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 */
803 * With new tasks being created, their initial util_avgs are extrapolated
804 * based on the cfs_rq's current util_avg:
806 * util_avg = cfs_rq->util_avg / (cfs_rq->load_avg + 1) * se.load.weight
808 * However, in many cases, the above util_avg does not give a desired
809 * value. Moreover, the sum of the util_avgs may be divergent, such
810 * as when the series is a harmonic series.
812 * To solve this problem, we also cap the util_avg of successive tasks to
813 * only 1/2 of the left utilization budget:
815 * util_avg_cap = (cpu_scale - cfs_rq->avg.util_avg) / 2^n
817 * where n denotes the nth task and cpu_scale the CPU capacity.
819 * For example, for a CPU with 1024 of capacity, a simplest series from
820 * the beginning would be like:
822 * task util_avg: 512, 256, 128, 64, 32, 16, 8, ...
823 * cfs_rq util_avg: 512, 768, 896, 960, 992, 1008, 1016, ...
825 * Finally, that extrapolated util_avg is clamped to the cap (util_avg_cap)
826 * if util_avg > util_avg_cap.
828 void post_init_entity_util_avg(struct task_struct *p)
830 struct sched_entity *se = &p->se;
831 struct cfs_rq *cfs_rq = cfs_rq_of(se);
832 struct sched_avg *sa = &se->avg;
833 long cpu_scale = arch_scale_cpu_capacity(cpu_of(rq_of(cfs_rq)));
834 long cap = (long)(cpu_scale - cfs_rq->avg.util_avg) / 2;
836 if (p->sched_class != &fair_sched_class) {
838 * For !fair tasks do:
840 update_cfs_rq_load_avg(now, cfs_rq);
841 attach_entity_load_avg(cfs_rq, se);
842 switched_from_fair(rq, p);
844 * such that the next switched_to_fair() has the
847 se->avg.last_update_time = cfs_rq_clock_pelt(cfs_rq);
852 if (cfs_rq->avg.util_avg != 0) {
853 sa->util_avg = cfs_rq->avg.util_avg * se->load.weight;
854 sa->util_avg /= (cfs_rq->avg.load_avg + 1);
856 if (sa->util_avg > cap)
863 sa->runnable_avg = sa->util_avg;
866 #else /* !CONFIG_SMP */
867 void init_entity_runnable_average(struct sched_entity *se)
870 void post_init_entity_util_avg(struct task_struct *p)
873 static void update_tg_load_avg(struct cfs_rq *cfs_rq)
876 #endif /* CONFIG_SMP */
879 * Update the current task's runtime statistics.
881 static void update_curr(struct cfs_rq *cfs_rq)
883 struct sched_entity *curr = cfs_rq->curr;
884 u64 now = rq_clock_task(rq_of(cfs_rq));
890 delta_exec = now - curr->exec_start;
891 if (unlikely((s64)delta_exec <= 0))
894 curr->exec_start = now;
896 if (schedstat_enabled()) {
897 struct sched_statistics *stats;
899 stats = __schedstats_from_se(curr);
900 __schedstat_set(stats->exec_max,
901 max(delta_exec, stats->exec_max));
904 curr->sum_exec_runtime += delta_exec;
905 schedstat_add(cfs_rq->exec_clock, delta_exec);
907 curr->vruntime += calc_delta_fair(delta_exec, curr);
908 update_min_vruntime(cfs_rq);
910 if (entity_is_task(curr)) {
911 struct task_struct *curtask = task_of(curr);
913 trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime);
914 cgroup_account_cputime(curtask, delta_exec);
915 account_group_exec_runtime(curtask, delta_exec);
918 account_cfs_rq_runtime(cfs_rq, delta_exec);
921 static void update_curr_fair(struct rq *rq)
923 update_curr(cfs_rq_of(&rq->curr->se));
927 update_stats_wait_start_fair(struct cfs_rq *cfs_rq, struct sched_entity *se)
929 struct sched_statistics *stats;
930 struct task_struct *p = NULL;
932 if (!schedstat_enabled())
935 stats = __schedstats_from_se(se);
937 if (entity_is_task(se))
940 __update_stats_wait_start(rq_of(cfs_rq), p, stats);
944 update_stats_wait_end_fair(struct cfs_rq *cfs_rq, struct sched_entity *se)
946 struct sched_statistics *stats;
947 struct task_struct *p = NULL;
949 if (!schedstat_enabled())
952 stats = __schedstats_from_se(se);
955 * When the sched_schedstat changes from 0 to 1, some sched se
956 * maybe already in the runqueue, the se->statistics.wait_start
957 * will be 0.So it will let the delta wrong. We need to avoid this
960 if (unlikely(!schedstat_val(stats->wait_start)))
963 if (entity_is_task(se))
966 __update_stats_wait_end(rq_of(cfs_rq), p, stats);
970 update_stats_enqueue_sleeper_fair(struct cfs_rq *cfs_rq, struct sched_entity *se)
972 struct sched_statistics *stats;
973 struct task_struct *tsk = NULL;
975 if (!schedstat_enabled())
978 stats = __schedstats_from_se(se);
980 if (entity_is_task(se))
983 __update_stats_enqueue_sleeper(rq_of(cfs_rq), tsk, stats);
987 * Task is being enqueued - update stats:
990 update_stats_enqueue_fair(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
992 if (!schedstat_enabled())
996 * Are we enqueueing a waiting task? (for current tasks
997 * a dequeue/enqueue event is a NOP)
999 if (se != cfs_rq->curr)
1000 update_stats_wait_start_fair(cfs_rq, se);
1002 if (flags & ENQUEUE_WAKEUP)
1003 update_stats_enqueue_sleeper_fair(cfs_rq, se);
1007 update_stats_dequeue_fair(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
1010 if (!schedstat_enabled())
1014 * Mark the end of the wait period if dequeueing a
1017 if (se != cfs_rq->curr)
1018 update_stats_wait_end_fair(cfs_rq, se);
1020 if ((flags & DEQUEUE_SLEEP) && entity_is_task(se)) {
1021 struct task_struct *tsk = task_of(se);
1024 /* XXX racy against TTWU */
1025 state = READ_ONCE(tsk->__state);
1026 if (state & TASK_INTERRUPTIBLE)
1027 __schedstat_set(tsk->stats.sleep_start,
1028 rq_clock(rq_of(cfs_rq)));
1029 if (state & TASK_UNINTERRUPTIBLE)
1030 __schedstat_set(tsk->stats.block_start,
1031 rq_clock(rq_of(cfs_rq)));
1036 * We are picking a new current task - update its stats:
1039 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
1042 * We are starting a new run period:
1044 se->exec_start = rq_clock_task(rq_of(cfs_rq));
1047 /**************************************************
1048 * Scheduling class queueing methods:
1052 #define NUMA_IMBALANCE_MIN 2
1055 adjust_numa_imbalance(int imbalance, int dst_running, int imb_numa_nr)
1058 * Allow a NUMA imbalance if busy CPUs is less than the maximum
1059 * threshold. Above this threshold, individual tasks may be contending
1060 * for both memory bandwidth and any shared HT resources. This is an
1061 * approximation as the number of running tasks may not be related to
1062 * the number of busy CPUs due to sched_setaffinity.
1064 if (dst_running > imb_numa_nr)
1068 * Allow a small imbalance based on a simple pair of communicating
1069 * tasks that remain local when the destination is lightly loaded.
1071 if (imbalance <= NUMA_IMBALANCE_MIN)
1076 #endif /* CONFIG_NUMA */
1078 #ifdef CONFIG_NUMA_BALANCING
1080 * Approximate time to scan a full NUMA task in ms. The task scan period is
1081 * calculated based on the tasks virtual memory size and
1082 * numa_balancing_scan_size.
1084 unsigned int sysctl_numa_balancing_scan_period_min = 1000;
1085 unsigned int sysctl_numa_balancing_scan_period_max = 60000;
1087 /* Portion of address space to scan in MB */
1088 unsigned int sysctl_numa_balancing_scan_size = 256;
1090 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
1091 unsigned int sysctl_numa_balancing_scan_delay = 1000;
1094 refcount_t refcount;
1096 spinlock_t lock; /* nr_tasks, tasks */
1101 struct rcu_head rcu;
1102 unsigned long total_faults;
1103 unsigned long max_faults_cpu;
1105 * faults[] array is split into two regions: faults_mem and faults_cpu.
1107 * Faults_cpu is used to decide whether memory should move
1108 * towards the CPU. As a consequence, these stats are weighted
1109 * more by CPU use than by memory faults.
1111 unsigned long faults[];
1115 * For functions that can be called in multiple contexts that permit reading
1116 * ->numa_group (see struct task_struct for locking rules).
1118 static struct numa_group *deref_task_numa_group(struct task_struct *p)
1120 return rcu_dereference_check(p->numa_group, p == current ||
1121 (lockdep_is_held(__rq_lockp(task_rq(p))) && !READ_ONCE(p->on_cpu)));
1124 static struct numa_group *deref_curr_numa_group(struct task_struct *p)
1126 return rcu_dereference_protected(p->numa_group, p == current);
1129 static inline unsigned long group_faults_priv(struct numa_group *ng);
1130 static inline unsigned long group_faults_shared(struct numa_group *ng);
1132 static unsigned int task_nr_scan_windows(struct task_struct *p)
1134 unsigned long rss = 0;
1135 unsigned long nr_scan_pages;
1138 * Calculations based on RSS as non-present and empty pages are skipped
1139 * by the PTE scanner and NUMA hinting faults should be trapped based
1142 nr_scan_pages = sysctl_numa_balancing_scan_size << (20 - PAGE_SHIFT);
1143 rss = get_mm_rss(p->mm);
1145 rss = nr_scan_pages;
1147 rss = round_up(rss, nr_scan_pages);
1148 return rss / nr_scan_pages;
1151 /* For sanity's sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */
1152 #define MAX_SCAN_WINDOW 2560
1154 static unsigned int task_scan_min(struct task_struct *p)
1156 unsigned int scan_size = READ_ONCE(sysctl_numa_balancing_scan_size);
1157 unsigned int scan, floor;
1158 unsigned int windows = 1;
1160 if (scan_size < MAX_SCAN_WINDOW)
1161 windows = MAX_SCAN_WINDOW / scan_size;
1162 floor = 1000 / windows;
1164 scan = sysctl_numa_balancing_scan_period_min / task_nr_scan_windows(p);
1165 return max_t(unsigned int, floor, scan);
1168 static unsigned int task_scan_start(struct task_struct *p)
1170 unsigned long smin = task_scan_min(p);
1171 unsigned long period = smin;
1172 struct numa_group *ng;
1174 /* Scale the maximum scan period with the amount of shared memory. */
1176 ng = rcu_dereference(p->numa_group);
1178 unsigned long shared = group_faults_shared(ng);
1179 unsigned long private = group_faults_priv(ng);
1181 period *= refcount_read(&ng->refcount);
1182 period *= shared + 1;
1183 period /= private + shared + 1;
1187 return max(smin, period);
1190 static unsigned int task_scan_max(struct task_struct *p)
1192 unsigned long smin = task_scan_min(p);
1194 struct numa_group *ng;
1196 /* Watch for min being lower than max due to floor calculations */
1197 smax = sysctl_numa_balancing_scan_period_max / task_nr_scan_windows(p);
1199 /* Scale the maximum scan period with the amount of shared memory. */
1200 ng = deref_curr_numa_group(p);
1202 unsigned long shared = group_faults_shared(ng);
1203 unsigned long private = group_faults_priv(ng);
1204 unsigned long period = smax;
1206 period *= refcount_read(&ng->refcount);
1207 period *= shared + 1;
1208 period /= private + shared + 1;
1210 smax = max(smax, period);
1213 return max(smin, smax);
1216 static void account_numa_enqueue(struct rq *rq, struct task_struct *p)
1218 rq->nr_numa_running += (p->numa_preferred_nid != NUMA_NO_NODE);
1219 rq->nr_preferred_running += (p->numa_preferred_nid == task_node(p));
1222 static void account_numa_dequeue(struct rq *rq, struct task_struct *p)
1224 rq->nr_numa_running -= (p->numa_preferred_nid != NUMA_NO_NODE);
1225 rq->nr_preferred_running -= (p->numa_preferred_nid == task_node(p));
1228 /* Shared or private faults. */
1229 #define NR_NUMA_HINT_FAULT_TYPES 2
1231 /* Memory and CPU locality */
1232 #define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2)
1234 /* Averaged statistics, and temporary buffers. */
1235 #define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2)
1237 pid_t task_numa_group_id(struct task_struct *p)
1239 struct numa_group *ng;
1243 ng = rcu_dereference(p->numa_group);
1252 * The averaged statistics, shared & private, memory & CPU,
1253 * occupy the first half of the array. The second half of the
1254 * array is for current counters, which are averaged into the
1255 * first set by task_numa_placement.
1257 static inline int task_faults_idx(enum numa_faults_stats s, int nid, int priv)
1259 return NR_NUMA_HINT_FAULT_TYPES * (s * nr_node_ids + nid) + priv;
1262 static inline unsigned long task_faults(struct task_struct *p, int nid)
1264 if (!p->numa_faults)
1267 return p->numa_faults[task_faults_idx(NUMA_MEM, nid, 0)] +
1268 p->numa_faults[task_faults_idx(NUMA_MEM, nid, 1)];
1271 static inline unsigned long group_faults(struct task_struct *p, int nid)
1273 struct numa_group *ng = deref_task_numa_group(p);
1278 return ng->faults[task_faults_idx(NUMA_MEM, nid, 0)] +
1279 ng->faults[task_faults_idx(NUMA_MEM, nid, 1)];
1282 static inline unsigned long group_faults_cpu(struct numa_group *group, int nid)
1284 return group->faults[task_faults_idx(NUMA_CPU, nid, 0)] +
1285 group->faults[task_faults_idx(NUMA_CPU, nid, 1)];
1288 static inline unsigned long group_faults_priv(struct numa_group *ng)
1290 unsigned long faults = 0;
1293 for_each_online_node(node) {
1294 faults += ng->faults[task_faults_idx(NUMA_MEM, node, 1)];
1300 static inline unsigned long group_faults_shared(struct numa_group *ng)
1302 unsigned long faults = 0;
1305 for_each_online_node(node) {
1306 faults += ng->faults[task_faults_idx(NUMA_MEM, node, 0)];
1313 * A node triggering more than 1/3 as many NUMA faults as the maximum is
1314 * considered part of a numa group's pseudo-interleaving set. Migrations
1315 * between these nodes are slowed down, to allow things to settle down.
1317 #define ACTIVE_NODE_FRACTION 3
1319 static bool numa_is_active_node(int nid, struct numa_group *ng)
1321 return group_faults_cpu(ng, nid) * ACTIVE_NODE_FRACTION > ng->max_faults_cpu;
1324 /* Handle placement on systems where not all nodes are directly connected. */
1325 static unsigned long score_nearby_nodes(struct task_struct *p, int nid,
1326 int lim_dist, bool task)
1328 unsigned long score = 0;
1332 * All nodes are directly connected, and the same distance
1333 * from each other. No need for fancy placement algorithms.
1335 if (sched_numa_topology_type == NUMA_DIRECT)
1338 /* sched_max_numa_distance may be changed in parallel. */
1339 max_dist = READ_ONCE(sched_max_numa_distance);
1341 * This code is called for each node, introducing N^2 complexity,
1342 * which should be ok given the number of nodes rarely exceeds 8.
1344 for_each_online_node(node) {
1345 unsigned long faults;
1346 int dist = node_distance(nid, node);
1349 * The furthest away nodes in the system are not interesting
1350 * for placement; nid was already counted.
1352 if (dist >= max_dist || node == nid)
1356 * On systems with a backplane NUMA topology, compare groups
1357 * of nodes, and move tasks towards the group with the most
1358 * memory accesses. When comparing two nodes at distance
1359 * "hoplimit", only nodes closer by than "hoplimit" are part
1360 * of each group. Skip other nodes.
1362 if (sched_numa_topology_type == NUMA_BACKPLANE && dist >= lim_dist)
1365 /* Add up the faults from nearby nodes. */
1367 faults = task_faults(p, node);
1369 faults = group_faults(p, node);
1372 * On systems with a glueless mesh NUMA topology, there are
1373 * no fixed "groups of nodes". Instead, nodes that are not
1374 * directly connected bounce traffic through intermediate
1375 * nodes; a numa_group can occupy any set of nodes.
1376 * The further away a node is, the less the faults count.
1377 * This seems to result in good task placement.
1379 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
1380 faults *= (max_dist - dist);
1381 faults /= (max_dist - LOCAL_DISTANCE);
1391 * These return the fraction of accesses done by a particular task, or
1392 * task group, on a particular numa node. The group weight is given a
1393 * larger multiplier, in order to group tasks together that are almost
1394 * evenly spread out between numa nodes.
1396 static inline unsigned long task_weight(struct task_struct *p, int nid,
1399 unsigned long faults, total_faults;
1401 if (!p->numa_faults)
1404 total_faults = p->total_numa_faults;
1409 faults = task_faults(p, nid);
1410 faults += score_nearby_nodes(p, nid, dist, true);
1412 return 1000 * faults / total_faults;
1415 static inline unsigned long group_weight(struct task_struct *p, int nid,
1418 struct numa_group *ng = deref_task_numa_group(p);
1419 unsigned long faults, total_faults;
1424 total_faults = ng->total_faults;
1429 faults = group_faults(p, nid);
1430 faults += score_nearby_nodes(p, nid, dist, false);
1432 return 1000 * faults / total_faults;
1435 bool should_numa_migrate_memory(struct task_struct *p, struct page * page,
1436 int src_nid, int dst_cpu)
1438 struct numa_group *ng = deref_curr_numa_group(p);
1439 int dst_nid = cpu_to_node(dst_cpu);
1440 int last_cpupid, this_cpupid;
1442 this_cpupid = cpu_pid_to_cpupid(dst_cpu, current->pid);
1443 last_cpupid = page_cpupid_xchg_last(page, this_cpupid);
1446 * Allow first faults or private faults to migrate immediately early in
1447 * the lifetime of a task. The magic number 4 is based on waiting for
1448 * two full passes of the "multi-stage node selection" test that is
1451 if ((p->numa_preferred_nid == NUMA_NO_NODE || p->numa_scan_seq <= 4) &&
1452 (cpupid_pid_unset(last_cpupid) || cpupid_match_pid(p, last_cpupid)))
1456 * Multi-stage node selection is used in conjunction with a periodic
1457 * migration fault to build a temporal task<->page relation. By using
1458 * a two-stage filter we remove short/unlikely relations.
1460 * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate
1461 * a task's usage of a particular page (n_p) per total usage of this
1462 * page (n_t) (in a given time-span) to a probability.
1464 * Our periodic faults will sample this probability and getting the
1465 * same result twice in a row, given these samples are fully
1466 * independent, is then given by P(n)^2, provided our sample period
1467 * is sufficiently short compared to the usage pattern.
1469 * This quadric squishes small probabilities, making it less likely we
1470 * act on an unlikely task<->page relation.
1472 if (!cpupid_pid_unset(last_cpupid) &&
1473 cpupid_to_nid(last_cpupid) != dst_nid)
1476 /* Always allow migrate on private faults */
1477 if (cpupid_match_pid(p, last_cpupid))
1480 /* A shared fault, but p->numa_group has not been set up yet. */
1485 * Destination node is much more heavily used than the source
1486 * node? Allow migration.
1488 if (group_faults_cpu(ng, dst_nid) > group_faults_cpu(ng, src_nid) *
1489 ACTIVE_NODE_FRACTION)
1493 * Distribute memory according to CPU & memory use on each node,
1494 * with 3/4 hysteresis to avoid unnecessary memory migrations:
1496 * faults_cpu(dst) 3 faults_cpu(src)
1497 * --------------- * - > ---------------
1498 * faults_mem(dst) 4 faults_mem(src)
1500 return group_faults_cpu(ng, dst_nid) * group_faults(p, src_nid) * 3 >
1501 group_faults_cpu(ng, src_nid) * group_faults(p, dst_nid) * 4;
1505 * 'numa_type' describes the node at the moment of load balancing.
1508 /* The node has spare capacity that can be used to run more tasks. */
1511 * The node is fully used and the tasks don't compete for more CPU
1512 * cycles. Nevertheless, some tasks might wait before running.
1516 * The node is overloaded and can't provide expected CPU cycles to all
1522 /* Cached statistics for all CPUs within a node */
1525 unsigned long runnable;
1527 /* Total compute capacity of CPUs on a node */
1528 unsigned long compute_capacity;
1529 unsigned int nr_running;
1530 unsigned int weight;
1531 enum numa_type node_type;
1535 static inline bool is_core_idle(int cpu)
1537 #ifdef CONFIG_SCHED_SMT
1540 for_each_cpu(sibling, cpu_smt_mask(cpu)) {
1544 if (!idle_cpu(sibling))
1552 struct task_numa_env {
1553 struct task_struct *p;
1555 int src_cpu, src_nid;
1556 int dst_cpu, dst_nid;
1559 struct numa_stats src_stats, dst_stats;
1564 struct task_struct *best_task;
1569 static unsigned long cpu_load(struct rq *rq);
1570 static unsigned long cpu_runnable(struct rq *rq);
1573 numa_type numa_classify(unsigned int imbalance_pct,
1574 struct numa_stats *ns)
1576 if ((ns->nr_running > ns->weight) &&
1577 (((ns->compute_capacity * 100) < (ns->util * imbalance_pct)) ||
1578 ((ns->compute_capacity * imbalance_pct) < (ns->runnable * 100))))
1579 return node_overloaded;
1581 if ((ns->nr_running < ns->weight) ||
1582 (((ns->compute_capacity * 100) > (ns->util * imbalance_pct)) &&
1583 ((ns->compute_capacity * imbalance_pct) > (ns->runnable * 100))))
1584 return node_has_spare;
1586 return node_fully_busy;
1589 #ifdef CONFIG_SCHED_SMT
1590 /* Forward declarations of select_idle_sibling helpers */
1591 static inline bool test_idle_cores(int cpu, bool def);
1592 static inline int numa_idle_core(int idle_core, int cpu)
1594 if (!static_branch_likely(&sched_smt_present) ||
1595 idle_core >= 0 || !test_idle_cores(cpu, false))
1599 * Prefer cores instead of packing HT siblings
1600 * and triggering future load balancing.
1602 if (is_core_idle(cpu))
1608 static inline int numa_idle_core(int idle_core, int cpu)
1615 * Gather all necessary information to make NUMA balancing placement
1616 * decisions that are compatible with standard load balancer. This
1617 * borrows code and logic from update_sg_lb_stats but sharing a
1618 * common implementation is impractical.
1620 static void update_numa_stats(struct task_numa_env *env,
1621 struct numa_stats *ns, int nid,
1624 int cpu, idle_core = -1;
1626 memset(ns, 0, sizeof(*ns));
1630 for_each_cpu(cpu, cpumask_of_node(nid)) {
1631 struct rq *rq = cpu_rq(cpu);
1633 ns->load += cpu_load(rq);
1634 ns->runnable += cpu_runnable(rq);
1635 ns->util += cpu_util_cfs(cpu);
1636 ns->nr_running += rq->cfs.h_nr_running;
1637 ns->compute_capacity += capacity_of(cpu);
1639 if (find_idle && !rq->nr_running && idle_cpu(cpu)) {
1640 if (READ_ONCE(rq->numa_migrate_on) ||
1641 !cpumask_test_cpu(cpu, env->p->cpus_ptr))
1644 if (ns->idle_cpu == -1)
1647 idle_core = numa_idle_core(idle_core, cpu);
1652 ns->weight = cpumask_weight(cpumask_of_node(nid));
1654 ns->node_type = numa_classify(env->imbalance_pct, ns);
1657 ns->idle_cpu = idle_core;
1660 static void task_numa_assign(struct task_numa_env *env,
1661 struct task_struct *p, long imp)
1663 struct rq *rq = cpu_rq(env->dst_cpu);
1665 /* Check if run-queue part of active NUMA balance. */
1666 if (env->best_cpu != env->dst_cpu && xchg(&rq->numa_migrate_on, 1)) {
1668 int start = env->dst_cpu;
1670 /* Find alternative idle CPU. */
1671 for_each_cpu_wrap(cpu, cpumask_of_node(env->dst_nid), start) {
1672 if (cpu == env->best_cpu || !idle_cpu(cpu) ||
1673 !cpumask_test_cpu(cpu, env->p->cpus_ptr)) {
1678 rq = cpu_rq(env->dst_cpu);
1679 if (!xchg(&rq->numa_migrate_on, 1))
1683 /* Failed to find an alternative idle CPU */
1689 * Clear previous best_cpu/rq numa-migrate flag, since task now
1690 * found a better CPU to move/swap.
1692 if (env->best_cpu != -1 && env->best_cpu != env->dst_cpu) {
1693 rq = cpu_rq(env->best_cpu);
1694 WRITE_ONCE(rq->numa_migrate_on, 0);
1698 put_task_struct(env->best_task);
1703 env->best_imp = imp;
1704 env->best_cpu = env->dst_cpu;
1707 static bool load_too_imbalanced(long src_load, long dst_load,
1708 struct task_numa_env *env)
1711 long orig_src_load, orig_dst_load;
1712 long src_capacity, dst_capacity;
1715 * The load is corrected for the CPU capacity available on each node.
1718 * ------------ vs ---------
1719 * src_capacity dst_capacity
1721 src_capacity = env->src_stats.compute_capacity;
1722 dst_capacity = env->dst_stats.compute_capacity;
1724 imb = abs(dst_load * src_capacity - src_load * dst_capacity);
1726 orig_src_load = env->src_stats.load;
1727 orig_dst_load = env->dst_stats.load;
1729 old_imb = abs(orig_dst_load * src_capacity - orig_src_load * dst_capacity);
1731 /* Would this change make things worse? */
1732 return (imb > old_imb);
1736 * Maximum NUMA importance can be 1998 (2*999);
1737 * SMALLIMP @ 30 would be close to 1998/64.
1738 * Used to deter task migration.
1743 * This checks if the overall compute and NUMA accesses of the system would
1744 * be improved if the source tasks was migrated to the target dst_cpu taking
1745 * into account that it might be best if task running on the dst_cpu should
1746 * be exchanged with the source task
1748 static bool task_numa_compare(struct task_numa_env *env,
1749 long taskimp, long groupimp, bool maymove)
1751 struct numa_group *cur_ng, *p_ng = deref_curr_numa_group(env->p);
1752 struct rq *dst_rq = cpu_rq(env->dst_cpu);
1753 long imp = p_ng ? groupimp : taskimp;
1754 struct task_struct *cur;
1755 long src_load, dst_load;
1756 int dist = env->dist;
1759 bool stopsearch = false;
1761 if (READ_ONCE(dst_rq->numa_migrate_on))
1765 cur = rcu_dereference(dst_rq->curr);
1766 if (cur && ((cur->flags & PF_EXITING) || is_idle_task(cur)))
1770 * Because we have preemption enabled we can get migrated around and
1771 * end try selecting ourselves (current == env->p) as a swap candidate.
1773 if (cur == env->p) {
1779 if (maymove && moveimp >= env->best_imp)
1785 /* Skip this swap candidate if cannot move to the source cpu. */
1786 if (!cpumask_test_cpu(env->src_cpu, cur->cpus_ptr))
1790 * Skip this swap candidate if it is not moving to its preferred
1791 * node and the best task is.
1793 if (env->best_task &&
1794 env->best_task->numa_preferred_nid == env->src_nid &&
1795 cur->numa_preferred_nid != env->src_nid) {
1800 * "imp" is the fault differential for the source task between the
1801 * source and destination node. Calculate the total differential for
1802 * the source task and potential destination task. The more negative
1803 * the value is, the more remote accesses that would be expected to
1804 * be incurred if the tasks were swapped.
1806 * If dst and source tasks are in the same NUMA group, or not
1807 * in any group then look only at task weights.
1809 cur_ng = rcu_dereference(cur->numa_group);
1810 if (cur_ng == p_ng) {
1812 * Do not swap within a group or between tasks that have
1813 * no group if there is spare capacity. Swapping does
1814 * not address the load imbalance and helps one task at
1815 * the cost of punishing another.
1817 if (env->dst_stats.node_type == node_has_spare)
1820 imp = taskimp + task_weight(cur, env->src_nid, dist) -
1821 task_weight(cur, env->dst_nid, dist);
1823 * Add some hysteresis to prevent swapping the
1824 * tasks within a group over tiny differences.
1830 * Compare the group weights. If a task is all by itself
1831 * (not part of a group), use the task weight instead.
1834 imp += group_weight(cur, env->src_nid, dist) -
1835 group_weight(cur, env->dst_nid, dist);
1837 imp += task_weight(cur, env->src_nid, dist) -
1838 task_weight(cur, env->dst_nid, dist);
1841 /* Discourage picking a task already on its preferred node */
1842 if (cur->numa_preferred_nid == env->dst_nid)
1846 * Encourage picking a task that moves to its preferred node.
1847 * This potentially makes imp larger than it's maximum of
1848 * 1998 (see SMALLIMP and task_weight for why) but in this
1849 * case, it does not matter.
1851 if (cur->numa_preferred_nid == env->src_nid)
1854 if (maymove && moveimp > imp && moveimp > env->best_imp) {
1861 * Prefer swapping with a task moving to its preferred node over a
1864 if (env->best_task && cur->numa_preferred_nid == env->src_nid &&
1865 env->best_task->numa_preferred_nid != env->src_nid) {
1870 * If the NUMA importance is less than SMALLIMP,
1871 * task migration might only result in ping pong
1872 * of tasks and also hurt performance due to cache
1875 if (imp < SMALLIMP || imp <= env->best_imp + SMALLIMP / 2)
1879 * In the overloaded case, try and keep the load balanced.
1881 load = task_h_load(env->p) - task_h_load(cur);
1885 dst_load = env->dst_stats.load + load;
1886 src_load = env->src_stats.load - load;
1888 if (load_too_imbalanced(src_load, dst_load, env))
1892 /* Evaluate an idle CPU for a task numa move. */
1894 int cpu = env->dst_stats.idle_cpu;
1896 /* Nothing cached so current CPU went idle since the search. */
1901 * If the CPU is no longer truly idle and the previous best CPU
1902 * is, keep using it.
1904 if (!idle_cpu(cpu) && env->best_cpu >= 0 &&
1905 idle_cpu(env->best_cpu)) {
1906 cpu = env->best_cpu;
1912 task_numa_assign(env, cur, imp);
1915 * If a move to idle is allowed because there is capacity or load
1916 * balance improves then stop the search. While a better swap
1917 * candidate may exist, a search is not free.
1919 if (maymove && !cur && env->best_cpu >= 0 && idle_cpu(env->best_cpu))
1923 * If a swap candidate must be identified and the current best task
1924 * moves its preferred node then stop the search.
1926 if (!maymove && env->best_task &&
1927 env->best_task->numa_preferred_nid == env->src_nid) {
1936 static void task_numa_find_cpu(struct task_numa_env *env,
1937 long taskimp, long groupimp)
1939 bool maymove = false;
1943 * If dst node has spare capacity, then check if there is an
1944 * imbalance that would be overruled by the load balancer.
1946 if (env->dst_stats.node_type == node_has_spare) {
1947 unsigned int imbalance;
1948 int src_running, dst_running;
1951 * Would movement cause an imbalance? Note that if src has
1952 * more running tasks that the imbalance is ignored as the
1953 * move improves the imbalance from the perspective of the
1954 * CPU load balancer.
1956 src_running = env->src_stats.nr_running - 1;
1957 dst_running = env->dst_stats.nr_running + 1;
1958 imbalance = max(0, dst_running - src_running);
1959 imbalance = adjust_numa_imbalance(imbalance, dst_running,
1962 /* Use idle CPU if there is no imbalance */
1965 if (env->dst_stats.idle_cpu >= 0) {
1966 env->dst_cpu = env->dst_stats.idle_cpu;
1967 task_numa_assign(env, NULL, 0);
1972 long src_load, dst_load, load;
1974 * If the improvement from just moving env->p direction is better
1975 * than swapping tasks around, check if a move is possible.
1977 load = task_h_load(env->p);
1978 dst_load = env->dst_stats.load + load;
1979 src_load = env->src_stats.load - load;
1980 maymove = !load_too_imbalanced(src_load, dst_load, env);
1983 for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) {
1984 /* Skip this CPU if the source task cannot migrate */
1985 if (!cpumask_test_cpu(cpu, env->p->cpus_ptr))
1989 if (task_numa_compare(env, taskimp, groupimp, maymove))
1994 static int task_numa_migrate(struct task_struct *p)
1996 struct task_numa_env env = {
1999 .src_cpu = task_cpu(p),
2000 .src_nid = task_node(p),
2002 .imbalance_pct = 112,
2008 unsigned long taskweight, groupweight;
2009 struct sched_domain *sd;
2010 long taskimp, groupimp;
2011 struct numa_group *ng;
2016 * Pick the lowest SD_NUMA domain, as that would have the smallest
2017 * imbalance and would be the first to start moving tasks about.
2019 * And we want to avoid any moving of tasks about, as that would create
2020 * random movement of tasks -- counter the numa conditions we're trying
2024 sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu));
2026 env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2;
2027 env.imb_numa_nr = sd->imb_numa_nr;
2032 * Cpusets can break the scheduler domain tree into smaller
2033 * balance domains, some of which do not cross NUMA boundaries.
2034 * Tasks that are "trapped" in such domains cannot be migrated
2035 * elsewhere, so there is no point in (re)trying.
2037 if (unlikely(!sd)) {
2038 sched_setnuma(p, task_node(p));
2042 env.dst_nid = p->numa_preferred_nid;
2043 dist = env.dist = node_distance(env.src_nid, env.dst_nid);
2044 taskweight = task_weight(p, env.src_nid, dist);
2045 groupweight = group_weight(p, env.src_nid, dist);
2046 update_numa_stats(&env, &env.src_stats, env.src_nid, false);
2047 taskimp = task_weight(p, env.dst_nid, dist) - taskweight;
2048 groupimp = group_weight(p, env.dst_nid, dist) - groupweight;
2049 update_numa_stats(&env, &env.dst_stats, env.dst_nid, true);
2051 /* Try to find a spot on the preferred nid. */
2052 task_numa_find_cpu(&env, taskimp, groupimp);
2055 * Look at other nodes in these cases:
2056 * - there is no space available on the preferred_nid
2057 * - the task is part of a numa_group that is interleaved across
2058 * multiple NUMA nodes; in order to better consolidate the group,
2059 * we need to check other locations.
2061 ng = deref_curr_numa_group(p);
2062 if (env.best_cpu == -1 || (ng && ng->active_nodes > 1)) {
2063 for_each_node_state(nid, N_CPU) {
2064 if (nid == env.src_nid || nid == p->numa_preferred_nid)
2067 dist = node_distance(env.src_nid, env.dst_nid);
2068 if (sched_numa_topology_type == NUMA_BACKPLANE &&
2070 taskweight = task_weight(p, env.src_nid, dist);
2071 groupweight = group_weight(p, env.src_nid, dist);
2074 /* Only consider nodes where both task and groups benefit */
2075 taskimp = task_weight(p, nid, dist) - taskweight;
2076 groupimp = group_weight(p, nid, dist) - groupweight;
2077 if (taskimp < 0 && groupimp < 0)
2082 update_numa_stats(&env, &env.dst_stats, env.dst_nid, true);
2083 task_numa_find_cpu(&env, taskimp, groupimp);
2088 * If the task is part of a workload that spans multiple NUMA nodes,
2089 * and is migrating into one of the workload's active nodes, remember
2090 * this node as the task's preferred numa node, so the workload can
2092 * A task that migrated to a second choice node will be better off
2093 * trying for a better one later. Do not set the preferred node here.
2096 if (env.best_cpu == -1)
2099 nid = cpu_to_node(env.best_cpu);
2101 if (nid != p->numa_preferred_nid)
2102 sched_setnuma(p, nid);
2105 /* No better CPU than the current one was found. */
2106 if (env.best_cpu == -1) {
2107 trace_sched_stick_numa(p, env.src_cpu, NULL, -1);
2111 best_rq = cpu_rq(env.best_cpu);
2112 if (env.best_task == NULL) {
2113 ret = migrate_task_to(p, env.best_cpu);
2114 WRITE_ONCE(best_rq->numa_migrate_on, 0);
2116 trace_sched_stick_numa(p, env.src_cpu, NULL, env.best_cpu);
2120 ret = migrate_swap(p, env.best_task, env.best_cpu, env.src_cpu);
2121 WRITE_ONCE(best_rq->numa_migrate_on, 0);
2124 trace_sched_stick_numa(p, env.src_cpu, env.best_task, env.best_cpu);
2125 put_task_struct(env.best_task);
2129 /* Attempt to migrate a task to a CPU on the preferred node. */
2130 static void numa_migrate_preferred(struct task_struct *p)
2132 unsigned long interval = HZ;
2134 /* This task has no NUMA fault statistics yet */
2135 if (unlikely(p->numa_preferred_nid == NUMA_NO_NODE || !p->numa_faults))
2138 /* Periodically retry migrating the task to the preferred node */
2139 interval = min(interval, msecs_to_jiffies(p->numa_scan_period) / 16);
2140 p->numa_migrate_retry = jiffies + interval;
2142 /* Success if task is already running on preferred CPU */
2143 if (task_node(p) == p->numa_preferred_nid)
2146 /* Otherwise, try migrate to a CPU on the preferred node */
2147 task_numa_migrate(p);
2151 * Find out how many nodes the workload is actively running on. Do this by
2152 * tracking the nodes from which NUMA hinting faults are triggered. This can
2153 * be different from the set of nodes where the workload's memory is currently
2156 static void numa_group_count_active_nodes(struct numa_group *numa_group)
2158 unsigned long faults, max_faults = 0;
2159 int nid, active_nodes = 0;
2161 for_each_node_state(nid, N_CPU) {
2162 faults = group_faults_cpu(numa_group, nid);
2163 if (faults > max_faults)
2164 max_faults = faults;
2167 for_each_node_state(nid, N_CPU) {
2168 faults = group_faults_cpu(numa_group, nid);
2169 if (faults * ACTIVE_NODE_FRACTION > max_faults)
2173 numa_group->max_faults_cpu = max_faults;
2174 numa_group->active_nodes = active_nodes;
2178 * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
2179 * increments. The more local the fault statistics are, the higher the scan
2180 * period will be for the next scan window. If local/(local+remote) ratio is
2181 * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS)
2182 * the scan period will decrease. Aim for 70% local accesses.
2184 #define NUMA_PERIOD_SLOTS 10
2185 #define NUMA_PERIOD_THRESHOLD 7
2188 * Increase the scan period (slow down scanning) if the majority of
2189 * our memory is already on our local node, or if the majority of
2190 * the page accesses are shared with other processes.
2191 * Otherwise, decrease the scan period.
2193 static void update_task_scan_period(struct task_struct *p,
2194 unsigned long shared, unsigned long private)
2196 unsigned int period_slot;
2197 int lr_ratio, ps_ratio;
2200 unsigned long remote = p->numa_faults_locality[0];
2201 unsigned long local = p->numa_faults_locality[1];
2204 * If there were no record hinting faults then either the task is
2205 * completely idle or all activity is in areas that are not of interest
2206 * to automatic numa balancing. Related to that, if there were failed
2207 * migration then it implies we are migrating too quickly or the local
2208 * node is overloaded. In either case, scan slower
2210 if (local + shared == 0 || p->numa_faults_locality[2]) {
2211 p->numa_scan_period = min(p->numa_scan_period_max,
2212 p->numa_scan_period << 1);
2214 p->mm->numa_next_scan = jiffies +
2215 msecs_to_jiffies(p->numa_scan_period);
2221 * Prepare to scale scan period relative to the current period.
2222 * == NUMA_PERIOD_THRESHOLD scan period stays the same
2223 * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
2224 * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
2226 period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS);
2227 lr_ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote);
2228 ps_ratio = (private * NUMA_PERIOD_SLOTS) / (private + shared);
2230 if (ps_ratio >= NUMA_PERIOD_THRESHOLD) {
2232 * Most memory accesses are local. There is no need to
2233 * do fast NUMA scanning, since memory is already local.
2235 int slot = ps_ratio - NUMA_PERIOD_THRESHOLD;
2238 diff = slot * period_slot;
2239 } else if (lr_ratio >= NUMA_PERIOD_THRESHOLD) {
2241 * Most memory accesses are shared with other tasks.
2242 * There is no point in continuing fast NUMA scanning,
2243 * since other tasks may just move the memory elsewhere.
2245 int slot = lr_ratio - NUMA_PERIOD_THRESHOLD;
2248 diff = slot * period_slot;
2251 * Private memory faults exceed (SLOTS-THRESHOLD)/SLOTS,
2252 * yet they are not on the local NUMA node. Speed up
2253 * NUMA scanning to get the memory moved over.
2255 int ratio = max(lr_ratio, ps_ratio);
2256 diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot;
2259 p->numa_scan_period = clamp(p->numa_scan_period + diff,
2260 task_scan_min(p), task_scan_max(p));
2261 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
2265 * Get the fraction of time the task has been running since the last
2266 * NUMA placement cycle. The scheduler keeps similar statistics, but
2267 * decays those on a 32ms period, which is orders of magnitude off
2268 * from the dozens-of-seconds NUMA balancing period. Use the scheduler
2269 * stats only if the task is so new there are no NUMA statistics yet.
2271 static u64 numa_get_avg_runtime(struct task_struct *p, u64 *period)
2273 u64 runtime, delta, now;
2274 /* Use the start of this time slice to avoid calculations. */
2275 now = p->se.exec_start;
2276 runtime = p->se.sum_exec_runtime;
2278 if (p->last_task_numa_placement) {
2279 delta = runtime - p->last_sum_exec_runtime;
2280 *period = now - p->last_task_numa_placement;
2282 /* Avoid time going backwards, prevent potential divide error: */
2283 if (unlikely((s64)*period < 0))
2286 delta = p->se.avg.load_sum;
2287 *period = LOAD_AVG_MAX;
2290 p->last_sum_exec_runtime = runtime;
2291 p->last_task_numa_placement = now;
2297 * Determine the preferred nid for a task in a numa_group. This needs to
2298 * be done in a way that produces consistent results with group_weight,
2299 * otherwise workloads might not converge.
2301 static int preferred_group_nid(struct task_struct *p, int nid)
2306 /* Direct connections between all NUMA nodes. */
2307 if (sched_numa_topology_type == NUMA_DIRECT)
2311 * On a system with glueless mesh NUMA topology, group_weight
2312 * scores nodes according to the number of NUMA hinting faults on
2313 * both the node itself, and on nearby nodes.
2315 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
2316 unsigned long score, max_score = 0;
2317 int node, max_node = nid;
2319 dist = sched_max_numa_distance;
2321 for_each_node_state(node, N_CPU) {
2322 score = group_weight(p, node, dist);
2323 if (score > max_score) {
2332 * Finding the preferred nid in a system with NUMA backplane
2333 * interconnect topology is more involved. The goal is to locate
2334 * tasks from numa_groups near each other in the system, and
2335 * untangle workloads from different sides of the system. This requires
2336 * searching down the hierarchy of node groups, recursively searching
2337 * inside the highest scoring group of nodes. The nodemask tricks
2338 * keep the complexity of the search down.
2340 nodes = node_states[N_CPU];
2341 for (dist = sched_max_numa_distance; dist > LOCAL_DISTANCE; dist--) {
2342 unsigned long max_faults = 0;
2343 nodemask_t max_group = NODE_MASK_NONE;
2346 /* Are there nodes at this distance from each other? */
2347 if (!find_numa_distance(dist))
2350 for_each_node_mask(a, nodes) {
2351 unsigned long faults = 0;
2352 nodemask_t this_group;
2353 nodes_clear(this_group);
2355 /* Sum group's NUMA faults; includes a==b case. */
2356 for_each_node_mask(b, nodes) {
2357 if (node_distance(a, b) < dist) {
2358 faults += group_faults(p, b);
2359 node_set(b, this_group);
2360 node_clear(b, nodes);
2364 /* Remember the top group. */
2365 if (faults > max_faults) {
2366 max_faults = faults;
2367 max_group = this_group;
2369 * subtle: at the smallest distance there is
2370 * just one node left in each "group", the
2371 * winner is the preferred nid.
2376 /* Next round, evaluate the nodes within max_group. */
2384 static void task_numa_placement(struct task_struct *p)
2386 int seq, nid, max_nid = NUMA_NO_NODE;
2387 unsigned long max_faults = 0;
2388 unsigned long fault_types[2] = { 0, 0 };
2389 unsigned long total_faults;
2390 u64 runtime, period;
2391 spinlock_t *group_lock = NULL;
2392 struct numa_group *ng;
2395 * The p->mm->numa_scan_seq field gets updated without
2396 * exclusive access. Use READ_ONCE() here to ensure
2397 * that the field is read in a single access:
2399 seq = READ_ONCE(p->mm->numa_scan_seq);
2400 if (p->numa_scan_seq == seq)
2402 p->numa_scan_seq = seq;
2403 p->numa_scan_period_max = task_scan_max(p);
2405 total_faults = p->numa_faults_locality[0] +
2406 p->numa_faults_locality[1];
2407 runtime = numa_get_avg_runtime(p, &period);
2409 /* If the task is part of a group prevent parallel updates to group stats */
2410 ng = deref_curr_numa_group(p);
2412 group_lock = &ng->lock;
2413 spin_lock_irq(group_lock);
2416 /* Find the node with the highest number of faults */
2417 for_each_online_node(nid) {
2418 /* Keep track of the offsets in numa_faults array */
2419 int mem_idx, membuf_idx, cpu_idx, cpubuf_idx;
2420 unsigned long faults = 0, group_faults = 0;
2423 for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) {
2424 long diff, f_diff, f_weight;
2426 mem_idx = task_faults_idx(NUMA_MEM, nid, priv);
2427 membuf_idx = task_faults_idx(NUMA_MEMBUF, nid, priv);
2428 cpu_idx = task_faults_idx(NUMA_CPU, nid, priv);
2429 cpubuf_idx = task_faults_idx(NUMA_CPUBUF, nid, priv);
2431 /* Decay existing window, copy faults since last scan */
2432 diff = p->numa_faults[membuf_idx] - p->numa_faults[mem_idx] / 2;
2433 fault_types[priv] += p->numa_faults[membuf_idx];
2434 p->numa_faults[membuf_idx] = 0;
2437 * Normalize the faults_from, so all tasks in a group
2438 * count according to CPU use, instead of by the raw
2439 * number of faults. Tasks with little runtime have
2440 * little over-all impact on throughput, and thus their
2441 * faults are less important.
2443 f_weight = div64_u64(runtime << 16, period + 1);
2444 f_weight = (f_weight * p->numa_faults[cpubuf_idx]) /
2446 f_diff = f_weight - p->numa_faults[cpu_idx] / 2;
2447 p->numa_faults[cpubuf_idx] = 0;
2449 p->numa_faults[mem_idx] += diff;
2450 p->numa_faults[cpu_idx] += f_diff;
2451 faults += p->numa_faults[mem_idx];
2452 p->total_numa_faults += diff;
2455 * safe because we can only change our own group
2457 * mem_idx represents the offset for a given
2458 * nid and priv in a specific region because it
2459 * is at the beginning of the numa_faults array.
2461 ng->faults[mem_idx] += diff;
2462 ng->faults[cpu_idx] += f_diff;
2463 ng->total_faults += diff;
2464 group_faults += ng->faults[mem_idx];
2469 if (faults > max_faults) {
2470 max_faults = faults;
2473 } else if (group_faults > max_faults) {
2474 max_faults = group_faults;
2479 /* Cannot migrate task to CPU-less node */
2480 if (max_nid != NUMA_NO_NODE && !node_state(max_nid, N_CPU)) {
2481 int near_nid = max_nid;
2482 int distance, near_distance = INT_MAX;
2484 for_each_node_state(nid, N_CPU) {
2485 distance = node_distance(max_nid, nid);
2486 if (distance < near_distance) {
2488 near_distance = distance;
2495 numa_group_count_active_nodes(ng);
2496 spin_unlock_irq(group_lock);
2497 max_nid = preferred_group_nid(p, max_nid);
2501 /* Set the new preferred node */
2502 if (max_nid != p->numa_preferred_nid)
2503 sched_setnuma(p, max_nid);
2506 update_task_scan_period(p, fault_types[0], fault_types[1]);
2509 static inline int get_numa_group(struct numa_group *grp)
2511 return refcount_inc_not_zero(&grp->refcount);
2514 static inline void put_numa_group(struct numa_group *grp)
2516 if (refcount_dec_and_test(&grp->refcount))
2517 kfree_rcu(grp, rcu);
2520 static void task_numa_group(struct task_struct *p, int cpupid, int flags,
2523 struct numa_group *grp, *my_grp;
2524 struct task_struct *tsk;
2526 int cpu = cpupid_to_cpu(cpupid);
2529 if (unlikely(!deref_curr_numa_group(p))) {
2530 unsigned int size = sizeof(struct numa_group) +
2531 NR_NUMA_HINT_FAULT_STATS *
2532 nr_node_ids * sizeof(unsigned long);
2534 grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
2538 refcount_set(&grp->refcount, 1);
2539 grp->active_nodes = 1;
2540 grp->max_faults_cpu = 0;
2541 spin_lock_init(&grp->lock);
2544 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2545 grp->faults[i] = p->numa_faults[i];
2547 grp->total_faults = p->total_numa_faults;
2550 rcu_assign_pointer(p->numa_group, grp);
2554 tsk = READ_ONCE(cpu_rq(cpu)->curr);
2556 if (!cpupid_match_pid(tsk, cpupid))
2559 grp = rcu_dereference(tsk->numa_group);
2563 my_grp = deref_curr_numa_group(p);
2568 * Only join the other group if its bigger; if we're the bigger group,
2569 * the other task will join us.
2571 if (my_grp->nr_tasks > grp->nr_tasks)
2575 * Tie-break on the grp address.
2577 if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp)
2580 /* Always join threads in the same process. */
2581 if (tsk->mm == current->mm)
2584 /* Simple filter to avoid false positives due to PID collisions */
2585 if (flags & TNF_SHARED)
2588 /* Update priv based on whether false sharing was detected */
2591 if (join && !get_numa_group(grp))
2599 WARN_ON_ONCE(irqs_disabled());
2600 double_lock_irq(&my_grp->lock, &grp->lock);
2602 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) {
2603 my_grp->faults[i] -= p->numa_faults[i];
2604 grp->faults[i] += p->numa_faults[i];
2606 my_grp->total_faults -= p->total_numa_faults;
2607 grp->total_faults += p->total_numa_faults;
2612 spin_unlock(&my_grp->lock);
2613 spin_unlock_irq(&grp->lock);
2615 rcu_assign_pointer(p->numa_group, grp);
2617 put_numa_group(my_grp);
2626 * Get rid of NUMA statistics associated with a task (either current or dead).
2627 * If @final is set, the task is dead and has reached refcount zero, so we can
2628 * safely free all relevant data structures. Otherwise, there might be
2629 * concurrent reads from places like load balancing and procfs, and we should
2630 * reset the data back to default state without freeing ->numa_faults.
2632 void task_numa_free(struct task_struct *p, bool final)
2634 /* safe: p either is current or is being freed by current */
2635 struct numa_group *grp = rcu_dereference_raw(p->numa_group);
2636 unsigned long *numa_faults = p->numa_faults;
2637 unsigned long flags;
2644 spin_lock_irqsave(&grp->lock, flags);
2645 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2646 grp->faults[i] -= p->numa_faults[i];
2647 grp->total_faults -= p->total_numa_faults;
2650 spin_unlock_irqrestore(&grp->lock, flags);
2651 RCU_INIT_POINTER(p->numa_group, NULL);
2652 put_numa_group(grp);
2656 p->numa_faults = NULL;
2659 p->total_numa_faults = 0;
2660 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2666 * Got a PROT_NONE fault for a page on @node.
2668 void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags)
2670 struct task_struct *p = current;
2671 bool migrated = flags & TNF_MIGRATED;
2672 int cpu_node = task_node(current);
2673 int local = !!(flags & TNF_FAULT_LOCAL);
2674 struct numa_group *ng;
2677 if (!static_branch_likely(&sched_numa_balancing))
2680 /* for example, ksmd faulting in a user's mm */
2684 /* Allocate buffer to track faults on a per-node basis */
2685 if (unlikely(!p->numa_faults)) {
2686 int size = sizeof(*p->numa_faults) *
2687 NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids;
2689 p->numa_faults = kzalloc(size, GFP_KERNEL|__GFP_NOWARN);
2690 if (!p->numa_faults)
2693 p->total_numa_faults = 0;
2694 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
2698 * First accesses are treated as private, otherwise consider accesses
2699 * to be private if the accessing pid has not changed
2701 if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) {
2704 priv = cpupid_match_pid(p, last_cpupid);
2705 if (!priv && !(flags & TNF_NO_GROUP))
2706 task_numa_group(p, last_cpupid, flags, &priv);
2710 * If a workload spans multiple NUMA nodes, a shared fault that
2711 * occurs wholly within the set of nodes that the workload is
2712 * actively using should be counted as local. This allows the
2713 * scan rate to slow down when a workload has settled down.
2715 ng = deref_curr_numa_group(p);
2716 if (!priv && !local && ng && ng->active_nodes > 1 &&
2717 numa_is_active_node(cpu_node, ng) &&
2718 numa_is_active_node(mem_node, ng))
2722 * Retry to migrate task to preferred node periodically, in case it
2723 * previously failed, or the scheduler moved us.
2725 if (time_after(jiffies, p->numa_migrate_retry)) {
2726 task_numa_placement(p);
2727 numa_migrate_preferred(p);
2731 p->numa_pages_migrated += pages;
2732 if (flags & TNF_MIGRATE_FAIL)
2733 p->numa_faults_locality[2] += pages;
2735 p->numa_faults[task_faults_idx(NUMA_MEMBUF, mem_node, priv)] += pages;
2736 p->numa_faults[task_faults_idx(NUMA_CPUBUF, cpu_node, priv)] += pages;
2737 p->numa_faults_locality[local] += pages;
2740 static void reset_ptenuma_scan(struct task_struct *p)
2743 * We only did a read acquisition of the mmap sem, so
2744 * p->mm->numa_scan_seq is written to without exclusive access
2745 * and the update is not guaranteed to be atomic. That's not
2746 * much of an issue though, since this is just used for
2747 * statistical sampling. Use READ_ONCE/WRITE_ONCE, which are not
2748 * expensive, to avoid any form of compiler optimizations:
2750 WRITE_ONCE(p->mm->numa_scan_seq, READ_ONCE(p->mm->numa_scan_seq) + 1);
2751 p->mm->numa_scan_offset = 0;
2755 * The expensive part of numa migration is done from task_work context.
2756 * Triggered from task_tick_numa().
2758 static void task_numa_work(struct callback_head *work)
2760 unsigned long migrate, next_scan, now = jiffies;
2761 struct task_struct *p = current;
2762 struct mm_struct *mm = p->mm;
2763 u64 runtime = p->se.sum_exec_runtime;
2764 struct vm_area_struct *vma;
2765 unsigned long start, end;
2766 unsigned long nr_pte_updates = 0;
2767 long pages, virtpages;
2769 SCHED_WARN_ON(p != container_of(work, struct task_struct, numa_work));
2773 * Who cares about NUMA placement when they're dying.
2775 * NOTE: make sure not to dereference p->mm before this check,
2776 * exit_task_work() happens _after_ exit_mm() so we could be called
2777 * without p->mm even though we still had it when we enqueued this
2780 if (p->flags & PF_EXITING)
2783 if (!mm->numa_next_scan) {
2784 mm->numa_next_scan = now +
2785 msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2789 * Enforce maximal scan/migration frequency..
2791 migrate = mm->numa_next_scan;
2792 if (time_before(now, migrate))
2795 if (p->numa_scan_period == 0) {
2796 p->numa_scan_period_max = task_scan_max(p);
2797 p->numa_scan_period = task_scan_start(p);
2800 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
2801 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
2805 * Delay this task enough that another task of this mm will likely win
2806 * the next time around.
2808 p->node_stamp += 2 * TICK_NSEC;
2810 start = mm->numa_scan_offset;
2811 pages = sysctl_numa_balancing_scan_size;
2812 pages <<= 20 - PAGE_SHIFT; /* MB in pages */
2813 virtpages = pages * 8; /* Scan up to this much virtual space */
2818 if (!mmap_read_trylock(mm))
2820 vma = find_vma(mm, start);
2822 reset_ptenuma_scan(p);
2826 for (; vma; vma = vma->vm_next) {
2827 if (!vma_migratable(vma) || !vma_policy_mof(vma) ||
2828 is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_MIXEDMAP)) {
2833 * Shared library pages mapped by multiple processes are not
2834 * migrated as it is expected they are cache replicated. Avoid
2835 * hinting faults in read-only file-backed mappings or the vdso
2836 * as migrating the pages will be of marginal benefit.
2839 (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ)))
2843 * Skip inaccessible VMAs to avoid any confusion between
2844 * PROT_NONE and NUMA hinting ptes
2846 if (!vma_is_accessible(vma))
2850 start = max(start, vma->vm_start);
2851 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
2852 end = min(end, vma->vm_end);
2853 nr_pte_updates = change_prot_numa(vma, start, end);
2856 * Try to scan sysctl_numa_balancing_size worth of
2857 * hpages that have at least one present PTE that
2858 * is not already pte-numa. If the VMA contains
2859 * areas that are unused or already full of prot_numa
2860 * PTEs, scan up to virtpages, to skip through those
2864 pages -= (end - start) >> PAGE_SHIFT;
2865 virtpages -= (end - start) >> PAGE_SHIFT;
2868 if (pages <= 0 || virtpages <= 0)
2872 } while (end != vma->vm_end);
2877 * It is possible to reach the end of the VMA list but the last few
2878 * VMAs are not guaranteed to the vma_migratable. If they are not, we
2879 * would find the !migratable VMA on the next scan but not reset the
2880 * scanner to the start so check it now.
2883 mm->numa_scan_offset = start;
2885 reset_ptenuma_scan(p);
2886 mmap_read_unlock(mm);
2889 * Make sure tasks use at least 32x as much time to run other code
2890 * than they used here, to limit NUMA PTE scanning overhead to 3% max.
2891 * Usually update_task_scan_period slows down scanning enough; on an
2892 * overloaded system we need to limit overhead on a per task basis.
2894 if (unlikely(p->se.sum_exec_runtime != runtime)) {
2895 u64 diff = p->se.sum_exec_runtime - runtime;
2896 p->node_stamp += 32 * diff;
2900 void init_numa_balancing(unsigned long clone_flags, struct task_struct *p)
2903 struct mm_struct *mm = p->mm;
2906 mm_users = atomic_read(&mm->mm_users);
2907 if (mm_users == 1) {
2908 mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2909 mm->numa_scan_seq = 0;
2913 p->numa_scan_seq = mm ? mm->numa_scan_seq : 0;
2914 p->numa_scan_period = sysctl_numa_balancing_scan_delay;
2915 p->numa_migrate_retry = 0;
2916 /* Protect against double add, see task_tick_numa and task_numa_work */
2917 p->numa_work.next = &p->numa_work;
2918 p->numa_faults = NULL;
2919 p->numa_pages_migrated = 0;
2920 p->total_numa_faults = 0;
2921 RCU_INIT_POINTER(p->numa_group, NULL);
2922 p->last_task_numa_placement = 0;
2923 p->last_sum_exec_runtime = 0;
2925 init_task_work(&p->numa_work, task_numa_work);
2927 /* New address space, reset the preferred nid */
2928 if (!(clone_flags & CLONE_VM)) {
2929 p->numa_preferred_nid = NUMA_NO_NODE;
2934 * New thread, keep existing numa_preferred_nid which should be copied
2935 * already by arch_dup_task_struct but stagger when scans start.
2940 delay = min_t(unsigned int, task_scan_max(current),
2941 current->numa_scan_period * mm_users * NSEC_PER_MSEC);
2942 delay += 2 * TICK_NSEC;
2943 p->node_stamp = delay;
2948 * Drive the periodic memory faults..
2950 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
2952 struct callback_head *work = &curr->numa_work;
2956 * We don't care about NUMA placement if we don't have memory.
2958 if (!curr->mm || (curr->flags & (PF_EXITING | PF_KTHREAD)) || work->next != work)
2962 * Using runtime rather than walltime has the dual advantage that
2963 * we (mostly) drive the selection from busy threads and that the
2964 * task needs to have done some actual work before we bother with
2967 now = curr->se.sum_exec_runtime;
2968 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
2970 if (now > curr->node_stamp + period) {
2971 if (!curr->node_stamp)
2972 curr->numa_scan_period = task_scan_start(curr);
2973 curr->node_stamp += period;
2975 if (!time_before(jiffies, curr->mm->numa_next_scan))
2976 task_work_add(curr, work, TWA_RESUME);
2980 static void update_scan_period(struct task_struct *p, int new_cpu)
2982 int src_nid = cpu_to_node(task_cpu(p));
2983 int dst_nid = cpu_to_node(new_cpu);
2985 if (!static_branch_likely(&sched_numa_balancing))
2988 if (!p->mm || !p->numa_faults || (p->flags & PF_EXITING))
2991 if (src_nid == dst_nid)
2995 * Allow resets if faults have been trapped before one scan
2996 * has completed. This is most likely due to a new task that
2997 * is pulled cross-node due to wakeups or load balancing.
2999 if (p->numa_scan_seq) {
3001 * Avoid scan adjustments if moving to the preferred
3002 * node or if the task was not previously running on
3003 * the preferred node.
3005 if (dst_nid == p->numa_preferred_nid ||
3006 (p->numa_preferred_nid != NUMA_NO_NODE &&
3007 src_nid != p->numa_preferred_nid))
3011 p->numa_scan_period = task_scan_start(p);
3015 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
3019 static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p)
3023 static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p)
3027 static inline void update_scan_period(struct task_struct *p, int new_cpu)
3031 #endif /* CONFIG_NUMA_BALANCING */
3034 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
3036 update_load_add(&cfs_rq->load, se->load.weight);
3038 if (entity_is_task(se)) {
3039 struct rq *rq = rq_of(cfs_rq);
3041 account_numa_enqueue(rq, task_of(se));
3042 list_add(&se->group_node, &rq->cfs_tasks);
3045 cfs_rq->nr_running++;
3047 cfs_rq->idle_nr_running++;
3051 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
3053 update_load_sub(&cfs_rq->load, se->load.weight);
3055 if (entity_is_task(se)) {
3056 account_numa_dequeue(rq_of(cfs_rq), task_of(se));
3057 list_del_init(&se->group_node);
3060 cfs_rq->nr_running--;
3062 cfs_rq->idle_nr_running--;
3066 * Signed add and clamp on underflow.
3068 * Explicitly do a load-store to ensure the intermediate value never hits
3069 * memory. This allows lockless observations without ever seeing the negative
3072 #define add_positive(_ptr, _val) do { \
3073 typeof(_ptr) ptr = (_ptr); \
3074 typeof(_val) val = (_val); \
3075 typeof(*ptr) res, var = READ_ONCE(*ptr); \
3079 if (val < 0 && res > var) \
3082 WRITE_ONCE(*ptr, res); \
3086 * Unsigned subtract and clamp on underflow.
3088 * Explicitly do a load-store to ensure the intermediate value never hits
3089 * memory. This allows lockless observations without ever seeing the negative
3092 #define sub_positive(_ptr, _val) do { \
3093 typeof(_ptr) ptr = (_ptr); \
3094 typeof(*ptr) val = (_val); \
3095 typeof(*ptr) res, var = READ_ONCE(*ptr); \
3099 WRITE_ONCE(*ptr, res); \
3103 * Remove and clamp on negative, from a local variable.
3105 * A variant of sub_positive(), which does not use explicit load-store
3106 * and is thus optimized for local variable updates.
3108 #define lsub_positive(_ptr, _val) do { \
3109 typeof(_ptr) ptr = (_ptr); \
3110 *ptr -= min_t(typeof(*ptr), *ptr, _val); \
3115 enqueue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3117 cfs_rq->avg.load_avg += se->avg.load_avg;
3118 cfs_rq->avg.load_sum += se_weight(se) * se->avg.load_sum;
3122 dequeue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3124 sub_positive(&cfs_rq->avg.load_avg, se->avg.load_avg);
3125 sub_positive(&cfs_rq->avg.load_sum, se_weight(se) * se->avg.load_sum);
3126 /* See update_cfs_rq_load_avg() */
3127 cfs_rq->avg.load_sum = max_t(u32, cfs_rq->avg.load_sum,
3128 cfs_rq->avg.load_avg * PELT_MIN_DIVIDER);
3132 enqueue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
3134 dequeue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
3137 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
3138 unsigned long weight)
3141 /* commit outstanding execution time */
3142 if (cfs_rq->curr == se)
3143 update_curr(cfs_rq);
3144 update_load_sub(&cfs_rq->load, se->load.weight);
3146 dequeue_load_avg(cfs_rq, se);
3148 update_load_set(&se->load, weight);
3152 u32 divider = get_pelt_divider(&se->avg);
3154 se->avg.load_avg = div_u64(se_weight(se) * se->avg.load_sum, divider);
3158 enqueue_load_avg(cfs_rq, se);
3160 update_load_add(&cfs_rq->load, se->load.weight);
3164 void reweight_task(struct task_struct *p, int prio)
3166 struct sched_entity *se = &p->se;
3167 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3168 struct load_weight *load = &se->load;
3169 unsigned long weight = scale_load(sched_prio_to_weight[prio]);
3171 reweight_entity(cfs_rq, se, weight);
3172 load->inv_weight = sched_prio_to_wmult[prio];
3175 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
3177 #ifdef CONFIG_FAIR_GROUP_SCHED
3180 * All this does is approximate the hierarchical proportion which includes that
3181 * global sum we all love to hate.
3183 * That is, the weight of a group entity, is the proportional share of the
3184 * group weight based on the group runqueue weights. That is:
3186 * tg->weight * grq->load.weight
3187 * ge->load.weight = ----------------------------- (1)
3188 * \Sum grq->load.weight
3190 * Now, because computing that sum is prohibitively expensive to compute (been
3191 * there, done that) we approximate it with this average stuff. The average
3192 * moves slower and therefore the approximation is cheaper and more stable.
3194 * So instead of the above, we substitute:
3196 * grq->load.weight -> grq->avg.load_avg (2)
3198 * which yields the following:
3200 * tg->weight * grq->avg.load_avg
3201 * ge->load.weight = ------------------------------ (3)
3204 * Where: tg->load_avg ~= \Sum grq->avg.load_avg
3206 * That is shares_avg, and it is right (given the approximation (2)).
3208 * The problem with it is that because the average is slow -- it was designed
3209 * to be exactly that of course -- this leads to transients in boundary
3210 * conditions. In specific, the case where the group was idle and we start the
3211 * one task. It takes time for our CPU's grq->avg.load_avg to build up,
3212 * yielding bad latency etc..
3214 * Now, in that special case (1) reduces to:
3216 * tg->weight * grq->load.weight
3217 * ge->load.weight = ----------------------------- = tg->weight (4)
3220 * That is, the sum collapses because all other CPUs are idle; the UP scenario.
3222 * So what we do is modify our approximation (3) to approach (4) in the (near)
3227 * tg->weight * grq->load.weight
3228 * --------------------------------------------------- (5)
3229 * tg->load_avg - grq->avg.load_avg + grq->load.weight
3231 * But because grq->load.weight can drop to 0, resulting in a divide by zero,
3232 * we need to use grq->avg.load_avg as its lower bound, which then gives:
3235 * tg->weight * grq->load.weight
3236 * ge->load.weight = ----------------------------- (6)
3241 * tg_load_avg' = tg->load_avg - grq->avg.load_avg +
3242 * max(grq->load.weight, grq->avg.load_avg)
3244 * And that is shares_weight and is icky. In the (near) UP case it approaches
3245 * (4) while in the normal case it approaches (3). It consistently
3246 * overestimates the ge->load.weight and therefore:
3248 * \Sum ge->load.weight >= tg->weight
3252 static long calc_group_shares(struct cfs_rq *cfs_rq)
3254 long tg_weight, tg_shares, load, shares;
3255 struct task_group *tg = cfs_rq->tg;
3257 tg_shares = READ_ONCE(tg->shares);
3259 load = max(scale_load_down(cfs_rq->load.weight), cfs_rq->avg.load_avg);
3261 tg_weight = atomic_long_read(&tg->load_avg);
3263 /* Ensure tg_weight >= load */
3264 tg_weight -= cfs_rq->tg_load_avg_contrib;
3267 shares = (tg_shares * load);
3269 shares /= tg_weight;
3272 * MIN_SHARES has to be unscaled here to support per-CPU partitioning
3273 * of a group with small tg->shares value. It is a floor value which is
3274 * assigned as a minimum load.weight to the sched_entity representing
3275 * the group on a CPU.
3277 * E.g. on 64-bit for a group with tg->shares of scale_load(15)=15*1024
3278 * on an 8-core system with 8 tasks each runnable on one CPU shares has
3279 * to be 15*1024*1/8=1920 instead of scale_load(MIN_SHARES)=2*1024. In
3280 * case no task is runnable on a CPU MIN_SHARES=2 should be returned
3283 return clamp_t(long, shares, MIN_SHARES, tg_shares);
3285 #endif /* CONFIG_SMP */
3288 * Recomputes the group entity based on the current state of its group
3291 static void update_cfs_group(struct sched_entity *se)
3293 struct cfs_rq *gcfs_rq = group_cfs_rq(se);
3299 if (throttled_hierarchy(gcfs_rq))
3303 shares = READ_ONCE(gcfs_rq->tg->shares);
3305 if (likely(se->load.weight == shares))
3308 shares = calc_group_shares(gcfs_rq);
3311 reweight_entity(cfs_rq_of(se), se, shares);
3314 #else /* CONFIG_FAIR_GROUP_SCHED */
3315 static inline void update_cfs_group(struct sched_entity *se)
3318 #endif /* CONFIG_FAIR_GROUP_SCHED */
3320 static inline void cfs_rq_util_change(struct cfs_rq *cfs_rq, int flags)
3322 struct rq *rq = rq_of(cfs_rq);
3324 if (&rq->cfs == cfs_rq) {
3326 * There are a few boundary cases this might miss but it should
3327 * get called often enough that that should (hopefully) not be
3330 * It will not get called when we go idle, because the idle
3331 * thread is a different class (!fair), nor will the utilization
3332 * number include things like RT tasks.
3334 * As is, the util number is not freq-invariant (we'd have to
3335 * implement arch_scale_freq_capacity() for that).
3337 * See cpu_util_cfs().
3339 cpufreq_update_util(rq, flags);
3344 static inline bool load_avg_is_decayed(struct sched_avg *sa)
3352 if (sa->runnable_sum)
3356 * _avg must be null when _sum are null because _avg = _sum / divider
3357 * Make sure that rounding and/or propagation of PELT values never
3360 SCHED_WARN_ON(sa->load_avg ||
3367 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
3369 return u64_u32_load_copy(cfs_rq->avg.last_update_time,
3370 cfs_rq->last_update_time_copy);
3372 #ifdef CONFIG_FAIR_GROUP_SCHED
3374 * Because list_add_leaf_cfs_rq always places a child cfs_rq on the list
3375 * immediately before a parent cfs_rq, and cfs_rqs are removed from the list
3376 * bottom-up, we only have to test whether the cfs_rq before us on the list
3378 * If cfs_rq is not on the list, test whether a child needs its to be added to
3379 * connect a branch to the tree * (see list_add_leaf_cfs_rq() for details).
3381 static inline bool child_cfs_rq_on_list(struct cfs_rq *cfs_rq)
3383 struct cfs_rq *prev_cfs_rq;
3384 struct list_head *prev;
3386 if (cfs_rq->on_list) {
3387 prev = cfs_rq->leaf_cfs_rq_list.prev;
3389 struct rq *rq = rq_of(cfs_rq);
3391 prev = rq->tmp_alone_branch;
3394 prev_cfs_rq = container_of(prev, struct cfs_rq, leaf_cfs_rq_list);
3396 return (prev_cfs_rq->tg->parent == cfs_rq->tg);
3399 static inline bool cfs_rq_is_decayed(struct cfs_rq *cfs_rq)
3401 if (cfs_rq->load.weight)
3404 if (!load_avg_is_decayed(&cfs_rq->avg))
3407 if (child_cfs_rq_on_list(cfs_rq))
3414 * update_tg_load_avg - update the tg's load avg
3415 * @cfs_rq: the cfs_rq whose avg changed
3417 * This function 'ensures': tg->load_avg := \Sum tg->cfs_rq[]->avg.load.
3418 * However, because tg->load_avg is a global value there are performance
3421 * In order to avoid having to look at the other cfs_rq's, we use a
3422 * differential update where we store the last value we propagated. This in
3423 * turn allows skipping updates if the differential is 'small'.
3425 * Updating tg's load_avg is necessary before update_cfs_share().
3427 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq)
3429 long delta = cfs_rq->avg.load_avg - cfs_rq->tg_load_avg_contrib;
3432 * No need to update load_avg for root_task_group as it is not used.
3434 if (cfs_rq->tg == &root_task_group)
3437 if (abs(delta) > cfs_rq->tg_load_avg_contrib / 64) {
3438 atomic_long_add(delta, &cfs_rq->tg->load_avg);
3439 cfs_rq->tg_load_avg_contrib = cfs_rq->avg.load_avg;
3444 * Called within set_task_rq() right before setting a task's CPU. The
3445 * caller only guarantees p->pi_lock is held; no other assumptions,
3446 * including the state of rq->lock, should be made.
3448 void set_task_rq_fair(struct sched_entity *se,
3449 struct cfs_rq *prev, struct cfs_rq *next)
3451 u64 p_last_update_time;
3452 u64 n_last_update_time;
3454 if (!sched_feat(ATTACH_AGE_LOAD))
3458 * We are supposed to update the task to "current" time, then its up to
3459 * date and ready to go to new CPU/cfs_rq. But we have difficulty in
3460 * getting what current time is, so simply throw away the out-of-date
3461 * time. This will result in the wakee task is less decayed, but giving
3462 * the wakee more load sounds not bad.
3464 if (!(se->avg.last_update_time && prev))
3467 p_last_update_time = cfs_rq_last_update_time(prev);
3468 n_last_update_time = cfs_rq_last_update_time(next);
3470 __update_load_avg_blocked_se(p_last_update_time, se);
3471 se->avg.last_update_time = n_last_update_time;
3475 * When on migration a sched_entity joins/leaves the PELT hierarchy, we need to
3476 * propagate its contribution. The key to this propagation is the invariant
3477 * that for each group:
3479 * ge->avg == grq->avg (1)
3481 * _IFF_ we look at the pure running and runnable sums. Because they
3482 * represent the very same entity, just at different points in the hierarchy.
3484 * Per the above update_tg_cfs_util() and update_tg_cfs_runnable() are trivial
3485 * and simply copies the running/runnable sum over (but still wrong, because
3486 * the group entity and group rq do not have their PELT windows aligned).
3488 * However, update_tg_cfs_load() is more complex. So we have:
3490 * ge->avg.load_avg = ge->load.weight * ge->avg.runnable_avg (2)
3492 * And since, like util, the runnable part should be directly transferable,
3493 * the following would _appear_ to be the straight forward approach:
3495 * grq->avg.load_avg = grq->load.weight * grq->avg.runnable_avg (3)
3497 * And per (1) we have:
3499 * ge->avg.runnable_avg == grq->avg.runnable_avg
3503 * ge->load.weight * grq->avg.load_avg
3504 * ge->avg.load_avg = ----------------------------------- (4)
3507 * Except that is wrong!
3509 * Because while for entities historical weight is not important and we
3510 * really only care about our future and therefore can consider a pure
3511 * runnable sum, runqueues can NOT do this.
3513 * We specifically want runqueues to have a load_avg that includes
3514 * historical weights. Those represent the blocked load, the load we expect
3515 * to (shortly) return to us. This only works by keeping the weights as
3516 * integral part of the sum. We therefore cannot decompose as per (3).
3518 * Another reason this doesn't work is that runnable isn't a 0-sum entity.
3519 * Imagine a rq with 2 tasks that each are runnable 2/3 of the time. Then the
3520 * rq itself is runnable anywhere between 2/3 and 1 depending on how the
3521 * runnable section of these tasks overlap (or not). If they were to perfectly
3522 * align the rq as a whole would be runnable 2/3 of the time. If however we
3523 * always have at least 1 runnable task, the rq as a whole is always runnable.
3525 * So we'll have to approximate.. :/
3527 * Given the constraint:
3529 * ge->avg.running_sum <= ge->avg.runnable_sum <= LOAD_AVG_MAX
3531 * We can construct a rule that adds runnable to a rq by assuming minimal
3534 * On removal, we'll assume each task is equally runnable; which yields:
3536 * grq->avg.runnable_sum = grq->avg.load_sum / grq->load.weight
3538 * XXX: only do this for the part of runnable > running ?
3542 update_tg_cfs_util(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq)
3544 long delta_sum, delta_avg = gcfs_rq->avg.util_avg - se->avg.util_avg;
3545 u32 new_sum, divider;
3547 /* Nothing to update */
3552 * cfs_rq->avg.period_contrib can be used for both cfs_rq and se.
3553 * See ___update_load_avg() for details.
3555 divider = get_pelt_divider(&cfs_rq->avg);
3558 /* Set new sched_entity's utilization */
3559 se->avg.util_avg = gcfs_rq->avg.util_avg;
3560 new_sum = se->avg.util_avg * divider;
3561 delta_sum = (long)new_sum - (long)se->avg.util_sum;
3562 se->avg.util_sum = new_sum;
3564 /* Update parent cfs_rq utilization */
3565 add_positive(&cfs_rq->avg.util_avg, delta_avg);
3566 add_positive(&cfs_rq->avg.util_sum, delta_sum);
3568 /* See update_cfs_rq_load_avg() */
3569 cfs_rq->avg.util_sum = max_t(u32, cfs_rq->avg.util_sum,
3570 cfs_rq->avg.util_avg * PELT_MIN_DIVIDER);
3574 update_tg_cfs_runnable(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq)
3576 long delta_sum, delta_avg = gcfs_rq->avg.runnable_avg - se->avg.runnable_avg;
3577 u32 new_sum, divider;
3579 /* Nothing to update */
3584 * cfs_rq->avg.period_contrib can be used for both cfs_rq and se.
3585 * See ___update_load_avg() for details.
3587 divider = get_pelt_divider(&cfs_rq->avg);
3589 /* Set new sched_entity's runnable */
3590 se->avg.runnable_avg = gcfs_rq->avg.runnable_avg;
3591 new_sum = se->avg.runnable_avg * divider;
3592 delta_sum = (long)new_sum - (long)se->avg.runnable_sum;
3593 se->avg.runnable_sum = new_sum;
3595 /* Update parent cfs_rq runnable */
3596 add_positive(&cfs_rq->avg.runnable_avg, delta_avg);
3597 add_positive(&cfs_rq->avg.runnable_sum, delta_sum);
3598 /* See update_cfs_rq_load_avg() */
3599 cfs_rq->avg.runnable_sum = max_t(u32, cfs_rq->avg.runnable_sum,
3600 cfs_rq->avg.runnable_avg * PELT_MIN_DIVIDER);
3604 update_tg_cfs_load(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq)
3606 long delta_avg, running_sum, runnable_sum = gcfs_rq->prop_runnable_sum;
3607 unsigned long load_avg;
3615 gcfs_rq->prop_runnable_sum = 0;
3618 * cfs_rq->avg.period_contrib can be used for both cfs_rq and se.
3619 * See ___update_load_avg() for details.
3621 divider = get_pelt_divider(&cfs_rq->avg);
3623 if (runnable_sum >= 0) {
3625 * Add runnable; clip at LOAD_AVG_MAX. Reflects that until
3626 * the CPU is saturated running == runnable.
3628 runnable_sum += se->avg.load_sum;
3629 runnable_sum = min_t(long, runnable_sum, divider);
3632 * Estimate the new unweighted runnable_sum of the gcfs_rq by
3633 * assuming all tasks are equally runnable.
3635 if (scale_load_down(gcfs_rq->load.weight)) {
3636 load_sum = div_u64(gcfs_rq->avg.load_sum,
3637 scale_load_down(gcfs_rq->load.weight));
3640 /* But make sure to not inflate se's runnable */
3641 runnable_sum = min(se->avg.load_sum, load_sum);
3645 * runnable_sum can't be lower than running_sum
3646 * Rescale running sum to be in the same range as runnable sum
3647 * running_sum is in [0 : LOAD_AVG_MAX << SCHED_CAPACITY_SHIFT]
3648 * runnable_sum is in [0 : LOAD_AVG_MAX]
3650 running_sum = se->avg.util_sum >> SCHED_CAPACITY_SHIFT;
3651 runnable_sum = max(runnable_sum, running_sum);
3653 load_sum = se_weight(se) * runnable_sum;
3654 load_avg = div_u64(load_sum, divider);
3656 delta_avg = load_avg - se->avg.load_avg;
3660 delta_sum = load_sum - (s64)se_weight(se) * se->avg.load_sum;
3662 se->avg.load_sum = runnable_sum;
3663 se->avg.load_avg = load_avg;
3664 add_positive(&cfs_rq->avg.load_avg, delta_avg);
3665 add_positive(&cfs_rq->avg.load_sum, delta_sum);
3666 /* See update_cfs_rq_load_avg() */
3667 cfs_rq->avg.load_sum = max_t(u32, cfs_rq->avg.load_sum,
3668 cfs_rq->avg.load_avg * PELT_MIN_DIVIDER);
3671 static inline void add_tg_cfs_propagate(struct cfs_rq *cfs_rq, long runnable_sum)
3673 cfs_rq->propagate = 1;
3674 cfs_rq->prop_runnable_sum += runnable_sum;
3677 /* Update task and its cfs_rq load average */
3678 static inline int propagate_entity_load_avg(struct sched_entity *se)
3680 struct cfs_rq *cfs_rq, *gcfs_rq;
3682 if (entity_is_task(se))
3685 gcfs_rq = group_cfs_rq(se);
3686 if (!gcfs_rq->propagate)
3689 gcfs_rq->propagate = 0;
3691 cfs_rq = cfs_rq_of(se);
3693 add_tg_cfs_propagate(cfs_rq, gcfs_rq->prop_runnable_sum);
3695 update_tg_cfs_util(cfs_rq, se, gcfs_rq);
3696 update_tg_cfs_runnable(cfs_rq, se, gcfs_rq);
3697 update_tg_cfs_load(cfs_rq, se, gcfs_rq);
3699 trace_pelt_cfs_tp(cfs_rq);
3700 trace_pelt_se_tp(se);
3706 * Check if we need to update the load and the utilization of a blocked
3709 static inline bool skip_blocked_update(struct sched_entity *se)
3711 struct cfs_rq *gcfs_rq = group_cfs_rq(se);
3714 * If sched_entity still have not zero load or utilization, we have to
3717 if (se->avg.load_avg || se->avg.util_avg)
3721 * If there is a pending propagation, we have to update the load and
3722 * the utilization of the sched_entity:
3724 if (gcfs_rq->propagate)
3728 * Otherwise, the load and the utilization of the sched_entity is
3729 * already zero and there is no pending propagation, so it will be a
3730 * waste of time to try to decay it:
3735 #else /* CONFIG_FAIR_GROUP_SCHED */
3737 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq) {}
3739 static inline int propagate_entity_load_avg(struct sched_entity *se)
3744 static inline void add_tg_cfs_propagate(struct cfs_rq *cfs_rq, long runnable_sum) {}
3746 #endif /* CONFIG_FAIR_GROUP_SCHED */
3748 #ifdef CONFIG_NO_HZ_COMMON
3749 static inline void migrate_se_pelt_lag(struct sched_entity *se)
3751 u64 throttled = 0, now, lut;
3752 struct cfs_rq *cfs_rq;
3756 if (load_avg_is_decayed(&se->avg))
3759 cfs_rq = cfs_rq_of(se);
3763 is_idle = is_idle_task(rcu_dereference(rq->curr));
3767 * The lag estimation comes with a cost we don't want to pay all the
3768 * time. Hence, limiting to the case where the source CPU is idle and
3769 * we know we are at the greatest risk to have an outdated clock.
3775 * Estimated "now" is: last_update_time + cfs_idle_lag + rq_idle_lag, where:
3777 * last_update_time (the cfs_rq's last_update_time)
3778 * = cfs_rq_clock_pelt()@cfs_rq_idle
3779 * = rq_clock_pelt()@cfs_rq_idle
3780 * - cfs->throttled_clock_pelt_time@cfs_rq_idle
3782 * cfs_idle_lag (delta between rq's update and cfs_rq's update)
3783 * = rq_clock_pelt()@rq_idle - rq_clock_pelt()@cfs_rq_idle
3785 * rq_idle_lag (delta between now and rq's update)
3786 * = sched_clock_cpu() - rq_clock()@rq_idle
3788 * We can then write:
3790 * now = rq_clock_pelt()@rq_idle - cfs->throttled_clock_pelt_time +
3791 * sched_clock_cpu() - rq_clock()@rq_idle
3793 * rq_clock_pelt()@rq_idle is rq->clock_pelt_idle
3794 * rq_clock()@rq_idle is rq->clock_idle
3795 * cfs->throttled_clock_pelt_time@cfs_rq_idle
3796 * is cfs_rq->throttled_pelt_idle
3799 #ifdef CONFIG_CFS_BANDWIDTH
3800 throttled = u64_u32_load(cfs_rq->throttled_pelt_idle);
3801 /* The clock has been stopped for throttling */
3802 if (throttled == U64_MAX)
3805 now = u64_u32_load(rq->clock_pelt_idle);
3807 * Paired with _update_idle_rq_clock_pelt(). It ensures at the worst case
3808 * is observed the old clock_pelt_idle value and the new clock_idle,
3809 * which lead to an underestimation. The opposite would lead to an
3813 lut = cfs_rq_last_update_time(cfs_rq);
3818 * cfs_rq->avg.last_update_time is more recent than our
3819 * estimation, let's use it.
3823 now += sched_clock_cpu(cpu_of(rq)) - u64_u32_load(rq->clock_idle);
3825 __update_load_avg_blocked_se(now, se);
3828 static void migrate_se_pelt_lag(struct sched_entity *se) {}
3832 * update_cfs_rq_load_avg - update the cfs_rq's load/util averages
3833 * @now: current time, as per cfs_rq_clock_pelt()
3834 * @cfs_rq: cfs_rq to update
3836 * The cfs_rq avg is the direct sum of all its entities (blocked and runnable)
3837 * avg. The immediate corollary is that all (fair) tasks must be attached.
3839 * cfs_rq->avg is used for task_h_load() and update_cfs_share() for example.
3841 * Return: true if the load decayed or we removed load.
3843 * Since both these conditions indicate a changed cfs_rq->avg.load we should
3844 * call update_tg_load_avg() when this function returns true.
3847 update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq)
3849 unsigned long removed_load = 0, removed_util = 0, removed_runnable = 0;
3850 struct sched_avg *sa = &cfs_rq->avg;
3853 if (cfs_rq->removed.nr) {
3855 u32 divider = get_pelt_divider(&cfs_rq->avg);
3857 raw_spin_lock(&cfs_rq->removed.lock);
3858 swap(cfs_rq->removed.util_avg, removed_util);
3859 swap(cfs_rq->removed.load_avg, removed_load);
3860 swap(cfs_rq->removed.runnable_avg, removed_runnable);
3861 cfs_rq->removed.nr = 0;
3862 raw_spin_unlock(&cfs_rq->removed.lock);
3865 sub_positive(&sa->load_avg, r);
3866 sub_positive(&sa->load_sum, r * divider);
3867 /* See sa->util_sum below */
3868 sa->load_sum = max_t(u32, sa->load_sum, sa->load_avg * PELT_MIN_DIVIDER);
3871 sub_positive(&sa->util_avg, r);
3872 sub_positive(&sa->util_sum, r * divider);
3874 * Because of rounding, se->util_sum might ends up being +1 more than
3875 * cfs->util_sum. Although this is not a problem by itself, detaching
3876 * a lot of tasks with the rounding problem between 2 updates of
3877 * util_avg (~1ms) can make cfs->util_sum becoming null whereas
3878 * cfs_util_avg is not.
3879 * Check that util_sum is still above its lower bound for the new
3880 * util_avg. Given that period_contrib might have moved since the last
3881 * sync, we are only sure that util_sum must be above or equal to
3882 * util_avg * minimum possible divider
3884 sa->util_sum = max_t(u32, sa->util_sum, sa->util_avg * PELT_MIN_DIVIDER);
3886 r = removed_runnable;
3887 sub_positive(&sa->runnable_avg, r);
3888 sub_positive(&sa->runnable_sum, r * divider);
3889 /* See sa->util_sum above */
3890 sa->runnable_sum = max_t(u32, sa->runnable_sum,
3891 sa->runnable_avg * PELT_MIN_DIVIDER);
3894 * removed_runnable is the unweighted version of removed_load so we
3895 * can use it to estimate removed_load_sum.
3897 add_tg_cfs_propagate(cfs_rq,
3898 -(long)(removed_runnable * divider) >> SCHED_CAPACITY_SHIFT);
3903 decayed |= __update_load_avg_cfs_rq(now, cfs_rq);
3904 u64_u32_store_copy(sa->last_update_time,
3905 cfs_rq->last_update_time_copy,
3906 sa->last_update_time);
3911 * attach_entity_load_avg - attach this entity to its cfs_rq load avg
3912 * @cfs_rq: cfs_rq to attach to
3913 * @se: sched_entity to attach
3915 * Must call update_cfs_rq_load_avg() before this, since we rely on
3916 * cfs_rq->avg.last_update_time being current.
3918 static void attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3921 * cfs_rq->avg.period_contrib can be used for both cfs_rq and se.
3922 * See ___update_load_avg() for details.
3924 u32 divider = get_pelt_divider(&cfs_rq->avg);
3927 * When we attach the @se to the @cfs_rq, we must align the decay
3928 * window because without that, really weird and wonderful things can
3933 se->avg.last_update_time = cfs_rq->avg.last_update_time;
3934 se->avg.period_contrib = cfs_rq->avg.period_contrib;
3937 * Hell(o) Nasty stuff.. we need to recompute _sum based on the new
3938 * period_contrib. This isn't strictly correct, but since we're
3939 * entirely outside of the PELT hierarchy, nobody cares if we truncate
3942 se->avg.util_sum = se->avg.util_avg * divider;
3944 se->avg.runnable_sum = se->avg.runnable_avg * divider;
3946 se->avg.load_sum = se->avg.load_avg * divider;
3947 if (se_weight(se) < se->avg.load_sum)
3948 se->avg.load_sum = div_u64(se->avg.load_sum, se_weight(se));
3950 se->avg.load_sum = 1;
3952 enqueue_load_avg(cfs_rq, se);
3953 cfs_rq->avg.util_avg += se->avg.util_avg;
3954 cfs_rq->avg.util_sum += se->avg.util_sum;
3955 cfs_rq->avg.runnable_avg += se->avg.runnable_avg;
3956 cfs_rq->avg.runnable_sum += se->avg.runnable_sum;
3958 add_tg_cfs_propagate(cfs_rq, se->avg.load_sum);
3960 cfs_rq_util_change(cfs_rq, 0);
3962 trace_pelt_cfs_tp(cfs_rq);
3966 * detach_entity_load_avg - detach this entity from its cfs_rq load avg
3967 * @cfs_rq: cfs_rq to detach from
3968 * @se: sched_entity to detach
3970 * Must call update_cfs_rq_load_avg() before this, since we rely on
3971 * cfs_rq->avg.last_update_time being current.
3973 static void detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3975 dequeue_load_avg(cfs_rq, se);
3976 sub_positive(&cfs_rq->avg.util_avg, se->avg.util_avg);
3977 sub_positive(&cfs_rq->avg.util_sum, se->avg.util_sum);
3978 /* See update_cfs_rq_load_avg() */
3979 cfs_rq->avg.util_sum = max_t(u32, cfs_rq->avg.util_sum,
3980 cfs_rq->avg.util_avg * PELT_MIN_DIVIDER);
3982 sub_positive(&cfs_rq->avg.runnable_avg, se->avg.runnable_avg);
3983 sub_positive(&cfs_rq->avg.runnable_sum, se->avg.runnable_sum);
3984 /* See update_cfs_rq_load_avg() */
3985 cfs_rq->avg.runnable_sum = max_t(u32, cfs_rq->avg.runnable_sum,
3986 cfs_rq->avg.runnable_avg * PELT_MIN_DIVIDER);
3988 add_tg_cfs_propagate(cfs_rq, -se->avg.load_sum);
3990 cfs_rq_util_change(cfs_rq, 0);
3992 trace_pelt_cfs_tp(cfs_rq);
3996 * Optional action to be done while updating the load average
3998 #define UPDATE_TG 0x1
3999 #define SKIP_AGE_LOAD 0x2
4000 #define DO_ATTACH 0x4
4001 #define DO_DETACH 0x8
4003 /* Update task and its cfs_rq load average */
4004 static inline void update_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
4006 u64 now = cfs_rq_clock_pelt(cfs_rq);
4010 * Track task load average for carrying it to new CPU after migrated, and
4011 * track group sched_entity load average for task_h_load calc in migration
4013 if (se->avg.last_update_time && !(flags & SKIP_AGE_LOAD))
4014 __update_load_avg_se(now, cfs_rq, se);
4016 decayed = update_cfs_rq_load_avg(now, cfs_rq);
4017 decayed |= propagate_entity_load_avg(se);
4019 if (!se->avg.last_update_time && (flags & DO_ATTACH)) {
4022 * DO_ATTACH means we're here from enqueue_entity().
4023 * !last_update_time means we've passed through
4024 * migrate_task_rq_fair() indicating we migrated.
4026 * IOW we're enqueueing a task on a new CPU.
4028 attach_entity_load_avg(cfs_rq, se);
4029 update_tg_load_avg(cfs_rq);
4031 } else if (flags & DO_DETACH) {
4033 * DO_DETACH means we're here from dequeue_entity()
4034 * and we are migrating task out of the CPU.
4036 detach_entity_load_avg(cfs_rq, se);
4037 update_tg_load_avg(cfs_rq);
4038 } else if (decayed) {
4039 cfs_rq_util_change(cfs_rq, 0);
4041 if (flags & UPDATE_TG)
4042 update_tg_load_avg(cfs_rq);
4047 * Synchronize entity load avg of dequeued entity without locking
4050 static void sync_entity_load_avg(struct sched_entity *se)
4052 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4053 u64 last_update_time;
4055 last_update_time = cfs_rq_last_update_time(cfs_rq);
4056 __update_load_avg_blocked_se(last_update_time, se);
4060 * Task first catches up with cfs_rq, and then subtract
4061 * itself from the cfs_rq (task must be off the queue now).
4063 static void remove_entity_load_avg(struct sched_entity *se)
4065 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4066 unsigned long flags;
4069 * tasks cannot exit without having gone through wake_up_new_task() ->
4070 * enqueue_task_fair() which will have added things to the cfs_rq,
4071 * so we can remove unconditionally.
4074 sync_entity_load_avg(se);
4076 raw_spin_lock_irqsave(&cfs_rq->removed.lock, flags);
4077 ++cfs_rq->removed.nr;
4078 cfs_rq->removed.util_avg += se->avg.util_avg;
4079 cfs_rq->removed.load_avg += se->avg.load_avg;
4080 cfs_rq->removed.runnable_avg += se->avg.runnable_avg;
4081 raw_spin_unlock_irqrestore(&cfs_rq->removed.lock, flags);
4084 static inline unsigned long cfs_rq_runnable_avg(struct cfs_rq *cfs_rq)
4086 return cfs_rq->avg.runnable_avg;
4089 static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq)
4091 return cfs_rq->avg.load_avg;
4094 static int newidle_balance(struct rq *this_rq, struct rq_flags *rf);
4096 static inline unsigned long task_util(struct task_struct *p)
4098 return READ_ONCE(p->se.avg.util_avg);
4101 static inline unsigned long _task_util_est(struct task_struct *p)
4103 struct util_est ue = READ_ONCE(p->se.avg.util_est);
4105 return max(ue.ewma, (ue.enqueued & ~UTIL_AVG_UNCHANGED));
4108 static inline unsigned long task_util_est(struct task_struct *p)
4110 return max(task_util(p), _task_util_est(p));
4113 #ifdef CONFIG_UCLAMP_TASK
4114 static inline unsigned long uclamp_task_util(struct task_struct *p)
4116 return clamp(task_util_est(p),
4117 uclamp_eff_value(p, UCLAMP_MIN),
4118 uclamp_eff_value(p, UCLAMP_MAX));
4121 static inline unsigned long uclamp_task_util(struct task_struct *p)
4123 return task_util_est(p);
4127 static inline void util_est_enqueue(struct cfs_rq *cfs_rq,
4128 struct task_struct *p)
4130 unsigned int enqueued;
4132 if (!sched_feat(UTIL_EST))
4135 /* Update root cfs_rq's estimated utilization */
4136 enqueued = cfs_rq->avg.util_est.enqueued;
4137 enqueued += _task_util_est(p);
4138 WRITE_ONCE(cfs_rq->avg.util_est.enqueued, enqueued);
4140 trace_sched_util_est_cfs_tp(cfs_rq);
4143 static inline void util_est_dequeue(struct cfs_rq *cfs_rq,
4144 struct task_struct *p)
4146 unsigned int enqueued;
4148 if (!sched_feat(UTIL_EST))
4151 /* Update root cfs_rq's estimated utilization */
4152 enqueued = cfs_rq->avg.util_est.enqueued;
4153 enqueued -= min_t(unsigned int, enqueued, _task_util_est(p));
4154 WRITE_ONCE(cfs_rq->avg.util_est.enqueued, enqueued);
4156 trace_sched_util_est_cfs_tp(cfs_rq);
4159 #define UTIL_EST_MARGIN (SCHED_CAPACITY_SCALE / 100)
4162 * Check if a (signed) value is within a specified (unsigned) margin,
4163 * based on the observation that:
4165 * abs(x) < y := (unsigned)(x + y - 1) < (2 * y - 1)
4167 * NOTE: this only works when value + margin < INT_MAX.
4169 static inline bool within_margin(int value, int margin)
4171 return ((unsigned int)(value + margin - 1) < (2 * margin - 1));
4174 static inline void util_est_update(struct cfs_rq *cfs_rq,
4175 struct task_struct *p,
4178 long last_ewma_diff, last_enqueued_diff;
4181 if (!sched_feat(UTIL_EST))
4185 * Skip update of task's estimated utilization when the task has not
4186 * yet completed an activation, e.g. being migrated.
4192 * If the PELT values haven't changed since enqueue time,
4193 * skip the util_est update.
4195 ue = p->se.avg.util_est;
4196 if (ue.enqueued & UTIL_AVG_UNCHANGED)
4199 last_enqueued_diff = ue.enqueued;
4202 * Reset EWMA on utilization increases, the moving average is used only
4203 * to smooth utilization decreases.
4205 ue.enqueued = task_util(p);
4206 if (sched_feat(UTIL_EST_FASTUP)) {
4207 if (ue.ewma < ue.enqueued) {
4208 ue.ewma = ue.enqueued;
4214 * Skip update of task's estimated utilization when its members are
4215 * already ~1% close to its last activation value.
4217 last_ewma_diff = ue.enqueued - ue.ewma;
4218 last_enqueued_diff -= ue.enqueued;
4219 if (within_margin(last_ewma_diff, UTIL_EST_MARGIN)) {
4220 if (!within_margin(last_enqueued_diff, UTIL_EST_MARGIN))
4227 * To avoid overestimation of actual task utilization, skip updates if
4228 * we cannot grant there is idle time in this CPU.
4230 if (task_util(p) > capacity_orig_of(cpu_of(rq_of(cfs_rq))))
4234 * Update Task's estimated utilization
4236 * When *p completes an activation we can consolidate another sample
4237 * of the task size. This is done by storing the current PELT value
4238 * as ue.enqueued and by using this value to update the Exponential
4239 * Weighted Moving Average (EWMA):
4241 * ewma(t) = w * task_util(p) + (1-w) * ewma(t-1)
4242 * = w * task_util(p) + ewma(t-1) - w * ewma(t-1)
4243 * = w * (task_util(p) - ewma(t-1)) + ewma(t-1)
4244 * = w * ( last_ewma_diff ) + ewma(t-1)
4245 * = w * (last_ewma_diff + ewma(t-1) / w)
4247 * Where 'w' is the weight of new samples, which is configured to be
4248 * 0.25, thus making w=1/4 ( >>= UTIL_EST_WEIGHT_SHIFT)
4250 ue.ewma <<= UTIL_EST_WEIGHT_SHIFT;
4251 ue.ewma += last_ewma_diff;
4252 ue.ewma >>= UTIL_EST_WEIGHT_SHIFT;
4254 ue.enqueued |= UTIL_AVG_UNCHANGED;
4255 WRITE_ONCE(p->se.avg.util_est, ue);
4257 trace_sched_util_est_se_tp(&p->se);
4260 static inline int task_fits_capacity(struct task_struct *p,
4261 unsigned long capacity)
4263 return fits_capacity(uclamp_task_util(p), capacity);
4266 static inline void update_misfit_status(struct task_struct *p, struct rq *rq)
4268 if (!sched_asym_cpucap_active())
4271 if (!p || p->nr_cpus_allowed == 1) {
4272 rq->misfit_task_load = 0;
4276 if (task_fits_capacity(p, capacity_of(cpu_of(rq)))) {
4277 rq->misfit_task_load = 0;
4282 * Make sure that misfit_task_load will not be null even if
4283 * task_h_load() returns 0.
4285 rq->misfit_task_load = max_t(unsigned long, task_h_load(p), 1);
4288 #else /* CONFIG_SMP */
4290 static inline bool cfs_rq_is_decayed(struct cfs_rq *cfs_rq)
4295 #define UPDATE_TG 0x0
4296 #define SKIP_AGE_LOAD 0x0
4297 #define DO_ATTACH 0x0
4298 #define DO_DETACH 0x0
4300 static inline void update_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int not_used1)
4302 cfs_rq_util_change(cfs_rq, 0);
4305 static inline void remove_entity_load_avg(struct sched_entity *se) {}
4308 attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
4310 detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
4312 static inline int newidle_balance(struct rq *rq, struct rq_flags *rf)
4318 util_est_enqueue(struct cfs_rq *cfs_rq, struct task_struct *p) {}
4321 util_est_dequeue(struct cfs_rq *cfs_rq, struct task_struct *p) {}
4324 util_est_update(struct cfs_rq *cfs_rq, struct task_struct *p,
4326 static inline void update_misfit_status(struct task_struct *p, struct rq *rq) {}
4328 #endif /* CONFIG_SMP */
4330 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
4332 #ifdef CONFIG_SCHED_DEBUG
4333 s64 d = se->vruntime - cfs_rq->min_vruntime;
4338 if (d > 3*sysctl_sched_latency)
4339 schedstat_inc(cfs_rq->nr_spread_over);
4344 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
4346 u64 vruntime = cfs_rq->min_vruntime;
4349 * The 'current' period is already promised to the current tasks,
4350 * however the extra weight of the new task will slow them down a
4351 * little, place the new task so that it fits in the slot that
4352 * stays open at the end.
4354 if (initial && sched_feat(START_DEBIT))
4355 vruntime += sched_vslice(cfs_rq, se);
4357 /* sleeps up to a single latency don't count. */
4359 unsigned long thresh;
4362 thresh = sysctl_sched_min_granularity;
4364 thresh = sysctl_sched_latency;
4367 * Halve their sleep time's effect, to allow
4368 * for a gentler effect of sleepers:
4370 if (sched_feat(GENTLE_FAIR_SLEEPERS))
4376 /* ensure we never gain time by being placed backwards. */
4377 se->vruntime = max_vruntime(se->vruntime, vruntime);
4380 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
4382 static inline bool cfs_bandwidth_used(void);
4389 * update_min_vruntime()
4390 * vruntime -= min_vruntime
4394 * update_min_vruntime()
4395 * vruntime += min_vruntime
4397 * this way the vruntime transition between RQs is done when both
4398 * min_vruntime are up-to-date.
4402 * ->migrate_task_rq_fair() (p->state == TASK_WAKING)
4403 * vruntime -= min_vruntime
4407 * update_min_vruntime()
4408 * vruntime += min_vruntime
4410 * this way we don't have the most up-to-date min_vruntime on the originating
4411 * CPU and an up-to-date min_vruntime on the destination CPU.
4415 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
4417 bool renorm = !(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_MIGRATED);
4418 bool curr = cfs_rq->curr == se;
4421 * If we're the current task, we must renormalise before calling
4425 se->vruntime += cfs_rq->min_vruntime;
4427 update_curr(cfs_rq);
4430 * Otherwise, renormalise after, such that we're placed at the current
4431 * moment in time, instead of some random moment in the past. Being
4432 * placed in the past could significantly boost this task to the
4433 * fairness detriment of existing tasks.
4435 if (renorm && !curr)
4436 se->vruntime += cfs_rq->min_vruntime;
4439 * When enqueuing a sched_entity, we must:
4440 * - Update loads to have both entity and cfs_rq synced with now.
4441 * - For group_entity, update its runnable_weight to reflect the new
4442 * h_nr_running of its group cfs_rq.
4443 * - For group_entity, update its weight to reflect the new share of
4445 * - Add its new weight to cfs_rq->load.weight
4447 update_load_avg(cfs_rq, se, UPDATE_TG | DO_ATTACH);
4448 se_update_runnable(se);
4449 update_cfs_group(se);
4450 account_entity_enqueue(cfs_rq, se);
4452 if (flags & ENQUEUE_WAKEUP)
4453 place_entity(cfs_rq, se, 0);
4455 check_schedstat_required();
4456 update_stats_enqueue_fair(cfs_rq, se, flags);
4457 check_spread(cfs_rq, se);
4459 __enqueue_entity(cfs_rq, se);
4462 if (cfs_rq->nr_running == 1) {
4463 check_enqueue_throttle(cfs_rq);
4464 if (!throttled_hierarchy(cfs_rq))
4465 list_add_leaf_cfs_rq(cfs_rq);
4469 static void __clear_buddies_last(struct sched_entity *se)
4471 for_each_sched_entity(se) {
4472 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4473 if (cfs_rq->last != se)
4476 cfs_rq->last = NULL;
4480 static void __clear_buddies_next(struct sched_entity *se)
4482 for_each_sched_entity(se) {
4483 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4484 if (cfs_rq->next != se)
4487 cfs_rq->next = NULL;
4491 static void __clear_buddies_skip(struct sched_entity *se)
4493 for_each_sched_entity(se) {
4494 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4495 if (cfs_rq->skip != se)
4498 cfs_rq->skip = NULL;
4502 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
4504 if (cfs_rq->last == se)
4505 __clear_buddies_last(se);
4507 if (cfs_rq->next == se)
4508 __clear_buddies_next(se);
4510 if (cfs_rq->skip == se)
4511 __clear_buddies_skip(se);
4514 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
4517 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
4519 int action = UPDATE_TG;
4521 if (entity_is_task(se) && task_on_rq_migrating(task_of(se)))
4522 action |= DO_DETACH;
4525 * Update run-time statistics of the 'current'.
4527 update_curr(cfs_rq);
4530 * When dequeuing a sched_entity, we must:
4531 * - Update loads to have both entity and cfs_rq synced with now.
4532 * - For group_entity, update its runnable_weight to reflect the new
4533 * h_nr_running of its group cfs_rq.
4534 * - Subtract its previous weight from cfs_rq->load.weight.
4535 * - For group entity, update its weight to reflect the new share
4536 * of its group cfs_rq.
4538 update_load_avg(cfs_rq, se, action);
4539 se_update_runnable(se);
4541 update_stats_dequeue_fair(cfs_rq, se, flags);
4543 clear_buddies(cfs_rq, se);
4545 if (se != cfs_rq->curr)
4546 __dequeue_entity(cfs_rq, se);
4548 account_entity_dequeue(cfs_rq, se);
4551 * Normalize after update_curr(); which will also have moved
4552 * min_vruntime if @se is the one holding it back. But before doing
4553 * update_min_vruntime() again, which will discount @se's position and
4554 * can move min_vruntime forward still more.
4556 if (!(flags & DEQUEUE_SLEEP))
4557 se->vruntime -= cfs_rq->min_vruntime;
4559 /* return excess runtime on last dequeue */
4560 return_cfs_rq_runtime(cfs_rq);
4562 update_cfs_group(se);
4565 * Now advance min_vruntime if @se was the entity holding it back,
4566 * except when: DEQUEUE_SAVE && !DEQUEUE_MOVE, in this case we'll be
4567 * put back on, and if we advance min_vruntime, we'll be placed back
4568 * further than we started -- ie. we'll be penalized.
4570 if ((flags & (DEQUEUE_SAVE | DEQUEUE_MOVE)) != DEQUEUE_SAVE)
4571 update_min_vruntime(cfs_rq);
4573 if (cfs_rq->nr_running == 0)
4574 update_idle_cfs_rq_clock_pelt(cfs_rq);
4578 * Preempt the current task with a newly woken task if needed:
4581 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
4583 unsigned long ideal_runtime, delta_exec;
4584 struct sched_entity *se;
4587 ideal_runtime = sched_slice(cfs_rq, curr);
4588 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
4589 if (delta_exec > ideal_runtime) {
4590 resched_curr(rq_of(cfs_rq));
4592 * The current task ran long enough, ensure it doesn't get
4593 * re-elected due to buddy favours.
4595 clear_buddies(cfs_rq, curr);
4600 * Ensure that a task that missed wakeup preemption by a
4601 * narrow margin doesn't have to wait for a full slice.
4602 * This also mitigates buddy induced latencies under load.
4604 if (delta_exec < sysctl_sched_min_granularity)
4607 se = __pick_first_entity(cfs_rq);
4608 delta = curr->vruntime - se->vruntime;
4613 if (delta > ideal_runtime)
4614 resched_curr(rq_of(cfs_rq));
4618 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
4620 clear_buddies(cfs_rq, se);
4622 /* 'current' is not kept within the tree. */
4625 * Any task has to be enqueued before it get to execute on
4626 * a CPU. So account for the time it spent waiting on the
4629 update_stats_wait_end_fair(cfs_rq, se);
4630 __dequeue_entity(cfs_rq, se);
4631 update_load_avg(cfs_rq, se, UPDATE_TG);
4634 update_stats_curr_start(cfs_rq, se);
4638 * Track our maximum slice length, if the CPU's load is at
4639 * least twice that of our own weight (i.e. dont track it
4640 * when there are only lesser-weight tasks around):
4642 if (schedstat_enabled() &&
4643 rq_of(cfs_rq)->cfs.load.weight >= 2*se->load.weight) {
4644 struct sched_statistics *stats;
4646 stats = __schedstats_from_se(se);
4647 __schedstat_set(stats->slice_max,
4648 max((u64)stats->slice_max,
4649 se->sum_exec_runtime - se->prev_sum_exec_runtime));
4652 se->prev_sum_exec_runtime = se->sum_exec_runtime;
4656 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
4659 * Pick the next process, keeping these things in mind, in this order:
4660 * 1) keep things fair between processes/task groups
4661 * 2) pick the "next" process, since someone really wants that to run
4662 * 3) pick the "last" process, for cache locality
4663 * 4) do not run the "skip" process, if something else is available
4665 static struct sched_entity *
4666 pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr)
4668 struct sched_entity *left = __pick_first_entity(cfs_rq);
4669 struct sched_entity *se;
4672 * If curr is set we have to see if its left of the leftmost entity
4673 * still in the tree, provided there was anything in the tree at all.
4675 if (!left || (curr && entity_before(curr, left)))
4678 se = left; /* ideally we run the leftmost entity */
4681 * Avoid running the skip buddy, if running something else can
4682 * be done without getting too unfair.
4684 if (cfs_rq->skip && cfs_rq->skip == se) {
4685 struct sched_entity *second;
4688 second = __pick_first_entity(cfs_rq);
4690 second = __pick_next_entity(se);
4691 if (!second || (curr && entity_before(curr, second)))
4695 if (second && wakeup_preempt_entity(second, left) < 1)
4699 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1) {
4701 * Someone really wants this to run. If it's not unfair, run it.
4704 } else if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1) {
4706 * Prefer last buddy, try to return the CPU to a preempted task.
4714 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
4716 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
4719 * If still on the runqueue then deactivate_task()
4720 * was not called and update_curr() has to be done:
4723 update_curr(cfs_rq);
4725 /* throttle cfs_rqs exceeding runtime */
4726 check_cfs_rq_runtime(cfs_rq);
4728 check_spread(cfs_rq, prev);
4731 update_stats_wait_start_fair(cfs_rq, prev);
4732 /* Put 'current' back into the tree. */
4733 __enqueue_entity(cfs_rq, prev);
4734 /* in !on_rq case, update occurred at dequeue */
4735 update_load_avg(cfs_rq, prev, 0);
4737 cfs_rq->curr = NULL;
4741 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
4744 * Update run-time statistics of the 'current'.
4746 update_curr(cfs_rq);
4749 * Ensure that runnable average is periodically updated.
4751 update_load_avg(cfs_rq, curr, UPDATE_TG);
4752 update_cfs_group(curr);
4754 #ifdef CONFIG_SCHED_HRTICK
4756 * queued ticks are scheduled to match the slice, so don't bother
4757 * validating it and just reschedule.
4760 resched_curr(rq_of(cfs_rq));
4764 * don't let the period tick interfere with the hrtick preemption
4766 if (!sched_feat(DOUBLE_TICK) &&
4767 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
4771 if (cfs_rq->nr_running > 1)
4772 check_preempt_tick(cfs_rq, curr);
4776 /**************************************************
4777 * CFS bandwidth control machinery
4780 #ifdef CONFIG_CFS_BANDWIDTH
4782 #ifdef CONFIG_JUMP_LABEL
4783 static struct static_key __cfs_bandwidth_used;
4785 static inline bool cfs_bandwidth_used(void)
4787 return static_key_false(&__cfs_bandwidth_used);
4790 void cfs_bandwidth_usage_inc(void)
4792 static_key_slow_inc_cpuslocked(&__cfs_bandwidth_used);
4795 void cfs_bandwidth_usage_dec(void)
4797 static_key_slow_dec_cpuslocked(&__cfs_bandwidth_used);
4799 #else /* CONFIG_JUMP_LABEL */
4800 static bool cfs_bandwidth_used(void)
4805 void cfs_bandwidth_usage_inc(void) {}
4806 void cfs_bandwidth_usage_dec(void) {}
4807 #endif /* CONFIG_JUMP_LABEL */
4810 * default period for cfs group bandwidth.
4811 * default: 0.1s, units: nanoseconds
4813 static inline u64 default_cfs_period(void)
4815 return 100000000ULL;
4818 static inline u64 sched_cfs_bandwidth_slice(void)
4820 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
4824 * Replenish runtime according to assigned quota. We use sched_clock_cpu
4825 * directly instead of rq->clock to avoid adding additional synchronization
4828 * requires cfs_b->lock
4830 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
4834 if (unlikely(cfs_b->quota == RUNTIME_INF))
4837 cfs_b->runtime += cfs_b->quota;
4838 runtime = cfs_b->runtime_snap - cfs_b->runtime;
4840 cfs_b->burst_time += runtime;
4844 cfs_b->runtime = min(cfs_b->runtime, cfs_b->quota + cfs_b->burst);
4845 cfs_b->runtime_snap = cfs_b->runtime;
4848 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
4850 return &tg->cfs_bandwidth;
4853 /* returns 0 on failure to allocate runtime */
4854 static int __assign_cfs_rq_runtime(struct cfs_bandwidth *cfs_b,
4855 struct cfs_rq *cfs_rq, u64 target_runtime)
4857 u64 min_amount, amount = 0;
4859 lockdep_assert_held(&cfs_b->lock);
4861 /* note: this is a positive sum as runtime_remaining <= 0 */
4862 min_amount = target_runtime - cfs_rq->runtime_remaining;
4864 if (cfs_b->quota == RUNTIME_INF)
4865 amount = min_amount;
4867 start_cfs_bandwidth(cfs_b);
4869 if (cfs_b->runtime > 0) {
4870 amount = min(cfs_b->runtime, min_amount);
4871 cfs_b->runtime -= amount;
4876 cfs_rq->runtime_remaining += amount;
4878 return cfs_rq->runtime_remaining > 0;
4881 /* returns 0 on failure to allocate runtime */
4882 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4884 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4887 raw_spin_lock(&cfs_b->lock);
4888 ret = __assign_cfs_rq_runtime(cfs_b, cfs_rq, sched_cfs_bandwidth_slice());
4889 raw_spin_unlock(&cfs_b->lock);
4894 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
4896 /* dock delta_exec before expiring quota (as it could span periods) */
4897 cfs_rq->runtime_remaining -= delta_exec;
4899 if (likely(cfs_rq->runtime_remaining > 0))
4902 if (cfs_rq->throttled)
4905 * if we're unable to extend our runtime we resched so that the active
4906 * hierarchy can be throttled
4908 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
4909 resched_curr(rq_of(cfs_rq));
4912 static __always_inline
4913 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
4915 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
4918 __account_cfs_rq_runtime(cfs_rq, delta_exec);
4921 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
4923 return cfs_bandwidth_used() && cfs_rq->throttled;
4926 /* check whether cfs_rq, or any parent, is throttled */
4927 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
4929 return cfs_bandwidth_used() && cfs_rq->throttle_count;
4933 * Ensure that neither of the group entities corresponding to src_cpu or
4934 * dest_cpu are members of a throttled hierarchy when performing group
4935 * load-balance operations.
4937 static inline int throttled_lb_pair(struct task_group *tg,
4938 int src_cpu, int dest_cpu)
4940 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
4942 src_cfs_rq = tg->cfs_rq[src_cpu];
4943 dest_cfs_rq = tg->cfs_rq[dest_cpu];
4945 return throttled_hierarchy(src_cfs_rq) ||
4946 throttled_hierarchy(dest_cfs_rq);
4949 static int tg_unthrottle_up(struct task_group *tg, void *data)
4951 struct rq *rq = data;
4952 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
4954 cfs_rq->throttle_count--;
4955 if (!cfs_rq->throttle_count) {
4956 cfs_rq->throttled_clock_pelt_time += rq_clock_pelt(rq) -
4957 cfs_rq->throttled_clock_pelt;
4959 /* Add cfs_rq with load or one or more already running entities to the list */
4960 if (!cfs_rq_is_decayed(cfs_rq))
4961 list_add_leaf_cfs_rq(cfs_rq);
4967 static int tg_throttle_down(struct task_group *tg, void *data)
4969 struct rq *rq = data;
4970 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
4972 /* group is entering throttled state, stop time */
4973 if (!cfs_rq->throttle_count) {
4974 cfs_rq->throttled_clock_pelt = rq_clock_pelt(rq);
4975 list_del_leaf_cfs_rq(cfs_rq);
4977 cfs_rq->throttle_count++;
4982 static bool throttle_cfs_rq(struct cfs_rq *cfs_rq)
4984 struct rq *rq = rq_of(cfs_rq);
4985 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4986 struct sched_entity *se;
4987 long task_delta, idle_task_delta, dequeue = 1;
4989 raw_spin_lock(&cfs_b->lock);
4990 /* This will start the period timer if necessary */
4991 if (__assign_cfs_rq_runtime(cfs_b, cfs_rq, 1)) {
4993 * We have raced with bandwidth becoming available, and if we
4994 * actually throttled the timer might not unthrottle us for an
4995 * entire period. We additionally needed to make sure that any
4996 * subsequent check_cfs_rq_runtime calls agree not to throttle
4997 * us, as we may commit to do cfs put_prev+pick_next, so we ask
4998 * for 1ns of runtime rather than just check cfs_b.
5002 list_add_tail_rcu(&cfs_rq->throttled_list,
5003 &cfs_b->throttled_cfs_rq);
5005 raw_spin_unlock(&cfs_b->lock);
5008 return false; /* Throttle no longer required. */
5010 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
5012 /* freeze hierarchy runnable averages while throttled */
5014 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
5017 task_delta = cfs_rq->h_nr_running;
5018 idle_task_delta = cfs_rq->idle_h_nr_running;
5019 for_each_sched_entity(se) {
5020 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
5021 /* throttled entity or throttle-on-deactivate */
5025 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
5027 if (cfs_rq_is_idle(group_cfs_rq(se)))
5028 idle_task_delta = cfs_rq->h_nr_running;
5030 qcfs_rq->h_nr_running -= task_delta;
5031 qcfs_rq->idle_h_nr_running -= idle_task_delta;
5033 if (qcfs_rq->load.weight) {
5034 /* Avoid re-evaluating load for this entity: */
5035 se = parent_entity(se);
5040 for_each_sched_entity(se) {
5041 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
5042 /* throttled entity or throttle-on-deactivate */
5046 update_load_avg(qcfs_rq, se, 0);
5047 se_update_runnable(se);
5049 if (cfs_rq_is_idle(group_cfs_rq(se)))
5050 idle_task_delta = cfs_rq->h_nr_running;
5052 qcfs_rq->h_nr_running -= task_delta;
5053 qcfs_rq->idle_h_nr_running -= idle_task_delta;
5056 /* At this point se is NULL and we are at root level*/
5057 sub_nr_running(rq, task_delta);
5061 * Note: distribution will already see us throttled via the
5062 * throttled-list. rq->lock protects completion.
5064 cfs_rq->throttled = 1;
5065 cfs_rq->throttled_clock = rq_clock(rq);
5069 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
5071 struct rq *rq = rq_of(cfs_rq);
5072 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
5073 struct sched_entity *se;
5074 long task_delta, idle_task_delta;
5076 se = cfs_rq->tg->se[cpu_of(rq)];
5078 cfs_rq->throttled = 0;
5080 update_rq_clock(rq);
5082 raw_spin_lock(&cfs_b->lock);
5083 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
5084 list_del_rcu(&cfs_rq->throttled_list);
5085 raw_spin_unlock(&cfs_b->lock);
5087 /* update hierarchical throttle state */
5088 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
5090 if (!cfs_rq->load.weight) {
5091 if (!cfs_rq->on_list)
5094 * Nothing to run but something to decay (on_list)?
5095 * Complete the branch.
5097 for_each_sched_entity(se) {
5098 if (list_add_leaf_cfs_rq(cfs_rq_of(se)))
5101 goto unthrottle_throttle;
5104 task_delta = cfs_rq->h_nr_running;
5105 idle_task_delta = cfs_rq->idle_h_nr_running;
5106 for_each_sched_entity(se) {
5107 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
5111 enqueue_entity(qcfs_rq, se, ENQUEUE_WAKEUP);
5113 if (cfs_rq_is_idle(group_cfs_rq(se)))
5114 idle_task_delta = cfs_rq->h_nr_running;
5116 qcfs_rq->h_nr_running += task_delta;
5117 qcfs_rq->idle_h_nr_running += idle_task_delta;
5119 /* end evaluation on encountering a throttled cfs_rq */
5120 if (cfs_rq_throttled(qcfs_rq))
5121 goto unthrottle_throttle;
5124 for_each_sched_entity(se) {
5125 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
5127 update_load_avg(qcfs_rq, se, UPDATE_TG);
5128 se_update_runnable(se);
5130 if (cfs_rq_is_idle(group_cfs_rq(se)))
5131 idle_task_delta = cfs_rq->h_nr_running;
5133 qcfs_rq->h_nr_running += task_delta;
5134 qcfs_rq->idle_h_nr_running += idle_task_delta;
5136 /* end evaluation on encountering a throttled cfs_rq */
5137 if (cfs_rq_throttled(qcfs_rq))
5138 goto unthrottle_throttle;
5141 /* At this point se is NULL and we are at root level*/
5142 add_nr_running(rq, task_delta);
5144 unthrottle_throttle:
5145 assert_list_leaf_cfs_rq(rq);
5147 /* Determine whether we need to wake up potentially idle CPU: */
5148 if (rq->curr == rq->idle && rq->cfs.nr_running)
5152 static void distribute_cfs_runtime(struct cfs_bandwidth *cfs_b)
5154 struct cfs_rq *cfs_rq;
5155 u64 runtime, remaining = 1;
5158 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
5160 struct rq *rq = rq_of(cfs_rq);
5163 rq_lock_irqsave(rq, &rf);
5164 if (!cfs_rq_throttled(cfs_rq))
5167 /* By the above check, this should never be true */
5168 SCHED_WARN_ON(cfs_rq->runtime_remaining > 0);
5170 raw_spin_lock(&cfs_b->lock);
5171 runtime = -cfs_rq->runtime_remaining + 1;
5172 if (runtime > cfs_b->runtime)
5173 runtime = cfs_b->runtime;
5174 cfs_b->runtime -= runtime;
5175 remaining = cfs_b->runtime;
5176 raw_spin_unlock(&cfs_b->lock);
5178 cfs_rq->runtime_remaining += runtime;
5180 /* we check whether we're throttled above */
5181 if (cfs_rq->runtime_remaining > 0)
5182 unthrottle_cfs_rq(cfs_rq);
5185 rq_unlock_irqrestore(rq, &rf);
5194 * Responsible for refilling a task_group's bandwidth and unthrottling its
5195 * cfs_rqs as appropriate. If there has been no activity within the last
5196 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
5197 * used to track this state.
5199 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun, unsigned long flags)
5203 /* no need to continue the timer with no bandwidth constraint */
5204 if (cfs_b->quota == RUNTIME_INF)
5205 goto out_deactivate;
5207 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
5208 cfs_b->nr_periods += overrun;
5210 /* Refill extra burst quota even if cfs_b->idle */
5211 __refill_cfs_bandwidth_runtime(cfs_b);
5214 * idle depends on !throttled (for the case of a large deficit), and if
5215 * we're going inactive then everything else can be deferred
5217 if (cfs_b->idle && !throttled)
5218 goto out_deactivate;
5221 /* mark as potentially idle for the upcoming period */
5226 /* account preceding periods in which throttling occurred */
5227 cfs_b->nr_throttled += overrun;
5230 * This check is repeated as we release cfs_b->lock while we unthrottle.
5232 while (throttled && cfs_b->runtime > 0) {
5233 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
5234 /* we can't nest cfs_b->lock while distributing bandwidth */
5235 distribute_cfs_runtime(cfs_b);
5236 raw_spin_lock_irqsave(&cfs_b->lock, flags);
5238 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
5242 * While we are ensured activity in the period following an
5243 * unthrottle, this also covers the case in which the new bandwidth is
5244 * insufficient to cover the existing bandwidth deficit. (Forcing the
5245 * timer to remain active while there are any throttled entities.)
5255 /* a cfs_rq won't donate quota below this amount */
5256 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
5257 /* minimum remaining period time to redistribute slack quota */
5258 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
5259 /* how long we wait to gather additional slack before distributing */
5260 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
5263 * Are we near the end of the current quota period?
5265 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
5266 * hrtimer base being cleared by hrtimer_start. In the case of
5267 * migrate_hrtimers, base is never cleared, so we are fine.
5269 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
5271 struct hrtimer *refresh_timer = &cfs_b->period_timer;
5274 /* if the call-back is running a quota refresh is already occurring */
5275 if (hrtimer_callback_running(refresh_timer))
5278 /* is a quota refresh about to occur? */
5279 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
5280 if (remaining < (s64)min_expire)
5286 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
5288 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
5290 /* if there's a quota refresh soon don't bother with slack */
5291 if (runtime_refresh_within(cfs_b, min_left))
5294 /* don't push forwards an existing deferred unthrottle */
5295 if (cfs_b->slack_started)
5297 cfs_b->slack_started = true;
5299 hrtimer_start(&cfs_b->slack_timer,
5300 ns_to_ktime(cfs_bandwidth_slack_period),
5304 /* we know any runtime found here is valid as update_curr() precedes return */
5305 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
5307 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
5308 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
5310 if (slack_runtime <= 0)
5313 raw_spin_lock(&cfs_b->lock);
5314 if (cfs_b->quota != RUNTIME_INF) {
5315 cfs_b->runtime += slack_runtime;
5317 /* we are under rq->lock, defer unthrottling using a timer */
5318 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
5319 !list_empty(&cfs_b->throttled_cfs_rq))
5320 start_cfs_slack_bandwidth(cfs_b);
5322 raw_spin_unlock(&cfs_b->lock);
5324 /* even if it's not valid for return we don't want to try again */
5325 cfs_rq->runtime_remaining -= slack_runtime;
5328 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
5330 if (!cfs_bandwidth_used())
5333 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
5336 __return_cfs_rq_runtime(cfs_rq);
5340 * This is done with a timer (instead of inline with bandwidth return) since
5341 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
5343 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
5345 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
5346 unsigned long flags;
5348 /* confirm we're still not at a refresh boundary */
5349 raw_spin_lock_irqsave(&cfs_b->lock, flags);
5350 cfs_b->slack_started = false;
5352 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
5353 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
5357 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice)
5358 runtime = cfs_b->runtime;
5360 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
5365 distribute_cfs_runtime(cfs_b);
5369 * When a group wakes up we want to make sure that its quota is not already
5370 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
5371 * runtime as update_curr() throttling can not trigger until it's on-rq.
5373 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
5375 if (!cfs_bandwidth_used())
5378 /* an active group must be handled by the update_curr()->put() path */
5379 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
5382 /* ensure the group is not already throttled */
5383 if (cfs_rq_throttled(cfs_rq))
5386 /* update runtime allocation */
5387 account_cfs_rq_runtime(cfs_rq, 0);
5388 if (cfs_rq->runtime_remaining <= 0)
5389 throttle_cfs_rq(cfs_rq);
5392 static void sync_throttle(struct task_group *tg, int cpu)
5394 struct cfs_rq *pcfs_rq, *cfs_rq;
5396 if (!cfs_bandwidth_used())
5402 cfs_rq = tg->cfs_rq[cpu];
5403 pcfs_rq = tg->parent->cfs_rq[cpu];
5405 cfs_rq->throttle_count = pcfs_rq->throttle_count;
5406 cfs_rq->throttled_clock_pelt = rq_clock_pelt(cpu_rq(cpu));
5409 /* conditionally throttle active cfs_rq's from put_prev_entity() */
5410 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
5412 if (!cfs_bandwidth_used())
5415 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
5419 * it's possible for a throttled entity to be forced into a running
5420 * state (e.g. set_curr_task), in this case we're finished.
5422 if (cfs_rq_throttled(cfs_rq))
5425 return throttle_cfs_rq(cfs_rq);
5428 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
5430 struct cfs_bandwidth *cfs_b =
5431 container_of(timer, struct cfs_bandwidth, slack_timer);
5433 do_sched_cfs_slack_timer(cfs_b);
5435 return HRTIMER_NORESTART;
5438 extern const u64 max_cfs_quota_period;
5440 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
5442 struct cfs_bandwidth *cfs_b =
5443 container_of(timer, struct cfs_bandwidth, period_timer);
5444 unsigned long flags;
5449 raw_spin_lock_irqsave(&cfs_b->lock, flags);
5451 overrun = hrtimer_forward_now(timer, cfs_b->period);
5455 idle = do_sched_cfs_period_timer(cfs_b, overrun, flags);
5458 u64 new, old = ktime_to_ns(cfs_b->period);
5461 * Grow period by a factor of 2 to avoid losing precision.
5462 * Precision loss in the quota/period ratio can cause __cfs_schedulable
5466 if (new < max_cfs_quota_period) {
5467 cfs_b->period = ns_to_ktime(new);
5471 pr_warn_ratelimited(
5472 "cfs_period_timer[cpu%d]: period too short, scaling up (new cfs_period_us = %lld, cfs_quota_us = %lld)\n",
5474 div_u64(new, NSEC_PER_USEC),
5475 div_u64(cfs_b->quota, NSEC_PER_USEC));
5477 pr_warn_ratelimited(
5478 "cfs_period_timer[cpu%d]: period too short, but cannot scale up without losing precision (cfs_period_us = %lld, cfs_quota_us = %lld)\n",
5480 div_u64(old, NSEC_PER_USEC),
5481 div_u64(cfs_b->quota, NSEC_PER_USEC));
5484 /* reset count so we don't come right back in here */
5489 cfs_b->period_active = 0;
5490 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
5492 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
5495 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
5497 raw_spin_lock_init(&cfs_b->lock);
5499 cfs_b->quota = RUNTIME_INF;
5500 cfs_b->period = ns_to_ktime(default_cfs_period());
5503 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
5504 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED);
5505 cfs_b->period_timer.function = sched_cfs_period_timer;
5506 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
5507 cfs_b->slack_timer.function = sched_cfs_slack_timer;
5508 cfs_b->slack_started = false;
5511 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
5513 cfs_rq->runtime_enabled = 0;
5514 INIT_LIST_HEAD(&cfs_rq->throttled_list);
5517 void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
5519 lockdep_assert_held(&cfs_b->lock);
5521 if (cfs_b->period_active)
5524 cfs_b->period_active = 1;
5525 hrtimer_forward_now(&cfs_b->period_timer, cfs_b->period);
5526 hrtimer_start_expires(&cfs_b->period_timer, HRTIMER_MODE_ABS_PINNED);
5529 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
5531 /* init_cfs_bandwidth() was not called */
5532 if (!cfs_b->throttled_cfs_rq.next)
5535 hrtimer_cancel(&cfs_b->period_timer);
5536 hrtimer_cancel(&cfs_b->slack_timer);
5540 * Both these CPU hotplug callbacks race against unregister_fair_sched_group()
5542 * The race is harmless, since modifying bandwidth settings of unhooked group
5543 * bits doesn't do much.
5546 /* cpu online callback */
5547 static void __maybe_unused update_runtime_enabled(struct rq *rq)
5549 struct task_group *tg;
5551 lockdep_assert_rq_held(rq);
5554 list_for_each_entry_rcu(tg, &task_groups, list) {
5555 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
5556 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
5558 raw_spin_lock(&cfs_b->lock);
5559 cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF;
5560 raw_spin_unlock(&cfs_b->lock);
5565 /* cpu offline callback */
5566 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
5568 struct task_group *tg;
5570 lockdep_assert_rq_held(rq);
5573 list_for_each_entry_rcu(tg, &task_groups, list) {
5574 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
5576 if (!cfs_rq->runtime_enabled)
5580 * clock_task is not advancing so we just need to make sure
5581 * there's some valid quota amount
5583 cfs_rq->runtime_remaining = 1;
5585 * Offline rq is schedulable till CPU is completely disabled
5586 * in take_cpu_down(), so we prevent new cfs throttling here.
5588 cfs_rq->runtime_enabled = 0;
5590 if (cfs_rq_throttled(cfs_rq))
5591 unthrottle_cfs_rq(cfs_rq);
5596 #else /* CONFIG_CFS_BANDWIDTH */
5598 static inline bool cfs_bandwidth_used(void)
5603 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {}
5604 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; }
5605 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
5606 static inline void sync_throttle(struct task_group *tg, int cpu) {}
5607 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
5609 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
5614 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
5619 static inline int throttled_lb_pair(struct task_group *tg,
5620 int src_cpu, int dest_cpu)
5625 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
5627 #ifdef CONFIG_FAIR_GROUP_SCHED
5628 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
5631 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
5635 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
5636 static inline void update_runtime_enabled(struct rq *rq) {}
5637 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
5639 #endif /* CONFIG_CFS_BANDWIDTH */
5641 /**************************************************
5642 * CFS operations on tasks:
5645 #ifdef CONFIG_SCHED_HRTICK
5646 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
5648 struct sched_entity *se = &p->se;
5649 struct cfs_rq *cfs_rq = cfs_rq_of(se);
5651 SCHED_WARN_ON(task_rq(p) != rq);
5653 if (rq->cfs.h_nr_running > 1) {
5654 u64 slice = sched_slice(cfs_rq, se);
5655 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
5656 s64 delta = slice - ran;
5659 if (task_current(rq, p))
5663 hrtick_start(rq, delta);
5668 * called from enqueue/dequeue and updates the hrtick when the
5669 * current task is from our class and nr_running is low enough
5672 static void hrtick_update(struct rq *rq)
5674 struct task_struct *curr = rq->curr;
5676 if (!hrtick_enabled_fair(rq) || curr->sched_class != &fair_sched_class)
5679 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
5680 hrtick_start_fair(rq, curr);
5682 #else /* !CONFIG_SCHED_HRTICK */
5684 hrtick_start_fair(struct rq *rq, struct task_struct *p)
5688 static inline void hrtick_update(struct rq *rq)
5694 static inline bool cpu_overutilized(int cpu)
5696 return !fits_capacity(cpu_util_cfs(cpu), capacity_of(cpu));
5699 static inline void update_overutilized_status(struct rq *rq)
5701 if (!READ_ONCE(rq->rd->overutilized) && cpu_overutilized(rq->cpu)) {
5702 WRITE_ONCE(rq->rd->overutilized, SG_OVERUTILIZED);
5703 trace_sched_overutilized_tp(rq->rd, SG_OVERUTILIZED);
5707 static inline void update_overutilized_status(struct rq *rq) { }
5710 /* Runqueue only has SCHED_IDLE tasks enqueued */
5711 static int sched_idle_rq(struct rq *rq)
5713 return unlikely(rq->nr_running == rq->cfs.idle_h_nr_running &&
5718 * Returns true if cfs_rq only has SCHED_IDLE entities enqueued. Note the use
5719 * of idle_nr_running, which does not consider idle descendants of normal
5722 static bool sched_idle_cfs_rq(struct cfs_rq *cfs_rq)
5724 return cfs_rq->nr_running &&
5725 cfs_rq->nr_running == cfs_rq->idle_nr_running;
5729 static int sched_idle_cpu(int cpu)
5731 return sched_idle_rq(cpu_rq(cpu));
5736 * The enqueue_task method is called before nr_running is
5737 * increased. Here we update the fair scheduling stats and
5738 * then put the task into the rbtree:
5741 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
5743 struct cfs_rq *cfs_rq;
5744 struct sched_entity *se = &p->se;
5745 int idle_h_nr_running = task_has_idle_policy(p);
5746 int task_new = !(flags & ENQUEUE_WAKEUP);
5749 * The code below (indirectly) updates schedutil which looks at
5750 * the cfs_rq utilization to select a frequency.
5751 * Let's add the task's estimated utilization to the cfs_rq's
5752 * estimated utilization, before we update schedutil.
5754 util_est_enqueue(&rq->cfs, p);
5757 * If in_iowait is set, the code below may not trigger any cpufreq
5758 * utilization updates, so do it here explicitly with the IOWAIT flag
5762 cpufreq_update_util(rq, SCHED_CPUFREQ_IOWAIT);
5764 for_each_sched_entity(se) {
5767 cfs_rq = cfs_rq_of(se);
5768 enqueue_entity(cfs_rq, se, flags);
5770 cfs_rq->h_nr_running++;
5771 cfs_rq->idle_h_nr_running += idle_h_nr_running;
5773 if (cfs_rq_is_idle(cfs_rq))
5774 idle_h_nr_running = 1;
5776 /* end evaluation on encountering a throttled cfs_rq */
5777 if (cfs_rq_throttled(cfs_rq))
5778 goto enqueue_throttle;
5780 flags = ENQUEUE_WAKEUP;
5783 for_each_sched_entity(se) {
5784 cfs_rq = cfs_rq_of(se);
5786 update_load_avg(cfs_rq, se, UPDATE_TG);
5787 se_update_runnable(se);
5788 update_cfs_group(se);
5790 cfs_rq->h_nr_running++;
5791 cfs_rq->idle_h_nr_running += idle_h_nr_running;
5793 if (cfs_rq_is_idle(cfs_rq))
5794 idle_h_nr_running = 1;
5796 /* end evaluation on encountering a throttled cfs_rq */
5797 if (cfs_rq_throttled(cfs_rq))
5798 goto enqueue_throttle;
5801 /* At this point se is NULL and we are at root level*/
5802 add_nr_running(rq, 1);
5805 * Since new tasks are assigned an initial util_avg equal to
5806 * half of the spare capacity of their CPU, tiny tasks have the
5807 * ability to cross the overutilized threshold, which will
5808 * result in the load balancer ruining all the task placement
5809 * done by EAS. As a way to mitigate that effect, do not account
5810 * for the first enqueue operation of new tasks during the
5811 * overutilized flag detection.
5813 * A better way of solving this problem would be to wait for
5814 * the PELT signals of tasks to converge before taking them
5815 * into account, but that is not straightforward to implement,
5816 * and the following generally works well enough in practice.
5819 update_overutilized_status(rq);
5822 assert_list_leaf_cfs_rq(rq);
5827 static void set_next_buddy(struct sched_entity *se);
5830 * The dequeue_task method is called before nr_running is
5831 * decreased. We remove the task from the rbtree and
5832 * update the fair scheduling stats:
5834 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
5836 struct cfs_rq *cfs_rq;
5837 struct sched_entity *se = &p->se;
5838 int task_sleep = flags & DEQUEUE_SLEEP;
5839 int idle_h_nr_running = task_has_idle_policy(p);
5840 bool was_sched_idle = sched_idle_rq(rq);
5842 util_est_dequeue(&rq->cfs, p);
5844 for_each_sched_entity(se) {
5845 cfs_rq = cfs_rq_of(se);
5846 dequeue_entity(cfs_rq, se, flags);
5848 cfs_rq->h_nr_running--;
5849 cfs_rq->idle_h_nr_running -= idle_h_nr_running;
5851 if (cfs_rq_is_idle(cfs_rq))
5852 idle_h_nr_running = 1;
5854 /* end evaluation on encountering a throttled cfs_rq */
5855 if (cfs_rq_throttled(cfs_rq))
5856 goto dequeue_throttle;
5858 /* Don't dequeue parent if it has other entities besides us */
5859 if (cfs_rq->load.weight) {
5860 /* Avoid re-evaluating load for this entity: */
5861 se = parent_entity(se);
5863 * Bias pick_next to pick a task from this cfs_rq, as
5864 * p is sleeping when it is within its sched_slice.
5866 if (task_sleep && se && !throttled_hierarchy(cfs_rq))
5870 flags |= DEQUEUE_SLEEP;
5873 for_each_sched_entity(se) {
5874 cfs_rq = cfs_rq_of(se);
5876 update_load_avg(cfs_rq, se, UPDATE_TG);
5877 se_update_runnable(se);
5878 update_cfs_group(se);
5880 cfs_rq->h_nr_running--;
5881 cfs_rq->idle_h_nr_running -= idle_h_nr_running;
5883 if (cfs_rq_is_idle(cfs_rq))
5884 idle_h_nr_running = 1;
5886 /* end evaluation on encountering a throttled cfs_rq */
5887 if (cfs_rq_throttled(cfs_rq))
5888 goto dequeue_throttle;
5892 /* At this point se is NULL and we are at root level*/
5893 sub_nr_running(rq, 1);
5895 /* balance early to pull high priority tasks */
5896 if (unlikely(!was_sched_idle && sched_idle_rq(rq)))
5897 rq->next_balance = jiffies;
5900 util_est_update(&rq->cfs, p, task_sleep);
5906 /* Working cpumask for: load_balance, load_balance_newidle. */
5907 static DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
5908 static DEFINE_PER_CPU(cpumask_var_t, select_rq_mask);
5910 #ifdef CONFIG_NO_HZ_COMMON
5913 cpumask_var_t idle_cpus_mask;
5915 int has_blocked; /* Idle CPUS has blocked load */
5916 int needs_update; /* Newly idle CPUs need their next_balance collated */
5917 unsigned long next_balance; /* in jiffy units */
5918 unsigned long next_blocked; /* Next update of blocked load in jiffies */
5919 } nohz ____cacheline_aligned;
5921 #endif /* CONFIG_NO_HZ_COMMON */
5923 static unsigned long cpu_load(struct rq *rq)
5925 return cfs_rq_load_avg(&rq->cfs);
5929 * cpu_load_without - compute CPU load without any contributions from *p
5930 * @cpu: the CPU which load is requested
5931 * @p: the task which load should be discounted
5933 * The load of a CPU is defined by the load of tasks currently enqueued on that
5934 * CPU as well as tasks which are currently sleeping after an execution on that
5937 * This method returns the load of the specified CPU by discounting the load of
5938 * the specified task, whenever the task is currently contributing to the CPU
5941 static unsigned long cpu_load_without(struct rq *rq, struct task_struct *p)
5943 struct cfs_rq *cfs_rq;
5946 /* Task has no contribution or is new */
5947 if (cpu_of(rq) != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time))
5948 return cpu_load(rq);
5951 load = READ_ONCE(cfs_rq->avg.load_avg);
5953 /* Discount task's util from CPU's util */
5954 lsub_positive(&load, task_h_load(p));
5959 static unsigned long cpu_runnable(struct rq *rq)
5961 return cfs_rq_runnable_avg(&rq->cfs);
5964 static unsigned long cpu_runnable_without(struct rq *rq, struct task_struct *p)
5966 struct cfs_rq *cfs_rq;
5967 unsigned int runnable;
5969 /* Task has no contribution or is new */
5970 if (cpu_of(rq) != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time))
5971 return cpu_runnable(rq);
5974 runnable = READ_ONCE(cfs_rq->avg.runnable_avg);
5976 /* Discount task's runnable from CPU's runnable */
5977 lsub_positive(&runnable, p->se.avg.runnable_avg);
5982 static unsigned long capacity_of(int cpu)
5984 return cpu_rq(cpu)->cpu_capacity;
5987 static void record_wakee(struct task_struct *p)
5990 * Only decay a single time; tasks that have less then 1 wakeup per
5991 * jiffy will not have built up many flips.
5993 if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) {
5994 current->wakee_flips >>= 1;
5995 current->wakee_flip_decay_ts = jiffies;
5998 if (current->last_wakee != p) {
5999 current->last_wakee = p;
6000 current->wakee_flips++;
6005 * Detect M:N waker/wakee relationships via a switching-frequency heuristic.
6007 * A waker of many should wake a different task than the one last awakened
6008 * at a frequency roughly N times higher than one of its wakees.
6010 * In order to determine whether we should let the load spread vs consolidating
6011 * to shared cache, we look for a minimum 'flip' frequency of llc_size in one
6012 * partner, and a factor of lls_size higher frequency in the other.
6014 * With both conditions met, we can be relatively sure that the relationship is
6015 * non-monogamous, with partner count exceeding socket size.
6017 * Waker/wakee being client/server, worker/dispatcher, interrupt source or
6018 * whatever is irrelevant, spread criteria is apparent partner count exceeds
6021 static int wake_wide(struct task_struct *p)
6023 unsigned int master = current->wakee_flips;
6024 unsigned int slave = p->wakee_flips;
6025 int factor = __this_cpu_read(sd_llc_size);
6028 swap(master, slave);
6029 if (slave < factor || master < slave * factor)
6035 * The purpose of wake_affine() is to quickly determine on which CPU we can run
6036 * soonest. For the purpose of speed we only consider the waking and previous
6039 * wake_affine_idle() - only considers 'now', it check if the waking CPU is
6040 * cache-affine and is (or will be) idle.
6042 * wake_affine_weight() - considers the weight to reflect the average
6043 * scheduling latency of the CPUs. This seems to work
6044 * for the overloaded case.
6047 wake_affine_idle(int this_cpu, int prev_cpu, int sync)
6050 * If this_cpu is idle, it implies the wakeup is from interrupt
6051 * context. Only allow the move if cache is shared. Otherwise an
6052 * interrupt intensive workload could force all tasks onto one
6053 * node depending on the IO topology or IRQ affinity settings.
6055 * If the prev_cpu is idle and cache affine then avoid a migration.
6056 * There is no guarantee that the cache hot data from an interrupt
6057 * is more important than cache hot data on the prev_cpu and from
6058 * a cpufreq perspective, it's better to have higher utilisation
6061 if (available_idle_cpu(this_cpu) && cpus_share_cache(this_cpu, prev_cpu))
6062 return available_idle_cpu(prev_cpu) ? prev_cpu : this_cpu;
6064 if (sync && cpu_rq(this_cpu)->nr_running == 1)
6067 if (available_idle_cpu(prev_cpu))
6070 return nr_cpumask_bits;
6074 wake_affine_weight(struct sched_domain *sd, struct task_struct *p,
6075 int this_cpu, int prev_cpu, int sync)
6077 s64 this_eff_load, prev_eff_load;
6078 unsigned long task_load;
6080 this_eff_load = cpu_load(cpu_rq(this_cpu));
6083 unsigned long current_load = task_h_load(current);
6085 if (current_load > this_eff_load)
6088 this_eff_load -= current_load;
6091 task_load = task_h_load(p);
6093 this_eff_load += task_load;
6094 if (sched_feat(WA_BIAS))
6095 this_eff_load *= 100;
6096 this_eff_load *= capacity_of(prev_cpu);
6098 prev_eff_load = cpu_load(cpu_rq(prev_cpu));
6099 prev_eff_load -= task_load;
6100 if (sched_feat(WA_BIAS))
6101 prev_eff_load *= 100 + (sd->imbalance_pct - 100) / 2;
6102 prev_eff_load *= capacity_of(this_cpu);
6105 * If sync, adjust the weight of prev_eff_load such that if
6106 * prev_eff == this_eff that select_idle_sibling() will consider
6107 * stacking the wakee on top of the waker if no other CPU is
6113 return this_eff_load < prev_eff_load ? this_cpu : nr_cpumask_bits;
6116 static int wake_affine(struct sched_domain *sd, struct task_struct *p,
6117 int this_cpu, int prev_cpu, int sync)
6119 int target = nr_cpumask_bits;
6121 if (sched_feat(WA_IDLE))
6122 target = wake_affine_idle(this_cpu, prev_cpu, sync);
6124 if (sched_feat(WA_WEIGHT) && target == nr_cpumask_bits)
6125 target = wake_affine_weight(sd, p, this_cpu, prev_cpu, sync);
6127 schedstat_inc(p->stats.nr_wakeups_affine_attempts);
6128 if (target == nr_cpumask_bits)
6131 schedstat_inc(sd->ttwu_move_affine);
6132 schedstat_inc(p->stats.nr_wakeups_affine);
6136 static struct sched_group *
6137 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu);
6140 * find_idlest_group_cpu - find the idlest CPU among the CPUs in the group.
6143 find_idlest_group_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
6145 unsigned long load, min_load = ULONG_MAX;
6146 unsigned int min_exit_latency = UINT_MAX;
6147 u64 latest_idle_timestamp = 0;
6148 int least_loaded_cpu = this_cpu;
6149 int shallowest_idle_cpu = -1;
6152 /* Check if we have any choice: */
6153 if (group->group_weight == 1)
6154 return cpumask_first(sched_group_span(group));
6156 /* Traverse only the allowed CPUs */
6157 for_each_cpu_and(i, sched_group_span(group), p->cpus_ptr) {
6158 struct rq *rq = cpu_rq(i);
6160 if (!sched_core_cookie_match(rq, p))
6163 if (sched_idle_cpu(i))
6166 if (available_idle_cpu(i)) {
6167 struct cpuidle_state *idle = idle_get_state(rq);
6168 if (idle && idle->exit_latency < min_exit_latency) {
6170 * We give priority to a CPU whose idle state
6171 * has the smallest exit latency irrespective
6172 * of any idle timestamp.
6174 min_exit_latency = idle->exit_latency;
6175 latest_idle_timestamp = rq->idle_stamp;
6176 shallowest_idle_cpu = i;
6177 } else if ((!idle || idle->exit_latency == min_exit_latency) &&
6178 rq->idle_stamp > latest_idle_timestamp) {
6180 * If equal or no active idle state, then
6181 * the most recently idled CPU might have
6184 latest_idle_timestamp = rq->idle_stamp;
6185 shallowest_idle_cpu = i;
6187 } else if (shallowest_idle_cpu == -1) {
6188 load = cpu_load(cpu_rq(i));
6189 if (load < min_load) {
6191 least_loaded_cpu = i;
6196 return shallowest_idle_cpu != -1 ? shallowest_idle_cpu : least_loaded_cpu;
6199 static inline int find_idlest_cpu(struct sched_domain *sd, struct task_struct *p,
6200 int cpu, int prev_cpu, int sd_flag)
6204 if (!cpumask_intersects(sched_domain_span(sd), p->cpus_ptr))
6208 * We need task's util for cpu_util_without, sync it up to
6209 * prev_cpu's last_update_time.
6211 if (!(sd_flag & SD_BALANCE_FORK))
6212 sync_entity_load_avg(&p->se);
6215 struct sched_group *group;
6216 struct sched_domain *tmp;
6219 if (!(sd->flags & sd_flag)) {
6224 group = find_idlest_group(sd, p, cpu);
6230 new_cpu = find_idlest_group_cpu(group, p, cpu);
6231 if (new_cpu == cpu) {
6232 /* Now try balancing at a lower domain level of 'cpu': */
6237 /* Now try balancing at a lower domain level of 'new_cpu': */
6239 weight = sd->span_weight;
6241 for_each_domain(cpu, tmp) {
6242 if (weight <= tmp->span_weight)
6244 if (tmp->flags & sd_flag)
6252 static inline int __select_idle_cpu(int cpu, struct task_struct *p)
6254 if ((available_idle_cpu(cpu) || sched_idle_cpu(cpu)) &&
6255 sched_cpu_cookie_match(cpu_rq(cpu), p))
6261 #ifdef CONFIG_SCHED_SMT
6262 DEFINE_STATIC_KEY_FALSE(sched_smt_present);
6263 EXPORT_SYMBOL_GPL(sched_smt_present);
6265 static inline void set_idle_cores(int cpu, int val)
6267 struct sched_domain_shared *sds;
6269 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
6271 WRITE_ONCE(sds->has_idle_cores, val);
6274 static inline bool test_idle_cores(int cpu, bool def)
6276 struct sched_domain_shared *sds;
6278 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
6280 return READ_ONCE(sds->has_idle_cores);
6286 * Scans the local SMT mask to see if the entire core is idle, and records this
6287 * information in sd_llc_shared->has_idle_cores.
6289 * Since SMT siblings share all cache levels, inspecting this limited remote
6290 * state should be fairly cheap.
6292 void __update_idle_core(struct rq *rq)
6294 int core = cpu_of(rq);
6298 if (test_idle_cores(core, true))
6301 for_each_cpu(cpu, cpu_smt_mask(core)) {
6305 if (!available_idle_cpu(cpu))
6309 set_idle_cores(core, 1);
6315 * Scan the entire LLC domain for idle cores; this dynamically switches off if
6316 * there are no idle cores left in the system; tracked through
6317 * sd_llc->shared->has_idle_cores and enabled through update_idle_core() above.
6319 static int select_idle_core(struct task_struct *p, int core, struct cpumask *cpus, int *idle_cpu)
6324 if (!static_branch_likely(&sched_smt_present))
6325 return __select_idle_cpu(core, p);
6327 for_each_cpu(cpu, cpu_smt_mask(core)) {
6328 if (!available_idle_cpu(cpu)) {
6330 if (*idle_cpu == -1) {
6331 if (sched_idle_cpu(cpu) && cpumask_test_cpu(cpu, p->cpus_ptr)) {
6339 if (*idle_cpu == -1 && cpumask_test_cpu(cpu, p->cpus_ptr))
6346 cpumask_andnot(cpus, cpus, cpu_smt_mask(core));
6351 * Scan the local SMT mask for idle CPUs.
6353 static int select_idle_smt(struct task_struct *p, struct sched_domain *sd, int target)
6357 for_each_cpu(cpu, cpu_smt_mask(target)) {
6358 if (!cpumask_test_cpu(cpu, p->cpus_ptr) ||
6359 !cpumask_test_cpu(cpu, sched_domain_span(sd)))
6361 if (available_idle_cpu(cpu) || sched_idle_cpu(cpu))
6368 #else /* CONFIG_SCHED_SMT */
6370 static inline void set_idle_cores(int cpu, int val)
6374 static inline bool test_idle_cores(int cpu, bool def)
6379 static inline int select_idle_core(struct task_struct *p, int core, struct cpumask *cpus, int *idle_cpu)
6381 return __select_idle_cpu(core, p);
6384 static inline int select_idle_smt(struct task_struct *p, struct sched_domain *sd, int target)
6389 #endif /* CONFIG_SCHED_SMT */
6392 * Scan the LLC domain for idle CPUs; this is dynamically regulated by
6393 * comparing the average scan cost (tracked in sd->avg_scan_cost) against the
6394 * average idle time for this rq (as found in rq->avg_idle).
6396 static int select_idle_cpu(struct task_struct *p, struct sched_domain *sd, bool has_idle_core, int target)
6398 struct cpumask *cpus = this_cpu_cpumask_var_ptr(select_rq_mask);
6399 int i, cpu, idle_cpu = -1, nr = INT_MAX;
6400 struct sched_domain_shared *sd_share;
6401 struct rq *this_rq = this_rq();
6402 int this = smp_processor_id();
6403 struct sched_domain *this_sd;
6406 this_sd = rcu_dereference(*this_cpu_ptr(&sd_llc));
6410 cpumask_and(cpus, sched_domain_span(sd), p->cpus_ptr);
6412 if (sched_feat(SIS_PROP) && !has_idle_core) {
6413 u64 avg_cost, avg_idle, span_avg;
6414 unsigned long now = jiffies;
6417 * If we're busy, the assumption that the last idle period
6418 * predicts the future is flawed; age away the remaining
6419 * predicted idle time.
6421 if (unlikely(this_rq->wake_stamp < now)) {
6422 while (this_rq->wake_stamp < now && this_rq->wake_avg_idle) {
6423 this_rq->wake_stamp++;
6424 this_rq->wake_avg_idle >>= 1;
6428 avg_idle = this_rq->wake_avg_idle;
6429 avg_cost = this_sd->avg_scan_cost + 1;
6431 span_avg = sd->span_weight * avg_idle;
6432 if (span_avg > 4*avg_cost)
6433 nr = div_u64(span_avg, avg_cost);
6437 time = cpu_clock(this);
6440 if (sched_feat(SIS_UTIL)) {
6441 sd_share = rcu_dereference(per_cpu(sd_llc_shared, target));
6443 /* because !--nr is the condition to stop scan */
6444 nr = READ_ONCE(sd_share->nr_idle_scan) + 1;
6445 /* overloaded LLC is unlikely to have idle cpu/core */
6451 for_each_cpu_wrap(cpu, cpus, target + 1) {
6452 if (has_idle_core) {
6453 i = select_idle_core(p, cpu, cpus, &idle_cpu);
6454 if ((unsigned int)i < nr_cpumask_bits)
6460 idle_cpu = __select_idle_cpu(cpu, p);
6461 if ((unsigned int)idle_cpu < nr_cpumask_bits)
6467 set_idle_cores(target, false);
6469 if (sched_feat(SIS_PROP) && !has_idle_core) {
6470 time = cpu_clock(this) - time;
6473 * Account for the scan cost of wakeups against the average
6476 this_rq->wake_avg_idle -= min(this_rq->wake_avg_idle, time);
6478 update_avg(&this_sd->avg_scan_cost, time);
6485 * Scan the asym_capacity domain for idle CPUs; pick the first idle one on which
6486 * the task fits. If no CPU is big enough, but there are idle ones, try to
6487 * maximize capacity.
6490 select_idle_capacity(struct task_struct *p, struct sched_domain *sd, int target)
6492 unsigned long task_util, best_cap = 0;
6493 int cpu, best_cpu = -1;
6494 struct cpumask *cpus;
6496 cpus = this_cpu_cpumask_var_ptr(select_rq_mask);
6497 cpumask_and(cpus, sched_domain_span(sd), p->cpus_ptr);
6499 task_util = uclamp_task_util(p);
6501 for_each_cpu_wrap(cpu, cpus, target) {
6502 unsigned long cpu_cap = capacity_of(cpu);
6504 if (!available_idle_cpu(cpu) && !sched_idle_cpu(cpu))
6506 if (fits_capacity(task_util, cpu_cap))
6509 if (cpu_cap > best_cap) {
6518 static inline bool asym_fits_capacity(unsigned long task_util, int cpu)
6520 if (sched_asym_cpucap_active())
6521 return fits_capacity(task_util, capacity_of(cpu));
6527 * Try and locate an idle core/thread in the LLC cache domain.
6529 static int select_idle_sibling(struct task_struct *p, int prev, int target)
6531 bool has_idle_core = false;
6532 struct sched_domain *sd;
6533 unsigned long task_util;
6534 int i, recent_used_cpu;
6537 * On asymmetric system, update task utilization because we will check
6538 * that the task fits with cpu's capacity.
6540 if (sched_asym_cpucap_active()) {
6541 sync_entity_load_avg(&p->se);
6542 task_util = uclamp_task_util(p);
6546 * per-cpu select_rq_mask usage
6548 lockdep_assert_irqs_disabled();
6550 if ((available_idle_cpu(target) || sched_idle_cpu(target)) &&
6551 asym_fits_capacity(task_util, target))
6555 * If the previous CPU is cache affine and idle, don't be stupid:
6557 if (prev != target && cpus_share_cache(prev, target) &&
6558 (available_idle_cpu(prev) || sched_idle_cpu(prev)) &&
6559 asym_fits_capacity(task_util, prev))
6563 * Allow a per-cpu kthread to stack with the wakee if the
6564 * kworker thread and the tasks previous CPUs are the same.
6565 * The assumption is that the wakee queued work for the
6566 * per-cpu kthread that is now complete and the wakeup is
6567 * essentially a sync wakeup. An obvious example of this
6568 * pattern is IO completions.
6570 if (is_per_cpu_kthread(current) &&
6572 prev == smp_processor_id() &&
6573 this_rq()->nr_running <= 1 &&
6574 asym_fits_capacity(task_util, prev)) {
6578 /* Check a recently used CPU as a potential idle candidate: */
6579 recent_used_cpu = p->recent_used_cpu;
6580 p->recent_used_cpu = prev;
6581 if (recent_used_cpu != prev &&
6582 recent_used_cpu != target &&
6583 cpus_share_cache(recent_used_cpu, target) &&
6584 (available_idle_cpu(recent_used_cpu) || sched_idle_cpu(recent_used_cpu)) &&
6585 cpumask_test_cpu(p->recent_used_cpu, p->cpus_ptr) &&
6586 asym_fits_capacity(task_util, recent_used_cpu)) {
6587 return recent_used_cpu;
6591 * For asymmetric CPU capacity systems, our domain of interest is
6592 * sd_asym_cpucapacity rather than sd_llc.
6594 if (sched_asym_cpucap_active()) {
6595 sd = rcu_dereference(per_cpu(sd_asym_cpucapacity, target));
6597 * On an asymmetric CPU capacity system where an exclusive
6598 * cpuset defines a symmetric island (i.e. one unique
6599 * capacity_orig value through the cpuset), the key will be set
6600 * but the CPUs within that cpuset will not have a domain with
6601 * SD_ASYM_CPUCAPACITY. These should follow the usual symmetric
6605 i = select_idle_capacity(p, sd, target);
6606 return ((unsigned)i < nr_cpumask_bits) ? i : target;
6610 sd = rcu_dereference(per_cpu(sd_llc, target));
6614 if (sched_smt_active()) {
6615 has_idle_core = test_idle_cores(target, false);
6617 if (!has_idle_core && cpus_share_cache(prev, target)) {
6618 i = select_idle_smt(p, sd, prev);
6619 if ((unsigned int)i < nr_cpumask_bits)
6624 i = select_idle_cpu(p, sd, has_idle_core, target);
6625 if ((unsigned)i < nr_cpumask_bits)
6632 * Predicts what cpu_util(@cpu) would return if @p was removed from @cpu
6633 * (@dst_cpu = -1) or migrated to @dst_cpu.
6635 static unsigned long cpu_util_next(int cpu, struct task_struct *p, int dst_cpu)
6637 struct cfs_rq *cfs_rq = &cpu_rq(cpu)->cfs;
6638 unsigned long util = READ_ONCE(cfs_rq->avg.util_avg);
6641 * If @dst_cpu is -1 or @p migrates from @cpu to @dst_cpu remove its
6642 * contribution. If @p migrates from another CPU to @cpu add its
6643 * contribution. In all the other cases @cpu is not impacted by the
6644 * migration so its util_avg is already correct.
6646 if (task_cpu(p) == cpu && dst_cpu != cpu)
6647 lsub_positive(&util, task_util(p));
6648 else if (task_cpu(p) != cpu && dst_cpu == cpu)
6649 util += task_util(p);
6651 if (sched_feat(UTIL_EST)) {
6652 unsigned long util_est;
6654 util_est = READ_ONCE(cfs_rq->avg.util_est.enqueued);
6657 * During wake-up @p isn't enqueued yet and doesn't contribute
6658 * to any cpu_rq(cpu)->cfs.avg.util_est.enqueued.
6659 * If @dst_cpu == @cpu add it to "simulate" cpu_util after @p
6660 * has been enqueued.
6662 * During exec (@dst_cpu = -1) @p is enqueued and does
6663 * contribute to cpu_rq(cpu)->cfs.util_est.enqueued.
6664 * Remove it to "simulate" cpu_util without @p's contribution.
6666 * Despite the task_on_rq_queued(@p) check there is still a
6667 * small window for a possible race when an exec
6668 * select_task_rq_fair() races with LB's detach_task().
6672 * p->on_rq = TASK_ON_RQ_MIGRATING;
6673 * -------------------------------- A
6675 * dequeue_task_fair() + Race Time
6676 * util_est_dequeue() /
6677 * -------------------------------- B
6679 * The additional check "current == p" is required to further
6680 * reduce the race window.
6683 util_est += _task_util_est(p);
6684 else if (unlikely(task_on_rq_queued(p) || current == p))
6685 lsub_positive(&util_est, _task_util_est(p));
6687 util = max(util, util_est);
6690 return min(util, capacity_orig_of(cpu));
6694 * cpu_util_without: compute cpu utilization without any contributions from *p
6695 * @cpu: the CPU which utilization is requested
6696 * @p: the task which utilization should be discounted
6698 * The utilization of a CPU is defined by the utilization of tasks currently
6699 * enqueued on that CPU as well as tasks which are currently sleeping after an
6700 * execution on that CPU.
6702 * This method returns the utilization of the specified CPU by discounting the
6703 * utilization of the specified task, whenever the task is currently
6704 * contributing to the CPU utilization.
6706 static unsigned long cpu_util_without(int cpu, struct task_struct *p)
6708 /* Task has no contribution or is new */
6709 if (cpu != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time))
6710 return cpu_util_cfs(cpu);
6712 return cpu_util_next(cpu, p, -1);
6716 * energy_env - Utilization landscape for energy estimation.
6717 * @task_busy_time: Utilization contribution by the task for which we test the
6718 * placement. Given by eenv_task_busy_time().
6719 * @pd_busy_time: Utilization of the whole perf domain without the task
6720 * contribution. Given by eenv_pd_busy_time().
6721 * @cpu_cap: Maximum CPU capacity for the perf domain.
6722 * @pd_cap: Entire perf domain capacity. (pd->nr_cpus * cpu_cap).
6725 unsigned long task_busy_time;
6726 unsigned long pd_busy_time;
6727 unsigned long cpu_cap;
6728 unsigned long pd_cap;
6732 * Compute the task busy time for compute_energy(). This time cannot be
6733 * injected directly into effective_cpu_util() because of the IRQ scaling.
6734 * The latter only makes sense with the most recent CPUs where the task has
6737 static inline void eenv_task_busy_time(struct energy_env *eenv,
6738 struct task_struct *p, int prev_cpu)
6740 unsigned long busy_time, max_cap = arch_scale_cpu_capacity(prev_cpu);
6741 unsigned long irq = cpu_util_irq(cpu_rq(prev_cpu));
6743 if (unlikely(irq >= max_cap))
6744 busy_time = max_cap;
6746 busy_time = scale_irq_capacity(task_util_est(p), irq, max_cap);
6748 eenv->task_busy_time = busy_time;
6752 * Compute the perf_domain (PD) busy time for compute_energy(). Based on the
6753 * utilization for each @pd_cpus, it however doesn't take into account
6754 * clamping since the ratio (utilization / cpu_capacity) is already enough to
6755 * scale the EM reported power consumption at the (eventually clamped)
6758 * The contribution of the task @p for which we want to estimate the
6759 * energy cost is removed (by cpu_util_next()) and must be calculated
6760 * separately (see eenv_task_busy_time). This ensures:
6762 * - A stable PD utilization, no matter which CPU of that PD we want to place
6765 * - A fair comparison between CPUs as the task contribution (task_util())
6766 * will always be the same no matter which CPU utilization we rely on
6767 * (util_avg or util_est).
6769 * Set @eenv busy time for the PD that spans @pd_cpus. This busy time can't
6770 * exceed @eenv->pd_cap.
6772 static inline void eenv_pd_busy_time(struct energy_env *eenv,
6773 struct cpumask *pd_cpus,
6774 struct task_struct *p)
6776 unsigned long busy_time = 0;
6779 for_each_cpu(cpu, pd_cpus) {
6780 unsigned long util = cpu_util_next(cpu, p, -1);
6782 busy_time += effective_cpu_util(cpu, util, ENERGY_UTIL, NULL);
6785 eenv->pd_busy_time = min(eenv->pd_cap, busy_time);
6789 * Compute the maximum utilization for compute_energy() when the task @p
6790 * is placed on the cpu @dst_cpu.
6792 * Returns the maximum utilization among @eenv->cpus. This utilization can't
6793 * exceed @eenv->cpu_cap.
6795 static inline unsigned long
6796 eenv_pd_max_util(struct energy_env *eenv, struct cpumask *pd_cpus,
6797 struct task_struct *p, int dst_cpu)
6799 unsigned long max_util = 0;
6802 for_each_cpu(cpu, pd_cpus) {
6803 struct task_struct *tsk = (cpu == dst_cpu) ? p : NULL;
6804 unsigned long util = cpu_util_next(cpu, p, dst_cpu);
6805 unsigned long cpu_util;
6808 * Performance domain frequency: utilization clamping
6809 * must be considered since it affects the selection
6810 * of the performance domain frequency.
6811 * NOTE: in case RT tasks are running, by default the
6812 * FREQUENCY_UTIL's utilization can be max OPP.
6814 cpu_util = effective_cpu_util(cpu, util, FREQUENCY_UTIL, tsk);
6815 max_util = max(max_util, cpu_util);
6818 return min(max_util, eenv->cpu_cap);
6822 * compute_energy(): Use the Energy Model to estimate the energy that @pd would
6823 * consume for a given utilization landscape @eenv. When @dst_cpu < 0, the task
6824 * contribution is ignored.
6826 static inline unsigned long
6827 compute_energy(struct energy_env *eenv, struct perf_domain *pd,
6828 struct cpumask *pd_cpus, struct task_struct *p, int dst_cpu)
6830 unsigned long max_util = eenv_pd_max_util(eenv, pd_cpus, p, dst_cpu);
6831 unsigned long busy_time = eenv->pd_busy_time;
6834 busy_time = min(eenv->pd_cap, busy_time + eenv->task_busy_time);
6836 return em_cpu_energy(pd->em_pd, max_util, busy_time, eenv->cpu_cap);
6840 * find_energy_efficient_cpu(): Find most energy-efficient target CPU for the
6841 * waking task. find_energy_efficient_cpu() looks for the CPU with maximum
6842 * spare capacity in each performance domain and uses it as a potential
6843 * candidate to execute the task. Then, it uses the Energy Model to figure
6844 * out which of the CPU candidates is the most energy-efficient.
6846 * The rationale for this heuristic is as follows. In a performance domain,
6847 * all the most energy efficient CPU candidates (according to the Energy
6848 * Model) are those for which we'll request a low frequency. When there are
6849 * several CPUs for which the frequency request will be the same, we don't
6850 * have enough data to break the tie between them, because the Energy Model
6851 * only includes active power costs. With this model, if we assume that
6852 * frequency requests follow utilization (e.g. using schedutil), the CPU with
6853 * the maximum spare capacity in a performance domain is guaranteed to be among
6854 * the best candidates of the performance domain.
6856 * In practice, it could be preferable from an energy standpoint to pack
6857 * small tasks on a CPU in order to let other CPUs go in deeper idle states,
6858 * but that could also hurt our chances to go cluster idle, and we have no
6859 * ways to tell with the current Energy Model if this is actually a good
6860 * idea or not. So, find_energy_efficient_cpu() basically favors
6861 * cluster-packing, and spreading inside a cluster. That should at least be
6862 * a good thing for latency, and this is consistent with the idea that most
6863 * of the energy savings of EAS come from the asymmetry of the system, and
6864 * not so much from breaking the tie between identical CPUs. That's also the
6865 * reason why EAS is enabled in the topology code only for systems where
6866 * SD_ASYM_CPUCAPACITY is set.
6868 * NOTE: Forkees are not accepted in the energy-aware wake-up path because
6869 * they don't have any useful utilization data yet and it's not possible to
6870 * forecast their impact on energy consumption. Consequently, they will be
6871 * placed by find_idlest_cpu() on the least loaded CPU, which might turn out
6872 * to be energy-inefficient in some use-cases. The alternative would be to
6873 * bias new tasks towards specific types of CPUs first, or to try to infer
6874 * their util_avg from the parent task, but those heuristics could hurt
6875 * other use-cases too. So, until someone finds a better way to solve this,
6876 * let's keep things simple by re-using the existing slow path.
6878 static int find_energy_efficient_cpu(struct task_struct *p, int prev_cpu)
6880 struct cpumask *cpus = this_cpu_cpumask_var_ptr(select_rq_mask);
6881 unsigned long prev_delta = ULONG_MAX, best_delta = ULONG_MAX;
6882 struct root_domain *rd = this_rq()->rd;
6883 int cpu, best_energy_cpu, target = -1;
6884 struct sched_domain *sd;
6885 struct perf_domain *pd;
6886 struct energy_env eenv;
6889 pd = rcu_dereference(rd->pd);
6890 if (!pd || READ_ONCE(rd->overutilized))
6894 * Energy-aware wake-up happens on the lowest sched_domain starting
6895 * from sd_asym_cpucapacity spanning over this_cpu and prev_cpu.
6897 sd = rcu_dereference(*this_cpu_ptr(&sd_asym_cpucapacity));
6898 while (sd && !cpumask_test_cpu(prev_cpu, sched_domain_span(sd)))
6905 sync_entity_load_avg(&p->se);
6906 if (!task_util_est(p))
6909 eenv_task_busy_time(&eenv, p, prev_cpu);
6911 for (; pd; pd = pd->next) {
6912 unsigned long cpu_cap, cpu_thermal_cap, util;
6913 unsigned long cur_delta, max_spare_cap = 0;
6914 bool compute_prev_delta = false;
6915 int max_spare_cap_cpu = -1;
6916 unsigned long base_energy;
6918 cpumask_and(cpus, perf_domain_span(pd), cpu_online_mask);
6920 if (cpumask_empty(cpus))
6923 /* Account thermal pressure for the energy estimation */
6924 cpu = cpumask_first(cpus);
6925 cpu_thermal_cap = arch_scale_cpu_capacity(cpu);
6926 cpu_thermal_cap -= arch_scale_thermal_pressure(cpu);
6928 eenv.cpu_cap = cpu_thermal_cap;
6931 for_each_cpu(cpu, cpus) {
6932 eenv.pd_cap += cpu_thermal_cap;
6934 if (!cpumask_test_cpu(cpu, sched_domain_span(sd)))
6937 if (!cpumask_test_cpu(cpu, p->cpus_ptr))
6940 util = cpu_util_next(cpu, p, cpu);
6941 cpu_cap = capacity_of(cpu);
6944 * Skip CPUs that cannot satisfy the capacity request.
6945 * IOW, placing the task there would make the CPU
6946 * overutilized. Take uclamp into account to see how
6947 * much capacity we can get out of the CPU; this is
6948 * aligned with sched_cpu_util().
6950 util = uclamp_rq_util_with(cpu_rq(cpu), util, p);
6951 if (!fits_capacity(util, cpu_cap))
6954 lsub_positive(&cpu_cap, util);
6956 if (cpu == prev_cpu) {
6957 /* Always use prev_cpu as a candidate. */
6958 compute_prev_delta = true;
6959 } else if (cpu_cap > max_spare_cap) {
6961 * Find the CPU with the maximum spare capacity
6962 * in the performance domain.
6964 max_spare_cap = cpu_cap;
6965 max_spare_cap_cpu = cpu;
6969 if (max_spare_cap_cpu < 0 && !compute_prev_delta)
6972 eenv_pd_busy_time(&eenv, cpus, p);
6973 /* Compute the 'base' energy of the pd, without @p */
6974 base_energy = compute_energy(&eenv, pd, cpus, p, -1);
6976 /* Evaluate the energy impact of using prev_cpu. */
6977 if (compute_prev_delta) {
6978 prev_delta = compute_energy(&eenv, pd, cpus, p,
6980 /* CPU utilization has changed */
6981 if (prev_delta < base_energy)
6983 prev_delta -= base_energy;
6984 best_delta = min(best_delta, prev_delta);
6987 /* Evaluate the energy impact of using max_spare_cap_cpu. */
6988 if (max_spare_cap_cpu >= 0) {
6989 cur_delta = compute_energy(&eenv, pd, cpus, p,
6991 /* CPU utilization has changed */
6992 if (cur_delta < base_energy)
6994 cur_delta -= base_energy;
6995 if (cur_delta < best_delta) {
6996 best_delta = cur_delta;
6997 best_energy_cpu = max_spare_cap_cpu;
7003 if (best_delta < prev_delta)
7004 target = best_energy_cpu;
7015 * select_task_rq_fair: Select target runqueue for the waking task in domains
7016 * that have the relevant SD flag set. In practice, this is SD_BALANCE_WAKE,
7017 * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
7019 * Balances load by selecting the idlest CPU in the idlest group, or under
7020 * certain conditions an idle sibling CPU if the domain has SD_WAKE_AFFINE set.
7022 * Returns the target CPU number.
7025 select_task_rq_fair(struct task_struct *p, int prev_cpu, int wake_flags)
7027 int sync = (wake_flags & WF_SYNC) && !(current->flags & PF_EXITING);
7028 struct sched_domain *tmp, *sd = NULL;
7029 int cpu = smp_processor_id();
7030 int new_cpu = prev_cpu;
7031 int want_affine = 0;
7032 /* SD_flags and WF_flags share the first nibble */
7033 int sd_flag = wake_flags & 0xF;
7036 * required for stable ->cpus_allowed
7038 lockdep_assert_held(&p->pi_lock);
7039 if (wake_flags & WF_TTWU) {
7042 if (sched_energy_enabled()) {
7043 new_cpu = find_energy_efficient_cpu(p, prev_cpu);
7049 want_affine = !wake_wide(p) && cpumask_test_cpu(cpu, p->cpus_ptr);
7053 for_each_domain(cpu, tmp) {
7055 * If both 'cpu' and 'prev_cpu' are part of this domain,
7056 * cpu is a valid SD_WAKE_AFFINE target.
7058 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
7059 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
7060 if (cpu != prev_cpu)
7061 new_cpu = wake_affine(tmp, p, cpu, prev_cpu, sync);
7063 sd = NULL; /* Prefer wake_affine over balance flags */
7068 * Usually only true for WF_EXEC and WF_FORK, as sched_domains
7069 * usually do not have SD_BALANCE_WAKE set. That means wakeup
7070 * will usually go to the fast path.
7072 if (tmp->flags & sd_flag)
7074 else if (!want_affine)
7080 new_cpu = find_idlest_cpu(sd, p, cpu, prev_cpu, sd_flag);
7081 } else if (wake_flags & WF_TTWU) { /* XXX always ? */
7083 new_cpu = select_idle_sibling(p, prev_cpu, new_cpu);
7091 * Called immediately before a task is migrated to a new CPU; task_cpu(p) and
7092 * cfs_rq_of(p) references at time of call are still valid and identify the
7093 * previous CPU. The caller guarantees p->pi_lock or task_rq(p)->lock is held.
7095 static void migrate_task_rq_fair(struct task_struct *p, int new_cpu)
7097 struct sched_entity *se = &p->se;
7100 * As blocked tasks retain absolute vruntime the migration needs to
7101 * deal with this by subtracting the old and adding the new
7102 * min_vruntime -- the latter is done by enqueue_entity() when placing
7103 * the task on the new runqueue.
7105 if (READ_ONCE(p->__state) == TASK_WAKING) {
7106 struct cfs_rq *cfs_rq = cfs_rq_of(se);
7108 se->vruntime -= u64_u32_load(cfs_rq->min_vruntime);
7111 if (!task_on_rq_migrating(p)) {
7112 remove_entity_load_avg(se);
7115 * Here, the task's PELT values have been updated according to
7116 * the current rq's clock. But if that clock hasn't been
7117 * updated in a while, a substantial idle time will be missed,
7118 * leading to an inflation after wake-up on the new rq.
7120 * Estimate the missing time from the cfs_rq last_update_time
7121 * and update sched_avg to improve the PELT continuity after
7124 migrate_se_pelt_lag(se);
7127 /* Tell new CPU we are migrated */
7128 se->avg.last_update_time = 0;
7130 /* We have migrated, no longer consider this task hot */
7133 update_scan_period(p, new_cpu);
7136 static void task_dead_fair(struct task_struct *p)
7138 remove_entity_load_avg(&p->se);
7142 balance_fair(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
7147 return newidle_balance(rq, rf) != 0;
7149 #endif /* CONFIG_SMP */
7151 static unsigned long wakeup_gran(struct sched_entity *se)
7153 unsigned long gran = sysctl_sched_wakeup_granularity;
7156 * Since its curr running now, convert the gran from real-time
7157 * to virtual-time in his units.
7159 * By using 'se' instead of 'curr' we penalize light tasks, so
7160 * they get preempted easier. That is, if 'se' < 'curr' then
7161 * the resulting gran will be larger, therefore penalizing the
7162 * lighter, if otoh 'se' > 'curr' then the resulting gran will
7163 * be smaller, again penalizing the lighter task.
7165 * This is especially important for buddies when the leftmost
7166 * task is higher priority than the buddy.
7168 return calc_delta_fair(gran, se);
7172 * Should 'se' preempt 'curr'.
7186 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
7188 s64 gran, vdiff = curr->vruntime - se->vruntime;
7193 gran = wakeup_gran(se);
7200 static void set_last_buddy(struct sched_entity *se)
7202 for_each_sched_entity(se) {
7203 if (SCHED_WARN_ON(!se->on_rq))
7207 cfs_rq_of(se)->last = se;
7211 static void set_next_buddy(struct sched_entity *se)
7213 for_each_sched_entity(se) {
7214 if (SCHED_WARN_ON(!se->on_rq))
7218 cfs_rq_of(se)->next = se;
7222 static void set_skip_buddy(struct sched_entity *se)
7224 for_each_sched_entity(se)
7225 cfs_rq_of(se)->skip = se;
7229 * Preempt the current task with a newly woken task if needed:
7231 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
7233 struct task_struct *curr = rq->curr;
7234 struct sched_entity *se = &curr->se, *pse = &p->se;
7235 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
7236 int scale = cfs_rq->nr_running >= sched_nr_latency;
7237 int next_buddy_marked = 0;
7238 int cse_is_idle, pse_is_idle;
7240 if (unlikely(se == pse))
7244 * This is possible from callers such as attach_tasks(), in which we
7245 * unconditionally check_preempt_curr() after an enqueue (which may have
7246 * lead to a throttle). This both saves work and prevents false
7247 * next-buddy nomination below.
7249 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
7252 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
7253 set_next_buddy(pse);
7254 next_buddy_marked = 1;
7258 * We can come here with TIF_NEED_RESCHED already set from new task
7261 * Note: this also catches the edge-case of curr being in a throttled
7262 * group (e.g. via set_curr_task), since update_curr() (in the
7263 * enqueue of curr) will have resulted in resched being set. This
7264 * prevents us from potentially nominating it as a false LAST_BUDDY
7267 if (test_tsk_need_resched(curr))
7270 /* Idle tasks are by definition preempted by non-idle tasks. */
7271 if (unlikely(task_has_idle_policy(curr)) &&
7272 likely(!task_has_idle_policy(p)))
7276 * Batch and idle tasks do not preempt non-idle tasks (their preemption
7277 * is driven by the tick):
7279 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
7282 find_matching_se(&se, &pse);
7285 cse_is_idle = se_is_idle(se);
7286 pse_is_idle = se_is_idle(pse);
7289 * Preempt an idle group in favor of a non-idle group (and don't preempt
7290 * in the inverse case).
7292 if (cse_is_idle && !pse_is_idle)
7294 if (cse_is_idle != pse_is_idle)
7297 update_curr(cfs_rq_of(se));
7298 if (wakeup_preempt_entity(se, pse) == 1) {
7300 * Bias pick_next to pick the sched entity that is
7301 * triggering this preemption.
7303 if (!next_buddy_marked)
7304 set_next_buddy(pse);
7313 * Only set the backward buddy when the current task is still
7314 * on the rq. This can happen when a wakeup gets interleaved
7315 * with schedule on the ->pre_schedule() or idle_balance()
7316 * point, either of which can * drop the rq lock.
7318 * Also, during early boot the idle thread is in the fair class,
7319 * for obvious reasons its a bad idea to schedule back to it.
7321 if (unlikely(!se->on_rq || curr == rq->idle))
7324 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
7329 static struct task_struct *pick_task_fair(struct rq *rq)
7331 struct sched_entity *se;
7332 struct cfs_rq *cfs_rq;
7336 if (!cfs_rq->nr_running)
7340 struct sched_entity *curr = cfs_rq->curr;
7342 /* When we pick for a remote RQ, we'll not have done put_prev_entity() */
7345 update_curr(cfs_rq);
7349 if (unlikely(check_cfs_rq_runtime(cfs_rq)))
7353 se = pick_next_entity(cfs_rq, curr);
7354 cfs_rq = group_cfs_rq(se);
7361 struct task_struct *
7362 pick_next_task_fair(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
7364 struct cfs_rq *cfs_rq = &rq->cfs;
7365 struct sched_entity *se;
7366 struct task_struct *p;
7370 if (!sched_fair_runnable(rq))
7373 #ifdef CONFIG_FAIR_GROUP_SCHED
7374 if (!prev || prev->sched_class != &fair_sched_class)
7378 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
7379 * likely that a next task is from the same cgroup as the current.
7381 * Therefore attempt to avoid putting and setting the entire cgroup
7382 * hierarchy, only change the part that actually changes.
7386 struct sched_entity *curr = cfs_rq->curr;
7389 * Since we got here without doing put_prev_entity() we also
7390 * have to consider cfs_rq->curr. If it is still a runnable
7391 * entity, update_curr() will update its vruntime, otherwise
7392 * forget we've ever seen it.
7396 update_curr(cfs_rq);
7401 * This call to check_cfs_rq_runtime() will do the
7402 * throttle and dequeue its entity in the parent(s).
7403 * Therefore the nr_running test will indeed
7406 if (unlikely(check_cfs_rq_runtime(cfs_rq))) {
7409 if (!cfs_rq->nr_running)
7416 se = pick_next_entity(cfs_rq, curr);
7417 cfs_rq = group_cfs_rq(se);
7423 * Since we haven't yet done put_prev_entity and if the selected task
7424 * is a different task than we started out with, try and touch the
7425 * least amount of cfs_rqs.
7428 struct sched_entity *pse = &prev->se;
7430 while (!(cfs_rq = is_same_group(se, pse))) {
7431 int se_depth = se->depth;
7432 int pse_depth = pse->depth;
7434 if (se_depth <= pse_depth) {
7435 put_prev_entity(cfs_rq_of(pse), pse);
7436 pse = parent_entity(pse);
7438 if (se_depth >= pse_depth) {
7439 set_next_entity(cfs_rq_of(se), se);
7440 se = parent_entity(se);
7444 put_prev_entity(cfs_rq, pse);
7445 set_next_entity(cfs_rq, se);
7452 put_prev_task(rq, prev);
7455 se = pick_next_entity(cfs_rq, NULL);
7456 set_next_entity(cfs_rq, se);
7457 cfs_rq = group_cfs_rq(se);
7462 done: __maybe_unused;
7465 * Move the next running task to the front of
7466 * the list, so our cfs_tasks list becomes MRU
7469 list_move(&p->se.group_node, &rq->cfs_tasks);
7472 if (hrtick_enabled_fair(rq))
7473 hrtick_start_fair(rq, p);
7475 update_misfit_status(p, rq);
7483 new_tasks = newidle_balance(rq, rf);
7486 * Because newidle_balance() releases (and re-acquires) rq->lock, it is
7487 * possible for any higher priority task to appear. In that case we
7488 * must re-start the pick_next_entity() loop.
7497 * rq is about to be idle, check if we need to update the
7498 * lost_idle_time of clock_pelt
7500 update_idle_rq_clock_pelt(rq);
7505 static struct task_struct *__pick_next_task_fair(struct rq *rq)
7507 return pick_next_task_fair(rq, NULL, NULL);
7511 * Account for a descheduled task:
7513 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
7515 struct sched_entity *se = &prev->se;
7516 struct cfs_rq *cfs_rq;
7518 for_each_sched_entity(se) {
7519 cfs_rq = cfs_rq_of(se);
7520 put_prev_entity(cfs_rq, se);
7525 * sched_yield() is very simple
7527 * The magic of dealing with the ->skip buddy is in pick_next_entity.
7529 static void yield_task_fair(struct rq *rq)
7531 struct task_struct *curr = rq->curr;
7532 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
7533 struct sched_entity *se = &curr->se;
7536 * Are we the only task in the tree?
7538 if (unlikely(rq->nr_running == 1))
7541 clear_buddies(cfs_rq, se);
7543 if (curr->policy != SCHED_BATCH) {
7544 update_rq_clock(rq);
7546 * Update run-time statistics of the 'current'.
7548 update_curr(cfs_rq);
7550 * Tell update_rq_clock() that we've just updated,
7551 * so we don't do microscopic update in schedule()
7552 * and double the fastpath cost.
7554 rq_clock_skip_update(rq);
7560 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p)
7562 struct sched_entity *se = &p->se;
7564 /* throttled hierarchies are not runnable */
7565 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
7568 /* Tell the scheduler that we'd really like pse to run next. */
7571 yield_task_fair(rq);
7577 /**************************************************
7578 * Fair scheduling class load-balancing methods.
7582 * The purpose of load-balancing is to achieve the same basic fairness the
7583 * per-CPU scheduler provides, namely provide a proportional amount of compute
7584 * time to each task. This is expressed in the following equation:
7586 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
7588 * Where W_i,n is the n-th weight average for CPU i. The instantaneous weight
7589 * W_i,0 is defined as:
7591 * W_i,0 = \Sum_j w_i,j (2)
7593 * Where w_i,j is the weight of the j-th runnable task on CPU i. This weight
7594 * is derived from the nice value as per sched_prio_to_weight[].
7596 * The weight average is an exponential decay average of the instantaneous
7599 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
7601 * C_i is the compute capacity of CPU i, typically it is the
7602 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
7603 * can also include other factors [XXX].
7605 * To achieve this balance we define a measure of imbalance which follows
7606 * directly from (1):
7608 * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4)
7610 * We them move tasks around to minimize the imbalance. In the continuous
7611 * function space it is obvious this converges, in the discrete case we get
7612 * a few fun cases generally called infeasible weight scenarios.
7615 * - infeasible weights;
7616 * - local vs global optima in the discrete case. ]
7621 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
7622 * for all i,j solution, we create a tree of CPUs that follows the hardware
7623 * topology where each level pairs two lower groups (or better). This results
7624 * in O(log n) layers. Furthermore we reduce the number of CPUs going up the
7625 * tree to only the first of the previous level and we decrease the frequency
7626 * of load-balance at each level inv. proportional to the number of CPUs in
7632 * \Sum { --- * --- * 2^i } = O(n) (5)
7634 * `- size of each group
7635 * | | `- number of CPUs doing load-balance
7637 * `- sum over all levels
7639 * Coupled with a limit on how many tasks we can migrate every balance pass,
7640 * this makes (5) the runtime complexity of the balancer.
7642 * An important property here is that each CPU is still (indirectly) connected
7643 * to every other CPU in at most O(log n) steps:
7645 * The adjacency matrix of the resulting graph is given by:
7648 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
7651 * And you'll find that:
7653 * A^(log_2 n)_i,j != 0 for all i,j (7)
7655 * Showing there's indeed a path between every CPU in at most O(log n) steps.
7656 * The task movement gives a factor of O(m), giving a convergence complexity
7659 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
7664 * In order to avoid CPUs going idle while there's still work to do, new idle
7665 * balancing is more aggressive and has the newly idle CPU iterate up the domain
7666 * tree itself instead of relying on other CPUs to bring it work.
7668 * This adds some complexity to both (5) and (8) but it reduces the total idle
7676 * Cgroups make a horror show out of (2), instead of a simple sum we get:
7679 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
7684 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
7686 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on CPU i.
7688 * The big problem is S_k, its a global sum needed to compute a local (W_i)
7691 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
7692 * rewrite all of this once again.]
7695 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
7697 enum fbq_type { regular, remote, all };
7700 * 'group_type' describes the group of CPUs at the moment of load balancing.
7702 * The enum is ordered by pulling priority, with the group with lowest priority
7703 * first so the group_type can simply be compared when selecting the busiest
7704 * group. See update_sd_pick_busiest().
7707 /* The group has spare capacity that can be used to run more tasks. */
7708 group_has_spare = 0,
7710 * The group is fully used and the tasks don't compete for more CPU
7711 * cycles. Nevertheless, some tasks might wait before running.
7715 * One task doesn't fit with CPU's capacity and must be migrated to a
7716 * more powerful CPU.
7720 * SD_ASYM_PACKING only: One local CPU with higher capacity is available,
7721 * and the task should be migrated to it instead of running on the
7726 * The tasks' affinity constraints previously prevented the scheduler
7727 * from balancing the load across the system.
7731 * The CPU is overloaded and can't provide expected CPU cycles to all
7737 enum migration_type {
7744 #define LBF_ALL_PINNED 0x01
7745 #define LBF_NEED_BREAK 0x02
7746 #define LBF_DST_PINNED 0x04
7747 #define LBF_SOME_PINNED 0x08
7748 #define LBF_ACTIVE_LB 0x10
7751 struct sched_domain *sd;
7759 struct cpumask *dst_grpmask;
7761 enum cpu_idle_type idle;
7763 /* The set of CPUs under consideration for load-balancing */
7764 struct cpumask *cpus;
7769 unsigned int loop_break;
7770 unsigned int loop_max;
7772 enum fbq_type fbq_type;
7773 enum migration_type migration_type;
7774 struct list_head tasks;
7778 * Is this task likely cache-hot:
7780 static int task_hot(struct task_struct *p, struct lb_env *env)
7784 lockdep_assert_rq_held(env->src_rq);
7786 if (p->sched_class != &fair_sched_class)
7789 if (unlikely(task_has_idle_policy(p)))
7792 /* SMT siblings share cache */
7793 if (env->sd->flags & SD_SHARE_CPUCAPACITY)
7797 * Buddy candidates are cache hot:
7799 if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running &&
7800 (&p->se == cfs_rq_of(&p->se)->next ||
7801 &p->se == cfs_rq_of(&p->se)->last))
7804 if (sysctl_sched_migration_cost == -1)
7808 * Don't migrate task if the task's cookie does not match
7809 * with the destination CPU's core cookie.
7811 if (!sched_core_cookie_match(cpu_rq(env->dst_cpu), p))
7814 if (sysctl_sched_migration_cost == 0)
7817 delta = rq_clock_task(env->src_rq) - p->se.exec_start;
7819 return delta < (s64)sysctl_sched_migration_cost;
7822 #ifdef CONFIG_NUMA_BALANCING
7824 * Returns 1, if task migration degrades locality
7825 * Returns 0, if task migration improves locality i.e migration preferred.
7826 * Returns -1, if task migration is not affected by locality.
7828 static int migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
7830 struct numa_group *numa_group = rcu_dereference(p->numa_group);
7831 unsigned long src_weight, dst_weight;
7832 int src_nid, dst_nid, dist;
7834 if (!static_branch_likely(&sched_numa_balancing))
7837 if (!p->numa_faults || !(env->sd->flags & SD_NUMA))
7840 src_nid = cpu_to_node(env->src_cpu);
7841 dst_nid = cpu_to_node(env->dst_cpu);
7843 if (src_nid == dst_nid)
7846 /* Migrating away from the preferred node is always bad. */
7847 if (src_nid == p->numa_preferred_nid) {
7848 if (env->src_rq->nr_running > env->src_rq->nr_preferred_running)
7854 /* Encourage migration to the preferred node. */
7855 if (dst_nid == p->numa_preferred_nid)
7858 /* Leaving a core idle is often worse than degrading locality. */
7859 if (env->idle == CPU_IDLE)
7862 dist = node_distance(src_nid, dst_nid);
7864 src_weight = group_weight(p, src_nid, dist);
7865 dst_weight = group_weight(p, dst_nid, dist);
7867 src_weight = task_weight(p, src_nid, dist);
7868 dst_weight = task_weight(p, dst_nid, dist);
7871 return dst_weight < src_weight;
7875 static inline int migrate_degrades_locality(struct task_struct *p,
7883 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
7886 int can_migrate_task(struct task_struct *p, struct lb_env *env)
7890 lockdep_assert_rq_held(env->src_rq);
7893 * We do not migrate tasks that are:
7894 * 1) throttled_lb_pair, or
7895 * 2) cannot be migrated to this CPU due to cpus_ptr, or
7896 * 3) running (obviously), or
7897 * 4) are cache-hot on their current CPU.
7899 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
7902 /* Disregard pcpu kthreads; they are where they need to be. */
7903 if (kthread_is_per_cpu(p))
7906 if (!cpumask_test_cpu(env->dst_cpu, p->cpus_ptr)) {
7909 schedstat_inc(p->stats.nr_failed_migrations_affine);
7911 env->flags |= LBF_SOME_PINNED;
7914 * Remember if this task can be migrated to any other CPU in
7915 * our sched_group. We may want to revisit it if we couldn't
7916 * meet load balance goals by pulling other tasks on src_cpu.
7918 * Avoid computing new_dst_cpu
7920 * - if we have already computed one in current iteration
7921 * - if it's an active balance
7923 if (env->idle == CPU_NEWLY_IDLE ||
7924 env->flags & (LBF_DST_PINNED | LBF_ACTIVE_LB))
7927 /* Prevent to re-select dst_cpu via env's CPUs: */
7928 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
7929 if (cpumask_test_cpu(cpu, p->cpus_ptr)) {
7930 env->flags |= LBF_DST_PINNED;
7931 env->new_dst_cpu = cpu;
7939 /* Record that we found at least one task that could run on dst_cpu */
7940 env->flags &= ~LBF_ALL_PINNED;
7942 if (task_running(env->src_rq, p)) {
7943 schedstat_inc(p->stats.nr_failed_migrations_running);
7948 * Aggressive migration if:
7950 * 2) destination numa is preferred
7951 * 3) task is cache cold, or
7952 * 4) too many balance attempts have failed.
7954 if (env->flags & LBF_ACTIVE_LB)
7957 tsk_cache_hot = migrate_degrades_locality(p, env);
7958 if (tsk_cache_hot == -1)
7959 tsk_cache_hot = task_hot(p, env);
7961 if (tsk_cache_hot <= 0 ||
7962 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
7963 if (tsk_cache_hot == 1) {
7964 schedstat_inc(env->sd->lb_hot_gained[env->idle]);
7965 schedstat_inc(p->stats.nr_forced_migrations);
7970 schedstat_inc(p->stats.nr_failed_migrations_hot);
7975 * detach_task() -- detach the task for the migration specified in env
7977 static void detach_task(struct task_struct *p, struct lb_env *env)
7979 lockdep_assert_rq_held(env->src_rq);
7981 deactivate_task(env->src_rq, p, DEQUEUE_NOCLOCK);
7982 set_task_cpu(p, env->dst_cpu);
7986 * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
7987 * part of active balancing operations within "domain".
7989 * Returns a task if successful and NULL otherwise.
7991 static struct task_struct *detach_one_task(struct lb_env *env)
7993 struct task_struct *p;
7995 lockdep_assert_rq_held(env->src_rq);
7997 list_for_each_entry_reverse(p,
7998 &env->src_rq->cfs_tasks, se.group_node) {
7999 if (!can_migrate_task(p, env))
8002 detach_task(p, env);
8005 * Right now, this is only the second place where
8006 * lb_gained[env->idle] is updated (other is detach_tasks)
8007 * so we can safely collect stats here rather than
8008 * inside detach_tasks().
8010 schedstat_inc(env->sd->lb_gained[env->idle]);
8016 static const unsigned int sched_nr_migrate_break = 32;
8019 * detach_tasks() -- tries to detach up to imbalance load/util/tasks from
8020 * busiest_rq, as part of a balancing operation within domain "sd".
8022 * Returns number of detached tasks if successful and 0 otherwise.
8024 static int detach_tasks(struct lb_env *env)
8026 struct list_head *tasks = &env->src_rq->cfs_tasks;
8027 unsigned long util, load;
8028 struct task_struct *p;
8031 lockdep_assert_rq_held(env->src_rq);
8034 * Source run queue has been emptied by another CPU, clear
8035 * LBF_ALL_PINNED flag as we will not test any task.
8037 if (env->src_rq->nr_running <= 1) {
8038 env->flags &= ~LBF_ALL_PINNED;
8042 if (env->imbalance <= 0)
8045 while (!list_empty(tasks)) {
8047 * We don't want to steal all, otherwise we may be treated likewise,
8048 * which could at worst lead to a livelock crash.
8050 if (env->idle != CPU_NOT_IDLE && env->src_rq->nr_running <= 1)
8053 p = list_last_entry(tasks, struct task_struct, se.group_node);
8056 /* We've more or less seen every task there is, call it quits */
8057 if (env->loop > env->loop_max)
8060 /* take a breather every nr_migrate tasks */
8061 if (env->loop > env->loop_break) {
8062 env->loop_break += sched_nr_migrate_break;
8063 env->flags |= LBF_NEED_BREAK;
8067 if (!can_migrate_task(p, env))
8070 switch (env->migration_type) {
8073 * Depending of the number of CPUs and tasks and the
8074 * cgroup hierarchy, task_h_load() can return a null
8075 * value. Make sure that env->imbalance decreases
8076 * otherwise detach_tasks() will stop only after
8077 * detaching up to loop_max tasks.
8079 load = max_t(unsigned long, task_h_load(p), 1);
8081 if (sched_feat(LB_MIN) &&
8082 load < 16 && !env->sd->nr_balance_failed)
8086 * Make sure that we don't migrate too much load.
8087 * Nevertheless, let relax the constraint if
8088 * scheduler fails to find a good waiting task to
8091 if (shr_bound(load, env->sd->nr_balance_failed) > env->imbalance)
8094 env->imbalance -= load;
8098 util = task_util_est(p);
8100 if (util > env->imbalance)
8103 env->imbalance -= util;
8110 case migrate_misfit:
8111 /* This is not a misfit task */
8112 if (task_fits_capacity(p, capacity_of(env->src_cpu)))
8119 detach_task(p, env);
8120 list_add(&p->se.group_node, &env->tasks);
8124 #ifdef CONFIG_PREEMPTION
8126 * NEWIDLE balancing is a source of latency, so preemptible
8127 * kernels will stop after the first task is detached to minimize
8128 * the critical section.
8130 if (env->idle == CPU_NEWLY_IDLE)
8135 * We only want to steal up to the prescribed amount of
8138 if (env->imbalance <= 0)
8143 list_move(&p->se.group_node, tasks);
8147 * Right now, this is one of only two places we collect this stat
8148 * so we can safely collect detach_one_task() stats here rather
8149 * than inside detach_one_task().
8151 schedstat_add(env->sd->lb_gained[env->idle], detached);
8157 * attach_task() -- attach the task detached by detach_task() to its new rq.
8159 static void attach_task(struct rq *rq, struct task_struct *p)
8161 lockdep_assert_rq_held(rq);
8163 WARN_ON_ONCE(task_rq(p) != rq);
8164 activate_task(rq, p, ENQUEUE_NOCLOCK);
8165 check_preempt_curr(rq, p, 0);
8169 * attach_one_task() -- attaches the task returned from detach_one_task() to
8172 static void attach_one_task(struct rq *rq, struct task_struct *p)
8177 update_rq_clock(rq);
8183 * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
8186 static void attach_tasks(struct lb_env *env)
8188 struct list_head *tasks = &env->tasks;
8189 struct task_struct *p;
8192 rq_lock(env->dst_rq, &rf);
8193 update_rq_clock(env->dst_rq);
8195 while (!list_empty(tasks)) {
8196 p = list_first_entry(tasks, struct task_struct, se.group_node);
8197 list_del_init(&p->se.group_node);
8199 attach_task(env->dst_rq, p);
8202 rq_unlock(env->dst_rq, &rf);
8205 #ifdef CONFIG_NO_HZ_COMMON
8206 static inline bool cfs_rq_has_blocked(struct cfs_rq *cfs_rq)
8208 if (cfs_rq->avg.load_avg)
8211 if (cfs_rq->avg.util_avg)
8217 static inline bool others_have_blocked(struct rq *rq)
8219 if (READ_ONCE(rq->avg_rt.util_avg))
8222 if (READ_ONCE(rq->avg_dl.util_avg))
8225 if (thermal_load_avg(rq))
8228 #ifdef CONFIG_HAVE_SCHED_AVG_IRQ
8229 if (READ_ONCE(rq->avg_irq.util_avg))
8236 static inline void update_blocked_load_tick(struct rq *rq)
8238 WRITE_ONCE(rq->last_blocked_load_update_tick, jiffies);
8241 static inline void update_blocked_load_status(struct rq *rq, bool has_blocked)
8244 rq->has_blocked_load = 0;
8247 static inline bool cfs_rq_has_blocked(struct cfs_rq *cfs_rq) { return false; }
8248 static inline bool others_have_blocked(struct rq *rq) { return false; }
8249 static inline void update_blocked_load_tick(struct rq *rq) {}
8250 static inline void update_blocked_load_status(struct rq *rq, bool has_blocked) {}
8253 static bool __update_blocked_others(struct rq *rq, bool *done)
8255 const struct sched_class *curr_class;
8256 u64 now = rq_clock_pelt(rq);
8257 unsigned long thermal_pressure;
8261 * update_load_avg() can call cpufreq_update_util(). Make sure that RT,
8262 * DL and IRQ signals have been updated before updating CFS.
8264 curr_class = rq->curr->sched_class;
8266 thermal_pressure = arch_scale_thermal_pressure(cpu_of(rq));
8268 decayed = update_rt_rq_load_avg(now, rq, curr_class == &rt_sched_class) |
8269 update_dl_rq_load_avg(now, rq, curr_class == &dl_sched_class) |
8270 update_thermal_load_avg(rq_clock_thermal(rq), rq, thermal_pressure) |
8271 update_irq_load_avg(rq, 0);
8273 if (others_have_blocked(rq))
8279 #ifdef CONFIG_FAIR_GROUP_SCHED
8281 static bool __update_blocked_fair(struct rq *rq, bool *done)
8283 struct cfs_rq *cfs_rq, *pos;
8284 bool decayed = false;
8285 int cpu = cpu_of(rq);
8288 * Iterates the task_group tree in a bottom up fashion, see
8289 * list_add_leaf_cfs_rq() for details.
8291 for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) {
8292 struct sched_entity *se;
8294 if (update_cfs_rq_load_avg(cfs_rq_clock_pelt(cfs_rq), cfs_rq)) {
8295 update_tg_load_avg(cfs_rq);
8297 if (cfs_rq->nr_running == 0)
8298 update_idle_cfs_rq_clock_pelt(cfs_rq);
8300 if (cfs_rq == &rq->cfs)
8304 /* Propagate pending load changes to the parent, if any: */
8305 se = cfs_rq->tg->se[cpu];
8306 if (se && !skip_blocked_update(se))
8307 update_load_avg(cfs_rq_of(se), se, UPDATE_TG);
8310 * There can be a lot of idle CPU cgroups. Don't let fully
8311 * decayed cfs_rqs linger on the list.
8313 if (cfs_rq_is_decayed(cfs_rq))
8314 list_del_leaf_cfs_rq(cfs_rq);
8316 /* Don't need periodic decay once load/util_avg are null */
8317 if (cfs_rq_has_blocked(cfs_rq))
8325 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
8326 * This needs to be done in a top-down fashion because the load of a child
8327 * group is a fraction of its parents load.
8329 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
8331 struct rq *rq = rq_of(cfs_rq);
8332 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
8333 unsigned long now = jiffies;
8336 if (cfs_rq->last_h_load_update == now)
8339 WRITE_ONCE(cfs_rq->h_load_next, NULL);
8340 for_each_sched_entity(se) {
8341 cfs_rq = cfs_rq_of(se);
8342 WRITE_ONCE(cfs_rq->h_load_next, se);
8343 if (cfs_rq->last_h_load_update == now)
8348 cfs_rq->h_load = cfs_rq_load_avg(cfs_rq);
8349 cfs_rq->last_h_load_update = now;
8352 while ((se = READ_ONCE(cfs_rq->h_load_next)) != NULL) {
8353 load = cfs_rq->h_load;
8354 load = div64_ul(load * se->avg.load_avg,
8355 cfs_rq_load_avg(cfs_rq) + 1);
8356 cfs_rq = group_cfs_rq(se);
8357 cfs_rq->h_load = load;
8358 cfs_rq->last_h_load_update = now;
8362 static unsigned long task_h_load(struct task_struct *p)
8364 struct cfs_rq *cfs_rq = task_cfs_rq(p);
8366 update_cfs_rq_h_load(cfs_rq);
8367 return div64_ul(p->se.avg.load_avg * cfs_rq->h_load,
8368 cfs_rq_load_avg(cfs_rq) + 1);
8371 static bool __update_blocked_fair(struct rq *rq, bool *done)
8373 struct cfs_rq *cfs_rq = &rq->cfs;
8376 decayed = update_cfs_rq_load_avg(cfs_rq_clock_pelt(cfs_rq), cfs_rq);
8377 if (cfs_rq_has_blocked(cfs_rq))
8383 static unsigned long task_h_load(struct task_struct *p)
8385 return p->se.avg.load_avg;
8389 static void update_blocked_averages(int cpu)
8391 bool decayed = false, done = true;
8392 struct rq *rq = cpu_rq(cpu);
8395 rq_lock_irqsave(rq, &rf);
8396 update_blocked_load_tick(rq);
8397 update_rq_clock(rq);
8399 decayed |= __update_blocked_others(rq, &done);
8400 decayed |= __update_blocked_fair(rq, &done);
8402 update_blocked_load_status(rq, !done);
8404 cpufreq_update_util(rq, 0);
8405 rq_unlock_irqrestore(rq, &rf);
8408 /********** Helpers for find_busiest_group ************************/
8411 * sg_lb_stats - stats of a sched_group required for load_balancing
8413 struct sg_lb_stats {
8414 unsigned long avg_load; /*Avg load across the CPUs of the group */
8415 unsigned long group_load; /* Total load over the CPUs of the group */
8416 unsigned long group_capacity;
8417 unsigned long group_util; /* Total utilization over the CPUs of the group */
8418 unsigned long group_runnable; /* Total runnable time over the CPUs of the group */
8419 unsigned int sum_nr_running; /* Nr of tasks running in the group */
8420 unsigned int sum_h_nr_running; /* Nr of CFS tasks running in the group */
8421 unsigned int idle_cpus;
8422 unsigned int group_weight;
8423 enum group_type group_type;
8424 unsigned int group_asym_packing; /* Tasks should be moved to preferred CPU */
8425 unsigned long group_misfit_task_load; /* A CPU has a task too big for its capacity */
8426 #ifdef CONFIG_NUMA_BALANCING
8427 unsigned int nr_numa_running;
8428 unsigned int nr_preferred_running;
8433 * sd_lb_stats - Structure to store the statistics of a sched_domain
8434 * during load balancing.
8436 struct sd_lb_stats {
8437 struct sched_group *busiest; /* Busiest group in this sd */
8438 struct sched_group *local; /* Local group in this sd */
8439 unsigned long total_load; /* Total load of all groups in sd */
8440 unsigned long total_capacity; /* Total capacity of all groups in sd */
8441 unsigned long avg_load; /* Average load across all groups in sd */
8442 unsigned int prefer_sibling; /* tasks should go to sibling first */
8444 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
8445 struct sg_lb_stats local_stat; /* Statistics of the local group */
8448 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
8451 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
8452 * local_stat because update_sg_lb_stats() does a full clear/assignment.
8453 * We must however set busiest_stat::group_type and
8454 * busiest_stat::idle_cpus to the worst busiest group because
8455 * update_sd_pick_busiest() reads these before assignment.
8457 *sds = (struct sd_lb_stats){
8461 .total_capacity = 0UL,
8463 .idle_cpus = UINT_MAX,
8464 .group_type = group_has_spare,
8469 static unsigned long scale_rt_capacity(int cpu)
8471 struct rq *rq = cpu_rq(cpu);
8472 unsigned long max = arch_scale_cpu_capacity(cpu);
8473 unsigned long used, free;
8476 irq = cpu_util_irq(rq);
8478 if (unlikely(irq >= max))
8482 * avg_rt.util_avg and avg_dl.util_avg track binary signals
8483 * (running and not running) with weights 0 and 1024 respectively.
8484 * avg_thermal.load_avg tracks thermal pressure and the weighted
8485 * average uses the actual delta max capacity(load).
8487 used = READ_ONCE(rq->avg_rt.util_avg);
8488 used += READ_ONCE(rq->avg_dl.util_avg);
8489 used += thermal_load_avg(rq);
8491 if (unlikely(used >= max))
8496 return scale_irq_capacity(free, irq, max);
8499 static void update_cpu_capacity(struct sched_domain *sd, int cpu)
8501 unsigned long capacity = scale_rt_capacity(cpu);
8502 struct sched_group *sdg = sd->groups;
8504 cpu_rq(cpu)->cpu_capacity_orig = arch_scale_cpu_capacity(cpu);
8509 cpu_rq(cpu)->cpu_capacity = capacity;
8510 trace_sched_cpu_capacity_tp(cpu_rq(cpu));
8512 sdg->sgc->capacity = capacity;
8513 sdg->sgc->min_capacity = capacity;
8514 sdg->sgc->max_capacity = capacity;
8517 void update_group_capacity(struct sched_domain *sd, int cpu)
8519 struct sched_domain *child = sd->child;
8520 struct sched_group *group, *sdg = sd->groups;
8521 unsigned long capacity, min_capacity, max_capacity;
8522 unsigned long interval;
8524 interval = msecs_to_jiffies(sd->balance_interval);
8525 interval = clamp(interval, 1UL, max_load_balance_interval);
8526 sdg->sgc->next_update = jiffies + interval;
8529 update_cpu_capacity(sd, cpu);
8534 min_capacity = ULONG_MAX;
8537 if (child->flags & SD_OVERLAP) {
8539 * SD_OVERLAP domains cannot assume that child groups
8540 * span the current group.
8543 for_each_cpu(cpu, sched_group_span(sdg)) {
8544 unsigned long cpu_cap = capacity_of(cpu);
8546 capacity += cpu_cap;
8547 min_capacity = min(cpu_cap, min_capacity);
8548 max_capacity = max(cpu_cap, max_capacity);
8552 * !SD_OVERLAP domains can assume that child groups
8553 * span the current group.
8556 group = child->groups;
8558 struct sched_group_capacity *sgc = group->sgc;
8560 capacity += sgc->capacity;
8561 min_capacity = min(sgc->min_capacity, min_capacity);
8562 max_capacity = max(sgc->max_capacity, max_capacity);
8563 group = group->next;
8564 } while (group != child->groups);
8567 sdg->sgc->capacity = capacity;
8568 sdg->sgc->min_capacity = min_capacity;
8569 sdg->sgc->max_capacity = max_capacity;
8573 * Check whether the capacity of the rq has been noticeably reduced by side
8574 * activity. The imbalance_pct is used for the threshold.
8575 * Return true is the capacity is reduced
8578 check_cpu_capacity(struct rq *rq, struct sched_domain *sd)
8580 return ((rq->cpu_capacity * sd->imbalance_pct) <
8581 (rq->cpu_capacity_orig * 100));
8585 * Check whether a rq has a misfit task and if it looks like we can actually
8586 * help that task: we can migrate the task to a CPU of higher capacity, or
8587 * the task's current CPU is heavily pressured.
8589 static inline int check_misfit_status(struct rq *rq, struct sched_domain *sd)
8591 return rq->misfit_task_load &&
8592 (rq->cpu_capacity_orig < rq->rd->max_cpu_capacity ||
8593 check_cpu_capacity(rq, sd));
8597 * Group imbalance indicates (and tries to solve) the problem where balancing
8598 * groups is inadequate due to ->cpus_ptr constraints.
8600 * Imagine a situation of two groups of 4 CPUs each and 4 tasks each with a
8601 * cpumask covering 1 CPU of the first group and 3 CPUs of the second group.
8604 * { 0 1 2 3 } { 4 5 6 7 }
8607 * If we were to balance group-wise we'd place two tasks in the first group and
8608 * two tasks in the second group. Clearly this is undesired as it will overload
8609 * cpu 3 and leave one of the CPUs in the second group unused.
8611 * The current solution to this issue is detecting the skew in the first group
8612 * by noticing the lower domain failed to reach balance and had difficulty
8613 * moving tasks due to affinity constraints.
8615 * When this is so detected; this group becomes a candidate for busiest; see
8616 * update_sd_pick_busiest(). And calculate_imbalance() and
8617 * find_busiest_group() avoid some of the usual balance conditions to allow it
8618 * to create an effective group imbalance.
8620 * This is a somewhat tricky proposition since the next run might not find the
8621 * group imbalance and decide the groups need to be balanced again. A most
8622 * subtle and fragile situation.
8625 static inline int sg_imbalanced(struct sched_group *group)
8627 return group->sgc->imbalance;
8631 * group_has_capacity returns true if the group has spare capacity that could
8632 * be used by some tasks.
8633 * We consider that a group has spare capacity if the number of task is
8634 * smaller than the number of CPUs or if the utilization is lower than the
8635 * available capacity for CFS tasks.
8636 * For the latter, we use a threshold to stabilize the state, to take into
8637 * account the variance of the tasks' load and to return true if the available
8638 * capacity in meaningful for the load balancer.
8639 * As an example, an available capacity of 1% can appear but it doesn't make
8640 * any benefit for the load balance.
8643 group_has_capacity(unsigned int imbalance_pct, struct sg_lb_stats *sgs)
8645 if (sgs->sum_nr_running < sgs->group_weight)
8648 if ((sgs->group_capacity * imbalance_pct) <
8649 (sgs->group_runnable * 100))
8652 if ((sgs->group_capacity * 100) >
8653 (sgs->group_util * imbalance_pct))
8660 * group_is_overloaded returns true if the group has more tasks than it can
8662 * group_is_overloaded is not equals to !group_has_capacity because a group
8663 * with the exact right number of tasks, has no more spare capacity but is not
8664 * overloaded so both group_has_capacity and group_is_overloaded return
8668 group_is_overloaded(unsigned int imbalance_pct, struct sg_lb_stats *sgs)
8670 if (sgs->sum_nr_running <= sgs->group_weight)
8673 if ((sgs->group_capacity * 100) <
8674 (sgs->group_util * imbalance_pct))
8677 if ((sgs->group_capacity * imbalance_pct) <
8678 (sgs->group_runnable * 100))
8685 group_type group_classify(unsigned int imbalance_pct,
8686 struct sched_group *group,
8687 struct sg_lb_stats *sgs)
8689 if (group_is_overloaded(imbalance_pct, sgs))
8690 return group_overloaded;
8692 if (sg_imbalanced(group))
8693 return group_imbalanced;
8695 if (sgs->group_asym_packing)
8696 return group_asym_packing;
8698 if (sgs->group_misfit_task_load)
8699 return group_misfit_task;
8701 if (!group_has_capacity(imbalance_pct, sgs))
8702 return group_fully_busy;
8704 return group_has_spare;
8708 * asym_smt_can_pull_tasks - Check whether the load balancing CPU can pull tasks
8709 * @dst_cpu: Destination CPU of the load balancing
8710 * @sds: Load-balancing data with statistics of the local group
8711 * @sgs: Load-balancing statistics of the candidate busiest group
8712 * @sg: The candidate busiest group
8714 * Check the state of the SMT siblings of both @sds::local and @sg and decide
8715 * if @dst_cpu can pull tasks.
8717 * If @dst_cpu does not have SMT siblings, it can pull tasks if two or more of
8718 * the SMT siblings of @sg are busy. If only one CPU in @sg is busy, pull tasks
8719 * only if @dst_cpu has higher priority.
8721 * If both @dst_cpu and @sg have SMT siblings, and @sg has exactly one more
8722 * busy CPU than @sds::local, let @dst_cpu pull tasks if it has higher priority.
8723 * Bigger imbalances in the number of busy CPUs will be dealt with in
8724 * update_sd_pick_busiest().
8726 * If @sg does not have SMT siblings, only pull tasks if all of the SMT siblings
8727 * of @dst_cpu are idle and @sg has lower priority.
8729 * Return: true if @dst_cpu can pull tasks, false otherwise.
8731 static bool asym_smt_can_pull_tasks(int dst_cpu, struct sd_lb_stats *sds,
8732 struct sg_lb_stats *sgs,
8733 struct sched_group *sg)
8735 #ifdef CONFIG_SCHED_SMT
8736 bool local_is_smt, sg_is_smt;
8739 local_is_smt = sds->local->flags & SD_SHARE_CPUCAPACITY;
8740 sg_is_smt = sg->flags & SD_SHARE_CPUCAPACITY;
8742 sg_busy_cpus = sgs->group_weight - sgs->idle_cpus;
8744 if (!local_is_smt) {
8746 * If we are here, @dst_cpu is idle and does not have SMT
8747 * siblings. Pull tasks if candidate group has two or more
8750 if (sg_busy_cpus >= 2) /* implies sg_is_smt */
8754 * @dst_cpu does not have SMT siblings. @sg may have SMT
8755 * siblings and only one is busy. In such case, @dst_cpu
8756 * can help if it has higher priority and is idle (i.e.,
8757 * it has no running tasks).
8759 return sched_asym_prefer(dst_cpu, sg->asym_prefer_cpu);
8762 /* @dst_cpu has SMT siblings. */
8765 int local_busy_cpus = sds->local->group_weight -
8766 sds->local_stat.idle_cpus;
8767 int busy_cpus_delta = sg_busy_cpus - local_busy_cpus;
8769 if (busy_cpus_delta == 1)
8770 return sched_asym_prefer(dst_cpu, sg->asym_prefer_cpu);
8776 * @sg does not have SMT siblings. Ensure that @sds::local does not end
8777 * up with more than one busy SMT sibling and only pull tasks if there
8778 * are not busy CPUs (i.e., no CPU has running tasks).
8780 if (!sds->local_stat.sum_nr_running)
8781 return sched_asym_prefer(dst_cpu, sg->asym_prefer_cpu);
8785 /* Always return false so that callers deal with non-SMT cases. */
8791 sched_asym(struct lb_env *env, struct sd_lb_stats *sds, struct sg_lb_stats *sgs,
8792 struct sched_group *group)
8794 /* Only do SMT checks if either local or candidate have SMT siblings */
8795 if ((sds->local->flags & SD_SHARE_CPUCAPACITY) ||
8796 (group->flags & SD_SHARE_CPUCAPACITY))
8797 return asym_smt_can_pull_tasks(env->dst_cpu, sds, sgs, group);
8799 return sched_asym_prefer(env->dst_cpu, group->asym_prefer_cpu);
8803 sched_reduced_capacity(struct rq *rq, struct sched_domain *sd)
8806 * When there is more than 1 task, the group_overloaded case already
8807 * takes care of cpu with reduced capacity
8809 if (rq->cfs.h_nr_running != 1)
8812 return check_cpu_capacity(rq, sd);
8816 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
8817 * @env: The load balancing environment.
8818 * @sds: Load-balancing data with statistics of the local group.
8819 * @group: sched_group whose statistics are to be updated.
8820 * @sgs: variable to hold the statistics for this group.
8821 * @sg_status: Holds flag indicating the status of the sched_group
8823 static inline void update_sg_lb_stats(struct lb_env *env,
8824 struct sd_lb_stats *sds,
8825 struct sched_group *group,
8826 struct sg_lb_stats *sgs,
8829 int i, nr_running, local_group;
8831 memset(sgs, 0, sizeof(*sgs));
8833 local_group = group == sds->local;
8835 for_each_cpu_and(i, sched_group_span(group), env->cpus) {
8836 struct rq *rq = cpu_rq(i);
8837 unsigned long load = cpu_load(rq);
8839 sgs->group_load += load;
8840 sgs->group_util += cpu_util_cfs(i);
8841 sgs->group_runnable += cpu_runnable(rq);
8842 sgs->sum_h_nr_running += rq->cfs.h_nr_running;
8844 nr_running = rq->nr_running;
8845 sgs->sum_nr_running += nr_running;
8848 *sg_status |= SG_OVERLOAD;
8850 if (cpu_overutilized(i))
8851 *sg_status |= SG_OVERUTILIZED;
8853 #ifdef CONFIG_NUMA_BALANCING
8854 sgs->nr_numa_running += rq->nr_numa_running;
8855 sgs->nr_preferred_running += rq->nr_preferred_running;
8858 * No need to call idle_cpu() if nr_running is not 0
8860 if (!nr_running && idle_cpu(i)) {
8862 /* Idle cpu can't have misfit task */
8869 if (env->sd->flags & SD_ASYM_CPUCAPACITY) {
8870 /* Check for a misfit task on the cpu */
8871 if (sgs->group_misfit_task_load < rq->misfit_task_load) {
8872 sgs->group_misfit_task_load = rq->misfit_task_load;
8873 *sg_status |= SG_OVERLOAD;
8875 } else if ((env->idle != CPU_NOT_IDLE) &&
8876 sched_reduced_capacity(rq, env->sd)) {
8877 /* Check for a task running on a CPU with reduced capacity */
8878 if (sgs->group_misfit_task_load < load)
8879 sgs->group_misfit_task_load = load;
8883 sgs->group_capacity = group->sgc->capacity;
8885 sgs->group_weight = group->group_weight;
8887 /* Check if dst CPU is idle and preferred to this group */
8888 if (!local_group && env->sd->flags & SD_ASYM_PACKING &&
8889 env->idle != CPU_NOT_IDLE && sgs->sum_h_nr_running &&
8890 sched_asym(env, sds, sgs, group)) {
8891 sgs->group_asym_packing = 1;
8894 sgs->group_type = group_classify(env->sd->imbalance_pct, group, sgs);
8896 /* Computing avg_load makes sense only when group is overloaded */
8897 if (sgs->group_type == group_overloaded)
8898 sgs->avg_load = (sgs->group_load * SCHED_CAPACITY_SCALE) /
8899 sgs->group_capacity;
8903 * update_sd_pick_busiest - return 1 on busiest group
8904 * @env: The load balancing environment.
8905 * @sds: sched_domain statistics
8906 * @sg: sched_group candidate to be checked for being the busiest
8907 * @sgs: sched_group statistics
8909 * Determine if @sg is a busier group than the previously selected
8912 * Return: %true if @sg is a busier group than the previously selected
8913 * busiest group. %false otherwise.
8915 static bool update_sd_pick_busiest(struct lb_env *env,
8916 struct sd_lb_stats *sds,
8917 struct sched_group *sg,
8918 struct sg_lb_stats *sgs)
8920 struct sg_lb_stats *busiest = &sds->busiest_stat;
8922 /* Make sure that there is at least one task to pull */
8923 if (!sgs->sum_h_nr_running)
8927 * Don't try to pull misfit tasks we can't help.
8928 * We can use max_capacity here as reduction in capacity on some
8929 * CPUs in the group should either be possible to resolve
8930 * internally or be covered by avg_load imbalance (eventually).
8932 if ((env->sd->flags & SD_ASYM_CPUCAPACITY) &&
8933 (sgs->group_type == group_misfit_task) &&
8934 (!capacity_greater(capacity_of(env->dst_cpu), sg->sgc->max_capacity) ||
8935 sds->local_stat.group_type != group_has_spare))
8938 if (sgs->group_type > busiest->group_type)
8941 if (sgs->group_type < busiest->group_type)
8945 * The candidate and the current busiest group are the same type of
8946 * group. Let check which one is the busiest according to the type.
8949 switch (sgs->group_type) {
8950 case group_overloaded:
8951 /* Select the overloaded group with highest avg_load. */
8952 if (sgs->avg_load <= busiest->avg_load)
8956 case group_imbalanced:
8958 * Select the 1st imbalanced group as we don't have any way to
8959 * choose one more than another.
8963 case group_asym_packing:
8964 /* Prefer to move from lowest priority CPU's work */
8965 if (sched_asym_prefer(sg->asym_prefer_cpu, sds->busiest->asym_prefer_cpu))
8969 case group_misfit_task:
8971 * If we have more than one misfit sg go with the biggest
8974 if (sgs->group_misfit_task_load < busiest->group_misfit_task_load)
8978 case group_fully_busy:
8980 * Select the fully busy group with highest avg_load. In
8981 * theory, there is no need to pull task from such kind of
8982 * group because tasks have all compute capacity that they need
8983 * but we can still improve the overall throughput by reducing
8984 * contention when accessing shared HW resources.
8986 * XXX for now avg_load is not computed and always 0 so we
8987 * select the 1st one.
8989 if (sgs->avg_load <= busiest->avg_load)
8993 case group_has_spare:
8995 * Select not overloaded group with lowest number of idle cpus
8996 * and highest number of running tasks. We could also compare
8997 * the spare capacity which is more stable but it can end up
8998 * that the group has less spare capacity but finally more idle
8999 * CPUs which means less opportunity to pull tasks.
9001 if (sgs->idle_cpus > busiest->idle_cpus)
9003 else if ((sgs->idle_cpus == busiest->idle_cpus) &&
9004 (sgs->sum_nr_running <= busiest->sum_nr_running))
9011 * Candidate sg has no more than one task per CPU and has higher
9012 * per-CPU capacity. Migrating tasks to less capable CPUs may harm
9013 * throughput. Maximize throughput, power/energy consequences are not
9016 if ((env->sd->flags & SD_ASYM_CPUCAPACITY) &&
9017 (sgs->group_type <= group_fully_busy) &&
9018 (capacity_greater(sg->sgc->min_capacity, capacity_of(env->dst_cpu))))
9024 #ifdef CONFIG_NUMA_BALANCING
9025 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
9027 if (sgs->sum_h_nr_running > sgs->nr_numa_running)
9029 if (sgs->sum_h_nr_running > sgs->nr_preferred_running)
9034 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
9036 if (rq->nr_running > rq->nr_numa_running)
9038 if (rq->nr_running > rq->nr_preferred_running)
9043 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
9048 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
9052 #endif /* CONFIG_NUMA_BALANCING */
9058 * task_running_on_cpu - return 1 if @p is running on @cpu.
9061 static unsigned int task_running_on_cpu(int cpu, struct task_struct *p)
9063 /* Task has no contribution or is new */
9064 if (cpu != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time))
9067 if (task_on_rq_queued(p))
9074 * idle_cpu_without - would a given CPU be idle without p ?
9075 * @cpu: the processor on which idleness is tested.
9076 * @p: task which should be ignored.
9078 * Return: 1 if the CPU would be idle. 0 otherwise.
9080 static int idle_cpu_without(int cpu, struct task_struct *p)
9082 struct rq *rq = cpu_rq(cpu);
9084 if (rq->curr != rq->idle && rq->curr != p)
9088 * rq->nr_running can't be used but an updated version without the
9089 * impact of p on cpu must be used instead. The updated nr_running
9090 * be computed and tested before calling idle_cpu_without().
9094 if (rq->ttwu_pending)
9102 * update_sg_wakeup_stats - Update sched_group's statistics for wakeup.
9103 * @sd: The sched_domain level to look for idlest group.
9104 * @group: sched_group whose statistics are to be updated.
9105 * @sgs: variable to hold the statistics for this group.
9106 * @p: The task for which we look for the idlest group/CPU.
9108 static inline void update_sg_wakeup_stats(struct sched_domain *sd,
9109 struct sched_group *group,
9110 struct sg_lb_stats *sgs,
9111 struct task_struct *p)
9115 memset(sgs, 0, sizeof(*sgs));
9117 for_each_cpu(i, sched_group_span(group)) {
9118 struct rq *rq = cpu_rq(i);
9121 sgs->group_load += cpu_load_without(rq, p);
9122 sgs->group_util += cpu_util_without(i, p);
9123 sgs->group_runnable += cpu_runnable_without(rq, p);
9124 local = task_running_on_cpu(i, p);
9125 sgs->sum_h_nr_running += rq->cfs.h_nr_running - local;
9127 nr_running = rq->nr_running - local;
9128 sgs->sum_nr_running += nr_running;
9131 * No need to call idle_cpu_without() if nr_running is not 0
9133 if (!nr_running && idle_cpu_without(i, p))
9138 /* Check if task fits in the group */
9139 if (sd->flags & SD_ASYM_CPUCAPACITY &&
9140 !task_fits_capacity(p, group->sgc->max_capacity)) {
9141 sgs->group_misfit_task_load = 1;
9144 sgs->group_capacity = group->sgc->capacity;
9146 sgs->group_weight = group->group_weight;
9148 sgs->group_type = group_classify(sd->imbalance_pct, group, sgs);
9151 * Computing avg_load makes sense only when group is fully busy or
9154 if (sgs->group_type == group_fully_busy ||
9155 sgs->group_type == group_overloaded)
9156 sgs->avg_load = (sgs->group_load * SCHED_CAPACITY_SCALE) /
9157 sgs->group_capacity;
9160 static bool update_pick_idlest(struct sched_group *idlest,
9161 struct sg_lb_stats *idlest_sgs,
9162 struct sched_group *group,
9163 struct sg_lb_stats *sgs)
9165 if (sgs->group_type < idlest_sgs->group_type)
9168 if (sgs->group_type > idlest_sgs->group_type)
9172 * The candidate and the current idlest group are the same type of
9173 * group. Let check which one is the idlest according to the type.
9176 switch (sgs->group_type) {
9177 case group_overloaded:
9178 case group_fully_busy:
9179 /* Select the group with lowest avg_load. */
9180 if (idlest_sgs->avg_load <= sgs->avg_load)
9184 case group_imbalanced:
9185 case group_asym_packing:
9186 /* Those types are not used in the slow wakeup path */
9189 case group_misfit_task:
9190 /* Select group with the highest max capacity */
9191 if (idlest->sgc->max_capacity >= group->sgc->max_capacity)
9195 case group_has_spare:
9196 /* Select group with most idle CPUs */
9197 if (idlest_sgs->idle_cpus > sgs->idle_cpus)
9200 /* Select group with lowest group_util */
9201 if (idlest_sgs->idle_cpus == sgs->idle_cpus &&
9202 idlest_sgs->group_util <= sgs->group_util)
9212 * find_idlest_group() finds and returns the least busy CPU group within the
9215 * Assumes p is allowed on at least one CPU in sd.
9217 static struct sched_group *
9218 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
9220 struct sched_group *idlest = NULL, *local = NULL, *group = sd->groups;
9221 struct sg_lb_stats local_sgs, tmp_sgs;
9222 struct sg_lb_stats *sgs;
9223 unsigned long imbalance;
9224 struct sg_lb_stats idlest_sgs = {
9225 .avg_load = UINT_MAX,
9226 .group_type = group_overloaded,
9232 /* Skip over this group if it has no CPUs allowed */
9233 if (!cpumask_intersects(sched_group_span(group),
9237 /* Skip over this group if no cookie matched */
9238 if (!sched_group_cookie_match(cpu_rq(this_cpu), p, group))
9241 local_group = cpumask_test_cpu(this_cpu,
9242 sched_group_span(group));
9251 update_sg_wakeup_stats(sd, group, sgs, p);
9253 if (!local_group && update_pick_idlest(idlest, &idlest_sgs, group, sgs)) {
9258 } while (group = group->next, group != sd->groups);
9261 /* There is no idlest group to push tasks to */
9265 /* The local group has been skipped because of CPU affinity */
9270 * If the local group is idler than the selected idlest group
9271 * don't try and push the task.
9273 if (local_sgs.group_type < idlest_sgs.group_type)
9277 * If the local group is busier than the selected idlest group
9278 * try and push the task.
9280 if (local_sgs.group_type > idlest_sgs.group_type)
9283 switch (local_sgs.group_type) {
9284 case group_overloaded:
9285 case group_fully_busy:
9287 /* Calculate allowed imbalance based on load */
9288 imbalance = scale_load_down(NICE_0_LOAD) *
9289 (sd->imbalance_pct-100) / 100;
9292 * When comparing groups across NUMA domains, it's possible for
9293 * the local domain to be very lightly loaded relative to the
9294 * remote domains but "imbalance" skews the comparison making
9295 * remote CPUs look much more favourable. When considering
9296 * cross-domain, add imbalance to the load on the remote node
9297 * and consider staying local.
9300 if ((sd->flags & SD_NUMA) &&
9301 ((idlest_sgs.avg_load + imbalance) >= local_sgs.avg_load))
9305 * If the local group is less loaded than the selected
9306 * idlest group don't try and push any tasks.
9308 if (idlest_sgs.avg_load >= (local_sgs.avg_load + imbalance))
9311 if (100 * local_sgs.avg_load <= sd->imbalance_pct * idlest_sgs.avg_load)
9315 case group_imbalanced:
9316 case group_asym_packing:
9317 /* Those type are not used in the slow wakeup path */
9320 case group_misfit_task:
9321 /* Select group with the highest max capacity */
9322 if (local->sgc->max_capacity >= idlest->sgc->max_capacity)
9326 case group_has_spare:
9328 if (sd->flags & SD_NUMA) {
9329 int imb_numa_nr = sd->imb_numa_nr;
9330 #ifdef CONFIG_NUMA_BALANCING
9333 * If there is spare capacity at NUMA, try to select
9334 * the preferred node
9336 if (cpu_to_node(this_cpu) == p->numa_preferred_nid)
9339 idlest_cpu = cpumask_first(sched_group_span(idlest));
9340 if (cpu_to_node(idlest_cpu) == p->numa_preferred_nid)
9342 #endif /* CONFIG_NUMA_BALANCING */
9344 * Otherwise, keep the task close to the wakeup source
9345 * and improve locality if the number of running tasks
9346 * would remain below threshold where an imbalance is
9347 * allowed while accounting for the possibility the
9348 * task is pinned to a subset of CPUs. If there is a
9349 * real need of migration, periodic load balance will
9352 if (p->nr_cpus_allowed != NR_CPUS) {
9353 struct cpumask *cpus = this_cpu_cpumask_var_ptr(select_rq_mask);
9355 cpumask_and(cpus, sched_group_span(local), p->cpus_ptr);
9356 imb_numa_nr = min(cpumask_weight(cpus), sd->imb_numa_nr);
9359 imbalance = abs(local_sgs.idle_cpus - idlest_sgs.idle_cpus);
9360 if (!adjust_numa_imbalance(imbalance,
9361 local_sgs.sum_nr_running + 1,
9366 #endif /* CONFIG_NUMA */
9369 * Select group with highest number of idle CPUs. We could also
9370 * compare the utilization which is more stable but it can end
9371 * up that the group has less spare capacity but finally more
9372 * idle CPUs which means more opportunity to run task.
9374 if (local_sgs.idle_cpus >= idlest_sgs.idle_cpus)
9382 static void update_idle_cpu_scan(struct lb_env *env,
9383 unsigned long sum_util)
9385 struct sched_domain_shared *sd_share;
9386 int llc_weight, pct;
9389 * Update the number of CPUs to scan in LLC domain, which could
9390 * be used as a hint in select_idle_cpu(). The update of sd_share
9391 * could be expensive because it is within a shared cache line.
9392 * So the write of this hint only occurs during periodic load
9393 * balancing, rather than CPU_NEWLY_IDLE, because the latter
9394 * can fire way more frequently than the former.
9396 if (!sched_feat(SIS_UTIL) || env->idle == CPU_NEWLY_IDLE)
9399 llc_weight = per_cpu(sd_llc_size, env->dst_cpu);
9400 if (env->sd->span_weight != llc_weight)
9403 sd_share = rcu_dereference(per_cpu(sd_llc_shared, env->dst_cpu));
9408 * The number of CPUs to search drops as sum_util increases, when
9409 * sum_util hits 85% or above, the scan stops.
9410 * The reason to choose 85% as the threshold is because this is the
9411 * imbalance_pct(117) when a LLC sched group is overloaded.
9413 * let y = SCHED_CAPACITY_SCALE - p * x^2 [1]
9414 * and y'= y / SCHED_CAPACITY_SCALE
9416 * x is the ratio of sum_util compared to the CPU capacity:
9417 * x = sum_util / (llc_weight * SCHED_CAPACITY_SCALE)
9418 * y' is the ratio of CPUs to be scanned in the LLC domain,
9419 * and the number of CPUs to scan is calculated by:
9421 * nr_scan = llc_weight * y' [2]
9423 * When x hits the threshold of overloaded, AKA, when
9424 * x = 100 / pct, y drops to 0. According to [1],
9425 * p should be SCHED_CAPACITY_SCALE * pct^2 / 10000
9427 * Scale x by SCHED_CAPACITY_SCALE:
9428 * x' = sum_util / llc_weight; [3]
9430 * and finally [1] becomes:
9431 * y = SCHED_CAPACITY_SCALE -
9432 * x'^2 * pct^2 / (10000 * SCHED_CAPACITY_SCALE) [4]
9437 do_div(x, llc_weight);
9440 pct = env->sd->imbalance_pct;
9441 tmp = x * x * pct * pct;
9442 do_div(tmp, 10000 * SCHED_CAPACITY_SCALE);
9443 tmp = min_t(long, tmp, SCHED_CAPACITY_SCALE);
9444 y = SCHED_CAPACITY_SCALE - tmp;
9448 do_div(y, SCHED_CAPACITY_SCALE);
9449 if ((int)y != sd_share->nr_idle_scan)
9450 WRITE_ONCE(sd_share->nr_idle_scan, (int)y);
9454 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
9455 * @env: The load balancing environment.
9456 * @sds: variable to hold the statistics for this sched_domain.
9459 static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
9461 struct sched_domain *child = env->sd->child;
9462 struct sched_group *sg = env->sd->groups;
9463 struct sg_lb_stats *local = &sds->local_stat;
9464 struct sg_lb_stats tmp_sgs;
9465 unsigned long sum_util = 0;
9469 struct sg_lb_stats *sgs = &tmp_sgs;
9472 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_span(sg));
9477 if (env->idle != CPU_NEWLY_IDLE ||
9478 time_after_eq(jiffies, sg->sgc->next_update))
9479 update_group_capacity(env->sd, env->dst_cpu);
9482 update_sg_lb_stats(env, sds, sg, sgs, &sg_status);
9488 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
9490 sds->busiest_stat = *sgs;
9494 /* Now, start updating sd_lb_stats */
9495 sds->total_load += sgs->group_load;
9496 sds->total_capacity += sgs->group_capacity;
9498 sum_util += sgs->group_util;
9500 } while (sg != env->sd->groups);
9502 /* Tag domain that child domain prefers tasks go to siblings first */
9503 sds->prefer_sibling = child && child->flags & SD_PREFER_SIBLING;
9506 if (env->sd->flags & SD_NUMA)
9507 env->fbq_type = fbq_classify_group(&sds->busiest_stat);
9509 if (!env->sd->parent) {
9510 struct root_domain *rd = env->dst_rq->rd;
9512 /* update overload indicator if we are at root domain */
9513 WRITE_ONCE(rd->overload, sg_status & SG_OVERLOAD);
9515 /* Update over-utilization (tipping point, U >= 0) indicator */
9516 WRITE_ONCE(rd->overutilized, sg_status & SG_OVERUTILIZED);
9517 trace_sched_overutilized_tp(rd, sg_status & SG_OVERUTILIZED);
9518 } else if (sg_status & SG_OVERUTILIZED) {
9519 struct root_domain *rd = env->dst_rq->rd;
9521 WRITE_ONCE(rd->overutilized, SG_OVERUTILIZED);
9522 trace_sched_overutilized_tp(rd, SG_OVERUTILIZED);
9525 update_idle_cpu_scan(env, sum_util);
9529 * calculate_imbalance - Calculate the amount of imbalance present within the
9530 * groups of a given sched_domain during load balance.
9531 * @env: load balance environment
9532 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
9534 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
9536 struct sg_lb_stats *local, *busiest;
9538 local = &sds->local_stat;
9539 busiest = &sds->busiest_stat;
9541 if (busiest->group_type == group_misfit_task) {
9542 if (env->sd->flags & SD_ASYM_CPUCAPACITY) {
9543 /* Set imbalance to allow misfit tasks to be balanced. */
9544 env->migration_type = migrate_misfit;
9548 * Set load imbalance to allow moving task from cpu
9549 * with reduced capacity.
9551 env->migration_type = migrate_load;
9552 env->imbalance = busiest->group_misfit_task_load;
9557 if (busiest->group_type == group_asym_packing) {
9559 * In case of asym capacity, we will try to migrate all load to
9560 * the preferred CPU.
9562 env->migration_type = migrate_task;
9563 env->imbalance = busiest->sum_h_nr_running;
9567 if (busiest->group_type == group_imbalanced) {
9569 * In the group_imb case we cannot rely on group-wide averages
9570 * to ensure CPU-load equilibrium, try to move any task to fix
9571 * the imbalance. The next load balance will take care of
9572 * balancing back the system.
9574 env->migration_type = migrate_task;
9580 * Try to use spare capacity of local group without overloading it or
9583 if (local->group_type == group_has_spare) {
9584 if ((busiest->group_type > group_fully_busy) &&
9585 !(env->sd->flags & SD_SHARE_PKG_RESOURCES)) {
9587 * If busiest is overloaded, try to fill spare
9588 * capacity. This might end up creating spare capacity
9589 * in busiest or busiest still being overloaded but
9590 * there is no simple way to directly compute the
9591 * amount of load to migrate in order to balance the
9594 env->migration_type = migrate_util;
9595 env->imbalance = max(local->group_capacity, local->group_util) -
9599 * In some cases, the group's utilization is max or even
9600 * higher than capacity because of migrations but the
9601 * local CPU is (newly) idle. There is at least one
9602 * waiting task in this overloaded busiest group. Let's
9605 if (env->idle != CPU_NOT_IDLE && env->imbalance == 0) {
9606 env->migration_type = migrate_task;
9613 if (busiest->group_weight == 1 || sds->prefer_sibling) {
9614 unsigned int nr_diff = busiest->sum_nr_running;
9616 * When prefer sibling, evenly spread running tasks on
9619 env->migration_type = migrate_task;
9620 lsub_positive(&nr_diff, local->sum_nr_running);
9621 env->imbalance = nr_diff;
9625 * If there is no overload, we just want to even the number of
9628 env->migration_type = migrate_task;
9629 env->imbalance = max_t(long, 0,
9630 (local->idle_cpus - busiest->idle_cpus));
9634 /* Consider allowing a small imbalance between NUMA groups */
9635 if (env->sd->flags & SD_NUMA) {
9636 env->imbalance = adjust_numa_imbalance(env->imbalance,
9637 local->sum_nr_running + 1,
9638 env->sd->imb_numa_nr);
9642 /* Number of tasks to move to restore balance */
9643 env->imbalance >>= 1;
9649 * Local is fully busy but has to take more load to relieve the
9652 if (local->group_type < group_overloaded) {
9654 * Local will become overloaded so the avg_load metrics are
9658 local->avg_load = (local->group_load * SCHED_CAPACITY_SCALE) /
9659 local->group_capacity;
9662 * If the local group is more loaded than the selected
9663 * busiest group don't try to pull any tasks.
9665 if (local->avg_load >= busiest->avg_load) {
9670 sds->avg_load = (sds->total_load * SCHED_CAPACITY_SCALE) /
9671 sds->total_capacity;
9675 * Both group are or will become overloaded and we're trying to get all
9676 * the CPUs to the average_load, so we don't want to push ourselves
9677 * above the average load, nor do we wish to reduce the max loaded CPU
9678 * below the average load. At the same time, we also don't want to
9679 * reduce the group load below the group capacity. Thus we look for
9680 * the minimum possible imbalance.
9682 env->migration_type = migrate_load;
9683 env->imbalance = min(
9684 (busiest->avg_load - sds->avg_load) * busiest->group_capacity,
9685 (sds->avg_load - local->avg_load) * local->group_capacity
9686 ) / SCHED_CAPACITY_SCALE;
9689 /******* find_busiest_group() helpers end here *********************/
9692 * Decision matrix according to the local and busiest group type:
9694 * busiest \ local has_spare fully_busy misfit asym imbalanced overloaded
9695 * has_spare nr_idle balanced N/A N/A balanced balanced
9696 * fully_busy nr_idle nr_idle N/A N/A balanced balanced
9697 * misfit_task force N/A N/A N/A N/A N/A
9698 * asym_packing force force N/A N/A force force
9699 * imbalanced force force N/A N/A force force
9700 * overloaded force force N/A N/A force avg_load
9702 * N/A : Not Applicable because already filtered while updating
9704 * balanced : The system is balanced for these 2 groups.
9705 * force : Calculate the imbalance as load migration is probably needed.
9706 * avg_load : Only if imbalance is significant enough.
9707 * nr_idle : dst_cpu is not busy and the number of idle CPUs is quite
9708 * different in groups.
9712 * find_busiest_group - Returns the busiest group within the sched_domain
9713 * if there is an imbalance.
9714 * @env: The load balancing environment.
9716 * Also calculates the amount of runnable load which should be moved
9717 * to restore balance.
9719 * Return: - The busiest group if imbalance exists.
9721 static struct sched_group *find_busiest_group(struct lb_env *env)
9723 struct sg_lb_stats *local, *busiest;
9724 struct sd_lb_stats sds;
9726 init_sd_lb_stats(&sds);
9729 * Compute the various statistics relevant for load balancing at
9732 update_sd_lb_stats(env, &sds);
9734 if (sched_energy_enabled()) {
9735 struct root_domain *rd = env->dst_rq->rd;
9737 if (rcu_dereference(rd->pd) && !READ_ONCE(rd->overutilized))
9741 local = &sds.local_stat;
9742 busiest = &sds.busiest_stat;
9744 /* There is no busy sibling group to pull tasks from */
9748 /* Misfit tasks should be dealt with regardless of the avg load */
9749 if (busiest->group_type == group_misfit_task)
9752 /* ASYM feature bypasses nice load balance check */
9753 if (busiest->group_type == group_asym_packing)
9757 * If the busiest group is imbalanced the below checks don't
9758 * work because they assume all things are equal, which typically
9759 * isn't true due to cpus_ptr constraints and the like.
9761 if (busiest->group_type == group_imbalanced)
9765 * If the local group is busier than the selected busiest group
9766 * don't try and pull any tasks.
9768 if (local->group_type > busiest->group_type)
9772 * When groups are overloaded, use the avg_load to ensure fairness
9775 if (local->group_type == group_overloaded) {
9777 * If the local group is more loaded than the selected
9778 * busiest group don't try to pull any tasks.
9780 if (local->avg_load >= busiest->avg_load)
9783 /* XXX broken for overlapping NUMA groups */
9784 sds.avg_load = (sds.total_load * SCHED_CAPACITY_SCALE) /
9788 * Don't pull any tasks if this group is already above the
9789 * domain average load.
9791 if (local->avg_load >= sds.avg_load)
9795 * If the busiest group is more loaded, use imbalance_pct to be
9798 if (100 * busiest->avg_load <=
9799 env->sd->imbalance_pct * local->avg_load)
9803 /* Try to move all excess tasks to child's sibling domain */
9804 if (sds.prefer_sibling && local->group_type == group_has_spare &&
9805 busiest->sum_nr_running > local->sum_nr_running + 1)
9808 if (busiest->group_type != group_overloaded) {
9809 if (env->idle == CPU_NOT_IDLE)
9811 * If the busiest group is not overloaded (and as a
9812 * result the local one too) but this CPU is already
9813 * busy, let another idle CPU try to pull task.
9817 if (busiest->group_weight > 1 &&
9818 local->idle_cpus <= (busiest->idle_cpus + 1))
9820 * If the busiest group is not overloaded
9821 * and there is no imbalance between this and busiest
9822 * group wrt idle CPUs, it is balanced. The imbalance
9823 * becomes significant if the diff is greater than 1
9824 * otherwise we might end up to just move the imbalance
9825 * on another group. Of course this applies only if
9826 * there is more than 1 CPU per group.
9830 if (busiest->sum_h_nr_running == 1)
9832 * busiest doesn't have any tasks waiting to run
9838 /* Looks like there is an imbalance. Compute it */
9839 calculate_imbalance(env, &sds);
9840 return env->imbalance ? sds.busiest : NULL;
9848 * find_busiest_queue - find the busiest runqueue among the CPUs in the group.
9850 static struct rq *find_busiest_queue(struct lb_env *env,
9851 struct sched_group *group)
9853 struct rq *busiest = NULL, *rq;
9854 unsigned long busiest_util = 0, busiest_load = 0, busiest_capacity = 1;
9855 unsigned int busiest_nr = 0;
9858 for_each_cpu_and(i, sched_group_span(group), env->cpus) {
9859 unsigned long capacity, load, util;
9860 unsigned int nr_running;
9864 rt = fbq_classify_rq(rq);
9867 * We classify groups/runqueues into three groups:
9868 * - regular: there are !numa tasks
9869 * - remote: there are numa tasks that run on the 'wrong' node
9870 * - all: there is no distinction
9872 * In order to avoid migrating ideally placed numa tasks,
9873 * ignore those when there's better options.
9875 * If we ignore the actual busiest queue to migrate another
9876 * task, the next balance pass can still reduce the busiest
9877 * queue by moving tasks around inside the node.
9879 * If we cannot move enough load due to this classification
9880 * the next pass will adjust the group classification and
9881 * allow migration of more tasks.
9883 * Both cases only affect the total convergence complexity.
9885 if (rt > env->fbq_type)
9888 nr_running = rq->cfs.h_nr_running;
9892 capacity = capacity_of(i);
9895 * For ASYM_CPUCAPACITY domains, don't pick a CPU that could
9896 * eventually lead to active_balancing high->low capacity.
9897 * Higher per-CPU capacity is considered better than balancing
9900 if (env->sd->flags & SD_ASYM_CPUCAPACITY &&
9901 !capacity_greater(capacity_of(env->dst_cpu), capacity) &&
9905 /* Make sure we only pull tasks from a CPU of lower priority */
9906 if ((env->sd->flags & SD_ASYM_PACKING) &&
9907 sched_asym_prefer(i, env->dst_cpu) &&
9911 switch (env->migration_type) {
9914 * When comparing with load imbalance, use cpu_load()
9915 * which is not scaled with the CPU capacity.
9917 load = cpu_load(rq);
9919 if (nr_running == 1 && load > env->imbalance &&
9920 !check_cpu_capacity(rq, env->sd))
9924 * For the load comparisons with the other CPUs,
9925 * consider the cpu_load() scaled with the CPU
9926 * capacity, so that the load can be moved away
9927 * from the CPU that is potentially running at a
9930 * Thus we're looking for max(load_i / capacity_i),
9931 * crosswise multiplication to rid ourselves of the
9932 * division works out to:
9933 * load_i * capacity_j > load_j * capacity_i;
9934 * where j is our previous maximum.
9936 if (load * busiest_capacity > busiest_load * capacity) {
9937 busiest_load = load;
9938 busiest_capacity = capacity;
9944 util = cpu_util_cfs(i);
9947 * Don't try to pull utilization from a CPU with one
9948 * running task. Whatever its utilization, we will fail
9951 if (nr_running <= 1)
9954 if (busiest_util < util) {
9955 busiest_util = util;
9961 if (busiest_nr < nr_running) {
9962 busiest_nr = nr_running;
9967 case migrate_misfit:
9969 * For ASYM_CPUCAPACITY domains with misfit tasks we
9970 * simply seek the "biggest" misfit task.
9972 if (rq->misfit_task_load > busiest_load) {
9973 busiest_load = rq->misfit_task_load;
9986 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
9987 * so long as it is large enough.
9989 #define MAX_PINNED_INTERVAL 512
9992 asym_active_balance(struct lb_env *env)
9995 * ASYM_PACKING needs to force migrate tasks from busy but
9996 * lower priority CPUs in order to pack all tasks in the
9997 * highest priority CPUs.
9999 return env->idle != CPU_NOT_IDLE && (env->sd->flags & SD_ASYM_PACKING) &&
10000 sched_asym_prefer(env->dst_cpu, env->src_cpu);
10004 imbalanced_active_balance(struct lb_env *env)
10006 struct sched_domain *sd = env->sd;
10009 * The imbalanced case includes the case of pinned tasks preventing a fair
10010 * distribution of the load on the system but also the even distribution of the
10011 * threads on a system with spare capacity
10013 if ((env->migration_type == migrate_task) &&
10014 (sd->nr_balance_failed > sd->cache_nice_tries+2))
10020 static int need_active_balance(struct lb_env *env)
10022 struct sched_domain *sd = env->sd;
10024 if (asym_active_balance(env))
10027 if (imbalanced_active_balance(env))
10031 * The dst_cpu is idle and the src_cpu CPU has only 1 CFS task.
10032 * It's worth migrating the task if the src_cpu's capacity is reduced
10033 * because of other sched_class or IRQs if more capacity stays
10034 * available on dst_cpu.
10036 if ((env->idle != CPU_NOT_IDLE) &&
10037 (env->src_rq->cfs.h_nr_running == 1)) {
10038 if ((check_cpu_capacity(env->src_rq, sd)) &&
10039 (capacity_of(env->src_cpu)*sd->imbalance_pct < capacity_of(env->dst_cpu)*100))
10043 if (env->migration_type == migrate_misfit)
10049 static int active_load_balance_cpu_stop(void *data);
10051 static int should_we_balance(struct lb_env *env)
10053 struct sched_group *sg = env->sd->groups;
10057 * Ensure the balancing environment is consistent; can happen
10058 * when the softirq triggers 'during' hotplug.
10060 if (!cpumask_test_cpu(env->dst_cpu, env->cpus))
10064 * In the newly idle case, we will allow all the CPUs
10065 * to do the newly idle load balance.
10067 * However, we bail out if we already have tasks or a wakeup pending,
10068 * to optimize wakeup latency.
10070 if (env->idle == CPU_NEWLY_IDLE) {
10071 if (env->dst_rq->nr_running > 0 || env->dst_rq->ttwu_pending)
10076 /* Try to find first idle CPU */
10077 for_each_cpu_and(cpu, group_balance_mask(sg), env->cpus) {
10078 if (!idle_cpu(cpu))
10081 /* Are we the first idle CPU? */
10082 return cpu == env->dst_cpu;
10085 /* Are we the first CPU of this group ? */
10086 return group_balance_cpu(sg) == env->dst_cpu;
10090 * Check this_cpu to ensure it is balanced within domain. Attempt to move
10091 * tasks if there is an imbalance.
10093 static int load_balance(int this_cpu, struct rq *this_rq,
10094 struct sched_domain *sd, enum cpu_idle_type idle,
10095 int *continue_balancing)
10097 int ld_moved, cur_ld_moved, active_balance = 0;
10098 struct sched_domain *sd_parent = sd->parent;
10099 struct sched_group *group;
10100 struct rq *busiest;
10101 struct rq_flags rf;
10102 struct cpumask *cpus = this_cpu_cpumask_var_ptr(load_balance_mask);
10104 struct lb_env env = {
10106 .dst_cpu = this_cpu,
10108 .dst_grpmask = sched_group_span(sd->groups),
10110 .loop_break = sched_nr_migrate_break,
10113 .tasks = LIST_HEAD_INIT(env.tasks),
10116 cpumask_and(cpus, sched_domain_span(sd), cpu_active_mask);
10118 schedstat_inc(sd->lb_count[idle]);
10121 if (!should_we_balance(&env)) {
10122 *continue_balancing = 0;
10126 group = find_busiest_group(&env);
10128 schedstat_inc(sd->lb_nobusyg[idle]);
10132 busiest = find_busiest_queue(&env, group);
10134 schedstat_inc(sd->lb_nobusyq[idle]);
10138 WARN_ON_ONCE(busiest == env.dst_rq);
10140 schedstat_add(sd->lb_imbalance[idle], env.imbalance);
10142 env.src_cpu = busiest->cpu;
10143 env.src_rq = busiest;
10146 /* Clear this flag as soon as we find a pullable task */
10147 env.flags |= LBF_ALL_PINNED;
10148 if (busiest->nr_running > 1) {
10150 * Attempt to move tasks. If find_busiest_group has found
10151 * an imbalance but busiest->nr_running <= 1, the group is
10152 * still unbalanced. ld_moved simply stays zero, so it is
10153 * correctly treated as an imbalance.
10155 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
10158 rq_lock_irqsave(busiest, &rf);
10159 update_rq_clock(busiest);
10162 * cur_ld_moved - load moved in current iteration
10163 * ld_moved - cumulative load moved across iterations
10165 cur_ld_moved = detach_tasks(&env);
10168 * We've detached some tasks from busiest_rq. Every
10169 * task is masked "TASK_ON_RQ_MIGRATING", so we can safely
10170 * unlock busiest->lock, and we are able to be sure
10171 * that nobody can manipulate the tasks in parallel.
10172 * See task_rq_lock() family for the details.
10175 rq_unlock(busiest, &rf);
10177 if (cur_ld_moved) {
10178 attach_tasks(&env);
10179 ld_moved += cur_ld_moved;
10182 local_irq_restore(rf.flags);
10184 if (env.flags & LBF_NEED_BREAK) {
10185 env.flags &= ~LBF_NEED_BREAK;
10190 * Revisit (affine) tasks on src_cpu that couldn't be moved to
10191 * us and move them to an alternate dst_cpu in our sched_group
10192 * where they can run. The upper limit on how many times we
10193 * iterate on same src_cpu is dependent on number of CPUs in our
10196 * This changes load balance semantics a bit on who can move
10197 * load to a given_cpu. In addition to the given_cpu itself
10198 * (or a ilb_cpu acting on its behalf where given_cpu is
10199 * nohz-idle), we now have balance_cpu in a position to move
10200 * load to given_cpu. In rare situations, this may cause
10201 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
10202 * _independently_ and at _same_ time to move some load to
10203 * given_cpu) causing excess load to be moved to given_cpu.
10204 * This however should not happen so much in practice and
10205 * moreover subsequent load balance cycles should correct the
10206 * excess load moved.
10208 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
10210 /* Prevent to re-select dst_cpu via env's CPUs */
10211 __cpumask_clear_cpu(env.dst_cpu, env.cpus);
10213 env.dst_rq = cpu_rq(env.new_dst_cpu);
10214 env.dst_cpu = env.new_dst_cpu;
10215 env.flags &= ~LBF_DST_PINNED;
10217 env.loop_break = sched_nr_migrate_break;
10220 * Go back to "more_balance" rather than "redo" since we
10221 * need to continue with same src_cpu.
10227 * We failed to reach balance because of affinity.
10230 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
10232 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0)
10233 *group_imbalance = 1;
10236 /* All tasks on this runqueue were pinned by CPU affinity */
10237 if (unlikely(env.flags & LBF_ALL_PINNED)) {
10238 __cpumask_clear_cpu(cpu_of(busiest), cpus);
10240 * Attempting to continue load balancing at the current
10241 * sched_domain level only makes sense if there are
10242 * active CPUs remaining as possible busiest CPUs to
10243 * pull load from which are not contained within the
10244 * destination group that is receiving any migrated
10247 if (!cpumask_subset(cpus, env.dst_grpmask)) {
10249 env.loop_break = sched_nr_migrate_break;
10252 goto out_all_pinned;
10257 schedstat_inc(sd->lb_failed[idle]);
10259 * Increment the failure counter only on periodic balance.
10260 * We do not want newidle balance, which can be very
10261 * frequent, pollute the failure counter causing
10262 * excessive cache_hot migrations and active balances.
10264 if (idle != CPU_NEWLY_IDLE)
10265 sd->nr_balance_failed++;
10267 if (need_active_balance(&env)) {
10268 unsigned long flags;
10270 raw_spin_rq_lock_irqsave(busiest, flags);
10273 * Don't kick the active_load_balance_cpu_stop,
10274 * if the curr task on busiest CPU can't be
10275 * moved to this_cpu:
10277 if (!cpumask_test_cpu(this_cpu, busiest->curr->cpus_ptr)) {
10278 raw_spin_rq_unlock_irqrestore(busiest, flags);
10279 goto out_one_pinned;
10282 /* Record that we found at least one task that could run on this_cpu */
10283 env.flags &= ~LBF_ALL_PINNED;
10286 * ->active_balance synchronizes accesses to
10287 * ->active_balance_work. Once set, it's cleared
10288 * only after active load balance is finished.
10290 if (!busiest->active_balance) {
10291 busiest->active_balance = 1;
10292 busiest->push_cpu = this_cpu;
10293 active_balance = 1;
10295 raw_spin_rq_unlock_irqrestore(busiest, flags);
10297 if (active_balance) {
10298 stop_one_cpu_nowait(cpu_of(busiest),
10299 active_load_balance_cpu_stop, busiest,
10300 &busiest->active_balance_work);
10304 sd->nr_balance_failed = 0;
10307 if (likely(!active_balance) || need_active_balance(&env)) {
10308 /* We were unbalanced, so reset the balancing interval */
10309 sd->balance_interval = sd->min_interval;
10316 * We reach balance although we may have faced some affinity
10317 * constraints. Clear the imbalance flag only if other tasks got
10318 * a chance to move and fix the imbalance.
10320 if (sd_parent && !(env.flags & LBF_ALL_PINNED)) {
10321 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
10323 if (*group_imbalance)
10324 *group_imbalance = 0;
10329 * We reach balance because all tasks are pinned at this level so
10330 * we can't migrate them. Let the imbalance flag set so parent level
10331 * can try to migrate them.
10333 schedstat_inc(sd->lb_balanced[idle]);
10335 sd->nr_balance_failed = 0;
10341 * newidle_balance() disregards balance intervals, so we could
10342 * repeatedly reach this code, which would lead to balance_interval
10343 * skyrocketing in a short amount of time. Skip the balance_interval
10344 * increase logic to avoid that.
10346 if (env.idle == CPU_NEWLY_IDLE)
10349 /* tune up the balancing interval */
10350 if ((env.flags & LBF_ALL_PINNED &&
10351 sd->balance_interval < MAX_PINNED_INTERVAL) ||
10352 sd->balance_interval < sd->max_interval)
10353 sd->balance_interval *= 2;
10358 static inline unsigned long
10359 get_sd_balance_interval(struct sched_domain *sd, int cpu_busy)
10361 unsigned long interval = sd->balance_interval;
10364 interval *= sd->busy_factor;
10366 /* scale ms to jiffies */
10367 interval = msecs_to_jiffies(interval);
10370 * Reduce likelihood of busy balancing at higher domains racing with
10371 * balancing at lower domains by preventing their balancing periods
10372 * from being multiples of each other.
10377 interval = clamp(interval, 1UL, max_load_balance_interval);
10383 update_next_balance(struct sched_domain *sd, unsigned long *next_balance)
10385 unsigned long interval, next;
10387 /* used by idle balance, so cpu_busy = 0 */
10388 interval = get_sd_balance_interval(sd, 0);
10389 next = sd->last_balance + interval;
10391 if (time_after(*next_balance, next))
10392 *next_balance = next;
10396 * active_load_balance_cpu_stop is run by the CPU stopper. It pushes
10397 * running tasks off the busiest CPU onto idle CPUs. It requires at
10398 * least 1 task to be running on each physical CPU where possible, and
10399 * avoids physical / logical imbalances.
10401 static int active_load_balance_cpu_stop(void *data)
10403 struct rq *busiest_rq = data;
10404 int busiest_cpu = cpu_of(busiest_rq);
10405 int target_cpu = busiest_rq->push_cpu;
10406 struct rq *target_rq = cpu_rq(target_cpu);
10407 struct sched_domain *sd;
10408 struct task_struct *p = NULL;
10409 struct rq_flags rf;
10411 rq_lock_irq(busiest_rq, &rf);
10413 * Between queueing the stop-work and running it is a hole in which
10414 * CPUs can become inactive. We should not move tasks from or to
10417 if (!cpu_active(busiest_cpu) || !cpu_active(target_cpu))
10420 /* Make sure the requested CPU hasn't gone down in the meantime: */
10421 if (unlikely(busiest_cpu != smp_processor_id() ||
10422 !busiest_rq->active_balance))
10425 /* Is there any task to move? */
10426 if (busiest_rq->nr_running <= 1)
10430 * This condition is "impossible", if it occurs
10431 * we need to fix it. Originally reported by
10432 * Bjorn Helgaas on a 128-CPU setup.
10434 WARN_ON_ONCE(busiest_rq == target_rq);
10436 /* Search for an sd spanning us and the target CPU. */
10438 for_each_domain(target_cpu, sd) {
10439 if (cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
10444 struct lb_env env = {
10446 .dst_cpu = target_cpu,
10447 .dst_rq = target_rq,
10448 .src_cpu = busiest_rq->cpu,
10449 .src_rq = busiest_rq,
10451 .flags = LBF_ACTIVE_LB,
10454 schedstat_inc(sd->alb_count);
10455 update_rq_clock(busiest_rq);
10457 p = detach_one_task(&env);
10459 schedstat_inc(sd->alb_pushed);
10460 /* Active balancing done, reset the failure counter. */
10461 sd->nr_balance_failed = 0;
10463 schedstat_inc(sd->alb_failed);
10468 busiest_rq->active_balance = 0;
10469 rq_unlock(busiest_rq, &rf);
10472 attach_one_task(target_rq, p);
10474 local_irq_enable();
10479 static DEFINE_SPINLOCK(balancing);
10482 * Scale the max load_balance interval with the number of CPUs in the system.
10483 * This trades load-balance latency on larger machines for less cross talk.
10485 void update_max_interval(void)
10487 max_load_balance_interval = HZ*num_online_cpus()/10;
10490 static inline bool update_newidle_cost(struct sched_domain *sd, u64 cost)
10492 if (cost > sd->max_newidle_lb_cost) {
10494 * Track max cost of a domain to make sure to not delay the
10495 * next wakeup on the CPU.
10497 sd->max_newidle_lb_cost = cost;
10498 sd->last_decay_max_lb_cost = jiffies;
10499 } else if (time_after(jiffies, sd->last_decay_max_lb_cost + HZ)) {
10501 * Decay the newidle max times by ~1% per second to ensure that
10502 * it is not outdated and the current max cost is actually
10505 sd->max_newidle_lb_cost = (sd->max_newidle_lb_cost * 253) / 256;
10506 sd->last_decay_max_lb_cost = jiffies;
10515 * It checks each scheduling domain to see if it is due to be balanced,
10516 * and initiates a balancing operation if so.
10518 * Balancing parameters are set up in init_sched_domains.
10520 static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle)
10522 int continue_balancing = 1;
10524 int busy = idle != CPU_IDLE && !sched_idle_cpu(cpu);
10525 unsigned long interval;
10526 struct sched_domain *sd;
10527 /* Earliest time when we have to do rebalance again */
10528 unsigned long next_balance = jiffies + 60*HZ;
10529 int update_next_balance = 0;
10530 int need_serialize, need_decay = 0;
10534 for_each_domain(cpu, sd) {
10536 * Decay the newidle max times here because this is a regular
10537 * visit to all the domains.
10539 need_decay = update_newidle_cost(sd, 0);
10540 max_cost += sd->max_newidle_lb_cost;
10543 * Stop the load balance at this level. There is another
10544 * CPU in our sched group which is doing load balancing more
10547 if (!continue_balancing) {
10553 interval = get_sd_balance_interval(sd, busy);
10555 need_serialize = sd->flags & SD_SERIALIZE;
10556 if (need_serialize) {
10557 if (!spin_trylock(&balancing))
10561 if (time_after_eq(jiffies, sd->last_balance + interval)) {
10562 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
10564 * The LBF_DST_PINNED logic could have changed
10565 * env->dst_cpu, so we can't know our idle
10566 * state even if we migrated tasks. Update it.
10568 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
10569 busy = idle != CPU_IDLE && !sched_idle_cpu(cpu);
10571 sd->last_balance = jiffies;
10572 interval = get_sd_balance_interval(sd, busy);
10574 if (need_serialize)
10575 spin_unlock(&balancing);
10577 if (time_after(next_balance, sd->last_balance + interval)) {
10578 next_balance = sd->last_balance + interval;
10579 update_next_balance = 1;
10584 * Ensure the rq-wide value also decays but keep it at a
10585 * reasonable floor to avoid funnies with rq->avg_idle.
10587 rq->max_idle_balance_cost =
10588 max((u64)sysctl_sched_migration_cost, max_cost);
10593 * next_balance will be updated only when there is a need.
10594 * When the cpu is attached to null domain for ex, it will not be
10597 if (likely(update_next_balance))
10598 rq->next_balance = next_balance;
10602 static inline int on_null_domain(struct rq *rq)
10604 return unlikely(!rcu_dereference_sched(rq->sd));
10607 #ifdef CONFIG_NO_HZ_COMMON
10609 * idle load balancing details
10610 * - When one of the busy CPUs notice that there may be an idle rebalancing
10611 * needed, they will kick the idle load balancer, which then does idle
10612 * load balancing for all the idle CPUs.
10613 * - HK_TYPE_MISC CPUs are used for this task, because HK_TYPE_SCHED not set
10617 static inline int find_new_ilb(void)
10620 const struct cpumask *hk_mask;
10622 hk_mask = housekeeping_cpumask(HK_TYPE_MISC);
10624 for_each_cpu_and(ilb, nohz.idle_cpus_mask, hk_mask) {
10626 if (ilb == smp_processor_id())
10637 * Kick a CPU to do the nohz balancing, if it is time for it. We pick any
10638 * idle CPU in the HK_TYPE_MISC housekeeping set (if there is one).
10640 static void kick_ilb(unsigned int flags)
10645 * Increase nohz.next_balance only when if full ilb is triggered but
10646 * not if we only update stats.
10648 if (flags & NOHZ_BALANCE_KICK)
10649 nohz.next_balance = jiffies+1;
10651 ilb_cpu = find_new_ilb();
10653 if (ilb_cpu >= nr_cpu_ids)
10657 * Access to rq::nohz_csd is serialized by NOHZ_KICK_MASK; he who sets
10658 * the first flag owns it; cleared by nohz_csd_func().
10660 flags = atomic_fetch_or(flags, nohz_flags(ilb_cpu));
10661 if (flags & NOHZ_KICK_MASK)
10665 * This way we generate an IPI on the target CPU which
10666 * is idle. And the softirq performing nohz idle load balance
10667 * will be run before returning from the IPI.
10669 smp_call_function_single_async(ilb_cpu, &cpu_rq(ilb_cpu)->nohz_csd);
10673 * Current decision point for kicking the idle load balancer in the presence
10674 * of idle CPUs in the system.
10676 static void nohz_balancer_kick(struct rq *rq)
10678 unsigned long now = jiffies;
10679 struct sched_domain_shared *sds;
10680 struct sched_domain *sd;
10681 int nr_busy, i, cpu = rq->cpu;
10682 unsigned int flags = 0;
10684 if (unlikely(rq->idle_balance))
10688 * We may be recently in ticked or tickless idle mode. At the first
10689 * busy tick after returning from idle, we will update the busy stats.
10691 nohz_balance_exit_idle(rq);
10694 * None are in tickless mode and hence no need for NOHZ idle load
10697 if (likely(!atomic_read(&nohz.nr_cpus)))
10700 if (READ_ONCE(nohz.has_blocked) &&
10701 time_after(now, READ_ONCE(nohz.next_blocked)))
10702 flags = NOHZ_STATS_KICK;
10704 if (time_before(now, nohz.next_balance))
10707 if (rq->nr_running >= 2) {
10708 flags = NOHZ_STATS_KICK | NOHZ_BALANCE_KICK;
10714 sd = rcu_dereference(rq->sd);
10717 * If there's a CFS task and the current CPU has reduced
10718 * capacity; kick the ILB to see if there's a better CPU to run
10721 if (rq->cfs.h_nr_running >= 1 && check_cpu_capacity(rq, sd)) {
10722 flags = NOHZ_STATS_KICK | NOHZ_BALANCE_KICK;
10727 sd = rcu_dereference(per_cpu(sd_asym_packing, cpu));
10730 * When ASYM_PACKING; see if there's a more preferred CPU
10731 * currently idle; in which case, kick the ILB to move tasks
10734 for_each_cpu_and(i, sched_domain_span(sd), nohz.idle_cpus_mask) {
10735 if (sched_asym_prefer(i, cpu)) {
10736 flags = NOHZ_STATS_KICK | NOHZ_BALANCE_KICK;
10742 sd = rcu_dereference(per_cpu(sd_asym_cpucapacity, cpu));
10745 * When ASYM_CPUCAPACITY; see if there's a higher capacity CPU
10746 * to run the misfit task on.
10748 if (check_misfit_status(rq, sd)) {
10749 flags = NOHZ_STATS_KICK | NOHZ_BALANCE_KICK;
10754 * For asymmetric systems, we do not want to nicely balance
10755 * cache use, instead we want to embrace asymmetry and only
10756 * ensure tasks have enough CPU capacity.
10758 * Skip the LLC logic because it's not relevant in that case.
10763 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
10766 * If there is an imbalance between LLC domains (IOW we could
10767 * increase the overall cache use), we need some less-loaded LLC
10768 * domain to pull some load. Likewise, we may need to spread
10769 * load within the current LLC domain (e.g. packed SMT cores but
10770 * other CPUs are idle). We can't really know from here how busy
10771 * the others are - so just get a nohz balance going if it looks
10772 * like this LLC domain has tasks we could move.
10774 nr_busy = atomic_read(&sds->nr_busy_cpus);
10776 flags = NOHZ_STATS_KICK | NOHZ_BALANCE_KICK;
10783 if (READ_ONCE(nohz.needs_update))
10784 flags |= NOHZ_NEXT_KICK;
10790 static void set_cpu_sd_state_busy(int cpu)
10792 struct sched_domain *sd;
10795 sd = rcu_dereference(per_cpu(sd_llc, cpu));
10797 if (!sd || !sd->nohz_idle)
10801 atomic_inc(&sd->shared->nr_busy_cpus);
10806 void nohz_balance_exit_idle(struct rq *rq)
10808 SCHED_WARN_ON(rq != this_rq());
10810 if (likely(!rq->nohz_tick_stopped))
10813 rq->nohz_tick_stopped = 0;
10814 cpumask_clear_cpu(rq->cpu, nohz.idle_cpus_mask);
10815 atomic_dec(&nohz.nr_cpus);
10817 set_cpu_sd_state_busy(rq->cpu);
10820 static void set_cpu_sd_state_idle(int cpu)
10822 struct sched_domain *sd;
10825 sd = rcu_dereference(per_cpu(sd_llc, cpu));
10827 if (!sd || sd->nohz_idle)
10831 atomic_dec(&sd->shared->nr_busy_cpus);
10837 * This routine will record that the CPU is going idle with tick stopped.
10838 * This info will be used in performing idle load balancing in the future.
10840 void nohz_balance_enter_idle(int cpu)
10842 struct rq *rq = cpu_rq(cpu);
10844 SCHED_WARN_ON(cpu != smp_processor_id());
10846 /* If this CPU is going down, then nothing needs to be done: */
10847 if (!cpu_active(cpu))
10850 /* Spare idle load balancing on CPUs that don't want to be disturbed: */
10851 if (!housekeeping_cpu(cpu, HK_TYPE_SCHED))
10855 * Can be set safely without rq->lock held
10856 * If a clear happens, it will have evaluated last additions because
10857 * rq->lock is held during the check and the clear
10859 rq->has_blocked_load = 1;
10862 * The tick is still stopped but load could have been added in the
10863 * meantime. We set the nohz.has_blocked flag to trig a check of the
10864 * *_avg. The CPU is already part of nohz.idle_cpus_mask so the clear
10865 * of nohz.has_blocked can only happen after checking the new load
10867 if (rq->nohz_tick_stopped)
10870 /* If we're a completely isolated CPU, we don't play: */
10871 if (on_null_domain(rq))
10874 rq->nohz_tick_stopped = 1;
10876 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
10877 atomic_inc(&nohz.nr_cpus);
10880 * Ensures that if nohz_idle_balance() fails to observe our
10881 * @idle_cpus_mask store, it must observe the @has_blocked
10882 * and @needs_update stores.
10884 smp_mb__after_atomic();
10886 set_cpu_sd_state_idle(cpu);
10888 WRITE_ONCE(nohz.needs_update, 1);
10891 * Each time a cpu enter idle, we assume that it has blocked load and
10892 * enable the periodic update of the load of idle cpus
10894 WRITE_ONCE(nohz.has_blocked, 1);
10897 static bool update_nohz_stats(struct rq *rq)
10899 unsigned int cpu = rq->cpu;
10901 if (!rq->has_blocked_load)
10904 if (!cpumask_test_cpu(cpu, nohz.idle_cpus_mask))
10907 if (!time_after(jiffies, READ_ONCE(rq->last_blocked_load_update_tick)))
10910 update_blocked_averages(cpu);
10912 return rq->has_blocked_load;
10916 * Internal function that runs load balance for all idle cpus. The load balance
10917 * can be a simple update of blocked load or a complete load balance with
10918 * tasks movement depending of flags.
10920 static void _nohz_idle_balance(struct rq *this_rq, unsigned int flags)
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);
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);
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);
11557 * In case the task sched_avg hasn't been attached:
11558 * - A forked task which hasn't been woken up by wake_up_new_task().
11559 * - A task which has been woken up by try_to_wake_up() but is
11560 * waiting for actually being woken up by sched_ttwu_pending().
11562 if (!se->avg.last_update_time)
11566 /* Catch up with the cfs_rq and remove our load when we leave */
11567 update_load_avg(cfs_rq, se, 0);
11568 detach_entity_load_avg(cfs_rq, se);
11569 update_tg_load_avg(cfs_rq);
11570 propagate_entity_cfs_rq(se);
11573 static void attach_entity_cfs_rq(struct sched_entity *se)
11575 struct cfs_rq *cfs_rq = cfs_rq_of(se);
11577 /* Synchronize entity with its cfs_rq */
11578 update_load_avg(cfs_rq, se, sched_feat(ATTACH_AGE_LOAD) ? 0 : SKIP_AGE_LOAD);
11579 attach_entity_load_avg(cfs_rq, se);
11580 update_tg_load_avg(cfs_rq);
11581 propagate_entity_cfs_rq(se);
11584 static void detach_task_cfs_rq(struct task_struct *p)
11586 struct sched_entity *se = &p->se;
11587 struct cfs_rq *cfs_rq = cfs_rq_of(se);
11589 if (!vruntime_normalized(p)) {
11591 * Fix up our vruntime so that the current sleep doesn't
11592 * cause 'unlimited' sleep bonus.
11594 place_entity(cfs_rq, se, 0);
11595 se->vruntime -= cfs_rq->min_vruntime;
11598 detach_entity_cfs_rq(se);
11601 static void attach_task_cfs_rq(struct task_struct *p)
11603 struct sched_entity *se = &p->se;
11604 struct cfs_rq *cfs_rq = cfs_rq_of(se);
11606 attach_entity_cfs_rq(se);
11608 if (!vruntime_normalized(p))
11609 se->vruntime += cfs_rq->min_vruntime;
11612 static void switched_from_fair(struct rq *rq, struct task_struct *p)
11614 detach_task_cfs_rq(p);
11617 static void switched_to_fair(struct rq *rq, struct task_struct *p)
11619 attach_task_cfs_rq(p);
11621 if (task_on_rq_queued(p)) {
11623 * We were most likely switched from sched_rt, so
11624 * kick off the schedule if running, otherwise just see
11625 * if we can still preempt the current task.
11627 if (task_current(rq, p))
11630 check_preempt_curr(rq, p, 0);
11634 /* Account for a task changing its policy or group.
11636 * This routine is mostly called to set cfs_rq->curr field when a task
11637 * migrates between groups/classes.
11639 static void set_next_task_fair(struct rq *rq, struct task_struct *p, bool first)
11641 struct sched_entity *se = &p->se;
11644 if (task_on_rq_queued(p)) {
11646 * Move the next running task to the front of the list, so our
11647 * cfs_tasks list becomes MRU one.
11649 list_move(&se->group_node, &rq->cfs_tasks);
11653 for_each_sched_entity(se) {
11654 struct cfs_rq *cfs_rq = cfs_rq_of(se);
11656 set_next_entity(cfs_rq, se);
11657 /* ensure bandwidth has been allocated on our new cfs_rq */
11658 account_cfs_rq_runtime(cfs_rq, 0);
11662 void init_cfs_rq(struct cfs_rq *cfs_rq)
11664 cfs_rq->tasks_timeline = RB_ROOT_CACHED;
11665 u64_u32_store(cfs_rq->min_vruntime, (u64)(-(1LL << 20)));
11667 raw_spin_lock_init(&cfs_rq->removed.lock);
11671 #ifdef CONFIG_FAIR_GROUP_SCHED
11672 static void task_change_group_fair(struct task_struct *p)
11675 * We couldn't detach or attach a forked task which
11676 * hasn't been woken up by wake_up_new_task().
11678 if (READ_ONCE(p->__state) == TASK_NEW)
11681 detach_task_cfs_rq(p);
11684 /* Tell se's cfs_rq has been changed -- migrated */
11685 p->se.avg.last_update_time = 0;
11687 set_task_rq(p, task_cpu(p));
11688 attach_task_cfs_rq(p);
11691 void free_fair_sched_group(struct task_group *tg)
11695 for_each_possible_cpu(i) {
11697 kfree(tg->cfs_rq[i]);
11706 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
11708 struct sched_entity *se;
11709 struct cfs_rq *cfs_rq;
11712 tg->cfs_rq = kcalloc(nr_cpu_ids, sizeof(cfs_rq), GFP_KERNEL);
11715 tg->se = kcalloc(nr_cpu_ids, sizeof(se), GFP_KERNEL);
11719 tg->shares = NICE_0_LOAD;
11721 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
11723 for_each_possible_cpu(i) {
11724 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
11725 GFP_KERNEL, cpu_to_node(i));
11729 se = kzalloc_node(sizeof(struct sched_entity_stats),
11730 GFP_KERNEL, cpu_to_node(i));
11734 init_cfs_rq(cfs_rq);
11735 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
11736 init_entity_runnable_average(se);
11747 void online_fair_sched_group(struct task_group *tg)
11749 struct sched_entity *se;
11750 struct rq_flags rf;
11754 for_each_possible_cpu(i) {
11757 rq_lock_irq(rq, &rf);
11758 update_rq_clock(rq);
11759 attach_entity_cfs_rq(se);
11760 sync_throttle(tg, i);
11761 rq_unlock_irq(rq, &rf);
11765 void unregister_fair_sched_group(struct task_group *tg)
11767 unsigned long flags;
11771 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
11773 for_each_possible_cpu(cpu) {
11775 remove_entity_load_avg(tg->se[cpu]);
11778 * Only empty task groups can be destroyed; so we can speculatively
11779 * check on_list without danger of it being re-added.
11781 if (!tg->cfs_rq[cpu]->on_list)
11786 raw_spin_rq_lock_irqsave(rq, flags);
11787 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
11788 raw_spin_rq_unlock_irqrestore(rq, flags);
11792 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
11793 struct sched_entity *se, int cpu,
11794 struct sched_entity *parent)
11796 struct rq *rq = cpu_rq(cpu);
11800 init_cfs_rq_runtime(cfs_rq);
11802 tg->cfs_rq[cpu] = cfs_rq;
11805 /* se could be NULL for root_task_group */
11810 se->cfs_rq = &rq->cfs;
11813 se->cfs_rq = parent->my_q;
11814 se->depth = parent->depth + 1;
11818 /* guarantee group entities always have weight */
11819 update_load_set(&se->load, NICE_0_LOAD);
11820 se->parent = parent;
11823 static DEFINE_MUTEX(shares_mutex);
11825 static int __sched_group_set_shares(struct task_group *tg, unsigned long shares)
11829 lockdep_assert_held(&shares_mutex);
11832 * We can't change the weight of the root cgroup.
11837 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
11839 if (tg->shares == shares)
11842 tg->shares = shares;
11843 for_each_possible_cpu(i) {
11844 struct rq *rq = cpu_rq(i);
11845 struct sched_entity *se = tg->se[i];
11846 struct rq_flags rf;
11848 /* Propagate contribution to hierarchy */
11849 rq_lock_irqsave(rq, &rf);
11850 update_rq_clock(rq);
11851 for_each_sched_entity(se) {
11852 update_load_avg(cfs_rq_of(se), se, UPDATE_TG);
11853 update_cfs_group(se);
11855 rq_unlock_irqrestore(rq, &rf);
11861 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
11865 mutex_lock(&shares_mutex);
11866 if (tg_is_idle(tg))
11869 ret = __sched_group_set_shares(tg, shares);
11870 mutex_unlock(&shares_mutex);
11875 int sched_group_set_idle(struct task_group *tg, long idle)
11879 if (tg == &root_task_group)
11882 if (idle < 0 || idle > 1)
11885 mutex_lock(&shares_mutex);
11887 if (tg->idle == idle) {
11888 mutex_unlock(&shares_mutex);
11894 for_each_possible_cpu(i) {
11895 struct rq *rq = cpu_rq(i);
11896 struct sched_entity *se = tg->se[i];
11897 struct cfs_rq *parent_cfs_rq, *grp_cfs_rq = tg->cfs_rq[i];
11898 bool was_idle = cfs_rq_is_idle(grp_cfs_rq);
11899 long idle_task_delta;
11900 struct rq_flags rf;
11902 rq_lock_irqsave(rq, &rf);
11904 grp_cfs_rq->idle = idle;
11905 if (WARN_ON_ONCE(was_idle == cfs_rq_is_idle(grp_cfs_rq)))
11909 parent_cfs_rq = cfs_rq_of(se);
11910 if (cfs_rq_is_idle(grp_cfs_rq))
11911 parent_cfs_rq->idle_nr_running++;
11913 parent_cfs_rq->idle_nr_running--;
11916 idle_task_delta = grp_cfs_rq->h_nr_running -
11917 grp_cfs_rq->idle_h_nr_running;
11918 if (!cfs_rq_is_idle(grp_cfs_rq))
11919 idle_task_delta *= -1;
11921 for_each_sched_entity(se) {
11922 struct cfs_rq *cfs_rq = cfs_rq_of(se);
11927 cfs_rq->idle_h_nr_running += idle_task_delta;
11929 /* Already accounted at parent level and above. */
11930 if (cfs_rq_is_idle(cfs_rq))
11935 rq_unlock_irqrestore(rq, &rf);
11938 /* Idle groups have minimum weight. */
11939 if (tg_is_idle(tg))
11940 __sched_group_set_shares(tg, scale_load(WEIGHT_IDLEPRIO));
11942 __sched_group_set_shares(tg, NICE_0_LOAD);
11944 mutex_unlock(&shares_mutex);
11948 #else /* CONFIG_FAIR_GROUP_SCHED */
11950 void free_fair_sched_group(struct task_group *tg) { }
11952 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
11957 void online_fair_sched_group(struct task_group *tg) { }
11959 void unregister_fair_sched_group(struct task_group *tg) { }
11961 #endif /* CONFIG_FAIR_GROUP_SCHED */
11964 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
11966 struct sched_entity *se = &task->se;
11967 unsigned int rr_interval = 0;
11970 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
11973 if (rq->cfs.load.weight)
11974 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
11976 return rr_interval;
11980 * All the scheduling class methods:
11982 DEFINE_SCHED_CLASS(fair) = {
11984 .enqueue_task = enqueue_task_fair,
11985 .dequeue_task = dequeue_task_fair,
11986 .yield_task = yield_task_fair,
11987 .yield_to_task = yield_to_task_fair,
11989 .check_preempt_curr = check_preempt_wakeup,
11991 .pick_next_task = __pick_next_task_fair,
11992 .put_prev_task = put_prev_task_fair,
11993 .set_next_task = set_next_task_fair,
11996 .balance = balance_fair,
11997 .pick_task = pick_task_fair,
11998 .select_task_rq = select_task_rq_fair,
11999 .migrate_task_rq = migrate_task_rq_fair,
12001 .rq_online = rq_online_fair,
12002 .rq_offline = rq_offline_fair,
12004 .task_dead = task_dead_fair,
12005 .set_cpus_allowed = set_cpus_allowed_common,
12008 .task_tick = task_tick_fair,
12009 .task_fork = task_fork_fair,
12011 .prio_changed = prio_changed_fair,
12012 .switched_from = switched_from_fair,
12013 .switched_to = switched_to_fair,
12015 .get_rr_interval = get_rr_interval_fair,
12017 .update_curr = update_curr_fair,
12019 #ifdef CONFIG_FAIR_GROUP_SCHED
12020 .task_change_group = task_change_group_fair,
12023 #ifdef CONFIG_UCLAMP_TASK
12024 .uclamp_enabled = 1,
12028 #ifdef CONFIG_SCHED_DEBUG
12029 void print_cfs_stats(struct seq_file *m, int cpu)
12031 struct cfs_rq *cfs_rq, *pos;
12034 for_each_leaf_cfs_rq_safe(cpu_rq(cpu), cfs_rq, pos)
12035 print_cfs_rq(m, cpu, cfs_rq);
12039 #ifdef CONFIG_NUMA_BALANCING
12040 void show_numa_stats(struct task_struct *p, struct seq_file *m)
12043 unsigned long tsf = 0, tpf = 0, gsf = 0, gpf = 0;
12044 struct numa_group *ng;
12047 ng = rcu_dereference(p->numa_group);
12048 for_each_online_node(node) {
12049 if (p->numa_faults) {
12050 tsf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 0)];
12051 tpf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 1)];
12054 gsf = ng->faults[task_faults_idx(NUMA_MEM, node, 0)],
12055 gpf = ng->faults[task_faults_idx(NUMA_MEM, node, 1)];
12057 print_numa_stats(m, node, tsf, tpf, gsf, gpf);
12061 #endif /* CONFIG_NUMA_BALANCING */
12062 #endif /* CONFIG_SCHED_DEBUG */
12064 __init void init_sched_fair_class(void)
12069 for_each_possible_cpu(i) {
12070 zalloc_cpumask_var_node(&per_cpu(load_balance_mask, i), GFP_KERNEL, cpu_to_node(i));
12071 zalloc_cpumask_var_node(&per_cpu(select_rq_mask, i), GFP_KERNEL, cpu_to_node(i));
12074 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
12076 #ifdef CONFIG_NO_HZ_COMMON
12077 nohz.next_balance = jiffies;
12078 nohz.next_blocked = jiffies;
12079 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);