {
6520 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
6521 u64 slice = se->slice;
6522 s64 delta = slice - ran;
6525 if (task_current(rq, p))
6529 hrtick_start(rq, delta);
6534 * called from enqueue/dequeue and updates the hrtick when the
6535 * current task is from our class and nr_running is low enough
6538 static void hrtick_update(struct rq *rq)
6540 struct task_struct *curr = rq->curr;
6542 if (!hrtick_enabled_fair(rq) || curr->sched_class != &fair_sched_class)
6545 hrtick_start_fair(rq, curr);
6547 #else /* !CONFIG_SCHED_HRTICK */
6549 hrtick_start_fair(struct rq *rq, struct task_struct *p)
6553 static inline void hrtick_update(struct rq *rq)
6559 static inline bool cpu_overutilized(int cpu)
6561 unsigned long rq_util_min = uclamp_rq_get(cpu_rq(cpu), UCLAMP_MIN);
6562 unsigned long rq_util_max = uclamp_rq_get(cpu_rq(cpu), UCLAMP_MAX);
6564 /* Return true only if the utilization doesn't fit CPU's capacity */
6565 return !util_fits_cpu(cpu_util_cfs(cpu), rq_util_min, rq_util_max, cpu);
6568 static inline void update_overutilized_status(struct rq *rq)
6570 if (!READ_ONCE(rq->rd->overutilized) && cpu_overutilized(rq->cpu)) {
6571 WRITE_ONCE(rq->rd->overutilized, SG_OVERUTILIZED);
6572 trace_sched_overutilized_tp(rq->rd, SG_OVERUTILIZED);
6576 static inline void update_overutilized_status(struct rq *rq) { }
6579 /* Runqueue only has SCHED_IDLE tasks enqueued */
6580 static int sched_idle_rq(struct rq *rq)
6582 return unlikely(rq->nr_running == rq->cfs.idle_h_nr_running &&
6587 static int sched_idle_cpu(int cpu)
6589 return sched_idle_rq(cpu_rq(cpu));
6594 * The enqueue_task method is called before nr_running is
6595 * increased. Here we update the fair scheduling stats and
6596 * then put the task into the rbtree:
6599 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
6601 struct cfs_rq *cfs_rq;
6602 struct sched_entity *se = &p->se;
6603 int idle_h_nr_running = task_has_idle_policy(p);
6604 int task_new = !(flags & ENQUEUE_WAKEUP);
6607 * The code below (indirectly) updates schedutil which looks at
6608 * the cfs_rq utilization to select a frequency.
6609 * Let's add the task's estimated utilization to the cfs_rq's
6610 * estimated utilization, before we update schedutil.
6612 util_est_enqueue(&rq->cfs, p);
6615 * If in_iowait is set, the code below may not trigger any cpufreq
6616 * utilization updates, so do it here explicitly with the IOWAIT flag
6620 cpufreq_update_util(rq, SCHED_CPUFREQ_IOWAIT);
6622 for_each_sched_entity(se) {
6625 cfs_rq = cfs_rq_of(se);
6626 enqueue_entity(cfs_rq, se, flags);
6628 cfs_rq->h_nr_running++;
6629 cfs_rq->idle_h_nr_running += idle_h_nr_running;
6631 if (cfs_rq_is_idle(cfs_rq))
6632 idle_h_nr_running = 1;
6634 /* end evaluation on encountering a throttled cfs_rq */
6635 if (cfs_rq_throttled(cfs_rq))
6636 goto enqueue_throttle;
6638 flags = ENQUEUE_WAKEUP;
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 enqueue_throttle;
6659 /* At this point se is NULL and we are at root level*/
6660 add_nr_running(rq, 1);
6663 * Since new tasks are assigned an initial util_avg equal to
6664 * half of the spare capacity of their CPU, tiny tasks have the
6665 * ability to cross the overutilized threshold, which will
6666 * result in the load balancer ruining all the task placement
6667 * done by EAS. As a way to mitigate that effect, do not account
6668 * for the first enqueue operation of new tasks during the
6669 * overutilized flag detection.
6671 * A better way of solving this problem would be to wait for
6672 * the PELT signals of tasks to converge before taking them
6673 * into account, but that is not straightforward to implement,
6674 * and the following generally works well enough in practice.
6677 update_overutilized_status(rq);
6680 assert_list_leaf_cfs_rq(rq);
6685 static void set_next_buddy(struct sched_entity *se);
6688 * The dequeue_task method is called before nr_running is
6689 * decreased. We remove the task from the rbtree and
6690 * update the fair scheduling stats:
6692 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
6694 struct cfs_rq *cfs_rq;
6695 struct sched_entity *se = &p->se;
6696 int task_sleep = flags & DEQUEUE_SLEEP;
6697 int idle_h_nr_running = task_has_idle_policy(p);
6698 bool was_sched_idle = sched_idle_rq(rq);
6700 util_est_dequeue(&rq->cfs, p);
6702 for_each_sched_entity(se) {
6703 cfs_rq = cfs_rq_of(se);
6704 dequeue_entity(cfs_rq, se, flags);
6706 cfs_rq->h_nr_running--;
6707 cfs_rq->idle_h_nr_running -= idle_h_nr_running;
6709 if (cfs_rq_is_idle(cfs_rq))
6710 idle_h_nr_running = 1;
6712 /* end evaluation on encountering a throttled cfs_rq */
6713 if (cfs_rq_throttled(cfs_rq))
6714 goto dequeue_throttle;
6716 /* Don't dequeue parent if it has other entities besides us */
6717 if (cfs_rq->load.weight) {
6718 /* Avoid re-evaluating load for this entity: */
6719 se = parent_entity(se);
6721 * Bias pick_next to pick a task from this cfs_rq, as
6722 * p is sleeping when it is within its sched_slice.
6724 if (task_sleep && se && !throttled_hierarchy(cfs_rq))
6728 flags |= DEQUEUE_SLEEP;
6731 for_each_sched_entity(se) {
6732 cfs_rq = cfs_rq_of(se);
6734 update_load_avg(cfs_rq, se, UPDATE_TG);
6735 se_update_runnable(se);
6736 update_cfs_group(se);
6738 cfs_rq->h_nr_running--;
6739 cfs_rq->idle_h_nr_running -= idle_h_nr_running;
6741 if (cfs_rq_is_idle(cfs_rq))
6742 idle_h_nr_running = 1;
6744 /* end evaluation on encountering a throttled cfs_rq */
6745 if (cfs_rq_throttled(cfs_rq))
6746 goto dequeue_throttle;
6750 /* At this point se is NULL and we are at root level*/
6751 sub_nr_running(rq, 1);
6753 /* balance early to pull high priority tasks */
6754 if (unlikely(!was_sched_idle && sched_idle_rq(rq)))
6755 rq->next_balance = jiffies;
6758 util_est_update(&rq->cfs, p, task_sleep);
6764 /* Working cpumask for: load_balance, load_balance_newidle. */
6765 static DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
6766 static DEFINE_PER_CPU(cpumask_var_t, select_rq_mask);
6767 static DEFINE_PER_CPU(cpumask_var_t, should_we_balance_tmpmask);
6769 #ifdef CONFIG_NO_HZ_COMMON
6772 cpumask_var_t idle_cpus_mask;
6774 int has_blocked; /* Idle CPUS has blocked load */
6775 int needs_update; /* Newly idle CPUs need their next_balance collated */
6776 unsigned long next_balance; /* in jiffy units */
6777 unsigned long next_blocked; /* Next update of blocked load in jiffies */
6778 } nohz ____cacheline_aligned;
6780 #endif /* CONFIG_NO_HZ_COMMON */
6782 static unsigned long cpu_load(struct rq *rq)
6784 return cfs_rq_load_avg(&rq->cfs);
6788 * cpu_load_without - compute CPU load without any contributions from *p
6789 * @cpu: the CPU which load is requested
6790 * @p: the task which load should be discounted
6792 * The load of a CPU is defined by the load of tasks currently enqueued on that
6793 * CPU as well as tasks which are currently sleeping after an execution on that
6796 * This method returns the load of the specified CPU by discounting the load of
6797 * the specified task, whenever the task is currently contributing to the CPU
6800 static unsigned long cpu_load_without(struct rq *rq, struct task_struct *p)
6802 struct cfs_rq *cfs_rq;
6805 /* Task has no contribution or is new */
6806 if (cpu_of(rq) != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time))
6807 return cpu_load(rq);
6810 load = READ_ONCE(cfs_rq->avg.load_avg);
6812 /* Discount task's util from CPU's util */
6813 lsub_positive(&load, task_h_load(p));
6818 static unsigned long cpu_runnable(struct rq *rq)
6820 return cfs_rq_runnable_avg(&rq->cfs);
6823 static unsigned long cpu_runnable_without(struct rq *rq, struct task_struct *p)
6825 struct cfs_rq *cfs_rq;
6826 unsigned int runnable;
6828 /* Task has no contribution or is new */
6829 if (cpu_of(rq) != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time))
6830 return cpu_runnable(rq);
6833 runnable = READ_ONCE(cfs_rq->avg.runnable_avg);
6835 /* Discount task's runnable from CPU's runnable */
6836 lsub_positive(&runnable, p->se.avg.runnable_avg);
6841 static unsigned long capacity_of(int cpu)
6843 return cpu_rq(cpu)->cpu_capacity;
6846 static void record_wakee(struct task_struct *p)
6849 * Only decay a single time; tasks that have less then 1 wakeup per
6850 * jiffy will not have built up many flips.
6852 if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) {
6853 current->wakee_flips >>= 1;
6854 current->wakee_flip_decay_ts = jiffies;
6857 if (current->last_wakee != p) {
6858 current->last_wakee = p;
6859 current->wakee_flips++;
6864 * Detect M:N waker/wakee relationships via a switching-frequency heuristic.
6866 * A waker of many should wake a different task than the one last awakened
6867 * at a frequency roughly N times higher than one of its wakees.
6869 * In order to determine whether we should let the load spread vs consolidating
6870 * to shared cache, we look for a minimum 'flip' frequency of llc_size in one
6871 * partner, and a factor of lls_size higher frequency in the other.
6873 * With both conditions met, we can be relatively sure that the relationship is
6874 * non-monogamous, with partner count exceeding socket size.
6876 * Waker/wakee being client/server, worker/dispatcher, interrupt source or
6877 * whatever is irrelevant, spread criteria is apparent partner count exceeds
6880 static int wake_wide(struct task_struct *p)
6882 unsigned int master = current->wakee_flips;
6883 unsigned int slave = p->wakee_flips;
6884 int factor = __this_cpu_read(sd_llc_size);
6887 swap(master, slave);
6888 if (slave < factor || master < slave * factor)
6894 * The purpose of wake_affine() is to quickly determine on which CPU we can run
6895 * soonest. For the purpose of speed we only consider the waking and previous
6898 * wake_affine_idle() - only considers 'now', it check if the waking CPU is
6899 * cache-affine and is (or will be) idle.
6901 * wake_affine_weight() - considers the weight to reflect the average
6902 * scheduling latency of the CPUs. This seems to work
6903 * for the overloaded case.
6906 wake_affine_idle(int this_cpu, int prev_cpu, int sync)
6909 * If this_cpu is idle, it implies the wakeup is from interrupt
6910 * context. Only allow the move if cache is shared. Otherwise an
6911 * interrupt intensive workload could force all tasks onto one
6912 * node depending on the IO topology or IRQ affinity settings.
6914 * If the prev_cpu is idle and cache affine then avoid a migration.
6915 * There is no guarantee that the cache hot data from an interrupt
6916 * is more important than cache hot data on the prev_cpu and from
6917 * a cpufreq perspective, it's better to have higher utilisation
6920 if (available_idle_cpu(this_cpu) && cpus_share_cache(this_cpu, prev_cpu))
6921 return available_idle_cpu(prev_cpu) ? prev_cpu : this_cpu;
6923 if (sync && cpu_rq(this_cpu)->nr_running == 1)
6926 if (available_idle_cpu(prev_cpu))
6929 return nr_cpumask_bits;
6933 wake_affine_weight(struct sched_domain *sd, struct task_struct *p,
6934 int this_cpu, int prev_cpu, int sync)
6936 s64 this_eff_load, prev_eff_load;
6937 unsigned long task_load;
6939 this_eff_load = cpu_load(cpu_rq(this_cpu));
6942 unsigned long current_load = task_h_load(current);
6944 if (current_load > this_eff_load)
6947 this_eff_load -= current_load;
6950 task_load = task_h_load(p);
6952 this_eff_load += task_load;
6953 if (sched_feat(WA_BIAS))
6954 this_eff_load *= 100;
6955 this_eff_load *= capacity_of(prev_cpu);
6957 prev_eff_load = cpu_load(cpu_rq(prev_cpu));
6958 prev_eff_load -= task_load;
6959 if (sched_feat(WA_BIAS))
6960 prev_eff_load *= 100 + (sd->imbalance_pct - 100) / 2;
6961 prev_eff_load *= capacity_of(this_cpu);
6964 * If sync, adjust the weight of prev_eff_load such that if
6965 * prev_eff == this_eff that select_idle_sibling() will consider
6966 * stacking the wakee on top of the waker if no other CPU is
6972 return this_eff_load < prev_eff_load ? this_cpu : nr_cpumask_bits;
6975 static int wake_affine(struct sched_domain *sd, struct task_struct *p,
6976 int this_cpu, int prev_cpu, int sync)
6978 int target = nr_cpumask_bits;
6980 if (sched_feat(WA_IDLE))
6981 target = wake_affine_idle(this_cpu, prev_cpu, sync);
6983 if (sched_feat(WA_WEIGHT) && target == nr_cpumask_bits)
6984 target = wake_affine_weight(sd, p, this_cpu, prev_cpu, sync);
6986 schedstat_inc(p->stats.nr_wakeups_affine_attempts);
6987 if (target != this_cpu)
6990 schedstat_inc(sd->ttwu_move_affine);
6991 schedstat_inc(p->stats.nr_wakeups_affine);
6995 static struct sched_group *
6996 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu);
6999 * find_idlest_group_cpu - find the idlest CPU among the CPUs in the group.
7002 find_idlest_group_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
7004 unsigned long load, min_load = ULONG_MAX;
7005 unsigned int min_exit_latency = UINT_MAX;
7006 u64 latest_idle_timestamp = 0;
7007 int least_loaded_cpu = this_cpu;
7008 int shallowest_idle_cpu = -1;
7011 /* Check if we have any choice: */
7012 if (group->group_weight == 1)
7013 return cpumask_first(sched_group_span(group));
7015 /* Traverse only the allowed CPUs */
7016 for_each_cpu_and(i, sched_group_span(group), p->cpus_ptr) {
7017 struct rq *rq = cpu_rq(i);
7019 if (!sched_core_cookie_match(rq, p))
7022 if (sched_idle_cpu(i))
7025 if (available_idle_cpu(i)) {
7026 struct cpuidle_state *idle = idle_get_state(rq);
7027 if (idle && idle->exit_latency < min_exit_latency) {
7029 * We give priority to a CPU whose idle state
7030 * has the smallest exit latency irrespective
7031 * of any idle timestamp.
7033 min_exit_latency = idle->exit_latency;
7034 latest_idle_timestamp = rq->idle_stamp;
7035 shallowest_idle_cpu = i;
7036 } else if ((!idle || idle->exit_latency == min_exit_latency) &&
7037 rq->idle_stamp > latest_idle_timestamp) {
7039 * If equal or no active idle state, then
7040 * the most recently idled CPU might have
7043 latest_idle_timestamp = rq->idle_stamp;
7044 shallowest_idle_cpu = i;
7046 } else if (shallowest_idle_cpu == -1) {
7047 load = cpu_load(cpu_rq(i));
7048 if (load < min_load) {
7050 least_loaded_cpu = i;
7055 return shallowest_idle_cpu != -1 ? shallowest_idle_cpu : least_loaded_cpu;
7058 static inline int find_idlest_cpu(struct sched_domain *sd, struct task_struct *p,
7059 int cpu, int prev_cpu, int sd_flag)
7063 if (!cpumask_intersects(sched_domain_span(sd), p->cpus_ptr))
7067 * We need task's util for cpu_util_without, sync it up to
7068 * prev_cpu's last_update_time.
7070 if (!(sd_flag & SD_BALANCE_FORK))
7071 sync_entity_load_avg(&p->se);
7074 struct sched_group *group;
7075 struct sched_domain *tmp;
7078 if (!(sd->flags & sd_flag)) {
7083 group = find_idlest_group(sd, p, cpu);
7089 new_cpu = find_idlest_group_cpu(group, p, cpu);
7090 if (new_cpu == cpu) {
7091 /* Now try balancing at a lower domain level of 'cpu': */
7096 /* Now try balancing at a lower domain level of 'new_cpu': */
7098 weight = sd->span_weight;
7100 for_each_domain(cpu, tmp) {
7101 if (weight <= tmp->span_weight)
7103 if (tmp->flags & sd_flag)
7111 static inline int __select_idle_cpu(int cpu, struct task_struct *p)
7113 if ((available_idle_cpu(cpu) || sched_idle_cpu(cpu)) &&
7114 sched_cpu_cookie_match(cpu_rq(cpu), p))
7120 #ifdef CONFIG_SCHED_SMT
7121 DEFINE_STATIC_KEY_FALSE(sched_smt_present);
7122 EXPORT_SYMBOL_GPL(sched_smt_present);
7124 static inline void set_idle_cores(int cpu, int val)
7126 struct sched_domain_shared *sds;
7128 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
7130 WRITE_ONCE(sds->has_idle_cores, val);
7133 static inline bool test_idle_cores(int cpu)
7135 struct sched_domain_shared *sds;
7137 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
7139 return READ_ONCE(sds->has_idle_cores);
7145 * Scans the local SMT mask to see if the entire core is idle, and records this
7146 * information in sd_llc_shared->has_idle_cores.
7148 * Since SMT siblings share all cache levels, inspecting this limited remote
7149 * state should be fairly cheap.
7151 void __update_idle_core(struct rq *rq)
7153 int core = cpu_of(rq);
7157 if (test_idle_cores(core))
7160 for_each_cpu(cpu, cpu_smt_mask(core)) {
7164 if (!available_idle_cpu(cpu))
7168 set_idle_cores(core, 1);
7174 * Scan the entire LLC domain for idle cores; this dynamically switches off if
7175 * there are no idle cores left in the system; tracked through
7176 * sd_llc->shared->has_idle_cores and enabled through update_idle_core() above.
7178 static int select_idle_core(struct task_struct *p, int core, struct cpumask *cpus, int *idle_cpu)
7183 for_each_cpu(cpu, cpu_smt_mask(core)) {
7184 if (!available_idle_cpu(cpu)) {
7186 if (*idle_cpu == -1) {
7187 if (sched_idle_cpu(cpu) && cpumask_test_cpu(cpu, p->cpus_ptr)) {
7195 if (*idle_cpu == -1 && cpumask_test_cpu(cpu, p->cpus_ptr))
7202 cpumask_andnot(cpus, cpus, cpu_smt_mask(core));
7207 * Scan the local SMT mask for idle CPUs.
7209 static int select_idle_smt(struct task_struct *p, int target)
7213 for_each_cpu_and(cpu, cpu_smt_mask(target), p->cpus_ptr) {
7216 if (available_idle_cpu(cpu) || sched_idle_cpu(cpu))
7223 #else /* CONFIG_SCHED_SMT */
7225 static inline void set_idle_cores(int cpu, int val)
7229 static inline bool test_idle_cores(int cpu)
7234 static inline int select_idle_core(struct task_struct *p, int core, struct cpumask *cpus, int *idle_cpu)
7236 return __select_idle_cpu(core, p);
7239 static inline int select_idle_smt(struct task_struct *p, int target)
7244 #endif /* CONFIG_SCHED_SMT */
7247 * Scan the LLC domain for idle CPUs; this is dynamically regulated by
7248 * comparing the average scan cost (tracked in sd->avg_scan_cost) against the
7249 * average idle time for this rq (as found in rq->avg_idle).
7251 static int select_idle_cpu(struct task_struct *p, struct sched_domain *sd, bool has_idle_core, int target)
7253 struct cpumask *cpus = this_cpu_cpumask_var_ptr(select_rq_mask);
7254 int i, cpu, idle_cpu = -1, nr = INT_MAX;
7255 struct sched_domain_shared *sd_share;
7256 struct rq *this_rq = this_rq();
7257 int this = smp_processor_id();
7258 struct sched_domain *this_sd = NULL;
7261 cpumask_and(cpus, sched_domain_span(sd), p->cpus_ptr);
7263 if (sched_feat(SIS_PROP) && !has_idle_core) {
7264 u64 avg_cost, avg_idle, span_avg;
7265 unsigned long now = jiffies;
7267 this_sd = rcu_dereference(*this_cpu_ptr(&sd_llc));
7272 * If we're busy, the assumption that the last idle period
7273 * predicts the future is flawed; age away the remaining
7274 * predicted idle time.
7276 if (unlikely(this_rq->wake_stamp < now)) {
7277 while (this_rq->wake_stamp < now && this_rq->wake_avg_idle) {
7278 this_rq->wake_stamp++;
7279 this_rq->wake_avg_idle >>= 1;
7283 avg_idle = this_rq->wake_avg_idle;
7284 avg_cost = this_sd->avg_scan_cost + 1;
7286 span_avg = sd->span_weight * avg_idle;
7287 if (span_avg > 4*avg_cost)
7288 nr = div_u64(span_avg, avg_cost);
7292 time = cpu_clock(this);
7295 if (sched_feat(SIS_UTIL)) {
7296 sd_share = rcu_dereference(per_cpu(sd_llc_shared, target));
7298 /* because !--nr is the condition to stop scan */
7299 nr = READ_ONCE(sd_share->nr_idle_scan) + 1;
7300 /* overloaded LLC is unlikely to have idle cpu/core */
7306 for_each_cpu_wrap(cpu, cpus, target + 1) {
7307 if (has_idle_core) {
7308 i = select_idle_core(p, cpu, cpus, &idle_cpu);
7309 if ((unsigned int)i < nr_cpumask_bits)
7315 idle_cpu = __select_idle_cpu(cpu, p);
7316 if ((unsigned int)idle_cpu < nr_cpumask_bits)
7322 set_idle_cores(target, false);
7324 if (sched_feat(SIS_PROP) && this_sd && !has_idle_core) {
7325 time = cpu_clock(this) - time;
7328 * Account for the scan cost of wakeups against the average
7331 this_rq->wake_avg_idle -= min(this_rq->wake_avg_idle, time);
7333 update_avg(&this_sd->avg_scan_cost, time);
7340 * Scan the asym_capacity domain for idle CPUs; pick the first idle one on which
7341 * the task fits. If no CPU is big enough, but there are idle ones, try to
7342 * maximize capacity.
7345 select_idle_capacity(struct task_struct *p, struct sched_domain *sd, int target)
7347 unsigned long task_util, util_min, util_max, best_cap = 0;
7348 int fits, best_fits = 0;
7349 int cpu, best_cpu = -1;
7350 struct cpumask *cpus;
7352 cpus = this_cpu_cpumask_var_ptr(select_rq_mask);
7353 cpumask_and(cpus, sched_domain_span(sd), p->cpus_ptr);
7355 task_util = task_util_est(p);
7356 util_min = uclamp_eff_value(p, UCLAMP_MIN);
7357 util_max = uclamp_eff_value(p, UCLAMP_MAX);
7359 for_each_cpu_wrap(cpu, cpus, target) {
7360 unsigned long cpu_cap = capacity_of(cpu);
7362 if (!available_idle_cpu(cpu) && !sched_idle_cpu(cpu))
7365 fits = util_fits_cpu(task_util, util_min, util_max, cpu);
7367 /* This CPU fits with all requirements */
7371 * Only the min performance hint (i.e. uclamp_min) doesn't fit.
7372 * Look for the CPU with best capacity.
7375 cpu_cap = capacity_orig_of(cpu) - thermal_load_avg(cpu_rq(cpu));
7378 * First, select CPU which fits better (-1 being better than 0).
7379 * Then, select the one with best capacity at same level.
7381 if ((fits < best_fits) ||
7382 ((fits == best_fits) && (cpu_cap > best_cap))) {
7392 static inline bool asym_fits_cpu(unsigned long util,
7393 unsigned long util_min,
7394 unsigned long util_max,
7397 if (sched_asym_cpucap_active())
7399 * Return true only if the cpu fully fits the task requirements
7400 * which include the utilization and the performance hints.
7402 return (util_fits_cpu(util, util_min, util_max, cpu) > 0);
7408 * Try and locate an idle core/thread in the LLC cache domain.
7410 static int select_idle_sibling(struct task_struct *p, int prev, int target)
7412 bool has_idle_core = false;
7413 struct sched_domain *sd;
7414 unsigned long task_util, util_min, util_max;
7415 int i, recent_used_cpu;
7418 * On asymmetric system, update task utilization because we will check
7419 * that the task fits with cpu's capacity.
7421 if (sched_asym_cpucap_active()) {
7422 sync_entity_load_avg(&p->se);
7423 task_util = task_util_est(p);
7424 util_min = uclamp_eff_value(p, UCLAMP_MIN);
7425 util_max = uclamp_eff_value(p, UCLAMP_MAX);
7429 * per-cpu select_rq_mask usage
7431 lockdep_assert_irqs_disabled();
7433 if ((available_idle_cpu(target) || sched_idle_cpu(target)) &&
7434 asym_fits_cpu(task_util, util_min, util_max, target))
7438 * If the previous CPU is cache affine and idle, don't be stupid:
7440 if (prev != target && cpus_share_cache(prev, target) &&
7441 (available_idle_cpu(prev) || sched_idle_cpu(prev)) &&
7442 asym_fits_cpu(task_util, util_min, util_max, prev))
7446 * Allow a per-cpu kthread to stack with the wakee if the
7447 * kworker thread and the tasks previous CPUs are the same.
7448 * The assumption is that the wakee queued work for the
7449 * per-cpu kthread that is now complete and the wakeup is
7450 * essentially a sync wakeup. An obvious example of this
7451 * pattern is IO completions.
7453 if (is_per_cpu_kthread(current) &&
7455 prev == smp_processor_id() &&
7456 this_rq()->nr_running <= 1 &&
7457 asym_fits_cpu(task_util, util_min, util_max, prev)) {
7461 /* Check a recently used CPU as a potential idle candidate: */
7462 recent_used_cpu = p->recent_used_cpu;
7463 p->recent_used_cpu = prev;
7464 if (recent_used_cpu != prev &&
7465 recent_used_cpu != target &&
7466 cpus_share_cache(recent_used_cpu, target) &&
7467 (available_idle_cpu(recent_used_cpu) || sched_idle_cpu(recent_used_cpu)) &&
7468 cpumask_test_cpu(recent_used_cpu, p->cpus_ptr) &&
7469 asym_fits_cpu(task_util, util_min, util_max, recent_used_cpu)) {
7470 return recent_used_cpu;
7474 * For asymmetric CPU capacity systems, our domain of interest is
7475 * sd_asym_cpucapacity rather than sd_llc.
7477 if (sched_asym_cpucap_active()) {
7478 sd = rcu_dereference(per_cpu(sd_asym_cpucapacity, target));
7480 * On an asymmetric CPU capacity system where an exclusive
7481 * cpuset defines a symmetric island (i.e. one unique
7482 * capacity_orig value through the cpuset), the key will be set
7483 * but the CPUs within that cpuset will not have a domain with
7484 * SD_ASYM_CPUCAPACITY. These should follow the usual symmetric
7488 i = select_idle_capacity(p, sd, target);
7489 return ((unsigned)i < nr_cpumask_bits) ? i : target;
7493 sd = rcu_dereference(per_cpu(sd_llc, target));
7497 if (sched_smt_active()) {
7498 has_idle_core = test_idle_cores(target);
7500 if (!has_idle_core && cpus_share_cache(prev, target)) {
7501 i = select_idle_smt(p, prev);
7502 if ((unsigned int)i < nr_cpumask_bits)
7507 i = select_idle_cpu(p, sd, has_idle_core, target);
7508 if ((unsigned)i < nr_cpumask_bits)
7515 * cpu_util() - Estimates the amount of CPU capacity used by CFS tasks.
7516 * @cpu: the CPU to get the utilization for
7517 * @p: task for which the CPU utilization should be predicted or NULL
7518 * @dst_cpu: CPU @p migrates to, -1 if @p moves from @cpu or @p == NULL
7519 * @boost: 1 to enable boosting, otherwise 0
7521 * The unit of the return value must be the same as the one of CPU capacity
7522 * so that CPU utilization can be compared with CPU capacity.
7524 * CPU utilization is the sum of running time of runnable tasks plus the
7525 * recent utilization of currently non-runnable tasks on that CPU.
7526 * It represents the amount of CPU capacity currently used by CFS tasks in
7527 * the range [0..max CPU capacity] with max CPU capacity being the CPU
7528 * capacity at f_max.
7530 * The estimated CPU utilization is defined as the maximum between CPU
7531 * utilization and sum of the estimated utilization of the currently
7532 * runnable tasks on that CPU. It preserves a utilization "snapshot" of
7533 * previously-executed tasks, which helps better deduce how busy a CPU will
7534 * be when a long-sleeping task wakes up. The contribution to CPU utilization
7535 * of such a task would be significantly decayed at this point of time.
7537 * Boosted CPU utilization is defined as max(CPU runnable, CPU utilization).
7538 * CPU contention for CFS tasks can be detected by CPU runnable > CPU
7539 * utilization. Boosting is implemented in cpu_util() so that internal
7540 * users (e.g. EAS) can use it next to external users (e.g. schedutil),
7541 * latter via cpu_util_cfs_boost().
7543 * CPU utilization can be higher than the current CPU capacity
7544 * (f_curr/f_max * max CPU capacity) or even the max CPU capacity because
7545 * of rounding errors as well as task migrations or wakeups of new tasks.
7546 * CPU utilization has to be capped to fit into the [0..max CPU capacity]
7547 * range. Otherwise a group of CPUs (CPU0 util = 121% + CPU1 util = 80%)
7548 * could be seen as over-utilized even though CPU1 has 20% of spare CPU
7549 * capacity. CPU utilization is allowed to overshoot current CPU capacity
7550 * though since this is useful for predicting the CPU capacity required
7551 * after task migrations (scheduler-driven DVFS).
7553 * Return: (Boosted) (estimated) utilization for the specified CPU.
7555 static unsigned long
7556 cpu_util(int cpu, struct task_struct *p, int dst_cpu, int boost)
7558 struct cfs_rq *cfs_rq = &cpu_rq(cpu)->cfs;
7559 unsigned long util = READ_ONCE(cfs_rq->avg.util_avg);
7560 unsigned long runnable;
7563 runnable = READ_ONCE(cfs_rq->avg.runnable_avg);
7564 util = max(util, runnable);
7568 * If @dst_cpu is -1 or @p migrates from @cpu to @dst_cpu remove its
7569 * contribution. If @p migrates from another CPU to @cpu add its
7570 * contribution. In all the other cases @cpu is not impacted by the
7571 * migration so its util_avg is already correct.
7573 if (p && task_cpu(p) == cpu && dst_cpu != cpu)
7574 lsub_positive(&util, task_util(p));
7575 else if (p && task_cpu(p) != cpu && dst_cpu == cpu)
7576 util += task_util(p);
7578 if (sched_feat(UTIL_EST)) {
7579 unsigned long util_est;
7581 util_est = READ_ONCE(cfs_rq->avg.util_est.enqueued);
7584 * During wake-up @p isn't enqueued yet and doesn't contribute
7585 * to any cpu_rq(cpu)->cfs.avg.util_est.enqueued.
7586 * If @dst_cpu == @cpu add it to "simulate" cpu_util after @p
7587 * has been enqueued.
7589 * During exec (@dst_cpu = -1) @p is enqueued and does
7590 * contribute to cpu_rq(cpu)->cfs.util_est.enqueued.
7591 * Remove it to "simulate" cpu_util without @p's contribution.
7593 * Despite the task_on_rq_queued(@p) check there is still a
7594 * small window for a possible race when an exec
7595 * select_task_rq_fair() races with LB's detach_task().
7599 * p->on_rq = TASK_ON_RQ_MIGRATING;
7600 * -------------------------------- A
7602 * dequeue_task_fair() + Race Time
7603 * util_est_dequeue() /
7604 * -------------------------------- B
7606 * The additional check "current == p" is required to further
7607 * reduce the race window.
7610 util_est += _task_util_est(p);
7611 else if (p && unlikely(task_on_rq_queued(p) || current == p))
7612 lsub_positive(&util_est, _task_util_est(p));
7614 util = max(util, util_est);
7617 return min(util, capacity_orig_of(cpu));
7620 unsigned long cpu_util_cfs(int cpu)
7622 return cpu_util(cpu, NULL, -1, 0);
7625 unsigned long cpu_util_cfs_boost(int cpu)
7627 return cpu_util(cpu, NULL, -1, 1);
7631 * cpu_util_without: compute cpu utilization without any contributions from *p
7632 * @cpu: the CPU which utilization is requested
7633 * @p: the task which utilization should be discounted
7635 * The utilization of a CPU is defined by the utilization of tasks currently
7636 * enqueued on that CPU as well as tasks which are currently sleeping after an
7637 * execution on that CPU.
7639 * This method returns the utilization of the specified CPU by discounting the
7640 * utilization of the specified task, whenever the task is currently
7641 * contributing to the CPU utilization.
7643 static unsigned long cpu_util_without(int cpu, struct task_struct *p)
7645 /* Task has no contribution or is new */
7646 if (cpu != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time))
7649 return cpu_util(cpu, p, -1, 0);
7653 * energy_env - Utilization landscape for energy estimation.
7654 * @task_busy_time: Utilization contribution by the task for which we test the
7655 * placement. Given by eenv_task_busy_time().
7656 * @pd_busy_time: Utilization of the whole perf domain without the task
7657 * contribution. Given by eenv_pd_busy_time().
7658 * @cpu_cap: Maximum CPU capacity for the perf domain.
7659 * @pd_cap: Entire perf domain capacity. (pd->nr_cpus * cpu_cap).
7662 unsigned long task_busy_time;
7663 unsigned long pd_busy_time;
7664 unsigned long cpu_cap;
7665 unsigned long pd_cap;
7669 * Compute the task busy time for compute_energy(). This time cannot be
7670 * injected directly into effective_cpu_util() because of the IRQ scaling.
7671 * The latter only makes sense with the most recent CPUs where the task has
7674 static inline void eenv_task_busy_time(struct energy_env *eenv,
7675 struct task_struct *p, int prev_cpu)
7677 unsigned long busy_time, max_cap = arch_scale_cpu_capacity(prev_cpu);
7678 unsigned long irq = cpu_util_irq(cpu_rq(prev_cpu));
7680 if (unlikely(irq >= max_cap))
7681 busy_time = max_cap;
7683 busy_time = scale_irq_capacity(task_util_est(p), irq, max_cap);
7685 eenv->task_busy_time = busy_time;
7689 * Compute the perf_domain (PD) busy time for compute_energy(). Based on the
7690 * utilization for each @pd_cpus, it however doesn't take into account
7691 * clamping since the ratio (utilization / cpu_capacity) is already enough to
7692 * scale the EM reported power consumption at the (eventually clamped)
7695 * The contribution of the task @p for which we want to estimate the
7696 * energy cost is removed (by cpu_util()) and must be calculated
7697 * separately (see eenv_task_busy_time). This ensures:
7699 * - A stable PD utilization, no matter which CPU of that PD we want to place
7702 * - A fair comparison between CPUs as the task contribution (task_util())
7703 * will always be the same no matter which CPU utilization we rely on
7704 * (util_avg or util_est).
7706 * Set @eenv busy time for the PD that spans @pd_cpus. This busy time can't
7707 * exceed @eenv->pd_cap.
7709 static inline void eenv_pd_busy_time(struct energy_env *eenv,
7710 struct cpumask *pd_cpus,
7711 struct task_struct *p)
7713 unsigned long busy_time = 0;
7716 for_each_cpu(cpu, pd_cpus) {
7717 unsigned long util = cpu_util(cpu, p, -1, 0);
7719 busy_time += effective_cpu_util(cpu, util, ENERGY_UTIL, NULL);
7722 eenv->pd_busy_time = min(eenv->pd_cap, busy_time);
7726 * Compute the maximum utilization for compute_energy() when the task @p
7727 * is placed on the cpu @dst_cpu.
7729 * Returns the maximum utilization among @eenv->cpus. This utilization can't
7730 * exceed @eenv->cpu_cap.
7732 static inline unsigned long
7733 eenv_pd_max_util(struct energy_env *eenv, struct cpumask *pd_cpus,
7734 struct task_struct *p, int dst_cpu)
7736 unsigned long max_util = 0;
7739 for_each_cpu(cpu, pd_cpus) {
7740 struct task_struct *tsk = (cpu == dst_cpu) ? p : NULL;
7741 unsigned long util = cpu_util(cpu, p, dst_cpu, 1);
7742 unsigned long eff_util;
7745 * Performance domain frequency: utilization clamping
7746 * must be considered since it affects the selection
7747 * of the performance domain frequency.
7748 * NOTE: in case RT tasks are running, by default the
7749 * FREQUENCY_UTIL's utilization can be max OPP.
7751 eff_util = effective_cpu_util(cpu, util, FREQUENCY_UTIL, tsk);
7752 max_util = max(max_util, eff_util);
7755 return min(max_util, eenv->cpu_cap);
7759 * compute_energy(): Use the Energy Model to estimate the energy that @pd would
7760 * consume for a given utilization landscape @eenv. When @dst_cpu < 0, the task
7761 * contribution is ignored.
7763 static inline unsigned long
7764 compute_energy(struct energy_env *eenv, struct perf_domain *pd,
7765 struct cpumask *pd_cpus, struct task_struct *p, int dst_cpu)
7767 unsigned long max_util = eenv_pd_max_util(eenv, pd_cpus, p, dst_cpu);
7768 unsigned long busy_time = eenv->pd_busy_time;
7771 busy_time = min(eenv->pd_cap, busy_time + eenv->task_busy_time);
7773 return em_cpu_energy(pd->em_pd, max_util, busy_time, eenv->cpu_cap);
7777 * find_energy_efficient_cpu(): Find most energy-efficient target CPU for the
7778 * waking task. find_energy_efficient_cpu() looks for the CPU with maximum
7779 * spare capacity in each performance domain and uses it as a potential
7780 * candidate to execute the task. Then, it uses the Energy Model to figure
7781 * out which of the CPU candidates is the most energy-efficient.
7783 * The rationale for this heuristic is as follows. In a performance domain,
7784 * all the most energy efficient CPU candidates (according to the Energy
7785 * Model) are those for which we'll request a low frequency. When there are
7786 * several CPUs for which the frequency request will be the same, we don't
7787 * have enough data to break the tie between them, because the Energy Model
7788 * only includes active power costs. With this model, if we assume that
7789 * frequency requests follow utilization (e.g. using schedutil), the CPU with
7790 * the maximum spare capacity in a performance domain is guaranteed to be among
7791 * the best candidates of the performance domain.
7793 * In practice, it could be preferable from an energy standpoint to pack
7794 * small tasks on a CPU in order to let other CPUs go in deeper idle states,
7795 * but that could also hurt our chances to go cluster idle, and we have no
7796 * ways to tell with the current Energy Model if this is actually a good
7797 * idea or not. So, find_energy_efficient_cpu() basically favors
7798 * cluster-packing, and spreading inside a cluster. That should at least be
7799 * a good thing for latency, and this is consistent with the idea that most
7800 * of the energy savings of EAS come from the asymmetry of the system, and
7801 * not so much from breaking the tie between identical CPUs. That's also the
7802 * reason why EAS is enabled in the topology code only for systems where
7803 * SD_ASYM_CPUCAPACITY is set.
7805 * NOTE: Forkees are not accepted in the energy-aware wake-up path because
7806 * they don't have any useful utilization data yet and it's not possible to
7807 * forecast their impact on energy consumption. Consequently, they will be
7808 * placed by find_idlest_cpu() on the least loaded CPU, which might turn out
7809 * to be energy-inefficient in some use-cases. The alternative would be to
7810 * bias new tasks towards specific types of CPUs first, or to try to infer
7811 * their util_avg from the parent task, but those heuristics could hurt
7812 * other use-cases too. So, until someone finds a better way to solve this,
7813 * let's keep things simple by re-using the existing slow path.
7815 static int find_energy_efficient_cpu(struct task_struct *p, int prev_cpu)
7817 struct cpumask *cpus = this_cpu_cpumask_var_ptr(select_rq_mask);
7818 unsigned long prev_delta = ULONG_MAX, best_delta = ULONG_MAX;
7819 unsigned long p_util_min = uclamp_is_used() ? uclamp_eff_value(p, UCLAMP_MIN) : 0;
7820 unsigned long p_util_max = uclamp_is_used() ? uclamp_eff_value(p, UCLAMP_MAX) : 1024;
7821 struct root_domain *rd = this_rq()->rd;
7822 int cpu, best_energy_cpu, target = -1;
7823 int prev_fits = -1, best_fits = -1;
7824 unsigned long best_thermal_cap = 0;
7825 unsigned long prev_thermal_cap = 0;
7826 struct sched_domain *sd;
7827 struct perf_domain *pd;
7828 struct energy_env eenv;
7831 pd = rcu_dereference(rd->pd);
7832 if (!pd || READ_ONCE(rd->overutilized))
7836 * Energy-aware wake-up happens on the lowest sched_domain starting
7837 * from sd_asym_cpucapacity spanning over this_cpu and prev_cpu.
7839 sd = rcu_dereference(*this_cpu_ptr(&sd_asym_cpucapacity));
7840 while (sd && !cpumask_test_cpu(prev_cpu, sched_domain_span(sd)))
7847 sync_entity_load_avg(&p->se);
7848 if (!task_util_est(p) && p_util_min == 0)
7851 eenv_task_busy_time(&eenv, p, prev_cpu);
7853 for (; pd; pd = pd->next) {
7854 unsigned long util_min = p_util_min, util_max = p_util_max;
7855 unsigned long cpu_cap, cpu_thermal_cap, util;
7856 long prev_spare_cap = -1, max_spare_cap = -1;
7857 unsigned long rq_util_min, rq_util_max;
7858 unsigned long cur_delta, base_energy;
7859 int max_spare_cap_cpu = -1;
7860 int fits, max_fits = -1;
7862 cpumask_and(cpus, perf_domain_span(pd), cpu_online_mask);
7864 if (cpumask_empty(cpus))
7867 /* Account thermal pressure for the energy estimation */
7868 cpu = cpumask_first(cpus);
7869 cpu_thermal_cap = arch_scale_cpu_capacity(cpu);
7870 cpu_thermal_cap -= arch_scale_thermal_pressure(cpu);
7872 eenv.cpu_cap = cpu_thermal_cap;
7875 for_each_cpu(cpu, cpus) {
7876 struct rq *rq = cpu_rq(cpu);
7878 eenv.pd_cap += cpu_thermal_cap;
7880 if (!cpumask_test_cpu(cpu, sched_domain_span(sd)))
7883 if (!cpumask_test_cpu(cpu, p->cpus_ptr))
7886 util = cpu_util(cpu, p, cpu, 0);
7887 cpu_cap = capacity_of(cpu);
7890 * Skip CPUs that cannot satisfy the capacity request.
7891 * IOW, placing the task there would make the CPU
7892 * overutilized. Take uclamp into account to see how
7893 * much capacity we can get out of the CPU; this is
7894 * aligned with sched_cpu_util().
7896 if (uclamp_is_used() && !uclamp_rq_is_idle(rq)) {
7898 * Open code uclamp_rq_util_with() except for
7899 * the clamp() part. Ie: apply max aggregation
7900 * only. util_fits_cpu() logic requires to
7901 * operate on non clamped util but must use the
7902 * max-aggregated uclamp_{min, max}.
7904 rq_util_min = uclamp_rq_get(rq, UCLAMP_MIN);
7905 rq_util_max = uclamp_rq_get(rq, UCLAMP_MAX);
7907 util_min = max(rq_util_min, p_util_min);
7908 util_max = max(rq_util_max, p_util_max);
7911 fits = util_fits_cpu(util, util_min, util_max, cpu);
7915 lsub_positive(&cpu_cap, util);
7917 if (cpu == prev_cpu) {
7918 /* Always use prev_cpu as a candidate. */
7919 prev_spare_cap = cpu_cap;
7921 } else if ((fits > max_fits) ||
7922 ((fits == max_fits) && ((long)cpu_cap > max_spare_cap))) {
7924 * Find the CPU with the maximum spare capacity
7925 * among the remaining CPUs in the performance
7928 max_spare_cap = cpu_cap;
7929 max_spare_cap_cpu = cpu;
7934 if (max_spare_cap_cpu < 0 && prev_spare_cap < 0)
7937 eenv_pd_busy_time(&eenv, cpus, p);
7938 /* Compute the 'base' energy of the pd, without @p */
7939 base_energy = compute_energy(&eenv, pd, cpus, p, -1);
7941 /* Evaluate the energy impact of using prev_cpu. */
7942 if (prev_spare_cap > -1) {
7943 prev_delta = compute_energy(&eenv, pd, cpus, p,
7945 /* CPU utilization has changed */
7946 if (prev_delta < base_energy)
7948 prev_delta -= base_energy;
7949 prev_thermal_cap = cpu_thermal_cap;
7950 best_delta = min(best_delta, prev_delta);
7953 /* Evaluate the energy impact of using max_spare_cap_cpu. */
7954 if (max_spare_cap_cpu >= 0 && max_spare_cap > prev_spare_cap) {
7955 /* Current best energy cpu fits better */
7956 if (max_fits < best_fits)
7960 * Both don't fit performance hint (i.e. uclamp_min)
7961 * but best energy cpu has better capacity.
7963 if ((max_fits < 0) &&
7964 (cpu_thermal_cap <= best_thermal_cap))
7967 cur_delta = compute_energy(&eenv, pd, cpus, p,
7969 /* CPU utilization has changed */
7970 if (cur_delta < base_energy)
7972 cur_delta -= base_energy;
7975 * Both fit for the task but best energy cpu has lower
7978 if ((max_fits > 0) && (best_fits > 0) &&
7979 (cur_delta >= best_delta))
7982 best_delta = cur_delta;
7983 best_energy_cpu = max_spare_cap_cpu;
7984 best_fits = max_fits;
7985 best_thermal_cap = cpu_thermal_cap;
7990 if ((best_fits > prev_fits) ||
7991 ((best_fits > 0) && (best_delta < prev_delta)) ||
7992 ((best_fits < 0) && (best_thermal_cap > prev_thermal_cap)))
7993 target = best_energy_cpu;
8004 * select_task_rq_fair: Select target runqueue for the waking task in domains
8005 * that have the relevant SD flag set. In practice, this is SD_BALANCE_WAKE,
8006 * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
8008 * Balances load by selecting the idlest CPU in the idlest group, or under
8009 * certain conditions an idle sibling CPU if the domain has SD_WAKE_AFFINE set.
8011 * Returns the target CPU number.
8014 select_task_rq_fair(struct task_struct *p, int prev_cpu, int wake_flags)
8016 int sync = (wake_flags & WF_SYNC) && !(current->flags & PF_EXITING);
8017 struct sched_domain *tmp, *sd = NULL;
8018 int cpu = smp_processor_id();
8019 int new_cpu = prev_cpu;
8020 int want_affine = 0;
8021 /* SD_flags and WF_flags share the first nibble */
8022 int sd_flag = wake_flags & 0xF;
8025 * required for stable ->cpus_allowed
8027 lockdep_assert_held(&p->pi_lock);
8028 if (wake_flags & WF_TTWU) {
8031 if ((wake_flags & WF_CURRENT_CPU) &&
8032 cpumask_test_cpu(cpu, p->cpus_ptr))
8035 if (sched_energy_enabled()) {
8036 new_cpu = find_energy_efficient_cpu(p, prev_cpu);
8042 want_affine = !wake_wide(p) && cpumask_test_cpu(cpu, p->cpus_ptr);
8046 for_each_domain(cpu, tmp) {
8048 * If both 'cpu' and 'prev_cpu' are part of this domain,
8049 * cpu is a valid SD_WAKE_AFFINE target.
8051 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
8052 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
8053 if (cpu != prev_cpu)
8054 new_cpu = wake_affine(tmp, p, cpu, prev_cpu, sync);
8056 sd = NULL; /* Prefer wake_affine over balance flags */
8061 * Usually only true for WF_EXEC and WF_FORK, as sched_domains
8062 * usually do not have SD_BALANCE_WAKE set. That means wakeup
8063 * will usually go to the fast path.
8065 if (tmp->flags & sd_flag)
8067 else if (!want_affine)
8073 new_cpu = find_idlest_cpu(sd, p, cpu, prev_cpu, sd_flag);
8074 } else if (wake_flags & WF_TTWU) { /* XXX always ? */
8076 new_cpu = select_idle_sibling(p, prev_cpu, new_cpu);
8084 * Called immediately before a task is migrated to a new CPU; task_cpu(p) and
8085 * cfs_rq_of(p) references at time of call are still valid and identify the
8086 * previous CPU. The caller guarantees p->pi_lock or task_rq(p)->lock is held.
8088 static void migrate_task_rq_fair(struct task_struct *p, int new_cpu)
8090 struct sched_entity *se = &p->se;
8092 if (!task_on_rq_migrating(p)) {
8093 remove_entity_load_avg(se);
8096 * Here, the task's PELT values have been updated according to
8097 * the current rq's clock. But if that clock hasn't been
8098 * updated in a while, a substantial idle time will be missed,
8099 * leading to an inflation after wake-up on the new rq.
8101 * Estimate the missing time from the cfs_rq last_update_time
8102 * and update sched_avg to improve the PELT continuity after
8105 migrate_se_pelt_lag(se);
8108 /* Tell new CPU we are migrated */
8109 se->avg.last_update_time = 0;
8111 update_scan_period(p, new_cpu);
8114 static void task_dead_fair(struct task_struct *p)
8116 remove_entity_load_avg(&p->se);
8120 balance_fair(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
8125 return newidle_balance(rq, rf) != 0;
8127 #endif /* CONFIG_SMP */
8129 static void set_next_buddy(struct sched_entity *se)
8131 for_each_sched_entity(se) {
8132 if (SCHED_WARN_ON(!se->on_rq))
8136 cfs_rq_of(se)->next = se;
8141 * Preempt the current task with a newly woken task if needed:
8143 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
8145 struct task_struct *curr = rq->curr;
8146 struct sched_entity *se = &curr->se, *pse = &p->se;
8147 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
8148 int next_buddy_marked = 0;
8149 int cse_is_idle, pse_is_idle;
8151 if (unlikely(se == pse))
8155 * This is possible from callers such as attach_tasks(), in which we
8156 * unconditionally check_preempt_curr() after an enqueue (which may have
8157 * lead to a throttle). This both saves work and prevents false
8158 * next-buddy nomination below.
8160 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
8163 if (sched_feat(NEXT_BUDDY) && !(wake_flags & WF_FORK)) {
8164 set_next_buddy(pse);
8165 next_buddy_marked = 1;
8169 * We can come here with TIF_NEED_RESCHED already set from new task
8172 * Note: this also catches the edge-case of curr being in a throttled
8173 * group (e.g. via set_curr_task), since update_curr() (in the
8174 * enqueue of curr) will have resulted in resched being set. This
8175 * prevents us from potentially nominating it as a false LAST_BUDDY
8178 if (test_tsk_need_resched(curr))
8181 /* Idle tasks are by definition preempted by non-idle tasks. */
8182 if (unlikely(task_has_idle_policy(curr)) &&
8183 likely(!task_has_idle_policy(p)))
8187 * Batch and idle tasks do not preempt non-idle tasks (their preemption
8188 * is driven by the tick):
8190 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
8193 find_matching_se(&se, &pse);
8196 cse_is_idle = se_is_idle(se);
8197 pse_is_idle = se_is_idle(pse);
8200 * Preempt an idle group in favor of a non-idle group (and don't preempt
8201 * in the inverse case).
8203 if (cse_is_idle && !pse_is_idle)
8205 if (cse_is_idle != pse_is_idle)
8208 cfs_rq = cfs_rq_of(se);
8209 update_curr(cfs_rq);
8212 * XXX pick_eevdf(cfs_rq) != se ?
8214 if (pick_eevdf(cfs_rq) == pse)
8224 static struct task_struct *pick_task_fair(struct rq *rq)
8226 struct sched_entity *se;
8227 struct cfs_rq *cfs_rq;
8231 if (!cfs_rq->nr_running)
8235 struct sched_entity *curr = cfs_rq->curr;
8237 /* When we pick for a remote RQ, we'll not have done put_prev_entity() */
8240 update_curr(cfs_rq);
8244 if (unlikely(check_cfs_rq_runtime(cfs_rq)))
8248 se = pick_next_entity(cfs_rq, curr);
8249 cfs_rq = group_cfs_rq(se);
8256 struct task_struct *
8257 pick_next_task_fair(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
8259 struct cfs_rq *cfs_rq = &rq->cfs;
8260 struct sched_entity *se;
8261 struct task_struct *p;
8265 if (!sched_fair_runnable(rq))
8268 #ifdef CONFIG_FAIR_GROUP_SCHED
8269 if (!prev || prev->sched_class != &fair_sched_class)
8273 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
8274 * likely that a next task is from the same cgroup as the current.
8276 * Therefore attempt to avoid putting and setting the entire cgroup
8277 * hierarchy, only change the part that actually changes.
8281 struct sched_entity *curr = cfs_rq->curr;
8284 * Since we got here without doing put_prev_entity() we also
8285 * have to consider cfs_rq->curr. If it is still a runnable
8286 * entity, update_curr() will update its vruntime, otherwise
8287 * forget we've ever seen it.
8291 update_curr(cfs_rq);
8296 * This call to check_cfs_rq_runtime() will do the
8297 * throttle and dequeue its entity in the parent(s).
8298 * Therefore the nr_running test will indeed
8301 if (unlikely(check_cfs_rq_runtime(cfs_rq))) {
8304 if (!cfs_rq->nr_running)
8311 se = pick_next_entity(cfs_rq, curr);
8312 cfs_rq = group_cfs_rq(se);
8318 * Since we haven't yet done put_prev_entity and if the selected task
8319 * is a different task than we started out with, try and touch the
8320 * least amount of cfs_rqs.
8323 struct sched_entity *pse = &prev->se;
8325 while (!(cfs_rq = is_same_group(se, pse))) {
8326 int se_depth = se->depth;
8327 int pse_depth = pse->depth;
8329 if (se_depth <= pse_depth) {
8330 put_prev_entity(cfs_rq_of(pse), pse);
8331 pse = parent_entity(pse);
8333 if (se_depth >= pse_depth) {
8334 set_next_entity(cfs_rq_of(se), se);
8335 se = parent_entity(se);
8339 put_prev_entity(cfs_rq, pse);
8340 set_next_entity(cfs_rq, se);
8347 put_prev_task(rq, prev);
8350 se = pick_next_entity(cfs_rq, NULL);
8351 set_next_entity(cfs_rq, se);
8352 cfs_rq = group_cfs_rq(se);
8357 done: __maybe_unused;
8360 * Move the next running task to the front of
8361 * the list, so our cfs_tasks list becomes MRU
8364 list_move(&p->se.group_node, &rq->cfs_tasks);
8367 if (hrtick_enabled_fair(rq))
8368 hrtick_start_fair(rq, p);
8370 update_misfit_status(p, rq);
8371 sched_fair_update_stop_tick(rq, p);
8379 new_tasks = newidle_balance(rq, rf);
8382 * Because newidle_balance() releases (and re-acquires) rq->lock, it is
8383 * possible for any higher priority task to appear. In that case we
8384 * must re-start the pick_next_entity() loop.
8393 * rq is about to be idle, check if we need to update the
8394 * lost_idle_time of clock_pelt
8396 update_idle_rq_clock_pelt(rq);
8401 static struct task_struct *__pick_next_task_fair(struct rq *rq)
8403 return pick_next_task_fair(rq, NULL, NULL);
8407 * Account for a descheduled task:
8409 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
8411 struct sched_entity *se = &prev->se;
8412 struct cfs_rq *cfs_rq;
8414 for_each_sched_entity(se) {
8415 cfs_rq = cfs_rq_of(se);
8416 put_prev_entity(cfs_rq, se);
8421 * sched_yield() is very simple
8423 static void yield_task_fair(struct rq *rq)
8425 struct task_struct *curr = rq->curr;
8426 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
8427 struct sched_entity *se = &curr->se;
8430 * Are we the only task in the tree?
8432 if (unlikely(rq->nr_running == 1))
8435 clear_buddies(cfs_rq, se);
8437 update_rq_clock(rq);
8439 * Update run-time statistics of the 'current'.
8441 update_curr(cfs_rq);
8443 * Tell update_rq_clock() that we've just updated,
8444 * so we don't do microscopic update in schedule()
8445 * and double the fastpath cost.
8447 rq_clock_skip_update(rq);
8449 se->deadline += calc_delta_fair(se->slice, se);
8452 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p)
8454 struct sched_entity *se = &p->se;
8456 /* throttled hierarchies are not runnable */
8457 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
8460 /* Tell the scheduler that we'd really like pse to run next. */
8463 yield_task_fair(rq);
8469 /**************************************************
8470 * Fair scheduling class load-balancing methods.
8474 * The purpose of load-balancing is to achieve the same basic fairness the
8475 * per-CPU scheduler provides, namely provide a proportional amount of compute
8476 * time to each task. This is expressed in the following equation:
8478 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
8480 * Where W_i,n is the n-th weight average for CPU i. The instantaneous weight
8481 * W_i,0 is defined as:
8483 * W_i,0 = \Sum_j w_i,j (2)
8485 * Where w_i,j is the weight of the j-th runnable task on CPU i. This weight
8486 * is derived from the nice value as per sched_prio_to_weight[].
8488 * The weight average is an exponential decay average of the instantaneous
8491 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
8493 * C_i is the compute capacity of CPU i, typically it is the
8494 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
8495 * can also include other factors [XXX].
8497 * To achieve this balance we define a measure of imbalance which follows
8498 * directly from (1):
8500 * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4)
8502 * We them move tasks around to minimize the imbalance. In the continuous
8503 * function space it is obvious this converges, in the discrete case we get
8504 * a few fun cases generally called infeasible weight scenarios.
8507 * - infeasible weights;
8508 * - local vs global optima in the discrete case. ]
8513 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
8514 * for all i,j solution, we create a tree of CPUs that follows the hardware
8515 * topology where each level pairs two lower groups (or better). This results
8516 * in O(log n) layers. Furthermore we reduce the number of CPUs going up the
8517 * tree to only the first of the previous level and we decrease the frequency
8518 * of load-balance at each level inv. proportional to the number of CPUs in
8524 * \Sum { --- * --- * 2^i } = O(n) (5)
8526 * `- size of each group
8527 * | | `- number of CPUs doing load-balance
8529 * `- sum over all levels
8531 * Coupled with a limit on how many tasks we can migrate every balance pass,
8532 * this makes (5) the runtime complexity of the balancer.
8534 * An important property here is that each CPU is still (indirectly) connected
8535 * to every other CPU in at most O(log n) steps:
8537 * The adjacency matrix of the resulting graph is given by:
8540 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
8543 * And you'll find that:
8545 * A^(log_2 n)_i,j != 0 for all i,j (7)
8547 * Showing there's indeed a path between every CPU in at most O(log n) steps.
8548 * The task movement gives a factor of O(m), giving a convergence complexity
8551 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
8556 * In order to avoid CPUs going idle while there's still work to do, new idle
8557 * balancing is more aggressive and has the newly idle CPU iterate up the domain
8558 * tree itself instead of relying on other CPUs to bring it work.
8560 * This adds some complexity to both (5) and (8) but it reduces the total idle
8568 * Cgroups make a horror show out of (2), instead of a simple sum we get:
8571 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
8576 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
8578 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on CPU i.
8580 * The big problem is S_k, its a global sum needed to compute a local (W_i)
8583 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
8584 * rewrite all of this once again.]
8587 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
8589 enum fbq_type { regular, remote, all };
8592 * 'group_type' describes the group of CPUs at the moment of load balancing.
8594 * The enum is ordered by pulling priority, with the group with lowest priority
8595 * first so the group_type can simply be compared when selecting the busiest
8596 * group. See update_sd_pick_busiest().
8599 /* The group has spare capacity that can be used to run more tasks. */
8600 group_has_spare = 0,
8602 * The group is fully used and the tasks don't compete for more CPU
8603 * cycles. Nevertheless, some tasks might wait before running.
8607 * One task doesn't fit with CPU's capacity and must be migrated to a
8608 * more powerful CPU.
8612 * Balance SMT group that's fully busy. Can benefit from migration
8613 * a task on SMT with busy sibling to another CPU on idle core.
8617 * SD_ASYM_PACKING only: One local CPU with higher capacity is available,
8618 * and the task should be migrated to it instead of running on the
8623 * The tasks' affinity constraints previously prevented the scheduler
8624 * from balancing the load across the system.
8628 * The CPU is overloaded and can't provide expected CPU cycles to all
8634 enum migration_type {
8641 #define LBF_ALL_PINNED 0x01
8642 #define LBF_NEED_BREAK 0x02
8643 #define LBF_DST_PINNED 0x04
8644 #define LBF_SOME_PINNED 0x08
8645 #define LBF_ACTIVE_LB 0x10
8648 struct sched_domain *sd;
8656 struct cpumask *dst_grpmask;
8658 enum cpu_idle_type idle;
8660 /* The set of CPUs under consideration for load-balancing */
8661 struct cpumask *cpus;
8666 unsigned int loop_break;
8667 unsigned int loop_max;
8669 enum fbq_type fbq_type;
8670 enum migration_type migration_type;
8671 struct list_head tasks;
8675 * Is this task likely cache-hot:
8677 static int task_hot(struct task_struct *p, struct lb_env *env)
8681 lockdep_assert_rq_held(env->src_rq);
8683 if (p->sched_class != &fair_sched_class)
8686 if (unlikely(task_has_idle_policy(p)))
8689 /* SMT siblings share cache */
8690 if (env->sd->flags & SD_SHARE_CPUCAPACITY)
8694 * Buddy candidates are cache hot:
8696 if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running &&
8697 (&p->se == cfs_rq_of(&p->se)->next))
8700 if (sysctl_sched_migration_cost == -1)
8704 * Don't migrate task if the task's cookie does not match
8705 * with the destination CPU's core cookie.
8707 if (!sched_core_cookie_match(cpu_rq(env->dst_cpu), p))
8710 if (sysctl_sched_migration_cost == 0)
8713 delta = rq_clock_task(env->src_rq) - p->se.exec_start;
8715 return delta < (s64)sysctl_sched_migration_cost;
8718 #ifdef CONFIG_NUMA_BALANCING
8720 * Returns 1, if task migration degrades locality
8721 * Returns 0, if task migration improves locality i.e migration preferred.
8722 * Returns -1, if task migration is not affected by locality.
8724 static int migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
8726 struct numa_group *numa_group = rcu_dereference(p->numa_group);
8727 unsigned long src_weight, dst_weight;
8728 int src_nid, dst_nid, dist;
8730 if (!static_branch_likely(&sched_numa_balancing))
8733 if (!p->numa_faults || !(env->sd->flags & SD_NUMA))
8736 src_nid = cpu_to_node(env->src_cpu);
8737 dst_nid = cpu_to_node(env->dst_cpu);
8739 if (src_nid == dst_nid)
8742 /* Migrating away from the preferred node is always bad. */
8743 if (src_nid == p->numa_preferred_nid) {
8744 if (env->src_rq->nr_running > env->src_rq->nr_preferred_running)
8750 /* Encourage migration to the preferred node. */
8751 if (dst_nid == p->numa_preferred_nid)
8754 /* Leaving a core idle is often worse than degrading locality. */
8755 if (env->idle == CPU_IDLE)
8758 dist = node_distance(src_nid, dst_nid);
8760 src_weight = group_weight(p, src_nid, dist);
8761 dst_weight = group_weight(p, dst_nid, dist);
8763 src_weight = task_weight(p, src_nid, dist);
8764 dst_weight = task_weight(p, dst_nid, dist);
8767 return dst_weight < src_weight;
8771 static inline int migrate_degrades_locality(struct task_struct *p,
8779 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
8782 int can_migrate_task(struct task_struct *p, struct lb_env *env)
8786 lockdep_assert_rq_held(env->src_rq);
8789 * We do not migrate tasks that are:
8790 * 1) throttled_lb_pair, or
8791 * 2) cannot be migrated to this CPU due to cpus_ptr, or
8792 * 3) running (obviously), or
8793 * 4) are cache-hot on their current CPU.
8795 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
8798 /* Disregard pcpu kthreads; they are where they need to be. */
8799 if (kthread_is_per_cpu(p))
8802 if (!cpumask_test_cpu(env->dst_cpu, p->cpus_ptr)) {
8805 schedstat_inc(p->stats.nr_failed_migrations_affine);
8807 env->flags |= LBF_SOME_PINNED;
8810 * Remember if this task can be migrated to any other CPU in
8811 * our sched_group. We may want to revisit it if we couldn't
8812 * meet load balance goals by pulling other tasks on src_cpu.
8814 * Avoid computing new_dst_cpu
8816 * - if we have already computed one in current iteration
8817 * - if it's an active balance
8819 if (env->idle == CPU_NEWLY_IDLE ||
8820 env->flags & (LBF_DST_PINNED | LBF_ACTIVE_LB))
8823 /* Prevent to re-select dst_cpu via env's CPUs: */
8824 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
8825 if (cpumask_test_cpu(cpu, p->cpus_ptr)) {
8826 env->flags |= LBF_DST_PINNED;
8827 env->new_dst_cpu = cpu;
8835 /* Record that we found at least one task that could run on dst_cpu */
8836 env->flags &= ~LBF_ALL_PINNED;
8838 if (task_on_cpu(env->src_rq, p)) {
8839 schedstat_inc(p->stats.nr_failed_migrations_running);
8844 * Aggressive migration if:
8846 * 2) destination numa is preferred
8847 * 3) task is cache cold, or
8848 * 4) too many balance attempts have failed.
8850 if (env->flags & LBF_ACTIVE_LB)
8853 tsk_cache_hot = migrate_degrades_locality(p, env);
8854 if (tsk_cache_hot == -1)
8855 tsk_cache_hot = task_hot(p, env);
8857 if (tsk_cache_hot <= 0 ||
8858 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
8859 if (tsk_cache_hot == 1) {
8860 schedstat_inc(env->sd->lb_hot_gained[env->idle]);
8861 schedstat_inc(p->stats.nr_forced_migrations);
8866 schedstat_inc(p->stats.nr_failed_migrations_hot);
8871 * detach_task() -- detach the task for the migration specified in env
8873 static void detach_task(struct task_struct *p, struct lb_env *env)
8875 lockdep_assert_rq_held(env->src_rq);
8877 deactivate_task(env->src_rq, p, DEQUEUE_NOCLOCK);
8878 set_task_cpu(p, env->dst_cpu);
8882 * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
8883 * part of active balancing operations within "domain".
8885 * Returns a task if successful and NULL otherwise.
8887 static struct task_struct *detach_one_task(struct lb_env *env)
8889 struct task_struct *p;
8891 lockdep_assert_rq_held(env->src_rq);
8893 list_for_each_entry_reverse(p,
8894 &env->src_rq->cfs_tasks, se.group_node) {
8895 if (!can_migrate_task(p, env))
8898 detach_task(p, env);
8901 * Right now, this is only the second place where
8902 * lb_gained[env->idle] is updated (other is detach_tasks)
8903 * so we can safely collect stats here rather than
8904 * inside detach_tasks().
8906 schedstat_inc(env->sd->lb_gained[env->idle]);
8913 * detach_tasks() -- tries to detach up to imbalance load/util/tasks from
8914 * busiest_rq, as part of a balancing operation within domain "sd".
8916 * Returns number of detached tasks if successful and 0 otherwise.
8918 static int detach_tasks(struct lb_env *env)
8920 struct list_head *tasks = &env->src_rq->cfs_tasks;
8921 unsigned long util, load;
8922 struct task_struct *p;
8925 lockdep_assert_rq_held(env->src_rq);
8928 * Source run queue has been emptied by another CPU, clear
8929 * LBF_ALL_PINNED flag as we will not test any task.
8931 if (env->src_rq->nr_running <= 1) {
8932 env->flags &= ~LBF_ALL_PINNED;
8936 if (env->imbalance <= 0)
8939 while (!list_empty(tasks)) {
8941 * We don't want to steal all, otherwise we may be treated likewise,
8942 * which could at worst lead to a livelock crash.
8944 if (env->idle != CPU_NOT_IDLE && env->src_rq->nr_running <= 1)
8949 * We've more or less seen every task there is, call it quits
8950 * unless we haven't found any movable task yet.
8952 if (env->loop > env->loop_max &&
8953 !(env->flags & LBF_ALL_PINNED))
8956 /* take a breather every nr_migrate tasks */
8957 if (env->loop > env->loop_break) {
8958 env->loop_break += SCHED_NR_MIGRATE_BREAK;
8959 env->flags |= LBF_NEED_BREAK;
8963 p = list_last_entry(tasks, struct task_struct, se.group_node);
8965 if (!can_migrate_task(p, env))
8968 switch (env->migration_type) {
8971 * Depending of the number of CPUs and tasks and the
8972 * cgroup hierarchy, task_h_load() can return a null
8973 * value. Make sure that env->imbalance decreases
8974 * otherwise detach_tasks() will stop only after
8975 * detaching up to loop_max tasks.
8977 load = max_t(unsigned long, task_h_load(p), 1);
8979 if (sched_feat(LB_MIN) &&
8980 load < 16 && !env->sd->nr_balance_failed)
8984 * Make sure that we don't migrate too much load.
8985 * Nevertheless, let relax the constraint if
8986 * scheduler fails to find a good waiting task to
8989 if (shr_bound(load, env->sd->nr_balance_failed) > env->imbalance)
8992 env->imbalance -= load;
8996 util = task_util_est(p);
8998 if (util > env->imbalance)
9001 env->imbalance -= util;
9008 case migrate_misfit:
9009 /* This is not a misfit task */
9010 if (task_fits_cpu(p, env->src_cpu))
9017 detach_task(p, env);
9018 list_add(&p->se.group_node, &env->tasks);
9022 #ifdef CONFIG_PREEMPTION
9024 * NEWIDLE balancing is a source of latency, so preemptible
9025 * kernels will stop after the first task is detached to minimize
9026 * the critical section.
9028 if (env->idle == CPU_NEWLY_IDLE)
9033 * We only want to steal up to the prescribed amount of
9036 if (env->imbalance <= 0)
9041 list_move(&p->se.group_node, tasks);
9045 * Right now, this is one of only two places we collect this stat
9046 * so we can safely collect detach_one_task() stats here rather
9047 * than inside detach_one_task().
9049 schedstat_add(env->sd->lb_gained[env->idle], detached);
9055 * attach_task() -- attach the task detached by detach_task() to its new rq.
9057 static void attach_task(struct rq *rq, struct task_struct *p)
9059 lockdep_assert_rq_held(rq);
9061 WARN_ON_ONCE(task_rq(p) != rq);
9062 activate_task(rq, p, ENQUEUE_NOCLOCK);
9063 check_preempt_curr(rq, p, 0);
9067 * attach_one_task() -- attaches the task returned from detach_one_task() to
9070 static void attach_one_task(struct rq *rq, struct task_struct *p)
9075 update_rq_clock(rq);
9081 * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
9084 static void attach_tasks(struct lb_env *env)
9086 struct list_head *tasks = &env->tasks;
9087 struct task_struct *p;
9090 rq_lock(env->dst_rq, &rf);
9091 update_rq_clock(env->dst_rq);
9093 while (!list_empty(tasks)) {
9094 p = list_first_entry(tasks, struct task_struct, se.group_node);
9095 list_del_init(&p->se.group_node);
9097 attach_task(env->dst_rq, p);
9100 rq_unlock(env->dst_rq, &rf);
9103 #ifdef CONFIG_NO_HZ_COMMON
9104 static inline bool cfs_rq_has_blocked(struct cfs_rq *cfs_rq)
9106 if (cfs_rq->avg.load_avg)
9109 if (cfs_rq->avg.util_avg)
9115 static inline bool others_have_blocked(struct rq *rq)
9117 if (READ_ONCE(rq->avg_rt.util_avg))
9120 if (READ_ONCE(rq->avg_dl.util_avg))
9123 if (thermal_load_avg(rq))
9126 #ifdef CONFIG_HAVE_SCHED_AVG_IRQ
9127 if (READ_ONCE(rq->avg_irq.util_avg))
9134 static inline void update_blocked_load_tick(struct rq *rq)
9136 WRITE_ONCE(rq->last_blocked_load_update_tick, jiffies);
9139 static inline void update_blocked_load_status(struct rq *rq, bool has_blocked)
9142 rq->has_blocked_load = 0;
9145 static inline bool cfs_rq_has_blocked(struct cfs_rq *cfs_rq) { return false; }
9146 static inline bool others_have_blocked(struct rq *rq) { return false; }
9147 static inline void update_blocked_load_tick(struct rq *rq) {}
9148 static inline void update_blocked_load_status(struct rq *rq, bool has_blocked) {}
9151 static bool __update_blocked_others(struct rq *rq, bool *done)
9153 const struct sched_class *curr_class;
9154 u64 now = rq_clock_pelt(rq);
9155 unsigned long thermal_pressure;
9159 * update_load_avg() can call cpufreq_update_util(). Make sure that RT,
9160 * DL and IRQ signals have been updated before updating CFS.
9162 curr_class = rq->curr->sched_class;
9164 thermal_pressure = arch_scale_thermal_pressure(cpu_of(rq));
9166 decayed = update_rt_rq_load_avg(now, rq, curr_class == &rt_sched_class) |
9167 update_dl_rq_load_avg(now, rq, curr_class == &dl_sched_class) |
9168 update_thermal_load_avg(rq_clock_thermal(rq), rq, thermal_pressure) |
9169 update_irq_load_avg(rq, 0);
9171 if (others_have_blocked(rq))
9177 #ifdef CONFIG_FAIR_GROUP_SCHED
9179 static bool __update_blocked_fair(struct rq *rq, bool *done)
9181 struct cfs_rq *cfs_rq, *pos;
9182 bool decayed = false;
9183 int cpu = cpu_of(rq);
9186 * Iterates the task_group tree in a bottom up fashion, see
9187 * list_add_leaf_cfs_rq() for details.
9189 for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) {
9190 struct sched_entity *se;
9192 if (update_cfs_rq_load_avg(cfs_rq_clock_pelt(cfs_rq), cfs_rq)) {
9193 update_tg_load_avg(cfs_rq);
9195 if (cfs_rq->nr_running == 0)
9196 update_idle_cfs_rq_clock_pelt(cfs_rq);
9198 if (cfs_rq == &rq->cfs)
9202 /* Propagate pending load changes to the parent, if any: */
9203 se = cfs_rq->tg->se[cpu];
9204 if (se && !skip_blocked_update(se))
9205 update_load_avg(cfs_rq_of(se), se, UPDATE_TG);
9208 * There can be a lot of idle CPU cgroups. Don't let fully
9209 * decayed cfs_rqs linger on the list.
9211 if (cfs_rq_is_decayed(cfs_rq))
9212 list_del_leaf_cfs_rq(cfs_rq);
9214 /* Don't need periodic decay once load/util_avg are null */
9215 if (cfs_rq_has_blocked(cfs_rq))
9223 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
9224 * This needs to be done in a top-down fashion because the load of a child
9225 * group is a fraction of its parents load.
9227 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
9229 struct rq *rq = rq_of(cfs_rq);
9230 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
9231 unsigned long now = jiffies;
9234 if (cfs_rq->last_h_load_update == now)
9237 WRITE_ONCE(cfs_rq->h_load_next, NULL);
9238 for_each_sched_entity(se) {
9239 cfs_rq = cfs_rq_of(se);
9240 WRITE_ONCE(cfs_rq->h_load_next, se);
9241 if (cfs_rq->last_h_load_update == now)
9246 cfs_rq->h_load = cfs_rq_load_avg(cfs_rq);
9247 cfs_rq->last_h_load_update = now;
9250 while ((se = READ_ONCE(cfs_rq->h_load_next)) != NULL) {
9251 load = cfs_rq->h_load;
9252 load = div64_ul(load * se->avg.load_avg,
9253 cfs_rq_load_avg(cfs_rq) + 1);
9254 cfs_rq = group_cfs_rq(se);
9255 cfs_rq->h_load = load;
9256 cfs_rq->last_h_load_update = now;
9260 static unsigned long task_h_load(struct task_struct *p)
9262 struct cfs_rq *cfs_rq = task_cfs_rq(p);
9264 update_cfs_rq_h_load(cfs_rq);
9265 return div64_ul(p->se.avg.load_avg * cfs_rq->h_load,
9266 cfs_rq_load_avg(cfs_rq) + 1);
9269 static bool __update_blocked_fair(struct rq *rq, bool *done)
9271 struct cfs_rq *cfs_rq = &rq->cfs;
9274 decayed = update_cfs_rq_load_avg(cfs_rq_clock_pelt(cfs_rq), cfs_rq);
9275 if (cfs_rq_has_blocked(cfs_rq))
9281 static unsigned long task_h_load(struct task_struct *p)
9283 return p->se.avg.load_avg;
9287 static void update_blocked_averages(int cpu)
9289 bool decayed = false, done = true;
9290 struct rq *rq = cpu_rq(cpu);
9293 rq_lock_irqsave(rq, &rf);
9294 update_blocked_load_tick(rq);
9295 update_rq_clock(rq);
9297 decayed |= __update_blocked_others(rq, &done);
9298 decayed |= __update_blocked_fair(rq, &done);
9300 update_blocked_load_status(rq, !done);
9302 cpufreq_update_util(rq, 0);
9303 rq_unlock_irqrestore(rq, &rf);
9306 /********** Helpers for find_busiest_group ************************/
9309 * sg_lb_stats - stats of a sched_group required for load_balancing
9311 struct sg_lb_stats {
9312 unsigned long avg_load; /*Avg load across the CPUs of the group */
9313 unsigned long group_load; /* Total load over the CPUs of the group */
9314 unsigned long group_capacity;
9315 unsigned long group_util; /* Total utilization over the CPUs of the group */
9316 unsigned long group_runnable; /* Total runnable time over the CPUs of the group */
9317 unsigned int sum_nr_running; /* Nr of tasks running in the group */
9318 unsigned int sum_h_nr_running; /* Nr of CFS tasks running in the group */
9319 unsigned int idle_cpus;
9320 unsigned int group_weight;
9321 enum group_type group_type;
9322 unsigned int group_asym_packing; /* Tasks should be moved to preferred CPU */
9323 unsigned int group_smt_balance; /* Task on busy SMT be moved */
9324 unsigned long group_misfit_task_load; /* A CPU has a task too big for its capacity */
9325 #ifdef CONFIG_NUMA_BALANCING
9326 unsigned int nr_numa_running;
9327 unsigned int nr_preferred_running;
9332 * sd_lb_stats - Structure to store the statistics of a sched_domain
9333 * during load balancing.
9335 struct sd_lb_stats {
9336 struct sched_group *busiest; /* Busiest group in this sd */
9337 struct sched_group *local; /* Local group in this sd */
9338 unsigned long total_load; /* Total load of all groups in sd */
9339 unsigned long total_capacity; /* Total capacity of all groups in sd */
9340 unsigned long avg_load; /* Average load across all groups in sd */
9341 unsigned int prefer_sibling; /* tasks should go to sibling first */
9343 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
9344 struct sg_lb_stats local_stat; /* Statistics of the local group */
9347 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
9350 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
9351 * local_stat because update_sg_lb_stats() does a full clear/assignment.
9352 * We must however set busiest_stat::group_type and
9353 * busiest_stat::idle_cpus to the worst busiest group because
9354 * update_sd_pick_busiest() reads these before assignment.
9356 *sds = (struct sd_lb_stats){
9360 .total_capacity = 0UL,
9362 .idle_cpus = UINT_MAX,
9363 .group_type = group_has_spare,
9368 static unsigned long scale_rt_capacity(int cpu)
9370 struct rq *rq = cpu_rq(cpu);
9371 unsigned long max = arch_scale_cpu_capacity(cpu);
9372 unsigned long used, free;
9375 irq = cpu_util_irq(rq);
9377 if (unlikely(irq >= max))
9381 * avg_rt.util_avg and avg_dl.util_avg track binary signals
9382 * (running and not running) with weights 0 and 1024 respectively.
9383 * avg_thermal.load_avg tracks thermal pressure and the weighted
9384 * average uses the actual delta max capacity(load).
9386 used = READ_ONCE(rq->avg_rt.util_avg);
9387 used += READ_ONCE(rq->avg_dl.util_avg);
9388 used += thermal_load_avg(rq);
9390 if (unlikely(used >= max))
9395 return scale_irq_capacity(free, irq, max);
9398 static void update_cpu_capacity(struct sched_domain *sd, int cpu)
9400 unsigned long capacity = scale_rt_capacity(cpu);
9401 struct sched_group *sdg = sd->groups;
9403 cpu_rq(cpu)->cpu_capacity_orig = arch_scale_cpu_capacity(cpu);
9408 cpu_rq(cpu)->cpu_capacity = capacity;
9409 trace_sched_cpu_capacity_tp(cpu_rq(cpu));
9411 sdg->sgc->capacity = capacity;
9412 sdg->sgc->min_capacity = capacity;
9413 sdg->sgc->max_capacity = capacity;
9416 void update_group_capacity(struct sched_domain *sd, int cpu)
9418 struct sched_domain *child = sd->child;
9419 struct sched_group *group, *sdg = sd->groups;
9420 unsigned long capacity, min_capacity, max_capacity;
9421 unsigned long interval;
9423 interval = msecs_to_jiffies(sd->balance_interval);
9424 interval = clamp(interval, 1UL, max_load_balance_interval);
9425 sdg->sgc->next_update = jiffies + interval;
9428 update_cpu_capacity(sd, cpu);
9433 min_capacity = ULONG_MAX;
9436 if (child->flags & SD_OVERLAP) {
9438 * SD_OVERLAP domains cannot assume that child groups
9439 * span the current group.
9442 for_each_cpu(cpu, sched_group_span(sdg)) {
9443 unsigned long cpu_cap = capacity_of(cpu);
9445 capacity += cpu_cap;
9446 min_capacity = min(cpu_cap, min_capacity);
9447 max_capacity = max(cpu_cap, max_capacity);
9451 * !SD_OVERLAP domains can assume that child groups
9452 * span the current group.
9455 group = child->groups;
9457 struct sched_group_capacity *sgc = group->sgc;
9459 capacity += sgc->capacity;
9460 min_capacity = min(sgc->min_capacity, min_capacity);
9461 max_capacity = max(sgc->max_capacity, max_capacity);
9462 group = group->next;
9463 } while (group != child->groups);
9466 sdg->sgc->capacity = capacity;
9467 sdg->sgc->min_capacity = min_capacity;
9468 sdg->sgc->max_capacity = max_capacity;
9472 * Check whether the capacity of the rq has been noticeably reduced by side
9473 * activity. The imbalance_pct is used for the threshold.
9474 * Return true is the capacity is reduced
9477 check_cpu_capacity(struct rq *rq, struct sched_domain *sd)
9479 return ((rq->cpu_capacity * sd->imbalance_pct) <
9480 (rq->cpu_capacity_orig * 100));
9484 * Check whether a rq has a misfit task and if it looks like we can actually
9485 * help that task: we can migrate the task to a CPU of higher capacity, or
9486 * the task's current CPU is heavily pressured.
9488 static inline int check_misfit_status(struct rq *rq, struct sched_domain *sd)
9490 return rq->misfit_task_load &&
9491 (rq->cpu_capacity_orig < rq->rd->max_cpu_capacity ||
9492 check_cpu_capacity(rq, sd));
9496 * Group imbalance indicates (and tries to solve) the problem where balancing
9497 * groups is inadequate due to ->cpus_ptr constraints.
9499 * Imagine a situation of two groups of 4 CPUs each and 4 tasks each with a
9500 * cpumask covering 1 CPU of the first group and 3 CPUs of the second group.
9503 * { 0 1 2 3 } { 4 5 6 7 }
9506 * If we were to balance group-wise we'd place two tasks in the first group and
9507 * two tasks in the second group. Clearly this is undesired as it will overload
9508 * cpu 3 and leave one of the CPUs in the second group unused.
9510 * The current solution to this issue is detecting the skew in the first group
9511 * by noticing the lower domain failed to reach balance and had difficulty
9512 * moving tasks due to affinity constraints.
9514 * When this is so detected; this group becomes a candidate for busiest; see
9515 * update_sd_pick_busiest(). And calculate_imbalance() and
9516 * find_busiest_group() avoid some of the usual balance conditions to allow it
9517 * to create an effective group imbalance.
9519 * This is a somewhat tricky proposition since the next run might not find the
9520 * group imbalance and decide the groups need to be balanced again. A most
9521 * subtle and fragile situation.
9524 static inline int sg_imbalanced(struct sched_group *group)
9526 return group->sgc->imbalance;
9530 * group_has_capacity returns true if the group has spare capacity that could
9531 * be used by some tasks.
9532 * We consider that a group has spare capacity if the number of task is
9533 * smaller than the number of CPUs or if the utilization is lower than the
9534 * available capacity for CFS tasks.
9535 * For the latter, we use a threshold to stabilize the state, to take into
9536 * account the variance of the tasks' load and to return true if the available
9537 * capacity in meaningful for the load balancer.
9538 * As an example, an available capacity of 1% can appear but it doesn't make
9539 * any benefit for the load balance.
9542 group_has_capacity(unsigned int imbalance_pct, struct sg_lb_stats *sgs)
9544 if (sgs->sum_nr_running < sgs->group_weight)
9547 if ((sgs->group_capacity * imbalance_pct) <
9548 (sgs->group_runnable * 100))
9551 if ((sgs->group_capacity * 100) >
9552 (sgs->group_util * imbalance_pct))
9559 * group_is_overloaded returns true if the group has more tasks than it can
9561 * group_is_overloaded is not equals to !group_has_capacity because a group
9562 * with the exact right number of tasks, has no more spare capacity but is not
9563 * overloaded so both group_has_capacity and group_is_overloaded return
9567 group_is_overloaded(unsigned int imbalance_pct, struct sg_lb_stats *sgs)
9569 if (sgs->sum_nr_running <= sgs->group_weight)
9572 if ((sgs->group_capacity * 100) <
9573 (sgs->group_util * imbalance_pct))
9576 if ((sgs->group_capacity * imbalance_pct) <
9577 (sgs->group_runnable * 100))
9584 group_type group_classify(unsigned int imbalance_pct,
9585 struct sched_group *group,
9586 struct sg_lb_stats *sgs)
9588 if (group_is_overloaded(imbalance_pct, sgs))
9589 return group_overloaded;
9591 if (sg_imbalanced(group))
9592 return group_imbalanced;
9594 if (sgs->group_asym_packing)
9595 return group_asym_packing;
9597 if (sgs->group_smt_balance)
9598 return group_smt_balance;
9600 if (sgs->group_misfit_task_load)
9601 return group_misfit_task;
9603 if (!group_has_capacity(imbalance_pct, sgs))
9604 return group_fully_busy;
9606 return group_has_spare;
9610 * sched_use_asym_prio - Check whether asym_packing priority must be used
9611 * @sd: The scheduling domain of the load balancing
9614 * Always use CPU priority when balancing load between SMT siblings. When
9615 * balancing load between cores, it is not sufficient that @cpu is idle. Only
9616 * use CPU priority if the whole core is idle.
9618 * Returns: True if the priority of @cpu must be followed. False otherwise.
9620 static bool sched_use_asym_prio(struct sched_domain *sd, int cpu)
9622 if (!sched_smt_active())
9625 return sd->flags & SD_SHARE_CPUCAPACITY || is_core_idle(cpu);
9629 * sched_asym - Check if the destination CPU can do asym_packing load balance
9630 * @env: The load balancing environment
9631 * @sds: Load-balancing data with statistics of the local group
9632 * @sgs: Load-balancing statistics of the candidate busiest group
9633 * @group: The candidate busiest group
9635 * @env::dst_cpu can do asym_packing if it has higher priority than the
9636 * preferred CPU of @group.
9638 * SMT is a special case. If we are balancing load between cores, @env::dst_cpu
9639 * can do asym_packing balance only if all its SMT siblings are idle. Also, it
9640 * can only do it if @group is an SMT group and has exactly on busy CPU. Larger
9641 * imbalances in the number of CPUS are dealt with in find_busiest_group().
9643 * If we are balancing load within an SMT core, or at DIE domain level, always
9646 * Return: true if @env::dst_cpu can do with asym_packing load balance. False
9650 sched_asym(struct lb_env *env, struct sd_lb_stats *sds, struct sg_lb_stats *sgs,
9651 struct sched_group *group)
9653 /* Ensure that the whole local core is idle, if applicable. */
9654 if (!sched_use_asym_prio(env->sd, env->dst_cpu))
9658 * CPU priorities does not make sense for SMT cores with more than one
9661 if (group->flags & SD_SHARE_CPUCAPACITY) {
9662 if (sgs->group_weight - sgs->idle_cpus != 1)
9666 return sched_asym_prefer(env->dst_cpu, group->asym_prefer_cpu);
9669 /* One group has more than one SMT CPU while the other group does not */
9670 static inline bool smt_vs_nonsmt_groups(struct sched_group *sg1,
9671 struct sched_group *sg2)
9676 return (sg1->flags & SD_SHARE_CPUCAPACITY) !=
9677 (sg2->flags & SD_SHARE_CPUCAPACITY);
9680 static inline bool smt_balance(struct lb_env *env, struct sg_lb_stats *sgs,
9681 struct sched_group *group)
9683 if (env->idle == CPU_NOT_IDLE)
9687 * For SMT source group, it is better to move a task
9688 * to a CPU that doesn't have multiple tasks sharing its CPU capacity.
9689 * Note that if a group has a single SMT, SD_SHARE_CPUCAPACITY
9692 if (group->flags & SD_SHARE_CPUCAPACITY &&
9693 sgs->sum_h_nr_running > 1)
9699 static inline long sibling_imbalance(struct lb_env *env,
9700 struct sd_lb_stats *sds,
9701 struct sg_lb_stats *busiest,
9702 struct sg_lb_stats *local)
9704 int ncores_busiest, ncores_local;
9707 if (env->idle == CPU_NOT_IDLE || !busiest->sum_nr_running)
9710 ncores_busiest = sds->busiest->cores;
9711 ncores_local = sds->local->cores;
9713 if (ncores_busiest == ncores_local) {
9714 imbalance = busiest->sum_nr_running;
9715 lsub_positive(&imbalance, local->sum_nr_running);
9719 /* Balance such that nr_running/ncores ratio are same on both groups */
9720 imbalance = ncores_local * busiest->sum_nr_running;
9721 lsub_positive(&imbalance, ncores_busiest * local->sum_nr_running);
9722 /* Normalize imbalance and do rounding on normalization */
9723 imbalance = 2 * imbalance + ncores_local + ncores_busiest;
9724 imbalance /= ncores_local + ncores_busiest;
9726 /* Take advantage of resource in an empty sched group */
9727 if (imbalance <= 1 && local->sum_nr_running == 0 &&
9728 busiest->sum_nr_running > 1)
9735 sched_reduced_capacity(struct rq *rq, struct sched_domain *sd)
9738 * When there is more than 1 task, the group_overloaded case already
9739 * takes care of cpu with reduced capacity
9741 if (rq->cfs.h_nr_running != 1)
9744 return check_cpu_capacity(rq, sd);
9748 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
9749 * @env: The load balancing environment.
9750 * @sds: Load-balancing data with statistics of the local group.
9751 * @group: sched_group whose statistics are to be updated.
9752 * @sgs: variable to hold the statistics for this group.
9753 * @sg_status: Holds flag indicating the status of the sched_group
9755 static inline void update_sg_lb_stats(struct lb_env *env,
9756 struct sd_lb_stats *sds,
9757 struct sched_group *group,
9758 struct sg_lb_stats *sgs,
9761 int i, nr_running, local_group;
9763 memset(sgs, 0, sizeof(*sgs));
9765 local_group = group == sds->local;
9767 for_each_cpu_and(i, sched_group_span(group), env->cpus) {
9768 struct rq *rq = cpu_rq(i);
9769 unsigned long load = cpu_load(rq);
9771 sgs->group_load += load;
9772 sgs->group_util += cpu_util_cfs(i);
9773 sgs->group_runnable += cpu_runnable(rq);
9774 sgs->sum_h_nr_running += rq->cfs.h_nr_running;
9776 nr_running = rq->nr_running;
9777 sgs->sum_nr_running += nr_running;
9780 *sg_status |= SG_OVERLOAD;
9782 if (cpu_overutilized(i))
9783 *sg_status |= SG_OVERUTILIZED;
9785 #ifdef CONFIG_NUMA_BALANCING
9786 sgs->nr_numa_running += rq->nr_numa_running;
9787 sgs->nr_preferred_running += rq->nr_preferred_running;
9790 * No need to call idle_cpu() if nr_running is not 0
9792 if (!nr_running && idle_cpu(i)) {
9794 /* Idle cpu can't have misfit task */
9801 if (env->sd->flags & SD_ASYM_CPUCAPACITY) {
9802 /* Check for a misfit task on the cpu */
9803 if (sgs->group_misfit_task_load < rq->misfit_task_load) {
9804 sgs->group_misfit_task_load = rq->misfit_task_load;
9805 *sg_status |= SG_OVERLOAD;
9807 } else if ((env->idle != CPU_NOT_IDLE) &&
9808 sched_reduced_capacity(rq, env->sd)) {
9809 /* Check for a task running on a CPU with reduced capacity */
9810 if (sgs->group_misfit_task_load < load)
9811 sgs->group_misfit_task_load = load;
9815 sgs->group_capacity = group->sgc->capacity;
9817 sgs->group_weight = group->group_weight;
9819 /* Check if dst CPU is idle and preferred to this group */
9820 if (!local_group && env->sd->flags & SD_ASYM_PACKING &&
9821 env->idle != CPU_NOT_IDLE && sgs->sum_h_nr_running &&
9822 sched_asym(env, sds, sgs, group)) {
9823 sgs->group_asym_packing = 1;
9826 /* Check for loaded SMT group to be balanced to dst CPU */
9827 if (!local_group && smt_balance(env, sgs, group))
9828 sgs->group_smt_balance = 1;
9830 sgs->group_type = group_classify(env->sd->imbalance_pct, group, sgs);
9832 /* Computing avg_load makes sense only when group is overloaded */
9833 if (sgs->group_type == group_overloaded)
9834 sgs->avg_load = (sgs->group_load * SCHED_CAPACITY_SCALE) /
9835 sgs->group_capacity;
9839 * update_sd_pick_busiest - return 1 on busiest group
9840 * @env: The load balancing environment.
9841 * @sds: sched_domain statistics
9842 * @sg: sched_group candidate to be checked for being the busiest
9843 * @sgs: sched_group statistics
9845 * Determine if @sg is a busier group than the previously selected
9848 * Return: %true if @sg is a busier group than the previously selected
9849 * busiest group. %false otherwise.
9851 static bool update_sd_pick_busiest(struct lb_env *env,
9852 struct sd_lb_stats *sds,
9853 struct sched_group *sg,
9854 struct sg_lb_stats *sgs)
9856 struct sg_lb_stats *busiest = &sds->busiest_stat;
9858 /* Make sure that there is at least one task to pull */
9859 if (!sgs->sum_h_nr_running)
9863 * Don't try to pull misfit tasks we can't help.
9864 * We can use max_capacity here as reduction in capacity on some
9865 * CPUs in the group should either be possible to resolve
9866 * internally or be covered by avg_load imbalance (eventually).
9868 if ((env->sd->flags & SD_ASYM_CPUCAPACITY) &&
9869 (sgs->group_type == group_misfit_task) &&
9870 (!capacity_greater(capacity_of(env->dst_cpu), sg->sgc->max_capacity) ||
9871 sds->local_stat.group_type != group_has_spare))
9874 if (sgs->group_type > busiest->group_type)
9877 if (sgs->group_type < busiest->group_type)
9881 * The candidate and the current busiest group are the same type of
9882 * group. Let check which one is the busiest according to the type.
9885 switch (sgs->group_type) {
9886 case group_overloaded:
9887 /* Select the overloaded group with highest avg_load. */
9888 if (sgs->avg_load <= busiest->avg_load)
9892 case group_imbalanced:
9894 * Select the 1st imbalanced group as we don't have any way to
9895 * choose one more than another.
9899 case group_asym_packing:
9900 /* Prefer to move from lowest priority CPU's work */
9901 if (sched_asym_prefer(sg->asym_prefer_cpu, sds->busiest->asym_prefer_cpu))
9905 case group_misfit_task:
9907 * If we have more than one misfit sg go with the biggest
9910 if (sgs->group_misfit_task_load < busiest->group_misfit_task_load)
9914 case group_smt_balance:
9916 * Check if we have spare CPUs on either SMT group to
9917 * choose has spare or fully busy handling.
9919 if (sgs->idle_cpus != 0 || busiest->idle_cpus != 0)
9924 case group_fully_busy:
9926 * Select the fully busy group with highest avg_load. In
9927 * theory, there is no need to pull task from such kind of
9928 * group because tasks have all compute capacity that they need
9929 * but we can still improve the overall throughput by reducing
9930 * contention when accessing shared HW resources.
9932 * XXX for now avg_load is not computed and always 0 so we
9933 * select the 1st one, except if @sg is composed of SMT
9937 if (sgs->avg_load < busiest->avg_load)
9940 if (sgs->avg_load == busiest->avg_load) {
9942 * SMT sched groups need more help than non-SMT groups.
9943 * If @sg happens to also be SMT, either choice is good.
9945 if (sds->busiest->flags & SD_SHARE_CPUCAPACITY)
9951 case group_has_spare:
9953 * Do not pick sg with SMT CPUs over sg with pure CPUs,
9954 * as we do not want to pull task off SMT core with one task
9955 * and make the core idle.
9957 if (smt_vs_nonsmt_groups(sds->busiest, sg)) {
9958 if (sg->flags & SD_SHARE_CPUCAPACITY && sgs->sum_h_nr_running <= 1)
9966 * Select not overloaded group with lowest number of idle cpus
9967 * and highest number of running tasks. We could also compare
9968 * the spare capacity which is more stable but it can end up
9969 * that the group has less spare capacity but finally more idle
9970 * CPUs which means less opportunity to pull tasks.
9972 if (sgs->idle_cpus > busiest->idle_cpus)
9974 else if ((sgs->idle_cpus == busiest->idle_cpus) &&
9975 (sgs->sum_nr_running <= busiest->sum_nr_running))
9982 * Candidate sg has no more than one task per CPU and has higher
9983 * per-CPU capacity. Migrating tasks to less capable CPUs may harm
9984 * throughput. Maximize throughput, power/energy consequences are not
9987 if ((env->sd->flags & SD_ASYM_CPUCAPACITY) &&
9988 (sgs->group_type <= group_fully_busy) &&
9989 (capacity_greater(sg->sgc->min_capacity, capacity_of(env->dst_cpu))))
9995 #ifdef CONFIG_NUMA_BALANCING
9996 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
9998 if (sgs->sum_h_nr_running > sgs->nr_numa_running)
10000 if (sgs->sum_h_nr_running > sgs->nr_preferred_running)
10005 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
10007 if (rq->nr_running > rq->nr_numa_running)
10009 if (rq->nr_running > rq->nr_preferred_running)
10014 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
10019 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
10023 #endif /* CONFIG_NUMA_BALANCING */
10026 struct sg_lb_stats;
10029 * task_running_on_cpu - return 1 if @p is running on @cpu.
10032 static unsigned int task_running_on_cpu(int cpu, struct task_struct *p)
10034 /* Task has no contribution or is new */
10035 if (cpu != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time))
10038 if (task_on_rq_queued(p))
10045 * idle_cpu_without - would a given CPU be idle without p ?
10046 * @cpu: the processor on which idleness is tested.
10047 * @p: task which should be ignored.
10049 * Return: 1 if the CPU would be idle. 0 otherwise.
10051 static int idle_cpu_without(int cpu, struct task_struct *p)
10053 struct rq *rq = cpu_rq(cpu);
10055 if (rq->curr != rq->idle && rq->curr != p)
10059 * rq->nr_running can't be used but an updated version without the
10060 * impact of p on cpu must be used instead. The updated nr_running
10061 * be computed and tested before calling idle_cpu_without().
10065 if (rq->ttwu_pending)
10073 * update_sg_wakeup_stats - Update sched_group's statistics for wakeup.
10074 * @sd: The sched_domain level to look for idlest group.
10075 * @group: sched_group whose statistics are to be updated.
10076 * @sgs: variable to hold the statistics for this group.
10077 * @p: The task for which we look for the idlest group/CPU.
10079 static inline void update_sg_wakeup_stats(struct sched_domain *sd,
10080 struct sched_group *group,
10081 struct sg_lb_stats *sgs,
10082 struct task_struct *p)
10086 memset(sgs, 0, sizeof(*sgs));
10088 /* Assume that task can't fit any CPU of the group */
10089 if (sd->flags & SD_ASYM_CPUCAPACITY)
10090 sgs->group_misfit_task_load = 1;
10092 for_each_cpu(i, sched_group_span(group)) {
10093 struct rq *rq = cpu_rq(i);
10094 unsigned int local;
10096 sgs->group_load += cpu_load_without(rq, p);
10097 sgs->group_util += cpu_util_without(i, p);
10098 sgs->group_runnable += cpu_runnable_without(rq, p);
10099 local = task_running_on_cpu(i, p);
10100 sgs->sum_h_nr_running += rq->cfs.h_nr_running - local;
10102 nr_running = rq->nr_running - local;
10103 sgs->sum_nr_running += nr_running;
10106 * No need to call idle_cpu_without() if nr_running is not 0
10108 if (!nr_running && idle_cpu_without(i, p))
10111 /* Check if task fits in the CPU */
10112 if (sd->flags & SD_ASYM_CPUCAPACITY &&
10113 sgs->group_misfit_task_load &&
10114 task_fits_cpu(p, i))
10115 sgs->group_misfit_task_load = 0;
10119 sgs->group_capacity = group->sgc->capacity;
10121 sgs->group_weight = group->group_weight;
10123 sgs->group_type = group_classify(sd->imbalance_pct, group, sgs);
10126 * Computing avg_load makes sense only when group is fully busy or
10129 if (sgs->group_type == group_fully_busy ||
10130 sgs->group_type == group_overloaded)
10131 sgs->avg_load = (sgs->group_load * SCHED_CAPACITY_SCALE) /
10132 sgs->group_capacity;
10135 static bool update_pick_idlest(struct sched_group *idlest,
10136 struct sg_lb_stats *idlest_sgs,
10137 struct sched_group *group,
10138 struct sg_lb_stats *sgs)
10140 if (sgs->group_type < idlest_sgs->group_type)
10143 if (sgs->group_type > idlest_sgs->group_type)
10147 * The candidate and the current idlest group are the same type of
10148 * group. Let check which one is the idlest according to the type.
10151 switch (sgs->group_type) {
10152 case group_overloaded:
10153 case group_fully_busy:
10154 /* Select the group with lowest avg_load. */
10155 if (idlest_sgs->avg_load <= sgs->avg_load)
10159 case group_imbalanced:
10160 case group_asym_packing:
10161 case group_smt_balance:
10162 /* Those types are not used in the slow wakeup path */
10165 case group_misfit_task:
10166 /* Select group with the highest max capacity */
10167 if (idlest->sgc->max_capacity >= group->sgc->max_capacity)
10171 case group_has_spare:
10172 /* Select group with most idle CPUs */
10173 if (idlest_sgs->idle_cpus > sgs->idle_cpus)
10176 /* Select group with lowest group_util */
10177 if (idlest_sgs->idle_cpus == sgs->idle_cpus &&
10178 idlest_sgs->group_util <= sgs->group_util)
10188 * find_idlest_group() finds and returns the least busy CPU group within the
10191 * Assumes p is allowed on at least one CPU in sd.
10193 static struct sched_group *
10194 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
10196 struct sched_group *idlest = NULL, *local = NULL, *group = sd->groups;
10197 struct sg_lb_stats local_sgs, tmp_sgs;
10198 struct sg_lb_stats *sgs;
10199 unsigned long imbalance;
10200 struct sg_lb_stats idlest_sgs = {
10201 .avg_load = UINT_MAX,
10202 .group_type = group_overloaded,
10208 /* Skip over this group if it has no CPUs allowed */
10209 if (!cpumask_intersects(sched_group_span(group),
10213 /* Skip over this group if no cookie matched */
10214 if (!sched_group_cookie_match(cpu_rq(this_cpu), p, group))
10217 local_group = cpumask_test_cpu(this_cpu,
10218 sched_group_span(group));
10227 update_sg_wakeup_stats(sd, group, sgs, p);
10229 if (!local_group && update_pick_idlest(idlest, &idlest_sgs, group, sgs)) {
10234 } while (group = group->next, group != sd->groups);
10237 /* There is no idlest group to push tasks to */
10241 /* The local group has been skipped because of CPU affinity */
10246 * If the local group is idler than the selected idlest group
10247 * don't try and push the task.
10249 if (local_sgs.group_type < idlest_sgs.group_type)
10253 * If the local group is busier than the selected idlest group
10254 * try and push the task.
10256 if (local_sgs.group_type > idlest_sgs.group_type)
10259 switch (local_sgs.group_type) {
10260 case group_overloaded:
10261 case group_fully_busy:
10263 /* Calculate allowed imbalance based on load */
10264 imbalance = scale_load_down(NICE_0_LOAD) *
10265 (sd->imbalance_pct-100) / 100;
10268 * When comparing groups across NUMA domains, it's possible for
10269 * the local domain to be very lightly loaded relative to the
10270 * remote domains but "imbalance" skews the comparison making
10271 * remote CPUs look much more favourable. When considering
10272 * cross-domain, add imbalance to the load on the remote node
10273 * and consider staying local.
10276 if ((sd->flags & SD_NUMA) &&
10277 ((idlest_sgs.avg_load + imbalance) >= local_sgs.avg_load))
10281 * If the local group is less loaded than the selected
10282 * idlest group don't try and push any tasks.
10284 if (idlest_sgs.avg_load >= (local_sgs.avg_load + imbalance))
10287 if (100 * local_sgs.avg_load <= sd->imbalance_pct * idlest_sgs.avg_load)
10291 case group_imbalanced:
10292 case group_asym_packing:
10293 case group_smt_balance:
10294 /* Those type are not used in the slow wakeup path */
10297 case group_misfit_task:
10298 /* Select group with the highest max capacity */
10299 if (local->sgc->max_capacity >= idlest->sgc->max_capacity)
10303 case group_has_spare:
10305 if (sd->flags & SD_NUMA) {
10306 int imb_numa_nr = sd->imb_numa_nr;
10307 #ifdef CONFIG_NUMA_BALANCING
10310 * If there is spare capacity at NUMA, try to select
10311 * the preferred node
10313 if (cpu_to_node(this_cpu) == p->numa_preferred_nid)
10316 idlest_cpu = cpumask_first(sched_group_span(idlest));
10317 if (cpu_to_node(idlest_cpu) == p->numa_preferred_nid)
10319 #endif /* CONFIG_NUMA_BALANCING */
10321 * Otherwise, keep the task close to the wakeup source
10322 * and improve locality if the number of running tasks
10323 * would remain below threshold where an imbalance is
10324 * allowed while accounting for the possibility the
10325 * task is pinned to a subset of CPUs. If there is a
10326 * real need of migration, periodic load balance will
10329 if (p->nr_cpus_allowed != NR_CPUS) {
10330 struct cpumask *cpus = this_cpu_cpumask_var_ptr(select_rq_mask);
10332 cpumask_and(cpus, sched_group_span(local), p->cpus_ptr);
10333 imb_numa_nr = min(cpumask_weight(cpus), sd->imb_numa_nr);
10336 imbalance = abs(local_sgs.idle_cpus - idlest_sgs.idle_cpus);
10337 if (!adjust_numa_imbalance(imbalance,
10338 local_sgs.sum_nr_running + 1,
10343 #endif /* CONFIG_NUMA */
10346 * Select group with highest number of idle CPUs. We could also
10347 * compare the utilization which is more stable but it can end
10348 * up that the group has less spare capacity but finally more
10349 * idle CPUs which means more opportunity to run task.
10351 if (local_sgs.idle_cpus >= idlest_sgs.idle_cpus)
10359 static void update_idle_cpu_scan(struct lb_env *env,
10360 unsigned long sum_util)
10362 struct sched_domain_shared *sd_share;
10363 int llc_weight, pct;
10366 * Update the number of CPUs to scan in LLC domain, which could
10367 * be used as a hint in select_idle_cpu(). The update of sd_share
10368 * could be expensive because it is within a shared cache line.
10369 * So the write of this hint only occurs during periodic load
10370 * balancing, rather than CPU_NEWLY_IDLE, because the latter
10371 * can fire way more frequently than the former.
10373 if (!sched_feat(SIS_UTIL) || env->idle == CPU_NEWLY_IDLE)
10376 llc_weight = per_cpu(sd_llc_size, env->dst_cpu);
10377 if (env->sd->span_weight != llc_weight)
10380 sd_share = rcu_dereference(per_cpu(sd_llc_shared, env->dst_cpu));
10385 * The number of CPUs to search drops as sum_util increases, when
10386 * sum_util hits 85% or above, the scan stops.
10387 * The reason to choose 85% as the threshold is because this is the
10388 * imbalance_pct(117) when a LLC sched group is overloaded.
10390 * let y = SCHED_CAPACITY_SCALE - p * x^2 [1]
10391 * and y'= y / SCHED_CAPACITY_SCALE
10393 * x is the ratio of sum_util compared to the CPU capacity:
10394 * x = sum_util / (llc_weight * SCHED_CAPACITY_SCALE)
10395 * y' is the ratio of CPUs to be scanned in the LLC domain,
10396 * and the number of CPUs to scan is calculated by:
10398 * nr_scan = llc_weight * y' [2]
10400 * When x hits the threshold of overloaded, AKA, when
10401 * x = 100 / pct, y drops to 0. According to [1],
10402 * p should be SCHED_CAPACITY_SCALE * pct^2 / 10000
10404 * Scale x by SCHED_CAPACITY_SCALE:
10405 * x' = sum_util / llc_weight; [3]
10407 * and finally [1] becomes:
10408 * y = SCHED_CAPACITY_SCALE -
10409 * x'^2 * pct^2 / (10000 * SCHED_CAPACITY_SCALE) [4]
10414 do_div(x, llc_weight);
10417 pct = env->sd->imbalance_pct;
10418 tmp = x * x * pct * pct;
10419 do_div(tmp, 10000 * SCHED_CAPACITY_SCALE);
10420 tmp = min_t(long, tmp, SCHED_CAPACITY_SCALE);
10421 y = SCHED_CAPACITY_SCALE - tmp;
10425 do_div(y, SCHED_CAPACITY_SCALE);
10426 if ((int)y != sd_share->nr_idle_scan)
10427 WRITE_ONCE(sd_share->nr_idle_scan, (int)y);
10431 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
10432 * @env: The load balancing environment.
10433 * @sds: variable to hold the statistics for this sched_domain.
10436 static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
10438 struct sched_group *sg = env->sd->groups;
10439 struct sg_lb_stats *local = &sds->local_stat;
10440 struct sg_lb_stats tmp_sgs;
10441 unsigned long sum_util = 0;
10445 struct sg_lb_stats *sgs = &tmp_sgs;
10448 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_span(sg));
10453 if (env->idle != CPU_NEWLY_IDLE ||
10454 time_after_eq(jiffies, sg->sgc->next_update))
10455 update_group_capacity(env->sd, env->dst_cpu);
10458 update_sg_lb_stats(env, sds, sg, sgs, &sg_status);
10464 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
10466 sds->busiest_stat = *sgs;
10470 /* Now, start updating sd_lb_stats */
10471 sds->total_load += sgs->group_load;
10472 sds->total_capacity += sgs->group_capacity;
10474 sum_util += sgs->group_util;
10476 } while (sg != env->sd->groups);
10479 * Indicate that the child domain of the busiest group prefers tasks
10480 * go to a child's sibling domains first. NB the flags of a sched group
10481 * are those of the child domain.
10484 sds->prefer_sibling = !!(sds->busiest->flags & SD_PREFER_SIBLING);
10487 if (env->sd->flags & SD_NUMA)
10488 env->fbq_type = fbq_classify_group(&sds->busiest_stat);
10490 if (!env->sd->parent) {
10491 struct root_domain *rd = env->dst_rq->rd;
10493 /* update overload indicator if we are at root domain */
10494 WRITE_ONCE(rd->overload, sg_status & SG_OVERLOAD);
10496 /* Update over-utilization (tipping point, U >= 0) indicator */
10497 WRITE_ONCE(rd->overutilized, sg_status & SG_OVERUTILIZED);
10498 trace_sched_overutilized_tp(rd, sg_status & SG_OVERUTILIZED);
10499 } else if (sg_status & SG_OVERUTILIZED) {
10500 struct root_domain *rd = env->dst_rq->rd;
10502 WRITE_ONCE(rd->overutilized, SG_OVERUTILIZED);
10503 trace_sched_overutilized_tp(rd, SG_OVERUTILIZED);
10506 update_idle_cpu_scan(env, sum_util);
10510 * calculate_imbalance - Calculate the amount of imbalance present within the
10511 * groups of a given sched_domain during load balance.
10512 * @env: load balance environment
10513 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
10515 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
10517 struct sg_lb_stats *local, *busiest;
10519 local = &sds->local_stat;
10520 busiest = &sds->busiest_stat;
10522 if (busiest->group_type == group_misfit_task) {
10523 if (env->sd->flags & SD_ASYM_CPUCAPACITY) {
10524 /* Set imbalance to allow misfit tasks to be balanced. */
10525 env->migration_type = migrate_misfit;
10526 env->imbalance = 1;
10529 * Set load imbalance to allow moving task from cpu
10530 * with reduced capacity.
10532 env->migration_type = migrate_load;
10533 env->imbalance = busiest->group_misfit_task_load;
10538 if (busiest->group_type == group_asym_packing) {
10540 * In case of asym capacity, we will try to migrate all load to
10541 * the preferred CPU.
10543 env->migration_type = migrate_task;
10544 env->imbalance = busiest->sum_h_nr_running;
10548 if (busiest->group_type == group_smt_balance) {
10549 /* Reduce number of tasks sharing CPU capacity */
10550 env->migration_type = migrate_task;
10551 env->imbalance = 1;
10555 if (busiest->group_type == group_imbalanced) {
10557 * In the group_imb case we cannot rely on group-wide averages
10558 * to ensure CPU-load equilibrium, try to move any task to fix
10559 * the imbalance. The next load balance will take care of
10560 * balancing back the system.
10562 env->migration_type = migrate_task;
10563 env->imbalance = 1;
10568 * Try to use spare capacity of local group without overloading it or
10569 * emptying busiest.
10571 if (local->group_type == group_has_spare) {
10572 if ((busiest->group_type > group_fully_busy) &&
10573 !(env->sd->flags & SD_SHARE_PKG_RESOURCES)) {
10575 * If busiest is overloaded, try to fill spare
10576 * capacity. This might end up creating spare capacity
10577 * in busiest or busiest still being overloaded but
10578 * there is no simple way to directly compute the
10579 * amount of load to migrate in order to balance the
10582 env->migration_type = migrate_util;
10583 env->imbalance = max(local->group_capacity, local->group_util) -
10587 * In some cases, the group's utilization is max or even
10588 * higher than capacity because of migrations but the
10589 * local CPU is (newly) idle. There is at least one
10590 * waiting task in this overloaded busiest group. Let's
10593 if (env->idle != CPU_NOT_IDLE && env->imbalance == 0) {
10594 env->migration_type = migrate_task;
10595 env->imbalance = 1;
10601 if (busiest->group_weight == 1 || sds->prefer_sibling) {
10603 * When prefer sibling, evenly spread running tasks on
10606 env->migration_type = migrate_task;
10607 env->imbalance = sibling_imbalance(env, sds, busiest, local);
10611 * If there is no overload, we just want to even the number of
10614 env->migration_type = migrate_task;
10615 env->imbalance = max_t(long, 0,
10616 (local->idle_cpus - busiest->idle_cpus));
10620 /* Consider allowing a small imbalance between NUMA groups */
10621 if (env->sd->flags & SD_NUMA) {
10622 env->imbalance = adjust_numa_imbalance(env->imbalance,
10623 local->sum_nr_running + 1,
10624 env->sd->imb_numa_nr);
10628 /* Number of tasks to move to restore balance */
10629 env->imbalance >>= 1;
10635 * Local is fully busy but has to take more load to relieve the
10638 if (local->group_type < group_overloaded) {
10640 * Local will become overloaded so the avg_load metrics are
10644 local->avg_load = (local->group_load * SCHED_CAPACITY_SCALE) /
10645 local->group_capacity;
10648 * If the local group is more loaded than the selected
10649 * busiest group don't try to pull any tasks.
10651 if (local->avg_load >= busiest->avg_load) {
10652 env->imbalance = 0;
10656 sds->avg_load = (sds->total_load * SCHED_CAPACITY_SCALE) /
10657 sds->total_capacity;
10660 * If the local group is more loaded than the average system
10661 * load, don't try to pull any tasks.
10663 if (local->avg_load >= sds->avg_load) {
10664 env->imbalance = 0;
10671 * Both group are or will become overloaded and we're trying to get all
10672 * the CPUs to the average_load, so we don't want to push ourselves
10673 * above the average load, nor do we wish to reduce the max loaded CPU
10674 * below the average load. At the same time, we also don't want to
10675 * reduce the group load below the group capacity. Thus we look for
10676 * the minimum possible imbalance.
10678 env->migration_type = migrate_load;
10679 env->imbalance = min(
10680 (busiest->avg_load - sds->avg_load) * busiest->group_capacity,
10681 (sds->avg_load - local->avg_load) * local->group_capacity
10682 ) / SCHED_CAPACITY_SCALE;
10685 /******* find_busiest_group() helpers end here *********************/
10688 * Decision matrix according to the local and busiest group type:
10690 * busiest \ local has_spare fully_busy misfit asym imbalanced overloaded
10691 * has_spare nr_idle balanced N/A N/A balanced balanced
10692 * fully_busy nr_idle nr_idle N/A N/A balanced balanced
10693 * misfit_task force N/A N/A N/A N/A N/A
10694 * asym_packing force force N/A N/A force force
10695 * imbalanced force force N/A N/A force force
10696 * overloaded force force N/A N/A force avg_load
10698 * N/A : Not Applicable because already filtered while updating
10700 * balanced : The system is balanced for these 2 groups.
10701 * force : Calculate the imbalance as load migration is probably needed.
10702 * avg_load : Only if imbalance is significant enough.
10703 * nr_idle : dst_cpu is not busy and the number of idle CPUs is quite
10704 * different in groups.
10708 * find_busiest_group - Returns the busiest group within the sched_domain
10709 * if there is an imbalance.
10710 * @env: The load balancing environment.
10712 * Also calculates the amount of runnable load which should be moved
10713 * to restore balance.
10715 * Return: - The busiest group if imbalance exists.
10717 static struct sched_group *find_busiest_group(struct lb_env *env)
10719 struct sg_lb_stats *local, *busiest;
10720 struct sd_lb_stats sds;
10722 init_sd_lb_stats(&sds);
10725 * Compute the various statistics relevant for load balancing at
10728 update_sd_lb_stats(env, &sds);
10730 /* There is no busy sibling group to pull tasks from */
10734 busiest = &sds.busiest_stat;
10736 /* Misfit tasks should be dealt with regardless of the avg load */
10737 if (busiest->group_type == group_misfit_task)
10738 goto force_balance;
10740 if (sched_energy_enabled()) {
10741 struct root_domain *rd = env->dst_rq->rd;
10743 if (rcu_dereference(rd->pd) && !READ_ONCE(rd->overutilized))
10747 /* ASYM feature bypasses nice load balance check */
10748 if (busiest->group_type == group_asym_packing)
10749 goto force_balance;
10752 * If the busiest group is imbalanced the below checks don't
10753 * work because they assume all things are equal, which typically
10754 * isn't true due to cpus_ptr constraints and the like.
10756 if (busiest->group_type == group_imbalanced)
10757 goto force_balance;
10759 local = &sds.local_stat;
10761 * If the local group is busier than the selected busiest group
10762 * don't try and pull any tasks.
10764 if (local->group_type > busiest->group_type)
10768 * When groups are overloaded, use the avg_load to ensure fairness
10771 if (local->group_type == group_overloaded) {
10773 * If the local group is more loaded than the selected
10774 * busiest group don't try to pull any tasks.
10776 if (local->avg_load >= busiest->avg_load)
10779 /* XXX broken for overlapping NUMA groups */
10780 sds.avg_load = (sds.total_load * SCHED_CAPACITY_SCALE) /
10781 sds.total_capacity;
10784 * Don't pull any tasks if this group is already above the
10785 * domain average load.
10787 if (local->avg_load >= sds.avg_load)
10791 * If the busiest group is more loaded, use imbalance_pct to be
10794 if (100 * busiest->avg_load <=
10795 env->sd->imbalance_pct * local->avg_load)
10800 * Try to move all excess tasks to a sibling domain of the busiest
10801 * group's child domain.
10803 if (sds.prefer_sibling && local->group_type == group_has_spare &&
10804 sibling_imbalance(env, &sds, busiest, local) > 1)
10805 goto force_balance;
10807 if (busiest->group_type != group_overloaded) {
10808 if (env->idle == CPU_NOT_IDLE) {
10810 * If the busiest group is not overloaded (and as a
10811 * result the local one too) but this CPU is already
10812 * busy, let another idle CPU try to pull task.
10817 if (busiest->group_type == group_smt_balance &&
10818 smt_vs_nonsmt_groups(sds.local, sds.busiest)) {
10819 /* Let non SMT CPU pull from SMT CPU sharing with sibling */
10820 goto force_balance;
10823 if (busiest->group_weight > 1 &&
10824 local->idle_cpus <= (busiest->idle_cpus + 1)) {
10826 * If the busiest group is not overloaded
10827 * and there is no imbalance between this and busiest
10828 * group wrt idle CPUs, it is balanced. The imbalance
10829 * becomes significant if the diff is greater than 1
10830 * otherwise we might end up to just move the imbalance
10831 * on another group. Of course this applies only if
10832 * there is more than 1 CPU per group.
10837 if (busiest->sum_h_nr_running == 1) {
10839 * busiest doesn't have any tasks waiting to run
10846 /* Looks like there is an imbalance. Compute it */
10847 calculate_imbalance(env, &sds);
10848 return env->imbalance ? sds.busiest : NULL;
10851 env->imbalance = 0;
10856 * find_busiest_queue - find the busiest runqueue among the CPUs in the group.
10858 static struct rq *find_busiest_queue(struct lb_env *env,
10859 struct sched_group *group)
10861 struct rq *busiest = NULL, *rq;
10862 unsigned long busiest_util = 0, busiest_load = 0, busiest_capacity = 1;
10863 unsigned int busiest_nr = 0;
10866 for_each_cpu_and(i, sched_group_span(group), env->cpus) {
10867 unsigned long capacity, load, util;
10868 unsigned int nr_running;
10872 rt = fbq_classify_rq(rq);
10875 * We classify groups/runqueues into three groups:
10876 * - regular: there are !numa tasks
10877 * - remote: there are numa tasks that run on the 'wrong' node
10878 * - all: there is no distinction
10880 * In order to avoid migrating ideally placed numa tasks,
10881 * ignore those when there's better options.
10883 * If we ignore the actual busiest queue to migrate another
10884 * task, the next balance pass can still reduce the busiest
10885 * queue by moving tasks around inside the node.
10887 * If we cannot move enough load due to this classification
10888 * the next pass will adjust the group classification and
10889 * allow migration of more tasks.
10891 * Both cases only affect the total convergence complexity.
10893 if (rt > env->fbq_type)
10896 nr_running = rq->cfs.h_nr_running;
10900 capacity = capacity_of(i);
10903 * For ASYM_CPUCAPACITY domains, don't pick a CPU that could
10904 * eventually lead to active_balancing high->low capacity.
10905 * Higher per-CPU capacity is considered better than balancing
10908 if (env->sd->flags & SD_ASYM_CPUCAPACITY &&
10909 !capacity_greater(capacity_of(env->dst_cpu), capacity) &&
10914 * Make sure we only pull tasks from a CPU of lower priority
10915 * when balancing between SMT siblings.
10917 * If balancing between cores, let lower priority CPUs help
10918 * SMT cores with more than one busy sibling.
10920 if ((env->sd->flags & SD_ASYM_PACKING) &&
10921 sched_use_asym_prio(env->sd, i) &&
10922 sched_asym_prefer(i, env->dst_cpu) &&
10926 switch (env->migration_type) {
10929 * When comparing with load imbalance, use cpu_load()
10930 * which is not scaled with the CPU capacity.
10932 load = cpu_load(rq);
10934 if (nr_running == 1 && load > env->imbalance &&
10935 !check_cpu_capacity(rq, env->sd))
10939 * For the load comparisons with the other CPUs,
10940 * consider the cpu_load() scaled with the CPU
10941 * capacity, so that the load can be moved away
10942 * from the CPU that is potentially running at a
10945 * Thus we're looking for max(load_i / capacity_i),
10946 * crosswise multiplication to rid ourselves of the
10947 * division works out to:
10948 * load_i * capacity_j > load_j * capacity_i;
10949 * where j is our previous maximum.
10951 if (load * busiest_capacity > busiest_load * capacity) {
10952 busiest_load = load;
10953 busiest_capacity = capacity;
10959 util = cpu_util_cfs_boost(i);
10962 * Don't try to pull utilization from a CPU with one
10963 * running task. Whatever its utilization, we will fail
10966 if (nr_running <= 1)
10969 if (busiest_util < util) {
10970 busiest_util = util;
10976 if (busiest_nr < nr_running) {
10977 busiest_nr = nr_running;
10982 case migrate_misfit:
10984 * For ASYM_CPUCAPACITY domains with misfit tasks we
10985 * simply seek the "biggest" misfit task.
10987 if (rq->misfit_task_load > busiest_load) {
10988 busiest_load = rq->misfit_task_load;
11001 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
11002 * so long as it is large enough.
11004 #define MAX_PINNED_INTERVAL 512
11007 asym_active_balance(struct lb_env *env)
11010 * ASYM_PACKING needs to force migrate tasks from busy but lower
11011 * priority CPUs in order to pack all tasks in the highest priority
11012 * CPUs. When done between cores, do it only if the whole core if the
11013 * whole core is idle.
11015 * If @env::src_cpu is an SMT core with busy siblings, let
11016 * the lower priority @env::dst_cpu help it. Do not follow
11019 return env->idle != CPU_NOT_IDLE && (env->sd->flags & SD_ASYM_PACKING) &&
11020 sched_use_asym_prio(env->sd, env->dst_cpu) &&
11021 (sched_asym_prefer(env->dst_cpu, env->src_cpu) ||
11022 !sched_use_asym_prio(env->sd, env->src_cpu));
11026 imbalanced_active_balance(struct lb_env *env)
11028 struct sched_domain *sd = env->sd;
11031 * The imbalanced case includes the case of pinned tasks preventing a fair
11032 * distribution of the load on the system but also the even distribution of the
11033 * threads on a system with spare capacity
11035 if ((env->migration_type == migrate_task) &&
11036 (sd->nr_balance_failed > sd->cache_nice_tries+2))
11042 static int need_active_balance(struct lb_env *env)
11044 struct sched_domain *sd = env->sd;
11046 if (asym_active_balance(env))
11049 if (imbalanced_active_balance(env))
11053 * The dst_cpu is idle and the src_cpu CPU has only 1 CFS task.
11054 * It's worth migrating the task if the src_cpu's capacity is reduced
11055 * because of other sched_class or IRQs if more capacity stays
11056 * available on dst_cpu.
11058 if ((env->idle != CPU_NOT_IDLE) &&
11059 (env->src_rq->cfs.h_nr_running == 1)) {
11060 if ((check_cpu_capacity(env->src_rq, sd)) &&
11061 (capacity_of(env->src_cpu)*sd->imbalance_pct < capacity_of(env->dst_cpu)*100))
11065 if (env->migration_type == migrate_misfit)
11071 static int active_load_balance_cpu_stop(void *data);
11073 static int should_we_balance(struct lb_env *env)
11075 struct cpumask *swb_cpus = this_cpu_cpumask_var_ptr(should_we_balance_tmpmask);
11076 struct sched_group *sg = env->sd->groups;
11077 int cpu, idle_smt = -1;
11080 * Ensure the balancing environment is consistent; can happen
11081 * when the softirq triggers 'during' hotplug.
11083 if (!cpumask_test_cpu(env->dst_cpu, env->cpus))
11087 * In the newly idle case, we will allow all the CPUs
11088 * to do the newly idle load balance.
11090 * However, we bail out if we already have tasks or a wakeup pending,
11091 * to optimize wakeup latency.
11093 if (env->idle == CPU_NEWLY_IDLE) {
11094 if (env->dst_rq->nr_running > 0 || env->dst_rq->ttwu_pending)
11099 cpumask_copy(swb_cpus, group_balance_mask(sg));
11100 /* Try to find first idle CPU */
11101 for_each_cpu_and(cpu, swb_cpus, env->cpus) {
11102 if (!idle_cpu(cpu))
11106 * Don't balance to idle SMT in busy core right away when
11107 * balancing cores, but remember the first idle SMT CPU for
11108 * later consideration. Find CPU on an idle core first.
11110 if (!(env->sd->flags & SD_SHARE_CPUCAPACITY) && !is_core_idle(cpu)) {
11111 if (idle_smt == -1)
11114 * If the core is not idle, and first SMT sibling which is
11115 * idle has been found, then its not needed to check other
11116 * SMT siblings for idleness:
11118 #ifdef CONFIG_SCHED_SMT
11119 cpumask_andnot(swb_cpus, swb_cpus, cpu_smt_mask(cpu));
11124 /* Are we the first idle CPU? */
11125 return cpu == env->dst_cpu;
11128 if (idle_smt == env->dst_cpu)
11131 /* Are we the first CPU of this group ? */
11132 return group_balance_cpu(sg) == env->dst_cpu;
11136 * Check this_cpu to ensure it is balanced within domain. Attempt to move
11137 * tasks if there is an imbalance.
11139 static int load_balance(int this_cpu, struct rq *this_rq,
11140 struct sched_domain *sd, enum cpu_idle_type idle,
11141 int *continue_balancing)
11143 int ld_moved, cur_ld_moved, active_balance = 0;
11144 struct sched_domain *sd_parent = sd->parent;
11145 struct sched_group *group;
11146 struct rq *busiest;
11147 struct rq_flags rf;
11148 struct cpumask *cpus = this_cpu_cpumask_var_ptr(load_balance_mask);
11149 struct lb_env env = {
11151 .dst_cpu = this_cpu,
11153 .dst_grpmask = group_balance_mask(sd->groups),
11155 .loop_break = SCHED_NR_MIGRATE_BREAK,
11158 .tasks = LIST_HEAD_INIT(env.tasks),
11161 cpumask_and(cpus, sched_domain_span(sd), cpu_active_mask);
11163 schedstat_inc(sd->lb_count[idle]);
11166 if (!should_we_balance(&env)) {
11167 *continue_balancing = 0;
11171 group = find_busiest_group(&env);
11173 schedstat_inc(sd->lb_nobusyg[idle]);
11177 busiest = find_busiest_queue(&env, group);
11179 schedstat_inc(sd->lb_nobusyq[idle]);
11183 WARN_ON_ONCE(busiest == env.dst_rq);
11185 schedstat_add(sd->lb_imbalance[idle], env.imbalance);
11187 env.src_cpu = busiest->cpu;
11188 env.src_rq = busiest;
11191 /* Clear this flag as soon as we find a pullable task */
11192 env.flags |= LBF_ALL_PINNED;
11193 if (busiest->nr_running > 1) {
11195 * Attempt to move tasks. If find_busiest_group has found
11196 * an imbalance but busiest->nr_running <= 1, the group is
11197 * still unbalanced. ld_moved simply stays zero, so it is
11198 * correctly treated as an imbalance.
11200 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
11203 rq_lock_irqsave(busiest, &rf);
11204 update_rq_clock(busiest);
11207 * cur_ld_moved - load moved in current iteration
11208 * ld_moved - cumulative load moved across iterations
11210 cur_ld_moved = detach_tasks(&env);
11213 * We've detached some tasks from busiest_rq. Every
11214 * task is masked "TASK_ON_RQ_MIGRATING", so we can safely
11215 * unlock busiest->lock, and we are able to be sure
11216 * that nobody can manipulate the tasks in parallel.
11217 * See task_rq_lock() family for the details.
11220 rq_unlock(busiest, &rf);
11222 if (cur_ld_moved) {
11223 attach_tasks(&env);
11224 ld_moved += cur_ld_moved;
11227 local_irq_restore(rf.flags);
11229 if (env.flags & LBF_NEED_BREAK) {
11230 env.flags &= ~LBF_NEED_BREAK;
11231 /* Stop if we tried all running tasks */
11232 if (env.loop < busiest->nr_running)
11237 * Revisit (affine) tasks on src_cpu that couldn't be moved to
11238 * us and move them to an alternate dst_cpu in our sched_group
11239 * where they can run. The upper limit on how many times we
11240 * iterate on same src_cpu is dependent on number of CPUs in our
11243 * This changes load balance semantics a bit on who can move
11244 * load to a given_cpu. In addition to the given_cpu itself
11245 * (or a ilb_cpu acting on its behalf where given_cpu is
11246 * nohz-idle), we now have balance_cpu in a position to move
11247 * load to given_cpu. In rare situations, this may cause
11248 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
11249 * _independently_ and at _same_ time to move some load to
11250 * given_cpu) causing excess load to be moved to given_cpu.
11251 * This however should not happen so much in practice and
11252 * moreover subsequent load balance cycles should correct the
11253 * excess load moved.
11255 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
11257 /* Prevent to re-select dst_cpu via env's CPUs */
11258 __cpumask_clear_cpu(env.dst_cpu, env.cpus);
11260 env.dst_rq = cpu_rq(env.new_dst_cpu);
11261 env.dst_cpu = env.new_dst_cpu;
11262 env.flags &= ~LBF_DST_PINNED;
11264 env.loop_break = SCHED_NR_MIGRATE_BREAK;
11267 * Go back to "more_balance" rather than "redo" since we
11268 * need to continue with same src_cpu.
11274 * We failed to reach balance because of affinity.
11277 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
11279 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0)
11280 *group_imbalance = 1;
11283 /* All tasks on this runqueue were pinned by CPU affinity */
11284 if (unlikely(env.flags & LBF_ALL_PINNED)) {
11285 __cpumask_clear_cpu(cpu_of(busiest), cpus);
11287 * Attempting to continue load balancing at the current
11288 * sched_domain level only makes sense if there are
11289 * active CPUs remaining as possible busiest CPUs to
11290 * pull load from which are not contained within the
11291 * destination group that is receiving any migrated
11294 if (!cpumask_subset(cpus, env.dst_grpmask)) {
11296 env.loop_break = SCHED_NR_MIGRATE_BREAK;
11299 goto out_all_pinned;
11304 schedstat_inc(sd->lb_failed[idle]);
11306 * Increment the failure counter only on periodic balance.
11307 * We do not want newidle balance, which can be very
11308 * frequent, pollute the failure counter causing
11309 * excessive cache_hot migrations and active balances.
11311 if (idle != CPU_NEWLY_IDLE)
11312 sd->nr_balance_failed++;
11314 if (need_active_balance(&env)) {
11315 unsigned long flags;
11317 raw_spin_rq_lock_irqsave(busiest, flags);
11320 * Don't kick the active_load_balance_cpu_stop,
11321 * if the curr task on busiest CPU can't be
11322 * moved to this_cpu:
11324 if (!cpumask_test_cpu(this_cpu, busiest->curr->cpus_ptr)) {
11325 raw_spin_rq_unlock_irqrestore(busiest, flags);
11326 goto out_one_pinned;
11329 /* Record that we found at least one task that could run on this_cpu */
11330 env.flags &= ~LBF_ALL_PINNED;
11333 * ->active_balance synchronizes accesses to
11334 * ->active_balance_work. Once set, it's cleared
11335 * only after active load balance is finished.
11337 if (!busiest->active_balance) {
11338 busiest->active_balance = 1;
11339 busiest->push_cpu = this_cpu;
11340 active_balance = 1;
11344 raw_spin_rq_unlock_irqrestore(busiest, flags);
11345 if (active_balance) {
11346 stop_one_cpu_nowait(cpu_of(busiest),
11347 active_load_balance_cpu_stop, busiest,
11348 &busiest->active_balance_work);
11353 sd->nr_balance_failed = 0;
11356 if (likely(!active_balance) || need_active_balance(&env)) {
11357 /* We were unbalanced, so reset the balancing interval */
11358 sd->balance_interval = sd->min_interval;
11365 * We reach balance although we may have faced some affinity
11366 * constraints. Clear the imbalance flag only if other tasks got
11367 * a chance to move and fix the imbalance.
11369 if (sd_parent && !(env.flags & LBF_ALL_PINNED)) {
11370 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
11372 if (*group_imbalance)
11373 *group_imbalance = 0;
11378 * We reach balance because all tasks are pinned at this level so
11379 * we can't migrate them. Let the imbalance flag set so parent level
11380 * can try to migrate them.
11382 schedstat_inc(sd->lb_balanced[idle]);
11384 sd->nr_balance_failed = 0;
11390 * newidle_balance() disregards balance intervals, so we could
11391 * repeatedly reach this code, which would lead to balance_interval
11392 * skyrocketing in a short amount of time. Skip the balance_interval
11393 * increase logic to avoid that.
11395 if (env.idle == CPU_NEWLY_IDLE)
11398 /* tune up the balancing interval */
11399 if ((env.flags & LBF_ALL_PINNED &&
11400 sd->balance_interval < MAX_PINNED_INTERVAL) ||
11401 sd->balance_interval < sd->max_interval)
11402 sd->balance_interval *= 2;
11407 static inline unsigned long
11408 get_sd_balance_interval(struct sched_domain *sd, int cpu_busy)
11410 unsigned long interval = sd->balance_interval;
11413 interval *= sd->busy_factor;
11415 /* scale ms to jiffies */
11416 interval = msecs_to_jiffies(interval);
11419 * Reduce likelihood of busy balancing at higher domains racing with
11420 * balancing at lower domains by preventing their balancing periods
11421 * from being multiples of each other.
11426 interval = clamp(interval, 1UL, max_load_balance_interval);
11432 update_next_balance(struct sched_domain *sd, unsigned long *next_balance)
11434 unsigned long interval, next;
11436 /* used by idle balance, so cpu_busy = 0 */
11437 interval = get_sd_balance_interval(sd, 0);
11438 next = sd->last_balance + interval;
11440 if (time_after(*next_balance, next))
11441 *next_balance = next;
11445 * active_load_balance_cpu_stop is run by the CPU stopper. It pushes
11446 * running tasks off the busiest CPU onto idle CPUs. It requires at
11447 * least 1 task to be running on each physical CPU where possible, and
11448 * avoids physical / logical imbalances.
11450 static int active_load_balance_cpu_stop(void *data)
11452 struct rq *busiest_rq = data;
11453 int busiest_cpu = cpu_of(busiest_rq);
11454 int target_cpu = busiest_rq->push_cpu;
11455 struct rq *target_rq = cpu_rq(target_cpu);
11456 struct sched_domain *sd;
11457 struct task_struct *p = NULL;
11458 struct rq_flags rf;
11460 rq_lock_irq(busiest_rq, &rf);
11462 * Between queueing the stop-work and running it is a hole in which
11463 * CPUs can become inactive. We should not move tasks from or to
11466 if (!cpu_active(busiest_cpu) || !cpu_active(target_cpu))
11469 /* Make sure the requested CPU hasn't gone down in the meantime: */
11470 if (unlikely(busiest_cpu != smp_processor_id() ||
11471 !busiest_rq->active_balance))
11474 /* Is there any task to move? */
11475 if (busiest_rq->nr_running <= 1)
11479 * This condition is "impossible", if it occurs
11480 * we need to fix it. Originally reported by
11481 * Bjorn Helgaas on a 128-CPU setup.
11483 WARN_ON_ONCE(busiest_rq == target_rq);
11485 /* Search for an sd spanning us and the target CPU. */
11487 for_each_domain(target_cpu, sd) {
11488 if (cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
11493 struct lb_env env = {
11495 .dst_cpu = target_cpu,
11496 .dst_rq = target_rq,
11497 .src_cpu = busiest_rq->cpu,
11498 .src_rq = busiest_rq,
11500 .flags = LBF_ACTIVE_LB,
11503 schedstat_inc(sd->alb_count);
11504 update_rq_clock(busiest_rq);
11506 p = detach_one_task(&env);
11508 schedstat_inc(sd->alb_pushed);
11509 /* Active balancing done, reset the failure counter. */
11510 sd->nr_balance_failed = 0;
11512 schedstat_inc(sd->alb_failed);
11517 busiest_rq->active_balance = 0;
11518 rq_unlock(busiest_rq, &rf);
11521 attach_one_task(target_rq, p);
11523 local_irq_enable();
11528 static DEFINE_SPINLOCK(balancing);
11531 * Scale the max load_balance interval with the number of CPUs in the system.
11532 * This trades load-balance latency on larger machines for less cross talk.
11534 void update_max_interval(void)
11536 max_load_balance_interval = HZ*num_online_cpus()/10;
11539 static inline bool update_newidle_cost(struct sched_domain *sd, u64 cost)
11541 if (cost > sd->max_newidle_lb_cost) {
11543 * Track max cost of a domain to make sure to not delay the
11544 * next wakeup on the CPU.
11546 sd->max_newidle_lb_cost = cost;
11547 sd->last_decay_max_lb_cost = jiffies;
11548 } else if (time_after(jiffies, sd->last_decay_max_lb_cost + HZ)) {
11550 * Decay the newidle max times by ~1% per second to ensure that
11551 * it is not outdated and the current max cost is actually
11554 sd->max_newidle_lb_cost = (sd->max_newidle_lb_cost * 253) / 256;
11555 sd->last_decay_max_lb_cost = jiffies;
11564 * It checks each scheduling domain to see if it is due to be balanced,
11565 * and initiates a balancing operation if so.
11567 * Balancing parameters are set up in init_sched_domains.
11569 static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle)
11571 int continue_balancing = 1;
11573 int busy = idle != CPU_IDLE && !sched_idle_cpu(cpu);
11574 unsigned long interval;
11575 struct sched_domain *sd;
11576 /* Earliest time when we have to do rebalance again */
11577 unsigned long next_balance = jiffies + 60*HZ;
11578 int update_next_balance = 0;
11579 int need_serialize, need_decay = 0;
11583 for_each_domain(cpu, sd) {
11585 * Decay the newidle max times here because this is a regular
11586 * visit to all the domains.
11588 need_decay = update_newidle_cost(sd, 0);
11589 max_cost += sd->max_newidle_lb_cost;
11592 * Stop the load balance at this level. There is another
11593 * CPU in our sched group which is doing load balancing more
11596 if (!continue_balancing) {
11602 interval = get_sd_balance_interval(sd, busy);
11604 need_serialize = sd->flags & SD_SERIALIZE;
11605 if (need_serialize) {
11606 if (!spin_trylock(&balancing))
11610 if (time_after_eq(jiffies, sd->last_balance + interval)) {
11611 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
11613 * The LBF_DST_PINNED logic could have changed
11614 * env->dst_cpu, so we can't know our idle
11615 * state even if we migrated tasks. Update it.
11617 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
11618 busy = idle != CPU_IDLE && !sched_idle_cpu(cpu);
11620 sd->last_balance = jiffies;
11621 interval = get_sd_balance_interval(sd, busy);
11623 if (need_serialize)
11624 spin_unlock(&balancing);
11626 if (time_after(next_balance, sd->last_balance + interval)) {
11627 next_balance = sd->last_balance + interval;
11628 update_next_balance = 1;
11633 * Ensure the rq-wide value also decays but keep it at a
11634 * reasonable floor to avoid funnies with rq->avg_idle.
11636 rq->max_idle_balance_cost =
11637 max((u64)sysctl_sched_migration_cost, max_cost);
11642 * next_balance will be updated only when there is a need.
11643 * When the cpu is attached to null domain for ex, it will not be
11646 if (likely(update_next_balance))
11647 rq->next_balance = next_balance;
11651 static inline int on_null_domain(struct rq *rq)
11653 return unlikely(!rcu_dereference_sched(rq->sd));
11656 #ifdef CONFIG_NO_HZ_COMMON
11658 * idle load balancing details
11659 * - When one of the busy CPUs notice that there may be an idle rebalancing
11660 * needed, they will kick the idle load balancer, which then does idle
11661 * load balancing for all the idle CPUs.
11662 * - HK_TYPE_MISC CPUs are used for this task, because HK_TYPE_SCHED not set
11666 static inline int find_new_ilb(void)
11669 const struct cpumask *hk_mask;
11671 hk_mask = housekeeping_cpumask(HK_TYPE_MISC);
11673 for_each_cpu_and(ilb, nohz.idle_cpus_mask, hk_mask) {
11675 if (ilb == smp_processor_id())
11686 * Kick a CPU to do the nohz balancing, if it is time for it. We pick any
11687 * idle CPU in the HK_TYPE_MISC housekeeping set (if there is one).
11689 static void kick_ilb(unsigned int flags)
11694 * Increase nohz.next_balance only when if full ilb is triggered but
11695 * not if we only update stats.
11697 if (flags & NOHZ_BALANCE_KICK)
11698 nohz.next_balance = jiffies+1;
11700 ilb_cpu = find_new_ilb();
11702 if (ilb_cpu >= nr_cpu_ids)
11706 * Access to rq::nohz_csd is serialized by NOHZ_KICK_MASK; he who sets
11707 * the first flag owns it; cleared by nohz_csd_func().
11709 flags = atomic_fetch_or(flags, nohz_flags(ilb_cpu));
11710 if (flags & NOHZ_KICK_MASK)
11714 * This way we generate an IPI on the target CPU which
11715 * is idle. And the softirq performing nohz idle load balance
11716 * will be run before returning from the IPI.
11718 smp_call_function_single_async(ilb_cpu, &cpu_rq(ilb_cpu)->nohz_csd);
11722 * Current decision point for kicking the idle load balancer in the presence
11723 * of idle CPUs in the system.
11725 static void nohz_balancer_kick(struct rq *rq)
11727 unsigned long now = jiffies;
11728 struct sched_domain_shared *sds;
11729 struct sched_domain *sd;
11730 int nr_busy, i, cpu = rq->cpu;
11731 unsigned int flags = 0;
11733 if (unlikely(rq->idle_balance))
11737 * We may be recently in ticked or tickless idle mode. At the first
11738 * busy tick after returning from idle, we will update the busy stats.
11740 nohz_balance_exit_idle(rq);
11743 * None are in tickless mode and hence no need for NOHZ idle load
11746 if (likely(!atomic_read(&nohz.nr_cpus)))
11749 if (READ_ONCE(nohz.has_blocked) &&
11750 time_after(now, READ_ONCE(nohz.next_blocked)))
11751 flags = NOHZ_STATS_KICK;
11753 if (time_before(now, nohz.next_balance))
11756 if (rq->nr_running >= 2) {
11757 flags = NOHZ_STATS_KICK | NOHZ_BALANCE_KICK;
11763 sd = rcu_dereference(rq->sd);
11766 * If there's a CFS task and the current CPU has reduced
11767 * capacity; kick the ILB to see if there's a better CPU to run
11770 if (rq->cfs.h_nr_running >= 1 && check_cpu_capacity(rq, sd)) {
11771 flags = NOHZ_STATS_KICK | NOHZ_BALANCE_KICK;
11776 sd = rcu_dereference(per_cpu(sd_asym_packing, cpu));
11779 * When ASYM_PACKING; see if there's a more preferred CPU
11780 * currently idle; in which case, kick the ILB to move tasks
11783 * When balancing betwen cores, all the SMT siblings of the
11784 * preferred CPU must be idle.
11786 for_each_cpu_and(i, sched_domain_span(sd), nohz.idle_cpus_mask) {
11787 if (sched_use_asym_prio(sd, i) &&
11788 sched_asym_prefer(i, cpu)) {
11789 flags = NOHZ_STATS_KICK | NOHZ_BALANCE_KICK;
11795 sd = rcu_dereference(per_cpu(sd_asym_cpucapacity, cpu));
11798 * When ASYM_CPUCAPACITY; see if there's a higher capacity CPU
11799 * to run the misfit task on.
11801 if (check_misfit_status(rq, sd)) {
11802 flags = NOHZ_STATS_KICK | NOHZ_BALANCE_KICK;
11807 * For asymmetric systems, we do not want to nicely balance
11808 * cache use, instead we want to embrace asymmetry and only
11809 * ensure tasks have enough CPU capacity.
11811 * Skip the LLC logic because it's not relevant in that case.
11816 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
11819 * If there is an imbalance between LLC domains (IOW we could
11820 * increase the overall cache use), we need some less-loaded LLC
11821 * domain to pull some load. Likewise, we may need to spread
11822 * load within the current LLC domain (e.g. packed SMT cores but
11823 * other CPUs are idle). We can't really know from here how busy
11824 * the others are - so just get a nohz balance going if it looks
11825 * like this LLC domain has tasks we could move.
11827 nr_busy = atomic_read(&sds->nr_busy_cpus);
11829 flags = NOHZ_STATS_KICK | NOHZ_BALANCE_KICK;
11836 if (READ_ONCE(nohz.needs_update))
11837 flags |= NOHZ_NEXT_KICK;
11843 static void set_cpu_sd_state_busy(int cpu)
11845 struct sched_domain *sd;
11848 sd = rcu_dereference(per_cpu(sd_llc, cpu));
11850 if (!sd || !sd->nohz_idle)
11854 atomic_inc(&sd->shared->nr_busy_cpus);
11859 void nohz_balance_exit_idle(struct rq *rq)
11861 SCHED_WARN_ON(rq != this_rq());
11863 if (likely(!rq->nohz_tick_stopped))
11866 rq->nohz_tick_stopped = 0;
11867 cpumask_clear_cpu(rq->cpu, nohz.idle_cpus_mask);
11868 atomic_dec(&nohz.nr_cpus);
11870 set_cpu_sd_state_busy(rq->cpu);
11873 static void set_cpu_sd_state_idle(int cpu)
11875 struct sched_domain *sd;
11878 sd = rcu_dereference(per_cpu(sd_llc, cpu));
11880 if (!sd || sd->nohz_idle)
11884 atomic_dec(&sd->shared->nr_busy_cpus);
11890 * This routine will record that the CPU is going idle with tick stopped.
11891 * This info will be used in performing idle load balancing in the future.
11893 void nohz_balance_enter_idle(int cpu)
11895 struct rq *rq = cpu_rq(cpu);
11897 SCHED_WARN_ON(cpu != smp_processor_id());
11899 /* If this CPU is going down, then nothing needs to be done: */
11900 if (!cpu_active(cpu))
11903 /* Spare idle load balancing on CPUs that don't want to be disturbed: */
11904 if (!housekeeping_cpu(cpu, HK_TYPE_SCHED))
11908 * Can be set safely without rq->lock held
11909 * If a clear happens, it will have evaluated last additions because
11910 * rq->lock is held during the check and the clear
11912 rq->has_blocked_load = 1;
11915 * The tick is still stopped but load could have been added in the
11916 * meantime. We set the nohz.has_blocked flag to trig a check of the
11917 * *_avg. The CPU is already part of nohz.idle_cpus_mask so the clear
11918 * of nohz.has_blocked can only happen after checking the new load
11920 if (rq->nohz_tick_stopped)
11923 /* If we're a completely isolated CPU, we don't play: */
11924 if (on_null_domain(rq))
11927 rq->nohz_tick_stopped = 1;
11929 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
11930 atomic_inc(&nohz.nr_cpus);
11933 * Ensures that if nohz_idle_balance() fails to observe our
11934 * @idle_cpus_mask store, it must observe the @has_blocked
11935 * and @needs_update stores.
11937 smp_mb__after_atomic();
11939 set_cpu_sd_state_idle(cpu);
11941 WRITE_ONCE(nohz.needs_update, 1);
11944 * Each time a cpu enter idle, we assume that it has blocked load and
11945 * enable the periodic update of the load of idle cpus
11947 WRITE_ONCE(nohz.has_blocked, 1);
11950 static bool update_nohz_stats(struct rq *rq)
11952 unsigned int cpu = rq->cpu;
11954 if (!rq->has_blocked_load)
11957 if (!cpumask_test_cpu(cpu, nohz.idle_cpus_mask))
11960 if (!time_after(jiffies, READ_ONCE(rq->last_blocked_load_update_tick)))
11963 update_blocked_averages(cpu);
11965 return rq->has_blocked_load;
11969 * Internal function that runs load balance for all idle cpus. The load balance
11970 * can be a simple update of blocked load or a complete load balance with
11971 * tasks movement depending of flags.
11973 static void _nohz_idle_balance(struct rq *this_rq, unsigned int flags)
11975 /* Earliest time when we have to do rebalance again */
11976 unsigned long now = jiffies;
11977 unsigned long next_balance = now + 60*HZ;
11978 bool has_blocked_load = false;
11979 int update_next_balance = 0;
11980 int this_cpu = this_rq->cpu;
11984 SCHED_WARN_ON((flags & NOHZ_KICK_MASK) == NOHZ_BALANCE_KICK);
11987 * We assume there will be no idle load after this update and clear
11988 * the has_blocked flag. If a cpu enters idle in the mean time, it will
11989 * set the has_blocked flag and trigger another update of idle load.
11990 * Because a cpu that becomes idle, is added to idle_cpus_mask before
11991 * setting the flag, we are sure to not clear the state and not
11992 * check the load of an idle cpu.
11994 * Same applies to idle_cpus_mask vs needs_update.
11996 if (flags & NOHZ_STATS_KICK)
11997 WRITE_ONCE(nohz.has_blocked, 0);
11998 if (flags & NOHZ_NEXT_KICK)
11999 WRITE_ONCE(nohz.needs_update, 0);
12002 * Ensures that if we miss the CPU, we must see the has_blocked
12003 * store from nohz_balance_enter_idle().
12008 * Start with the next CPU after this_cpu so we will end with this_cpu and let a
12009 * chance for other idle cpu to pull load.
12011 for_each_cpu_wrap(balance_cpu, nohz.idle_cpus_mask, this_cpu+1) {
12012 if (!idle_cpu(balance_cpu))
12016 * If this CPU gets work to do, stop the load balancing
12017 * work being done for other CPUs. Next load
12018 * balancing owner will pick it up.
12020 if (need_resched()) {
12021 if (flags & NOHZ_STATS_KICK)
12022 has_blocked_load = true;
12023 if (flags & NOHZ_NEXT_KICK)
12024 WRITE_ONCE(nohz.needs_update, 1);
12028 rq = cpu_rq(balance_cpu);
12030 if (flags & NOHZ_STATS_KICK)
12031 has_blocked_load |= update_nohz_stats(rq);
12034 * If time for next balance is due,
12037 if (time_after_eq(jiffies, rq->next_balance)) {
12038 struct rq_flags rf;
12040 rq_lock_irqsave(rq, &rf);
12041 update_rq_clock(rq);
12042 rq_unlock_irqrestore(rq, &rf);
12044 if (flags & NOHZ_BALANCE_KICK)
12045 rebalance_domains(rq, CPU_IDLE);
12048 if (time_after(next_balance, rq->next_balance)) {
12049 next_balance = rq->next_balance;
12050 update_next_balance = 1;
12055 * next_balance will be updated only when there is a need.
12056 * When the CPU is attached to null domain for ex, it will not be
12059 if (likely(update_next_balance))
12060 nohz.next_balance = next_balance;
12062 if (flags & NOHZ_STATS_KICK)
12063 WRITE_ONCE(nohz.next_blocked,
12064 now + msecs_to_jiffies(LOAD_AVG_PERIOD));
12067 /* There is still blocked load, enable periodic update */
12068 if (has_blocked_load)
12069 WRITE_ONCE(nohz.has_blocked, 1);
12073 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
12074 * rebalancing for all the cpus for whom scheduler ticks are stopped.
12076 static bool nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
12078 unsigned int flags = this_rq->nohz_idle_balance;
12083 this_rq->nohz_idle_balance = 0;
12085 if (idle != CPU_IDLE)
12088 _nohz_idle_balance(this_rq, flags);
12094 * Check if we need to run the ILB for updating blocked load before entering
12097 void nohz_run_idle_balance(int cpu)
12099 unsigned int flags;
12101 flags = atomic_fetch_andnot(NOHZ_NEWILB_KICK, nohz_flags(cpu));
12104 * Update the blocked load only if no SCHED_SOFTIRQ is about to happen
12105 * (ie NOHZ_STATS_KICK set) and will do the same.
12107 if ((flags == NOHZ_NEWILB_KICK) && !need_resched())
12108 _nohz_idle_balance(cpu_rq(cpu), NOHZ_STATS_KICK);
12111 static void nohz_newidle_balance(struct rq *this_rq)
12113 int this_cpu = this_rq->cpu;
12116 * This CPU doesn't want to be disturbed by scheduler
12119 if (!housekeeping_cpu(this_cpu, HK_TYPE_SCHED))
12122 /* Will wake up very soon. No time for doing anything else*/
12123 if (this_rq->avg_idle < sysctl_sched_migration_cost)
12126 /* Don't need to update blocked load of idle CPUs*/
12127 if (!READ_ONCE(nohz.has_blocked) ||
12128 time_before(jiffies, READ_ONCE(nohz.next_blocked)))
12132 * Set the need to trigger ILB in order to update blocked load
12133 * before entering idle state.
12135 atomic_or(NOHZ_NEWILB_KICK, nohz_flags(this_cpu));
12138 #else /* !CONFIG_NO_HZ_COMMON */
12139 static inline void nohz_balancer_kick(struct rq *rq) { }
12141 static inline bool nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
12146 static inline void nohz_newidle_balance(struct rq *this_rq) { }
12147 #endif /* CONFIG_NO_HZ_COMMON */
12150 * newidle_balance is called by schedule() if this_cpu is about to become
12151 * idle. Attempts to pull tasks from other CPUs.
12154 * < 0 - we released the lock and there are !fair tasks present
12155 * 0 - failed, no new tasks
12156 * > 0 - success, new (fair) tasks present
12158 static int newidle_balance(struct rq *this_rq, struct rq_flags *rf)
12160 unsigned long next_balance = jiffies + HZ;
12161 int this_cpu = this_rq->cpu;
12162 u64 t0, t1, curr_cost = 0;
12163 struct sched_domain *sd;
12164 int pulled_task = 0;
12166 update_misfit_status(NULL, this_rq);
12169 * There is a task waiting to run. No need to search for one.
12170 * Return 0; the task will be enqueued when switching to idle.
12172 if (this_rq->ttwu_pending)
12176 * We must set idle_stamp _before_ calling idle_balance(), such that we
12177 * measure the duration of idle_balance() as idle time.
12179 this_rq->idle_stamp = rq_clock(this_rq);
12182 * Do not pull tasks towards !active CPUs...
12184 if (!cpu_active(this_cpu))
12188 * This is OK, because current is on_cpu, which avoids it being picked
12189 * for load-balance and preemption/IRQs are still disabled avoiding
12190 * further scheduler activity on it and we're being very careful to
12191 * re-start the picking loop.
12193 rq_unpin_lock(this_rq, rf);
12196 sd = rcu_dereference_check_sched_domain(this_rq->sd);
12198 if (!READ_ONCE(this_rq->rd->overload) ||
12199 (sd && this_rq->avg_idle < sd->max_newidle_lb_cost)) {
12202 update_next_balance(sd, &next_balance);
12209 raw_spin_rq_unlock(this_rq);
12211 t0 = sched_clock_cpu(this_cpu);
12212 update_blocked_averages(this_cpu);
12215 for_each_domain(this_cpu, sd) {
12216 int continue_balancing = 1;
12219 update_next_balance(sd, &next_balance);
12221 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost)
12224 if (sd->flags & SD_BALANCE_NEWIDLE) {
12226 pulled_task = load_balance(this_cpu, this_rq,
12227 sd, CPU_NEWLY_IDLE,
12228 &continue_balancing);
12230 t1 = sched_clock_cpu(this_cpu);
12231 domain_cost = t1 - t0;
12232 update_newidle_cost(sd, domain_cost);
12234 curr_cost += domain_cost;
12239 * Stop searching for tasks to pull if there are
12240 * now runnable tasks on this rq.
12242 if (pulled_task || this_rq->nr_running > 0 ||
12243 this_rq->ttwu_pending)
12248 raw_spin_rq_lock(this_rq);
12250 if (curr_cost > this_rq->max_idle_balance_cost)
12251 this_rq->max_idle_balance_cost = curr_cost;
12254 * While browsing the domains, we released the rq lock, a task could
12255 * have been enqueued in the meantime. Since we're not going idle,
12256 * pretend we pulled a task.
12258 if (this_rq->cfs.h_nr_running && !pulled_task)
12261 /* Is there a task of a high priority class? */
12262 if (this_rq->nr_running != this_rq->cfs.h_nr_running)
12266 /* Move the next balance forward */
12267 if (time_after(this_rq->next_balance, next_balance))
12268 this_rq->next_balance = next_balance;
12271 this_rq->idle_stamp = 0;
12273 nohz_newidle_balance(this_rq);
12275 rq_repin_lock(this_rq, rf);
12277 return pulled_task;
12281 * run_rebalance_domains is triggered when needed from the scheduler tick.
12282 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
12284 static __latent_entropy void run_rebalance_domains(struct softirq_action *h)
12286 struct rq *this_rq = this_rq();
12287 enum cpu_idle_type idle = this_rq->idle_balance ?
12288 CPU_IDLE : CPU_NOT_IDLE;
12291 * If this CPU has a pending nohz_balance_kick, then do the
12292 * balancing on behalf of the other idle CPUs whose ticks are
12293 * stopped. Do nohz_idle_balance *before* rebalance_domains to
12294 * give the idle CPUs a chance to load balance. Else we may
12295 * load balance only within the local sched_domain hierarchy
12296 * and abort nohz_idle_balance altogether if we pull some load.
12298 if (nohz_idle_balance(this_rq, idle))
12301 /* normal load balance */
12302 update_blocked_averages(this_rq->cpu);
12303 rebalance_domains(this_rq, idle);
12307 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
12309 void trigger_load_balance(struct rq *rq)
12312 * Don't need to rebalance while attached to NULL domain or
12313 * runqueue CPU is not active
12315 if (unlikely(on_null_domain(rq) || !cpu_active(cpu_of(rq))))
12318 if (time_after_eq(jiffies, rq->next_balance))
12319 raise_softirq(SCHED_SOFTIRQ);
12321 nohz_balancer_kick(rq);
12324 static void rq_online_fair(struct rq *rq)
12328 update_runtime_enabled(rq);
12331 static void rq_offline_fair(struct rq *rq)
12335 /* Ensure any throttled groups are reachable by pick_next_task */
12336 unthrottle_offline_cfs_rqs(rq);
12339 #endif /* CONFIG_SMP */
12341 #ifdef CONFIG_SCHED_CORE
12343 __entity_slice_used(struct sched_entity *se, int min_nr_tasks)
12345 u64 rtime = se->sum_exec_runtime - se->prev_sum_exec_runtime;
12346 u64 slice = se->slice;
12348 return (rtime * min_nr_tasks > slice);
12351 #define MIN_NR_TASKS_DURING_FORCEIDLE 2
12352 static inline void task_tick_core(struct rq *rq, struct task_struct *curr)
12354 if (!sched_core_enabled(rq))
12358 * If runqueue has only one task which used up its slice and
12359 * if the sibling is forced idle, then trigger schedule to
12360 * give forced idle task a chance.
12362 * sched_slice() considers only this active rq and it gets the
12363 * whole slice. But during force idle, we have siblings acting
12364 * like a single runqueue and hence we need to consider runnable
12365 * tasks on this CPU and the forced idle CPU. Ideally, we should
12366 * go through the forced idle rq, but that would be a perf hit.
12367 * We can assume that the forced idle CPU has at least
12368 * MIN_NR_TASKS_DURING_FORCEIDLE - 1 tasks and use that to check
12369 * if we need to give up the CPU.
12371 if (rq->core->core_forceidle_count && rq->cfs.nr_running == 1 &&
12372 __entity_slice_used(&curr->se, MIN_NR_TASKS_DURING_FORCEIDLE))
12377 * se_fi_update - Update the cfs_rq->min_vruntime_fi in a CFS hierarchy if needed.
12379 static void se_fi_update(const struct sched_entity *se, unsigned int fi_seq,
12382 for_each_sched_entity(se) {
12383 struct cfs_rq *cfs_rq = cfs_rq_of(se);
12386 if (cfs_rq->forceidle_seq == fi_seq)
12388 cfs_rq->forceidle_seq = fi_seq;
12391 cfs_rq->min_vruntime_fi = cfs_rq->min_vruntime;
12395 void task_vruntime_update(struct rq *rq, struct task_struct *p, bool in_fi)
12397 struct sched_entity *se = &p->se;
12399 if (p->sched_class != &fair_sched_class)
12402 se_fi_update(se, rq->core->core_forceidle_seq, in_fi);
12405 bool cfs_prio_less(const struct task_struct *a, const struct task_struct *b,
12408 struct rq *rq = task_rq(a);
12409 const struct sched_entity *sea = &a->se;
12410 const struct sched_entity *seb = &b->se;
12411 struct cfs_rq *cfs_rqa;
12412 struct cfs_rq *cfs_rqb;
12415 SCHED_WARN_ON(task_rq(b)->core != rq->core);
12417 #ifdef CONFIG_FAIR_GROUP_SCHED
12419 * Find an se in the hierarchy for tasks a and b, such that the se's
12420 * are immediate siblings.
12422 while (sea->cfs_rq->tg != seb->cfs_rq->tg) {
12423 int sea_depth = sea->depth;
12424 int seb_depth = seb->depth;
12426 if (sea_depth >= seb_depth)
12427 sea = parent_entity(sea);
12428 if (sea_depth <= seb_depth)
12429 seb = parent_entity(seb);
12432 se_fi_update(sea, rq->core->core_forceidle_seq, in_fi);
12433 se_fi_update(seb, rq->core->core_forceidle_seq, in_fi);
12435 cfs_rqa = sea->cfs_rq;
12436 cfs_rqb = seb->cfs_rq;
12438 cfs_rqa = &task_rq(a)->cfs;
12439 cfs_rqb = &task_rq(b)->cfs;
12443 * Find delta after normalizing se's vruntime with its cfs_rq's
12444 * min_vruntime_fi, which would have been updated in prior calls
12445 * to se_fi_update().
12447 delta = (s64)(sea->vruntime - seb->vruntime) +
12448 (s64)(cfs_rqb->min_vruntime_fi - cfs_rqa->min_vruntime_fi);
12453 static int task_is_throttled_fair(struct task_struct *p, int cpu)
12455 struct cfs_rq *cfs_rq;
12457 #ifdef CONFIG_FAIR_GROUP_SCHED
12458 cfs_rq = task_group(p)->cfs_rq[cpu];
12460 cfs_rq = &cpu_rq(cpu)->cfs;
12462 return throttled_hierarchy(cfs_rq);
12465 static inline void task_tick_core(struct rq *rq, struct task_struct *curr) {}
12469 * scheduler tick hitting a task of our scheduling class.
12471 * NOTE: This function can be called remotely by the tick offload that
12472 * goes along full dynticks. Therefore no local assumption can be made
12473 * and everything must be accessed through the @rq and @curr passed in
12476 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
12478 struct cfs_rq *cfs_rq;
12479 struct sched_entity *se = &curr->se;
12481 for_each_sched_entity(se) {
12482 cfs_rq = cfs_rq_of(se);
12483 entity_tick(cfs_rq, se, queued);
12486 if (static_branch_unlikely(&sched_numa_balancing))
12487 task_tick_numa(rq, curr);
12489 update_misfit_status(curr, rq);
12490 update_overutilized_status(task_rq(curr));
12492 task_tick_core(rq, curr);
12496 * called on fork with the child task as argument from the parent's context
12497 * - child not yet on the tasklist
12498 * - preemption disabled
12500 static void task_fork_fair(struct task_struct *p)
12502 struct sched_entity *se = &p->se, *curr;
12503 struct cfs_rq *cfs_rq;
12504 struct rq *rq = this_rq();
12505 struct rq_flags rf;
12508 update_rq_clock(rq);
12510 cfs_rq = task_cfs_rq(current);
12511 curr = cfs_rq->curr;
12513 update_curr(cfs_rq);
12514 place_entity(cfs_rq, se, ENQUEUE_INITIAL);
12515 rq_unlock(rq, &rf);
12519 * Priority of the task has changed. Check to see if we preempt
12520 * the current task.
12523 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
12525 if (!task_on_rq_queued(p))
12528 if (rq->cfs.nr_running == 1)
12532 * Reschedule if we are currently running on this runqueue and
12533 * our priority decreased, or if we are not currently running on
12534 * this runqueue and our priority is higher than the current's
12536 if (task_current(rq, p)) {
12537 if (p->prio > oldprio)
12540 check_preempt_curr(rq, p, 0);
12543 #ifdef CONFIG_FAIR_GROUP_SCHED
12545 * Propagate the changes of the sched_entity across the tg tree to make it
12546 * visible to the root
12548 static void propagate_entity_cfs_rq(struct sched_entity *se)
12550 struct cfs_rq *cfs_rq = cfs_rq_of(se);
12552 if (cfs_rq_throttled(cfs_rq))
12555 if (!throttled_hierarchy(cfs_rq))
12556 list_add_leaf_cfs_rq(cfs_rq);
12558 /* Start to propagate at parent */
12561 for_each_sched_entity(se) {
12562 cfs_rq = cfs_rq_of(se);
12564 update_load_avg(cfs_rq, se, UPDATE_TG);
12566 if (cfs_rq_throttled(cfs_rq))
12569 if (!throttled_hierarchy(cfs_rq))
12570 list_add_leaf_cfs_rq(cfs_rq);
12574 static void propagate_entity_cfs_rq(struct sched_entity *se) { }
12577 static void detach_entity_cfs_rq(struct sched_entity *se)
12579 struct cfs_rq *cfs_rq = cfs_rq_of(se);
12583 * In case the task sched_avg hasn't been attached:
12584 * - A forked task which hasn't been woken up by wake_up_new_task().
12585 * - A task which has been woken up by try_to_wake_up() but is
12586 * waiting for actually being woken up by sched_ttwu_pending().
12588 if (!se->avg.last_update_time)
12592 /* Catch up with the cfs_rq and remove our load when we leave */
12593 update_load_avg(cfs_rq, se, 0);
12594 detach_entity_load_avg(cfs_rq, se);
12595 update_tg_load_avg(cfs_rq);
12596 propagate_entity_cfs_rq(se);
12599 static void attach_entity_cfs_rq(struct sched_entity *se)
12601 struct cfs_rq *cfs_rq = cfs_rq_of(se);
12603 /* Synchronize entity with its cfs_rq */
12604 update_load_avg(cfs_rq, se, sched_feat(ATTACH_AGE_LOAD) ? 0 : SKIP_AGE_LOAD);
12605 attach_entity_load_avg(cfs_rq, se);
12606 update_tg_load_avg(cfs_rq);
12607 propagate_entity_cfs_rq(se);
12610 static void detach_task_cfs_rq(struct task_struct *p)
12612 struct sched_entity *se = &p->se;
12614 detach_entity_cfs_rq(se);
12617 static void attach_task_cfs_rq(struct task_struct *p)
12619 struct sched_entity *se = &p->se;
12621 attach_entity_cfs_rq(se);
12624 static void switched_from_fair(struct rq *rq, struct task_struct *p)
12626 detach_task_cfs_rq(p);
12629 static void switched_to_fair(struct rq *rq, struct task_struct *p)
12631 attach_task_cfs_rq(p);
12633 if (task_on_rq_queued(p)) {
12635 * We were most likely switched from sched_rt, so
12636 * kick off the schedule if running, otherwise just see
12637 * if we can still preempt the current task.
12639 if (task_current(rq, p))
12642 check_preempt_curr(rq, p, 0);
12646 /* Account for a task changing its policy or group.
12648 * This routine is mostly called to set cfs_rq->curr field when a task
12649 * migrates between groups/classes.
12651 static void set_next_task_fair(struct rq *rq, struct task_struct *p, bool first)
12653 struct sched_entity *se = &p->se;
12656 if (task_on_rq_queued(p)) {
12658 * Move the next running task to the front of the list, so our
12659 * cfs_tasks list becomes MRU one.
12661 list_move(&se->group_node, &rq->cfs_tasks);
12665 for_each_sched_entity(se) {
12666 struct cfs_rq *cfs_rq = cfs_rq_of(se);
12668 set_next_entity(cfs_rq, se);
12669 /* ensure bandwidth has been allocated on our new cfs_rq */
12670 account_cfs_rq_runtime(cfs_rq, 0);
12674 void init_cfs_rq(struct cfs_rq *cfs_rq)
12676 cfs_rq->tasks_timeline = RB_ROOT_CACHED;
12677 u64_u32_store(cfs_rq->min_vruntime, (u64)(-(1LL << 20)));
12679 raw_spin_lock_init(&cfs_rq->removed.lock);
12683 #ifdef CONFIG_FAIR_GROUP_SCHED
12684 static void task_change_group_fair(struct task_struct *p)
12687 * We couldn't detach or attach a forked task which
12688 * hasn't been woken up by wake_up_new_task().
12690 if (READ_ONCE(p->__state) == TASK_NEW)
12693 detach_task_cfs_rq(p);
12696 /* Tell se's cfs_rq has been changed -- migrated */
12697 p->se.avg.last_update_time = 0;
12699 set_task_rq(p, task_cpu(p));
12700 attach_task_cfs_rq(p);
12703 void free_fair_sched_group(struct task_group *tg)
12707 for_each_possible_cpu(i) {
12709 kfree(tg->cfs_rq[i]);
12718 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
12720 struct sched_entity *se;
12721 struct cfs_rq *cfs_rq;
12724 tg->cfs_rq = kcalloc(nr_cpu_ids, sizeof(cfs_rq), GFP_KERNEL);
12727 tg->se = kcalloc(nr_cpu_ids, sizeof(se), GFP_KERNEL);
12731 tg->shares = NICE_0_LOAD;
12733 init_cfs_bandwidth(tg_cfs_bandwidth(tg), tg_cfs_bandwidth(parent));
12735 for_each_possible_cpu(i) {
12736 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
12737 GFP_KERNEL, cpu_to_node(i));
12741 se = kzalloc_node(sizeof(struct sched_entity_stats),
12742 GFP_KERNEL, cpu_to_node(i));
12746 init_cfs_rq(cfs_rq);
12747 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
12748 init_entity_runnable_average(se);
12759 void online_fair_sched_group(struct task_group *tg)
12761 struct sched_entity *se;
12762 struct rq_flags rf;
12766 for_each_possible_cpu(i) {
12769 rq_lock_irq(rq, &rf);
12770 update_rq_clock(rq);
12771 attach_entity_cfs_rq(se);
12772 sync_throttle(tg, i);
12773 rq_unlock_irq(rq, &rf);
12777 void unregister_fair_sched_group(struct task_group *tg)
12779 unsigned long flags;
12783 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
12785 for_each_possible_cpu(cpu) {
12787 remove_entity_load_avg(tg->se[cpu]);
12790 * Only empty task groups can be destroyed; so we can speculatively
12791 * check on_list without danger of it being re-added.
12793 if (!tg->cfs_rq[cpu]->on_list)
12798 raw_spin_rq_lock_irqsave(rq, flags);
12799 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
12800 raw_spin_rq_unlock_irqrestore(rq, flags);
12804 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
12805 struct sched_entity *se, int cpu,
12806 struct sched_entity *parent)
12808 struct rq *rq = cpu_rq(cpu);
12812 init_cfs_rq_runtime(cfs_rq);
12814 tg->cfs_rq[cpu] = cfs_rq;
12817 /* se could be NULL for root_task_group */
12822 se->cfs_rq = &rq->cfs;
12825 se->cfs_rq = parent->my_q;
12826 se->depth = parent->depth + 1;
12830 /* guarantee group entities always have weight */
12831 update_load_set(&se->load, NICE_0_LOAD);
12832 se->parent = parent;
12835 static DEFINE_MUTEX(shares_mutex);
12837 static int __sched_group_set_shares(struct task_group *tg, unsigned long shares)
12841 lockdep_assert_held(&shares_mutex);
12844 * We can't change the weight of the root cgroup.
12849 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
12851 if (tg->shares == shares)
12854 tg->shares = shares;
12855 for_each_possible_cpu(i) {
12856 struct rq *rq = cpu_rq(i);
12857 struct sched_entity *se = tg->se[i];
12858 struct rq_flags rf;
12860 /* Propagate contribution to hierarchy */
12861 rq_lock_irqsave(rq, &rf);
12862 update_rq_clock(rq);
12863 for_each_sched_entity(se) {
12864 update_load_avg(cfs_rq_of(se), se, UPDATE_TG);
12865 update_cfs_group(se);
12867 rq_unlock_irqrestore(rq, &rf);
12873 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
12877 mutex_lock(&shares_mutex);
12878 if (tg_is_idle(tg))
12881 ret = __sched_group_set_shares(tg, shares);
12882 mutex_unlock(&shares_mutex);
12887 int sched_group_set_idle(struct task_group *tg, long idle)
12891 if (tg == &root_task_group)
12894 if (idle < 0 || idle > 1)
12897 mutex_lock(&shares_mutex);
12899 if (tg->idle == idle) {
12900 mutex_unlock(&shares_mutex);
12906 for_each_possible_cpu(i) {
12907 struct rq *rq = cpu_rq(i);
12908 struct sched_entity *se = tg->se[i];
12909 struct cfs_rq *parent_cfs_rq, *grp_cfs_rq = tg->cfs_rq[i];
12910 bool was_idle = cfs_rq_is_idle(grp_cfs_rq);
12911 long idle_task_delta;
12912 struct rq_flags rf;
12914 rq_lock_irqsave(rq, &rf);
12916 grp_cfs_rq->idle = idle;
12917 if (WARN_ON_ONCE(was_idle == cfs_rq_is_idle(grp_cfs_rq)))
12921 parent_cfs_rq = cfs_rq_of(se);
12922 if (cfs_rq_is_idle(grp_cfs_rq))
12923 parent_cfs_rq->idle_nr_running++;
12925 parent_cfs_rq->idle_nr_running--;
12928 idle_task_delta = grp_cfs_rq->h_nr_running -
12929 grp_cfs_rq->idle_h_nr_running;
12930 if (!cfs_rq_is_idle(grp_cfs_rq))
12931 idle_task_delta *= -1;
12933 for_each_sched_entity(se) {
12934 struct cfs_rq *cfs_rq = cfs_rq_of(se);
12939 cfs_rq->idle_h_nr_running += idle_task_delta;
12941 /* Already accounted at parent level and above. */
12942 if (cfs_rq_is_idle(cfs_rq))
12947 rq_unlock_irqrestore(rq, &rf);
12950 /* Idle groups have minimum weight. */
12951 if (tg_is_idle(tg))
12952 __sched_group_set_shares(tg, scale_load(WEIGHT_IDLEPRIO));
12954 __sched_group_set_shares(tg, NICE_0_LOAD);
12956 mutex_unlock(&shares_mutex);
12960 #else /* CONFIG_FAIR_GROUP_SCHED */
12962 void free_fair_sched_group(struct task_group *tg) { }
12964 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
12969 void online_fair_sched_group(struct task_group *tg) { }
12971 void unregister_fair_sched_group(struct task_group *tg) { }
12973 #endif /* CONFIG_FAIR_GROUP_SCHED */
12976 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
12978 struct sched_entity *se = &task->se;
12979 unsigned int rr_interval = 0;
12982 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
12985 if (rq->cfs.load.weight)
12986 rr_interval = NS_TO_JIFFIES(se->slice);
12988 return rr_interval;
12992 * All the scheduling class methods:
12994 DEFINE_SCHED_CLASS(fair) = {
12996 .enqueue_task = enqueue_task_fair,
12997 .dequeue_task = dequeue_task_fair,
12998 .yield_task = yield_task_fair,
12999 .yield_to_task = yield_to_task_fair,
13001 .check_preempt_curr = check_preempt_wakeup,
13003 .pick_next_task = __pick_next_task_fair,
13004 .put_prev_task = put_prev_task_fair,
13005 .set_next_task = set_next_task_fair,
13008 .balance = balance_fair,
13009 .pick_task = pick_task_fair,
13010 .select_task_rq = select_task_rq_fair,
13011 .migrate_task_rq = migrate_task_rq_fair,
13013 .rq_online = rq_online_fair,
13014 .rq_offline = rq_offline_fair,
13016 .task_dead = task_dead_fair,
13017 .set_cpus_allowed = set_cpus_allowed_common,
13020 .task_tick = task_tick_fair,
13021 .task_fork = task_fork_fair,
13023 .prio_changed = prio_changed_fair,
13024 .switched_from = switched_from_fair,
13025 .switched_to = switched_to_fair,
13027 .get_rr_interval = get_rr_interval_fair,
13029 .update_curr = update_curr_fair,
13031 #ifdef CONFIG_FAIR_GROUP_SCHED
13032 .task_change_group = task_change_group_fair,
13035 #ifdef CONFIG_SCHED_CORE
13036 .task_is_throttled = task_is_throttled_fair,
13039 #ifdef CONFIG_UCLAMP_TASK
13040 .uclamp_enabled = 1,
13044 #ifdef CONFIG_SCHED_DEBUG
13045 void print_cfs_stats(struct seq_file *m, int cpu)
13047 struct cfs_rq *cfs_rq, *pos;
13050 for_each_leaf_cfs_rq_safe(cpu_rq(cpu), cfs_rq, pos)
13051 print_cfs_rq(m, cpu, cfs_rq);
13055 #ifdef CONFIG_NUMA_BALANCING
13056 void show_numa_stats(struct task_struct *p, struct seq_file *m)
13059 unsigned long tsf = 0, tpf = 0, gsf = 0, gpf = 0;
13060 struct numa_group *ng;
13063 ng = rcu_dereference(p->numa_group);
13064 for_each_online_node(node) {
13065 if (p->numa_faults) {
13066 tsf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 0)];
13067 tpf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 1)];
13070 gsf = ng->faults[task_faults_idx(NUMA_MEM, node, 0)],
13071 gpf = ng->faults[task_faults_idx(NUMA_MEM, node, 1)];
13073 print_numa_stats(m, node, tsf, tpf, gsf, gpf);
13077 #endif /* CONFIG_NUMA_BALANCING */
13078 #endif /* CONFIG_SCHED_DEBUG */
13080 __init void init_sched_fair_class(void)
13085 for_each_possible_cpu(i) {
13086 zalloc_cpumask_var_node(&per_cpu(load_balance_mask, i), GFP_KERNEL, cpu_to_node(i));
13087 zalloc_cpumask_var_node(&per_cpu(select_rq_mask, i), GFP_KERNEL, cpu_to_node(i));
13088 zalloc_cpumask_var_node(&per_cpu(should_we_balance_tmpmask, i),
13089 GFP_KERNEL, cpu_to_node(i));
13091 #ifdef CONFIG_CFS_BANDWIDTH
13092 INIT_CSD(&cpu_rq(i)->cfsb_csd, __cfsb_csd_unthrottle, cpu_rq(i));
13093 INIT_LIST_HEAD(&cpu_rq(i)->cfsb_csd_list);
13097 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
13099 #ifdef CONFIG_NO_HZ_COMMON
13100 nohz.next_balance = jiffies;
13101 nohz.next_blocked = jiffies;
13102 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);