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/mempolicy.h>
44 #include <linux/mutex_api.h>
45 #include <linux/profile.h>
46 #include <linux/psi.h>
47 #include <linux/ratelimit.h>
48 #include <linux/task_work.h>
50 #include <asm/switch_to.h>
52 #include <linux/sched/cond_resched.h>
56 #include "autogroup.h"
59 * Targeted preemption latency for CPU-bound tasks:
61 * NOTE: this latency value is not the same as the concept of
62 * 'timeslice length' - timeslices in CFS are of variable length
63 * and have no persistent notion like in traditional, time-slice
64 * based scheduling concepts.
66 * (to see the precise effective timeslice length of your workload,
67 * run vmstat and monitor the context-switches (cs) field)
69 * (default: 6ms * (1 + ilog(ncpus)), units: nanoseconds)
71 unsigned int sysctl_sched_latency = 6000000ULL;
72 static unsigned int normalized_sysctl_sched_latency = 6000000ULL;
75 * The initial- and re-scaling of tunables is configurable
79 * SCHED_TUNABLESCALING_NONE - unscaled, always *1
80 * SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus)
81 * SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus
83 * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus))
85 unsigned int sysctl_sched_tunable_scaling = SCHED_TUNABLESCALING_LOG;
88 * Minimal preemption granularity for CPU-bound tasks:
90 * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds)
92 unsigned int sysctl_sched_min_granularity = 750000ULL;
93 static unsigned int normalized_sysctl_sched_min_granularity = 750000ULL;
96 * Minimal preemption granularity for CPU-bound SCHED_IDLE tasks.
97 * Applies only when SCHED_IDLE tasks compete with normal tasks.
99 * (default: 0.75 msec)
101 unsigned int sysctl_sched_idle_min_granularity = 750000ULL;
104 * This value is kept at sysctl_sched_latency/sysctl_sched_min_granularity
106 static unsigned int sched_nr_latency = 8;
109 * After fork, child runs first. If set to 0 (default) then
110 * parent will (try to) run first.
112 unsigned int sysctl_sched_child_runs_first __read_mostly;
115 * SCHED_OTHER wake-up granularity.
117 * This option delays the preemption effects of decoupled workloads
118 * and reduces their over-scheduling. Synchronous workloads will still
119 * have immediate wakeup/sleep latencies.
121 * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds)
123 unsigned int sysctl_sched_wakeup_granularity = 1000000UL;
124 static unsigned int normalized_sysctl_sched_wakeup_granularity = 1000000UL;
126 const_debug unsigned int sysctl_sched_migration_cost = 500000UL;
128 int sched_thermal_decay_shift;
129 static int __init setup_sched_thermal_decay_shift(char *str)
133 if (kstrtoint(str, 0, &_shift))
134 pr_warn("Unable to set scheduler thermal pressure decay shift parameter\n");
136 sched_thermal_decay_shift = clamp(_shift, 0, 10);
139 __setup("sched_thermal_decay_shift=", setup_sched_thermal_decay_shift);
143 * For asym packing, by default the lower numbered CPU has higher priority.
145 int __weak arch_asym_cpu_priority(int cpu)
151 * The margin used when comparing utilization with CPU capacity.
155 #define fits_capacity(cap, max) ((cap) * 1280 < (max) * 1024)
158 * The margin used when comparing CPU capacities.
159 * is 'cap1' noticeably greater than 'cap2'
163 #define capacity_greater(cap1, cap2) ((cap1) * 1024 > (cap2) * 1078)
166 #ifdef CONFIG_CFS_BANDWIDTH
168 * Amount of runtime to allocate from global (tg) to local (per-cfs_rq) pool
169 * each time a cfs_rq requests quota.
171 * Note: in the case that the slice exceeds the runtime remaining (either due
172 * to consumption or the quota being specified to be smaller than the slice)
173 * we will always only issue the remaining available time.
175 * (default: 5 msec, units: microseconds)
177 static unsigned int sysctl_sched_cfs_bandwidth_slice = 5000UL;
181 static struct ctl_table sched_fair_sysctls[] = {
183 .procname = "sched_child_runs_first",
184 .data = &sysctl_sched_child_runs_first,
185 .maxlen = sizeof(unsigned int),
187 .proc_handler = proc_dointvec,
189 #ifdef CONFIG_CFS_BANDWIDTH
191 .procname = "sched_cfs_bandwidth_slice_us",
192 .data = &sysctl_sched_cfs_bandwidth_slice,
193 .maxlen = sizeof(unsigned int),
195 .proc_handler = proc_dointvec_minmax,
196 .extra1 = SYSCTL_ONE,
202 static int __init sched_fair_sysctl_init(void)
204 register_sysctl_init("kernel", sched_fair_sysctls);
207 late_initcall(sched_fair_sysctl_init);
210 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
216 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
222 static inline void update_load_set(struct load_weight *lw, unsigned long w)
229 * Increase the granularity value when there are more CPUs,
230 * because with more CPUs the 'effective latency' as visible
231 * to users decreases. But the relationship is not linear,
232 * so pick a second-best guess by going with the log2 of the
235 * This idea comes from the SD scheduler of Con Kolivas:
237 static unsigned int get_update_sysctl_factor(void)
239 unsigned int cpus = min_t(unsigned int, num_online_cpus(), 8);
242 switch (sysctl_sched_tunable_scaling) {
243 case SCHED_TUNABLESCALING_NONE:
246 case SCHED_TUNABLESCALING_LINEAR:
249 case SCHED_TUNABLESCALING_LOG:
251 factor = 1 + ilog2(cpus);
258 static void update_sysctl(void)
260 unsigned int factor = get_update_sysctl_factor();
262 #define SET_SYSCTL(name) \
263 (sysctl_##name = (factor) * normalized_sysctl_##name)
264 SET_SYSCTL(sched_min_granularity);
265 SET_SYSCTL(sched_latency);
266 SET_SYSCTL(sched_wakeup_granularity);
270 void __init sched_init_granularity(void)
275 #define WMULT_CONST (~0U)
276 #define WMULT_SHIFT 32
278 static void __update_inv_weight(struct load_weight *lw)
282 if (likely(lw->inv_weight))
285 w = scale_load_down(lw->weight);
287 if (BITS_PER_LONG > 32 && unlikely(w >= WMULT_CONST))
289 else if (unlikely(!w))
290 lw->inv_weight = WMULT_CONST;
292 lw->inv_weight = WMULT_CONST / w;
296 * delta_exec * weight / lw.weight
298 * (delta_exec * (weight * lw->inv_weight)) >> WMULT_SHIFT
300 * Either weight := NICE_0_LOAD and lw \e sched_prio_to_wmult[], in which case
301 * we're guaranteed shift stays positive because inv_weight is guaranteed to
302 * fit 32 bits, and NICE_0_LOAD gives another 10 bits; therefore shift >= 22.
304 * Or, weight =< lw.weight (because lw.weight is the runqueue weight), thus
305 * weight/lw.weight <= 1, and therefore our shift will also be positive.
307 static u64 __calc_delta(u64 delta_exec, unsigned long weight, struct load_weight *lw)
309 u64 fact = scale_load_down(weight);
310 u32 fact_hi = (u32)(fact >> 32);
311 int shift = WMULT_SHIFT;
314 __update_inv_weight(lw);
316 if (unlikely(fact_hi)) {
322 fact = mul_u32_u32(fact, lw->inv_weight);
324 fact_hi = (u32)(fact >> 32);
331 return mul_u64_u32_shr(delta_exec, fact, shift);
335 const struct sched_class fair_sched_class;
337 /**************************************************************
338 * CFS operations on generic schedulable entities:
341 #ifdef CONFIG_FAIR_GROUP_SCHED
343 /* Walk up scheduling entities hierarchy */
344 #define for_each_sched_entity(se) \
345 for (; se; se = se->parent)
347 static inline bool list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
349 struct rq *rq = rq_of(cfs_rq);
350 int cpu = cpu_of(rq);
353 return rq->tmp_alone_branch == &rq->leaf_cfs_rq_list;
358 * Ensure we either appear before our parent (if already
359 * enqueued) or force our parent to appear after us when it is
360 * enqueued. The fact that we always enqueue bottom-up
361 * reduces this to two cases and a special case for the root
362 * cfs_rq. Furthermore, it also means that we will always reset
363 * tmp_alone_branch either when the branch is connected
364 * to a tree or when we reach the top of the tree
366 if (cfs_rq->tg->parent &&
367 cfs_rq->tg->parent->cfs_rq[cpu]->on_list) {
369 * If parent is already on the list, we add the child
370 * just before. Thanks to circular linked property of
371 * the list, this means to put the child at the tail
372 * of the list that starts by parent.
374 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
375 &(cfs_rq->tg->parent->cfs_rq[cpu]->leaf_cfs_rq_list));
377 * The branch is now connected to its tree so we can
378 * reset tmp_alone_branch to the beginning of the
381 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
385 if (!cfs_rq->tg->parent) {
387 * cfs rq without parent should be put
388 * at the tail of the list.
390 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
391 &rq->leaf_cfs_rq_list);
393 * We have reach the top of a tree so we can reset
394 * tmp_alone_branch to the beginning of the list.
396 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
401 * The parent has not already been added so we want to
402 * make sure that it will be put after us.
403 * tmp_alone_branch points to the begin of the branch
404 * where we will add parent.
406 list_add_rcu(&cfs_rq->leaf_cfs_rq_list, rq->tmp_alone_branch);
408 * update tmp_alone_branch to points to the new begin
411 rq->tmp_alone_branch = &cfs_rq->leaf_cfs_rq_list;
415 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
417 if (cfs_rq->on_list) {
418 struct rq *rq = rq_of(cfs_rq);
421 * With cfs_rq being unthrottled/throttled during an enqueue,
422 * it can happen the tmp_alone_branch points the a leaf that
423 * we finally want to del. In this case, tmp_alone_branch moves
424 * to the prev element but it will point to rq->leaf_cfs_rq_list
425 * at the end of the enqueue.
427 if (rq->tmp_alone_branch == &cfs_rq->leaf_cfs_rq_list)
428 rq->tmp_alone_branch = cfs_rq->leaf_cfs_rq_list.prev;
430 list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
435 static inline void assert_list_leaf_cfs_rq(struct rq *rq)
437 SCHED_WARN_ON(rq->tmp_alone_branch != &rq->leaf_cfs_rq_list);
440 /* Iterate thr' all leaf cfs_rq's on a runqueue */
441 #define for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) \
442 list_for_each_entry_safe(cfs_rq, pos, &rq->leaf_cfs_rq_list, \
445 /* Do the two (enqueued) entities belong to the same group ? */
446 static inline struct cfs_rq *
447 is_same_group(struct sched_entity *se, struct sched_entity *pse)
449 if (se->cfs_rq == pse->cfs_rq)
455 static inline struct sched_entity *parent_entity(struct sched_entity *se)
461 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
463 int se_depth, pse_depth;
466 * preemption test can be made between sibling entities who are in the
467 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
468 * both tasks until we find their ancestors who are siblings of common
472 /* First walk up until both entities are at same depth */
473 se_depth = (*se)->depth;
474 pse_depth = (*pse)->depth;
476 while (se_depth > pse_depth) {
478 *se = parent_entity(*se);
481 while (pse_depth > se_depth) {
483 *pse = parent_entity(*pse);
486 while (!is_same_group(*se, *pse)) {
487 *se = parent_entity(*se);
488 *pse = parent_entity(*pse);
492 static int tg_is_idle(struct task_group *tg)
497 static int cfs_rq_is_idle(struct cfs_rq *cfs_rq)
499 return cfs_rq->idle > 0;
502 static int se_is_idle(struct sched_entity *se)
504 if (entity_is_task(se))
505 return task_has_idle_policy(task_of(se));
506 return cfs_rq_is_idle(group_cfs_rq(se));
509 #else /* !CONFIG_FAIR_GROUP_SCHED */
511 #define for_each_sched_entity(se) \
512 for (; se; se = NULL)
514 static inline bool list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
519 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
523 static inline void assert_list_leaf_cfs_rq(struct rq *rq)
527 #define for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) \
528 for (cfs_rq = &rq->cfs, pos = NULL; cfs_rq; cfs_rq = pos)
530 static inline struct sched_entity *parent_entity(struct sched_entity *se)
536 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
540 static inline int tg_is_idle(struct task_group *tg)
545 static int cfs_rq_is_idle(struct cfs_rq *cfs_rq)
550 static int se_is_idle(struct sched_entity *se)
555 #endif /* CONFIG_FAIR_GROUP_SCHED */
557 static __always_inline
558 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec);
560 /**************************************************************
561 * Scheduling class tree data structure manipulation methods:
564 static inline u64 max_vruntime(u64 max_vruntime, u64 vruntime)
566 s64 delta = (s64)(vruntime - max_vruntime);
568 max_vruntime = vruntime;
573 static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime)
575 s64 delta = (s64)(vruntime - min_vruntime);
577 min_vruntime = vruntime;
582 static inline bool entity_before(struct sched_entity *a,
583 struct sched_entity *b)
585 return (s64)(a->vruntime - b->vruntime) < 0;
588 #define __node_2_se(node) \
589 rb_entry((node), struct sched_entity, run_node)
591 static void update_min_vruntime(struct cfs_rq *cfs_rq)
593 struct sched_entity *curr = cfs_rq->curr;
594 struct rb_node *leftmost = rb_first_cached(&cfs_rq->tasks_timeline);
596 u64 vruntime = cfs_rq->min_vruntime;
600 vruntime = curr->vruntime;
605 if (leftmost) { /* non-empty tree */
606 struct sched_entity *se = __node_2_se(leftmost);
609 vruntime = se->vruntime;
611 vruntime = min_vruntime(vruntime, se->vruntime);
614 /* ensure we never gain time by being placed backwards. */
615 u64_u32_store(cfs_rq->min_vruntime,
616 max_vruntime(cfs_rq->min_vruntime, vruntime));
619 static inline bool __entity_less(struct rb_node *a, const struct rb_node *b)
621 return entity_before(__node_2_se(a), __node_2_se(b));
625 * Enqueue an entity into the rb-tree:
627 static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
629 rb_add_cached(&se->run_node, &cfs_rq->tasks_timeline, __entity_less);
632 static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
634 rb_erase_cached(&se->run_node, &cfs_rq->tasks_timeline);
637 struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq)
639 struct rb_node *left = rb_first_cached(&cfs_rq->tasks_timeline);
644 return __node_2_se(left);
647 static struct sched_entity *__pick_next_entity(struct sched_entity *se)
649 struct rb_node *next = rb_next(&se->run_node);
654 return __node_2_se(next);
657 #ifdef CONFIG_SCHED_DEBUG
658 struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
660 struct rb_node *last = rb_last(&cfs_rq->tasks_timeline.rb_root);
665 return __node_2_se(last);
668 /**************************************************************
669 * Scheduling class statistics methods:
672 int sched_update_scaling(void)
674 unsigned int factor = get_update_sysctl_factor();
676 sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency,
677 sysctl_sched_min_granularity);
679 #define WRT_SYSCTL(name) \
680 (normalized_sysctl_##name = sysctl_##name / (factor))
681 WRT_SYSCTL(sched_min_granularity);
682 WRT_SYSCTL(sched_latency);
683 WRT_SYSCTL(sched_wakeup_granularity);
693 static inline u64 calc_delta_fair(u64 delta, struct sched_entity *se)
695 if (unlikely(se->load.weight != NICE_0_LOAD))
696 delta = __calc_delta(delta, NICE_0_LOAD, &se->load);
702 * The idea is to set a period in which each task runs once.
704 * When there are too many tasks (sched_nr_latency) we have to stretch
705 * this period because otherwise the slices get too small.
707 * p = (nr <= nl) ? l : l*nr/nl
709 static u64 __sched_period(unsigned long nr_running)
711 if (unlikely(nr_running > sched_nr_latency))
712 return nr_running * sysctl_sched_min_granularity;
714 return sysctl_sched_latency;
717 static bool sched_idle_cfs_rq(struct cfs_rq *cfs_rq);
720 * We calculate the wall-time slice from the period by taking a part
721 * proportional to the weight.
725 static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
727 unsigned int nr_running = cfs_rq->nr_running;
728 struct sched_entity *init_se = se;
729 unsigned int min_gran;
732 if (sched_feat(ALT_PERIOD))
733 nr_running = rq_of(cfs_rq)->cfs.h_nr_running;
735 slice = __sched_period(nr_running + !se->on_rq);
737 for_each_sched_entity(se) {
738 struct load_weight *load;
739 struct load_weight lw;
740 struct cfs_rq *qcfs_rq;
742 qcfs_rq = cfs_rq_of(se);
743 load = &qcfs_rq->load;
745 if (unlikely(!se->on_rq)) {
748 update_load_add(&lw, se->load.weight);
751 slice = __calc_delta(slice, se->load.weight, load);
754 if (sched_feat(BASE_SLICE)) {
755 if (se_is_idle(init_se) && !sched_idle_cfs_rq(cfs_rq))
756 min_gran = sysctl_sched_idle_min_granularity;
758 min_gran = sysctl_sched_min_granularity;
760 slice = max_t(u64, slice, min_gran);
767 * We calculate the vruntime slice of a to-be-inserted task.
771 static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se)
773 return calc_delta_fair(sched_slice(cfs_rq, se), se);
779 static int select_idle_sibling(struct task_struct *p, int prev_cpu, int cpu);
780 static unsigned long task_h_load(struct task_struct *p);
781 static unsigned long capacity_of(int cpu);
783 /* Give new sched_entity start runnable values to heavy its load in infant time */
784 void init_entity_runnable_average(struct sched_entity *se)
786 struct sched_avg *sa = &se->avg;
788 memset(sa, 0, sizeof(*sa));
791 * Tasks are initialized with full load to be seen as heavy tasks until
792 * they get a chance to stabilize to their real load level.
793 * Group entities are initialized with zero load to reflect the fact that
794 * nothing has been attached to the task group yet.
796 if (entity_is_task(se))
797 sa->load_avg = scale_load_down(se->load.weight);
799 /* when this task enqueue'ed, it will contribute to its cfs_rq's load_avg */
802 static void attach_entity_cfs_rq(struct sched_entity *se);
805 * With new tasks being created, their initial util_avgs are extrapolated
806 * based on the cfs_rq's current util_avg:
808 * util_avg = cfs_rq->util_avg / (cfs_rq->load_avg + 1) * se.load.weight
810 * However, in many cases, the above util_avg does not give a desired
811 * value. Moreover, the sum of the util_avgs may be divergent, such
812 * as when the series is a harmonic series.
814 * To solve this problem, we also cap the util_avg of successive tasks to
815 * only 1/2 of the left utilization budget:
817 * util_avg_cap = (cpu_scale - cfs_rq->avg.util_avg) / 2^n
819 * where n denotes the nth task and cpu_scale the CPU capacity.
821 * For example, for a CPU with 1024 of capacity, a simplest series from
822 * the beginning would be like:
824 * task util_avg: 512, 256, 128, 64, 32, 16, 8, ...
825 * cfs_rq util_avg: 512, 768, 896, 960, 992, 1008, 1016, ...
827 * Finally, that extrapolated util_avg is clamped to the cap (util_avg_cap)
828 * if util_avg > util_avg_cap.
830 void post_init_entity_util_avg(struct task_struct *p)
832 struct sched_entity *se = &p->se;
833 struct cfs_rq *cfs_rq = cfs_rq_of(se);
834 struct sched_avg *sa = &se->avg;
835 long cpu_scale = arch_scale_cpu_capacity(cpu_of(rq_of(cfs_rq)));
836 long cap = (long)(cpu_scale - cfs_rq->avg.util_avg) / 2;
839 if (cfs_rq->avg.util_avg != 0) {
840 sa->util_avg = cfs_rq->avg.util_avg * se->load.weight;
841 sa->util_avg /= (cfs_rq->avg.load_avg + 1);
843 if (sa->util_avg > cap)
850 sa->runnable_avg = sa->util_avg;
852 if (p->sched_class != &fair_sched_class) {
854 * For !fair tasks do:
856 update_cfs_rq_load_avg(now, cfs_rq);
857 attach_entity_load_avg(cfs_rq, se);
858 switched_from_fair(rq, p);
860 * such that the next switched_to_fair() has the
863 se->avg.last_update_time = cfs_rq_clock_pelt(cfs_rq);
867 attach_entity_cfs_rq(se);
870 #else /* !CONFIG_SMP */
871 void init_entity_runnable_average(struct sched_entity *se)
874 void post_init_entity_util_avg(struct task_struct *p)
877 static void update_tg_load_avg(struct cfs_rq *cfs_rq)
880 #endif /* CONFIG_SMP */
883 * Update the current task's runtime statistics.
885 static void update_curr(struct cfs_rq *cfs_rq)
887 struct sched_entity *curr = cfs_rq->curr;
888 u64 now = rq_clock_task(rq_of(cfs_rq));
894 delta_exec = now - curr->exec_start;
895 if (unlikely((s64)delta_exec <= 0))
898 curr->exec_start = now;
900 if (schedstat_enabled()) {
901 struct sched_statistics *stats;
903 stats = __schedstats_from_se(curr);
904 __schedstat_set(stats->exec_max,
905 max(delta_exec, stats->exec_max));
908 curr->sum_exec_runtime += delta_exec;
909 schedstat_add(cfs_rq->exec_clock, delta_exec);
911 curr->vruntime += calc_delta_fair(delta_exec, curr);
912 update_min_vruntime(cfs_rq);
914 if (entity_is_task(curr)) {
915 struct task_struct *curtask = task_of(curr);
917 trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime);
918 cgroup_account_cputime(curtask, delta_exec);
919 account_group_exec_runtime(curtask, delta_exec);
922 account_cfs_rq_runtime(cfs_rq, delta_exec);
925 static void update_curr_fair(struct rq *rq)
927 update_curr(cfs_rq_of(&rq->curr->se));
931 update_stats_wait_start_fair(struct cfs_rq *cfs_rq, struct sched_entity *se)
933 struct sched_statistics *stats;
934 struct task_struct *p = NULL;
936 if (!schedstat_enabled())
939 stats = __schedstats_from_se(se);
941 if (entity_is_task(se))
944 __update_stats_wait_start(rq_of(cfs_rq), p, stats);
948 update_stats_wait_end_fair(struct cfs_rq *cfs_rq, struct sched_entity *se)
950 struct sched_statistics *stats;
951 struct task_struct *p = NULL;
953 if (!schedstat_enabled())
956 stats = __schedstats_from_se(se);
959 * When the sched_schedstat changes from 0 to 1, some sched se
960 * maybe already in the runqueue, the se->statistics.wait_start
961 * will be 0.So it will let the delta wrong. We need to avoid this
964 if (unlikely(!schedstat_val(stats->wait_start)))
967 if (entity_is_task(se))
970 __update_stats_wait_end(rq_of(cfs_rq), p, stats);
974 update_stats_enqueue_sleeper_fair(struct cfs_rq *cfs_rq, struct sched_entity *se)
976 struct sched_statistics *stats;
977 struct task_struct *tsk = NULL;
979 if (!schedstat_enabled())
982 stats = __schedstats_from_se(se);
984 if (entity_is_task(se))
987 __update_stats_enqueue_sleeper(rq_of(cfs_rq), tsk, stats);
991 * Task is being enqueued - update stats:
994 update_stats_enqueue_fair(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
996 if (!schedstat_enabled())
1000 * Are we enqueueing a waiting task? (for current tasks
1001 * a dequeue/enqueue event is a NOP)
1003 if (se != cfs_rq->curr)
1004 update_stats_wait_start_fair(cfs_rq, se);
1006 if (flags & ENQUEUE_WAKEUP)
1007 update_stats_enqueue_sleeper_fair(cfs_rq, se);
1011 update_stats_dequeue_fair(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
1014 if (!schedstat_enabled())
1018 * Mark the end of the wait period if dequeueing a
1021 if (se != cfs_rq->curr)
1022 update_stats_wait_end_fair(cfs_rq, se);
1024 if ((flags & DEQUEUE_SLEEP) && entity_is_task(se)) {
1025 struct task_struct *tsk = task_of(se);
1028 /* XXX racy against TTWU */
1029 state = READ_ONCE(tsk->__state);
1030 if (state & TASK_INTERRUPTIBLE)
1031 __schedstat_set(tsk->stats.sleep_start,
1032 rq_clock(rq_of(cfs_rq)));
1033 if (state & TASK_UNINTERRUPTIBLE)
1034 __schedstat_set(tsk->stats.block_start,
1035 rq_clock(rq_of(cfs_rq)));
1040 * We are picking a new current task - update its stats:
1043 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
1046 * We are starting a new run period:
1048 se->exec_start = rq_clock_task(rq_of(cfs_rq));
1051 /**************************************************
1052 * Scheduling class queueing methods:
1056 #define NUMA_IMBALANCE_MIN 2
1059 adjust_numa_imbalance(int imbalance, int dst_running, int imb_numa_nr)
1062 * Allow a NUMA imbalance if busy CPUs is less than the maximum
1063 * threshold. Above this threshold, individual tasks may be contending
1064 * for both memory bandwidth and any shared HT resources. This is an
1065 * approximation as the number of running tasks may not be related to
1066 * the number of busy CPUs due to sched_setaffinity.
1068 if (dst_running > imb_numa_nr)
1072 * Allow a small imbalance based on a simple pair of communicating
1073 * tasks that remain local when the destination is lightly loaded.
1075 if (imbalance <= NUMA_IMBALANCE_MIN)
1080 #endif /* CONFIG_NUMA */
1082 #ifdef CONFIG_NUMA_BALANCING
1084 * Approximate time to scan a full NUMA task in ms. The task scan period is
1085 * calculated based on the tasks virtual memory size and
1086 * numa_balancing_scan_size.
1088 unsigned int sysctl_numa_balancing_scan_period_min = 1000;
1089 unsigned int sysctl_numa_balancing_scan_period_max = 60000;
1091 /* Portion of address space to scan in MB */
1092 unsigned int sysctl_numa_balancing_scan_size = 256;
1094 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
1095 unsigned int sysctl_numa_balancing_scan_delay = 1000;
1098 refcount_t refcount;
1100 spinlock_t lock; /* nr_tasks, tasks */
1105 struct rcu_head rcu;
1106 unsigned long total_faults;
1107 unsigned long max_faults_cpu;
1109 * faults[] array is split into two regions: faults_mem and faults_cpu.
1111 * Faults_cpu is used to decide whether memory should move
1112 * towards the CPU. As a consequence, these stats are weighted
1113 * more by CPU use than by memory faults.
1115 unsigned long faults[];
1119 * For functions that can be called in multiple contexts that permit reading
1120 * ->numa_group (see struct task_struct for locking rules).
1122 static struct numa_group *deref_task_numa_group(struct task_struct *p)
1124 return rcu_dereference_check(p->numa_group, p == current ||
1125 (lockdep_is_held(__rq_lockp(task_rq(p))) && !READ_ONCE(p->on_cpu)));
1128 static struct numa_group *deref_curr_numa_group(struct task_struct *p)
1130 return rcu_dereference_protected(p->numa_group, p == current);
1133 static inline unsigned long group_faults_priv(struct numa_group *ng);
1134 static inline unsigned long group_faults_shared(struct numa_group *ng);
1136 static unsigned int task_nr_scan_windows(struct task_struct *p)
1138 unsigned long rss = 0;
1139 unsigned long nr_scan_pages;
1142 * Calculations based on RSS as non-present and empty pages are skipped
1143 * by the PTE scanner and NUMA hinting faults should be trapped based
1146 nr_scan_pages = sysctl_numa_balancing_scan_size << (20 - PAGE_SHIFT);
1147 rss = get_mm_rss(p->mm);
1149 rss = nr_scan_pages;
1151 rss = round_up(rss, nr_scan_pages);
1152 return rss / nr_scan_pages;
1155 /* For sanity's sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */
1156 #define MAX_SCAN_WINDOW 2560
1158 static unsigned int task_scan_min(struct task_struct *p)
1160 unsigned int scan_size = READ_ONCE(sysctl_numa_balancing_scan_size);
1161 unsigned int scan, floor;
1162 unsigned int windows = 1;
1164 if (scan_size < MAX_SCAN_WINDOW)
1165 windows = MAX_SCAN_WINDOW / scan_size;
1166 floor = 1000 / windows;
1168 scan = sysctl_numa_balancing_scan_period_min / task_nr_scan_windows(p);
1169 return max_t(unsigned int, floor, scan);
1172 static unsigned int task_scan_start(struct task_struct *p)
1174 unsigned long smin = task_scan_min(p);
1175 unsigned long period = smin;
1176 struct numa_group *ng;
1178 /* Scale the maximum scan period with the amount of shared memory. */
1180 ng = rcu_dereference(p->numa_group);
1182 unsigned long shared = group_faults_shared(ng);
1183 unsigned long private = group_faults_priv(ng);
1185 period *= refcount_read(&ng->refcount);
1186 period *= shared + 1;
1187 period /= private + shared + 1;
1191 return max(smin, period);
1194 static unsigned int task_scan_max(struct task_struct *p)
1196 unsigned long smin = task_scan_min(p);
1198 struct numa_group *ng;
1200 /* Watch for min being lower than max due to floor calculations */
1201 smax = sysctl_numa_balancing_scan_period_max / task_nr_scan_windows(p);
1203 /* Scale the maximum scan period with the amount of shared memory. */
1204 ng = deref_curr_numa_group(p);
1206 unsigned long shared = group_faults_shared(ng);
1207 unsigned long private = group_faults_priv(ng);
1208 unsigned long period = smax;
1210 period *= refcount_read(&ng->refcount);
1211 period *= shared + 1;
1212 period /= private + shared + 1;
1214 smax = max(smax, period);
1217 return max(smin, smax);
1220 static void account_numa_enqueue(struct rq *rq, struct task_struct *p)
1222 rq->nr_numa_running += (p->numa_preferred_nid != NUMA_NO_NODE);
1223 rq->nr_preferred_running += (p->numa_preferred_nid == task_node(p));
1226 static void account_numa_dequeue(struct rq *rq, struct task_struct *p)
1228 rq->nr_numa_running -= (p->numa_preferred_nid != NUMA_NO_NODE);
1229 rq->nr_preferred_running -= (p->numa_preferred_nid == task_node(p));
1232 /* Shared or private faults. */
1233 #define NR_NUMA_HINT_FAULT_TYPES 2
1235 /* Memory and CPU locality */
1236 #define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2)
1238 /* Averaged statistics, and temporary buffers. */
1239 #define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2)
1241 pid_t task_numa_group_id(struct task_struct *p)
1243 struct numa_group *ng;
1247 ng = rcu_dereference(p->numa_group);
1256 * The averaged statistics, shared & private, memory & CPU,
1257 * occupy the first half of the array. The second half of the
1258 * array is for current counters, which are averaged into the
1259 * first set by task_numa_placement.
1261 static inline int task_faults_idx(enum numa_faults_stats s, int nid, int priv)
1263 return NR_NUMA_HINT_FAULT_TYPES * (s * nr_node_ids + nid) + priv;
1266 static inline unsigned long task_faults(struct task_struct *p, int nid)
1268 if (!p->numa_faults)
1271 return p->numa_faults[task_faults_idx(NUMA_MEM, nid, 0)] +
1272 p->numa_faults[task_faults_idx(NUMA_MEM, nid, 1)];
1275 static inline unsigned long group_faults(struct task_struct *p, int nid)
1277 struct numa_group *ng = deref_task_numa_group(p);
1282 return ng->faults[task_faults_idx(NUMA_MEM, nid, 0)] +
1283 ng->faults[task_faults_idx(NUMA_MEM, nid, 1)];
1286 static inline unsigned long group_faults_cpu(struct numa_group *group, int nid)
1288 return group->faults[task_faults_idx(NUMA_CPU, nid, 0)] +
1289 group->faults[task_faults_idx(NUMA_CPU, nid, 1)];
1292 static inline unsigned long group_faults_priv(struct numa_group *ng)
1294 unsigned long faults = 0;
1297 for_each_online_node(node) {
1298 faults += ng->faults[task_faults_idx(NUMA_MEM, node, 1)];
1304 static inline unsigned long group_faults_shared(struct numa_group *ng)
1306 unsigned long faults = 0;
1309 for_each_online_node(node) {
1310 faults += ng->faults[task_faults_idx(NUMA_MEM, node, 0)];
1317 * A node triggering more than 1/3 as many NUMA faults as the maximum is
1318 * considered part of a numa group's pseudo-interleaving set. Migrations
1319 * between these nodes are slowed down, to allow things to settle down.
1321 #define ACTIVE_NODE_FRACTION 3
1323 static bool numa_is_active_node(int nid, struct numa_group *ng)
1325 return group_faults_cpu(ng, nid) * ACTIVE_NODE_FRACTION > ng->max_faults_cpu;
1328 /* Handle placement on systems where not all nodes are directly connected. */
1329 static unsigned long score_nearby_nodes(struct task_struct *p, int nid,
1330 int lim_dist, bool task)
1332 unsigned long score = 0;
1336 * All nodes are directly connected, and the same distance
1337 * from each other. No need for fancy placement algorithms.
1339 if (sched_numa_topology_type == NUMA_DIRECT)
1342 /* sched_max_numa_distance may be changed in parallel. */
1343 max_dist = READ_ONCE(sched_max_numa_distance);
1345 * This code is called for each node, introducing N^2 complexity,
1346 * which should be ok given the number of nodes rarely exceeds 8.
1348 for_each_online_node(node) {
1349 unsigned long faults;
1350 int dist = node_distance(nid, node);
1353 * The furthest away nodes in the system are not interesting
1354 * for placement; nid was already counted.
1356 if (dist >= max_dist || node == nid)
1360 * On systems with a backplane NUMA topology, compare groups
1361 * of nodes, and move tasks towards the group with the most
1362 * memory accesses. When comparing two nodes at distance
1363 * "hoplimit", only nodes closer by than "hoplimit" are part
1364 * of each group. Skip other nodes.
1366 if (sched_numa_topology_type == NUMA_BACKPLANE && dist >= lim_dist)
1369 /* Add up the faults from nearby nodes. */
1371 faults = task_faults(p, node);
1373 faults = group_faults(p, node);
1376 * On systems with a glueless mesh NUMA topology, there are
1377 * no fixed "groups of nodes". Instead, nodes that are not
1378 * directly connected bounce traffic through intermediate
1379 * nodes; a numa_group can occupy any set of nodes.
1380 * The further away a node is, the less the faults count.
1381 * This seems to result in good task placement.
1383 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
1384 faults *= (max_dist - dist);
1385 faults /= (max_dist - LOCAL_DISTANCE);
1395 * These return the fraction of accesses done by a particular task, or
1396 * task group, on a particular numa node. The group weight is given a
1397 * larger multiplier, in order to group tasks together that are almost
1398 * evenly spread out between numa nodes.
1400 static inline unsigned long task_weight(struct task_struct *p, int nid,
1403 unsigned long faults, total_faults;
1405 if (!p->numa_faults)
1408 total_faults = p->total_numa_faults;
1413 faults = task_faults(p, nid);
1414 faults += score_nearby_nodes(p, nid, dist, true);
1416 return 1000 * faults / total_faults;
1419 static inline unsigned long group_weight(struct task_struct *p, int nid,
1422 struct numa_group *ng = deref_task_numa_group(p);
1423 unsigned long faults, total_faults;
1428 total_faults = ng->total_faults;
1433 faults = group_faults(p, nid);
1434 faults += score_nearby_nodes(p, nid, dist, false);
1436 return 1000 * faults / total_faults;
1439 bool should_numa_migrate_memory(struct task_struct *p, struct page * page,
1440 int src_nid, int dst_cpu)
1442 struct numa_group *ng = deref_curr_numa_group(p);
1443 int dst_nid = cpu_to_node(dst_cpu);
1444 int last_cpupid, this_cpupid;
1446 this_cpupid = cpu_pid_to_cpupid(dst_cpu, current->pid);
1447 last_cpupid = page_cpupid_xchg_last(page, this_cpupid);
1450 * Allow first faults or private faults to migrate immediately early in
1451 * the lifetime of a task. The magic number 4 is based on waiting for
1452 * two full passes of the "multi-stage node selection" test that is
1455 if ((p->numa_preferred_nid == NUMA_NO_NODE || p->numa_scan_seq <= 4) &&
1456 (cpupid_pid_unset(last_cpupid) || cpupid_match_pid(p, last_cpupid)))
1460 * Multi-stage node selection is used in conjunction with a periodic
1461 * migration fault to build a temporal task<->page relation. By using
1462 * a two-stage filter we remove short/unlikely relations.
1464 * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate
1465 * a task's usage of a particular page (n_p) per total usage of this
1466 * page (n_t) (in a given time-span) to a probability.
1468 * Our periodic faults will sample this probability and getting the
1469 * same result twice in a row, given these samples are fully
1470 * independent, is then given by P(n)^2, provided our sample period
1471 * is sufficiently short compared to the usage pattern.
1473 * This quadric squishes small probabilities, making it less likely we
1474 * act on an unlikely task<->page relation.
1476 if (!cpupid_pid_unset(last_cpupid) &&
1477 cpupid_to_nid(last_cpupid) != dst_nid)
1480 /* Always allow migrate on private faults */
1481 if (cpupid_match_pid(p, last_cpupid))
1484 /* A shared fault, but p->numa_group has not been set up yet. */
1489 * Destination node is much more heavily used than the source
1490 * node? Allow migration.
1492 if (group_faults_cpu(ng, dst_nid) > group_faults_cpu(ng, src_nid) *
1493 ACTIVE_NODE_FRACTION)
1497 * Distribute memory according to CPU & memory use on each node,
1498 * with 3/4 hysteresis to avoid unnecessary memory migrations:
1500 * faults_cpu(dst) 3 faults_cpu(src)
1501 * --------------- * - > ---------------
1502 * faults_mem(dst) 4 faults_mem(src)
1504 return group_faults_cpu(ng, dst_nid) * group_faults(p, src_nid) * 3 >
1505 group_faults_cpu(ng, src_nid) * group_faults(p, dst_nid) * 4;
1509 * 'numa_type' describes the node at the moment of load balancing.
1512 /* The node has spare capacity that can be used to run more tasks. */
1515 * The node is fully used and the tasks don't compete for more CPU
1516 * cycles. Nevertheless, some tasks might wait before running.
1520 * The node is overloaded and can't provide expected CPU cycles to all
1526 /* Cached statistics for all CPUs within a node */
1529 unsigned long runnable;
1531 /* Total compute capacity of CPUs on a node */
1532 unsigned long compute_capacity;
1533 unsigned int nr_running;
1534 unsigned int weight;
1535 enum numa_type node_type;
1539 static inline bool is_core_idle(int cpu)
1541 #ifdef CONFIG_SCHED_SMT
1544 for_each_cpu(sibling, cpu_smt_mask(cpu)) {
1548 if (!idle_cpu(sibling))
1556 struct task_numa_env {
1557 struct task_struct *p;
1559 int src_cpu, src_nid;
1560 int dst_cpu, dst_nid;
1563 struct numa_stats src_stats, dst_stats;
1568 struct task_struct *best_task;
1573 static unsigned long cpu_load(struct rq *rq);
1574 static unsigned long cpu_runnable(struct rq *rq);
1577 numa_type numa_classify(unsigned int imbalance_pct,
1578 struct numa_stats *ns)
1580 if ((ns->nr_running > ns->weight) &&
1581 (((ns->compute_capacity * 100) < (ns->util * imbalance_pct)) ||
1582 ((ns->compute_capacity * imbalance_pct) < (ns->runnable * 100))))
1583 return node_overloaded;
1585 if ((ns->nr_running < ns->weight) ||
1586 (((ns->compute_capacity * 100) > (ns->util * imbalance_pct)) &&
1587 ((ns->compute_capacity * imbalance_pct) > (ns->runnable * 100))))
1588 return node_has_spare;
1590 return node_fully_busy;
1593 #ifdef CONFIG_SCHED_SMT
1594 /* Forward declarations of select_idle_sibling helpers */
1595 static inline bool test_idle_cores(int cpu, bool def);
1596 static inline int numa_idle_core(int idle_core, int cpu)
1598 if (!static_branch_likely(&sched_smt_present) ||
1599 idle_core >= 0 || !test_idle_cores(cpu, false))
1603 * Prefer cores instead of packing HT siblings
1604 * and triggering future load balancing.
1606 if (is_core_idle(cpu))
1612 static inline int numa_idle_core(int idle_core, int cpu)
1619 * Gather all necessary information to make NUMA balancing placement
1620 * decisions that are compatible with standard load balancer. This
1621 * borrows code and logic from update_sg_lb_stats but sharing a
1622 * common implementation is impractical.
1624 static void update_numa_stats(struct task_numa_env *env,
1625 struct numa_stats *ns, int nid,
1628 int cpu, idle_core = -1;
1630 memset(ns, 0, sizeof(*ns));
1634 for_each_cpu(cpu, cpumask_of_node(nid)) {
1635 struct rq *rq = cpu_rq(cpu);
1637 ns->load += cpu_load(rq);
1638 ns->runnable += cpu_runnable(rq);
1639 ns->util += cpu_util_cfs(cpu);
1640 ns->nr_running += rq->cfs.h_nr_running;
1641 ns->compute_capacity += capacity_of(cpu);
1643 if (find_idle && !rq->nr_running && idle_cpu(cpu)) {
1644 if (READ_ONCE(rq->numa_migrate_on) ||
1645 !cpumask_test_cpu(cpu, env->p->cpus_ptr))
1648 if (ns->idle_cpu == -1)
1651 idle_core = numa_idle_core(idle_core, cpu);
1656 ns->weight = cpumask_weight(cpumask_of_node(nid));
1658 ns->node_type = numa_classify(env->imbalance_pct, ns);
1661 ns->idle_cpu = idle_core;
1664 static void task_numa_assign(struct task_numa_env *env,
1665 struct task_struct *p, long imp)
1667 struct rq *rq = cpu_rq(env->dst_cpu);
1669 /* Check if run-queue part of active NUMA balance. */
1670 if (env->best_cpu != env->dst_cpu && xchg(&rq->numa_migrate_on, 1)) {
1672 int start = env->dst_cpu;
1674 /* Find alternative idle CPU. */
1675 for_each_cpu_wrap(cpu, cpumask_of_node(env->dst_nid), start) {
1676 if (cpu == env->best_cpu || !idle_cpu(cpu) ||
1677 !cpumask_test_cpu(cpu, env->p->cpus_ptr)) {
1682 rq = cpu_rq(env->dst_cpu);
1683 if (!xchg(&rq->numa_migrate_on, 1))
1687 /* Failed to find an alternative idle CPU */
1693 * Clear previous best_cpu/rq numa-migrate flag, since task now
1694 * found a better CPU to move/swap.
1696 if (env->best_cpu != -1 && env->best_cpu != env->dst_cpu) {
1697 rq = cpu_rq(env->best_cpu);
1698 WRITE_ONCE(rq->numa_migrate_on, 0);
1702 put_task_struct(env->best_task);
1707 env->best_imp = imp;
1708 env->best_cpu = env->dst_cpu;
1711 static bool load_too_imbalanced(long src_load, long dst_load,
1712 struct task_numa_env *env)
1715 long orig_src_load, orig_dst_load;
1716 long src_capacity, dst_capacity;
1719 * The load is corrected for the CPU capacity available on each node.
1722 * ------------ vs ---------
1723 * src_capacity dst_capacity
1725 src_capacity = env->src_stats.compute_capacity;
1726 dst_capacity = env->dst_stats.compute_capacity;
1728 imb = abs(dst_load * src_capacity - src_load * dst_capacity);
1730 orig_src_load = env->src_stats.load;
1731 orig_dst_load = env->dst_stats.load;
1733 old_imb = abs(orig_dst_load * src_capacity - orig_src_load * dst_capacity);
1735 /* Would this change make things worse? */
1736 return (imb > old_imb);
1740 * Maximum NUMA importance can be 1998 (2*999);
1741 * SMALLIMP @ 30 would be close to 1998/64.
1742 * Used to deter task migration.
1747 * This checks if the overall compute and NUMA accesses of the system would
1748 * be improved if the source tasks was migrated to the target dst_cpu taking
1749 * into account that it might be best if task running on the dst_cpu should
1750 * be exchanged with the source task
1752 static bool task_numa_compare(struct task_numa_env *env,
1753 long taskimp, long groupimp, bool maymove)
1755 struct numa_group *cur_ng, *p_ng = deref_curr_numa_group(env->p);
1756 struct rq *dst_rq = cpu_rq(env->dst_cpu);
1757 long imp = p_ng ? groupimp : taskimp;
1758 struct task_struct *cur;
1759 long src_load, dst_load;
1760 int dist = env->dist;
1763 bool stopsearch = false;
1765 if (READ_ONCE(dst_rq->numa_migrate_on))
1769 cur = rcu_dereference(dst_rq->curr);
1770 if (cur && ((cur->flags & PF_EXITING) || is_idle_task(cur)))
1774 * Because we have preemption enabled we can get migrated around and
1775 * end try selecting ourselves (current == env->p) as a swap candidate.
1777 if (cur == env->p) {
1783 if (maymove && moveimp >= env->best_imp)
1789 /* Skip this swap candidate if cannot move to the source cpu. */
1790 if (!cpumask_test_cpu(env->src_cpu, cur->cpus_ptr))
1794 * Skip this swap candidate if it is not moving to its preferred
1795 * node and the best task is.
1797 if (env->best_task &&
1798 env->best_task->numa_preferred_nid == env->src_nid &&
1799 cur->numa_preferred_nid != env->src_nid) {
1804 * "imp" is the fault differential for the source task between the
1805 * source and destination node. Calculate the total differential for
1806 * the source task and potential destination task. The more negative
1807 * the value is, the more remote accesses that would be expected to
1808 * be incurred if the tasks were swapped.
1810 * If dst and source tasks are in the same NUMA group, or not
1811 * in any group then look only at task weights.
1813 cur_ng = rcu_dereference(cur->numa_group);
1814 if (cur_ng == p_ng) {
1816 * Do not swap within a group or between tasks that have
1817 * no group if there is spare capacity. Swapping does
1818 * not address the load imbalance and helps one task at
1819 * the cost of punishing another.
1821 if (env->dst_stats.node_type == node_has_spare)
1824 imp = taskimp + task_weight(cur, env->src_nid, dist) -
1825 task_weight(cur, env->dst_nid, dist);
1827 * Add some hysteresis to prevent swapping the
1828 * tasks within a group over tiny differences.
1834 * Compare the group weights. If a task is all by itself
1835 * (not part of a group), use the task weight instead.
1838 imp += group_weight(cur, env->src_nid, dist) -
1839 group_weight(cur, env->dst_nid, dist);
1841 imp += task_weight(cur, env->src_nid, dist) -
1842 task_weight(cur, env->dst_nid, dist);
1845 /* Discourage picking a task already on its preferred node */
1846 if (cur->numa_preferred_nid == env->dst_nid)
1850 * Encourage picking a task that moves to its preferred node.
1851 * This potentially makes imp larger than it's maximum of
1852 * 1998 (see SMALLIMP and task_weight for why) but in this
1853 * case, it does not matter.
1855 if (cur->numa_preferred_nid == env->src_nid)
1858 if (maymove && moveimp > imp && moveimp > env->best_imp) {
1865 * Prefer swapping with a task moving to its preferred node over a
1868 if (env->best_task && cur->numa_preferred_nid == env->src_nid &&
1869 env->best_task->numa_preferred_nid != env->src_nid) {
1874 * If the NUMA importance is less than SMALLIMP,
1875 * task migration might only result in ping pong
1876 * of tasks and also hurt performance due to cache
1879 if (imp < SMALLIMP || imp <= env->best_imp + SMALLIMP / 2)
1883 * In the overloaded case, try and keep the load balanced.
1885 load = task_h_load(env->p) - task_h_load(cur);
1889 dst_load = env->dst_stats.load + load;
1890 src_load = env->src_stats.load - load;
1892 if (load_too_imbalanced(src_load, dst_load, env))
1896 /* Evaluate an idle CPU for a task numa move. */
1898 int cpu = env->dst_stats.idle_cpu;
1900 /* Nothing cached so current CPU went idle since the search. */
1905 * If the CPU is no longer truly idle and the previous best CPU
1906 * is, keep using it.
1908 if (!idle_cpu(cpu) && env->best_cpu >= 0 &&
1909 idle_cpu(env->best_cpu)) {
1910 cpu = env->best_cpu;
1916 task_numa_assign(env, cur, imp);
1919 * If a move to idle is allowed because there is capacity or load
1920 * balance improves then stop the search. While a better swap
1921 * candidate may exist, a search is not free.
1923 if (maymove && !cur && env->best_cpu >= 0 && idle_cpu(env->best_cpu))
1927 * If a swap candidate must be identified and the current best task
1928 * moves its preferred node then stop the search.
1930 if (!maymove && env->best_task &&
1931 env->best_task->numa_preferred_nid == env->src_nid) {
1940 static void task_numa_find_cpu(struct task_numa_env *env,
1941 long taskimp, long groupimp)
1943 bool maymove = false;
1947 * If dst node has spare capacity, then check if there is an
1948 * imbalance that would be overruled by the load balancer.
1950 if (env->dst_stats.node_type == node_has_spare) {
1951 unsigned int imbalance;
1952 int src_running, dst_running;
1955 * Would movement cause an imbalance? Note that if src has
1956 * more running tasks that the imbalance is ignored as the
1957 * move improves the imbalance from the perspective of the
1958 * CPU load balancer.
1960 src_running = env->src_stats.nr_running - 1;
1961 dst_running = env->dst_stats.nr_running + 1;
1962 imbalance = max(0, dst_running - src_running);
1963 imbalance = adjust_numa_imbalance(imbalance, dst_running,
1966 /* Use idle CPU if there is no imbalance */
1969 if (env->dst_stats.idle_cpu >= 0) {
1970 env->dst_cpu = env->dst_stats.idle_cpu;
1971 task_numa_assign(env, NULL, 0);
1976 long src_load, dst_load, load;
1978 * If the improvement from just moving env->p direction is better
1979 * than swapping tasks around, check if a move is possible.
1981 load = task_h_load(env->p);
1982 dst_load = env->dst_stats.load + load;
1983 src_load = env->src_stats.load - load;
1984 maymove = !load_too_imbalanced(src_load, dst_load, env);
1987 for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) {
1988 /* Skip this CPU if the source task cannot migrate */
1989 if (!cpumask_test_cpu(cpu, env->p->cpus_ptr))
1993 if (task_numa_compare(env, taskimp, groupimp, maymove))
1998 static int task_numa_migrate(struct task_struct *p)
2000 struct task_numa_env env = {
2003 .src_cpu = task_cpu(p),
2004 .src_nid = task_node(p),
2006 .imbalance_pct = 112,
2012 unsigned long taskweight, groupweight;
2013 struct sched_domain *sd;
2014 long taskimp, groupimp;
2015 struct numa_group *ng;
2020 * Pick the lowest SD_NUMA domain, as that would have the smallest
2021 * imbalance and would be the first to start moving tasks about.
2023 * And we want to avoid any moving of tasks about, as that would create
2024 * random movement of tasks -- counter the numa conditions we're trying
2028 sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu));
2030 env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2;
2031 env.imb_numa_nr = sd->imb_numa_nr;
2036 * Cpusets can break the scheduler domain tree into smaller
2037 * balance domains, some of which do not cross NUMA boundaries.
2038 * Tasks that are "trapped" in such domains cannot be migrated
2039 * elsewhere, so there is no point in (re)trying.
2041 if (unlikely(!sd)) {
2042 sched_setnuma(p, task_node(p));
2046 env.dst_nid = p->numa_preferred_nid;
2047 dist = env.dist = node_distance(env.src_nid, env.dst_nid);
2048 taskweight = task_weight(p, env.src_nid, dist);
2049 groupweight = group_weight(p, env.src_nid, dist);
2050 update_numa_stats(&env, &env.src_stats, env.src_nid, false);
2051 taskimp = task_weight(p, env.dst_nid, dist) - taskweight;
2052 groupimp = group_weight(p, env.dst_nid, dist) - groupweight;
2053 update_numa_stats(&env, &env.dst_stats, env.dst_nid, true);
2055 /* Try to find a spot on the preferred nid. */
2056 task_numa_find_cpu(&env, taskimp, groupimp);
2059 * Look at other nodes in these cases:
2060 * - there is no space available on the preferred_nid
2061 * - the task is part of a numa_group that is interleaved across
2062 * multiple NUMA nodes; in order to better consolidate the group,
2063 * we need to check other locations.
2065 ng = deref_curr_numa_group(p);
2066 if (env.best_cpu == -1 || (ng && ng->active_nodes > 1)) {
2067 for_each_node_state(nid, N_CPU) {
2068 if (nid == env.src_nid || nid == p->numa_preferred_nid)
2071 dist = node_distance(env.src_nid, env.dst_nid);
2072 if (sched_numa_topology_type == NUMA_BACKPLANE &&
2074 taskweight = task_weight(p, env.src_nid, dist);
2075 groupweight = group_weight(p, env.src_nid, dist);
2078 /* Only consider nodes where both task and groups benefit */
2079 taskimp = task_weight(p, nid, dist) - taskweight;
2080 groupimp = group_weight(p, nid, dist) - groupweight;
2081 if (taskimp < 0 && groupimp < 0)
2086 update_numa_stats(&env, &env.dst_stats, env.dst_nid, true);
2087 task_numa_find_cpu(&env, taskimp, groupimp);
2092 * If the task is part of a workload that spans multiple NUMA nodes,
2093 * and is migrating into one of the workload's active nodes, remember
2094 * this node as the task's preferred numa node, so the workload can
2096 * A task that migrated to a second choice node will be better off
2097 * trying for a better one later. Do not set the preferred node here.
2100 if (env.best_cpu == -1)
2103 nid = cpu_to_node(env.best_cpu);
2105 if (nid != p->numa_preferred_nid)
2106 sched_setnuma(p, nid);
2109 /* No better CPU than the current one was found. */
2110 if (env.best_cpu == -1) {
2111 trace_sched_stick_numa(p, env.src_cpu, NULL, -1);
2115 best_rq = cpu_rq(env.best_cpu);
2116 if (env.best_task == NULL) {
2117 ret = migrate_task_to(p, env.best_cpu);
2118 WRITE_ONCE(best_rq->numa_migrate_on, 0);
2120 trace_sched_stick_numa(p, env.src_cpu, NULL, env.best_cpu);
2124 ret = migrate_swap(p, env.best_task, env.best_cpu, env.src_cpu);
2125 WRITE_ONCE(best_rq->numa_migrate_on, 0);
2128 trace_sched_stick_numa(p, env.src_cpu, env.best_task, env.best_cpu);
2129 put_task_struct(env.best_task);
2133 /* Attempt to migrate a task to a CPU on the preferred node. */
2134 static void numa_migrate_preferred(struct task_struct *p)
2136 unsigned long interval = HZ;
2138 /* This task has no NUMA fault statistics yet */
2139 if (unlikely(p->numa_preferred_nid == NUMA_NO_NODE || !p->numa_faults))
2142 /* Periodically retry migrating the task to the preferred node */
2143 interval = min(interval, msecs_to_jiffies(p->numa_scan_period) / 16);
2144 p->numa_migrate_retry = jiffies + interval;
2146 /* Success if task is already running on preferred CPU */
2147 if (task_node(p) == p->numa_preferred_nid)
2150 /* Otherwise, try migrate to a CPU on the preferred node */
2151 task_numa_migrate(p);
2155 * Find out how many nodes the workload is actively running on. Do this by
2156 * tracking the nodes from which NUMA hinting faults are triggered. This can
2157 * be different from the set of nodes where the workload's memory is currently
2160 static void numa_group_count_active_nodes(struct numa_group *numa_group)
2162 unsigned long faults, max_faults = 0;
2163 int nid, active_nodes = 0;
2165 for_each_node_state(nid, N_CPU) {
2166 faults = group_faults_cpu(numa_group, nid);
2167 if (faults > max_faults)
2168 max_faults = faults;
2171 for_each_node_state(nid, N_CPU) {
2172 faults = group_faults_cpu(numa_group, nid);
2173 if (faults * ACTIVE_NODE_FRACTION > max_faults)
2177 numa_group->max_faults_cpu = max_faults;
2178 numa_group->active_nodes = active_nodes;
2182 * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
2183 * increments. The more local the fault statistics are, the higher the scan
2184 * period will be for the next scan window. If local/(local+remote) ratio is
2185 * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS)
2186 * the scan period will decrease. Aim for 70% local accesses.
2188 #define NUMA_PERIOD_SLOTS 10
2189 #define NUMA_PERIOD_THRESHOLD 7
2192 * Increase the scan period (slow down scanning) if the majority of
2193 * our memory is already on our local node, or if the majority of
2194 * the page accesses are shared with other processes.
2195 * Otherwise, decrease the scan period.
2197 static void update_task_scan_period(struct task_struct *p,
2198 unsigned long shared, unsigned long private)
2200 unsigned int period_slot;
2201 int lr_ratio, ps_ratio;
2204 unsigned long remote = p->numa_faults_locality[0];
2205 unsigned long local = p->numa_faults_locality[1];
2208 * If there were no record hinting faults then either the task is
2209 * completely idle or all activity is in areas that are not of interest
2210 * to automatic numa balancing. Related to that, if there were failed
2211 * migration then it implies we are migrating too quickly or the local
2212 * node is overloaded. In either case, scan slower
2214 if (local + shared == 0 || p->numa_faults_locality[2]) {
2215 p->numa_scan_period = min(p->numa_scan_period_max,
2216 p->numa_scan_period << 1);
2218 p->mm->numa_next_scan = jiffies +
2219 msecs_to_jiffies(p->numa_scan_period);
2225 * Prepare to scale scan period relative to the current period.
2226 * == NUMA_PERIOD_THRESHOLD scan period stays the same
2227 * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
2228 * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
2230 period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS);
2231 lr_ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote);
2232 ps_ratio = (private * NUMA_PERIOD_SLOTS) / (private + shared);
2234 if (ps_ratio >= NUMA_PERIOD_THRESHOLD) {
2236 * Most memory accesses are local. There is no need to
2237 * do fast NUMA scanning, since memory is already local.
2239 int slot = ps_ratio - NUMA_PERIOD_THRESHOLD;
2242 diff = slot * period_slot;
2243 } else if (lr_ratio >= NUMA_PERIOD_THRESHOLD) {
2245 * Most memory accesses are shared with other tasks.
2246 * There is no point in continuing fast NUMA scanning,
2247 * since other tasks may just move the memory elsewhere.
2249 int slot = lr_ratio - NUMA_PERIOD_THRESHOLD;
2252 diff = slot * period_slot;
2255 * Private memory faults exceed (SLOTS-THRESHOLD)/SLOTS,
2256 * yet they are not on the local NUMA node. Speed up
2257 * NUMA scanning to get the memory moved over.
2259 int ratio = max(lr_ratio, ps_ratio);
2260 diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot;
2263 p->numa_scan_period = clamp(p->numa_scan_period + diff,
2264 task_scan_min(p), task_scan_max(p));
2265 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
2269 * Get the fraction of time the task has been running since the last
2270 * NUMA placement cycle. The scheduler keeps similar statistics, but
2271 * decays those on a 32ms period, which is orders of magnitude off
2272 * from the dozens-of-seconds NUMA balancing period. Use the scheduler
2273 * stats only if the task is so new there are no NUMA statistics yet.
2275 static u64 numa_get_avg_runtime(struct task_struct *p, u64 *period)
2277 u64 runtime, delta, now;
2278 /* Use the start of this time slice to avoid calculations. */
2279 now = p->se.exec_start;
2280 runtime = p->se.sum_exec_runtime;
2282 if (p->last_task_numa_placement) {
2283 delta = runtime - p->last_sum_exec_runtime;
2284 *period = now - p->last_task_numa_placement;
2286 /* Avoid time going backwards, prevent potential divide error: */
2287 if (unlikely((s64)*period < 0))
2290 delta = p->se.avg.load_sum;
2291 *period = LOAD_AVG_MAX;
2294 p->last_sum_exec_runtime = runtime;
2295 p->last_task_numa_placement = now;
2301 * Determine the preferred nid for a task in a numa_group. This needs to
2302 * be done in a way that produces consistent results with group_weight,
2303 * otherwise workloads might not converge.
2305 static int preferred_group_nid(struct task_struct *p, int nid)
2310 /* Direct connections between all NUMA nodes. */
2311 if (sched_numa_topology_type == NUMA_DIRECT)
2315 * On a system with glueless mesh NUMA topology, group_weight
2316 * scores nodes according to the number of NUMA hinting faults on
2317 * both the node itself, and on nearby nodes.
2319 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
2320 unsigned long score, max_score = 0;
2321 int node, max_node = nid;
2323 dist = sched_max_numa_distance;
2325 for_each_node_state(node, N_CPU) {
2326 score = group_weight(p, node, dist);
2327 if (score > max_score) {
2336 * Finding the preferred nid in a system with NUMA backplane
2337 * interconnect topology is more involved. The goal is to locate
2338 * tasks from numa_groups near each other in the system, and
2339 * untangle workloads from different sides of the system. This requires
2340 * searching down the hierarchy of node groups, recursively searching
2341 * inside the highest scoring group of nodes. The nodemask tricks
2342 * keep the complexity of the search down.
2344 nodes = node_states[N_CPU];
2345 for (dist = sched_max_numa_distance; dist > LOCAL_DISTANCE; dist--) {
2346 unsigned long max_faults = 0;
2347 nodemask_t max_group = NODE_MASK_NONE;
2350 /* Are there nodes at this distance from each other? */
2351 if (!find_numa_distance(dist))
2354 for_each_node_mask(a, nodes) {
2355 unsigned long faults = 0;
2356 nodemask_t this_group;
2357 nodes_clear(this_group);
2359 /* Sum group's NUMA faults; includes a==b case. */
2360 for_each_node_mask(b, nodes) {
2361 if (node_distance(a, b) < dist) {
2362 faults += group_faults(p, b);
2363 node_set(b, this_group);
2364 node_clear(b, nodes);
2368 /* Remember the top group. */
2369 if (faults > max_faults) {
2370 max_faults = faults;
2371 max_group = this_group;
2373 * subtle: at the smallest distance there is
2374 * just one node left in each "group", the
2375 * winner is the preferred nid.
2380 /* Next round, evaluate the nodes within max_group. */
2388 static void task_numa_placement(struct task_struct *p)
2390 int seq, nid, max_nid = NUMA_NO_NODE;
2391 unsigned long max_faults = 0;
2392 unsigned long fault_types[2] = { 0, 0 };
2393 unsigned long total_faults;
2394 u64 runtime, period;
2395 spinlock_t *group_lock = NULL;
2396 struct numa_group *ng;
2399 * The p->mm->numa_scan_seq field gets updated without
2400 * exclusive access. Use READ_ONCE() here to ensure
2401 * that the field is read in a single access:
2403 seq = READ_ONCE(p->mm->numa_scan_seq);
2404 if (p->numa_scan_seq == seq)
2406 p->numa_scan_seq = seq;
2407 p->numa_scan_period_max = task_scan_max(p);
2409 total_faults = p->numa_faults_locality[0] +
2410 p->numa_faults_locality[1];
2411 runtime = numa_get_avg_runtime(p, &period);
2413 /* If the task is part of a group prevent parallel updates to group stats */
2414 ng = deref_curr_numa_group(p);
2416 group_lock = &ng->lock;
2417 spin_lock_irq(group_lock);
2420 /* Find the node with the highest number of faults */
2421 for_each_online_node(nid) {
2422 /* Keep track of the offsets in numa_faults array */
2423 int mem_idx, membuf_idx, cpu_idx, cpubuf_idx;
2424 unsigned long faults = 0, group_faults = 0;
2427 for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) {
2428 long diff, f_diff, f_weight;
2430 mem_idx = task_faults_idx(NUMA_MEM, nid, priv);
2431 membuf_idx = task_faults_idx(NUMA_MEMBUF, nid, priv);
2432 cpu_idx = task_faults_idx(NUMA_CPU, nid, priv);
2433 cpubuf_idx = task_faults_idx(NUMA_CPUBUF, nid, priv);
2435 /* Decay existing window, copy faults since last scan */
2436 diff = p->numa_faults[membuf_idx] - p->numa_faults[mem_idx] / 2;
2437 fault_types[priv] += p->numa_faults[membuf_idx];
2438 p->numa_faults[membuf_idx] = 0;
2441 * Normalize the faults_from, so all tasks in a group
2442 * count according to CPU use, instead of by the raw
2443 * number of faults. Tasks with little runtime have
2444 * little over-all impact on throughput, and thus their
2445 * faults are less important.
2447 f_weight = div64_u64(runtime << 16, period + 1);
2448 f_weight = (f_weight * p->numa_faults[cpubuf_idx]) /
2450 f_diff = f_weight - p->numa_faults[cpu_idx] / 2;
2451 p->numa_faults[cpubuf_idx] = 0;
2453 p->numa_faults[mem_idx] += diff;
2454 p->numa_faults[cpu_idx] += f_diff;
2455 faults += p->numa_faults[mem_idx];
2456 p->total_numa_faults += diff;
2459 * safe because we can only change our own group
2461 * mem_idx represents the offset for a given
2462 * nid and priv in a specific region because it
2463 * is at the beginning of the numa_faults array.
2465 ng->faults[mem_idx] += diff;
2466 ng->faults[cpu_idx] += f_diff;
2467 ng->total_faults += diff;
2468 group_faults += ng->faults[mem_idx];
2473 if (faults > max_faults) {
2474 max_faults = faults;
2477 } else if (group_faults > max_faults) {
2478 max_faults = group_faults;
2483 /* Cannot migrate task to CPU-less node */
2484 if (max_nid != NUMA_NO_NODE && !node_state(max_nid, N_CPU)) {
2485 int near_nid = max_nid;
2486 int distance, near_distance = INT_MAX;
2488 for_each_node_state(nid, N_CPU) {
2489 distance = node_distance(max_nid, nid);
2490 if (distance < near_distance) {
2492 near_distance = distance;
2499 numa_group_count_active_nodes(ng);
2500 spin_unlock_irq(group_lock);
2501 max_nid = preferred_group_nid(p, max_nid);
2505 /* Set the new preferred node */
2506 if (max_nid != p->numa_preferred_nid)
2507 sched_setnuma(p, max_nid);
2510 update_task_scan_period(p, fault_types[0], fault_types[1]);
2513 static inline int get_numa_group(struct numa_group *grp)
2515 return refcount_inc_not_zero(&grp->refcount);
2518 static inline void put_numa_group(struct numa_group *grp)
2520 if (refcount_dec_and_test(&grp->refcount))
2521 kfree_rcu(grp, rcu);
2524 static void task_numa_group(struct task_struct *p, int cpupid, int flags,
2527 struct numa_group *grp, *my_grp;
2528 struct task_struct *tsk;
2530 int cpu = cpupid_to_cpu(cpupid);
2533 if (unlikely(!deref_curr_numa_group(p))) {
2534 unsigned int size = sizeof(struct numa_group) +
2535 NR_NUMA_HINT_FAULT_STATS *
2536 nr_node_ids * sizeof(unsigned long);
2538 grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
2542 refcount_set(&grp->refcount, 1);
2543 grp->active_nodes = 1;
2544 grp->max_faults_cpu = 0;
2545 spin_lock_init(&grp->lock);
2548 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2549 grp->faults[i] = p->numa_faults[i];
2551 grp->total_faults = p->total_numa_faults;
2554 rcu_assign_pointer(p->numa_group, grp);
2558 tsk = READ_ONCE(cpu_rq(cpu)->curr);
2560 if (!cpupid_match_pid(tsk, cpupid))
2563 grp = rcu_dereference(tsk->numa_group);
2567 my_grp = deref_curr_numa_group(p);
2572 * Only join the other group if its bigger; if we're the bigger group,
2573 * the other task will join us.
2575 if (my_grp->nr_tasks > grp->nr_tasks)
2579 * Tie-break on the grp address.
2581 if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp)
2584 /* Always join threads in the same process. */
2585 if (tsk->mm == current->mm)
2588 /* Simple filter to avoid false positives due to PID collisions */
2589 if (flags & TNF_SHARED)
2592 /* Update priv based on whether false sharing was detected */
2595 if (join && !get_numa_group(grp))
2603 BUG_ON(irqs_disabled());
2604 double_lock_irq(&my_grp->lock, &grp->lock);
2606 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) {
2607 my_grp->faults[i] -= p->numa_faults[i];
2608 grp->faults[i] += p->numa_faults[i];
2610 my_grp->total_faults -= p->total_numa_faults;
2611 grp->total_faults += p->total_numa_faults;
2616 spin_unlock(&my_grp->lock);
2617 spin_unlock_irq(&grp->lock);
2619 rcu_assign_pointer(p->numa_group, grp);
2621 put_numa_group(my_grp);
2630 * Get rid of NUMA statistics associated with a task (either current or dead).
2631 * If @final is set, the task is dead and has reached refcount zero, so we can
2632 * safely free all relevant data structures. Otherwise, there might be
2633 * concurrent reads from places like load balancing and procfs, and we should
2634 * reset the data back to default state without freeing ->numa_faults.
2636 void task_numa_free(struct task_struct *p, bool final)
2638 /* safe: p either is current or is being freed by current */
2639 struct numa_group *grp = rcu_dereference_raw(p->numa_group);
2640 unsigned long *numa_faults = p->numa_faults;
2641 unsigned long flags;
2648 spin_lock_irqsave(&grp->lock, flags);
2649 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2650 grp->faults[i] -= p->numa_faults[i];
2651 grp->total_faults -= p->total_numa_faults;
2654 spin_unlock_irqrestore(&grp->lock, flags);
2655 RCU_INIT_POINTER(p->numa_group, NULL);
2656 put_numa_group(grp);
2660 p->numa_faults = NULL;
2663 p->total_numa_faults = 0;
2664 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2670 * Got a PROT_NONE fault for a page on @node.
2672 void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags)
2674 struct task_struct *p = current;
2675 bool migrated = flags & TNF_MIGRATED;
2676 int cpu_node = task_node(current);
2677 int local = !!(flags & TNF_FAULT_LOCAL);
2678 struct numa_group *ng;
2681 if (!static_branch_likely(&sched_numa_balancing))
2684 /* for example, ksmd faulting in a user's mm */
2688 /* Allocate buffer to track faults on a per-node basis */
2689 if (unlikely(!p->numa_faults)) {
2690 int size = sizeof(*p->numa_faults) *
2691 NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids;
2693 p->numa_faults = kzalloc(size, GFP_KERNEL|__GFP_NOWARN);
2694 if (!p->numa_faults)
2697 p->total_numa_faults = 0;
2698 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
2702 * First accesses are treated as private, otherwise consider accesses
2703 * to be private if the accessing pid has not changed
2705 if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) {
2708 priv = cpupid_match_pid(p, last_cpupid);
2709 if (!priv && !(flags & TNF_NO_GROUP))
2710 task_numa_group(p, last_cpupid, flags, &priv);
2714 * If a workload spans multiple NUMA nodes, a shared fault that
2715 * occurs wholly within the set of nodes that the workload is
2716 * actively using should be counted as local. This allows the
2717 * scan rate to slow down when a workload has settled down.
2719 ng = deref_curr_numa_group(p);
2720 if (!priv && !local && ng && ng->active_nodes > 1 &&
2721 numa_is_active_node(cpu_node, ng) &&
2722 numa_is_active_node(mem_node, ng))
2726 * Retry to migrate task to preferred node periodically, in case it
2727 * previously failed, or the scheduler moved us.
2729 if (time_after(jiffies, p->numa_migrate_retry)) {
2730 task_numa_placement(p);
2731 numa_migrate_preferred(p);
2735 p->numa_pages_migrated += pages;
2736 if (flags & TNF_MIGRATE_FAIL)
2737 p->numa_faults_locality[2] += pages;
2739 p->numa_faults[task_faults_idx(NUMA_MEMBUF, mem_node, priv)] += pages;
2740 p->numa_faults[task_faults_idx(NUMA_CPUBUF, cpu_node, priv)] += pages;
2741 p->numa_faults_locality[local] += pages;
2744 static void reset_ptenuma_scan(struct task_struct *p)
2747 * We only did a read acquisition of the mmap sem, so
2748 * p->mm->numa_scan_seq is written to without exclusive access
2749 * and the update is not guaranteed to be atomic. That's not
2750 * much of an issue though, since this is just used for
2751 * statistical sampling. Use READ_ONCE/WRITE_ONCE, which are not
2752 * expensive, to avoid any form of compiler optimizations:
2754 WRITE_ONCE(p->mm->numa_scan_seq, READ_ONCE(p->mm->numa_scan_seq) + 1);
2755 p->mm->numa_scan_offset = 0;
2759 * The expensive part of numa migration is done from task_work context.
2760 * Triggered from task_tick_numa().
2762 static void task_numa_work(struct callback_head *work)
2764 unsigned long migrate, next_scan, now = jiffies;
2765 struct task_struct *p = current;
2766 struct mm_struct *mm = p->mm;
2767 u64 runtime = p->se.sum_exec_runtime;
2768 struct vm_area_struct *vma;
2769 unsigned long start, end;
2770 unsigned long nr_pte_updates = 0;
2771 long pages, virtpages;
2773 SCHED_WARN_ON(p != container_of(work, struct task_struct, numa_work));
2777 * Who cares about NUMA placement when they're dying.
2779 * NOTE: make sure not to dereference p->mm before this check,
2780 * exit_task_work() happens _after_ exit_mm() so we could be called
2781 * without p->mm even though we still had it when we enqueued this
2784 if (p->flags & PF_EXITING)
2787 if (!mm->numa_next_scan) {
2788 mm->numa_next_scan = now +
2789 msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2793 * Enforce maximal scan/migration frequency..
2795 migrate = mm->numa_next_scan;
2796 if (time_before(now, migrate))
2799 if (p->numa_scan_period == 0) {
2800 p->numa_scan_period_max = task_scan_max(p);
2801 p->numa_scan_period = task_scan_start(p);
2804 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
2805 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
2809 * Delay this task enough that another task of this mm will likely win
2810 * the next time around.
2812 p->node_stamp += 2 * TICK_NSEC;
2814 start = mm->numa_scan_offset;
2815 pages = sysctl_numa_balancing_scan_size;
2816 pages <<= 20 - PAGE_SHIFT; /* MB in pages */
2817 virtpages = pages * 8; /* Scan up to this much virtual space */
2822 if (!mmap_read_trylock(mm))
2824 vma = find_vma(mm, start);
2826 reset_ptenuma_scan(p);
2830 for (; vma; vma = vma->vm_next) {
2831 if (!vma_migratable(vma) || !vma_policy_mof(vma) ||
2832 is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_MIXEDMAP)) {
2837 * Shared library pages mapped by multiple processes are not
2838 * migrated as it is expected they are cache replicated. Avoid
2839 * hinting faults in read-only file-backed mappings or the vdso
2840 * as migrating the pages will be of marginal benefit.
2843 (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ)))
2847 * Skip inaccessible VMAs to avoid any confusion between
2848 * PROT_NONE and NUMA hinting ptes
2850 if (!vma_is_accessible(vma))
2854 start = max(start, vma->vm_start);
2855 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
2856 end = min(end, vma->vm_end);
2857 nr_pte_updates = change_prot_numa(vma, start, end);
2860 * Try to scan sysctl_numa_balancing_size worth of
2861 * hpages that have at least one present PTE that
2862 * is not already pte-numa. If the VMA contains
2863 * areas that are unused or already full of prot_numa
2864 * PTEs, scan up to virtpages, to skip through those
2868 pages -= (end - start) >> PAGE_SHIFT;
2869 virtpages -= (end - start) >> PAGE_SHIFT;
2872 if (pages <= 0 || virtpages <= 0)
2876 } while (end != vma->vm_end);
2881 * It is possible to reach the end of the VMA list but the last few
2882 * VMAs are not guaranteed to the vma_migratable. If they are not, we
2883 * would find the !migratable VMA on the next scan but not reset the
2884 * scanner to the start so check it now.
2887 mm->numa_scan_offset = start;
2889 reset_ptenuma_scan(p);
2890 mmap_read_unlock(mm);
2893 * Make sure tasks use at least 32x as much time to run other code
2894 * than they used here, to limit NUMA PTE scanning overhead to 3% max.
2895 * Usually update_task_scan_period slows down scanning enough; on an
2896 * overloaded system we need to limit overhead on a per task basis.
2898 if (unlikely(p->se.sum_exec_runtime != runtime)) {
2899 u64 diff = p->se.sum_exec_runtime - runtime;
2900 p->node_stamp += 32 * diff;
2904 void init_numa_balancing(unsigned long clone_flags, struct task_struct *p)
2907 struct mm_struct *mm = p->mm;
2910 mm_users = atomic_read(&mm->mm_users);
2911 if (mm_users == 1) {
2912 mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2913 mm->numa_scan_seq = 0;
2917 p->numa_scan_seq = mm ? mm->numa_scan_seq : 0;
2918 p->numa_scan_period = sysctl_numa_balancing_scan_delay;
2919 p->numa_migrate_retry = 0;
2920 /* Protect against double add, see task_tick_numa and task_numa_work */
2921 p->numa_work.next = &p->numa_work;
2922 p->numa_faults = NULL;
2923 p->numa_pages_migrated = 0;
2924 p->total_numa_faults = 0;
2925 RCU_INIT_POINTER(p->numa_group, NULL);
2926 p->last_task_numa_placement = 0;
2927 p->last_sum_exec_runtime = 0;
2929 init_task_work(&p->numa_work, task_numa_work);
2931 /* New address space, reset the preferred nid */
2932 if (!(clone_flags & CLONE_VM)) {
2933 p->numa_preferred_nid = NUMA_NO_NODE;
2938 * New thread, keep existing numa_preferred_nid which should be copied
2939 * already by arch_dup_task_struct but stagger when scans start.
2944 delay = min_t(unsigned int, task_scan_max(current),
2945 current->numa_scan_period * mm_users * NSEC_PER_MSEC);
2946 delay += 2 * TICK_NSEC;
2947 p->node_stamp = delay;
2952 * Drive the periodic memory faults..
2954 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
2956 struct callback_head *work = &curr->numa_work;
2960 * We don't care about NUMA placement if we don't have memory.
2962 if (!curr->mm || (curr->flags & (PF_EXITING | PF_KTHREAD)) || work->next != work)
2966 * Using runtime rather than walltime has the dual advantage that
2967 * we (mostly) drive the selection from busy threads and that the
2968 * task needs to have done some actual work before we bother with
2971 now = curr->se.sum_exec_runtime;
2972 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
2974 if (now > curr->node_stamp + period) {
2975 if (!curr->node_stamp)
2976 curr->numa_scan_period = task_scan_start(curr);
2977 curr->node_stamp += period;
2979 if (!time_before(jiffies, curr->mm->numa_next_scan))
2980 task_work_add(curr, work, TWA_RESUME);
2984 static void update_scan_period(struct task_struct *p, int new_cpu)
2986 int src_nid = cpu_to_node(task_cpu(p));
2987 int dst_nid = cpu_to_node(new_cpu);
2989 if (!static_branch_likely(&sched_numa_balancing))
2992 if (!p->mm || !p->numa_faults || (p->flags & PF_EXITING))
2995 if (src_nid == dst_nid)
2999 * Allow resets if faults have been trapped before one scan
3000 * has completed. This is most likely due to a new task that
3001 * is pulled cross-node due to wakeups or load balancing.
3003 if (p->numa_scan_seq) {
3005 * Avoid scan adjustments if moving to the preferred
3006 * node or if the task was not previously running on
3007 * the preferred node.
3009 if (dst_nid == p->numa_preferred_nid ||
3010 (p->numa_preferred_nid != NUMA_NO_NODE &&
3011 src_nid != p->numa_preferred_nid))
3015 p->numa_scan_period = task_scan_start(p);
3019 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
3023 static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p)
3027 static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p)
3031 static inline void update_scan_period(struct task_struct *p, int new_cpu)
3035 #endif /* CONFIG_NUMA_BALANCING */
3038 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
3040 update_load_add(&cfs_rq->load, se->load.weight);
3042 if (entity_is_task(se)) {
3043 struct rq *rq = rq_of(cfs_rq);
3045 account_numa_enqueue(rq, task_of(se));
3046 list_add(&se->group_node, &rq->cfs_tasks);
3049 cfs_rq->nr_running++;
3051 cfs_rq->idle_nr_running++;
3055 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
3057 update_load_sub(&cfs_rq->load, se->load.weight);
3059 if (entity_is_task(se)) {
3060 account_numa_dequeue(rq_of(cfs_rq), task_of(se));
3061 list_del_init(&se->group_node);
3064 cfs_rq->nr_running--;
3066 cfs_rq->idle_nr_running--;
3070 * Signed add and clamp on underflow.
3072 * Explicitly do a load-store to ensure the intermediate value never hits
3073 * memory. This allows lockless observations without ever seeing the negative
3076 #define add_positive(_ptr, _val) do { \
3077 typeof(_ptr) ptr = (_ptr); \
3078 typeof(_val) val = (_val); \
3079 typeof(*ptr) res, var = READ_ONCE(*ptr); \
3083 if (val < 0 && res > var) \
3086 WRITE_ONCE(*ptr, res); \
3090 * Unsigned subtract and clamp on underflow.
3092 * Explicitly do a load-store to ensure the intermediate value never hits
3093 * memory. This allows lockless observations without ever seeing the negative
3096 #define sub_positive(_ptr, _val) do { \
3097 typeof(_ptr) ptr = (_ptr); \
3098 typeof(*ptr) val = (_val); \
3099 typeof(*ptr) res, var = READ_ONCE(*ptr); \
3103 WRITE_ONCE(*ptr, res); \
3107 * Remove and clamp on negative, from a local variable.
3109 * A variant of sub_positive(), which does not use explicit load-store
3110 * and is thus optimized for local variable updates.
3112 #define lsub_positive(_ptr, _val) do { \
3113 typeof(_ptr) ptr = (_ptr); \
3114 *ptr -= min_t(typeof(*ptr), *ptr, _val); \
3119 enqueue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3121 cfs_rq->avg.load_avg += se->avg.load_avg;
3122 cfs_rq->avg.load_sum += se_weight(se) * se->avg.load_sum;
3126 dequeue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3128 sub_positive(&cfs_rq->avg.load_avg, se->avg.load_avg);
3129 sub_positive(&cfs_rq->avg.load_sum, se_weight(se) * se->avg.load_sum);
3130 /* See update_cfs_rq_load_avg() */
3131 cfs_rq->avg.load_sum = max_t(u32, cfs_rq->avg.load_sum,
3132 cfs_rq->avg.load_avg * PELT_MIN_DIVIDER);
3136 enqueue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
3138 dequeue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
3141 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
3142 unsigned long weight)
3145 /* commit outstanding execution time */
3146 if (cfs_rq->curr == se)
3147 update_curr(cfs_rq);
3148 update_load_sub(&cfs_rq->load, se->load.weight);
3150 dequeue_load_avg(cfs_rq, se);
3152 update_load_set(&se->load, weight);
3156 u32 divider = get_pelt_divider(&se->avg);
3158 se->avg.load_avg = div_u64(se_weight(se) * se->avg.load_sum, divider);
3162 enqueue_load_avg(cfs_rq, se);
3164 update_load_add(&cfs_rq->load, se->load.weight);
3168 void reweight_task(struct task_struct *p, int prio)
3170 struct sched_entity *se = &p->se;
3171 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3172 struct load_weight *load = &se->load;
3173 unsigned long weight = scale_load(sched_prio_to_weight[prio]);
3175 reweight_entity(cfs_rq, se, weight);
3176 load->inv_weight = sched_prio_to_wmult[prio];
3179 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
3181 #ifdef CONFIG_FAIR_GROUP_SCHED
3184 * All this does is approximate the hierarchical proportion which includes that
3185 * global sum we all love to hate.
3187 * That is, the weight of a group entity, is the proportional share of the
3188 * group weight based on the group runqueue weights. That is:
3190 * tg->weight * grq->load.weight
3191 * ge->load.weight = ----------------------------- (1)
3192 * \Sum grq->load.weight
3194 * Now, because computing that sum is prohibitively expensive to compute (been
3195 * there, done that) we approximate it with this average stuff. The average
3196 * moves slower and therefore the approximation is cheaper and more stable.
3198 * So instead of the above, we substitute:
3200 * grq->load.weight -> grq->avg.load_avg (2)
3202 * which yields the following:
3204 * tg->weight * grq->avg.load_avg
3205 * ge->load.weight = ------------------------------ (3)
3208 * Where: tg->load_avg ~= \Sum grq->avg.load_avg
3210 * That is shares_avg, and it is right (given the approximation (2)).
3212 * The problem with it is that because the average is slow -- it was designed
3213 * to be exactly that of course -- this leads to transients in boundary
3214 * conditions. In specific, the case where the group was idle and we start the
3215 * one task. It takes time for our CPU's grq->avg.load_avg to build up,
3216 * yielding bad latency etc..
3218 * Now, in that special case (1) reduces to:
3220 * tg->weight * grq->load.weight
3221 * ge->load.weight = ----------------------------- = tg->weight (4)
3224 * That is, the sum collapses because all other CPUs are idle; the UP scenario.
3226 * So what we do is modify our approximation (3) to approach (4) in the (near)
3231 * tg->weight * grq->load.weight
3232 * --------------------------------------------------- (5)
3233 * tg->load_avg - grq->avg.load_avg + grq->load.weight
3235 * But because grq->load.weight can drop to 0, resulting in a divide by zero,
3236 * we need to use grq->avg.load_avg as its lower bound, which then gives:
3239 * tg->weight * grq->load.weight
3240 * ge->load.weight = ----------------------------- (6)
3245 * tg_load_avg' = tg->load_avg - grq->avg.load_avg +
3246 * max(grq->load.weight, grq->avg.load_avg)
3248 * And that is shares_weight and is icky. In the (near) UP case it approaches
3249 * (4) while in the normal case it approaches (3). It consistently
3250 * overestimates the ge->load.weight and therefore:
3252 * \Sum ge->load.weight >= tg->weight
3256 static long calc_group_shares(struct cfs_rq *cfs_rq)
3258 long tg_weight, tg_shares, load, shares;
3259 struct task_group *tg = cfs_rq->tg;
3261 tg_shares = READ_ONCE(tg->shares);
3263 load = max(scale_load_down(cfs_rq->load.weight), cfs_rq->avg.load_avg);
3265 tg_weight = atomic_long_read(&tg->load_avg);
3267 /* Ensure tg_weight >= load */
3268 tg_weight -= cfs_rq->tg_load_avg_contrib;
3271 shares = (tg_shares * load);
3273 shares /= tg_weight;
3276 * MIN_SHARES has to be unscaled here to support per-CPU partitioning
3277 * of a group with small tg->shares value. It is a floor value which is
3278 * assigned as a minimum load.weight to the sched_entity representing
3279 * the group on a CPU.
3281 * E.g. on 64-bit for a group with tg->shares of scale_load(15)=15*1024
3282 * on an 8-core system with 8 tasks each runnable on one CPU shares has
3283 * to be 15*1024*1/8=1920 instead of scale_load(MIN_SHARES)=2*1024. In
3284 * case no task is runnable on a CPU MIN_SHARES=2 should be returned
3287 return clamp_t(long, shares, MIN_SHARES, tg_shares);
3289 #endif /* CONFIG_SMP */
3292 * Recomputes the group entity based on the current state of its group
3295 static void update_cfs_group(struct sched_entity *se)
3297 struct cfs_rq *gcfs_rq = group_cfs_rq(se);
3303 if (throttled_hierarchy(gcfs_rq))
3307 shares = READ_ONCE(gcfs_rq->tg->shares);
3309 if (likely(se->load.weight == shares))
3312 shares = calc_group_shares(gcfs_rq);
3315 reweight_entity(cfs_rq_of(se), se, shares);
3318 #else /* CONFIG_FAIR_GROUP_SCHED */
3319 static inline void update_cfs_group(struct sched_entity *se)
3322 #endif /* CONFIG_FAIR_GROUP_SCHED */
3324 static inline void cfs_rq_util_change(struct cfs_rq *cfs_rq, int flags)
3326 struct rq *rq = rq_of(cfs_rq);
3328 if (&rq->cfs == cfs_rq) {
3330 * There are a few boundary cases this might miss but it should
3331 * get called often enough that that should (hopefully) not be
3334 * It will not get called when we go idle, because the idle
3335 * thread is a different class (!fair), nor will the utilization
3336 * number include things like RT tasks.
3338 * As is, the util number is not freq-invariant (we'd have to
3339 * implement arch_scale_freq_capacity() for that).
3341 * See cpu_util_cfs().
3343 cpufreq_update_util(rq, flags);
3348 static inline bool load_avg_is_decayed(struct sched_avg *sa)
3356 if (sa->runnable_sum)
3360 * _avg must be null when _sum are null because _avg = _sum / divider
3361 * Make sure that rounding and/or propagation of PELT values never
3364 SCHED_WARN_ON(sa->load_avg ||
3371 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
3373 return u64_u32_load_copy(cfs_rq->avg.last_update_time,
3374 cfs_rq->last_update_time_copy);
3376 #ifdef CONFIG_FAIR_GROUP_SCHED
3378 * Because list_add_leaf_cfs_rq always places a child cfs_rq on the list
3379 * immediately before a parent cfs_rq, and cfs_rqs are removed from the list
3380 * bottom-up, we only have to test whether the cfs_rq before us on the list
3382 * If cfs_rq is not on the list, test whether a child needs its to be added to
3383 * connect a branch to the tree * (see list_add_leaf_cfs_rq() for details).
3385 static inline bool child_cfs_rq_on_list(struct cfs_rq *cfs_rq)
3387 struct cfs_rq *prev_cfs_rq;
3388 struct list_head *prev;
3390 if (cfs_rq->on_list) {
3391 prev = cfs_rq->leaf_cfs_rq_list.prev;
3393 struct rq *rq = rq_of(cfs_rq);
3395 prev = rq->tmp_alone_branch;
3398 prev_cfs_rq = container_of(prev, struct cfs_rq, leaf_cfs_rq_list);
3400 return (prev_cfs_rq->tg->parent == cfs_rq->tg);
3403 static inline bool cfs_rq_is_decayed(struct cfs_rq *cfs_rq)
3405 if (cfs_rq->load.weight)
3408 if (!load_avg_is_decayed(&cfs_rq->avg))
3411 if (child_cfs_rq_on_list(cfs_rq))
3418 * update_tg_load_avg - update the tg's load avg
3419 * @cfs_rq: the cfs_rq whose avg changed
3421 * This function 'ensures': tg->load_avg := \Sum tg->cfs_rq[]->avg.load.
3422 * However, because tg->load_avg is a global value there are performance
3425 * In order to avoid having to look at the other cfs_rq's, we use a
3426 * differential update where we store the last value we propagated. This in
3427 * turn allows skipping updates if the differential is 'small'.
3429 * Updating tg's load_avg is necessary before update_cfs_share().
3431 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq)
3433 long delta = cfs_rq->avg.load_avg - cfs_rq->tg_load_avg_contrib;
3436 * No need to update load_avg for root_task_group as it is not used.
3438 if (cfs_rq->tg == &root_task_group)
3441 if (abs(delta) > cfs_rq->tg_load_avg_contrib / 64) {
3442 atomic_long_add(delta, &cfs_rq->tg->load_avg);
3443 cfs_rq->tg_load_avg_contrib = cfs_rq->avg.load_avg;
3448 * Called within set_task_rq() right before setting a task's CPU. The
3449 * caller only guarantees p->pi_lock is held; no other assumptions,
3450 * including the state of rq->lock, should be made.
3452 void set_task_rq_fair(struct sched_entity *se,
3453 struct cfs_rq *prev, struct cfs_rq *next)
3455 u64 p_last_update_time;
3456 u64 n_last_update_time;
3458 if (!sched_feat(ATTACH_AGE_LOAD))
3462 * We are supposed to update the task to "current" time, then its up to
3463 * date and ready to go to new CPU/cfs_rq. But we have difficulty in
3464 * getting what current time is, so simply throw away the out-of-date
3465 * time. This will result in the wakee task is less decayed, but giving
3466 * the wakee more load sounds not bad.
3468 if (!(se->avg.last_update_time && prev))
3471 p_last_update_time = cfs_rq_last_update_time(prev);
3472 n_last_update_time = cfs_rq_last_update_time(next);
3474 __update_load_avg_blocked_se(p_last_update_time, se);
3475 se->avg.last_update_time = n_last_update_time;
3479 * When on migration a sched_entity joins/leaves the PELT hierarchy, we need to
3480 * propagate its contribution. The key to this propagation is the invariant
3481 * that for each group:
3483 * ge->avg == grq->avg (1)
3485 * _IFF_ we look at the pure running and runnable sums. Because they
3486 * represent the very same entity, just at different points in the hierarchy.
3488 * Per the above update_tg_cfs_util() and update_tg_cfs_runnable() are trivial
3489 * and simply copies the running/runnable sum over (but still wrong, because
3490 * the group entity and group rq do not have their PELT windows aligned).
3492 * However, update_tg_cfs_load() is more complex. So we have:
3494 * ge->avg.load_avg = ge->load.weight * ge->avg.runnable_avg (2)
3496 * And since, like util, the runnable part should be directly transferable,
3497 * the following would _appear_ to be the straight forward approach:
3499 * grq->avg.load_avg = grq->load.weight * grq->avg.runnable_avg (3)
3501 * And per (1) we have:
3503 * ge->avg.runnable_avg == grq->avg.runnable_avg
3507 * ge->load.weight * grq->avg.load_avg
3508 * ge->avg.load_avg = ----------------------------------- (4)
3511 * Except that is wrong!
3513 * Because while for entities historical weight is not important and we
3514 * really only care about our future and therefore can consider a pure
3515 * runnable sum, runqueues can NOT do this.
3517 * We specifically want runqueues to have a load_avg that includes
3518 * historical weights. Those represent the blocked load, the load we expect
3519 * to (shortly) return to us. This only works by keeping the weights as
3520 * integral part of the sum. We therefore cannot decompose as per (3).
3522 * Another reason this doesn't work is that runnable isn't a 0-sum entity.
3523 * Imagine a rq with 2 tasks that each are runnable 2/3 of the time. Then the
3524 * rq itself is runnable anywhere between 2/3 and 1 depending on how the
3525 * runnable section of these tasks overlap (or not). If they were to perfectly
3526 * align the rq as a whole would be runnable 2/3 of the time. If however we
3527 * always have at least 1 runnable task, the rq as a whole is always runnable.
3529 * So we'll have to approximate.. :/
3531 * Given the constraint:
3533 * ge->avg.running_sum <= ge->avg.runnable_sum <= LOAD_AVG_MAX
3535 * We can construct a rule that adds runnable to a rq by assuming minimal
3538 * On removal, we'll assume each task is equally runnable; which yields:
3540 * grq->avg.runnable_sum = grq->avg.load_sum / grq->load.weight
3542 * XXX: only do this for the part of runnable > running ?
3546 update_tg_cfs_util(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq)
3548 long delta_sum, delta_avg = gcfs_rq->avg.util_avg - se->avg.util_avg;
3549 u32 new_sum, divider;
3551 /* Nothing to update */
3556 * cfs_rq->avg.period_contrib can be used for both cfs_rq and se.
3557 * See ___update_load_avg() for details.
3559 divider = get_pelt_divider(&cfs_rq->avg);
3562 /* Set new sched_entity's utilization */
3563 se->avg.util_avg = gcfs_rq->avg.util_avg;
3564 new_sum = se->avg.util_avg * divider;
3565 delta_sum = (long)new_sum - (long)se->avg.util_sum;
3566 se->avg.util_sum = new_sum;
3568 /* Update parent cfs_rq utilization */
3569 add_positive(&cfs_rq->avg.util_avg, delta_avg);
3570 add_positive(&cfs_rq->avg.util_sum, delta_sum);
3572 /* See update_cfs_rq_load_avg() */
3573 cfs_rq->avg.util_sum = max_t(u32, cfs_rq->avg.util_sum,
3574 cfs_rq->avg.util_avg * PELT_MIN_DIVIDER);
3578 update_tg_cfs_runnable(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq)
3580 long delta_sum, delta_avg = gcfs_rq->avg.runnable_avg - se->avg.runnable_avg;
3581 u32 new_sum, divider;
3583 /* Nothing to update */
3588 * cfs_rq->avg.period_contrib can be used for both cfs_rq and se.
3589 * See ___update_load_avg() for details.
3591 divider = get_pelt_divider(&cfs_rq->avg);
3593 /* Set new sched_entity's runnable */
3594 se->avg.runnable_avg = gcfs_rq->avg.runnable_avg;
3595 new_sum = se->avg.runnable_avg * divider;
3596 delta_sum = (long)new_sum - (long)se->avg.runnable_sum;
3597 se->avg.runnable_sum = new_sum;
3599 /* Update parent cfs_rq runnable */
3600 add_positive(&cfs_rq->avg.runnable_avg, delta_avg);
3601 add_positive(&cfs_rq->avg.runnable_sum, delta_sum);
3602 /* See update_cfs_rq_load_avg() */
3603 cfs_rq->avg.runnable_sum = max_t(u32, cfs_rq->avg.runnable_sum,
3604 cfs_rq->avg.runnable_avg * PELT_MIN_DIVIDER);
3608 update_tg_cfs_load(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq)
3610 long delta_avg, running_sum, runnable_sum = gcfs_rq->prop_runnable_sum;
3611 unsigned long load_avg;
3619 gcfs_rq->prop_runnable_sum = 0;
3622 * cfs_rq->avg.period_contrib can be used for both cfs_rq and se.
3623 * See ___update_load_avg() for details.
3625 divider = get_pelt_divider(&cfs_rq->avg);
3627 if (runnable_sum >= 0) {
3629 * Add runnable; clip at LOAD_AVG_MAX. Reflects that until
3630 * the CPU is saturated running == runnable.
3632 runnable_sum += se->avg.load_sum;
3633 runnable_sum = min_t(long, runnable_sum, divider);
3636 * Estimate the new unweighted runnable_sum of the gcfs_rq by
3637 * assuming all tasks are equally runnable.
3639 if (scale_load_down(gcfs_rq->load.weight)) {
3640 load_sum = div_u64(gcfs_rq->avg.load_sum,
3641 scale_load_down(gcfs_rq->load.weight));
3644 /* But make sure to not inflate se's runnable */
3645 runnable_sum = min(se->avg.load_sum, load_sum);
3649 * runnable_sum can't be lower than running_sum
3650 * Rescale running sum to be in the same range as runnable sum
3651 * running_sum is in [0 : LOAD_AVG_MAX << SCHED_CAPACITY_SHIFT]
3652 * runnable_sum is in [0 : LOAD_AVG_MAX]
3654 running_sum = se->avg.util_sum >> SCHED_CAPACITY_SHIFT;
3655 runnable_sum = max(runnable_sum, running_sum);
3657 load_sum = se_weight(se) * runnable_sum;
3658 load_avg = div_u64(load_sum, divider);
3660 delta_avg = load_avg - se->avg.load_avg;
3664 delta_sum = load_sum - (s64)se_weight(se) * se->avg.load_sum;
3666 se->avg.load_sum = runnable_sum;
3667 se->avg.load_avg = load_avg;
3668 add_positive(&cfs_rq->avg.load_avg, delta_avg);
3669 add_positive(&cfs_rq->avg.load_sum, delta_sum);
3670 /* See update_cfs_rq_load_avg() */
3671 cfs_rq->avg.load_sum = max_t(u32, cfs_rq->avg.load_sum,
3672 cfs_rq->avg.load_avg * PELT_MIN_DIVIDER);
3675 static inline void add_tg_cfs_propagate(struct cfs_rq *cfs_rq, long runnable_sum)
3677 cfs_rq->propagate = 1;
3678 cfs_rq->prop_runnable_sum += runnable_sum;
3681 /* Update task and its cfs_rq load average */
3682 static inline int propagate_entity_load_avg(struct sched_entity *se)
3684 struct cfs_rq *cfs_rq, *gcfs_rq;
3686 if (entity_is_task(se))
3689 gcfs_rq = group_cfs_rq(se);
3690 if (!gcfs_rq->propagate)
3693 gcfs_rq->propagate = 0;
3695 cfs_rq = cfs_rq_of(se);
3697 add_tg_cfs_propagate(cfs_rq, gcfs_rq->prop_runnable_sum);
3699 update_tg_cfs_util(cfs_rq, se, gcfs_rq);
3700 update_tg_cfs_runnable(cfs_rq, se, gcfs_rq);
3701 update_tg_cfs_load(cfs_rq, se, gcfs_rq);
3703 trace_pelt_cfs_tp(cfs_rq);
3704 trace_pelt_se_tp(se);
3710 * Check if we need to update the load and the utilization of a blocked
3713 static inline bool skip_blocked_update(struct sched_entity *se)
3715 struct cfs_rq *gcfs_rq = group_cfs_rq(se);
3718 * If sched_entity still have not zero load or utilization, we have to
3721 if (se->avg.load_avg || se->avg.util_avg)
3725 * If there is a pending propagation, we have to update the load and
3726 * the utilization of the sched_entity:
3728 if (gcfs_rq->propagate)
3732 * Otherwise, the load and the utilization of the sched_entity is
3733 * already zero and there is no pending propagation, so it will be a
3734 * waste of time to try to decay it:
3739 #else /* CONFIG_FAIR_GROUP_SCHED */
3741 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq) {}
3743 static inline int propagate_entity_load_avg(struct sched_entity *se)
3748 static inline void add_tg_cfs_propagate(struct cfs_rq *cfs_rq, long runnable_sum) {}
3750 #endif /* CONFIG_FAIR_GROUP_SCHED */
3752 #ifdef CONFIG_NO_HZ_COMMON
3753 static inline void migrate_se_pelt_lag(struct sched_entity *se)
3755 u64 throttled = 0, now, lut;
3756 struct cfs_rq *cfs_rq;
3760 if (load_avg_is_decayed(&se->avg))
3763 cfs_rq = cfs_rq_of(se);
3767 is_idle = is_idle_task(rcu_dereference(rq->curr));
3771 * The lag estimation comes with a cost we don't want to pay all the
3772 * time. Hence, limiting to the case where the source CPU is idle and
3773 * we know we are at the greatest risk to have an outdated clock.
3779 * Estimated "now" is: last_update_time + cfs_idle_lag + rq_idle_lag, where:
3781 * last_update_time (the cfs_rq's last_update_time)
3782 * = cfs_rq_clock_pelt()@cfs_rq_idle
3783 * = rq_clock_pelt()@cfs_rq_idle
3784 * - cfs->throttled_clock_pelt_time@cfs_rq_idle
3786 * cfs_idle_lag (delta between rq's update and cfs_rq's update)
3787 * = rq_clock_pelt()@rq_idle - rq_clock_pelt()@cfs_rq_idle
3789 * rq_idle_lag (delta between now and rq's update)
3790 * = sched_clock_cpu() - rq_clock()@rq_idle
3792 * We can then write:
3794 * now = rq_clock_pelt()@rq_idle - cfs->throttled_clock_pelt_time +
3795 * sched_clock_cpu() - rq_clock()@rq_idle
3797 * rq_clock_pelt()@rq_idle is rq->clock_pelt_idle
3798 * rq_clock()@rq_idle is rq->clock_idle
3799 * cfs->throttled_clock_pelt_time@cfs_rq_idle
3800 * is cfs_rq->throttled_pelt_idle
3803 #ifdef CONFIG_CFS_BANDWIDTH
3804 throttled = u64_u32_load(cfs_rq->throttled_pelt_idle);
3805 /* The clock has been stopped for throttling */
3806 if (throttled == U64_MAX)
3809 now = u64_u32_load(rq->clock_pelt_idle);
3811 * Paired with _update_idle_rq_clock_pelt(). It ensures at the worst case
3812 * is observed the old clock_pelt_idle value and the new clock_idle,
3813 * which lead to an underestimation. The opposite would lead to an
3817 lut = cfs_rq_last_update_time(cfs_rq);
3822 * cfs_rq->avg.last_update_time is more recent than our
3823 * estimation, let's use it.
3827 now += sched_clock_cpu(cpu_of(rq)) - u64_u32_load(rq->clock_idle);
3829 __update_load_avg_blocked_se(now, se);
3832 static void migrate_se_pelt_lag(struct sched_entity *se) {}
3836 * update_cfs_rq_load_avg - update the cfs_rq's load/util averages
3837 * @now: current time, as per cfs_rq_clock_pelt()
3838 * @cfs_rq: cfs_rq to update
3840 * The cfs_rq avg is the direct sum of all its entities (blocked and runnable)
3841 * avg. The immediate corollary is that all (fair) tasks must be attached, see
3842 * post_init_entity_util_avg().
3844 * cfs_rq->avg is used for task_h_load() and update_cfs_share() for example.
3846 * Return: true if the load decayed or we removed load.
3848 * Since both these conditions indicate a changed cfs_rq->avg.load we should
3849 * call update_tg_load_avg() when this function returns true.
3852 update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq)
3854 unsigned long removed_load = 0, removed_util = 0, removed_runnable = 0;
3855 struct sched_avg *sa = &cfs_rq->avg;
3858 if (cfs_rq->removed.nr) {
3860 u32 divider = get_pelt_divider(&cfs_rq->avg);
3862 raw_spin_lock(&cfs_rq->removed.lock);
3863 swap(cfs_rq->removed.util_avg, removed_util);
3864 swap(cfs_rq->removed.load_avg, removed_load);
3865 swap(cfs_rq->removed.runnable_avg, removed_runnable);
3866 cfs_rq->removed.nr = 0;
3867 raw_spin_unlock(&cfs_rq->removed.lock);
3870 sub_positive(&sa->load_avg, r);
3871 sub_positive(&sa->load_sum, r * divider);
3872 /* See sa->util_sum below */
3873 sa->load_sum = max_t(u32, sa->load_sum, sa->load_avg * PELT_MIN_DIVIDER);
3876 sub_positive(&sa->util_avg, r);
3877 sub_positive(&sa->util_sum, r * divider);
3879 * Because of rounding, se->util_sum might ends up being +1 more than
3880 * cfs->util_sum. Although this is not a problem by itself, detaching
3881 * a lot of tasks with the rounding problem between 2 updates of
3882 * util_avg (~1ms) can make cfs->util_sum becoming null whereas
3883 * cfs_util_avg is not.
3884 * Check that util_sum is still above its lower bound for the new
3885 * util_avg. Given that period_contrib might have moved since the last
3886 * sync, we are only sure that util_sum must be above or equal to
3887 * util_avg * minimum possible divider
3889 sa->util_sum = max_t(u32, sa->util_sum, sa->util_avg * PELT_MIN_DIVIDER);
3891 r = removed_runnable;
3892 sub_positive(&sa->runnable_avg, r);
3893 sub_positive(&sa->runnable_sum, r * divider);
3894 /* See sa->util_sum above */
3895 sa->runnable_sum = max_t(u32, sa->runnable_sum,
3896 sa->runnable_avg * PELT_MIN_DIVIDER);
3899 * removed_runnable is the unweighted version of removed_load so we
3900 * can use it to estimate removed_load_sum.
3902 add_tg_cfs_propagate(cfs_rq,
3903 -(long)(removed_runnable * divider) >> SCHED_CAPACITY_SHIFT);
3908 decayed |= __update_load_avg_cfs_rq(now, cfs_rq);
3909 u64_u32_store_copy(sa->last_update_time,
3910 cfs_rq->last_update_time_copy,
3911 sa->last_update_time);
3916 * attach_entity_load_avg - attach this entity to its cfs_rq load avg
3917 * @cfs_rq: cfs_rq to attach to
3918 * @se: sched_entity to attach
3920 * Must call update_cfs_rq_load_avg() before this, since we rely on
3921 * cfs_rq->avg.last_update_time being current.
3923 static void attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3926 * cfs_rq->avg.period_contrib can be used for both cfs_rq and se.
3927 * See ___update_load_avg() for details.
3929 u32 divider = get_pelt_divider(&cfs_rq->avg);
3932 * When we attach the @se to the @cfs_rq, we must align the decay
3933 * window because without that, really weird and wonderful things can
3938 se->avg.last_update_time = cfs_rq->avg.last_update_time;
3939 se->avg.period_contrib = cfs_rq->avg.period_contrib;
3942 * Hell(o) Nasty stuff.. we need to recompute _sum based on the new
3943 * period_contrib. This isn't strictly correct, but since we're
3944 * entirely outside of the PELT hierarchy, nobody cares if we truncate
3947 se->avg.util_sum = se->avg.util_avg * divider;
3949 se->avg.runnable_sum = se->avg.runnable_avg * divider;
3951 se->avg.load_sum = se->avg.load_avg * divider;
3952 if (se_weight(se) < se->avg.load_sum)
3953 se->avg.load_sum = div_u64(se->avg.load_sum, se_weight(se));
3955 se->avg.load_sum = 1;
3957 enqueue_load_avg(cfs_rq, se);
3958 cfs_rq->avg.util_avg += se->avg.util_avg;
3959 cfs_rq->avg.util_sum += se->avg.util_sum;
3960 cfs_rq->avg.runnable_avg += se->avg.runnable_avg;
3961 cfs_rq->avg.runnable_sum += se->avg.runnable_sum;
3963 add_tg_cfs_propagate(cfs_rq, se->avg.load_sum);
3965 cfs_rq_util_change(cfs_rq, 0);
3967 trace_pelt_cfs_tp(cfs_rq);
3971 * detach_entity_load_avg - detach this entity from its cfs_rq load avg
3972 * @cfs_rq: cfs_rq to detach from
3973 * @se: sched_entity to detach
3975 * Must call update_cfs_rq_load_avg() before this, since we rely on
3976 * cfs_rq->avg.last_update_time being current.
3978 static void detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3980 dequeue_load_avg(cfs_rq, se);
3981 sub_positive(&cfs_rq->avg.util_avg, se->avg.util_avg);
3982 sub_positive(&cfs_rq->avg.util_sum, se->avg.util_sum);
3983 /* See update_cfs_rq_load_avg() */
3984 cfs_rq->avg.util_sum = max_t(u32, cfs_rq->avg.util_sum,
3985 cfs_rq->avg.util_avg * PELT_MIN_DIVIDER);
3987 sub_positive(&cfs_rq->avg.runnable_avg, se->avg.runnable_avg);
3988 sub_positive(&cfs_rq->avg.runnable_sum, se->avg.runnable_sum);
3989 /* See update_cfs_rq_load_avg() */
3990 cfs_rq->avg.runnable_sum = max_t(u32, cfs_rq->avg.runnable_sum,
3991 cfs_rq->avg.runnable_avg * PELT_MIN_DIVIDER);
3993 add_tg_cfs_propagate(cfs_rq, -se->avg.load_sum);
3995 cfs_rq_util_change(cfs_rq, 0);
3997 trace_pelt_cfs_tp(cfs_rq);
4001 * Optional action to be done while updating the load average
4003 #define UPDATE_TG 0x1
4004 #define SKIP_AGE_LOAD 0x2
4005 #define DO_ATTACH 0x4
4007 /* Update task and its cfs_rq load average */
4008 static inline void update_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
4010 u64 now = cfs_rq_clock_pelt(cfs_rq);
4014 * Track task load average for carrying it to new CPU after migrated, and
4015 * track group sched_entity load average for task_h_load calc in migration
4017 if (se->avg.last_update_time && !(flags & SKIP_AGE_LOAD))
4018 __update_load_avg_se(now, cfs_rq, se);
4020 decayed = update_cfs_rq_load_avg(now, cfs_rq);
4021 decayed |= propagate_entity_load_avg(se);
4023 if (!se->avg.last_update_time && (flags & DO_ATTACH)) {
4026 * DO_ATTACH means we're here from enqueue_entity().
4027 * !last_update_time means we've passed through
4028 * migrate_task_rq_fair() indicating we migrated.
4030 * IOW we're enqueueing a task on a new CPU.
4032 attach_entity_load_avg(cfs_rq, se);
4033 update_tg_load_avg(cfs_rq);
4035 } else if (decayed) {
4036 cfs_rq_util_change(cfs_rq, 0);
4038 if (flags & UPDATE_TG)
4039 update_tg_load_avg(cfs_rq);
4044 * Synchronize entity load avg of dequeued entity without locking
4047 static void sync_entity_load_avg(struct sched_entity *se)
4049 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4050 u64 last_update_time;
4052 last_update_time = cfs_rq_last_update_time(cfs_rq);
4053 __update_load_avg_blocked_se(last_update_time, se);
4057 * Task first catches up with cfs_rq, and then subtract
4058 * itself from the cfs_rq (task must be off the queue now).
4060 static void remove_entity_load_avg(struct sched_entity *se)
4062 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4063 unsigned long flags;
4066 * tasks cannot exit without having gone through wake_up_new_task() ->
4067 * post_init_entity_util_avg() which will have added things to the
4068 * cfs_rq, so we can remove unconditionally.
4071 sync_entity_load_avg(se);
4073 raw_spin_lock_irqsave(&cfs_rq->removed.lock, flags);
4074 ++cfs_rq->removed.nr;
4075 cfs_rq->removed.util_avg += se->avg.util_avg;
4076 cfs_rq->removed.load_avg += se->avg.load_avg;
4077 cfs_rq->removed.runnable_avg += se->avg.runnable_avg;
4078 raw_spin_unlock_irqrestore(&cfs_rq->removed.lock, flags);
4081 static inline unsigned long cfs_rq_runnable_avg(struct cfs_rq *cfs_rq)
4083 return cfs_rq->avg.runnable_avg;
4086 static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq)
4088 return cfs_rq->avg.load_avg;
4091 static int newidle_balance(struct rq *this_rq, struct rq_flags *rf);
4093 static inline unsigned long task_util(struct task_struct *p)
4095 return READ_ONCE(p->se.avg.util_avg);
4098 static inline unsigned long _task_util_est(struct task_struct *p)
4100 struct util_est ue = READ_ONCE(p->se.avg.util_est);
4102 return max(ue.ewma, (ue.enqueued & ~UTIL_AVG_UNCHANGED));
4105 static inline unsigned long task_util_est(struct task_struct *p)
4107 return max(task_util(p), _task_util_est(p));
4110 #ifdef CONFIG_UCLAMP_TASK
4111 static inline unsigned long uclamp_task_util(struct task_struct *p)
4113 return clamp(task_util_est(p),
4114 uclamp_eff_value(p, UCLAMP_MIN),
4115 uclamp_eff_value(p, UCLAMP_MAX));
4118 static inline unsigned long uclamp_task_util(struct task_struct *p)
4120 return task_util_est(p);
4124 static inline void util_est_enqueue(struct cfs_rq *cfs_rq,
4125 struct task_struct *p)
4127 unsigned int enqueued;
4129 if (!sched_feat(UTIL_EST))
4132 /* Update root cfs_rq's estimated utilization */
4133 enqueued = cfs_rq->avg.util_est.enqueued;
4134 enqueued += _task_util_est(p);
4135 WRITE_ONCE(cfs_rq->avg.util_est.enqueued, enqueued);
4137 trace_sched_util_est_cfs_tp(cfs_rq);
4140 static inline void util_est_dequeue(struct cfs_rq *cfs_rq,
4141 struct task_struct *p)
4143 unsigned int enqueued;
4145 if (!sched_feat(UTIL_EST))
4148 /* Update root cfs_rq's estimated utilization */
4149 enqueued = cfs_rq->avg.util_est.enqueued;
4150 enqueued -= min_t(unsigned int, enqueued, _task_util_est(p));
4151 WRITE_ONCE(cfs_rq->avg.util_est.enqueued, enqueued);
4153 trace_sched_util_est_cfs_tp(cfs_rq);
4156 #define UTIL_EST_MARGIN (SCHED_CAPACITY_SCALE / 100)
4159 * Check if a (signed) value is within a specified (unsigned) margin,
4160 * based on the observation that:
4162 * abs(x) < y := (unsigned)(x + y - 1) < (2 * y - 1)
4164 * NOTE: this only works when value + margin < INT_MAX.
4166 static inline bool within_margin(int value, int margin)
4168 return ((unsigned int)(value + margin - 1) < (2 * margin - 1));
4171 static inline void util_est_update(struct cfs_rq *cfs_rq,
4172 struct task_struct *p,
4175 long last_ewma_diff, last_enqueued_diff;
4178 if (!sched_feat(UTIL_EST))
4182 * Skip update of task's estimated utilization when the task has not
4183 * yet completed an activation, e.g. being migrated.
4189 * If the PELT values haven't changed since enqueue time,
4190 * skip the util_est update.
4192 ue = p->se.avg.util_est;
4193 if (ue.enqueued & UTIL_AVG_UNCHANGED)
4196 last_enqueued_diff = ue.enqueued;
4199 * Reset EWMA on utilization increases, the moving average is used only
4200 * to smooth utilization decreases.
4202 ue.enqueued = task_util(p);
4203 if (sched_feat(UTIL_EST_FASTUP)) {
4204 if (ue.ewma < ue.enqueued) {
4205 ue.ewma = ue.enqueued;
4211 * Skip update of task's estimated utilization when its members are
4212 * already ~1% close to its last activation value.
4214 last_ewma_diff = ue.enqueued - ue.ewma;
4215 last_enqueued_diff -= ue.enqueued;
4216 if (within_margin(last_ewma_diff, UTIL_EST_MARGIN)) {
4217 if (!within_margin(last_enqueued_diff, UTIL_EST_MARGIN))
4224 * To avoid overestimation of actual task utilization, skip updates if
4225 * we cannot grant there is idle time in this CPU.
4227 if (task_util(p) > capacity_orig_of(cpu_of(rq_of(cfs_rq))))
4231 * Update Task's estimated utilization
4233 * When *p completes an activation we can consolidate another sample
4234 * of the task size. This is done by storing the current PELT value
4235 * as ue.enqueued and by using this value to update the Exponential
4236 * Weighted Moving Average (EWMA):
4238 * ewma(t) = w * task_util(p) + (1-w) * ewma(t-1)
4239 * = w * task_util(p) + ewma(t-1) - w * ewma(t-1)
4240 * = w * (task_util(p) - ewma(t-1)) + ewma(t-1)
4241 * = w * ( last_ewma_diff ) + ewma(t-1)
4242 * = w * (last_ewma_diff + ewma(t-1) / w)
4244 * Where 'w' is the weight of new samples, which is configured to be
4245 * 0.25, thus making w=1/4 ( >>= UTIL_EST_WEIGHT_SHIFT)
4247 ue.ewma <<= UTIL_EST_WEIGHT_SHIFT;
4248 ue.ewma += last_ewma_diff;
4249 ue.ewma >>= UTIL_EST_WEIGHT_SHIFT;
4251 ue.enqueued |= UTIL_AVG_UNCHANGED;
4252 WRITE_ONCE(p->se.avg.util_est, ue);
4254 trace_sched_util_est_se_tp(&p->se);
4257 static inline int task_fits_capacity(struct task_struct *p,
4258 unsigned long capacity)
4260 return fits_capacity(uclamp_task_util(p), capacity);
4263 static inline void update_misfit_status(struct task_struct *p, struct rq *rq)
4265 if (!static_branch_unlikely(&sched_asym_cpucapacity))
4268 if (!p || p->nr_cpus_allowed == 1) {
4269 rq->misfit_task_load = 0;
4273 if (task_fits_capacity(p, capacity_of(cpu_of(rq)))) {
4274 rq->misfit_task_load = 0;
4279 * Make sure that misfit_task_load will not be null even if
4280 * task_h_load() returns 0.
4282 rq->misfit_task_load = max_t(unsigned long, task_h_load(p), 1);
4285 #else /* CONFIG_SMP */
4287 static inline bool cfs_rq_is_decayed(struct cfs_rq *cfs_rq)
4292 #define UPDATE_TG 0x0
4293 #define SKIP_AGE_LOAD 0x0
4294 #define DO_ATTACH 0x0
4296 static inline void update_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int not_used1)
4298 cfs_rq_util_change(cfs_rq, 0);
4301 static inline void remove_entity_load_avg(struct sched_entity *se) {}
4304 attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
4306 detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
4308 static inline int newidle_balance(struct rq *rq, struct rq_flags *rf)
4314 util_est_enqueue(struct cfs_rq *cfs_rq, struct task_struct *p) {}
4317 util_est_dequeue(struct cfs_rq *cfs_rq, struct task_struct *p) {}
4320 util_est_update(struct cfs_rq *cfs_rq, struct task_struct *p,
4322 static inline void update_misfit_status(struct task_struct *p, struct rq *rq) {}
4324 #endif /* CONFIG_SMP */
4326 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
4328 #ifdef CONFIG_SCHED_DEBUG
4329 s64 d = se->vruntime - cfs_rq->min_vruntime;
4334 if (d > 3*sysctl_sched_latency)
4335 schedstat_inc(cfs_rq->nr_spread_over);
4340 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
4342 u64 vruntime = cfs_rq->min_vruntime;
4345 * The 'current' period is already promised to the current tasks,
4346 * however the extra weight of the new task will slow them down a
4347 * little, place the new task so that it fits in the slot that
4348 * stays open at the end.
4350 if (initial && sched_feat(START_DEBIT))
4351 vruntime += sched_vslice(cfs_rq, se);
4353 /* sleeps up to a single latency don't count. */
4355 unsigned long thresh;
4358 thresh = sysctl_sched_min_granularity;
4360 thresh = sysctl_sched_latency;
4363 * Halve their sleep time's effect, to allow
4364 * for a gentler effect of sleepers:
4366 if (sched_feat(GENTLE_FAIR_SLEEPERS))
4372 /* ensure we never gain time by being placed backwards. */
4373 se->vruntime = max_vruntime(se->vruntime, vruntime);
4376 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
4378 static inline bool cfs_bandwidth_used(void);
4385 * update_min_vruntime()
4386 * vruntime -= min_vruntime
4390 * update_min_vruntime()
4391 * vruntime += min_vruntime
4393 * this way the vruntime transition between RQs is done when both
4394 * min_vruntime are up-to-date.
4398 * ->migrate_task_rq_fair() (p->state == TASK_WAKING)
4399 * vruntime -= min_vruntime
4403 * update_min_vruntime()
4404 * vruntime += min_vruntime
4406 * this way we don't have the most up-to-date min_vruntime on the originating
4407 * CPU and an up-to-date min_vruntime on the destination CPU.
4411 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
4413 bool renorm = !(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_MIGRATED);
4414 bool curr = cfs_rq->curr == se;
4417 * If we're the current task, we must renormalise before calling
4421 se->vruntime += cfs_rq->min_vruntime;
4423 update_curr(cfs_rq);
4426 * Otherwise, renormalise after, such that we're placed at the current
4427 * moment in time, instead of some random moment in the past. Being
4428 * placed in the past could significantly boost this task to the
4429 * fairness detriment of existing tasks.
4431 if (renorm && !curr)
4432 se->vruntime += cfs_rq->min_vruntime;
4435 * When enqueuing a sched_entity, we must:
4436 * - Update loads to have both entity and cfs_rq synced with now.
4437 * - Add its load to cfs_rq->runnable_avg
4438 * - For group_entity, update its weight to reflect the new share of
4440 * - Add its new weight to cfs_rq->load.weight
4442 update_load_avg(cfs_rq, se, UPDATE_TG | DO_ATTACH);
4443 se_update_runnable(se);
4444 update_cfs_group(se);
4445 account_entity_enqueue(cfs_rq, se);
4447 if (flags & ENQUEUE_WAKEUP)
4448 place_entity(cfs_rq, se, 0);
4450 check_schedstat_required();
4451 update_stats_enqueue_fair(cfs_rq, se, flags);
4452 check_spread(cfs_rq, se);
4454 __enqueue_entity(cfs_rq, se);
4457 if (cfs_rq->nr_running == 1) {
4458 check_enqueue_throttle(cfs_rq);
4459 if (!throttled_hierarchy(cfs_rq))
4460 list_add_leaf_cfs_rq(cfs_rq);
4464 static void __clear_buddies_last(struct sched_entity *se)
4466 for_each_sched_entity(se) {
4467 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4468 if (cfs_rq->last != se)
4471 cfs_rq->last = NULL;
4475 static void __clear_buddies_next(struct sched_entity *se)
4477 for_each_sched_entity(se) {
4478 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4479 if (cfs_rq->next != se)
4482 cfs_rq->next = NULL;
4486 static void __clear_buddies_skip(struct sched_entity *se)
4488 for_each_sched_entity(se) {
4489 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4490 if (cfs_rq->skip != se)
4493 cfs_rq->skip = NULL;
4497 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
4499 if (cfs_rq->last == se)
4500 __clear_buddies_last(se);
4502 if (cfs_rq->next == se)
4503 __clear_buddies_next(se);
4505 if (cfs_rq->skip == se)
4506 __clear_buddies_skip(se);
4509 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
4512 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
4515 * Update run-time statistics of the 'current'.
4517 update_curr(cfs_rq);
4520 * When dequeuing a sched_entity, we must:
4521 * - Update loads to have both entity and cfs_rq synced with now.
4522 * - Subtract its load from the cfs_rq->runnable_avg.
4523 * - Subtract its previous weight from cfs_rq->load.weight.
4524 * - For group entity, update its weight to reflect the new share
4525 * of its group cfs_rq.
4527 update_load_avg(cfs_rq, se, UPDATE_TG);
4528 se_update_runnable(se);
4530 update_stats_dequeue_fair(cfs_rq, se, flags);
4532 clear_buddies(cfs_rq, se);
4534 if (se != cfs_rq->curr)
4535 __dequeue_entity(cfs_rq, se);
4537 account_entity_dequeue(cfs_rq, se);
4540 * Normalize after update_curr(); which will also have moved
4541 * min_vruntime if @se is the one holding it back. But before doing
4542 * update_min_vruntime() again, which will discount @se's position and
4543 * can move min_vruntime forward still more.
4545 if (!(flags & DEQUEUE_SLEEP))
4546 se->vruntime -= cfs_rq->min_vruntime;
4548 /* return excess runtime on last dequeue */
4549 return_cfs_rq_runtime(cfs_rq);
4551 update_cfs_group(se);
4554 * Now advance min_vruntime if @se was the entity holding it back,
4555 * except when: DEQUEUE_SAVE && !DEQUEUE_MOVE, in this case we'll be
4556 * put back on, and if we advance min_vruntime, we'll be placed back
4557 * further than we started -- ie. we'll be penalized.
4559 if ((flags & (DEQUEUE_SAVE | DEQUEUE_MOVE)) != DEQUEUE_SAVE)
4560 update_min_vruntime(cfs_rq);
4562 if (cfs_rq->nr_running == 0)
4563 update_idle_cfs_rq_clock_pelt(cfs_rq);
4567 * Preempt the current task with a newly woken task if needed:
4570 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
4572 unsigned long ideal_runtime, delta_exec;
4573 struct sched_entity *se;
4576 ideal_runtime = sched_slice(cfs_rq, curr);
4577 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
4578 if (delta_exec > ideal_runtime) {
4579 resched_curr(rq_of(cfs_rq));
4581 * The current task ran long enough, ensure it doesn't get
4582 * re-elected due to buddy favours.
4584 clear_buddies(cfs_rq, curr);
4589 * Ensure that a task that missed wakeup preemption by a
4590 * narrow margin doesn't have to wait for a full slice.
4591 * This also mitigates buddy induced latencies under load.
4593 if (delta_exec < sysctl_sched_min_granularity)
4596 se = __pick_first_entity(cfs_rq);
4597 delta = curr->vruntime - se->vruntime;
4602 if (delta > ideal_runtime)
4603 resched_curr(rq_of(cfs_rq));
4607 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
4609 clear_buddies(cfs_rq, se);
4611 /* 'current' is not kept within the tree. */
4614 * Any task has to be enqueued before it get to execute on
4615 * a CPU. So account for the time it spent waiting on the
4618 update_stats_wait_end_fair(cfs_rq, se);
4619 __dequeue_entity(cfs_rq, se);
4620 update_load_avg(cfs_rq, se, UPDATE_TG);
4623 update_stats_curr_start(cfs_rq, se);
4627 * Track our maximum slice length, if the CPU's load is at
4628 * least twice that of our own weight (i.e. dont track it
4629 * when there are only lesser-weight tasks around):
4631 if (schedstat_enabled() &&
4632 rq_of(cfs_rq)->cfs.load.weight >= 2*se->load.weight) {
4633 struct sched_statistics *stats;
4635 stats = __schedstats_from_se(se);
4636 __schedstat_set(stats->slice_max,
4637 max((u64)stats->slice_max,
4638 se->sum_exec_runtime - se->prev_sum_exec_runtime));
4641 se->prev_sum_exec_runtime = se->sum_exec_runtime;
4645 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
4648 * Pick the next process, keeping these things in mind, in this order:
4649 * 1) keep things fair between processes/task groups
4650 * 2) pick the "next" process, since someone really wants that to run
4651 * 3) pick the "last" process, for cache locality
4652 * 4) do not run the "skip" process, if something else is available
4654 static struct sched_entity *
4655 pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr)
4657 struct sched_entity *left = __pick_first_entity(cfs_rq);
4658 struct sched_entity *se;
4661 * If curr is set we have to see if its left of the leftmost entity
4662 * still in the tree, provided there was anything in the tree at all.
4664 if (!left || (curr && entity_before(curr, left)))
4667 se = left; /* ideally we run the leftmost entity */
4670 * Avoid running the skip buddy, if running something else can
4671 * be done without getting too unfair.
4673 if (cfs_rq->skip && cfs_rq->skip == se) {
4674 struct sched_entity *second;
4677 second = __pick_first_entity(cfs_rq);
4679 second = __pick_next_entity(se);
4680 if (!second || (curr && entity_before(curr, second)))
4684 if (second && wakeup_preempt_entity(second, left) < 1)
4688 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1) {
4690 * Someone really wants this to run. If it's not unfair, run it.
4693 } else if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1) {
4695 * Prefer last buddy, try to return the CPU to a preempted task.
4703 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
4705 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
4708 * If still on the runqueue then deactivate_task()
4709 * was not called and update_curr() has to be done:
4712 update_curr(cfs_rq);
4714 /* throttle cfs_rqs exceeding runtime */
4715 check_cfs_rq_runtime(cfs_rq);
4717 check_spread(cfs_rq, prev);
4720 update_stats_wait_start_fair(cfs_rq, prev);
4721 /* Put 'current' back into the tree. */
4722 __enqueue_entity(cfs_rq, prev);
4723 /* in !on_rq case, update occurred at dequeue */
4724 update_load_avg(cfs_rq, prev, 0);
4726 cfs_rq->curr = NULL;
4730 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
4733 * Update run-time statistics of the 'current'.
4735 update_curr(cfs_rq);
4738 * Ensure that runnable average is periodically updated.
4740 update_load_avg(cfs_rq, curr, UPDATE_TG);
4741 update_cfs_group(curr);
4743 #ifdef CONFIG_SCHED_HRTICK
4745 * queued ticks are scheduled to match the slice, so don't bother
4746 * validating it and just reschedule.
4749 resched_curr(rq_of(cfs_rq));
4753 * don't let the period tick interfere with the hrtick preemption
4755 if (!sched_feat(DOUBLE_TICK) &&
4756 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
4760 if (cfs_rq->nr_running > 1)
4761 check_preempt_tick(cfs_rq, curr);
4765 /**************************************************
4766 * CFS bandwidth control machinery
4769 #ifdef CONFIG_CFS_BANDWIDTH
4771 #ifdef CONFIG_JUMP_LABEL
4772 static struct static_key __cfs_bandwidth_used;
4774 static inline bool cfs_bandwidth_used(void)
4776 return static_key_false(&__cfs_bandwidth_used);
4779 void cfs_bandwidth_usage_inc(void)
4781 static_key_slow_inc_cpuslocked(&__cfs_bandwidth_used);
4784 void cfs_bandwidth_usage_dec(void)
4786 static_key_slow_dec_cpuslocked(&__cfs_bandwidth_used);
4788 #else /* CONFIG_JUMP_LABEL */
4789 static bool cfs_bandwidth_used(void)
4794 void cfs_bandwidth_usage_inc(void) {}
4795 void cfs_bandwidth_usage_dec(void) {}
4796 #endif /* CONFIG_JUMP_LABEL */
4799 * default period for cfs group bandwidth.
4800 * default: 0.1s, units: nanoseconds
4802 static inline u64 default_cfs_period(void)
4804 return 100000000ULL;
4807 static inline u64 sched_cfs_bandwidth_slice(void)
4809 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
4813 * Replenish runtime according to assigned quota. We use sched_clock_cpu
4814 * directly instead of rq->clock to avoid adding additional synchronization
4817 * requires cfs_b->lock
4819 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
4823 if (unlikely(cfs_b->quota == RUNTIME_INF))
4826 cfs_b->runtime += cfs_b->quota;
4827 runtime = cfs_b->runtime_snap - cfs_b->runtime;
4829 cfs_b->burst_time += runtime;
4833 cfs_b->runtime = min(cfs_b->runtime, cfs_b->quota + cfs_b->burst);
4834 cfs_b->runtime_snap = cfs_b->runtime;
4837 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
4839 return &tg->cfs_bandwidth;
4842 /* returns 0 on failure to allocate runtime */
4843 static int __assign_cfs_rq_runtime(struct cfs_bandwidth *cfs_b,
4844 struct cfs_rq *cfs_rq, u64 target_runtime)
4846 u64 min_amount, amount = 0;
4848 lockdep_assert_held(&cfs_b->lock);
4850 /* note: this is a positive sum as runtime_remaining <= 0 */
4851 min_amount = target_runtime - cfs_rq->runtime_remaining;
4853 if (cfs_b->quota == RUNTIME_INF)
4854 amount = min_amount;
4856 start_cfs_bandwidth(cfs_b);
4858 if (cfs_b->runtime > 0) {
4859 amount = min(cfs_b->runtime, min_amount);
4860 cfs_b->runtime -= amount;
4865 cfs_rq->runtime_remaining += amount;
4867 return cfs_rq->runtime_remaining > 0;
4870 /* returns 0 on failure to allocate runtime */
4871 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4873 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4876 raw_spin_lock(&cfs_b->lock);
4877 ret = __assign_cfs_rq_runtime(cfs_b, cfs_rq, sched_cfs_bandwidth_slice());
4878 raw_spin_unlock(&cfs_b->lock);
4883 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
4885 /* dock delta_exec before expiring quota (as it could span periods) */
4886 cfs_rq->runtime_remaining -= delta_exec;
4888 if (likely(cfs_rq->runtime_remaining > 0))
4891 if (cfs_rq->throttled)
4894 * if we're unable to extend our runtime we resched so that the active
4895 * hierarchy can be throttled
4897 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
4898 resched_curr(rq_of(cfs_rq));
4901 static __always_inline
4902 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
4904 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
4907 __account_cfs_rq_runtime(cfs_rq, delta_exec);
4910 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
4912 return cfs_bandwidth_used() && cfs_rq->throttled;
4915 /* check whether cfs_rq, or any parent, is throttled */
4916 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
4918 return cfs_bandwidth_used() && cfs_rq->throttle_count;
4922 * Ensure that neither of the group entities corresponding to src_cpu or
4923 * dest_cpu are members of a throttled hierarchy when performing group
4924 * load-balance operations.
4926 static inline int throttled_lb_pair(struct task_group *tg,
4927 int src_cpu, int dest_cpu)
4929 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
4931 src_cfs_rq = tg->cfs_rq[src_cpu];
4932 dest_cfs_rq = tg->cfs_rq[dest_cpu];
4934 return throttled_hierarchy(src_cfs_rq) ||
4935 throttled_hierarchy(dest_cfs_rq);
4938 static int tg_unthrottle_up(struct task_group *tg, void *data)
4940 struct rq *rq = data;
4941 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
4943 cfs_rq->throttle_count--;
4944 if (!cfs_rq->throttle_count) {
4945 cfs_rq->throttled_clock_pelt_time += rq_clock_pelt(rq) -
4946 cfs_rq->throttled_clock_pelt;
4948 /* Add cfs_rq with load or one or more already running entities to the list */
4949 if (!cfs_rq_is_decayed(cfs_rq))
4950 list_add_leaf_cfs_rq(cfs_rq);
4956 static int tg_throttle_down(struct task_group *tg, void *data)
4958 struct rq *rq = data;
4959 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
4961 /* group is entering throttled state, stop time */
4962 if (!cfs_rq->throttle_count) {
4963 cfs_rq->throttled_clock_pelt = rq_clock_pelt(rq);
4964 list_del_leaf_cfs_rq(cfs_rq);
4966 cfs_rq->throttle_count++;
4971 static bool throttle_cfs_rq(struct cfs_rq *cfs_rq)
4973 struct rq *rq = rq_of(cfs_rq);
4974 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4975 struct sched_entity *se;
4976 long task_delta, idle_task_delta, dequeue = 1;
4978 raw_spin_lock(&cfs_b->lock);
4979 /* This will start the period timer if necessary */
4980 if (__assign_cfs_rq_runtime(cfs_b, cfs_rq, 1)) {
4982 * We have raced with bandwidth becoming available, and if we
4983 * actually throttled the timer might not unthrottle us for an
4984 * entire period. We additionally needed to make sure that any
4985 * subsequent check_cfs_rq_runtime calls agree not to throttle
4986 * us, as we may commit to do cfs put_prev+pick_next, so we ask
4987 * for 1ns of runtime rather than just check cfs_b.
4991 list_add_tail_rcu(&cfs_rq->throttled_list,
4992 &cfs_b->throttled_cfs_rq);
4994 raw_spin_unlock(&cfs_b->lock);
4997 return false; /* Throttle no longer required. */
4999 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
5001 /* freeze hierarchy runnable averages while throttled */
5003 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
5006 task_delta = cfs_rq->h_nr_running;
5007 idle_task_delta = cfs_rq->idle_h_nr_running;
5008 for_each_sched_entity(se) {
5009 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
5010 /* throttled entity or throttle-on-deactivate */
5014 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
5016 if (cfs_rq_is_idle(group_cfs_rq(se)))
5017 idle_task_delta = cfs_rq->h_nr_running;
5019 qcfs_rq->h_nr_running -= task_delta;
5020 qcfs_rq->idle_h_nr_running -= idle_task_delta;
5022 if (qcfs_rq->load.weight) {
5023 /* Avoid re-evaluating load for this entity: */
5024 se = parent_entity(se);
5029 for_each_sched_entity(se) {
5030 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
5031 /* throttled entity or throttle-on-deactivate */
5035 update_load_avg(qcfs_rq, se, 0);
5036 se_update_runnable(se);
5038 if (cfs_rq_is_idle(group_cfs_rq(se)))
5039 idle_task_delta = cfs_rq->h_nr_running;
5041 qcfs_rq->h_nr_running -= task_delta;
5042 qcfs_rq->idle_h_nr_running -= idle_task_delta;
5045 /* At this point se is NULL and we are at root level*/
5046 sub_nr_running(rq, task_delta);
5050 * Note: distribution will already see us throttled via the
5051 * throttled-list. rq->lock protects completion.
5053 cfs_rq->throttled = 1;
5054 cfs_rq->throttled_clock = rq_clock(rq);
5058 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
5060 struct rq *rq = rq_of(cfs_rq);
5061 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
5062 struct sched_entity *se;
5063 long task_delta, idle_task_delta;
5065 se = cfs_rq->tg->se[cpu_of(rq)];
5067 cfs_rq->throttled = 0;
5069 update_rq_clock(rq);
5071 raw_spin_lock(&cfs_b->lock);
5072 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
5073 list_del_rcu(&cfs_rq->throttled_list);
5074 raw_spin_unlock(&cfs_b->lock);
5076 /* update hierarchical throttle state */
5077 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
5079 if (!cfs_rq->load.weight) {
5080 if (!cfs_rq->on_list)
5083 * Nothing to run but something to decay (on_list)?
5084 * Complete the branch.
5086 for_each_sched_entity(se) {
5087 if (list_add_leaf_cfs_rq(cfs_rq_of(se)))
5090 goto unthrottle_throttle;
5093 task_delta = cfs_rq->h_nr_running;
5094 idle_task_delta = cfs_rq->idle_h_nr_running;
5095 for_each_sched_entity(se) {
5096 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
5100 enqueue_entity(qcfs_rq, se, ENQUEUE_WAKEUP);
5102 if (cfs_rq_is_idle(group_cfs_rq(se)))
5103 idle_task_delta = cfs_rq->h_nr_running;
5105 qcfs_rq->h_nr_running += task_delta;
5106 qcfs_rq->idle_h_nr_running += idle_task_delta;
5108 /* end evaluation on encountering a throttled cfs_rq */
5109 if (cfs_rq_throttled(qcfs_rq))
5110 goto unthrottle_throttle;
5113 for_each_sched_entity(se) {
5114 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
5116 update_load_avg(qcfs_rq, se, UPDATE_TG);
5117 se_update_runnable(se);
5119 if (cfs_rq_is_idle(group_cfs_rq(se)))
5120 idle_task_delta = cfs_rq->h_nr_running;
5122 qcfs_rq->h_nr_running += task_delta;
5123 qcfs_rq->idle_h_nr_running += idle_task_delta;
5125 /* end evaluation on encountering a throttled cfs_rq */
5126 if (cfs_rq_throttled(qcfs_rq))
5127 goto unthrottle_throttle;
5130 /* At this point se is NULL and we are at root level*/
5131 add_nr_running(rq, task_delta);
5133 unthrottle_throttle:
5134 assert_list_leaf_cfs_rq(rq);
5136 /* Determine whether we need to wake up potentially idle CPU: */
5137 if (rq->curr == rq->idle && rq->cfs.nr_running)
5141 static void distribute_cfs_runtime(struct cfs_bandwidth *cfs_b)
5143 struct cfs_rq *cfs_rq;
5144 u64 runtime, remaining = 1;
5147 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
5149 struct rq *rq = rq_of(cfs_rq);
5152 rq_lock_irqsave(rq, &rf);
5153 if (!cfs_rq_throttled(cfs_rq))
5156 /* By the above check, this should never be true */
5157 SCHED_WARN_ON(cfs_rq->runtime_remaining > 0);
5159 raw_spin_lock(&cfs_b->lock);
5160 runtime = -cfs_rq->runtime_remaining + 1;
5161 if (runtime > cfs_b->runtime)
5162 runtime = cfs_b->runtime;
5163 cfs_b->runtime -= runtime;
5164 remaining = cfs_b->runtime;
5165 raw_spin_unlock(&cfs_b->lock);
5167 cfs_rq->runtime_remaining += runtime;
5169 /* we check whether we're throttled above */
5170 if (cfs_rq->runtime_remaining > 0)
5171 unthrottle_cfs_rq(cfs_rq);
5174 rq_unlock_irqrestore(rq, &rf);
5183 * Responsible for refilling a task_group's bandwidth and unthrottling its
5184 * cfs_rqs as appropriate. If there has been no activity within the last
5185 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
5186 * used to track this state.
5188 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun, unsigned long flags)
5192 /* no need to continue the timer with no bandwidth constraint */
5193 if (cfs_b->quota == RUNTIME_INF)
5194 goto out_deactivate;
5196 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
5197 cfs_b->nr_periods += overrun;
5199 /* Refill extra burst quota even if cfs_b->idle */
5200 __refill_cfs_bandwidth_runtime(cfs_b);
5203 * idle depends on !throttled (for the case of a large deficit), and if
5204 * we're going inactive then everything else can be deferred
5206 if (cfs_b->idle && !throttled)
5207 goto out_deactivate;
5210 /* mark as potentially idle for the upcoming period */
5215 /* account preceding periods in which throttling occurred */
5216 cfs_b->nr_throttled += overrun;
5219 * This check is repeated as we release cfs_b->lock while we unthrottle.
5221 while (throttled && cfs_b->runtime > 0) {
5222 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
5223 /* we can't nest cfs_b->lock while distributing bandwidth */
5224 distribute_cfs_runtime(cfs_b);
5225 raw_spin_lock_irqsave(&cfs_b->lock, flags);
5227 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
5231 * While we are ensured activity in the period following an
5232 * unthrottle, this also covers the case in which the new bandwidth is
5233 * insufficient to cover the existing bandwidth deficit. (Forcing the
5234 * timer to remain active while there are any throttled entities.)
5244 /* a cfs_rq won't donate quota below this amount */
5245 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
5246 /* minimum remaining period time to redistribute slack quota */
5247 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
5248 /* how long we wait to gather additional slack before distributing */
5249 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
5252 * Are we near the end of the current quota period?
5254 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
5255 * hrtimer base being cleared by hrtimer_start. In the case of
5256 * migrate_hrtimers, base is never cleared, so we are fine.
5258 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
5260 struct hrtimer *refresh_timer = &cfs_b->period_timer;
5263 /* if the call-back is running a quota refresh is already occurring */
5264 if (hrtimer_callback_running(refresh_timer))
5267 /* is a quota refresh about to occur? */
5268 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
5269 if (remaining < (s64)min_expire)
5275 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
5277 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
5279 /* if there's a quota refresh soon don't bother with slack */
5280 if (runtime_refresh_within(cfs_b, min_left))
5283 /* don't push forwards an existing deferred unthrottle */
5284 if (cfs_b->slack_started)
5286 cfs_b->slack_started = true;
5288 hrtimer_start(&cfs_b->slack_timer,
5289 ns_to_ktime(cfs_bandwidth_slack_period),
5293 /* we know any runtime found here is valid as update_curr() precedes return */
5294 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
5296 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
5297 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
5299 if (slack_runtime <= 0)
5302 raw_spin_lock(&cfs_b->lock);
5303 if (cfs_b->quota != RUNTIME_INF) {
5304 cfs_b->runtime += slack_runtime;
5306 /* we are under rq->lock, defer unthrottling using a timer */
5307 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
5308 !list_empty(&cfs_b->throttled_cfs_rq))
5309 start_cfs_slack_bandwidth(cfs_b);
5311 raw_spin_unlock(&cfs_b->lock);
5313 /* even if it's not valid for return we don't want to try again */
5314 cfs_rq->runtime_remaining -= slack_runtime;
5317 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
5319 if (!cfs_bandwidth_used())
5322 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
5325 __return_cfs_rq_runtime(cfs_rq);
5329 * This is done with a timer (instead of inline with bandwidth return) since
5330 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
5332 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
5334 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
5335 unsigned long flags;
5337 /* confirm we're still not at a refresh boundary */
5338 raw_spin_lock_irqsave(&cfs_b->lock, flags);
5339 cfs_b->slack_started = false;
5341 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
5342 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
5346 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice)
5347 runtime = cfs_b->runtime;
5349 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
5354 distribute_cfs_runtime(cfs_b);
5358 * When a group wakes up we want to make sure that its quota is not already
5359 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
5360 * runtime as update_curr() throttling can not trigger until it's on-rq.
5362 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
5364 if (!cfs_bandwidth_used())
5367 /* an active group must be handled by the update_curr()->put() path */
5368 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
5371 /* ensure the group is not already throttled */
5372 if (cfs_rq_throttled(cfs_rq))
5375 /* update runtime allocation */
5376 account_cfs_rq_runtime(cfs_rq, 0);
5377 if (cfs_rq->runtime_remaining <= 0)
5378 throttle_cfs_rq(cfs_rq);
5381 static void sync_throttle(struct task_group *tg, int cpu)
5383 struct cfs_rq *pcfs_rq, *cfs_rq;
5385 if (!cfs_bandwidth_used())
5391 cfs_rq = tg->cfs_rq[cpu];
5392 pcfs_rq = tg->parent->cfs_rq[cpu];
5394 cfs_rq->throttle_count = pcfs_rq->throttle_count;
5395 cfs_rq->throttled_clock_pelt = rq_clock_pelt(cpu_rq(cpu));
5398 /* conditionally throttle active cfs_rq's from put_prev_entity() */
5399 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
5401 if (!cfs_bandwidth_used())
5404 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
5408 * it's possible for a throttled entity to be forced into a running
5409 * state (e.g. set_curr_task), in this case we're finished.
5411 if (cfs_rq_throttled(cfs_rq))
5414 return throttle_cfs_rq(cfs_rq);
5417 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
5419 struct cfs_bandwidth *cfs_b =
5420 container_of(timer, struct cfs_bandwidth, slack_timer);
5422 do_sched_cfs_slack_timer(cfs_b);
5424 return HRTIMER_NORESTART;
5427 extern const u64 max_cfs_quota_period;
5429 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
5431 struct cfs_bandwidth *cfs_b =
5432 container_of(timer, struct cfs_bandwidth, period_timer);
5433 unsigned long flags;
5438 raw_spin_lock_irqsave(&cfs_b->lock, flags);
5440 overrun = hrtimer_forward_now(timer, cfs_b->period);
5444 idle = do_sched_cfs_period_timer(cfs_b, overrun, flags);
5447 u64 new, old = ktime_to_ns(cfs_b->period);
5450 * Grow period by a factor of 2 to avoid losing precision.
5451 * Precision loss in the quota/period ratio can cause __cfs_schedulable
5455 if (new < max_cfs_quota_period) {
5456 cfs_b->period = ns_to_ktime(new);
5460 pr_warn_ratelimited(
5461 "cfs_period_timer[cpu%d]: period too short, scaling up (new cfs_period_us = %lld, cfs_quota_us = %lld)\n",
5463 div_u64(new, NSEC_PER_USEC),
5464 div_u64(cfs_b->quota, NSEC_PER_USEC));
5466 pr_warn_ratelimited(
5467 "cfs_period_timer[cpu%d]: period too short, but cannot scale up without losing precision (cfs_period_us = %lld, cfs_quota_us = %lld)\n",
5469 div_u64(old, NSEC_PER_USEC),
5470 div_u64(cfs_b->quota, NSEC_PER_USEC));
5473 /* reset count so we don't come right back in here */
5478 cfs_b->period_active = 0;
5479 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
5481 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
5484 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
5486 raw_spin_lock_init(&cfs_b->lock);
5488 cfs_b->quota = RUNTIME_INF;
5489 cfs_b->period = ns_to_ktime(default_cfs_period());
5492 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
5493 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED);
5494 cfs_b->period_timer.function = sched_cfs_period_timer;
5495 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
5496 cfs_b->slack_timer.function = sched_cfs_slack_timer;
5497 cfs_b->slack_started = false;
5500 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
5502 cfs_rq->runtime_enabled = 0;
5503 INIT_LIST_HEAD(&cfs_rq->throttled_list);
5506 void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
5508 lockdep_assert_held(&cfs_b->lock);
5510 if (cfs_b->period_active)
5513 cfs_b->period_active = 1;
5514 hrtimer_forward_now(&cfs_b->period_timer, cfs_b->period);
5515 hrtimer_start_expires(&cfs_b->period_timer, HRTIMER_MODE_ABS_PINNED);
5518 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
5520 /* init_cfs_bandwidth() was not called */
5521 if (!cfs_b->throttled_cfs_rq.next)
5524 hrtimer_cancel(&cfs_b->period_timer);
5525 hrtimer_cancel(&cfs_b->slack_timer);
5529 * Both these CPU hotplug callbacks race against unregister_fair_sched_group()
5531 * The race is harmless, since modifying bandwidth settings of unhooked group
5532 * bits doesn't do much.
5535 /* cpu online callback */
5536 static void __maybe_unused update_runtime_enabled(struct rq *rq)
5538 struct task_group *tg;
5540 lockdep_assert_rq_held(rq);
5543 list_for_each_entry_rcu(tg, &task_groups, list) {
5544 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
5545 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
5547 raw_spin_lock(&cfs_b->lock);
5548 cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF;
5549 raw_spin_unlock(&cfs_b->lock);
5554 /* cpu offline callback */
5555 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
5557 struct task_group *tg;
5559 lockdep_assert_rq_held(rq);
5562 list_for_each_entry_rcu(tg, &task_groups, list) {
5563 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
5565 if (!cfs_rq->runtime_enabled)
5569 * clock_task is not advancing so we just need to make sure
5570 * there's some valid quota amount
5572 cfs_rq->runtime_remaining = 1;
5574 * Offline rq is schedulable till CPU is completely disabled
5575 * in take_cpu_down(), so we prevent new cfs throttling here.
5577 cfs_rq->runtime_enabled = 0;
5579 if (cfs_rq_throttled(cfs_rq))
5580 unthrottle_cfs_rq(cfs_rq);
5585 #else /* CONFIG_CFS_BANDWIDTH */
5587 static inline bool cfs_bandwidth_used(void)
5592 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {}
5593 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; }
5594 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
5595 static inline void sync_throttle(struct task_group *tg, int cpu) {}
5596 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
5598 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
5603 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
5608 static inline int throttled_lb_pair(struct task_group *tg,
5609 int src_cpu, int dest_cpu)
5614 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
5616 #ifdef CONFIG_FAIR_GROUP_SCHED
5617 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
5620 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
5624 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
5625 static inline void update_runtime_enabled(struct rq *rq) {}
5626 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
5628 #endif /* CONFIG_CFS_BANDWIDTH */
5630 /**************************************************
5631 * CFS operations on tasks:
5634 #ifdef CONFIG_SCHED_HRTICK
5635 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
5637 struct sched_entity *se = &p->se;
5638 struct cfs_rq *cfs_rq = cfs_rq_of(se);
5640 SCHED_WARN_ON(task_rq(p) != rq);
5642 if (rq->cfs.h_nr_running > 1) {
5643 u64 slice = sched_slice(cfs_rq, se);
5644 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
5645 s64 delta = slice - ran;
5648 if (task_current(rq, p))
5652 hrtick_start(rq, delta);
5657 * called from enqueue/dequeue and updates the hrtick when the
5658 * current task is from our class and nr_running is low enough
5661 static void hrtick_update(struct rq *rq)
5663 struct task_struct *curr = rq->curr;
5665 if (!hrtick_enabled_fair(rq) || curr->sched_class != &fair_sched_class)
5668 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
5669 hrtick_start_fair(rq, curr);
5671 #else /* !CONFIG_SCHED_HRTICK */
5673 hrtick_start_fair(struct rq *rq, struct task_struct *p)
5677 static inline void hrtick_update(struct rq *rq)
5683 static inline bool cpu_overutilized(int cpu)
5685 return !fits_capacity(cpu_util_cfs(cpu), capacity_of(cpu));
5688 static inline void update_overutilized_status(struct rq *rq)
5690 if (!READ_ONCE(rq->rd->overutilized) && cpu_overutilized(rq->cpu)) {
5691 WRITE_ONCE(rq->rd->overutilized, SG_OVERUTILIZED);
5692 trace_sched_overutilized_tp(rq->rd, SG_OVERUTILIZED);
5696 static inline void update_overutilized_status(struct rq *rq) { }
5699 /* Runqueue only has SCHED_IDLE tasks enqueued */
5700 static int sched_idle_rq(struct rq *rq)
5702 return unlikely(rq->nr_running == rq->cfs.idle_h_nr_running &&
5707 * Returns true if cfs_rq only has SCHED_IDLE entities enqueued. Note the use
5708 * of idle_nr_running, which does not consider idle descendants of normal
5711 static bool sched_idle_cfs_rq(struct cfs_rq *cfs_rq)
5713 return cfs_rq->nr_running &&
5714 cfs_rq->nr_running == cfs_rq->idle_nr_running;
5718 static int sched_idle_cpu(int cpu)
5720 return sched_idle_rq(cpu_rq(cpu));
5725 * The enqueue_task method is called before nr_running is
5726 * increased. Here we update the fair scheduling stats and
5727 * then put the task into the rbtree:
5730 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
5732 struct cfs_rq *cfs_rq;
5733 struct sched_entity *se = &p->se;
5734 int idle_h_nr_running = task_has_idle_policy(p);
5735 int task_new = !(flags & ENQUEUE_WAKEUP);
5738 * The code below (indirectly) updates schedutil which looks at
5739 * the cfs_rq utilization to select a frequency.
5740 * Let's add the task's estimated utilization to the cfs_rq's
5741 * estimated utilization, before we update schedutil.
5743 util_est_enqueue(&rq->cfs, p);
5746 * If in_iowait is set, the code below may not trigger any cpufreq
5747 * utilization updates, so do it here explicitly with the IOWAIT flag
5751 cpufreq_update_util(rq, SCHED_CPUFREQ_IOWAIT);
5753 for_each_sched_entity(se) {
5756 cfs_rq = cfs_rq_of(se);
5757 enqueue_entity(cfs_rq, se, flags);
5759 cfs_rq->h_nr_running++;
5760 cfs_rq->idle_h_nr_running += idle_h_nr_running;
5762 if (cfs_rq_is_idle(cfs_rq))
5763 idle_h_nr_running = 1;
5765 /* end evaluation on encountering a throttled cfs_rq */
5766 if (cfs_rq_throttled(cfs_rq))
5767 goto enqueue_throttle;
5769 flags = ENQUEUE_WAKEUP;
5772 for_each_sched_entity(se) {
5773 cfs_rq = cfs_rq_of(se);
5775 update_load_avg(cfs_rq, se, UPDATE_TG);
5776 se_update_runnable(se);
5777 update_cfs_group(se);
5779 cfs_rq->h_nr_running++;
5780 cfs_rq->idle_h_nr_running += idle_h_nr_running;
5782 if (cfs_rq_is_idle(cfs_rq))
5783 idle_h_nr_running = 1;
5785 /* end evaluation on encountering a throttled cfs_rq */
5786 if (cfs_rq_throttled(cfs_rq))
5787 goto enqueue_throttle;
5790 /* At this point se is NULL and we are at root level*/
5791 add_nr_running(rq, 1);
5794 * Since new tasks are assigned an initial util_avg equal to
5795 * half of the spare capacity of their CPU, tiny tasks have the
5796 * ability to cross the overutilized threshold, which will
5797 * result in the load balancer ruining all the task placement
5798 * done by EAS. As a way to mitigate that effect, do not account
5799 * for the first enqueue operation of new tasks during the
5800 * overutilized flag detection.
5802 * A better way of solving this problem would be to wait for
5803 * the PELT signals of tasks to converge before taking them
5804 * into account, but that is not straightforward to implement,
5805 * and the following generally works well enough in practice.
5808 update_overutilized_status(rq);
5811 assert_list_leaf_cfs_rq(rq);
5816 static void set_next_buddy(struct sched_entity *se);
5819 * The dequeue_task method is called before nr_running is
5820 * decreased. We remove the task from the rbtree and
5821 * update the fair scheduling stats:
5823 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
5825 struct cfs_rq *cfs_rq;
5826 struct sched_entity *se = &p->se;
5827 int task_sleep = flags & DEQUEUE_SLEEP;
5828 int idle_h_nr_running = task_has_idle_policy(p);
5829 bool was_sched_idle = sched_idle_rq(rq);
5831 util_est_dequeue(&rq->cfs, p);
5833 for_each_sched_entity(se) {
5834 cfs_rq = cfs_rq_of(se);
5835 dequeue_entity(cfs_rq, se, flags);
5837 cfs_rq->h_nr_running--;
5838 cfs_rq->idle_h_nr_running -= idle_h_nr_running;
5840 if (cfs_rq_is_idle(cfs_rq))
5841 idle_h_nr_running = 1;
5843 /* end evaluation on encountering a throttled cfs_rq */
5844 if (cfs_rq_throttled(cfs_rq))
5845 goto dequeue_throttle;
5847 /* Don't dequeue parent if it has other entities besides us */
5848 if (cfs_rq->load.weight) {
5849 /* Avoid re-evaluating load for this entity: */
5850 se = parent_entity(se);
5852 * Bias pick_next to pick a task from this cfs_rq, as
5853 * p is sleeping when it is within its sched_slice.
5855 if (task_sleep && se && !throttled_hierarchy(cfs_rq))
5859 flags |= DEQUEUE_SLEEP;
5862 for_each_sched_entity(se) {
5863 cfs_rq = cfs_rq_of(se);
5865 update_load_avg(cfs_rq, se, UPDATE_TG);
5866 se_update_runnable(se);
5867 update_cfs_group(se);
5869 cfs_rq->h_nr_running--;
5870 cfs_rq->idle_h_nr_running -= idle_h_nr_running;
5872 if (cfs_rq_is_idle(cfs_rq))
5873 idle_h_nr_running = 1;
5875 /* end evaluation on encountering a throttled cfs_rq */
5876 if (cfs_rq_throttled(cfs_rq))
5877 goto dequeue_throttle;
5881 /* At this point se is NULL and we are at root level*/
5882 sub_nr_running(rq, 1);
5884 /* balance early to pull high priority tasks */
5885 if (unlikely(!was_sched_idle && sched_idle_rq(rq)))
5886 rq->next_balance = jiffies;
5889 util_est_update(&rq->cfs, p, task_sleep);
5895 /* Working cpumask for: load_balance, load_balance_newidle. */
5896 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
5897 DEFINE_PER_CPU(cpumask_var_t, select_rq_mask);
5899 #ifdef CONFIG_NO_HZ_COMMON
5902 cpumask_var_t idle_cpus_mask;
5904 int has_blocked; /* Idle CPUS has blocked load */
5905 int needs_update; /* Newly idle CPUs need their next_balance collated */
5906 unsigned long next_balance; /* in jiffy units */
5907 unsigned long next_blocked; /* Next update of blocked load in jiffies */
5908 } nohz ____cacheline_aligned;
5910 #endif /* CONFIG_NO_HZ_COMMON */
5912 static unsigned long cpu_load(struct rq *rq)
5914 return cfs_rq_load_avg(&rq->cfs);
5918 * cpu_load_without - compute CPU load without any contributions from *p
5919 * @cpu: the CPU which load is requested
5920 * @p: the task which load should be discounted
5922 * The load of a CPU is defined by the load of tasks currently enqueued on that
5923 * CPU as well as tasks which are currently sleeping after an execution on that
5926 * This method returns the load of the specified CPU by discounting the load of
5927 * the specified task, whenever the task is currently contributing to the CPU
5930 static unsigned long cpu_load_without(struct rq *rq, struct task_struct *p)
5932 struct cfs_rq *cfs_rq;
5935 /* Task has no contribution or is new */
5936 if (cpu_of(rq) != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time))
5937 return cpu_load(rq);
5940 load = READ_ONCE(cfs_rq->avg.load_avg);
5942 /* Discount task's util from CPU's util */
5943 lsub_positive(&load, task_h_load(p));
5948 static unsigned long cpu_runnable(struct rq *rq)
5950 return cfs_rq_runnable_avg(&rq->cfs);
5953 static unsigned long cpu_runnable_without(struct rq *rq, struct task_struct *p)
5955 struct cfs_rq *cfs_rq;
5956 unsigned int runnable;
5958 /* Task has no contribution or is new */
5959 if (cpu_of(rq) != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time))
5960 return cpu_runnable(rq);
5963 runnable = READ_ONCE(cfs_rq->avg.runnable_avg);
5965 /* Discount task's runnable from CPU's runnable */
5966 lsub_positive(&runnable, p->se.avg.runnable_avg);
5971 static unsigned long capacity_of(int cpu)
5973 return cpu_rq(cpu)->cpu_capacity;
5976 static void record_wakee(struct task_struct *p)
5979 * Only decay a single time; tasks that have less then 1 wakeup per
5980 * jiffy will not have built up many flips.
5982 if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) {
5983 current->wakee_flips >>= 1;
5984 current->wakee_flip_decay_ts = jiffies;
5987 if (current->last_wakee != p) {
5988 current->last_wakee = p;
5989 current->wakee_flips++;
5994 * Detect M:N waker/wakee relationships via a switching-frequency heuristic.
5996 * A waker of many should wake a different task than the one last awakened
5997 * at a frequency roughly N times higher than one of its wakees.
5999 * In order to determine whether we should let the load spread vs consolidating
6000 * to shared cache, we look for a minimum 'flip' frequency of llc_size in one
6001 * partner, and a factor of lls_size higher frequency in the other.
6003 * With both conditions met, we can be relatively sure that the relationship is
6004 * non-monogamous, with partner count exceeding socket size.
6006 * Waker/wakee being client/server, worker/dispatcher, interrupt source or
6007 * whatever is irrelevant, spread criteria is apparent partner count exceeds
6010 static int wake_wide(struct task_struct *p)
6012 unsigned int master = current->wakee_flips;
6013 unsigned int slave = p->wakee_flips;
6014 int factor = __this_cpu_read(sd_llc_size);
6017 swap(master, slave);
6018 if (slave < factor || master < slave * factor)
6024 * The purpose of wake_affine() is to quickly determine on which CPU we can run
6025 * soonest. For the purpose of speed we only consider the waking and previous
6028 * wake_affine_idle() - only considers 'now', it check if the waking CPU is
6029 * cache-affine and is (or will be) idle.
6031 * wake_affine_weight() - considers the weight to reflect the average
6032 * scheduling latency of the CPUs. This seems to work
6033 * for the overloaded case.
6036 wake_affine_idle(int this_cpu, int prev_cpu, int sync)
6039 * If this_cpu is idle, it implies the wakeup is from interrupt
6040 * context. Only allow the move if cache is shared. Otherwise an
6041 * interrupt intensive workload could force all tasks onto one
6042 * node depending on the IO topology or IRQ affinity settings.
6044 * If the prev_cpu is idle and cache affine then avoid a migration.
6045 * There is no guarantee that the cache hot data from an interrupt
6046 * is more important than cache hot data on the prev_cpu and from
6047 * a cpufreq perspective, it's better to have higher utilisation
6050 if (available_idle_cpu(this_cpu) && cpus_share_cache(this_cpu, prev_cpu))
6051 return available_idle_cpu(prev_cpu) ? prev_cpu : this_cpu;
6053 if (sync && cpu_rq(this_cpu)->nr_running == 1)
6056 if (available_idle_cpu(prev_cpu))
6059 return nr_cpumask_bits;
6063 wake_affine_weight(struct sched_domain *sd, struct task_struct *p,
6064 int this_cpu, int prev_cpu, int sync)
6066 s64 this_eff_load, prev_eff_load;
6067 unsigned long task_load;
6069 this_eff_load = cpu_load(cpu_rq(this_cpu));
6072 unsigned long current_load = task_h_load(current);
6074 if (current_load > this_eff_load)
6077 this_eff_load -= current_load;
6080 task_load = task_h_load(p);
6082 this_eff_load += task_load;
6083 if (sched_feat(WA_BIAS))
6084 this_eff_load *= 100;
6085 this_eff_load *= capacity_of(prev_cpu);
6087 prev_eff_load = cpu_load(cpu_rq(prev_cpu));
6088 prev_eff_load -= task_load;
6089 if (sched_feat(WA_BIAS))
6090 prev_eff_load *= 100 + (sd->imbalance_pct - 100) / 2;
6091 prev_eff_load *= capacity_of(this_cpu);
6094 * If sync, adjust the weight of prev_eff_load such that if
6095 * prev_eff == this_eff that select_idle_sibling() will consider
6096 * stacking the wakee on top of the waker if no other CPU is
6102 return this_eff_load < prev_eff_load ? this_cpu : nr_cpumask_bits;
6105 static int wake_affine(struct sched_domain *sd, struct task_struct *p,
6106 int this_cpu, int prev_cpu, int sync)
6108 int target = nr_cpumask_bits;
6110 if (sched_feat(WA_IDLE))
6111 target = wake_affine_idle(this_cpu, prev_cpu, sync);
6113 if (sched_feat(WA_WEIGHT) && target == nr_cpumask_bits)
6114 target = wake_affine_weight(sd, p, this_cpu, prev_cpu, sync);
6116 schedstat_inc(p->stats.nr_wakeups_affine_attempts);
6117 if (target == nr_cpumask_bits)
6120 schedstat_inc(sd->ttwu_move_affine);
6121 schedstat_inc(p->stats.nr_wakeups_affine);
6125 static struct sched_group *
6126 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu);
6129 * find_idlest_group_cpu - find the idlest CPU among the CPUs in the group.
6132 find_idlest_group_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
6134 unsigned long load, min_load = ULONG_MAX;
6135 unsigned int min_exit_latency = UINT_MAX;
6136 u64 latest_idle_timestamp = 0;
6137 int least_loaded_cpu = this_cpu;
6138 int shallowest_idle_cpu = -1;
6141 /* Check if we have any choice: */
6142 if (group->group_weight == 1)
6143 return cpumask_first(sched_group_span(group));
6145 /* Traverse only the allowed CPUs */
6146 for_each_cpu_and(i, sched_group_span(group), p->cpus_ptr) {
6147 struct rq *rq = cpu_rq(i);
6149 if (!sched_core_cookie_match(rq, p))
6152 if (sched_idle_cpu(i))
6155 if (available_idle_cpu(i)) {
6156 struct cpuidle_state *idle = idle_get_state(rq);
6157 if (idle && idle->exit_latency < min_exit_latency) {
6159 * We give priority to a CPU whose idle state
6160 * has the smallest exit latency irrespective
6161 * of any idle timestamp.
6163 min_exit_latency = idle->exit_latency;
6164 latest_idle_timestamp = rq->idle_stamp;
6165 shallowest_idle_cpu = i;
6166 } else if ((!idle || idle->exit_latency == min_exit_latency) &&
6167 rq->idle_stamp > latest_idle_timestamp) {
6169 * If equal or no active idle state, then
6170 * the most recently idled CPU might have
6173 latest_idle_timestamp = rq->idle_stamp;
6174 shallowest_idle_cpu = i;
6176 } else if (shallowest_idle_cpu == -1) {
6177 load = cpu_load(cpu_rq(i));
6178 if (load < min_load) {
6180 least_loaded_cpu = i;
6185 return shallowest_idle_cpu != -1 ? shallowest_idle_cpu : least_loaded_cpu;
6188 static inline int find_idlest_cpu(struct sched_domain *sd, struct task_struct *p,
6189 int cpu, int prev_cpu, int sd_flag)
6193 if (!cpumask_intersects(sched_domain_span(sd), p->cpus_ptr))
6197 * We need task's util for cpu_util_without, sync it up to
6198 * prev_cpu's last_update_time.
6200 if (!(sd_flag & SD_BALANCE_FORK))
6201 sync_entity_load_avg(&p->se);
6204 struct sched_group *group;
6205 struct sched_domain *tmp;
6208 if (!(sd->flags & sd_flag)) {
6213 group = find_idlest_group(sd, p, cpu);
6219 new_cpu = find_idlest_group_cpu(group, p, cpu);
6220 if (new_cpu == cpu) {
6221 /* Now try balancing at a lower domain level of 'cpu': */
6226 /* Now try balancing at a lower domain level of 'new_cpu': */
6228 weight = sd->span_weight;
6230 for_each_domain(cpu, tmp) {
6231 if (weight <= tmp->span_weight)
6233 if (tmp->flags & sd_flag)
6241 static inline int __select_idle_cpu(int cpu, struct task_struct *p)
6243 if ((available_idle_cpu(cpu) || sched_idle_cpu(cpu)) &&
6244 sched_cpu_cookie_match(cpu_rq(cpu), p))
6250 #ifdef CONFIG_SCHED_SMT
6251 DEFINE_STATIC_KEY_FALSE(sched_smt_present);
6252 EXPORT_SYMBOL_GPL(sched_smt_present);
6254 static inline void set_idle_cores(int cpu, int val)
6256 struct sched_domain_shared *sds;
6258 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
6260 WRITE_ONCE(sds->has_idle_cores, val);
6263 static inline bool test_idle_cores(int cpu, bool def)
6265 struct sched_domain_shared *sds;
6267 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
6269 return READ_ONCE(sds->has_idle_cores);
6275 * Scans the local SMT mask to see if the entire core is idle, and records this
6276 * information in sd_llc_shared->has_idle_cores.
6278 * Since SMT siblings share all cache levels, inspecting this limited remote
6279 * state should be fairly cheap.
6281 void __update_idle_core(struct rq *rq)
6283 int core = cpu_of(rq);
6287 if (test_idle_cores(core, true))
6290 for_each_cpu(cpu, cpu_smt_mask(core)) {
6294 if (!available_idle_cpu(cpu))
6298 set_idle_cores(core, 1);
6304 * Scan the entire LLC domain for idle cores; this dynamically switches off if
6305 * there are no idle cores left in the system; tracked through
6306 * sd_llc->shared->has_idle_cores and enabled through update_idle_core() above.
6308 static int select_idle_core(struct task_struct *p, int core, struct cpumask *cpus, int *idle_cpu)
6313 if (!static_branch_likely(&sched_smt_present))
6314 return __select_idle_cpu(core, p);
6316 for_each_cpu(cpu, cpu_smt_mask(core)) {
6317 if (!available_idle_cpu(cpu)) {
6319 if (*idle_cpu == -1) {
6320 if (sched_idle_cpu(cpu) && cpumask_test_cpu(cpu, p->cpus_ptr)) {
6328 if (*idle_cpu == -1 && cpumask_test_cpu(cpu, p->cpus_ptr))
6335 cpumask_andnot(cpus, cpus, cpu_smt_mask(core));
6340 * Scan the local SMT mask for idle CPUs.
6342 static int select_idle_smt(struct task_struct *p, struct sched_domain *sd, int target)
6346 for_each_cpu(cpu, cpu_smt_mask(target)) {
6347 if (!cpumask_test_cpu(cpu, p->cpus_ptr) ||
6348 !cpumask_test_cpu(cpu, sched_domain_span(sd)))
6350 if (available_idle_cpu(cpu) || sched_idle_cpu(cpu))
6357 #else /* CONFIG_SCHED_SMT */
6359 static inline void set_idle_cores(int cpu, int val)
6363 static inline bool test_idle_cores(int cpu, bool def)
6368 static inline int select_idle_core(struct task_struct *p, int core, struct cpumask *cpus, int *idle_cpu)
6370 return __select_idle_cpu(core, p);
6373 static inline int select_idle_smt(struct task_struct *p, struct sched_domain *sd, int target)
6378 #endif /* CONFIG_SCHED_SMT */
6381 * Scan the LLC domain for idle CPUs; this is dynamically regulated by
6382 * comparing the average scan cost (tracked in sd->avg_scan_cost) against the
6383 * average idle time for this rq (as found in rq->avg_idle).
6385 static int select_idle_cpu(struct task_struct *p, struct sched_domain *sd, bool has_idle_core, int target)
6387 struct cpumask *cpus = this_cpu_cpumask_var_ptr(select_rq_mask);
6388 int i, cpu, idle_cpu = -1, nr = INT_MAX;
6389 struct sched_domain_shared *sd_share;
6390 struct rq *this_rq = this_rq();
6391 int this = smp_processor_id();
6392 struct sched_domain *this_sd;
6395 this_sd = rcu_dereference(*this_cpu_ptr(&sd_llc));
6399 cpumask_and(cpus, sched_domain_span(sd), p->cpus_ptr);
6401 if (sched_feat(SIS_PROP) && !has_idle_core) {
6402 u64 avg_cost, avg_idle, span_avg;
6403 unsigned long now = jiffies;
6406 * If we're busy, the assumption that the last idle period
6407 * predicts the future is flawed; age away the remaining
6408 * predicted idle time.
6410 if (unlikely(this_rq->wake_stamp < now)) {
6411 while (this_rq->wake_stamp < now && this_rq->wake_avg_idle) {
6412 this_rq->wake_stamp++;
6413 this_rq->wake_avg_idle >>= 1;
6417 avg_idle = this_rq->wake_avg_idle;
6418 avg_cost = this_sd->avg_scan_cost + 1;
6420 span_avg = sd->span_weight * avg_idle;
6421 if (span_avg > 4*avg_cost)
6422 nr = div_u64(span_avg, avg_cost);
6426 time = cpu_clock(this);
6429 if (sched_feat(SIS_UTIL)) {
6430 sd_share = rcu_dereference(per_cpu(sd_llc_shared, target));
6432 /* because !--nr is the condition to stop scan */
6433 nr = READ_ONCE(sd_share->nr_idle_scan) + 1;
6434 /* overloaded LLC is unlikely to have idle cpu/core */
6440 for_each_cpu_wrap(cpu, cpus, target + 1) {
6441 if (has_idle_core) {
6442 i = select_idle_core(p, cpu, cpus, &idle_cpu);
6443 if ((unsigned int)i < nr_cpumask_bits)
6449 idle_cpu = __select_idle_cpu(cpu, p);
6450 if ((unsigned int)idle_cpu < nr_cpumask_bits)
6456 set_idle_cores(target, false);
6458 if (sched_feat(SIS_PROP) && !has_idle_core) {
6459 time = cpu_clock(this) - time;
6462 * Account for the scan cost of wakeups against the average
6465 this_rq->wake_avg_idle -= min(this_rq->wake_avg_idle, time);
6467 update_avg(&this_sd->avg_scan_cost, time);
6474 * Scan the asym_capacity domain for idle CPUs; pick the first idle one on which
6475 * the task fits. If no CPU is big enough, but there are idle ones, try to
6476 * maximize capacity.
6479 select_idle_capacity(struct task_struct *p, struct sched_domain *sd, int target)
6481 unsigned long task_util, best_cap = 0;
6482 int cpu, best_cpu = -1;
6483 struct cpumask *cpus;
6485 cpus = this_cpu_cpumask_var_ptr(select_rq_mask);
6486 cpumask_and(cpus, sched_domain_span(sd), p->cpus_ptr);
6488 task_util = uclamp_task_util(p);
6490 for_each_cpu_wrap(cpu, cpus, target) {
6491 unsigned long cpu_cap = capacity_of(cpu);
6493 if (!available_idle_cpu(cpu) && !sched_idle_cpu(cpu))
6495 if (fits_capacity(task_util, cpu_cap))
6498 if (cpu_cap > best_cap) {
6507 static inline bool asym_fits_capacity(unsigned long task_util, int cpu)
6509 if (static_branch_unlikely(&sched_asym_cpucapacity))
6510 return fits_capacity(task_util, capacity_of(cpu));
6516 * Try and locate an idle core/thread in the LLC cache domain.
6518 static int select_idle_sibling(struct task_struct *p, int prev, int target)
6520 bool has_idle_core = false;
6521 struct sched_domain *sd;
6522 unsigned long task_util;
6523 int i, recent_used_cpu;
6526 * On asymmetric system, update task utilization because we will check
6527 * that the task fits with cpu's capacity.
6529 if (static_branch_unlikely(&sched_asym_cpucapacity)) {
6530 sync_entity_load_avg(&p->se);
6531 task_util = uclamp_task_util(p);
6535 * per-cpu select_rq_mask usage
6537 lockdep_assert_irqs_disabled();
6539 if ((available_idle_cpu(target) || sched_idle_cpu(target)) &&
6540 asym_fits_capacity(task_util, target))
6544 * If the previous CPU is cache affine and idle, don't be stupid:
6546 if (prev != target && cpus_share_cache(prev, target) &&
6547 (available_idle_cpu(prev) || sched_idle_cpu(prev)) &&
6548 asym_fits_capacity(task_util, prev))
6552 * Allow a per-cpu kthread to stack with the wakee if the
6553 * kworker thread and the tasks previous CPUs are the same.
6554 * The assumption is that the wakee queued work for the
6555 * per-cpu kthread that is now complete and the wakeup is
6556 * essentially a sync wakeup. An obvious example of this
6557 * pattern is IO completions.
6559 if (is_per_cpu_kthread(current) &&
6561 prev == smp_processor_id() &&
6562 this_rq()->nr_running <= 1 &&
6563 asym_fits_capacity(task_util, prev)) {
6567 /* Check a recently used CPU as a potential idle candidate: */
6568 recent_used_cpu = p->recent_used_cpu;
6569 p->recent_used_cpu = prev;
6570 if (recent_used_cpu != prev &&
6571 recent_used_cpu != target &&
6572 cpus_share_cache(recent_used_cpu, target) &&
6573 (available_idle_cpu(recent_used_cpu) || sched_idle_cpu(recent_used_cpu)) &&
6574 cpumask_test_cpu(p->recent_used_cpu, p->cpus_ptr) &&
6575 asym_fits_capacity(task_util, recent_used_cpu)) {
6576 return recent_used_cpu;
6580 * For asymmetric CPU capacity systems, our domain of interest is
6581 * sd_asym_cpucapacity rather than sd_llc.
6583 if (static_branch_unlikely(&sched_asym_cpucapacity)) {
6584 sd = rcu_dereference(per_cpu(sd_asym_cpucapacity, target));
6586 * On an asymmetric CPU capacity system where an exclusive
6587 * cpuset defines a symmetric island (i.e. one unique
6588 * capacity_orig value through the cpuset), the key will be set
6589 * but the CPUs within that cpuset will not have a domain with
6590 * SD_ASYM_CPUCAPACITY. These should follow the usual symmetric
6594 i = select_idle_capacity(p, sd, target);
6595 return ((unsigned)i < nr_cpumask_bits) ? i : target;
6599 sd = rcu_dereference(per_cpu(sd_llc, target));
6603 if (sched_smt_active()) {
6604 has_idle_core = test_idle_cores(target, false);
6606 if (!has_idle_core && cpus_share_cache(prev, target)) {
6607 i = select_idle_smt(p, sd, prev);
6608 if ((unsigned int)i < nr_cpumask_bits)
6613 i = select_idle_cpu(p, sd, has_idle_core, target);
6614 if ((unsigned)i < nr_cpumask_bits)
6621 * Predicts what cpu_util(@cpu) would return if @p was removed from @cpu
6622 * (@dst_cpu = -1) or migrated to @dst_cpu.
6624 static unsigned long cpu_util_next(int cpu, struct task_struct *p, int dst_cpu)
6626 struct cfs_rq *cfs_rq = &cpu_rq(cpu)->cfs;
6627 unsigned long util = READ_ONCE(cfs_rq->avg.util_avg);
6630 * If @dst_cpu is -1 or @p migrates from @cpu to @dst_cpu remove its
6631 * contribution. If @p migrates from another CPU to @cpu add its
6632 * contribution. In all the other cases @cpu is not impacted by the
6633 * migration so its util_avg is already correct.
6635 if (task_cpu(p) == cpu && dst_cpu != cpu)
6636 lsub_positive(&util, task_util(p));
6637 else if (task_cpu(p) != cpu && dst_cpu == cpu)
6638 util += task_util(p);
6640 if (sched_feat(UTIL_EST)) {
6641 unsigned long util_est;
6643 util_est = READ_ONCE(cfs_rq->avg.util_est.enqueued);
6646 * During wake-up @p isn't enqueued yet and doesn't contribute
6647 * to any cpu_rq(cpu)->cfs.avg.util_est.enqueued.
6648 * If @dst_cpu == @cpu add it to "simulate" cpu_util after @p
6649 * has been enqueued.
6651 * During exec (@dst_cpu = -1) @p is enqueued and does
6652 * contribute to cpu_rq(cpu)->cfs.util_est.enqueued.
6653 * Remove it to "simulate" cpu_util without @p's contribution.
6655 * Despite the task_on_rq_queued(@p) check there is still a
6656 * small window for a possible race when an exec
6657 * select_task_rq_fair() races with LB's detach_task().
6661 * p->on_rq = TASK_ON_RQ_MIGRATING;
6662 * -------------------------------- A
6664 * dequeue_task_fair() + Race Time
6665 * util_est_dequeue() /
6666 * -------------------------------- B
6668 * The additional check "current == p" is required to further
6669 * reduce the race window.
6672 util_est += _task_util_est(p);
6673 else if (unlikely(task_on_rq_queued(p) || current == p))
6674 lsub_positive(&util_est, _task_util_est(p));
6676 util = max(util, util_est);
6679 return min(util, capacity_orig_of(cpu));
6683 * cpu_util_without: compute cpu utilization without any contributions from *p
6684 * @cpu: the CPU which utilization is requested
6685 * @p: the task which utilization should be discounted
6687 * The utilization of a CPU is defined by the utilization of tasks currently
6688 * enqueued on that CPU as well as tasks which are currently sleeping after an
6689 * execution on that CPU.
6691 * This method returns the utilization of the specified CPU by discounting the
6692 * utilization of the specified task, whenever the task is currently
6693 * contributing to the CPU utilization.
6695 static unsigned long cpu_util_without(int cpu, struct task_struct *p)
6697 /* Task has no contribution or is new */
6698 if (cpu != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time))
6699 return cpu_util_cfs(cpu);
6701 return cpu_util_next(cpu, p, -1);
6705 * energy_env - Utilization landscape for energy estimation.
6706 * @task_busy_time: Utilization contribution by the task for which we test the
6707 * placement. Given by eenv_task_busy_time().
6708 * @pd_busy_time: Utilization of the whole perf domain without the task
6709 * contribution. Given by eenv_pd_busy_time().
6710 * @cpu_cap: Maximum CPU capacity for the perf domain.
6711 * @pd_cap: Entire perf domain capacity. (pd->nr_cpus * cpu_cap).
6714 unsigned long task_busy_time;
6715 unsigned long pd_busy_time;
6716 unsigned long cpu_cap;
6717 unsigned long pd_cap;
6721 * Compute the task busy time for compute_energy(). This time cannot be
6722 * injected directly into effective_cpu_util() because of the IRQ scaling.
6723 * The latter only makes sense with the most recent CPUs where the task has
6726 static inline void eenv_task_busy_time(struct energy_env *eenv,
6727 struct task_struct *p, int prev_cpu)
6729 unsigned long busy_time, max_cap = arch_scale_cpu_capacity(prev_cpu);
6730 unsigned long irq = cpu_util_irq(cpu_rq(prev_cpu));
6732 if (unlikely(irq >= max_cap))
6733 busy_time = max_cap;
6735 busy_time = scale_irq_capacity(task_util_est(p), irq, max_cap);
6737 eenv->task_busy_time = busy_time;
6741 * Compute the perf_domain (PD) busy time for compute_energy(). Based on the
6742 * utilization for each @pd_cpus, it however doesn't take into account
6743 * clamping since the ratio (utilization / cpu_capacity) is already enough to
6744 * scale the EM reported power consumption at the (eventually clamped)
6747 * The contribution of the task @p for which we want to estimate the
6748 * energy cost is removed (by cpu_util_next()) and must be calculated
6749 * separately (see eenv_task_busy_time). This ensures:
6751 * - A stable PD utilization, no matter which CPU of that PD we want to place
6754 * - A fair comparison between CPUs as the task contribution (task_util())
6755 * will always be the same no matter which CPU utilization we rely on
6756 * (util_avg or util_est).
6758 * Set @eenv busy time for the PD that spans @pd_cpus. This busy time can't
6759 * exceed @eenv->pd_cap.
6761 static inline void eenv_pd_busy_time(struct energy_env *eenv,
6762 struct cpumask *pd_cpus,
6763 struct task_struct *p)
6765 unsigned long busy_time = 0;
6768 for_each_cpu(cpu, pd_cpus) {
6769 unsigned long util = cpu_util_next(cpu, p, -1);
6771 busy_time += effective_cpu_util(cpu, util, ENERGY_UTIL, NULL);
6774 eenv->pd_busy_time = min(eenv->pd_cap, busy_time);
6778 * Compute the maximum utilization for compute_energy() when the task @p
6779 * is placed on the cpu @dst_cpu.
6781 * Returns the maximum utilization among @eenv->cpus. This utilization can't
6782 * exceed @eenv->cpu_cap.
6784 static inline unsigned long
6785 eenv_pd_max_util(struct energy_env *eenv, struct cpumask *pd_cpus,
6786 struct task_struct *p, int dst_cpu)
6788 unsigned long max_util = 0;
6791 for_each_cpu(cpu, pd_cpus) {
6792 struct task_struct *tsk = (cpu == dst_cpu) ? p : NULL;
6793 unsigned long util = cpu_util_next(cpu, p, dst_cpu);
6794 unsigned long cpu_util;
6797 * Performance domain frequency: utilization clamping
6798 * must be considered since it affects the selection
6799 * of the performance domain frequency.
6800 * NOTE: in case RT tasks are running, by default the
6801 * FREQUENCY_UTIL's utilization can be max OPP.
6803 cpu_util = effective_cpu_util(cpu, util, FREQUENCY_UTIL, tsk);
6804 max_util = max(max_util, cpu_util);
6807 return min(max_util, eenv->cpu_cap);
6811 * compute_energy(): Use the Energy Model to estimate the energy that @pd would
6812 * consume for a given utilization landscape @eenv. When @dst_cpu < 0, the task
6813 * contribution is ignored.
6815 static inline unsigned long
6816 compute_energy(struct energy_env *eenv, struct perf_domain *pd,
6817 struct cpumask *pd_cpus, struct task_struct *p, int dst_cpu)
6819 unsigned long max_util = eenv_pd_max_util(eenv, pd_cpus, p, dst_cpu);
6820 unsigned long busy_time = eenv->pd_busy_time;
6823 busy_time = min(eenv->pd_cap, busy_time + eenv->task_busy_time);
6825 return em_cpu_energy(pd->em_pd, max_util, busy_time, eenv->cpu_cap);
6829 * find_energy_efficient_cpu(): Find most energy-efficient target CPU for the
6830 * waking task. find_energy_efficient_cpu() looks for the CPU with maximum
6831 * spare capacity in each performance domain and uses it as a potential
6832 * candidate to execute the task. Then, it uses the Energy Model to figure
6833 * out which of the CPU candidates is the most energy-efficient.
6835 * The rationale for this heuristic is as follows. In a performance domain,
6836 * all the most energy efficient CPU candidates (according to the Energy
6837 * Model) are those for which we'll request a low frequency. When there are
6838 * several CPUs for which the frequency request will be the same, we don't
6839 * have enough data to break the tie between them, because the Energy Model
6840 * only includes active power costs. With this model, if we assume that
6841 * frequency requests follow utilization (e.g. using schedutil), the CPU with
6842 * the maximum spare capacity in a performance domain is guaranteed to be among
6843 * the best candidates of the performance domain.
6845 * In practice, it could be preferable from an energy standpoint to pack
6846 * small tasks on a CPU in order to let other CPUs go in deeper idle states,
6847 * but that could also hurt our chances to go cluster idle, and we have no
6848 * ways to tell with the current Energy Model if this is actually a good
6849 * idea or not. So, find_energy_efficient_cpu() basically favors
6850 * cluster-packing, and spreading inside a cluster. That should at least be
6851 * a good thing for latency, and this is consistent with the idea that most
6852 * of the energy savings of EAS come from the asymmetry of the system, and
6853 * not so much from breaking the tie between identical CPUs. That's also the
6854 * reason why EAS is enabled in the topology code only for systems where
6855 * SD_ASYM_CPUCAPACITY is set.
6857 * NOTE: Forkees are not accepted in the energy-aware wake-up path because
6858 * they don't have any useful utilization data yet and it's not possible to
6859 * forecast their impact on energy consumption. Consequently, they will be
6860 * placed by find_idlest_cpu() on the least loaded CPU, which might turn out
6861 * to be energy-inefficient in some use-cases. The alternative would be to
6862 * bias new tasks towards specific types of CPUs first, or to try to infer
6863 * their util_avg from the parent task, but those heuristics could hurt
6864 * other use-cases too. So, until someone finds a better way to solve this,
6865 * let's keep things simple by re-using the existing slow path.
6867 static int find_energy_efficient_cpu(struct task_struct *p, int prev_cpu)
6869 struct cpumask *cpus = this_cpu_cpumask_var_ptr(select_rq_mask);
6870 unsigned long prev_delta = ULONG_MAX, best_delta = ULONG_MAX;
6871 struct root_domain *rd = this_rq()->rd;
6872 int cpu, best_energy_cpu, target = -1;
6873 struct sched_domain *sd;
6874 struct perf_domain *pd;
6875 struct energy_env eenv;
6878 pd = rcu_dereference(rd->pd);
6879 if (!pd || READ_ONCE(rd->overutilized))
6883 * Energy-aware wake-up happens on the lowest sched_domain starting
6884 * from sd_asym_cpucapacity spanning over this_cpu and prev_cpu.
6886 sd = rcu_dereference(*this_cpu_ptr(&sd_asym_cpucapacity));
6887 while (sd && !cpumask_test_cpu(prev_cpu, sched_domain_span(sd)))
6894 sync_entity_load_avg(&p->se);
6895 if (!task_util_est(p))
6898 eenv_task_busy_time(&eenv, p, prev_cpu);
6900 for (; pd; pd = pd->next) {
6901 unsigned long cpu_cap, cpu_thermal_cap, util;
6902 unsigned long cur_delta, max_spare_cap = 0;
6903 bool compute_prev_delta = false;
6904 int max_spare_cap_cpu = -1;
6905 unsigned long base_energy;
6907 cpumask_and(cpus, perf_domain_span(pd), cpu_online_mask);
6909 if (cpumask_empty(cpus))
6912 /* Account thermal pressure for the energy estimation */
6913 cpu = cpumask_first(cpus);
6914 cpu_thermal_cap = arch_scale_cpu_capacity(cpu);
6915 cpu_thermal_cap -= arch_scale_thermal_pressure(cpu);
6917 eenv.cpu_cap = cpu_thermal_cap;
6920 for_each_cpu(cpu, cpus) {
6921 eenv.pd_cap += cpu_thermal_cap;
6923 if (!cpumask_test_cpu(cpu, sched_domain_span(sd)))
6926 if (!cpumask_test_cpu(cpu, p->cpus_ptr))
6929 util = cpu_util_next(cpu, p, cpu);
6930 cpu_cap = capacity_of(cpu);
6933 * Skip CPUs that cannot satisfy the capacity request.
6934 * IOW, placing the task there would make the CPU
6935 * overutilized. Take uclamp into account to see how
6936 * much capacity we can get out of the CPU; this is
6937 * aligned with sched_cpu_util().
6939 util = uclamp_rq_util_with(cpu_rq(cpu), util, p);
6940 if (!fits_capacity(util, cpu_cap))
6943 lsub_positive(&cpu_cap, util);
6945 if (cpu == prev_cpu) {
6946 /* Always use prev_cpu as a candidate. */
6947 compute_prev_delta = true;
6948 } else if (cpu_cap > max_spare_cap) {
6950 * Find the CPU with the maximum spare capacity
6951 * in the performance domain.
6953 max_spare_cap = cpu_cap;
6954 max_spare_cap_cpu = cpu;
6958 if (max_spare_cap_cpu < 0 && !compute_prev_delta)
6961 eenv_pd_busy_time(&eenv, cpus, p);
6962 /* Compute the 'base' energy of the pd, without @p */
6963 base_energy = compute_energy(&eenv, pd, cpus, p, -1);
6965 /* Evaluate the energy impact of using prev_cpu. */
6966 if (compute_prev_delta) {
6967 prev_delta = compute_energy(&eenv, pd, cpus, p,
6969 /* CPU utilization has changed */
6970 if (prev_delta < base_energy)
6972 prev_delta -= base_energy;
6973 best_delta = min(best_delta, prev_delta);
6976 /* Evaluate the energy impact of using max_spare_cap_cpu. */
6977 if (max_spare_cap_cpu >= 0) {
6978 cur_delta = compute_energy(&eenv, pd, cpus, p,
6980 /* CPU utilization has changed */
6981 if (cur_delta < base_energy)
6983 cur_delta -= base_energy;
6984 if (cur_delta < best_delta) {
6985 best_delta = cur_delta;
6986 best_energy_cpu = max_spare_cap_cpu;
6992 if (best_delta < prev_delta)
6993 target = best_energy_cpu;
7004 * select_task_rq_fair: Select target runqueue for the waking task in domains
7005 * that have the relevant SD flag set. In practice, this is SD_BALANCE_WAKE,
7006 * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
7008 * Balances load by selecting the idlest CPU in the idlest group, or under
7009 * certain conditions an idle sibling CPU if the domain has SD_WAKE_AFFINE set.
7011 * Returns the target CPU number.
7014 select_task_rq_fair(struct task_struct *p, int prev_cpu, int wake_flags)
7016 int sync = (wake_flags & WF_SYNC) && !(current->flags & PF_EXITING);
7017 struct sched_domain *tmp, *sd = NULL;
7018 int cpu = smp_processor_id();
7019 int new_cpu = prev_cpu;
7020 int want_affine = 0;
7021 /* SD_flags and WF_flags share the first nibble */
7022 int sd_flag = wake_flags & 0xF;
7025 * required for stable ->cpus_allowed
7027 lockdep_assert_held(&p->pi_lock);
7028 if (wake_flags & WF_TTWU) {
7031 if (sched_energy_enabled()) {
7032 new_cpu = find_energy_efficient_cpu(p, prev_cpu);
7038 want_affine = !wake_wide(p) && cpumask_test_cpu(cpu, p->cpus_ptr);
7042 for_each_domain(cpu, tmp) {
7044 * If both 'cpu' and 'prev_cpu' are part of this domain,
7045 * cpu is a valid SD_WAKE_AFFINE target.
7047 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
7048 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
7049 if (cpu != prev_cpu)
7050 new_cpu = wake_affine(tmp, p, cpu, prev_cpu, sync);
7052 sd = NULL; /* Prefer wake_affine over balance flags */
7057 * Usually only true for WF_EXEC and WF_FORK, as sched_domains
7058 * usually do not have SD_BALANCE_WAKE set. That means wakeup
7059 * will usually go to the fast path.
7061 if (tmp->flags & sd_flag)
7063 else if (!want_affine)
7069 new_cpu = find_idlest_cpu(sd, p, cpu, prev_cpu, sd_flag);
7070 } else if (wake_flags & WF_TTWU) { /* XXX always ? */
7072 new_cpu = select_idle_sibling(p, prev_cpu, new_cpu);
7079 static void detach_entity_cfs_rq(struct sched_entity *se);
7082 * Called immediately before a task is migrated to a new CPU; task_cpu(p) and
7083 * cfs_rq_of(p) references at time of call are still valid and identify the
7084 * previous CPU. The caller guarantees p->pi_lock or task_rq(p)->lock is held.
7086 static void migrate_task_rq_fair(struct task_struct *p, int new_cpu)
7088 struct sched_entity *se = &p->se;
7091 * As blocked tasks retain absolute vruntime the migration needs to
7092 * deal with this by subtracting the old and adding the new
7093 * min_vruntime -- the latter is done by enqueue_entity() when placing
7094 * the task on the new runqueue.
7096 if (READ_ONCE(p->__state) == TASK_WAKING) {
7097 struct cfs_rq *cfs_rq = cfs_rq_of(se);
7099 se->vruntime -= u64_u32_load(cfs_rq->min_vruntime);
7102 if (p->on_rq == TASK_ON_RQ_MIGRATING) {
7104 * In case of TASK_ON_RQ_MIGRATING we in fact hold the 'old'
7105 * rq->lock and can modify state directly.
7107 lockdep_assert_rq_held(task_rq(p));
7108 detach_entity_cfs_rq(se);
7111 remove_entity_load_avg(se);
7114 * Here, the task's PELT values have been updated according to
7115 * the current rq's clock. But if that clock hasn't been
7116 * updated in a while, a substantial idle time will be missed,
7117 * leading to an inflation after wake-up on the new rq.
7119 * Estimate the missing time from the cfs_rq last_update_time
7120 * and update sched_avg to improve the PELT continuity after
7123 migrate_se_pelt_lag(se);
7126 /* Tell new CPU we are migrated */
7127 se->avg.last_update_time = 0;
7129 /* We have migrated, no longer consider this task hot */
7132 update_scan_period(p, new_cpu);
7135 static void task_dead_fair(struct task_struct *p)
7137 remove_entity_load_avg(&p->se);
7141 balance_fair(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
7146 return newidle_balance(rq, rf) != 0;
7148 #endif /* CONFIG_SMP */
7150 static unsigned long wakeup_gran(struct sched_entity *se)
7152 unsigned long gran = sysctl_sched_wakeup_granularity;
7155 * Since its curr running now, convert the gran from real-time
7156 * to virtual-time in his units.
7158 * By using 'se' instead of 'curr' we penalize light tasks, so
7159 * they get preempted easier. That is, if 'se' < 'curr' then
7160 * the resulting gran will be larger, therefore penalizing the
7161 * lighter, if otoh 'se' > 'curr' then the resulting gran will
7162 * be smaller, again penalizing the lighter task.
7164 * This is especially important for buddies when the leftmost
7165 * task is higher priority than the buddy.
7167 return calc_delta_fair(gran, se);
7171 * Should 'se' preempt 'curr'.
7185 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
7187 s64 gran, vdiff = curr->vruntime - se->vruntime;
7192 gran = wakeup_gran(se);
7199 static void set_last_buddy(struct sched_entity *se)
7201 for_each_sched_entity(se) {
7202 if (SCHED_WARN_ON(!se->on_rq))
7206 cfs_rq_of(se)->last = se;
7210 static void set_next_buddy(struct sched_entity *se)
7212 for_each_sched_entity(se) {
7213 if (SCHED_WARN_ON(!se->on_rq))
7217 cfs_rq_of(se)->next = se;
7221 static void set_skip_buddy(struct sched_entity *se)
7223 for_each_sched_entity(se)
7224 cfs_rq_of(se)->skip = se;
7228 * Preempt the current task with a newly woken task if needed:
7230 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
7232 struct task_struct *curr = rq->curr;
7233 struct sched_entity *se = &curr->se, *pse = &p->se;
7234 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
7235 int scale = cfs_rq->nr_running >= sched_nr_latency;
7236 int next_buddy_marked = 0;
7237 int cse_is_idle, pse_is_idle;
7239 if (unlikely(se == pse))
7243 * This is possible from callers such as attach_tasks(), in which we
7244 * unconditionally check_preempt_curr() after an enqueue (which may have
7245 * lead to a throttle). This both saves work and prevents false
7246 * next-buddy nomination below.
7248 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
7251 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
7252 set_next_buddy(pse);
7253 next_buddy_marked = 1;
7257 * We can come here with TIF_NEED_RESCHED already set from new task
7260 * Note: this also catches the edge-case of curr being in a throttled
7261 * group (e.g. via set_curr_task), since update_curr() (in the
7262 * enqueue of curr) will have resulted in resched being set. This
7263 * prevents us from potentially nominating it as a false LAST_BUDDY
7266 if (test_tsk_need_resched(curr))
7269 /* Idle tasks are by definition preempted by non-idle tasks. */
7270 if (unlikely(task_has_idle_policy(curr)) &&
7271 likely(!task_has_idle_policy(p)))
7275 * Batch and idle tasks do not preempt non-idle tasks (their preemption
7276 * is driven by the tick):
7278 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
7281 find_matching_se(&se, &pse);
7284 cse_is_idle = se_is_idle(se);
7285 pse_is_idle = se_is_idle(pse);
7288 * Preempt an idle group in favor of a non-idle group (and don't preempt
7289 * in the inverse case).
7291 if (cse_is_idle && !pse_is_idle)
7293 if (cse_is_idle != pse_is_idle)
7296 update_curr(cfs_rq_of(se));
7297 if (wakeup_preempt_entity(se, pse) == 1) {
7299 * Bias pick_next to pick the sched entity that is
7300 * triggering this preemption.
7302 if (!next_buddy_marked)
7303 set_next_buddy(pse);
7312 * Only set the backward buddy when the current task is still
7313 * on the rq. This can happen when a wakeup gets interleaved
7314 * with schedule on the ->pre_schedule() or idle_balance()
7315 * point, either of which can * drop the rq lock.
7317 * Also, during early boot the idle thread is in the fair class,
7318 * for obvious reasons its a bad idea to schedule back to it.
7320 if (unlikely(!se->on_rq || curr == rq->idle))
7323 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
7328 static struct task_struct *pick_task_fair(struct rq *rq)
7330 struct sched_entity *se;
7331 struct cfs_rq *cfs_rq;
7335 if (!cfs_rq->nr_running)
7339 struct sched_entity *curr = cfs_rq->curr;
7341 /* When we pick for a remote RQ, we'll not have done put_prev_entity() */
7344 update_curr(cfs_rq);
7348 if (unlikely(check_cfs_rq_runtime(cfs_rq)))
7352 se = pick_next_entity(cfs_rq, curr);
7353 cfs_rq = group_cfs_rq(se);
7360 struct task_struct *
7361 pick_next_task_fair(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
7363 struct cfs_rq *cfs_rq = &rq->cfs;
7364 struct sched_entity *se;
7365 struct task_struct *p;
7369 if (!sched_fair_runnable(rq))
7372 #ifdef CONFIG_FAIR_GROUP_SCHED
7373 if (!prev || prev->sched_class != &fair_sched_class)
7377 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
7378 * likely that a next task is from the same cgroup as the current.
7380 * Therefore attempt to avoid putting and setting the entire cgroup
7381 * hierarchy, only change the part that actually changes.
7385 struct sched_entity *curr = cfs_rq->curr;
7388 * Since we got here without doing put_prev_entity() we also
7389 * have to consider cfs_rq->curr. If it is still a runnable
7390 * entity, update_curr() will update its vruntime, otherwise
7391 * forget we've ever seen it.
7395 update_curr(cfs_rq);
7400 * This call to check_cfs_rq_runtime() will do the
7401 * throttle and dequeue its entity in the parent(s).
7402 * Therefore the nr_running test will indeed
7405 if (unlikely(check_cfs_rq_runtime(cfs_rq))) {
7408 if (!cfs_rq->nr_running)
7415 se = pick_next_entity(cfs_rq, curr);
7416 cfs_rq = group_cfs_rq(se);
7422 * Since we haven't yet done put_prev_entity and if the selected task
7423 * is a different task than we started out with, try and touch the
7424 * least amount of cfs_rqs.
7427 struct sched_entity *pse = &prev->se;
7429 while (!(cfs_rq = is_same_group(se, pse))) {
7430 int se_depth = se->depth;
7431 int pse_depth = pse->depth;
7433 if (se_depth <= pse_depth) {
7434 put_prev_entity(cfs_rq_of(pse), pse);
7435 pse = parent_entity(pse);
7437 if (se_depth >= pse_depth) {
7438 set_next_entity(cfs_rq_of(se), se);
7439 se = parent_entity(se);
7443 put_prev_entity(cfs_rq, pse);
7444 set_next_entity(cfs_rq, se);
7451 put_prev_task(rq, prev);
7454 se = pick_next_entity(cfs_rq, NULL);
7455 set_next_entity(cfs_rq, se);
7456 cfs_rq = group_cfs_rq(se);
7461 done: __maybe_unused;
7464 * Move the next running task to the front of
7465 * the list, so our cfs_tasks list becomes MRU
7468 list_move(&p->se.group_node, &rq->cfs_tasks);
7471 if (hrtick_enabled_fair(rq))
7472 hrtick_start_fair(rq, p);
7474 update_misfit_status(p, rq);
7482 new_tasks = newidle_balance(rq, rf);
7485 * Because newidle_balance() releases (and re-acquires) rq->lock, it is
7486 * possible for any higher priority task to appear. In that case we
7487 * must re-start the pick_next_entity() loop.
7496 * rq is about to be idle, check if we need to update the
7497 * lost_idle_time of clock_pelt
7499 update_idle_rq_clock_pelt(rq);
7504 static struct task_struct *__pick_next_task_fair(struct rq *rq)
7506 return pick_next_task_fair(rq, NULL, NULL);
7510 * Account for a descheduled task:
7512 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
7514 struct sched_entity *se = &prev->se;
7515 struct cfs_rq *cfs_rq;
7517 for_each_sched_entity(se) {
7518 cfs_rq = cfs_rq_of(se);
7519 put_prev_entity(cfs_rq, se);
7524 * sched_yield() is very simple
7526 * The magic of dealing with the ->skip buddy is in pick_next_entity.
7528 static void yield_task_fair(struct rq *rq)
7530 struct task_struct *curr = rq->curr;
7531 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
7532 struct sched_entity *se = &curr->se;
7535 * Are we the only task in the tree?
7537 if (unlikely(rq->nr_running == 1))
7540 clear_buddies(cfs_rq, se);
7542 if (curr->policy != SCHED_BATCH) {
7543 update_rq_clock(rq);
7545 * Update run-time statistics of the 'current'.
7547 update_curr(cfs_rq);
7549 * Tell update_rq_clock() that we've just updated,
7550 * so we don't do microscopic update in schedule()
7551 * and double the fastpath cost.
7553 rq_clock_skip_update(rq);
7559 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p)
7561 struct sched_entity *se = &p->se;
7563 /* throttled hierarchies are not runnable */
7564 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
7567 /* Tell the scheduler that we'd really like pse to run next. */
7570 yield_task_fair(rq);
7576 /**************************************************
7577 * Fair scheduling class load-balancing methods.
7581 * The purpose of load-balancing is to achieve the same basic fairness the
7582 * per-CPU scheduler provides, namely provide a proportional amount of compute
7583 * time to each task. This is expressed in the following equation:
7585 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
7587 * Where W_i,n is the n-th weight average for CPU i. The instantaneous weight
7588 * W_i,0 is defined as:
7590 * W_i,0 = \Sum_j w_i,j (2)
7592 * Where w_i,j is the weight of the j-th runnable task on CPU i. This weight
7593 * is derived from the nice value as per sched_prio_to_weight[].
7595 * The weight average is an exponential decay average of the instantaneous
7598 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
7600 * C_i is the compute capacity of CPU i, typically it is the
7601 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
7602 * can also include other factors [XXX].
7604 * To achieve this balance we define a measure of imbalance which follows
7605 * directly from (1):
7607 * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4)
7609 * We them move tasks around to minimize the imbalance. In the continuous
7610 * function space it is obvious this converges, in the discrete case we get
7611 * a few fun cases generally called infeasible weight scenarios.
7614 * - infeasible weights;
7615 * - local vs global optima in the discrete case. ]
7620 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
7621 * for all i,j solution, we create a tree of CPUs that follows the hardware
7622 * topology where each level pairs two lower groups (or better). This results
7623 * in O(log n) layers. Furthermore we reduce the number of CPUs going up the
7624 * tree to only the first of the previous level and we decrease the frequency
7625 * of load-balance at each level inv. proportional to the number of CPUs in
7631 * \Sum { --- * --- * 2^i } = O(n) (5)
7633 * `- size of each group
7634 * | | `- number of CPUs doing load-balance
7636 * `- sum over all levels
7638 * Coupled with a limit on how many tasks we can migrate every balance pass,
7639 * this makes (5) the runtime complexity of the balancer.
7641 * An important property here is that each CPU is still (indirectly) connected
7642 * to every other CPU in at most O(log n) steps:
7644 * The adjacency matrix of the resulting graph is given by:
7647 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
7650 * And you'll find that:
7652 * A^(log_2 n)_i,j != 0 for all i,j (7)
7654 * Showing there's indeed a path between every CPU in at most O(log n) steps.
7655 * The task movement gives a factor of O(m), giving a convergence complexity
7658 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
7663 * In order to avoid CPUs going idle while there's still work to do, new idle
7664 * balancing is more aggressive and has the newly idle CPU iterate up the domain
7665 * tree itself instead of relying on other CPUs to bring it work.
7667 * This adds some complexity to both (5) and (8) but it reduces the total idle
7675 * Cgroups make a horror show out of (2), instead of a simple sum we get:
7678 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
7683 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
7685 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on CPU i.
7687 * The big problem is S_k, its a global sum needed to compute a local (W_i)
7690 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
7691 * rewrite all of this once again.]
7694 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
7696 enum fbq_type { regular, remote, all };
7699 * 'group_type' describes the group of CPUs at the moment of load balancing.
7701 * The enum is ordered by pulling priority, with the group with lowest priority
7702 * first so the group_type can simply be compared when selecting the busiest
7703 * group. See update_sd_pick_busiest().
7706 /* The group has spare capacity that can be used to run more tasks. */
7707 group_has_spare = 0,
7709 * The group is fully used and the tasks don't compete for more CPU
7710 * cycles. Nevertheless, some tasks might wait before running.
7714 * One task doesn't fit with CPU's capacity and must be migrated to a
7715 * more powerful CPU.
7719 * SD_ASYM_PACKING only: One local CPU with higher capacity is available,
7720 * and the task should be migrated to it instead of running on the
7725 * The tasks' affinity constraints previously prevented the scheduler
7726 * from balancing the load across the system.
7730 * The CPU is overloaded and can't provide expected CPU cycles to all
7736 enum migration_type {
7743 #define LBF_ALL_PINNED 0x01
7744 #define LBF_NEED_BREAK 0x02
7745 #define LBF_DST_PINNED 0x04
7746 #define LBF_SOME_PINNED 0x08
7747 #define LBF_ACTIVE_LB 0x10
7750 struct sched_domain *sd;
7758 struct cpumask *dst_grpmask;
7760 enum cpu_idle_type idle;
7762 /* The set of CPUs under consideration for load-balancing */
7763 struct cpumask *cpus;
7768 unsigned int loop_break;
7769 unsigned int loop_max;
7771 enum fbq_type fbq_type;
7772 enum migration_type migration_type;
7773 struct list_head tasks;
7777 * Is this task likely cache-hot:
7779 static int task_hot(struct task_struct *p, struct lb_env *env)
7783 lockdep_assert_rq_held(env->src_rq);
7785 if (p->sched_class != &fair_sched_class)
7788 if (unlikely(task_has_idle_policy(p)))
7791 /* SMT siblings share cache */
7792 if (env->sd->flags & SD_SHARE_CPUCAPACITY)
7796 * Buddy candidates are cache hot:
7798 if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running &&
7799 (&p->se == cfs_rq_of(&p->se)->next ||
7800 &p->se == cfs_rq_of(&p->se)->last))
7803 if (sysctl_sched_migration_cost == -1)
7807 * Don't migrate task if the task's cookie does not match
7808 * with the destination CPU's core cookie.
7810 if (!sched_core_cookie_match(cpu_rq(env->dst_cpu), p))
7813 if (sysctl_sched_migration_cost == 0)
7816 delta = rq_clock_task(env->src_rq) - p->se.exec_start;
7818 return delta < (s64)sysctl_sched_migration_cost;
7821 #ifdef CONFIG_NUMA_BALANCING
7823 * Returns 1, if task migration degrades locality
7824 * Returns 0, if task migration improves locality i.e migration preferred.
7825 * Returns -1, if task migration is not affected by locality.
7827 static int migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
7829 struct numa_group *numa_group = rcu_dereference(p->numa_group);
7830 unsigned long src_weight, dst_weight;
7831 int src_nid, dst_nid, dist;
7833 if (!static_branch_likely(&sched_numa_balancing))
7836 if (!p->numa_faults || !(env->sd->flags & SD_NUMA))
7839 src_nid = cpu_to_node(env->src_cpu);
7840 dst_nid = cpu_to_node(env->dst_cpu);
7842 if (src_nid == dst_nid)
7845 /* Migrating away from the preferred node is always bad. */
7846 if (src_nid == p->numa_preferred_nid) {
7847 if (env->src_rq->nr_running > env->src_rq->nr_preferred_running)
7853 /* Encourage migration to the preferred node. */
7854 if (dst_nid == p->numa_preferred_nid)
7857 /* Leaving a core idle is often worse than degrading locality. */
7858 if (env->idle == CPU_IDLE)
7861 dist = node_distance(src_nid, dst_nid);
7863 src_weight = group_weight(p, src_nid, dist);
7864 dst_weight = group_weight(p, dst_nid, dist);
7866 src_weight = task_weight(p, src_nid, dist);
7867 dst_weight = task_weight(p, dst_nid, dist);
7870 return dst_weight < src_weight;
7874 static inline int migrate_degrades_locality(struct task_struct *p,
7882 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
7885 int can_migrate_task(struct task_struct *p, struct lb_env *env)
7889 lockdep_assert_rq_held(env->src_rq);
7892 * We do not migrate tasks that are:
7893 * 1) throttled_lb_pair, or
7894 * 2) cannot be migrated to this CPU due to cpus_ptr, or
7895 * 3) running (obviously), or
7896 * 4) are cache-hot on their current CPU.
7898 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
7901 /* Disregard pcpu kthreads; they are where they need to be. */
7902 if (kthread_is_per_cpu(p))
7905 if (!cpumask_test_cpu(env->dst_cpu, p->cpus_ptr)) {
7908 schedstat_inc(p->stats.nr_failed_migrations_affine);
7910 env->flags |= LBF_SOME_PINNED;
7913 * Remember if this task can be migrated to any other CPU in
7914 * our sched_group. We may want to revisit it if we couldn't
7915 * meet load balance goals by pulling other tasks on src_cpu.
7917 * Avoid computing new_dst_cpu
7919 * - if we have already computed one in current iteration
7920 * - if it's an active balance
7922 if (env->idle == CPU_NEWLY_IDLE ||
7923 env->flags & (LBF_DST_PINNED | LBF_ACTIVE_LB))
7926 /* Prevent to re-select dst_cpu via env's CPUs: */
7927 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
7928 if (cpumask_test_cpu(cpu, p->cpus_ptr)) {
7929 env->flags |= LBF_DST_PINNED;
7930 env->new_dst_cpu = cpu;
7938 /* Record that we found at least one task that could run on dst_cpu */
7939 env->flags &= ~LBF_ALL_PINNED;
7941 if (task_running(env->src_rq, p)) {
7942 schedstat_inc(p->stats.nr_failed_migrations_running);
7947 * Aggressive migration if:
7949 * 2) destination numa is preferred
7950 * 3) task is cache cold, or
7951 * 4) too many balance attempts have failed.
7953 if (env->flags & LBF_ACTIVE_LB)
7956 tsk_cache_hot = migrate_degrades_locality(p, env);
7957 if (tsk_cache_hot == -1)
7958 tsk_cache_hot = task_hot(p, env);
7960 if (tsk_cache_hot <= 0 ||
7961 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
7962 if (tsk_cache_hot == 1) {
7963 schedstat_inc(env->sd->lb_hot_gained[env->idle]);
7964 schedstat_inc(p->stats.nr_forced_migrations);
7969 schedstat_inc(p->stats.nr_failed_migrations_hot);
7974 * detach_task() -- detach the task for the migration specified in env
7976 static void detach_task(struct task_struct *p, struct lb_env *env)
7978 lockdep_assert_rq_held(env->src_rq);
7980 deactivate_task(env->src_rq, p, DEQUEUE_NOCLOCK);
7981 set_task_cpu(p, env->dst_cpu);
7985 * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
7986 * part of active balancing operations within "domain".
7988 * Returns a task if successful and NULL otherwise.
7990 static struct task_struct *detach_one_task(struct lb_env *env)
7992 struct task_struct *p;
7994 lockdep_assert_rq_held(env->src_rq);
7996 list_for_each_entry_reverse(p,
7997 &env->src_rq->cfs_tasks, se.group_node) {
7998 if (!can_migrate_task(p, env))
8001 detach_task(p, env);
8004 * Right now, this is only the second place where
8005 * lb_gained[env->idle] is updated (other is detach_tasks)
8006 * so we can safely collect stats here rather than
8007 * inside detach_tasks().
8009 schedstat_inc(env->sd->lb_gained[env->idle]);
8015 static const unsigned int sched_nr_migrate_break = 32;
8018 * detach_tasks() -- tries to detach up to imbalance load/util/tasks from
8019 * busiest_rq, as part of a balancing operation within domain "sd".
8021 * Returns number of detached tasks if successful and 0 otherwise.
8023 static int detach_tasks(struct lb_env *env)
8025 struct list_head *tasks = &env->src_rq->cfs_tasks;
8026 unsigned long util, load;
8027 struct task_struct *p;
8030 lockdep_assert_rq_held(env->src_rq);
8033 * Source run queue has been emptied by another CPU, clear
8034 * LBF_ALL_PINNED flag as we will not test any task.
8036 if (env->src_rq->nr_running <= 1) {
8037 env->flags &= ~LBF_ALL_PINNED;
8041 if (env->imbalance <= 0)
8044 while (!list_empty(tasks)) {
8046 * We don't want to steal all, otherwise we may be treated likewise,
8047 * which could at worst lead to a livelock crash.
8049 if (env->idle != CPU_NOT_IDLE && env->src_rq->nr_running <= 1)
8052 p = list_last_entry(tasks, struct task_struct, se.group_node);
8055 /* We've more or less seen every task there is, call it quits */
8056 if (env->loop > env->loop_max)
8059 /* take a breather every nr_migrate tasks */
8060 if (env->loop > env->loop_break) {
8061 env->loop_break += sched_nr_migrate_break;
8062 env->flags |= LBF_NEED_BREAK;
8066 if (!can_migrate_task(p, env))
8069 switch (env->migration_type) {
8072 * Depending of the number of CPUs and tasks and the
8073 * cgroup hierarchy, task_h_load() can return a null
8074 * value. Make sure that env->imbalance decreases
8075 * otherwise detach_tasks() will stop only after
8076 * detaching up to loop_max tasks.
8078 load = max_t(unsigned long, task_h_load(p), 1);
8080 if (sched_feat(LB_MIN) &&
8081 load < 16 && !env->sd->nr_balance_failed)
8085 * Make sure that we don't migrate too much load.
8086 * Nevertheless, let relax the constraint if
8087 * scheduler fails to find a good waiting task to
8090 if (shr_bound(load, env->sd->nr_balance_failed) > env->imbalance)
8093 env->imbalance -= load;
8097 util = task_util_est(p);
8099 if (util > env->imbalance)
8102 env->imbalance -= util;
8109 case migrate_misfit:
8110 /* This is not a misfit task */
8111 if (task_fits_capacity(p, capacity_of(env->src_cpu)))
8118 detach_task(p, env);
8119 list_add(&p->se.group_node, &env->tasks);
8123 #ifdef CONFIG_PREEMPTION
8125 * NEWIDLE balancing is a source of latency, so preemptible
8126 * kernels will stop after the first task is detached to minimize
8127 * the critical section.
8129 if (env->idle == CPU_NEWLY_IDLE)
8134 * We only want to steal up to the prescribed amount of
8137 if (env->imbalance <= 0)
8142 list_move(&p->se.group_node, tasks);
8146 * Right now, this is one of only two places we collect this stat
8147 * so we can safely collect detach_one_task() stats here rather
8148 * than inside detach_one_task().
8150 schedstat_add(env->sd->lb_gained[env->idle], detached);
8156 * attach_task() -- attach the task detached by detach_task() to its new rq.
8158 static void attach_task(struct rq *rq, struct task_struct *p)
8160 lockdep_assert_rq_held(rq);
8162 BUG_ON(task_rq(p) != rq);
8163 activate_task(rq, p, ENQUEUE_NOCLOCK);
8164 check_preempt_curr(rq, p, 0);
8168 * attach_one_task() -- attaches the task returned from detach_one_task() to
8171 static void attach_one_task(struct rq *rq, struct task_struct *p)
8176 update_rq_clock(rq);
8182 * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
8185 static void attach_tasks(struct lb_env *env)
8187 struct list_head *tasks = &env->tasks;
8188 struct task_struct *p;
8191 rq_lock(env->dst_rq, &rf);
8192 update_rq_clock(env->dst_rq);
8194 while (!list_empty(tasks)) {
8195 p = list_first_entry(tasks, struct task_struct, se.group_node);
8196 list_del_init(&p->se.group_node);
8198 attach_task(env->dst_rq, p);
8201 rq_unlock(env->dst_rq, &rf);
8204 #ifdef CONFIG_NO_HZ_COMMON
8205 static inline bool cfs_rq_has_blocked(struct cfs_rq *cfs_rq)
8207 if (cfs_rq->avg.load_avg)
8210 if (cfs_rq->avg.util_avg)
8216 static inline bool others_have_blocked(struct rq *rq)
8218 if (READ_ONCE(rq->avg_rt.util_avg))
8221 if (READ_ONCE(rq->avg_dl.util_avg))
8224 if (thermal_load_avg(rq))
8227 #ifdef CONFIG_HAVE_SCHED_AVG_IRQ
8228 if (READ_ONCE(rq->avg_irq.util_avg))
8235 static inline void update_blocked_load_tick(struct rq *rq)
8237 WRITE_ONCE(rq->last_blocked_load_update_tick, jiffies);
8240 static inline void update_blocked_load_status(struct rq *rq, bool has_blocked)
8243 rq->has_blocked_load = 0;
8246 static inline bool cfs_rq_has_blocked(struct cfs_rq *cfs_rq) { return false; }
8247 static inline bool others_have_blocked(struct rq *rq) { return false; }
8248 static inline void update_blocked_load_tick(struct rq *rq) {}
8249 static inline void update_blocked_load_status(struct rq *rq, bool has_blocked) {}
8252 static bool __update_blocked_others(struct rq *rq, bool *done)
8254 const struct sched_class *curr_class;
8255 u64 now = rq_clock_pelt(rq);
8256 unsigned long thermal_pressure;
8260 * update_load_avg() can call cpufreq_update_util(). Make sure that RT,
8261 * DL and IRQ signals have been updated before updating CFS.
8263 curr_class = rq->curr->sched_class;
8265 thermal_pressure = arch_scale_thermal_pressure(cpu_of(rq));
8267 decayed = update_rt_rq_load_avg(now, rq, curr_class == &rt_sched_class) |
8268 update_dl_rq_load_avg(now, rq, curr_class == &dl_sched_class) |
8269 update_thermal_load_avg(rq_clock_thermal(rq), rq, thermal_pressure) |
8270 update_irq_load_avg(rq, 0);
8272 if (others_have_blocked(rq))
8278 #ifdef CONFIG_FAIR_GROUP_SCHED
8280 static bool __update_blocked_fair(struct rq *rq, bool *done)
8282 struct cfs_rq *cfs_rq, *pos;
8283 bool decayed = false;
8284 int cpu = cpu_of(rq);
8287 * Iterates the task_group tree in a bottom up fashion, see
8288 * list_add_leaf_cfs_rq() for details.
8290 for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) {
8291 struct sched_entity *se;
8293 if (update_cfs_rq_load_avg(cfs_rq_clock_pelt(cfs_rq), cfs_rq)) {
8294 update_tg_load_avg(cfs_rq);
8296 if (cfs_rq->nr_running == 0)
8297 update_idle_cfs_rq_clock_pelt(cfs_rq);
8299 if (cfs_rq == &rq->cfs)
8303 /* Propagate pending load changes to the parent, if any: */
8304 se = cfs_rq->tg->se[cpu];
8305 if (se && !skip_blocked_update(se))
8306 update_load_avg(cfs_rq_of(se), se, UPDATE_TG);
8309 * There can be a lot of idle CPU cgroups. Don't let fully
8310 * decayed cfs_rqs linger on the list.
8312 if (cfs_rq_is_decayed(cfs_rq))
8313 list_del_leaf_cfs_rq(cfs_rq);
8315 /* Don't need periodic decay once load/util_avg are null */
8316 if (cfs_rq_has_blocked(cfs_rq))
8324 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
8325 * This needs to be done in a top-down fashion because the load of a child
8326 * group is a fraction of its parents load.
8328 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
8330 struct rq *rq = rq_of(cfs_rq);
8331 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
8332 unsigned long now = jiffies;
8335 if (cfs_rq->last_h_load_update == now)
8338 WRITE_ONCE(cfs_rq->h_load_next, NULL);
8339 for_each_sched_entity(se) {
8340 cfs_rq = cfs_rq_of(se);
8341 WRITE_ONCE(cfs_rq->h_load_next, se);
8342 if (cfs_rq->last_h_load_update == now)
8347 cfs_rq->h_load = cfs_rq_load_avg(cfs_rq);
8348 cfs_rq->last_h_load_update = now;
8351 while ((se = READ_ONCE(cfs_rq->h_load_next)) != NULL) {
8352 load = cfs_rq->h_load;
8353 load = div64_ul(load * se->avg.load_avg,
8354 cfs_rq_load_avg(cfs_rq) + 1);
8355 cfs_rq = group_cfs_rq(se);
8356 cfs_rq->h_load = load;
8357 cfs_rq->last_h_load_update = now;
8361 static unsigned long task_h_load(struct task_struct *p)
8363 struct cfs_rq *cfs_rq = task_cfs_rq(p);
8365 update_cfs_rq_h_load(cfs_rq);
8366 return div64_ul(p->se.avg.load_avg * cfs_rq->h_load,
8367 cfs_rq_load_avg(cfs_rq) + 1);
8370 static bool __update_blocked_fair(struct rq *rq, bool *done)
8372 struct cfs_rq *cfs_rq = &rq->cfs;
8375 decayed = update_cfs_rq_load_avg(cfs_rq_clock_pelt(cfs_rq), cfs_rq);
8376 if (cfs_rq_has_blocked(cfs_rq))
8382 static unsigned long task_h_load(struct task_struct *p)
8384 return p->se.avg.load_avg;
8388 static void update_blocked_averages(int cpu)
8390 bool decayed = false, done = true;
8391 struct rq *rq = cpu_rq(cpu);
8394 rq_lock_irqsave(rq, &rf);
8395 update_blocked_load_tick(rq);
8396 update_rq_clock(rq);
8398 decayed |= __update_blocked_others(rq, &done);
8399 decayed |= __update_blocked_fair(rq, &done);
8401 update_blocked_load_status(rq, !done);
8403 cpufreq_update_util(rq, 0);
8404 rq_unlock_irqrestore(rq, &rf);
8407 /********** Helpers for find_busiest_group ************************/
8410 * sg_lb_stats - stats of a sched_group required for load_balancing
8412 struct sg_lb_stats {
8413 unsigned long avg_load; /*Avg load across the CPUs of the group */
8414 unsigned long group_load; /* Total load over the CPUs of the group */
8415 unsigned long group_capacity;
8416 unsigned long group_util; /* Total utilization over the CPUs of the group */
8417 unsigned long group_runnable; /* Total runnable time over the CPUs of the group */
8418 unsigned int sum_nr_running; /* Nr of tasks running in the group */
8419 unsigned int sum_h_nr_running; /* Nr of CFS tasks running in the group */
8420 unsigned int idle_cpus;
8421 unsigned int group_weight;
8422 enum group_type group_type;
8423 unsigned int group_asym_packing; /* Tasks should be moved to preferred CPU */
8424 unsigned long group_misfit_task_load; /* A CPU has a task too big for its capacity */
8425 #ifdef CONFIG_NUMA_BALANCING
8426 unsigned int nr_numa_running;
8427 unsigned int nr_preferred_running;
8432 * sd_lb_stats - Structure to store the statistics of a sched_domain
8433 * during load balancing.
8435 struct sd_lb_stats {
8436 struct sched_group *busiest; /* Busiest group in this sd */
8437 struct sched_group *local; /* Local group in this sd */
8438 unsigned long total_load; /* Total load of all groups in sd */
8439 unsigned long total_capacity; /* Total capacity of all groups in sd */
8440 unsigned long avg_load; /* Average load across all groups in sd */
8441 unsigned int prefer_sibling; /* tasks should go to sibling first */
8443 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
8444 struct sg_lb_stats local_stat; /* Statistics of the local group */
8447 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
8450 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
8451 * local_stat because update_sg_lb_stats() does a full clear/assignment.
8452 * We must however set busiest_stat::group_type and
8453 * busiest_stat::idle_cpus to the worst busiest group because
8454 * update_sd_pick_busiest() reads these before assignment.
8456 *sds = (struct sd_lb_stats){
8460 .total_capacity = 0UL,
8462 .idle_cpus = UINT_MAX,
8463 .group_type = group_has_spare,
8468 static unsigned long scale_rt_capacity(int cpu)
8470 struct rq *rq = cpu_rq(cpu);
8471 unsigned long max = arch_scale_cpu_capacity(cpu);
8472 unsigned long used, free;
8475 irq = cpu_util_irq(rq);
8477 if (unlikely(irq >= max))
8481 * avg_rt.util_avg and avg_dl.util_avg track binary signals
8482 * (running and not running) with weights 0 and 1024 respectively.
8483 * avg_thermal.load_avg tracks thermal pressure and the weighted
8484 * average uses the actual delta max capacity(load).
8486 used = READ_ONCE(rq->avg_rt.util_avg);
8487 used += READ_ONCE(rq->avg_dl.util_avg);
8488 used += thermal_load_avg(rq);
8490 if (unlikely(used >= max))
8495 return scale_irq_capacity(free, irq, max);
8498 static void update_cpu_capacity(struct sched_domain *sd, int cpu)
8500 unsigned long capacity = scale_rt_capacity(cpu);
8501 struct sched_group *sdg = sd->groups;
8503 cpu_rq(cpu)->cpu_capacity_orig = arch_scale_cpu_capacity(cpu);
8508 cpu_rq(cpu)->cpu_capacity = capacity;
8509 trace_sched_cpu_capacity_tp(cpu_rq(cpu));
8511 sdg->sgc->capacity = capacity;
8512 sdg->sgc->min_capacity = capacity;
8513 sdg->sgc->max_capacity = capacity;
8516 void update_group_capacity(struct sched_domain *sd, int cpu)
8518 struct sched_domain *child = sd->child;
8519 struct sched_group *group, *sdg = sd->groups;
8520 unsigned long capacity, min_capacity, max_capacity;
8521 unsigned long interval;
8523 interval = msecs_to_jiffies(sd->balance_interval);
8524 interval = clamp(interval, 1UL, max_load_balance_interval);
8525 sdg->sgc->next_update = jiffies + interval;
8528 update_cpu_capacity(sd, cpu);
8533 min_capacity = ULONG_MAX;
8536 if (child->flags & SD_OVERLAP) {
8538 * SD_OVERLAP domains cannot assume that child groups
8539 * span the current group.
8542 for_each_cpu(cpu, sched_group_span(sdg)) {
8543 unsigned long cpu_cap = capacity_of(cpu);
8545 capacity += cpu_cap;
8546 min_capacity = min(cpu_cap, min_capacity);
8547 max_capacity = max(cpu_cap, max_capacity);
8551 * !SD_OVERLAP domains can assume that child groups
8552 * span the current group.
8555 group = child->groups;
8557 struct sched_group_capacity *sgc = group->sgc;
8559 capacity += sgc->capacity;
8560 min_capacity = min(sgc->min_capacity, min_capacity);
8561 max_capacity = max(sgc->max_capacity, max_capacity);
8562 group = group->next;
8563 } while (group != child->groups);
8566 sdg->sgc->capacity = capacity;
8567 sdg->sgc->min_capacity = min_capacity;
8568 sdg->sgc->max_capacity = max_capacity;
8572 * Check whether the capacity of the rq has been noticeably reduced by side
8573 * activity. The imbalance_pct is used for the threshold.
8574 * Return true is the capacity is reduced
8577 check_cpu_capacity(struct rq *rq, struct sched_domain *sd)
8579 return ((rq->cpu_capacity * sd->imbalance_pct) <
8580 (rq->cpu_capacity_orig * 100));
8584 * Check whether a rq has a misfit task and if it looks like we can actually
8585 * help that task: we can migrate the task to a CPU of higher capacity, or
8586 * the task's current CPU is heavily pressured.
8588 static inline int check_misfit_status(struct rq *rq, struct sched_domain *sd)
8590 return rq->misfit_task_load &&
8591 (rq->cpu_capacity_orig < rq->rd->max_cpu_capacity ||
8592 check_cpu_capacity(rq, sd));
8596 * Group imbalance indicates (and tries to solve) the problem where balancing
8597 * groups is inadequate due to ->cpus_ptr constraints.
8599 * Imagine a situation of two groups of 4 CPUs each and 4 tasks each with a
8600 * cpumask covering 1 CPU of the first group and 3 CPUs of the second group.
8603 * { 0 1 2 3 } { 4 5 6 7 }
8606 * If we were to balance group-wise we'd place two tasks in the first group and
8607 * two tasks in the second group. Clearly this is undesired as it will overload
8608 * cpu 3 and leave one of the CPUs in the second group unused.
8610 * The current solution to this issue is detecting the skew in the first group
8611 * by noticing the lower domain failed to reach balance and had difficulty
8612 * moving tasks due to affinity constraints.
8614 * When this is so detected; this group becomes a candidate for busiest; see
8615 * update_sd_pick_busiest(). And calculate_imbalance() and
8616 * find_busiest_group() avoid some of the usual balance conditions to allow it
8617 * to create an effective group imbalance.
8619 * This is a somewhat tricky proposition since the next run might not find the
8620 * group imbalance and decide the groups need to be balanced again. A most
8621 * subtle and fragile situation.
8624 static inline int sg_imbalanced(struct sched_group *group)
8626 return group->sgc->imbalance;
8630 * group_has_capacity returns true if the group has spare capacity that could
8631 * be used by some tasks.
8632 * We consider that a group has spare capacity if the number of task is
8633 * smaller than the number of CPUs or if the utilization is lower than the
8634 * available capacity for CFS tasks.
8635 * For the latter, we use a threshold to stabilize the state, to take into
8636 * account the variance of the tasks' load and to return true if the available
8637 * capacity in meaningful for the load balancer.
8638 * As an example, an available capacity of 1% can appear but it doesn't make
8639 * any benefit for the load balance.
8642 group_has_capacity(unsigned int imbalance_pct, struct sg_lb_stats *sgs)
8644 if (sgs->sum_nr_running < sgs->group_weight)
8647 if ((sgs->group_capacity * imbalance_pct) <
8648 (sgs->group_runnable * 100))
8651 if ((sgs->group_capacity * 100) >
8652 (sgs->group_util * imbalance_pct))
8659 * group_is_overloaded returns true if the group has more tasks than it can
8661 * group_is_overloaded is not equals to !group_has_capacity because a group
8662 * with the exact right number of tasks, has no more spare capacity but is not
8663 * overloaded so both group_has_capacity and group_is_overloaded return
8667 group_is_overloaded(unsigned int imbalance_pct, struct sg_lb_stats *sgs)
8669 if (sgs->sum_nr_running <= sgs->group_weight)
8672 if ((sgs->group_capacity * 100) <
8673 (sgs->group_util * imbalance_pct))
8676 if ((sgs->group_capacity * imbalance_pct) <
8677 (sgs->group_runnable * 100))
8684 group_type group_classify(unsigned int imbalance_pct,
8685 struct sched_group *group,
8686 struct sg_lb_stats *sgs)
8688 if (group_is_overloaded(imbalance_pct, sgs))
8689 return group_overloaded;
8691 if (sg_imbalanced(group))
8692 return group_imbalanced;
8694 if (sgs->group_asym_packing)
8695 return group_asym_packing;
8697 if (sgs->group_misfit_task_load)
8698 return group_misfit_task;
8700 if (!group_has_capacity(imbalance_pct, sgs))
8701 return group_fully_busy;
8703 return group_has_spare;
8707 * asym_smt_can_pull_tasks - Check whether the load balancing CPU can pull tasks
8708 * @dst_cpu: Destination CPU of the load balancing
8709 * @sds: Load-balancing data with statistics of the local group
8710 * @sgs: Load-balancing statistics of the candidate busiest group
8711 * @sg: The candidate busiest group
8713 * Check the state of the SMT siblings of both @sds::local and @sg and decide
8714 * if @dst_cpu can pull tasks.
8716 * If @dst_cpu does not have SMT siblings, it can pull tasks if two or more of
8717 * the SMT siblings of @sg are busy. If only one CPU in @sg is busy, pull tasks
8718 * only if @dst_cpu has higher priority.
8720 * If both @dst_cpu and @sg have SMT siblings, and @sg has exactly one more
8721 * busy CPU than @sds::local, let @dst_cpu pull tasks if it has higher priority.
8722 * Bigger imbalances in the number of busy CPUs will be dealt with in
8723 * update_sd_pick_busiest().
8725 * If @sg does not have SMT siblings, only pull tasks if all of the SMT siblings
8726 * of @dst_cpu are idle and @sg has lower priority.
8728 * Return: true if @dst_cpu can pull tasks, false otherwise.
8730 static bool asym_smt_can_pull_tasks(int dst_cpu, struct sd_lb_stats *sds,
8731 struct sg_lb_stats *sgs,
8732 struct sched_group *sg)
8734 #ifdef CONFIG_SCHED_SMT
8735 bool local_is_smt, sg_is_smt;
8738 local_is_smt = sds->local->flags & SD_SHARE_CPUCAPACITY;
8739 sg_is_smt = sg->flags & SD_SHARE_CPUCAPACITY;
8741 sg_busy_cpus = sgs->group_weight - sgs->idle_cpus;
8743 if (!local_is_smt) {
8745 * If we are here, @dst_cpu is idle and does not have SMT
8746 * siblings. Pull tasks if candidate group has two or more
8749 if (sg_busy_cpus >= 2) /* implies sg_is_smt */
8753 * @dst_cpu does not have SMT siblings. @sg may have SMT
8754 * siblings and only one is busy. In such case, @dst_cpu
8755 * can help if it has higher priority and is idle (i.e.,
8756 * it has no running tasks).
8758 return sched_asym_prefer(dst_cpu, sg->asym_prefer_cpu);
8761 /* @dst_cpu has SMT siblings. */
8764 int local_busy_cpus = sds->local->group_weight -
8765 sds->local_stat.idle_cpus;
8766 int busy_cpus_delta = sg_busy_cpus - local_busy_cpus;
8768 if (busy_cpus_delta == 1)
8769 return sched_asym_prefer(dst_cpu, sg->asym_prefer_cpu);
8775 * @sg does not have SMT siblings. Ensure that @sds::local does not end
8776 * up with more than one busy SMT sibling and only pull tasks if there
8777 * are not busy CPUs (i.e., no CPU has running tasks).
8779 if (!sds->local_stat.sum_nr_running)
8780 return sched_asym_prefer(dst_cpu, sg->asym_prefer_cpu);
8784 /* Always return false so that callers deal with non-SMT cases. */
8790 sched_asym(struct lb_env *env, struct sd_lb_stats *sds, struct sg_lb_stats *sgs,
8791 struct sched_group *group)
8793 /* Only do SMT checks if either local or candidate have SMT siblings */
8794 if ((sds->local->flags & SD_SHARE_CPUCAPACITY) ||
8795 (group->flags & SD_SHARE_CPUCAPACITY))
8796 return asym_smt_can_pull_tasks(env->dst_cpu, sds, sgs, group);
8798 return sched_asym_prefer(env->dst_cpu, group->asym_prefer_cpu);
8802 sched_reduced_capacity(struct rq *rq, struct sched_domain *sd)
8805 * When there is more than 1 task, the group_overloaded case already
8806 * takes care of cpu with reduced capacity
8808 if (rq->cfs.h_nr_running != 1)
8811 return check_cpu_capacity(rq, sd);
8815 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
8816 * @env: The load balancing environment.
8817 * @sds: Load-balancing data with statistics of the local group.
8818 * @group: sched_group whose statistics are to be updated.
8819 * @sgs: variable to hold the statistics for this group.
8820 * @sg_status: Holds flag indicating the status of the sched_group
8822 static inline void update_sg_lb_stats(struct lb_env *env,
8823 struct sd_lb_stats *sds,
8824 struct sched_group *group,
8825 struct sg_lb_stats *sgs,
8828 int i, nr_running, local_group;
8830 memset(sgs, 0, sizeof(*sgs));
8832 local_group = group == sds->local;
8834 for_each_cpu_and(i, sched_group_span(group), env->cpus) {
8835 struct rq *rq = cpu_rq(i);
8836 unsigned long load = cpu_load(rq);
8838 sgs->group_load += load;
8839 sgs->group_util += cpu_util_cfs(i);
8840 sgs->group_runnable += cpu_runnable(rq);
8841 sgs->sum_h_nr_running += rq->cfs.h_nr_running;
8843 nr_running = rq->nr_running;
8844 sgs->sum_nr_running += nr_running;
8847 *sg_status |= SG_OVERLOAD;
8849 if (cpu_overutilized(i))
8850 *sg_status |= SG_OVERUTILIZED;
8852 #ifdef CONFIG_NUMA_BALANCING
8853 sgs->nr_numa_running += rq->nr_numa_running;
8854 sgs->nr_preferred_running += rq->nr_preferred_running;
8857 * No need to call idle_cpu() if nr_running is not 0
8859 if (!nr_running && idle_cpu(i)) {
8861 /* Idle cpu can't have misfit task */
8868 if (env->sd->flags & SD_ASYM_CPUCAPACITY) {
8869 /* Check for a misfit task on the cpu */
8870 if (sgs->group_misfit_task_load < rq->misfit_task_load) {
8871 sgs->group_misfit_task_load = rq->misfit_task_load;
8872 *sg_status |= SG_OVERLOAD;
8874 } else if ((env->idle != CPU_NOT_IDLE) &&
8875 sched_reduced_capacity(rq, env->sd)) {
8876 /* Check for a task running on a CPU with reduced capacity */
8877 if (sgs->group_misfit_task_load < load)
8878 sgs->group_misfit_task_load = load;
8882 sgs->group_capacity = group->sgc->capacity;
8884 sgs->group_weight = group->group_weight;
8886 /* Check if dst CPU is idle and preferred to this group */
8887 if (!local_group && env->sd->flags & SD_ASYM_PACKING &&
8888 env->idle != CPU_NOT_IDLE && sgs->sum_h_nr_running &&
8889 sched_asym(env, sds, sgs, group)) {
8890 sgs->group_asym_packing = 1;
8893 sgs->group_type = group_classify(env->sd->imbalance_pct, group, sgs);
8895 /* Computing avg_load makes sense only when group is overloaded */
8896 if (sgs->group_type == group_overloaded)
8897 sgs->avg_load = (sgs->group_load * SCHED_CAPACITY_SCALE) /
8898 sgs->group_capacity;
8902 * update_sd_pick_busiest - return 1 on busiest group
8903 * @env: The load balancing environment.
8904 * @sds: sched_domain statistics
8905 * @sg: sched_group candidate to be checked for being the busiest
8906 * @sgs: sched_group statistics
8908 * Determine if @sg is a busier group than the previously selected
8911 * Return: %true if @sg is a busier group than the previously selected
8912 * busiest group. %false otherwise.
8914 static bool update_sd_pick_busiest(struct lb_env *env,
8915 struct sd_lb_stats *sds,
8916 struct sched_group *sg,
8917 struct sg_lb_stats *sgs)
8919 struct sg_lb_stats *busiest = &sds->busiest_stat;
8921 /* Make sure that there is at least one task to pull */
8922 if (!sgs->sum_h_nr_running)
8926 * Don't try to pull misfit tasks we can't help.
8927 * We can use max_capacity here as reduction in capacity on some
8928 * CPUs in the group should either be possible to resolve
8929 * internally or be covered by avg_load imbalance (eventually).
8931 if ((env->sd->flags & SD_ASYM_CPUCAPACITY) &&
8932 (sgs->group_type == group_misfit_task) &&
8933 (!capacity_greater(capacity_of(env->dst_cpu), sg->sgc->max_capacity) ||
8934 sds->local_stat.group_type != group_has_spare))
8937 if (sgs->group_type > busiest->group_type)
8940 if (sgs->group_type < busiest->group_type)
8944 * The candidate and the current busiest group are the same type of
8945 * group. Let check which one is the busiest according to the type.
8948 switch (sgs->group_type) {
8949 case group_overloaded:
8950 /* Select the overloaded group with highest avg_load. */
8951 if (sgs->avg_load <= busiest->avg_load)
8955 case group_imbalanced:
8957 * Select the 1st imbalanced group as we don't have any way to
8958 * choose one more than another.
8962 case group_asym_packing:
8963 /* Prefer to move from lowest priority CPU's work */
8964 if (sched_asym_prefer(sg->asym_prefer_cpu, sds->busiest->asym_prefer_cpu))
8968 case group_misfit_task:
8970 * If we have more than one misfit sg go with the biggest
8973 if (sgs->group_misfit_task_load < busiest->group_misfit_task_load)
8977 case group_fully_busy:
8979 * Select the fully busy group with highest avg_load. In
8980 * theory, there is no need to pull task from such kind of
8981 * group because tasks have all compute capacity that they need
8982 * but we can still improve the overall throughput by reducing
8983 * contention when accessing shared HW resources.
8985 * XXX for now avg_load is not computed and always 0 so we
8986 * select the 1st one.
8988 if (sgs->avg_load <= busiest->avg_load)
8992 case group_has_spare:
8994 * Select not overloaded group with lowest number of idle cpus
8995 * and highest number of running tasks. We could also compare
8996 * the spare capacity which is more stable but it can end up
8997 * that the group has less spare capacity but finally more idle
8998 * CPUs which means less opportunity to pull tasks.
9000 if (sgs->idle_cpus > busiest->idle_cpus)
9002 else if ((sgs->idle_cpus == busiest->idle_cpus) &&
9003 (sgs->sum_nr_running <= busiest->sum_nr_running))
9010 * Candidate sg has no more than one task per CPU and has higher
9011 * per-CPU capacity. Migrating tasks to less capable CPUs may harm
9012 * throughput. Maximize throughput, power/energy consequences are not
9015 if ((env->sd->flags & SD_ASYM_CPUCAPACITY) &&
9016 (sgs->group_type <= group_fully_busy) &&
9017 (capacity_greater(sg->sgc->min_capacity, capacity_of(env->dst_cpu))))
9023 #ifdef CONFIG_NUMA_BALANCING
9024 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
9026 if (sgs->sum_h_nr_running > sgs->nr_numa_running)
9028 if (sgs->sum_h_nr_running > sgs->nr_preferred_running)
9033 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
9035 if (rq->nr_running > rq->nr_numa_running)
9037 if (rq->nr_running > rq->nr_preferred_running)
9042 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
9047 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
9051 #endif /* CONFIG_NUMA_BALANCING */
9057 * task_running_on_cpu - return 1 if @p is running on @cpu.
9060 static unsigned int task_running_on_cpu(int cpu, struct task_struct *p)
9062 /* Task has no contribution or is new */
9063 if (cpu != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time))
9066 if (task_on_rq_queued(p))
9073 * idle_cpu_without - would a given CPU be idle without p ?
9074 * @cpu: the processor on which idleness is tested.
9075 * @p: task which should be ignored.
9077 * Return: 1 if the CPU would be idle. 0 otherwise.
9079 static int idle_cpu_without(int cpu, struct task_struct *p)
9081 struct rq *rq = cpu_rq(cpu);
9083 if (rq->curr != rq->idle && rq->curr != p)
9087 * rq->nr_running can't be used but an updated version without the
9088 * impact of p on cpu must be used instead. The updated nr_running
9089 * be computed and tested before calling idle_cpu_without().
9093 if (rq->ttwu_pending)
9101 * update_sg_wakeup_stats - Update sched_group's statistics for wakeup.
9102 * @sd: The sched_domain level to look for idlest group.
9103 * @group: sched_group whose statistics are to be updated.
9104 * @sgs: variable to hold the statistics for this group.
9105 * @p: The task for which we look for the idlest group/CPU.
9107 static inline void update_sg_wakeup_stats(struct sched_domain *sd,
9108 struct sched_group *group,
9109 struct sg_lb_stats *sgs,
9110 struct task_struct *p)
9114 memset(sgs, 0, sizeof(*sgs));
9116 for_each_cpu(i, sched_group_span(group)) {
9117 struct rq *rq = cpu_rq(i);
9120 sgs->group_load += cpu_load_without(rq, p);
9121 sgs->group_util += cpu_util_without(i, p);
9122 sgs->group_runnable += cpu_runnable_without(rq, p);
9123 local = task_running_on_cpu(i, p);
9124 sgs->sum_h_nr_running += rq->cfs.h_nr_running - local;
9126 nr_running = rq->nr_running - local;
9127 sgs->sum_nr_running += nr_running;
9130 * No need to call idle_cpu_without() if nr_running is not 0
9132 if (!nr_running && idle_cpu_without(i, p))
9137 /* Check if task fits in the group */
9138 if (sd->flags & SD_ASYM_CPUCAPACITY &&
9139 !task_fits_capacity(p, group->sgc->max_capacity)) {
9140 sgs->group_misfit_task_load = 1;
9143 sgs->group_capacity = group->sgc->capacity;
9145 sgs->group_weight = group->group_weight;
9147 sgs->group_type = group_classify(sd->imbalance_pct, group, sgs);
9150 * Computing avg_load makes sense only when group is fully busy or
9153 if (sgs->group_type == group_fully_busy ||
9154 sgs->group_type == group_overloaded)
9155 sgs->avg_load = (sgs->group_load * SCHED_CAPACITY_SCALE) /
9156 sgs->group_capacity;
9159 static bool update_pick_idlest(struct sched_group *idlest,
9160 struct sg_lb_stats *idlest_sgs,
9161 struct sched_group *group,
9162 struct sg_lb_stats *sgs)
9164 if (sgs->group_type < idlest_sgs->group_type)
9167 if (sgs->group_type > idlest_sgs->group_type)
9171 * The candidate and the current idlest group are the same type of
9172 * group. Let check which one is the idlest according to the type.
9175 switch (sgs->group_type) {
9176 case group_overloaded:
9177 case group_fully_busy:
9178 /* Select the group with lowest avg_load. */
9179 if (idlest_sgs->avg_load <= sgs->avg_load)
9183 case group_imbalanced:
9184 case group_asym_packing:
9185 /* Those types are not used in the slow wakeup path */
9188 case group_misfit_task:
9189 /* Select group with the highest max capacity */
9190 if (idlest->sgc->max_capacity >= group->sgc->max_capacity)
9194 case group_has_spare:
9195 /* Select group with most idle CPUs */
9196 if (idlest_sgs->idle_cpus > sgs->idle_cpus)
9199 /* Select group with lowest group_util */
9200 if (idlest_sgs->idle_cpus == sgs->idle_cpus &&
9201 idlest_sgs->group_util <= sgs->group_util)
9211 * find_idlest_group() finds and returns the least busy CPU group within the
9214 * Assumes p is allowed on at least one CPU in sd.
9216 static struct sched_group *
9217 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
9219 struct sched_group *idlest = NULL, *local = NULL, *group = sd->groups;
9220 struct sg_lb_stats local_sgs, tmp_sgs;
9221 struct sg_lb_stats *sgs;
9222 unsigned long imbalance;
9223 struct sg_lb_stats idlest_sgs = {
9224 .avg_load = UINT_MAX,
9225 .group_type = group_overloaded,
9231 /* Skip over this group if it has no CPUs allowed */
9232 if (!cpumask_intersects(sched_group_span(group),
9236 /* Skip over this group if no cookie matched */
9237 if (!sched_group_cookie_match(cpu_rq(this_cpu), p, group))
9240 local_group = cpumask_test_cpu(this_cpu,
9241 sched_group_span(group));
9250 update_sg_wakeup_stats(sd, group, sgs, p);
9252 if (!local_group && update_pick_idlest(idlest, &idlest_sgs, group, sgs)) {
9257 } while (group = group->next, group != sd->groups);
9260 /* There is no idlest group to push tasks to */
9264 /* The local group has been skipped because of CPU affinity */
9269 * If the local group is idler than the selected idlest group
9270 * don't try and push the task.
9272 if (local_sgs.group_type < idlest_sgs.group_type)
9276 * If the local group is busier than the selected idlest group
9277 * try and push the task.
9279 if (local_sgs.group_type > idlest_sgs.group_type)
9282 switch (local_sgs.group_type) {
9283 case group_overloaded:
9284 case group_fully_busy:
9286 /* Calculate allowed imbalance based on load */
9287 imbalance = scale_load_down(NICE_0_LOAD) *
9288 (sd->imbalance_pct-100) / 100;
9291 * When comparing groups across NUMA domains, it's possible for
9292 * the local domain to be very lightly loaded relative to the
9293 * remote domains but "imbalance" skews the comparison making
9294 * remote CPUs look much more favourable. When considering
9295 * cross-domain, add imbalance to the load on the remote node
9296 * and consider staying local.
9299 if ((sd->flags & SD_NUMA) &&
9300 ((idlest_sgs.avg_load + imbalance) >= local_sgs.avg_load))
9304 * If the local group is less loaded than the selected
9305 * idlest group don't try and push any tasks.
9307 if (idlest_sgs.avg_load >= (local_sgs.avg_load + imbalance))
9310 if (100 * local_sgs.avg_load <= sd->imbalance_pct * idlest_sgs.avg_load)
9314 case group_imbalanced:
9315 case group_asym_packing:
9316 /* Those type are not used in the slow wakeup path */
9319 case group_misfit_task:
9320 /* Select group with the highest max capacity */
9321 if (local->sgc->max_capacity >= idlest->sgc->max_capacity)
9325 case group_has_spare:
9327 if (sd->flags & SD_NUMA) {
9328 int imb_numa_nr = sd->imb_numa_nr;
9329 #ifdef CONFIG_NUMA_BALANCING
9332 * If there is spare capacity at NUMA, try to select
9333 * the preferred node
9335 if (cpu_to_node(this_cpu) == p->numa_preferred_nid)
9338 idlest_cpu = cpumask_first(sched_group_span(idlest));
9339 if (cpu_to_node(idlest_cpu) == p->numa_preferred_nid)
9341 #endif /* CONFIG_NUMA_BALANCING */
9343 * Otherwise, keep the task close to the wakeup source
9344 * and improve locality if the number of running tasks
9345 * would remain below threshold where an imbalance is
9346 * allowed while accounting for the possibility the
9347 * task is pinned to a subset of CPUs. If there is a
9348 * real need of migration, periodic load balance will
9351 if (p->nr_cpus_allowed != NR_CPUS) {
9352 struct cpumask *cpus = this_cpu_cpumask_var_ptr(select_rq_mask);
9354 cpumask_and(cpus, sched_group_span(local), p->cpus_ptr);
9355 imb_numa_nr = min(cpumask_weight(cpus), sd->imb_numa_nr);
9358 imbalance = abs(local_sgs.idle_cpus - idlest_sgs.idle_cpus);
9359 if (!adjust_numa_imbalance(imbalance,
9360 local_sgs.sum_nr_running + 1,
9365 #endif /* CONFIG_NUMA */
9368 * Select group with highest number of idle CPUs. We could also
9369 * compare the utilization which is more stable but it can end
9370 * up that the group has less spare capacity but finally more
9371 * idle CPUs which means more opportunity to run task.
9373 if (local_sgs.idle_cpus >= idlest_sgs.idle_cpus)
9381 static void update_idle_cpu_scan(struct lb_env *env,
9382 unsigned long sum_util)
9384 struct sched_domain_shared *sd_share;
9385 int llc_weight, pct;
9388 * Update the number of CPUs to scan in LLC domain, which could
9389 * be used as a hint in select_idle_cpu(). The update of sd_share
9390 * could be expensive because it is within a shared cache line.
9391 * So the write of this hint only occurs during periodic load
9392 * balancing, rather than CPU_NEWLY_IDLE, because the latter
9393 * can fire way more frequently than the former.
9395 if (!sched_feat(SIS_UTIL) || env->idle == CPU_NEWLY_IDLE)
9398 llc_weight = per_cpu(sd_llc_size, env->dst_cpu);
9399 if (env->sd->span_weight != llc_weight)
9402 sd_share = rcu_dereference(per_cpu(sd_llc_shared, env->dst_cpu));
9407 * The number of CPUs to search drops as sum_util increases, when
9408 * sum_util hits 85% or above, the scan stops.
9409 * The reason to choose 85% as the threshold is because this is the
9410 * imbalance_pct(117) when a LLC sched group is overloaded.
9412 * let y = SCHED_CAPACITY_SCALE - p * x^2 [1]
9413 * and y'= y / SCHED_CAPACITY_SCALE
9415 * x is the ratio of sum_util compared to the CPU capacity:
9416 * x = sum_util / (llc_weight * SCHED_CAPACITY_SCALE)
9417 * y' is the ratio of CPUs to be scanned in the LLC domain,
9418 * and the number of CPUs to scan is calculated by:
9420 * nr_scan = llc_weight * y' [2]
9422 * When x hits the threshold of overloaded, AKA, when
9423 * x = 100 / pct, y drops to 0. According to [1],
9424 * p should be SCHED_CAPACITY_SCALE * pct^2 / 10000
9426 * Scale x by SCHED_CAPACITY_SCALE:
9427 * x' = sum_util / llc_weight; [3]
9429 * and finally [1] becomes:
9430 * y = SCHED_CAPACITY_SCALE -
9431 * x'^2 * pct^2 / (10000 * SCHED_CAPACITY_SCALE) [4]
9436 do_div(x, llc_weight);
9439 pct = env->sd->imbalance_pct;
9440 tmp = x * x * pct * pct;
9441 do_div(tmp, 10000 * SCHED_CAPACITY_SCALE);
9442 tmp = min_t(long, tmp, SCHED_CAPACITY_SCALE);
9443 y = SCHED_CAPACITY_SCALE - tmp;
9447 do_div(y, SCHED_CAPACITY_SCALE);
9448 if ((int)y != sd_share->nr_idle_scan)
9449 WRITE_ONCE(sd_share->nr_idle_scan, (int)y);
9453 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
9454 * @env: The load balancing environment.
9455 * @sds: variable to hold the statistics for this sched_domain.
9458 static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
9460 struct sched_domain *child = env->sd->child;
9461 struct sched_group *sg = env->sd->groups;
9462 struct sg_lb_stats *local = &sds->local_stat;
9463 struct sg_lb_stats tmp_sgs;
9464 unsigned long sum_util = 0;
9468 struct sg_lb_stats *sgs = &tmp_sgs;
9471 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_span(sg));
9476 if (env->idle != CPU_NEWLY_IDLE ||
9477 time_after_eq(jiffies, sg->sgc->next_update))
9478 update_group_capacity(env->sd, env->dst_cpu);
9481 update_sg_lb_stats(env, sds, sg, sgs, &sg_status);
9487 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
9489 sds->busiest_stat = *sgs;
9493 /* Now, start updating sd_lb_stats */
9494 sds->total_load += sgs->group_load;
9495 sds->total_capacity += sgs->group_capacity;
9497 sum_util += sgs->group_util;
9499 } while (sg != env->sd->groups);
9501 /* Tag domain that child domain prefers tasks go to siblings first */
9502 sds->prefer_sibling = child && child->flags & SD_PREFER_SIBLING;
9505 if (env->sd->flags & SD_NUMA)
9506 env->fbq_type = fbq_classify_group(&sds->busiest_stat);
9508 if (!env->sd->parent) {
9509 struct root_domain *rd = env->dst_rq->rd;
9511 /* update overload indicator if we are at root domain */
9512 WRITE_ONCE(rd->overload, sg_status & SG_OVERLOAD);
9514 /* Update over-utilization (tipping point, U >= 0) indicator */
9515 WRITE_ONCE(rd->overutilized, sg_status & SG_OVERUTILIZED);
9516 trace_sched_overutilized_tp(rd, sg_status & SG_OVERUTILIZED);
9517 } else if (sg_status & SG_OVERUTILIZED) {
9518 struct root_domain *rd = env->dst_rq->rd;
9520 WRITE_ONCE(rd->overutilized, SG_OVERUTILIZED);
9521 trace_sched_overutilized_tp(rd, SG_OVERUTILIZED);
9524 update_idle_cpu_scan(env, sum_util);
9528 * calculate_imbalance - Calculate the amount of imbalance present within the
9529 * groups of a given sched_domain during load balance.
9530 * @env: load balance environment
9531 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
9533 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
9535 struct sg_lb_stats *local, *busiest;
9537 local = &sds->local_stat;
9538 busiest = &sds->busiest_stat;
9540 if (busiest->group_type == group_misfit_task) {
9541 if (env->sd->flags & SD_ASYM_CPUCAPACITY) {
9542 /* Set imbalance to allow misfit tasks to be balanced. */
9543 env->migration_type = migrate_misfit;
9547 * Set load imbalance to allow moving task from cpu
9548 * with reduced capacity.
9550 env->migration_type = migrate_load;
9551 env->imbalance = busiest->group_misfit_task_load;
9556 if (busiest->group_type == group_asym_packing) {
9558 * In case of asym capacity, we will try to migrate all load to
9559 * the preferred CPU.
9561 env->migration_type = migrate_task;
9562 env->imbalance = busiest->sum_h_nr_running;
9566 if (busiest->group_type == group_imbalanced) {
9568 * In the group_imb case we cannot rely on group-wide averages
9569 * to ensure CPU-load equilibrium, try to move any task to fix
9570 * the imbalance. The next load balance will take care of
9571 * balancing back the system.
9573 env->migration_type = migrate_task;
9579 * Try to use spare capacity of local group without overloading it or
9582 if (local->group_type == group_has_spare) {
9583 if ((busiest->group_type > group_fully_busy) &&
9584 !(env->sd->flags & SD_SHARE_PKG_RESOURCES)) {
9586 * If busiest is overloaded, try to fill spare
9587 * capacity. This might end up creating spare capacity
9588 * in busiest or busiest still being overloaded but
9589 * there is no simple way to directly compute the
9590 * amount of load to migrate in order to balance the
9593 env->migration_type = migrate_util;
9594 env->imbalance = max(local->group_capacity, local->group_util) -
9598 * In some cases, the group's utilization is max or even
9599 * higher than capacity because of migrations but the
9600 * local CPU is (newly) idle. There is at least one
9601 * waiting task in this overloaded busiest group. Let's
9604 if (env->idle != CPU_NOT_IDLE && env->imbalance == 0) {
9605 env->migration_type = migrate_task;
9612 if (busiest->group_weight == 1 || sds->prefer_sibling) {
9613 unsigned int nr_diff = busiest->sum_nr_running;
9615 * When prefer sibling, evenly spread running tasks on
9618 env->migration_type = migrate_task;
9619 lsub_positive(&nr_diff, local->sum_nr_running);
9620 env->imbalance = nr_diff;
9624 * If there is no overload, we just want to even the number of
9627 env->migration_type = migrate_task;
9628 env->imbalance = max_t(long, 0,
9629 (local->idle_cpus - busiest->idle_cpus));
9633 /* Consider allowing a small imbalance between NUMA groups */
9634 if (env->sd->flags & SD_NUMA) {
9635 env->imbalance = adjust_numa_imbalance(env->imbalance,
9636 local->sum_nr_running + 1,
9637 env->sd->imb_numa_nr);
9641 /* Number of tasks to move to restore balance */
9642 env->imbalance >>= 1;
9648 * Local is fully busy but has to take more load to relieve the
9651 if (local->group_type < group_overloaded) {
9653 * Local will become overloaded so the avg_load metrics are
9657 local->avg_load = (local->group_load * SCHED_CAPACITY_SCALE) /
9658 local->group_capacity;
9661 * If the local group is more loaded than the selected
9662 * busiest group don't try to pull any tasks.
9664 if (local->avg_load >= busiest->avg_load) {
9669 sds->avg_load = (sds->total_load * SCHED_CAPACITY_SCALE) /
9670 sds->total_capacity;
9674 * Both group are or will become overloaded and we're trying to get all
9675 * the CPUs to the average_load, so we don't want to push ourselves
9676 * above the average load, nor do we wish to reduce the max loaded CPU
9677 * below the average load. At the same time, we also don't want to
9678 * reduce the group load below the group capacity. Thus we look for
9679 * the minimum possible imbalance.
9681 env->migration_type = migrate_load;
9682 env->imbalance = min(
9683 (busiest->avg_load - sds->avg_load) * busiest->group_capacity,
9684 (sds->avg_load - local->avg_load) * local->group_capacity
9685 ) / SCHED_CAPACITY_SCALE;
9688 /******* find_busiest_group() helpers end here *********************/
9691 * Decision matrix according to the local and busiest group type:
9693 * busiest \ local has_spare fully_busy misfit asym imbalanced overloaded
9694 * has_spare nr_idle balanced N/A N/A balanced balanced
9695 * fully_busy nr_idle nr_idle N/A N/A balanced balanced
9696 * misfit_task force N/A N/A N/A N/A N/A
9697 * asym_packing force force N/A N/A force force
9698 * imbalanced force force N/A N/A force force
9699 * overloaded force force N/A N/A force avg_load
9701 * N/A : Not Applicable because already filtered while updating
9703 * balanced : The system is balanced for these 2 groups.
9704 * force : Calculate the imbalance as load migration is probably needed.
9705 * avg_load : Only if imbalance is significant enough.
9706 * nr_idle : dst_cpu is not busy and the number of idle CPUs is quite
9707 * different in groups.
9711 * find_busiest_group - Returns the busiest group within the sched_domain
9712 * if there is an imbalance.
9713 * @env: The load balancing environment.
9715 * Also calculates the amount of runnable load which should be moved
9716 * to restore balance.
9718 * Return: - The busiest group if imbalance exists.
9720 static struct sched_group *find_busiest_group(struct lb_env *env)
9722 struct sg_lb_stats *local, *busiest;
9723 struct sd_lb_stats sds;
9725 init_sd_lb_stats(&sds);
9728 * Compute the various statistics relevant for load balancing at
9731 update_sd_lb_stats(env, &sds);
9733 if (sched_energy_enabled()) {
9734 struct root_domain *rd = env->dst_rq->rd;
9736 if (rcu_dereference(rd->pd) && !READ_ONCE(rd->overutilized))
9740 local = &sds.local_stat;
9741 busiest = &sds.busiest_stat;
9743 /* There is no busy sibling group to pull tasks from */
9747 /* Misfit tasks should be dealt with regardless of the avg load */
9748 if (busiest->group_type == group_misfit_task)
9751 /* ASYM feature bypasses nice load balance check */
9752 if (busiest->group_type == group_asym_packing)
9756 * If the busiest group is imbalanced the below checks don't
9757 * work because they assume all things are equal, which typically
9758 * isn't true due to cpus_ptr constraints and the like.
9760 if (busiest->group_type == group_imbalanced)
9764 * If the local group is busier than the selected busiest group
9765 * don't try and pull any tasks.
9767 if (local->group_type > busiest->group_type)
9771 * When groups are overloaded, use the avg_load to ensure fairness
9774 if (local->group_type == group_overloaded) {
9776 * If the local group is more loaded than the selected
9777 * busiest group don't try to pull any tasks.
9779 if (local->avg_load >= busiest->avg_load)
9782 /* XXX broken for overlapping NUMA groups */
9783 sds.avg_load = (sds.total_load * SCHED_CAPACITY_SCALE) /
9787 * Don't pull any tasks if this group is already above the
9788 * domain average load.
9790 if (local->avg_load >= sds.avg_load)
9794 * If the busiest group is more loaded, use imbalance_pct to be
9797 if (100 * busiest->avg_load <=
9798 env->sd->imbalance_pct * local->avg_load)
9802 /* Try to move all excess tasks to child's sibling domain */
9803 if (sds.prefer_sibling && local->group_type == group_has_spare &&
9804 busiest->sum_nr_running > local->sum_nr_running + 1)
9807 if (busiest->group_type != group_overloaded) {
9808 if (env->idle == CPU_NOT_IDLE)
9810 * If the busiest group is not overloaded (and as a
9811 * result the local one too) but this CPU is already
9812 * busy, let another idle CPU try to pull task.
9816 if (busiest->group_weight > 1 &&
9817 local->idle_cpus <= (busiest->idle_cpus + 1))
9819 * If the busiest group is not overloaded
9820 * and there is no imbalance between this and busiest
9821 * group wrt idle CPUs, it is balanced. The imbalance
9822 * becomes significant if the diff is greater than 1
9823 * otherwise we might end up to just move the imbalance
9824 * on another group. Of course this applies only if
9825 * there is more than 1 CPU per group.
9829 if (busiest->sum_h_nr_running == 1)
9831 * busiest doesn't have any tasks waiting to run
9837 /* Looks like there is an imbalance. Compute it */
9838 calculate_imbalance(env, &sds);
9839 return env->imbalance ? sds.busiest : NULL;
9847 * find_busiest_queue - find the busiest runqueue among the CPUs in the group.
9849 static struct rq *find_busiest_queue(struct lb_env *env,
9850 struct sched_group *group)
9852 struct rq *busiest = NULL, *rq;
9853 unsigned long busiest_util = 0, busiest_load = 0, busiest_capacity = 1;
9854 unsigned int busiest_nr = 0;
9857 for_each_cpu_and(i, sched_group_span(group), env->cpus) {
9858 unsigned long capacity, load, util;
9859 unsigned int nr_running;
9863 rt = fbq_classify_rq(rq);
9866 * We classify groups/runqueues into three groups:
9867 * - regular: there are !numa tasks
9868 * - remote: there are numa tasks that run on the 'wrong' node
9869 * - all: there is no distinction
9871 * In order to avoid migrating ideally placed numa tasks,
9872 * ignore those when there's better options.
9874 * If we ignore the actual busiest queue to migrate another
9875 * task, the next balance pass can still reduce the busiest
9876 * queue by moving tasks around inside the node.
9878 * If we cannot move enough load due to this classification
9879 * the next pass will adjust the group classification and
9880 * allow migration of more tasks.
9882 * Both cases only affect the total convergence complexity.
9884 if (rt > env->fbq_type)
9887 nr_running = rq->cfs.h_nr_running;
9891 capacity = capacity_of(i);
9894 * For ASYM_CPUCAPACITY domains, don't pick a CPU that could
9895 * eventually lead to active_balancing high->low capacity.
9896 * Higher per-CPU capacity is considered better than balancing
9899 if (env->sd->flags & SD_ASYM_CPUCAPACITY &&
9900 !capacity_greater(capacity_of(env->dst_cpu), capacity) &&
9904 /* Make sure we only pull tasks from a CPU of lower priority */
9905 if ((env->sd->flags & SD_ASYM_PACKING) &&
9906 sched_asym_prefer(i, env->dst_cpu) &&
9910 switch (env->migration_type) {
9913 * When comparing with load imbalance, use cpu_load()
9914 * which is not scaled with the CPU capacity.
9916 load = cpu_load(rq);
9918 if (nr_running == 1 && load > env->imbalance &&
9919 !check_cpu_capacity(rq, env->sd))
9923 * For the load comparisons with the other CPUs,
9924 * consider the cpu_load() scaled with the CPU
9925 * capacity, so that the load can be moved away
9926 * from the CPU that is potentially running at a
9929 * Thus we're looking for max(load_i / capacity_i),
9930 * crosswise multiplication to rid ourselves of the
9931 * division works out to:
9932 * load_i * capacity_j > load_j * capacity_i;
9933 * where j is our previous maximum.
9935 if (load * busiest_capacity > busiest_load * capacity) {
9936 busiest_load = load;
9937 busiest_capacity = capacity;
9943 util = cpu_util_cfs(i);
9946 * Don't try to pull utilization from a CPU with one
9947 * running task. Whatever its utilization, we will fail
9950 if (nr_running <= 1)
9953 if (busiest_util < util) {
9954 busiest_util = util;
9960 if (busiest_nr < nr_running) {
9961 busiest_nr = nr_running;
9966 case migrate_misfit:
9968 * For ASYM_CPUCAPACITY domains with misfit tasks we
9969 * simply seek the "biggest" misfit task.
9971 if (rq->misfit_task_load > busiest_load) {
9972 busiest_load = rq->misfit_task_load;
9985 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
9986 * so long as it is large enough.
9988 #define MAX_PINNED_INTERVAL 512
9991 asym_active_balance(struct lb_env *env)
9994 * ASYM_PACKING needs to force migrate tasks from busy but
9995 * lower priority CPUs in order to pack all tasks in the
9996 * highest priority CPUs.
9998 return env->idle != CPU_NOT_IDLE && (env->sd->flags & SD_ASYM_PACKING) &&
9999 sched_asym_prefer(env->dst_cpu, env->src_cpu);
10003 imbalanced_active_balance(struct lb_env *env)
10005 struct sched_domain *sd = env->sd;
10008 * The imbalanced case includes the case of pinned tasks preventing a fair
10009 * distribution of the load on the system but also the even distribution of the
10010 * threads on a system with spare capacity
10012 if ((env->migration_type == migrate_task) &&
10013 (sd->nr_balance_failed > sd->cache_nice_tries+2))
10019 static int need_active_balance(struct lb_env *env)
10021 struct sched_domain *sd = env->sd;
10023 if (asym_active_balance(env))
10026 if (imbalanced_active_balance(env))
10030 * The dst_cpu is idle and the src_cpu CPU has only 1 CFS task.
10031 * It's worth migrating the task if the src_cpu's capacity is reduced
10032 * because of other sched_class or IRQs if more capacity stays
10033 * available on dst_cpu.
10035 if ((env->idle != CPU_NOT_IDLE) &&
10036 (env->src_rq->cfs.h_nr_running == 1)) {
10037 if ((check_cpu_capacity(env->src_rq, sd)) &&
10038 (capacity_of(env->src_cpu)*sd->imbalance_pct < capacity_of(env->dst_cpu)*100))
10042 if (env->migration_type == migrate_misfit)
10048 static int active_load_balance_cpu_stop(void *data);
10050 static int should_we_balance(struct lb_env *env)
10052 struct sched_group *sg = env->sd->groups;
10056 * Ensure the balancing environment is consistent; can happen
10057 * when the softirq triggers 'during' hotplug.
10059 if (!cpumask_test_cpu(env->dst_cpu, env->cpus))
10063 * In the newly idle case, we will allow all the CPUs
10064 * to do the newly idle load balance.
10066 * However, we bail out if we already have tasks or a wakeup pending,
10067 * to optimize wakeup latency.
10069 if (env->idle == CPU_NEWLY_IDLE) {
10070 if (env->dst_rq->nr_running > 0 || env->dst_rq->ttwu_pending)
10075 /* Try to find first idle CPU */
10076 for_each_cpu_and(cpu, group_balance_mask(sg), env->cpus) {
10077 if (!idle_cpu(cpu))
10080 /* Are we the first idle CPU? */
10081 return cpu == env->dst_cpu;
10084 /* Are we the first CPU of this group ? */
10085 return group_balance_cpu(sg) == env->dst_cpu;
10089 * Check this_cpu to ensure it is balanced within domain. Attempt to move
10090 * tasks if there is an imbalance.
10092 static int load_balance(int this_cpu, struct rq *this_rq,
10093 struct sched_domain *sd, enum cpu_idle_type idle,
10094 int *continue_balancing)
10096 int ld_moved, cur_ld_moved, active_balance = 0;
10097 struct sched_domain *sd_parent = sd->parent;
10098 struct sched_group *group;
10099 struct rq *busiest;
10100 struct rq_flags rf;
10101 struct cpumask *cpus = this_cpu_cpumask_var_ptr(load_balance_mask);
10103 struct lb_env env = {
10105 .dst_cpu = this_cpu,
10107 .dst_grpmask = sched_group_span(sd->groups),
10109 .loop_break = sched_nr_migrate_break,
10112 .tasks = LIST_HEAD_INIT(env.tasks),
10115 cpumask_and(cpus, sched_domain_span(sd), cpu_active_mask);
10117 schedstat_inc(sd->lb_count[idle]);
10120 if (!should_we_balance(&env)) {
10121 *continue_balancing = 0;
10125 group = find_busiest_group(&env);
10127 schedstat_inc(sd->lb_nobusyg[idle]);
10131 busiest = find_busiest_queue(&env, group);
10133 schedstat_inc(sd->lb_nobusyq[idle]);
10137 BUG_ON(busiest == env.dst_rq);
10139 schedstat_add(sd->lb_imbalance[idle], env.imbalance);
10141 env.src_cpu = busiest->cpu;
10142 env.src_rq = busiest;
10145 /* Clear this flag as soon as we find a pullable task */
10146 env.flags |= LBF_ALL_PINNED;
10147 if (busiest->nr_running > 1) {
10149 * Attempt to move tasks. If find_busiest_group has found
10150 * an imbalance but busiest->nr_running <= 1, the group is
10151 * still unbalanced. ld_moved simply stays zero, so it is
10152 * correctly treated as an imbalance.
10154 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
10157 rq_lock_irqsave(busiest, &rf);
10158 update_rq_clock(busiest);
10161 * cur_ld_moved - load moved in current iteration
10162 * ld_moved - cumulative load moved across iterations
10164 cur_ld_moved = detach_tasks(&env);
10167 * We've detached some tasks from busiest_rq. Every
10168 * task is masked "TASK_ON_RQ_MIGRATING", so we can safely
10169 * unlock busiest->lock, and we are able to be sure
10170 * that nobody can manipulate the tasks in parallel.
10171 * See task_rq_lock() family for the details.
10174 rq_unlock(busiest, &rf);
10176 if (cur_ld_moved) {
10177 attach_tasks(&env);
10178 ld_moved += cur_ld_moved;
10181 local_irq_restore(rf.flags);
10183 if (env.flags & LBF_NEED_BREAK) {
10184 env.flags &= ~LBF_NEED_BREAK;
10189 * Revisit (affine) tasks on src_cpu that couldn't be moved to
10190 * us and move them to an alternate dst_cpu in our sched_group
10191 * where they can run. The upper limit on how many times we
10192 * iterate on same src_cpu is dependent on number of CPUs in our
10195 * This changes load balance semantics a bit on who can move
10196 * load to a given_cpu. In addition to the given_cpu itself
10197 * (or a ilb_cpu acting on its behalf where given_cpu is
10198 * nohz-idle), we now have balance_cpu in a position to move
10199 * load to given_cpu. In rare situations, this may cause
10200 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
10201 * _independently_ and at _same_ time to move some load to
10202 * given_cpu) causing excess load to be moved to given_cpu.
10203 * This however should not happen so much in practice and
10204 * moreover subsequent load balance cycles should correct the
10205 * excess load moved.
10207 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
10209 /* Prevent to re-select dst_cpu via env's CPUs */
10210 __cpumask_clear_cpu(env.dst_cpu, env.cpus);
10212 env.dst_rq = cpu_rq(env.new_dst_cpu);
10213 env.dst_cpu = env.new_dst_cpu;
10214 env.flags &= ~LBF_DST_PINNED;
10216 env.loop_break = sched_nr_migrate_break;
10219 * Go back to "more_balance" rather than "redo" since we
10220 * need to continue with same src_cpu.
10226 * We failed to reach balance because of affinity.
10229 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
10231 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0)
10232 *group_imbalance = 1;
10235 /* All tasks on this runqueue were pinned by CPU affinity */
10236 if (unlikely(env.flags & LBF_ALL_PINNED)) {
10237 __cpumask_clear_cpu(cpu_of(busiest), cpus);
10239 * Attempting to continue load balancing at the current
10240 * sched_domain level only makes sense if there are
10241 * active CPUs remaining as possible busiest CPUs to
10242 * pull load from which are not contained within the
10243 * destination group that is receiving any migrated
10246 if (!cpumask_subset(cpus, env.dst_grpmask)) {
10248 env.loop_break = sched_nr_migrate_break;
10251 goto out_all_pinned;
10256 schedstat_inc(sd->lb_failed[idle]);
10258 * Increment the failure counter only on periodic balance.
10259 * We do not want newidle balance, which can be very
10260 * frequent, pollute the failure counter causing
10261 * excessive cache_hot migrations and active balances.
10263 if (idle != CPU_NEWLY_IDLE)
10264 sd->nr_balance_failed++;
10266 if (need_active_balance(&env)) {
10267 unsigned long flags;
10269 raw_spin_rq_lock_irqsave(busiest, flags);
10272 * Don't kick the active_load_balance_cpu_stop,
10273 * if the curr task on busiest CPU can't be
10274 * moved to this_cpu:
10276 if (!cpumask_test_cpu(this_cpu, busiest->curr->cpus_ptr)) {
10277 raw_spin_rq_unlock_irqrestore(busiest, flags);
10278 goto out_one_pinned;
10281 /* Record that we found at least one task that could run on this_cpu */
10282 env.flags &= ~LBF_ALL_PINNED;
10285 * ->active_balance synchronizes accesses to
10286 * ->active_balance_work. Once set, it's cleared
10287 * only after active load balance is finished.
10289 if (!busiest->active_balance) {
10290 busiest->active_balance = 1;
10291 busiest->push_cpu = this_cpu;
10292 active_balance = 1;
10294 raw_spin_rq_unlock_irqrestore(busiest, flags);
10296 if (active_balance) {
10297 stop_one_cpu_nowait(cpu_of(busiest),
10298 active_load_balance_cpu_stop, busiest,
10299 &busiest->active_balance_work);
10303 sd->nr_balance_failed = 0;
10306 if (likely(!active_balance) || need_active_balance(&env)) {
10307 /* We were unbalanced, so reset the balancing interval */
10308 sd->balance_interval = sd->min_interval;
10315 * We reach balance although we may have faced some affinity
10316 * constraints. Clear the imbalance flag only if other tasks got
10317 * a chance to move and fix the imbalance.
10319 if (sd_parent && !(env.flags & LBF_ALL_PINNED)) {
10320 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
10322 if (*group_imbalance)
10323 *group_imbalance = 0;
10328 * We reach balance because all tasks are pinned at this level so
10329 * we can't migrate them. Let the imbalance flag set so parent level
10330 * can try to migrate them.
10332 schedstat_inc(sd->lb_balanced[idle]);
10334 sd->nr_balance_failed = 0;
10340 * newidle_balance() disregards balance intervals, so we could
10341 * repeatedly reach this code, which would lead to balance_interval
10342 * skyrocketing in a short amount of time. Skip the balance_interval
10343 * increase logic to avoid that.
10345 if (env.idle == CPU_NEWLY_IDLE)
10348 /* tune up the balancing interval */
10349 if ((env.flags & LBF_ALL_PINNED &&
10350 sd->balance_interval < MAX_PINNED_INTERVAL) ||
10351 sd->balance_interval < sd->max_interval)
10352 sd->balance_interval *= 2;
10357 static inline unsigned long
10358 get_sd_balance_interval(struct sched_domain *sd, int cpu_busy)
10360 unsigned long interval = sd->balance_interval;
10363 interval *= sd->busy_factor;
10365 /* scale ms to jiffies */
10366 interval = msecs_to_jiffies(interval);
10369 * Reduce likelihood of busy balancing at higher domains racing with
10370 * balancing at lower domains by preventing their balancing periods
10371 * from being multiples of each other.
10376 interval = clamp(interval, 1UL, max_load_balance_interval);
10382 update_next_balance(struct sched_domain *sd, unsigned long *next_balance)
10384 unsigned long interval, next;
10386 /* used by idle balance, so cpu_busy = 0 */
10387 interval = get_sd_balance_interval(sd, 0);
10388 next = sd->last_balance + interval;
10390 if (time_after(*next_balance, next))
10391 *next_balance = next;
10395 * active_load_balance_cpu_stop is run by the CPU stopper. It pushes
10396 * running tasks off the busiest CPU onto idle CPUs. It requires at
10397 * least 1 task to be running on each physical CPU where possible, and
10398 * avoids physical / logical imbalances.
10400 static int active_load_balance_cpu_stop(void *data)
10402 struct rq *busiest_rq = data;
10403 int busiest_cpu = cpu_of(busiest_rq);
10404 int target_cpu = busiest_rq->push_cpu;
10405 struct rq *target_rq = cpu_rq(target_cpu);
10406 struct sched_domain *sd;
10407 struct task_struct *p = NULL;
10408 struct rq_flags rf;
10410 rq_lock_irq(busiest_rq, &rf);
10412 * Between queueing the stop-work and running it is a hole in which
10413 * CPUs can become inactive. We should not move tasks from or to
10416 if (!cpu_active(busiest_cpu) || !cpu_active(target_cpu))
10419 /* Make sure the requested CPU hasn't gone down in the meantime: */
10420 if (unlikely(busiest_cpu != smp_processor_id() ||
10421 !busiest_rq->active_balance))
10424 /* Is there any task to move? */
10425 if (busiest_rq->nr_running <= 1)
10429 * This condition is "impossible", if it occurs
10430 * we need to fix it. Originally reported by
10431 * Bjorn Helgaas on a 128-CPU setup.
10433 BUG_ON(busiest_rq == target_rq);
10435 /* Search for an sd spanning us and the target CPU. */
10437 for_each_domain(target_cpu, sd) {
10438 if (cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
10443 struct lb_env env = {
10445 .dst_cpu = target_cpu,
10446 .dst_rq = target_rq,
10447 .src_cpu = busiest_rq->cpu,
10448 .src_rq = busiest_rq,
10450 .flags = LBF_ACTIVE_LB,
10453 schedstat_inc(sd->alb_count);
10454 update_rq_clock(busiest_rq);
10456 p = detach_one_task(&env);
10458 schedstat_inc(sd->alb_pushed);
10459 /* Active balancing done, reset the failure counter. */
10460 sd->nr_balance_failed = 0;
10462 schedstat_inc(sd->alb_failed);
10467 busiest_rq->active_balance = 0;
10468 rq_unlock(busiest_rq, &rf);
10471 attach_one_task(target_rq, p);
10473 local_irq_enable();
10478 static DEFINE_SPINLOCK(balancing);
10481 * Scale the max load_balance interval with the number of CPUs in the system.
10482 * This trades load-balance latency on larger machines for less cross talk.
10484 void update_max_interval(void)
10486 max_load_balance_interval = HZ*num_online_cpus()/10;
10489 static inline bool update_newidle_cost(struct sched_domain *sd, u64 cost)
10491 if (cost > sd->max_newidle_lb_cost) {
10493 * Track max cost of a domain to make sure to not delay the
10494 * next wakeup on the CPU.
10496 sd->max_newidle_lb_cost = cost;
10497 sd->last_decay_max_lb_cost = jiffies;
10498 } else if (time_after(jiffies, sd->last_decay_max_lb_cost + HZ)) {
10500 * Decay the newidle max times by ~1% per second to ensure that
10501 * it is not outdated and the current max cost is actually
10504 sd->max_newidle_lb_cost = (sd->max_newidle_lb_cost * 253) / 256;
10505 sd->last_decay_max_lb_cost = jiffies;
10514 * It checks each scheduling domain to see if it is due to be balanced,
10515 * and initiates a balancing operation if so.
10517 * Balancing parameters are set up in init_sched_domains.
10519 static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle)
10521 int continue_balancing = 1;
10523 int busy = idle != CPU_IDLE && !sched_idle_cpu(cpu);
10524 unsigned long interval;
10525 struct sched_domain *sd;
10526 /* Earliest time when we have to do rebalance again */
10527 unsigned long next_balance = jiffies + 60*HZ;
10528 int update_next_balance = 0;
10529 int need_serialize, need_decay = 0;
10533 for_each_domain(cpu, sd) {
10535 * Decay the newidle max times here because this is a regular
10536 * visit to all the domains.
10538 need_decay = update_newidle_cost(sd, 0);
10539 max_cost += sd->max_newidle_lb_cost;
10542 * Stop the load balance at this level. There is another
10543 * CPU in our sched group which is doing load balancing more
10546 if (!continue_balancing) {
10552 interval = get_sd_balance_interval(sd, busy);
10554 need_serialize = sd->flags & SD_SERIALIZE;
10555 if (need_serialize) {
10556 if (!spin_trylock(&balancing))
10560 if (time_after_eq(jiffies, sd->last_balance + interval)) {
10561 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
10563 * The LBF_DST_PINNED logic could have changed
10564 * env->dst_cpu, so we can't know our idle
10565 * state even if we migrated tasks. Update it.
10567 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
10568 busy = idle != CPU_IDLE && !sched_idle_cpu(cpu);
10570 sd->last_balance = jiffies;
10571 interval = get_sd_balance_interval(sd, busy);
10573 if (need_serialize)
10574 spin_unlock(&balancing);
10576 if (time_after(next_balance, sd->last_balance + interval)) {
10577 next_balance = sd->last_balance + interval;
10578 update_next_balance = 1;
10583 * Ensure the rq-wide value also decays but keep it at a
10584 * reasonable floor to avoid funnies with rq->avg_idle.
10586 rq->max_idle_balance_cost =
10587 max((u64)sysctl_sched_migration_cost, max_cost);
10592 * next_balance will be updated only when there is a need.
10593 * When the cpu is attached to null domain for ex, it will not be
10596 if (likely(update_next_balance))
10597 rq->next_balance = next_balance;
10601 static inline int on_null_domain(struct rq *rq)
10603 return unlikely(!rcu_dereference_sched(rq->sd));
10606 #ifdef CONFIG_NO_HZ_COMMON
10608 * idle load balancing details
10609 * - When one of the busy CPUs notice that there may be an idle rebalancing
10610 * needed, they will kick the idle load balancer, which then does idle
10611 * load balancing for all the idle CPUs.
10612 * - HK_TYPE_MISC CPUs are used for this task, because HK_TYPE_SCHED not set
10616 static inline int find_new_ilb(void)
10619 const struct cpumask *hk_mask;
10621 hk_mask = housekeeping_cpumask(HK_TYPE_MISC);
10623 for_each_cpu_and(ilb, nohz.idle_cpus_mask, hk_mask) {
10625 if (ilb == smp_processor_id())
10636 * Kick a CPU to do the nohz balancing, if it is time for it. We pick any
10637 * idle CPU in the HK_TYPE_MISC housekeeping set (if there is one).
10639 static void kick_ilb(unsigned int flags)
10644 * Increase nohz.next_balance only when if full ilb is triggered but
10645 * not if we only update stats.
10647 if (flags & NOHZ_BALANCE_KICK)
10648 nohz.next_balance = jiffies+1;
10650 ilb_cpu = find_new_ilb();
10652 if (ilb_cpu >= nr_cpu_ids)
10656 * Access to rq::nohz_csd is serialized by NOHZ_KICK_MASK; he who sets
10657 * the first flag owns it; cleared by nohz_csd_func().
10659 flags = atomic_fetch_or(flags, nohz_flags(ilb_cpu));
10660 if (flags & NOHZ_KICK_MASK)
10664 * This way we generate an IPI on the target CPU which
10665 * is idle. And the softirq performing nohz idle load balance
10666 * will be run before returning from the IPI.
10668 smp_call_function_single_async(ilb_cpu, &cpu_rq(ilb_cpu)->nohz_csd);
10672 * Current decision point for kicking the idle load balancer in the presence
10673 * of idle CPUs in the system.
10675 static void nohz_balancer_kick(struct rq *rq)
10677 unsigned long now = jiffies;
10678 struct sched_domain_shared *sds;
10679 struct sched_domain *sd;
10680 int nr_busy, i, cpu = rq->cpu;
10681 unsigned int flags = 0;
10683 if (unlikely(rq->idle_balance))
10687 * We may be recently in ticked or tickless idle mode. At the first
10688 * busy tick after returning from idle, we will update the busy stats.
10690 nohz_balance_exit_idle(rq);
10693 * None are in tickless mode and hence no need for NOHZ idle load
10696 if (likely(!atomic_read(&nohz.nr_cpus)))
10699 if (READ_ONCE(nohz.has_blocked) &&
10700 time_after(now, READ_ONCE(nohz.next_blocked)))
10701 flags = NOHZ_STATS_KICK;
10703 if (time_before(now, nohz.next_balance))
10706 if (rq->nr_running >= 2) {
10707 flags = NOHZ_STATS_KICK | NOHZ_BALANCE_KICK;
10713 sd = rcu_dereference(rq->sd);
10716 * If there's a CFS task and the current CPU has reduced
10717 * capacity; kick the ILB to see if there's a better CPU to run
10720 if (rq->cfs.h_nr_running >= 1 && check_cpu_capacity(rq, sd)) {
10721 flags = NOHZ_STATS_KICK | NOHZ_BALANCE_KICK;
10726 sd = rcu_dereference(per_cpu(sd_asym_packing, cpu));
10729 * When ASYM_PACKING; see if there's a more preferred CPU
10730 * currently idle; in which case, kick the ILB to move tasks
10733 for_each_cpu_and(i, sched_domain_span(sd), nohz.idle_cpus_mask) {
10734 if (sched_asym_prefer(i, cpu)) {
10735 flags = NOHZ_STATS_KICK | NOHZ_BALANCE_KICK;
10741 sd = rcu_dereference(per_cpu(sd_asym_cpucapacity, cpu));
10744 * When ASYM_CPUCAPACITY; see if there's a higher capacity CPU
10745 * to run the misfit task on.
10747 if (check_misfit_status(rq, sd)) {
10748 flags = NOHZ_STATS_KICK | NOHZ_BALANCE_KICK;
10753 * For asymmetric systems, we do not want to nicely balance
10754 * cache use, instead we want to embrace asymmetry and only
10755 * ensure tasks have enough CPU capacity.
10757 * Skip the LLC logic because it's not relevant in that case.
10762 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
10765 * If there is an imbalance between LLC domains (IOW we could
10766 * increase the overall cache use), we need some less-loaded LLC
10767 * domain to pull some load. Likewise, we may need to spread
10768 * load within the current LLC domain (e.g. packed SMT cores but
10769 * other CPUs are idle). We can't really know from here how busy
10770 * the others are - so just get a nohz balance going if it looks
10771 * like this LLC domain has tasks we could move.
10773 nr_busy = atomic_read(&sds->nr_busy_cpus);
10775 flags = NOHZ_STATS_KICK | NOHZ_BALANCE_KICK;
10782 if (READ_ONCE(nohz.needs_update))
10783 flags |= NOHZ_NEXT_KICK;
10789 static void set_cpu_sd_state_busy(int cpu)
10791 struct sched_domain *sd;
10794 sd = rcu_dereference(per_cpu(sd_llc, cpu));
10796 if (!sd || !sd->nohz_idle)
10800 atomic_inc(&sd->shared->nr_busy_cpus);
10805 void nohz_balance_exit_idle(struct rq *rq)
10807 SCHED_WARN_ON(rq != this_rq());
10809 if (likely(!rq->nohz_tick_stopped))
10812 rq->nohz_tick_stopped = 0;
10813 cpumask_clear_cpu(rq->cpu, nohz.idle_cpus_mask);
10814 atomic_dec(&nohz.nr_cpus);
10816 set_cpu_sd_state_busy(rq->cpu);
10819 static void set_cpu_sd_state_idle(int cpu)
10821 struct sched_domain *sd;
10824 sd = rcu_dereference(per_cpu(sd_llc, cpu));
10826 if (!sd || sd->nohz_idle)
10830 atomic_dec(&sd->shared->nr_busy_cpus);
10836 * This routine will record that the CPU is going idle with tick stopped.
10837 * This info will be used in performing idle load balancing in the future.
10839 void nohz_balance_enter_idle(int cpu)
10841 struct rq *rq = cpu_rq(cpu);
10843 SCHED_WARN_ON(cpu != smp_processor_id());
10845 /* If this CPU is going down, then nothing needs to be done: */
10846 if (!cpu_active(cpu))
10849 /* Spare idle load balancing on CPUs that don't want to be disturbed: */
10850 if (!housekeeping_cpu(cpu, HK_TYPE_SCHED))
10854 * Can be set safely without rq->lock held
10855 * If a clear happens, it will have evaluated last additions because
10856 * rq->lock is held during the check and the clear
10858 rq->has_blocked_load = 1;
10861 * The tick is still stopped but load could have been added in the
10862 * meantime. We set the nohz.has_blocked flag to trig a check of the
10863 * *_avg. The CPU is already part of nohz.idle_cpus_mask so the clear
10864 * of nohz.has_blocked can only happen after checking the new load
10866 if (rq->nohz_tick_stopped)
10869 /* If we're a completely isolated CPU, we don't play: */
10870 if (on_null_domain(rq))
10873 rq->nohz_tick_stopped = 1;
10875 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
10876 atomic_inc(&nohz.nr_cpus);
10879 * Ensures that if nohz_idle_balance() fails to observe our
10880 * @idle_cpus_mask store, it must observe the @has_blocked
10881 * and @needs_update stores.
10883 smp_mb__after_atomic();
10885 set_cpu_sd_state_idle(cpu);
10887 WRITE_ONCE(nohz.needs_update, 1);
10890 * Each time a cpu enter idle, we assume that it has blocked load and
10891 * enable the periodic update of the load of idle cpus
10893 WRITE_ONCE(nohz.has_blocked, 1);
10896 static bool update_nohz_stats(struct rq *rq)
10898 unsigned int cpu = rq->cpu;
10900 if (!rq->has_blocked_load)
10903 if (!cpumask_test_cpu(cpu, nohz.idle_cpus_mask))
10906 if (!time_after(jiffies, READ_ONCE(rq->last_blocked_load_update_tick)))
10909 update_blocked_averages(cpu);
10911 return rq->has_blocked_load;
10915 * Internal function that runs load balance for all idle cpus. The load balance
10916 * can be a simple update of blocked load or a complete load balance with
10917 * tasks movement depending of flags.
10919 static void _nohz_idle_balance(struct rq *this_rq, unsigned int flags,
10920 enum cpu_idle_type idle)
10922 /* Earliest time when we have to do rebalance again */
10923 unsigned long now = jiffies;
10924 unsigned long next_balance = now + 60*HZ;
10925 bool has_blocked_load = false;
10926 int update_next_balance = 0;
10927 int this_cpu = this_rq->cpu;
10931 SCHED_WARN_ON((flags & NOHZ_KICK_MASK) == NOHZ_BALANCE_KICK);
10934 * We assume there will be no idle load after this update and clear
10935 * the has_blocked flag. If a cpu enters idle in the mean time, it will
10936 * set the has_blocked flag and trigger another update of idle load.
10937 * Because a cpu that becomes idle, is added to idle_cpus_mask before
10938 * setting the flag, we are sure to not clear the state and not
10939 * check the load of an idle cpu.
10941 * Same applies to idle_cpus_mask vs needs_update.
10943 if (flags & NOHZ_STATS_KICK)
10944 WRITE_ONCE(nohz.has_blocked, 0);
10945 if (flags & NOHZ_NEXT_KICK)
10946 WRITE_ONCE(nohz.needs_update, 0);
10949 * Ensures that if we miss the CPU, we must see the has_blocked
10950 * store from nohz_balance_enter_idle().
10955 * Start with the next CPU after this_cpu so we will end with this_cpu and let a
10956 * chance for other idle cpu to pull load.
10958 for_each_cpu_wrap(balance_cpu, nohz.idle_cpus_mask, this_cpu+1) {
10959 if (!idle_cpu(balance_cpu))
10963 * If this CPU gets work to do, stop the load balancing
10964 * work being done for other CPUs. Next load
10965 * balancing owner will pick it up.
10967 if (need_resched()) {
10968 if (flags & NOHZ_STATS_KICK)
10969 has_blocked_load = true;
10970 if (flags & NOHZ_NEXT_KICK)
10971 WRITE_ONCE(nohz.needs_update, 1);
10975 rq = cpu_rq(balance_cpu);
10977 if (flags & NOHZ_STATS_KICK)
10978 has_blocked_load |= update_nohz_stats(rq);
10981 * If time for next balance is due,
10984 if (time_after_eq(jiffies, rq->next_balance)) {
10985 struct rq_flags rf;
10987 rq_lock_irqsave(rq, &rf);
10988 update_rq_clock(rq);
10989 rq_unlock_irqrestore(rq, &rf);
10991 if (flags & NOHZ_BALANCE_KICK)
10992 rebalance_domains(rq, CPU_IDLE);
10995 if (time_after(next_balance, rq->next_balance)) {
10996 next_balance = rq->next_balance;
10997 update_next_balance = 1;
11002 * next_balance will be updated only when there is a need.
11003 * When the CPU is attached to null domain for ex, it will not be
11006 if (likely(update_next_balance))
11007 nohz.next_balance = next_balance;
11009 if (flags & NOHZ_STATS_KICK)
11010 WRITE_ONCE(nohz.next_blocked,
11011 now + msecs_to_jiffies(LOAD_AVG_PERIOD));
11014 /* There is still blocked load, enable periodic update */
11015 if (has_blocked_load)
11016 WRITE_ONCE(nohz.has_blocked, 1);
11020 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
11021 * rebalancing for all the cpus for whom scheduler ticks are stopped.
11023 static bool nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
11025 unsigned int flags = this_rq->nohz_idle_balance;
11030 this_rq->nohz_idle_balance = 0;
11032 if (idle != CPU_IDLE)
11035 _nohz_idle_balance(this_rq, flags, idle);
11041 * Check if we need to run the ILB for updating blocked load before entering
11044 void nohz_run_idle_balance(int cpu)
11046 unsigned int flags;
11048 flags = atomic_fetch_andnot(NOHZ_NEWILB_KICK, nohz_flags(cpu));
11051 * Update the blocked load only if no SCHED_SOFTIRQ is about to happen
11052 * (ie NOHZ_STATS_KICK set) and will do the same.
11054 if ((flags == NOHZ_NEWILB_KICK) && !need_resched())
11055 _nohz_idle_balance(cpu_rq(cpu), NOHZ_STATS_KICK, CPU_IDLE);
11058 static void nohz_newidle_balance(struct rq *this_rq)
11060 int this_cpu = this_rq->cpu;
11063 * This CPU doesn't want to be disturbed by scheduler
11066 if (!housekeeping_cpu(this_cpu, HK_TYPE_SCHED))
11069 /* Will wake up very soon. No time for doing anything else*/
11070 if (this_rq->avg_idle < sysctl_sched_migration_cost)
11073 /* Don't need to update blocked load of idle CPUs*/
11074 if (!READ_ONCE(nohz.has_blocked) ||
11075 time_before(jiffies, READ_ONCE(nohz.next_blocked)))
11079 * Set the need to trigger ILB in order to update blocked load
11080 * before entering idle state.
11082 atomic_or(NOHZ_NEWILB_KICK, nohz_flags(this_cpu));
11085 #else /* !CONFIG_NO_HZ_COMMON */
11086 static inline void nohz_balancer_kick(struct rq *rq) { }
11088 static inline bool nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
11093 static inline void nohz_newidle_balance(struct rq *this_rq) { }
11094 #endif /* CONFIG_NO_HZ_COMMON */
11097 * newidle_balance is called by schedule() if this_cpu is about to become
11098 * idle. Attempts to pull tasks from other CPUs.
11101 * < 0 - we released the lock and there are !fair tasks present
11102 * 0 - failed, no new tasks
11103 * > 0 - success, new (fair) tasks present
11105 static int newidle_balance(struct rq *this_rq, struct rq_flags *rf)
11107 unsigned long next_balance = jiffies + HZ;
11108 int this_cpu = this_rq->cpu;
11109 u64 t0, t1, curr_cost = 0;
11110 struct sched_domain *sd;
11111 int pulled_task = 0;
11113 update_misfit_status(NULL, this_rq);
11116 * There is a task waiting to run. No need to search for one.
11117 * Return 0; the task will be enqueued when switching to idle.
11119 if (this_rq->ttwu_pending)
11123 * We must set idle_stamp _before_ calling idle_balance(), such that we
11124 * measure the duration of idle_balance() as idle time.
11126 this_rq->idle_stamp = rq_clock(this_rq);
11129 * Do not pull tasks towards !active CPUs...
11131 if (!cpu_active(this_cpu))
11135 * This is OK, because current is on_cpu, which avoids it being picked
11136 * for load-balance and preemption/IRQs are still disabled avoiding
11137 * further scheduler activity on it and we're being very careful to
11138 * re-start the picking loop.
11140 rq_unpin_lock(this_rq, rf);
11143 sd = rcu_dereference_check_sched_domain(this_rq->sd);
11145 if (!READ_ONCE(this_rq->rd->overload) ||
11146 (sd && this_rq->avg_idle < sd->max_newidle_lb_cost)) {
11149 update_next_balance(sd, &next_balance);
11156 raw_spin_rq_unlock(this_rq);
11158 t0 = sched_clock_cpu(this_cpu);
11159 update_blocked_averages(this_cpu);
11162 for_each_domain(this_cpu, sd) {
11163 int continue_balancing = 1;
11166 update_next_balance(sd, &next_balance);
11168 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost)
11171 if (sd->flags & SD_BALANCE_NEWIDLE) {
11173 pulled_task = load_balance(this_cpu, this_rq,
11174 sd, CPU_NEWLY_IDLE,
11175 &continue_balancing);
11177 t1 = sched_clock_cpu(this_cpu);
11178 domain_cost = t1 - t0;
11179 update_newidle_cost(sd, domain_cost);
11181 curr_cost += domain_cost;
11186 * Stop searching for tasks to pull if there are
11187 * now runnable tasks on this rq.
11189 if (pulled_task || this_rq->nr_running > 0 ||
11190 this_rq->ttwu_pending)
11195 raw_spin_rq_lock(this_rq);
11197 if (curr_cost > this_rq->max_idle_balance_cost)
11198 this_rq->max_idle_balance_cost = curr_cost;
11201 * While browsing the domains, we released the rq lock, a task could
11202 * have been enqueued in the meantime. Since we're not going idle,
11203 * pretend we pulled a task.
11205 if (this_rq->cfs.h_nr_running && !pulled_task)
11208 /* Is there a task of a high priority class? */
11209 if (this_rq->nr_running != this_rq->cfs.h_nr_running)
11213 /* Move the next balance forward */
11214 if (time_after(this_rq->next_balance, next_balance))
11215 this_rq->next_balance = next_balance;
11218 this_rq->idle_stamp = 0;
11220 nohz_newidle_balance(this_rq);
11222 rq_repin_lock(this_rq, rf);
11224 return pulled_task;
11228 * run_rebalance_domains is triggered when needed from the scheduler tick.
11229 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
11231 static __latent_entropy void run_rebalance_domains(struct softirq_action *h)
11233 struct rq *this_rq = this_rq();
11234 enum cpu_idle_type idle = this_rq->idle_balance ?
11235 CPU_IDLE : CPU_NOT_IDLE;
11238 * If this CPU has a pending nohz_balance_kick, then do the
11239 * balancing on behalf of the other idle CPUs whose ticks are
11240 * stopped. Do nohz_idle_balance *before* rebalance_domains to
11241 * give the idle CPUs a chance to load balance. Else we may
11242 * load balance only within the local sched_domain hierarchy
11243 * and abort nohz_idle_balance altogether if we pull some load.
11245 if (nohz_idle_balance(this_rq, idle))
11248 /* normal load balance */
11249 update_blocked_averages(this_rq->cpu);
11250 rebalance_domains(this_rq, idle);
11254 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
11256 void trigger_load_balance(struct rq *rq)
11259 * Don't need to rebalance while attached to NULL domain or
11260 * runqueue CPU is not active
11262 if (unlikely(on_null_domain(rq) || !cpu_active(cpu_of(rq))))
11265 if (time_after_eq(jiffies, rq->next_balance))
11266 raise_softirq(SCHED_SOFTIRQ);
11268 nohz_balancer_kick(rq);
11271 static void rq_online_fair(struct rq *rq)
11275 update_runtime_enabled(rq);
11278 static void rq_offline_fair(struct rq *rq)
11282 /* Ensure any throttled groups are reachable by pick_next_task */
11283 unthrottle_offline_cfs_rqs(rq);
11286 #endif /* CONFIG_SMP */
11288 #ifdef CONFIG_SCHED_CORE
11290 __entity_slice_used(struct sched_entity *se, int min_nr_tasks)
11292 u64 slice = sched_slice(cfs_rq_of(se), se);
11293 u64 rtime = se->sum_exec_runtime - se->prev_sum_exec_runtime;
11295 return (rtime * min_nr_tasks > slice);
11298 #define MIN_NR_TASKS_DURING_FORCEIDLE 2
11299 static inline void task_tick_core(struct rq *rq, struct task_struct *curr)
11301 if (!sched_core_enabled(rq))
11305 * If runqueue has only one task which used up its slice and
11306 * if the sibling is forced idle, then trigger schedule to
11307 * give forced idle task a chance.
11309 * sched_slice() considers only this active rq and it gets the
11310 * whole slice. But during force idle, we have siblings acting
11311 * like a single runqueue and hence we need to consider runnable
11312 * tasks on this CPU and the forced idle CPU. Ideally, we should
11313 * go through the forced idle rq, but that would be a perf hit.
11314 * We can assume that the forced idle CPU has at least
11315 * MIN_NR_TASKS_DURING_FORCEIDLE - 1 tasks and use that to check
11316 * if we need to give up the CPU.
11318 if (rq->core->core_forceidle_count && rq->cfs.nr_running == 1 &&
11319 __entity_slice_used(&curr->se, MIN_NR_TASKS_DURING_FORCEIDLE))
11324 * se_fi_update - Update the cfs_rq->min_vruntime_fi in a CFS hierarchy if needed.
11326 static void se_fi_update(struct sched_entity *se, unsigned int fi_seq, bool forceidle)
11328 for_each_sched_entity(se) {
11329 struct cfs_rq *cfs_rq = cfs_rq_of(se);
11332 if (cfs_rq->forceidle_seq == fi_seq)
11334 cfs_rq->forceidle_seq = fi_seq;
11337 cfs_rq->min_vruntime_fi = cfs_rq->min_vruntime;
11341 void task_vruntime_update(struct rq *rq, struct task_struct *p, bool in_fi)
11343 struct sched_entity *se = &p->se;
11345 if (p->sched_class != &fair_sched_class)
11348 se_fi_update(se, rq->core->core_forceidle_seq, in_fi);
11351 bool cfs_prio_less(struct task_struct *a, struct task_struct *b, bool in_fi)
11353 struct rq *rq = task_rq(a);
11354 struct sched_entity *sea = &a->se;
11355 struct sched_entity *seb = &b->se;
11356 struct cfs_rq *cfs_rqa;
11357 struct cfs_rq *cfs_rqb;
11360 SCHED_WARN_ON(task_rq(b)->core != rq->core);
11362 #ifdef CONFIG_FAIR_GROUP_SCHED
11364 * Find an se in the hierarchy for tasks a and b, such that the se's
11365 * are immediate siblings.
11367 while (sea->cfs_rq->tg != seb->cfs_rq->tg) {
11368 int sea_depth = sea->depth;
11369 int seb_depth = seb->depth;
11371 if (sea_depth >= seb_depth)
11372 sea = parent_entity(sea);
11373 if (sea_depth <= seb_depth)
11374 seb = parent_entity(seb);
11377 se_fi_update(sea, rq->core->core_forceidle_seq, in_fi);
11378 se_fi_update(seb, rq->core->core_forceidle_seq, in_fi);
11380 cfs_rqa = sea->cfs_rq;
11381 cfs_rqb = seb->cfs_rq;
11383 cfs_rqa = &task_rq(a)->cfs;
11384 cfs_rqb = &task_rq(b)->cfs;
11388 * Find delta after normalizing se's vruntime with its cfs_rq's
11389 * min_vruntime_fi, which would have been updated in prior calls
11390 * to se_fi_update().
11392 delta = (s64)(sea->vruntime - seb->vruntime) +
11393 (s64)(cfs_rqb->min_vruntime_fi - cfs_rqa->min_vruntime_fi);
11398 static inline void task_tick_core(struct rq *rq, struct task_struct *curr) {}
11402 * scheduler tick hitting a task of our scheduling class.
11404 * NOTE: This function can be called remotely by the tick offload that
11405 * goes along full dynticks. Therefore no local assumption can be made
11406 * and everything must be accessed through the @rq and @curr passed in
11409 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
11411 struct cfs_rq *cfs_rq;
11412 struct sched_entity *se = &curr->se;
11414 for_each_sched_entity(se) {
11415 cfs_rq = cfs_rq_of(se);
11416 entity_tick(cfs_rq, se, queued);
11419 if (static_branch_unlikely(&sched_numa_balancing))
11420 task_tick_numa(rq, curr);
11422 update_misfit_status(curr, rq);
11423 update_overutilized_status(task_rq(curr));
11425 task_tick_core(rq, curr);
11429 * called on fork with the child task as argument from the parent's context
11430 * - child not yet on the tasklist
11431 * - preemption disabled
11433 static void task_fork_fair(struct task_struct *p)
11435 struct cfs_rq *cfs_rq;
11436 struct sched_entity *se = &p->se, *curr;
11437 struct rq *rq = this_rq();
11438 struct rq_flags rf;
11441 update_rq_clock(rq);
11443 cfs_rq = task_cfs_rq(current);
11444 curr = cfs_rq->curr;
11446 update_curr(cfs_rq);
11447 se->vruntime = curr->vruntime;
11449 place_entity(cfs_rq, se, 1);
11451 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
11453 * Upon rescheduling, sched_class::put_prev_task() will place
11454 * 'current' within the tree based on its new key value.
11456 swap(curr->vruntime, se->vruntime);
11460 se->vruntime -= cfs_rq->min_vruntime;
11461 rq_unlock(rq, &rf);
11465 * Priority of the task has changed. Check to see if we preempt
11466 * the current task.
11469 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
11471 if (!task_on_rq_queued(p))
11474 if (rq->cfs.nr_running == 1)
11478 * Reschedule if we are currently running on this runqueue and
11479 * our priority decreased, or if we are not currently running on
11480 * this runqueue and our priority is higher than the current's
11482 if (task_current(rq, p)) {
11483 if (p->prio > oldprio)
11486 check_preempt_curr(rq, p, 0);
11489 static inline bool vruntime_normalized(struct task_struct *p)
11491 struct sched_entity *se = &p->se;
11494 * In both the TASK_ON_RQ_QUEUED and TASK_ON_RQ_MIGRATING cases,
11495 * the dequeue_entity(.flags=0) will already have normalized the
11502 * When !on_rq, vruntime of the task has usually NOT been normalized.
11503 * But there are some cases where it has already been normalized:
11505 * - A forked child which is waiting for being woken up by
11506 * wake_up_new_task().
11507 * - A task which has been woken up by try_to_wake_up() and
11508 * waiting for actually being woken up by sched_ttwu_pending().
11510 if (!se->sum_exec_runtime ||
11511 (READ_ONCE(p->__state) == TASK_WAKING && p->sched_remote_wakeup))
11517 #ifdef CONFIG_FAIR_GROUP_SCHED
11519 * Propagate the changes of the sched_entity across the tg tree to make it
11520 * visible to the root
11522 static void propagate_entity_cfs_rq(struct sched_entity *se)
11524 struct cfs_rq *cfs_rq = cfs_rq_of(se);
11526 if (cfs_rq_throttled(cfs_rq))
11529 if (!throttled_hierarchy(cfs_rq))
11530 list_add_leaf_cfs_rq(cfs_rq);
11532 /* Start to propagate at parent */
11535 for_each_sched_entity(se) {
11536 cfs_rq = cfs_rq_of(se);
11538 update_load_avg(cfs_rq, se, UPDATE_TG);
11540 if (cfs_rq_throttled(cfs_rq))
11543 if (!throttled_hierarchy(cfs_rq))
11544 list_add_leaf_cfs_rq(cfs_rq);
11548 static void propagate_entity_cfs_rq(struct sched_entity *se) { }
11551 static void detach_entity_cfs_rq(struct sched_entity *se)
11553 struct cfs_rq *cfs_rq = cfs_rq_of(se);
11555 /* Catch up with the cfs_rq and remove our load when we leave */
11556 update_load_avg(cfs_rq, se, 0);
11557 detach_entity_load_avg(cfs_rq, se);
11558 update_tg_load_avg(cfs_rq);
11559 propagate_entity_cfs_rq(se);
11562 static void attach_entity_cfs_rq(struct sched_entity *se)
11564 struct cfs_rq *cfs_rq = cfs_rq_of(se);
11566 #ifdef CONFIG_FAIR_GROUP_SCHED
11568 * Since the real-depth could have been changed (only FAIR
11569 * class maintain depth value), reset depth properly.
11571 se->depth = se->parent ? se->parent->depth + 1 : 0;
11574 /* Synchronize entity with its cfs_rq */
11575 update_load_avg(cfs_rq, se, sched_feat(ATTACH_AGE_LOAD) ? 0 : SKIP_AGE_LOAD);
11576 attach_entity_load_avg(cfs_rq, se);
11577 update_tg_load_avg(cfs_rq);
11578 propagate_entity_cfs_rq(se);
11581 static void detach_task_cfs_rq(struct task_struct *p)
11583 struct sched_entity *se = &p->se;
11584 struct cfs_rq *cfs_rq = cfs_rq_of(se);
11586 if (!vruntime_normalized(p)) {
11588 * Fix up our vruntime so that the current sleep doesn't
11589 * cause 'unlimited' sleep bonus.
11591 place_entity(cfs_rq, se, 0);
11592 se->vruntime -= cfs_rq->min_vruntime;
11595 detach_entity_cfs_rq(se);
11598 static void attach_task_cfs_rq(struct task_struct *p)
11600 struct sched_entity *se = &p->se;
11601 struct cfs_rq *cfs_rq = cfs_rq_of(se);
11603 attach_entity_cfs_rq(se);
11605 if (!vruntime_normalized(p))
11606 se->vruntime += cfs_rq->min_vruntime;
11609 static void switched_from_fair(struct rq *rq, struct task_struct *p)
11611 detach_task_cfs_rq(p);
11614 static void switched_to_fair(struct rq *rq, struct task_struct *p)
11616 attach_task_cfs_rq(p);
11618 if (task_on_rq_queued(p)) {
11620 * We were most likely switched from sched_rt, so
11621 * kick off the schedule if running, otherwise just see
11622 * if we can still preempt the current task.
11624 if (task_current(rq, p))
11627 check_preempt_curr(rq, p, 0);
11631 /* Account for a task changing its policy or group.
11633 * This routine is mostly called to set cfs_rq->curr field when a task
11634 * migrates between groups/classes.
11636 static void set_next_task_fair(struct rq *rq, struct task_struct *p, bool first)
11638 struct sched_entity *se = &p->se;
11641 if (task_on_rq_queued(p)) {
11643 * Move the next running task to the front of the list, so our
11644 * cfs_tasks list becomes MRU one.
11646 list_move(&se->group_node, &rq->cfs_tasks);
11650 for_each_sched_entity(se) {
11651 struct cfs_rq *cfs_rq = cfs_rq_of(se);
11653 set_next_entity(cfs_rq, se);
11654 /* ensure bandwidth has been allocated on our new cfs_rq */
11655 account_cfs_rq_runtime(cfs_rq, 0);
11659 void init_cfs_rq(struct cfs_rq *cfs_rq)
11661 cfs_rq->tasks_timeline = RB_ROOT_CACHED;
11662 u64_u32_store(cfs_rq->min_vruntime, (u64)(-(1LL << 20)));
11664 raw_spin_lock_init(&cfs_rq->removed.lock);
11668 #ifdef CONFIG_FAIR_GROUP_SCHED
11669 static void task_set_group_fair(struct task_struct *p)
11671 struct sched_entity *se = &p->se;
11673 set_task_rq(p, task_cpu(p));
11674 se->depth = se->parent ? se->parent->depth + 1 : 0;
11677 static void task_move_group_fair(struct task_struct *p)
11679 detach_task_cfs_rq(p);
11680 set_task_rq(p, task_cpu(p));
11683 /* Tell se's cfs_rq has been changed -- migrated */
11684 p->se.avg.last_update_time = 0;
11686 attach_task_cfs_rq(p);
11689 static void task_change_group_fair(struct task_struct *p, int type)
11692 case TASK_SET_GROUP:
11693 task_set_group_fair(p);
11696 case TASK_MOVE_GROUP:
11697 task_move_group_fair(p);
11702 void free_fair_sched_group(struct task_group *tg)
11706 for_each_possible_cpu(i) {
11708 kfree(tg->cfs_rq[i]);
11717 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
11719 struct sched_entity *se;
11720 struct cfs_rq *cfs_rq;
11723 tg->cfs_rq = kcalloc(nr_cpu_ids, sizeof(cfs_rq), GFP_KERNEL);
11726 tg->se = kcalloc(nr_cpu_ids, sizeof(se), GFP_KERNEL);
11730 tg->shares = NICE_0_LOAD;
11732 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
11734 for_each_possible_cpu(i) {
11735 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
11736 GFP_KERNEL, cpu_to_node(i));
11740 se = kzalloc_node(sizeof(struct sched_entity_stats),
11741 GFP_KERNEL, cpu_to_node(i));
11745 init_cfs_rq(cfs_rq);
11746 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
11747 init_entity_runnable_average(se);
11758 void online_fair_sched_group(struct task_group *tg)
11760 struct sched_entity *se;
11761 struct rq_flags rf;
11765 for_each_possible_cpu(i) {
11768 rq_lock_irq(rq, &rf);
11769 update_rq_clock(rq);
11770 attach_entity_cfs_rq(se);
11771 sync_throttle(tg, i);
11772 rq_unlock_irq(rq, &rf);
11776 void unregister_fair_sched_group(struct task_group *tg)
11778 unsigned long flags;
11782 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
11784 for_each_possible_cpu(cpu) {
11786 remove_entity_load_avg(tg->se[cpu]);
11789 * Only empty task groups can be destroyed; so we can speculatively
11790 * check on_list without danger of it being re-added.
11792 if (!tg->cfs_rq[cpu]->on_list)
11797 raw_spin_rq_lock_irqsave(rq, flags);
11798 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
11799 raw_spin_rq_unlock_irqrestore(rq, flags);
11803 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
11804 struct sched_entity *se, int cpu,
11805 struct sched_entity *parent)
11807 struct rq *rq = cpu_rq(cpu);
11811 init_cfs_rq_runtime(cfs_rq);
11813 tg->cfs_rq[cpu] = cfs_rq;
11816 /* se could be NULL for root_task_group */
11821 se->cfs_rq = &rq->cfs;
11824 se->cfs_rq = parent->my_q;
11825 se->depth = parent->depth + 1;
11829 /* guarantee group entities always have weight */
11830 update_load_set(&se->load, NICE_0_LOAD);
11831 se->parent = parent;
11834 static DEFINE_MUTEX(shares_mutex);
11836 static int __sched_group_set_shares(struct task_group *tg, unsigned long shares)
11840 lockdep_assert_held(&shares_mutex);
11843 * We can't change the weight of the root cgroup.
11848 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
11850 if (tg->shares == shares)
11853 tg->shares = shares;
11854 for_each_possible_cpu(i) {
11855 struct rq *rq = cpu_rq(i);
11856 struct sched_entity *se = tg->se[i];
11857 struct rq_flags rf;
11859 /* Propagate contribution to hierarchy */
11860 rq_lock_irqsave(rq, &rf);
11861 update_rq_clock(rq);
11862 for_each_sched_entity(se) {
11863 update_load_avg(cfs_rq_of(se), se, UPDATE_TG);
11864 update_cfs_group(se);
11866 rq_unlock_irqrestore(rq, &rf);
11872 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
11876 mutex_lock(&shares_mutex);
11877 if (tg_is_idle(tg))
11880 ret = __sched_group_set_shares(tg, shares);
11881 mutex_unlock(&shares_mutex);
11886 int sched_group_set_idle(struct task_group *tg, long idle)
11890 if (tg == &root_task_group)
11893 if (idle < 0 || idle > 1)
11896 mutex_lock(&shares_mutex);
11898 if (tg->idle == idle) {
11899 mutex_unlock(&shares_mutex);
11905 for_each_possible_cpu(i) {
11906 struct rq *rq = cpu_rq(i);
11907 struct sched_entity *se = tg->se[i];
11908 struct cfs_rq *parent_cfs_rq, *grp_cfs_rq = tg->cfs_rq[i];
11909 bool was_idle = cfs_rq_is_idle(grp_cfs_rq);
11910 long idle_task_delta;
11911 struct rq_flags rf;
11913 rq_lock_irqsave(rq, &rf);
11915 grp_cfs_rq->idle = idle;
11916 if (WARN_ON_ONCE(was_idle == cfs_rq_is_idle(grp_cfs_rq)))
11920 parent_cfs_rq = cfs_rq_of(se);
11921 if (cfs_rq_is_idle(grp_cfs_rq))
11922 parent_cfs_rq->idle_nr_running++;
11924 parent_cfs_rq->idle_nr_running--;
11927 idle_task_delta = grp_cfs_rq->h_nr_running -
11928 grp_cfs_rq->idle_h_nr_running;
11929 if (!cfs_rq_is_idle(grp_cfs_rq))
11930 idle_task_delta *= -1;
11932 for_each_sched_entity(se) {
11933 struct cfs_rq *cfs_rq = cfs_rq_of(se);
11938 cfs_rq->idle_h_nr_running += idle_task_delta;
11940 /* Already accounted at parent level and above. */
11941 if (cfs_rq_is_idle(cfs_rq))
11946 rq_unlock_irqrestore(rq, &rf);
11949 /* Idle groups have minimum weight. */
11950 if (tg_is_idle(tg))
11951 __sched_group_set_shares(tg, scale_load(WEIGHT_IDLEPRIO));
11953 __sched_group_set_shares(tg, NICE_0_LOAD);
11955 mutex_unlock(&shares_mutex);
11959 #else /* CONFIG_FAIR_GROUP_SCHED */
11961 void free_fair_sched_group(struct task_group *tg) { }
11963 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
11968 void online_fair_sched_group(struct task_group *tg) { }
11970 void unregister_fair_sched_group(struct task_group *tg) { }
11972 #endif /* CONFIG_FAIR_GROUP_SCHED */
11975 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
11977 struct sched_entity *se = &task->se;
11978 unsigned int rr_interval = 0;
11981 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
11984 if (rq->cfs.load.weight)
11985 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
11987 return rr_interval;
11991 * All the scheduling class methods:
11993 DEFINE_SCHED_CLASS(fair) = {
11995 .enqueue_task = enqueue_task_fair,
11996 .dequeue_task = dequeue_task_fair,
11997 .yield_task = yield_task_fair,
11998 .yield_to_task = yield_to_task_fair,
12000 .check_preempt_curr = check_preempt_wakeup,
12002 .pick_next_task = __pick_next_task_fair,
12003 .put_prev_task = put_prev_task_fair,
12004 .set_next_task = set_next_task_fair,
12007 .balance = balance_fair,
12008 .pick_task = pick_task_fair,
12009 .select_task_rq = select_task_rq_fair,
12010 .migrate_task_rq = migrate_task_rq_fair,
12012 .rq_online = rq_online_fair,
12013 .rq_offline = rq_offline_fair,
12015 .task_dead = task_dead_fair,
12016 .set_cpus_allowed = set_cpus_allowed_common,
12019 .task_tick = task_tick_fair,
12020 .task_fork = task_fork_fair,
12022 .prio_changed = prio_changed_fair,
12023 .switched_from = switched_from_fair,
12024 .switched_to = switched_to_fair,
12026 .get_rr_interval = get_rr_interval_fair,
12028 .update_curr = update_curr_fair,
12030 #ifdef CONFIG_FAIR_GROUP_SCHED
12031 .task_change_group = task_change_group_fair,
12034 #ifdef CONFIG_UCLAMP_TASK
12035 .uclamp_enabled = 1,
12039 #ifdef CONFIG_SCHED_DEBUG
12040 void print_cfs_stats(struct seq_file *m, int cpu)
12042 struct cfs_rq *cfs_rq, *pos;
12045 for_each_leaf_cfs_rq_safe(cpu_rq(cpu), cfs_rq, pos)
12046 print_cfs_rq(m, cpu, cfs_rq);
12050 #ifdef CONFIG_NUMA_BALANCING
12051 void show_numa_stats(struct task_struct *p, struct seq_file *m)
12054 unsigned long tsf = 0, tpf = 0, gsf = 0, gpf = 0;
12055 struct numa_group *ng;
12058 ng = rcu_dereference(p->numa_group);
12059 for_each_online_node(node) {
12060 if (p->numa_faults) {
12061 tsf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 0)];
12062 tpf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 1)];
12065 gsf = ng->faults[task_faults_idx(NUMA_MEM, node, 0)],
12066 gpf = ng->faults[task_faults_idx(NUMA_MEM, node, 1)];
12068 print_numa_stats(m, node, tsf, tpf, gsf, gpf);
12072 #endif /* CONFIG_NUMA_BALANCING */
12073 #endif /* CONFIG_SCHED_DEBUG */
12075 __init void init_sched_fair_class(void)
12078 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
12080 #ifdef CONFIG_NO_HZ_COMMON
12081 nohz.next_balance = jiffies;
12082 nohz.next_blocked = jiffies;
12083 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);