1 // SPDX-License-Identifier: GPL-2.0
3 * Completely Fair Scheduling (CFS) Class (SCHED_NORMAL/SCHED_BATCH)
5 * Copyright (C) 2007 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
7 * Interactivity improvements by Mike Galbraith
8 * (C) 2007 Mike Galbraith <efault@gmx.de>
10 * Various enhancements by Dmitry Adamushko.
11 * (C) 2007 Dmitry Adamushko <dmitry.adamushko@gmail.com>
13 * Group scheduling enhancements by Srivatsa Vaddagiri
14 * Copyright IBM Corporation, 2007
15 * Author: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
17 * Scaled math optimizations by Thomas Gleixner
18 * Copyright (C) 2007, Thomas Gleixner <tglx@linutronix.de>
20 * Adaptive scheduling granularity, math enhancements by Peter Zijlstra
21 * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra
23 #include <linux/energy_model.h>
24 #include <linux/mmap_lock.h>
25 #include <linux/hugetlb_inline.h>
26 #include <linux/jiffies.h>
27 #include <linux/mm_api.h>
28 #include <linux/highmem.h>
29 #include <linux/spinlock_api.h>
30 #include <linux/cpumask_api.h>
31 #include <linux/lockdep_api.h>
32 #include <linux/softirq.h>
33 #include <linux/refcount_api.h>
34 #include <linux/topology.h>
35 #include <linux/sched/clock.h>
36 #include <linux/sched/cond_resched.h>
37 #include <linux/sched/cputime.h>
38 #include <linux/sched/isolation.h>
39 #include <linux/sched/nohz.h>
41 #include <linux/cpuidle.h>
42 #include <linux/interrupt.h>
43 #include <linux/memory-tiers.h>
44 #include <linux/mempolicy.h>
45 #include <linux/mutex_api.h>
46 #include <linux/profile.h>
47 #include <linux/psi.h>
48 #include <linux/ratelimit.h>
49 #include <linux/task_work.h>
51 #include <asm/switch_to.h>
53 #include <linux/sched/cond_resched.h>
57 #include "autogroup.h"
60 * Targeted preemption latency for CPU-bound tasks:
62 * NOTE: this latency value is not the same as the concept of
63 * 'timeslice length' - timeslices in CFS are of variable length
64 * and have no persistent notion like in traditional, time-slice
65 * based scheduling concepts.
67 * (to see the precise effective timeslice length of your workload,
68 * run vmstat and monitor the context-switches (cs) field)
70 * (default: 6ms * (1 + ilog(ncpus)), units: nanoseconds)
72 unsigned int sysctl_sched_latency = 6000000ULL;
73 static unsigned int normalized_sysctl_sched_latency = 6000000ULL;
76 * The initial- and re-scaling of tunables is configurable
80 * SCHED_TUNABLESCALING_NONE - unscaled, always *1
81 * SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus)
82 * SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus
84 * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus))
86 unsigned int sysctl_sched_tunable_scaling = SCHED_TUNABLESCALING_LOG;
89 * Minimal preemption granularity for CPU-bound tasks:
91 * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds)
93 unsigned int sysctl_sched_min_granularity = 750000ULL;
94 static unsigned int normalized_sysctl_sched_min_granularity = 750000ULL;
97 * Minimal preemption granularity for CPU-bound SCHED_IDLE tasks.
98 * Applies only when SCHED_IDLE tasks compete with normal tasks.
100 * (default: 0.75 msec)
102 unsigned int sysctl_sched_idle_min_granularity = 750000ULL;
105 * This value is kept at sysctl_sched_latency/sysctl_sched_min_granularity
107 static unsigned int sched_nr_latency = 8;
110 * After fork, child runs first. If set to 0 (default) then
111 * parent will (try to) run first.
113 unsigned int sysctl_sched_child_runs_first __read_mostly;
116 * SCHED_OTHER wake-up granularity.
118 * This option delays the preemption effects of decoupled workloads
119 * and reduces their over-scheduling. Synchronous workloads will still
120 * have immediate wakeup/sleep latencies.
122 * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds)
124 unsigned int sysctl_sched_wakeup_granularity = 1000000UL;
125 static unsigned int normalized_sysctl_sched_wakeup_granularity = 1000000UL;
127 const_debug unsigned int sysctl_sched_migration_cost = 500000UL;
129 int sched_thermal_decay_shift;
130 static int __init setup_sched_thermal_decay_shift(char *str)
134 if (kstrtoint(str, 0, &_shift))
135 pr_warn("Unable to set scheduler thermal pressure decay shift parameter\n");
137 sched_thermal_decay_shift = clamp(_shift, 0, 10);
140 __setup("sched_thermal_decay_shift=", setup_sched_thermal_decay_shift);
144 * For asym packing, by default the lower numbered CPU has higher priority.
146 int __weak arch_asym_cpu_priority(int cpu)
152 * The margin used when comparing utilization with CPU capacity.
156 #define fits_capacity(cap, max) ((cap) * 1280 < (max) * 1024)
159 * The margin used when comparing CPU capacities.
160 * is 'cap1' noticeably greater than 'cap2'
164 #define capacity_greater(cap1, cap2) ((cap1) * 1024 > (cap2) * 1078)
167 #ifdef CONFIG_CFS_BANDWIDTH
169 * Amount of runtime to allocate from global (tg) to local (per-cfs_rq) pool
170 * each time a cfs_rq requests quota.
172 * Note: in the case that the slice exceeds the runtime remaining (either due
173 * to consumption or the quota being specified to be smaller than the slice)
174 * we will always only issue the remaining available time.
176 * (default: 5 msec, units: microseconds)
178 static unsigned int sysctl_sched_cfs_bandwidth_slice = 5000UL;
181 #ifdef CONFIG_NUMA_BALANCING
182 /* Restrict the NUMA promotion throughput (MB/s) for each target node. */
183 static unsigned int sysctl_numa_balancing_promote_rate_limit = 65536;
187 static struct ctl_table sched_fair_sysctls[] = {
189 .procname = "sched_child_runs_first",
190 .data = &sysctl_sched_child_runs_first,
191 .maxlen = sizeof(unsigned int),
193 .proc_handler = proc_dointvec,
195 #ifdef CONFIG_CFS_BANDWIDTH
197 .procname = "sched_cfs_bandwidth_slice_us",
198 .data = &sysctl_sched_cfs_bandwidth_slice,
199 .maxlen = sizeof(unsigned int),
201 .proc_handler = proc_dointvec_minmax,
202 .extra1 = SYSCTL_ONE,
205 #ifdef CONFIG_NUMA_BALANCING
207 .procname = "numa_balancing_promote_rate_limit_MBps",
208 .data = &sysctl_numa_balancing_promote_rate_limit,
209 .maxlen = sizeof(unsigned int),
211 .proc_handler = proc_dointvec_minmax,
212 .extra1 = SYSCTL_ZERO,
214 #endif /* CONFIG_NUMA_BALANCING */
218 static int __init sched_fair_sysctl_init(void)
220 register_sysctl_init("kernel", sched_fair_sysctls);
223 late_initcall(sched_fair_sysctl_init);
226 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
232 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
238 static inline void update_load_set(struct load_weight *lw, unsigned long w)
245 * Increase the granularity value when there are more CPUs,
246 * because with more CPUs the 'effective latency' as visible
247 * to users decreases. But the relationship is not linear,
248 * so pick a second-best guess by going with the log2 of the
251 * This idea comes from the SD scheduler of Con Kolivas:
253 static unsigned int get_update_sysctl_factor(void)
255 unsigned int cpus = min_t(unsigned int, num_online_cpus(), 8);
258 switch (sysctl_sched_tunable_scaling) {
259 case SCHED_TUNABLESCALING_NONE:
262 case SCHED_TUNABLESCALING_LINEAR:
265 case SCHED_TUNABLESCALING_LOG:
267 factor = 1 + ilog2(cpus);
274 static void update_sysctl(void)
276 unsigned int factor = get_update_sysctl_factor();
278 #define SET_SYSCTL(name) \
279 (sysctl_##name = (factor) * normalized_sysctl_##name)
280 SET_SYSCTL(sched_min_granularity);
281 SET_SYSCTL(sched_latency);
282 SET_SYSCTL(sched_wakeup_granularity);
286 void __init sched_init_granularity(void)
291 #define WMULT_CONST (~0U)
292 #define WMULT_SHIFT 32
294 static void __update_inv_weight(struct load_weight *lw)
298 if (likely(lw->inv_weight))
301 w = scale_load_down(lw->weight);
303 if (BITS_PER_LONG > 32 && unlikely(w >= WMULT_CONST))
305 else if (unlikely(!w))
306 lw->inv_weight = WMULT_CONST;
308 lw->inv_weight = WMULT_CONST / w;
312 * delta_exec * weight / lw.weight
314 * (delta_exec * (weight * lw->inv_weight)) >> WMULT_SHIFT
316 * Either weight := NICE_0_LOAD and lw \e sched_prio_to_wmult[], in which case
317 * we're guaranteed shift stays positive because inv_weight is guaranteed to
318 * fit 32 bits, and NICE_0_LOAD gives another 10 bits; therefore shift >= 22.
320 * Or, weight =< lw.weight (because lw.weight is the runqueue weight), thus
321 * weight/lw.weight <= 1, and therefore our shift will also be positive.
323 static u64 __calc_delta(u64 delta_exec, unsigned long weight, struct load_weight *lw)
325 u64 fact = scale_load_down(weight);
326 u32 fact_hi = (u32)(fact >> 32);
327 int shift = WMULT_SHIFT;
330 __update_inv_weight(lw);
332 if (unlikely(fact_hi)) {
338 fact = mul_u32_u32(fact, lw->inv_weight);
340 fact_hi = (u32)(fact >> 32);
347 return mul_u64_u32_shr(delta_exec, fact, shift);
351 const struct sched_class fair_sched_class;
353 /**************************************************************
354 * CFS operations on generic schedulable entities:
357 #ifdef CONFIG_FAIR_GROUP_SCHED
359 /* Walk up scheduling entities hierarchy */
360 #define for_each_sched_entity(se) \
361 for (; se; se = se->parent)
363 static inline bool list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
365 struct rq *rq = rq_of(cfs_rq);
366 int cpu = cpu_of(rq);
369 return rq->tmp_alone_branch == &rq->leaf_cfs_rq_list;
374 * Ensure we either appear before our parent (if already
375 * enqueued) or force our parent to appear after us when it is
376 * enqueued. The fact that we always enqueue bottom-up
377 * reduces this to two cases and a special case for the root
378 * cfs_rq. Furthermore, it also means that we will always reset
379 * tmp_alone_branch either when the branch is connected
380 * to a tree or when we reach the top of the tree
382 if (cfs_rq->tg->parent &&
383 cfs_rq->tg->parent->cfs_rq[cpu]->on_list) {
385 * If parent is already on the list, we add the child
386 * just before. Thanks to circular linked property of
387 * the list, this means to put the child at the tail
388 * of the list that starts by parent.
390 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
391 &(cfs_rq->tg->parent->cfs_rq[cpu]->leaf_cfs_rq_list));
393 * The branch is now connected to its tree so we can
394 * reset tmp_alone_branch to the beginning of the
397 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
401 if (!cfs_rq->tg->parent) {
403 * cfs rq without parent should be put
404 * at the tail of the list.
406 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
407 &rq->leaf_cfs_rq_list);
409 * We have reach the top of a tree so we can reset
410 * tmp_alone_branch to the beginning of the list.
412 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
417 * The parent has not already been added so we want to
418 * make sure that it will be put after us.
419 * tmp_alone_branch points to the begin of the branch
420 * where we will add parent.
422 list_add_rcu(&cfs_rq->leaf_cfs_rq_list, rq->tmp_alone_branch);
424 * update tmp_alone_branch to points to the new begin
427 rq->tmp_alone_branch = &cfs_rq->leaf_cfs_rq_list;
431 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
433 if (cfs_rq->on_list) {
434 struct rq *rq = rq_of(cfs_rq);
437 * With cfs_rq being unthrottled/throttled during an enqueue,
438 * it can happen the tmp_alone_branch points the a leaf that
439 * we finally want to del. In this case, tmp_alone_branch moves
440 * to the prev element but it will point to rq->leaf_cfs_rq_list
441 * at the end of the enqueue.
443 if (rq->tmp_alone_branch == &cfs_rq->leaf_cfs_rq_list)
444 rq->tmp_alone_branch = cfs_rq->leaf_cfs_rq_list.prev;
446 list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
451 static inline void assert_list_leaf_cfs_rq(struct rq *rq)
453 SCHED_WARN_ON(rq->tmp_alone_branch != &rq->leaf_cfs_rq_list);
456 /* Iterate thr' all leaf cfs_rq's on a runqueue */
457 #define for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) \
458 list_for_each_entry_safe(cfs_rq, pos, &rq->leaf_cfs_rq_list, \
461 /* Do the two (enqueued) entities belong to the same group ? */
462 static inline struct cfs_rq *
463 is_same_group(struct sched_entity *se, struct sched_entity *pse)
465 if (se->cfs_rq == pse->cfs_rq)
471 static inline struct sched_entity *parent_entity(struct sched_entity *se)
477 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
479 int se_depth, pse_depth;
482 * preemption test can be made between sibling entities who are in the
483 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
484 * both tasks until we find their ancestors who are siblings of common
488 /* First walk up until both entities are at same depth */
489 se_depth = (*se)->depth;
490 pse_depth = (*pse)->depth;
492 while (se_depth > pse_depth) {
494 *se = parent_entity(*se);
497 while (pse_depth > se_depth) {
499 *pse = parent_entity(*pse);
502 while (!is_same_group(*se, *pse)) {
503 *se = parent_entity(*se);
504 *pse = parent_entity(*pse);
508 static int tg_is_idle(struct task_group *tg)
513 static int cfs_rq_is_idle(struct cfs_rq *cfs_rq)
515 return cfs_rq->idle > 0;
518 static int se_is_idle(struct sched_entity *se)
520 if (entity_is_task(se))
521 return task_has_idle_policy(task_of(se));
522 return cfs_rq_is_idle(group_cfs_rq(se));
525 #else /* !CONFIG_FAIR_GROUP_SCHED */
527 #define for_each_sched_entity(se) \
528 for (; se; se = NULL)
530 static inline bool list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
535 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
539 static inline void assert_list_leaf_cfs_rq(struct rq *rq)
543 #define for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) \
544 for (cfs_rq = &rq->cfs, pos = NULL; cfs_rq; cfs_rq = pos)
546 static inline struct sched_entity *parent_entity(struct sched_entity *se)
552 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
556 static inline int tg_is_idle(struct task_group *tg)
561 static int cfs_rq_is_idle(struct cfs_rq *cfs_rq)
566 static int se_is_idle(struct sched_entity *se)
571 #endif /* CONFIG_FAIR_GROUP_SCHED */
573 static __always_inline
574 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec);
576 /**************************************************************
577 * Scheduling class tree data structure manipulation methods:
580 static inline u64 max_vruntime(u64 max_vruntime, u64 vruntime)
582 s64 delta = (s64)(vruntime - max_vruntime);
584 max_vruntime = vruntime;
589 static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime)
591 s64 delta = (s64)(vruntime - min_vruntime);
593 min_vruntime = vruntime;
598 static inline bool entity_before(struct sched_entity *a,
599 struct sched_entity *b)
601 return (s64)(a->vruntime - b->vruntime) < 0;
604 #define __node_2_se(node) \
605 rb_entry((node), struct sched_entity, run_node)
607 static void update_min_vruntime(struct cfs_rq *cfs_rq)
609 struct sched_entity *curr = cfs_rq->curr;
610 struct rb_node *leftmost = rb_first_cached(&cfs_rq->tasks_timeline);
612 u64 vruntime = cfs_rq->min_vruntime;
616 vruntime = curr->vruntime;
621 if (leftmost) { /* non-empty tree */
622 struct sched_entity *se = __node_2_se(leftmost);
625 vruntime = se->vruntime;
627 vruntime = min_vruntime(vruntime, se->vruntime);
630 /* ensure we never gain time by being placed backwards. */
631 u64_u32_store(cfs_rq->min_vruntime,
632 max_vruntime(cfs_rq->min_vruntime, vruntime));
635 static inline bool __entity_less(struct rb_node *a, const struct rb_node *b)
637 return entity_before(__node_2_se(a), __node_2_se(b));
641 * Enqueue an entity into the rb-tree:
643 static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
645 rb_add_cached(&se->run_node, &cfs_rq->tasks_timeline, __entity_less);
648 static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
650 rb_erase_cached(&se->run_node, &cfs_rq->tasks_timeline);
653 struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq)
655 struct rb_node *left = rb_first_cached(&cfs_rq->tasks_timeline);
660 return __node_2_se(left);
663 static struct sched_entity *__pick_next_entity(struct sched_entity *se)
665 struct rb_node *next = rb_next(&se->run_node);
670 return __node_2_se(next);
673 #ifdef CONFIG_SCHED_DEBUG
674 struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
676 struct rb_node *last = rb_last(&cfs_rq->tasks_timeline.rb_root);
681 return __node_2_se(last);
684 /**************************************************************
685 * Scheduling class statistics methods:
688 int sched_update_scaling(void)
690 unsigned int factor = get_update_sysctl_factor();
692 sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency,
693 sysctl_sched_min_granularity);
695 #define WRT_SYSCTL(name) \
696 (normalized_sysctl_##name = sysctl_##name / (factor))
697 WRT_SYSCTL(sched_min_granularity);
698 WRT_SYSCTL(sched_latency);
699 WRT_SYSCTL(sched_wakeup_granularity);
709 static inline u64 calc_delta_fair(u64 delta, struct sched_entity *se)
711 if (unlikely(se->load.weight != NICE_0_LOAD))
712 delta = __calc_delta(delta, NICE_0_LOAD, &se->load);
718 * The idea is to set a period in which each task runs once.
720 * When there are too many tasks (sched_nr_latency) we have to stretch
721 * this period because otherwise the slices get too small.
723 * p = (nr <= nl) ? l : l*nr/nl
725 static u64 __sched_period(unsigned long nr_running)
727 if (unlikely(nr_running > sched_nr_latency))
728 return nr_running * sysctl_sched_min_granularity;
730 return sysctl_sched_latency;
733 static bool sched_idle_cfs_rq(struct cfs_rq *cfs_rq);
736 * We calculate the wall-time slice from the period by taking a part
737 * proportional to the weight.
741 static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
743 unsigned int nr_running = cfs_rq->nr_running;
744 struct sched_entity *init_se = se;
745 unsigned int min_gran;
748 if (sched_feat(ALT_PERIOD))
749 nr_running = rq_of(cfs_rq)->cfs.h_nr_running;
751 slice = __sched_period(nr_running + !se->on_rq);
753 for_each_sched_entity(se) {
754 struct load_weight *load;
755 struct load_weight lw;
756 struct cfs_rq *qcfs_rq;
758 qcfs_rq = cfs_rq_of(se);
759 load = &qcfs_rq->load;
761 if (unlikely(!se->on_rq)) {
764 update_load_add(&lw, se->load.weight);
767 slice = __calc_delta(slice, se->load.weight, load);
770 if (sched_feat(BASE_SLICE)) {
771 if (se_is_idle(init_se) && !sched_idle_cfs_rq(cfs_rq))
772 min_gran = sysctl_sched_idle_min_granularity;
774 min_gran = sysctl_sched_min_granularity;
776 slice = max_t(u64, slice, min_gran);
783 * We calculate the vruntime slice of a to-be-inserted task.
787 static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se)
789 return calc_delta_fair(sched_slice(cfs_rq, se), se);
795 static int select_idle_sibling(struct task_struct *p, int prev_cpu, int cpu);
796 static unsigned long task_h_load(struct task_struct *p);
797 static unsigned long capacity_of(int cpu);
799 /* Give new sched_entity start runnable values to heavy its load in infant time */
800 void init_entity_runnable_average(struct sched_entity *se)
802 struct sched_avg *sa = &se->avg;
804 memset(sa, 0, sizeof(*sa));
807 * Tasks are initialized with full load to be seen as heavy tasks until
808 * they get a chance to stabilize to their real load level.
809 * Group entities are initialized with zero load to reflect the fact that
810 * nothing has been attached to the task group yet.
812 if (entity_is_task(se))
813 sa->load_avg = scale_load_down(se->load.weight);
815 /* when this task enqueue'ed, it will contribute to its cfs_rq's load_avg */
819 * With new tasks being created, their initial util_avgs are extrapolated
820 * based on the cfs_rq's current util_avg:
822 * util_avg = cfs_rq->util_avg / (cfs_rq->load_avg + 1) * se.load.weight
824 * However, in many cases, the above util_avg does not give a desired
825 * value. Moreover, the sum of the util_avgs may be divergent, such
826 * as when the series is a harmonic series.
828 * To solve this problem, we also cap the util_avg of successive tasks to
829 * only 1/2 of the left utilization budget:
831 * util_avg_cap = (cpu_scale - cfs_rq->avg.util_avg) / 2^n
833 * where n denotes the nth task and cpu_scale the CPU capacity.
835 * For example, for a CPU with 1024 of capacity, a simplest series from
836 * the beginning would be like:
838 * task util_avg: 512, 256, 128, 64, 32, 16, 8, ...
839 * cfs_rq util_avg: 512, 768, 896, 960, 992, 1008, 1016, ...
841 * Finally, that extrapolated util_avg is clamped to the cap (util_avg_cap)
842 * if util_avg > util_avg_cap.
844 void post_init_entity_util_avg(struct task_struct *p)
846 struct sched_entity *se = &p->se;
847 struct cfs_rq *cfs_rq = cfs_rq_of(se);
848 struct sched_avg *sa = &se->avg;
849 long cpu_scale = arch_scale_cpu_capacity(cpu_of(rq_of(cfs_rq)));
850 long cap = (long)(cpu_scale - cfs_rq->avg.util_avg) / 2;
852 if (p->sched_class != &fair_sched_class) {
854 * For !fair tasks do:
856 update_cfs_rq_load_avg(now, cfs_rq);
857 attach_entity_load_avg(cfs_rq, se);
858 switched_from_fair(rq, p);
860 * such that the next switched_to_fair() has the
863 se->avg.last_update_time = cfs_rq_clock_pelt(cfs_rq);
868 if (cfs_rq->avg.util_avg != 0) {
869 sa->util_avg = cfs_rq->avg.util_avg * se->load.weight;
870 sa->util_avg /= (cfs_rq->avg.load_avg + 1);
872 if (sa->util_avg > cap)
879 sa->runnable_avg = sa->util_avg;
882 #else /* !CONFIG_SMP */
883 void init_entity_runnable_average(struct sched_entity *se)
886 void post_init_entity_util_avg(struct task_struct *p)
889 static void update_tg_load_avg(struct cfs_rq *cfs_rq)
892 #endif /* CONFIG_SMP */
895 * Update the current task's runtime statistics.
897 static void update_curr(struct cfs_rq *cfs_rq)
899 struct sched_entity *curr = cfs_rq->curr;
900 u64 now = rq_clock_task(rq_of(cfs_rq));
906 delta_exec = now - curr->exec_start;
907 if (unlikely((s64)delta_exec <= 0))
910 curr->exec_start = now;
912 if (schedstat_enabled()) {
913 struct sched_statistics *stats;
915 stats = __schedstats_from_se(curr);
916 __schedstat_set(stats->exec_max,
917 max(delta_exec, stats->exec_max));
920 curr->sum_exec_runtime += delta_exec;
921 schedstat_add(cfs_rq->exec_clock, delta_exec);
923 curr->vruntime += calc_delta_fair(delta_exec, curr);
924 update_min_vruntime(cfs_rq);
926 if (entity_is_task(curr)) {
927 struct task_struct *curtask = task_of(curr);
929 trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime);
930 cgroup_account_cputime(curtask, delta_exec);
931 account_group_exec_runtime(curtask, delta_exec);
934 account_cfs_rq_runtime(cfs_rq, delta_exec);
937 static void update_curr_fair(struct rq *rq)
939 update_curr(cfs_rq_of(&rq->curr->se));
943 update_stats_wait_start_fair(struct cfs_rq *cfs_rq, struct sched_entity *se)
945 struct sched_statistics *stats;
946 struct task_struct *p = NULL;
948 if (!schedstat_enabled())
951 stats = __schedstats_from_se(se);
953 if (entity_is_task(se))
956 __update_stats_wait_start(rq_of(cfs_rq), p, stats);
960 update_stats_wait_end_fair(struct cfs_rq *cfs_rq, struct sched_entity *se)
962 struct sched_statistics *stats;
963 struct task_struct *p = NULL;
965 if (!schedstat_enabled())
968 stats = __schedstats_from_se(se);
971 * When the sched_schedstat changes from 0 to 1, some sched se
972 * maybe already in the runqueue, the se->statistics.wait_start
973 * will be 0.So it will let the delta wrong. We need to avoid this
976 if (unlikely(!schedstat_val(stats->wait_start)))
979 if (entity_is_task(se))
982 __update_stats_wait_end(rq_of(cfs_rq), p, stats);
986 update_stats_enqueue_sleeper_fair(struct cfs_rq *cfs_rq, struct sched_entity *se)
988 struct sched_statistics *stats;
989 struct task_struct *tsk = NULL;
991 if (!schedstat_enabled())
994 stats = __schedstats_from_se(se);
996 if (entity_is_task(se))
999 __update_stats_enqueue_sleeper(rq_of(cfs_rq), tsk, stats);
1003 * Task is being enqueued - update stats:
1006 update_stats_enqueue_fair(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
1008 if (!schedstat_enabled())
1012 * Are we enqueueing a waiting task? (for current tasks
1013 * a dequeue/enqueue event is a NOP)
1015 if (se != cfs_rq->curr)
1016 update_stats_wait_start_fair(cfs_rq, se);
1018 if (flags & ENQUEUE_WAKEUP)
1019 update_stats_enqueue_sleeper_fair(cfs_rq, se);
1023 update_stats_dequeue_fair(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
1026 if (!schedstat_enabled())
1030 * Mark the end of the wait period if dequeueing a
1033 if (se != cfs_rq->curr)
1034 update_stats_wait_end_fair(cfs_rq, se);
1036 if ((flags & DEQUEUE_SLEEP) && entity_is_task(se)) {
1037 struct task_struct *tsk = task_of(se);
1040 /* XXX racy against TTWU */
1041 state = READ_ONCE(tsk->__state);
1042 if (state & TASK_INTERRUPTIBLE)
1043 __schedstat_set(tsk->stats.sleep_start,
1044 rq_clock(rq_of(cfs_rq)));
1045 if (state & TASK_UNINTERRUPTIBLE)
1046 __schedstat_set(tsk->stats.block_start,
1047 rq_clock(rq_of(cfs_rq)));
1052 * We are picking a new current task - update its stats:
1055 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
1058 * We are starting a new run period:
1060 se->exec_start = rq_clock_task(rq_of(cfs_rq));
1063 /**************************************************
1064 * Scheduling class queueing methods:
1068 #define NUMA_IMBALANCE_MIN 2
1071 adjust_numa_imbalance(int imbalance, int dst_running, int imb_numa_nr)
1074 * Allow a NUMA imbalance if busy CPUs is less than the maximum
1075 * threshold. Above this threshold, individual tasks may be contending
1076 * for both memory bandwidth and any shared HT resources. This is an
1077 * approximation as the number of running tasks may not be related to
1078 * the number of busy CPUs due to sched_setaffinity.
1080 if (dst_running > imb_numa_nr)
1084 * Allow a small imbalance based on a simple pair of communicating
1085 * tasks that remain local when the destination is lightly loaded.
1087 if (imbalance <= NUMA_IMBALANCE_MIN)
1092 #endif /* CONFIG_NUMA */
1094 #ifdef CONFIG_NUMA_BALANCING
1096 * Approximate time to scan a full NUMA task in ms. The task scan period is
1097 * calculated based on the tasks virtual memory size and
1098 * numa_balancing_scan_size.
1100 unsigned int sysctl_numa_balancing_scan_period_min = 1000;
1101 unsigned int sysctl_numa_balancing_scan_period_max = 60000;
1103 /* Portion of address space to scan in MB */
1104 unsigned int sysctl_numa_balancing_scan_size = 256;
1106 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
1107 unsigned int sysctl_numa_balancing_scan_delay = 1000;
1109 /* The page with hint page fault latency < threshold in ms is considered hot */
1110 unsigned int sysctl_numa_balancing_hot_threshold = MSEC_PER_SEC;
1113 refcount_t refcount;
1115 spinlock_t lock; /* nr_tasks, tasks */
1120 struct rcu_head rcu;
1121 unsigned long total_faults;
1122 unsigned long max_faults_cpu;
1124 * faults[] array is split into two regions: faults_mem and faults_cpu.
1126 * Faults_cpu is used to decide whether memory should move
1127 * towards the CPU. As a consequence, these stats are weighted
1128 * more by CPU use than by memory faults.
1130 unsigned long faults[];
1134 * For functions that can be called in multiple contexts that permit reading
1135 * ->numa_group (see struct task_struct for locking rules).
1137 static struct numa_group *deref_task_numa_group(struct task_struct *p)
1139 return rcu_dereference_check(p->numa_group, p == current ||
1140 (lockdep_is_held(__rq_lockp(task_rq(p))) && !READ_ONCE(p->on_cpu)));
1143 static struct numa_group *deref_curr_numa_group(struct task_struct *p)
1145 return rcu_dereference_protected(p->numa_group, p == current);
1148 static inline unsigned long group_faults_priv(struct numa_group *ng);
1149 static inline unsigned long group_faults_shared(struct numa_group *ng);
1151 static unsigned int task_nr_scan_windows(struct task_struct *p)
1153 unsigned long rss = 0;
1154 unsigned long nr_scan_pages;
1157 * Calculations based on RSS as non-present and empty pages are skipped
1158 * by the PTE scanner and NUMA hinting faults should be trapped based
1161 nr_scan_pages = sysctl_numa_balancing_scan_size << (20 - PAGE_SHIFT);
1162 rss = get_mm_rss(p->mm);
1164 rss = nr_scan_pages;
1166 rss = round_up(rss, nr_scan_pages);
1167 return rss / nr_scan_pages;
1170 /* For sanity's sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */
1171 #define MAX_SCAN_WINDOW 2560
1173 static unsigned int task_scan_min(struct task_struct *p)
1175 unsigned int scan_size = READ_ONCE(sysctl_numa_balancing_scan_size);
1176 unsigned int scan, floor;
1177 unsigned int windows = 1;
1179 if (scan_size < MAX_SCAN_WINDOW)
1180 windows = MAX_SCAN_WINDOW / scan_size;
1181 floor = 1000 / windows;
1183 scan = sysctl_numa_balancing_scan_period_min / task_nr_scan_windows(p);
1184 return max_t(unsigned int, floor, scan);
1187 static unsigned int task_scan_start(struct task_struct *p)
1189 unsigned long smin = task_scan_min(p);
1190 unsigned long period = smin;
1191 struct numa_group *ng;
1193 /* Scale the maximum scan period with the amount of shared memory. */
1195 ng = rcu_dereference(p->numa_group);
1197 unsigned long shared = group_faults_shared(ng);
1198 unsigned long private = group_faults_priv(ng);
1200 period *= refcount_read(&ng->refcount);
1201 period *= shared + 1;
1202 period /= private + shared + 1;
1206 return max(smin, period);
1209 static unsigned int task_scan_max(struct task_struct *p)
1211 unsigned long smin = task_scan_min(p);
1213 struct numa_group *ng;
1215 /* Watch for min being lower than max due to floor calculations */
1216 smax = sysctl_numa_balancing_scan_period_max / task_nr_scan_windows(p);
1218 /* Scale the maximum scan period with the amount of shared memory. */
1219 ng = deref_curr_numa_group(p);
1221 unsigned long shared = group_faults_shared(ng);
1222 unsigned long private = group_faults_priv(ng);
1223 unsigned long period = smax;
1225 period *= refcount_read(&ng->refcount);
1226 period *= shared + 1;
1227 period /= private + shared + 1;
1229 smax = max(smax, period);
1232 return max(smin, smax);
1235 static void account_numa_enqueue(struct rq *rq, struct task_struct *p)
1237 rq->nr_numa_running += (p->numa_preferred_nid != NUMA_NO_NODE);
1238 rq->nr_preferred_running += (p->numa_preferred_nid == task_node(p));
1241 static void account_numa_dequeue(struct rq *rq, struct task_struct *p)
1243 rq->nr_numa_running -= (p->numa_preferred_nid != NUMA_NO_NODE);
1244 rq->nr_preferred_running -= (p->numa_preferred_nid == task_node(p));
1247 /* Shared or private faults. */
1248 #define NR_NUMA_HINT_FAULT_TYPES 2
1250 /* Memory and CPU locality */
1251 #define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2)
1253 /* Averaged statistics, and temporary buffers. */
1254 #define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2)
1256 pid_t task_numa_group_id(struct task_struct *p)
1258 struct numa_group *ng;
1262 ng = rcu_dereference(p->numa_group);
1271 * The averaged statistics, shared & private, memory & CPU,
1272 * occupy the first half of the array. The second half of the
1273 * array is for current counters, which are averaged into the
1274 * first set by task_numa_placement.
1276 static inline int task_faults_idx(enum numa_faults_stats s, int nid, int priv)
1278 return NR_NUMA_HINT_FAULT_TYPES * (s * nr_node_ids + nid) + priv;
1281 static inline unsigned long task_faults(struct task_struct *p, int nid)
1283 if (!p->numa_faults)
1286 return p->numa_faults[task_faults_idx(NUMA_MEM, nid, 0)] +
1287 p->numa_faults[task_faults_idx(NUMA_MEM, nid, 1)];
1290 static inline unsigned long group_faults(struct task_struct *p, int nid)
1292 struct numa_group *ng = deref_task_numa_group(p);
1297 return ng->faults[task_faults_idx(NUMA_MEM, nid, 0)] +
1298 ng->faults[task_faults_idx(NUMA_MEM, nid, 1)];
1301 static inline unsigned long group_faults_cpu(struct numa_group *group, int nid)
1303 return group->faults[task_faults_idx(NUMA_CPU, nid, 0)] +
1304 group->faults[task_faults_idx(NUMA_CPU, nid, 1)];
1307 static inline unsigned long group_faults_priv(struct numa_group *ng)
1309 unsigned long faults = 0;
1312 for_each_online_node(node) {
1313 faults += ng->faults[task_faults_idx(NUMA_MEM, node, 1)];
1319 static inline unsigned long group_faults_shared(struct numa_group *ng)
1321 unsigned long faults = 0;
1324 for_each_online_node(node) {
1325 faults += ng->faults[task_faults_idx(NUMA_MEM, node, 0)];
1332 * A node triggering more than 1/3 as many NUMA faults as the maximum is
1333 * considered part of a numa group's pseudo-interleaving set. Migrations
1334 * between these nodes are slowed down, to allow things to settle down.
1336 #define ACTIVE_NODE_FRACTION 3
1338 static bool numa_is_active_node(int nid, struct numa_group *ng)
1340 return group_faults_cpu(ng, nid) * ACTIVE_NODE_FRACTION > ng->max_faults_cpu;
1343 /* Handle placement on systems where not all nodes are directly connected. */
1344 static unsigned long score_nearby_nodes(struct task_struct *p, int nid,
1345 int lim_dist, bool task)
1347 unsigned long score = 0;
1351 * All nodes are directly connected, and the same distance
1352 * from each other. No need for fancy placement algorithms.
1354 if (sched_numa_topology_type == NUMA_DIRECT)
1357 /* sched_max_numa_distance may be changed in parallel. */
1358 max_dist = READ_ONCE(sched_max_numa_distance);
1360 * This code is called for each node, introducing N^2 complexity,
1361 * which should be ok given the number of nodes rarely exceeds 8.
1363 for_each_online_node(node) {
1364 unsigned long faults;
1365 int dist = node_distance(nid, node);
1368 * The furthest away nodes in the system are not interesting
1369 * for placement; nid was already counted.
1371 if (dist >= max_dist || node == nid)
1375 * On systems with a backplane NUMA topology, compare groups
1376 * of nodes, and move tasks towards the group with the most
1377 * memory accesses. When comparing two nodes at distance
1378 * "hoplimit", only nodes closer by than "hoplimit" are part
1379 * of each group. Skip other nodes.
1381 if (sched_numa_topology_type == NUMA_BACKPLANE && dist >= lim_dist)
1384 /* Add up the faults from nearby nodes. */
1386 faults = task_faults(p, node);
1388 faults = group_faults(p, node);
1391 * On systems with a glueless mesh NUMA topology, there are
1392 * no fixed "groups of nodes". Instead, nodes that are not
1393 * directly connected bounce traffic through intermediate
1394 * nodes; a numa_group can occupy any set of nodes.
1395 * The further away a node is, the less the faults count.
1396 * This seems to result in good task placement.
1398 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
1399 faults *= (max_dist - dist);
1400 faults /= (max_dist - LOCAL_DISTANCE);
1410 * These return the fraction of accesses done by a particular task, or
1411 * task group, on a particular numa node. The group weight is given a
1412 * larger multiplier, in order to group tasks together that are almost
1413 * evenly spread out between numa nodes.
1415 static inline unsigned long task_weight(struct task_struct *p, int nid,
1418 unsigned long faults, total_faults;
1420 if (!p->numa_faults)
1423 total_faults = p->total_numa_faults;
1428 faults = task_faults(p, nid);
1429 faults += score_nearby_nodes(p, nid, dist, true);
1431 return 1000 * faults / total_faults;
1434 static inline unsigned long group_weight(struct task_struct *p, int nid,
1437 struct numa_group *ng = deref_task_numa_group(p);
1438 unsigned long faults, total_faults;
1443 total_faults = ng->total_faults;
1448 faults = group_faults(p, nid);
1449 faults += score_nearby_nodes(p, nid, dist, false);
1451 return 1000 * faults / total_faults;
1455 * If memory tiering mode is enabled, cpupid of slow memory page is
1456 * used to record scan time instead of CPU and PID. When tiering mode
1457 * is disabled at run time, the scan time (in cpupid) will be
1458 * interpreted as CPU and PID. So CPU needs to be checked to avoid to
1459 * access out of array bound.
1461 static inline bool cpupid_valid(int cpupid)
1463 return cpupid_to_cpu(cpupid) < nr_cpu_ids;
1467 * For memory tiering mode, if there are enough free pages (more than
1468 * enough watermark defined here) in fast memory node, to take full
1469 * advantage of fast memory capacity, all recently accessed slow
1470 * memory pages will be migrated to fast memory node without
1471 * considering hot threshold.
1473 static bool pgdat_free_space_enough(struct pglist_data *pgdat)
1476 unsigned long enough_wmark;
1478 enough_wmark = max(1UL * 1024 * 1024 * 1024 >> PAGE_SHIFT,
1479 pgdat->node_present_pages >> 4);
1480 for (z = pgdat->nr_zones - 1; z >= 0; z--) {
1481 struct zone *zone = pgdat->node_zones + z;
1483 if (!populated_zone(zone))
1486 if (zone_watermark_ok(zone, 0,
1487 wmark_pages(zone, WMARK_PROMO) + enough_wmark,
1495 * For memory tiering mode, when page tables are scanned, the scan
1496 * time will be recorded in struct page in addition to make page
1497 * PROT_NONE for slow memory page. So when the page is accessed, in
1498 * hint page fault handler, the hint page fault latency is calculated
1501 * hint page fault latency = hint page fault time - scan time
1503 * The smaller the hint page fault latency, the higher the possibility
1504 * for the page to be hot.
1506 static int numa_hint_fault_latency(struct page *page)
1508 int last_time, time;
1510 time = jiffies_to_msecs(jiffies);
1511 last_time = xchg_page_access_time(page, time);
1513 return (time - last_time) & PAGE_ACCESS_TIME_MASK;
1517 * For memory tiering mode, too high promotion/demotion throughput may
1518 * hurt application latency. So we provide a mechanism to rate limit
1519 * the number of pages that are tried to be promoted.
1521 static bool numa_promotion_rate_limit(struct pglist_data *pgdat,
1522 unsigned long rate_limit, int nr)
1524 unsigned long nr_cand;
1525 unsigned int now, start;
1527 now = jiffies_to_msecs(jiffies);
1528 mod_node_page_state(pgdat, PGPROMOTE_CANDIDATE, nr);
1529 nr_cand = node_page_state(pgdat, PGPROMOTE_CANDIDATE);
1530 start = pgdat->nbp_rl_start;
1531 if (now - start > MSEC_PER_SEC &&
1532 cmpxchg(&pgdat->nbp_rl_start, start, now) == start)
1533 pgdat->nbp_rl_nr_cand = nr_cand;
1534 if (nr_cand - pgdat->nbp_rl_nr_cand >= rate_limit)
1539 #define NUMA_MIGRATION_ADJUST_STEPS 16
1541 static void numa_promotion_adjust_threshold(struct pglist_data *pgdat,
1542 unsigned long rate_limit,
1543 unsigned int ref_th)
1545 unsigned int now, start, th_period, unit_th, th;
1546 unsigned long nr_cand, ref_cand, diff_cand;
1548 now = jiffies_to_msecs(jiffies);
1549 th_period = sysctl_numa_balancing_scan_period_max;
1550 start = pgdat->nbp_th_start;
1551 if (now - start > th_period &&
1552 cmpxchg(&pgdat->nbp_th_start, start, now) == start) {
1553 ref_cand = rate_limit *
1554 sysctl_numa_balancing_scan_period_max / MSEC_PER_SEC;
1555 nr_cand = node_page_state(pgdat, PGPROMOTE_CANDIDATE);
1556 diff_cand = nr_cand - pgdat->nbp_th_nr_cand;
1557 unit_th = ref_th * 2 / NUMA_MIGRATION_ADJUST_STEPS;
1558 th = pgdat->nbp_threshold ? : ref_th;
1559 if (diff_cand > ref_cand * 11 / 10)
1560 th = max(th - unit_th, unit_th);
1561 else if (diff_cand < ref_cand * 9 / 10)
1562 th = min(th + unit_th, ref_th * 2);
1563 pgdat->nbp_th_nr_cand = nr_cand;
1564 pgdat->nbp_threshold = th;
1568 bool should_numa_migrate_memory(struct task_struct *p, struct page * page,
1569 int src_nid, int dst_cpu)
1571 struct numa_group *ng = deref_curr_numa_group(p);
1572 int dst_nid = cpu_to_node(dst_cpu);
1573 int last_cpupid, this_cpupid;
1576 * The pages in slow memory node should be migrated according
1577 * to hot/cold instead of private/shared.
1579 if (sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING &&
1580 !node_is_toptier(src_nid)) {
1581 struct pglist_data *pgdat;
1582 unsigned long rate_limit;
1583 unsigned int latency, th, def_th;
1585 pgdat = NODE_DATA(dst_nid);
1586 if (pgdat_free_space_enough(pgdat)) {
1587 /* workload changed, reset hot threshold */
1588 pgdat->nbp_threshold = 0;
1592 def_th = sysctl_numa_balancing_hot_threshold;
1593 rate_limit = sysctl_numa_balancing_promote_rate_limit << \
1595 numa_promotion_adjust_threshold(pgdat, rate_limit, def_th);
1597 th = pgdat->nbp_threshold ? : def_th;
1598 latency = numa_hint_fault_latency(page);
1602 return !numa_promotion_rate_limit(pgdat, rate_limit,
1603 thp_nr_pages(page));
1606 this_cpupid = cpu_pid_to_cpupid(dst_cpu, current->pid);
1607 last_cpupid = page_cpupid_xchg_last(page, this_cpupid);
1609 if (!(sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING) &&
1610 !node_is_toptier(src_nid) && !cpupid_valid(last_cpupid))
1614 * Allow first faults or private faults to migrate immediately early in
1615 * the lifetime of a task. The magic number 4 is based on waiting for
1616 * two full passes of the "multi-stage node selection" test that is
1619 if ((p->numa_preferred_nid == NUMA_NO_NODE || p->numa_scan_seq <= 4) &&
1620 (cpupid_pid_unset(last_cpupid) || cpupid_match_pid(p, last_cpupid)))
1624 * Multi-stage node selection is used in conjunction with a periodic
1625 * migration fault to build a temporal task<->page relation. By using
1626 * a two-stage filter we remove short/unlikely relations.
1628 * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate
1629 * a task's usage of a particular page (n_p) per total usage of this
1630 * page (n_t) (in a given time-span) to a probability.
1632 * Our periodic faults will sample this probability and getting the
1633 * same result twice in a row, given these samples are fully
1634 * independent, is then given by P(n)^2, provided our sample period
1635 * is sufficiently short compared to the usage pattern.
1637 * This quadric squishes small probabilities, making it less likely we
1638 * act on an unlikely task<->page relation.
1640 if (!cpupid_pid_unset(last_cpupid) &&
1641 cpupid_to_nid(last_cpupid) != dst_nid)
1644 /* Always allow migrate on private faults */
1645 if (cpupid_match_pid(p, last_cpupid))
1648 /* A shared fault, but p->numa_group has not been set up yet. */
1653 * Destination node is much more heavily used than the source
1654 * node? Allow migration.
1656 if (group_faults_cpu(ng, dst_nid) > group_faults_cpu(ng, src_nid) *
1657 ACTIVE_NODE_FRACTION)
1661 * Distribute memory according to CPU & memory use on each node,
1662 * with 3/4 hysteresis to avoid unnecessary memory migrations:
1664 * faults_cpu(dst) 3 faults_cpu(src)
1665 * --------------- * - > ---------------
1666 * faults_mem(dst) 4 faults_mem(src)
1668 return group_faults_cpu(ng, dst_nid) * group_faults(p, src_nid) * 3 >
1669 group_faults_cpu(ng, src_nid) * group_faults(p, dst_nid) * 4;
1673 * 'numa_type' describes the node at the moment of load balancing.
1676 /* The node has spare capacity that can be used to run more tasks. */
1679 * The node is fully used and the tasks don't compete for more CPU
1680 * cycles. Nevertheless, some tasks might wait before running.
1684 * The node is overloaded and can't provide expected CPU cycles to all
1690 /* Cached statistics for all CPUs within a node */
1693 unsigned long runnable;
1695 /* Total compute capacity of CPUs on a node */
1696 unsigned long compute_capacity;
1697 unsigned int nr_running;
1698 unsigned int weight;
1699 enum numa_type node_type;
1703 static inline bool is_core_idle(int cpu)
1705 #ifdef CONFIG_SCHED_SMT
1708 for_each_cpu(sibling, cpu_smt_mask(cpu)) {
1712 if (!idle_cpu(sibling))
1720 struct task_numa_env {
1721 struct task_struct *p;
1723 int src_cpu, src_nid;
1724 int dst_cpu, dst_nid;
1727 struct numa_stats src_stats, dst_stats;
1732 struct task_struct *best_task;
1737 static unsigned long cpu_load(struct rq *rq);
1738 static unsigned long cpu_runnable(struct rq *rq);
1741 numa_type numa_classify(unsigned int imbalance_pct,
1742 struct numa_stats *ns)
1744 if ((ns->nr_running > ns->weight) &&
1745 (((ns->compute_capacity * 100) < (ns->util * imbalance_pct)) ||
1746 ((ns->compute_capacity * imbalance_pct) < (ns->runnable * 100))))
1747 return node_overloaded;
1749 if ((ns->nr_running < ns->weight) ||
1750 (((ns->compute_capacity * 100) > (ns->util * imbalance_pct)) &&
1751 ((ns->compute_capacity * imbalance_pct) > (ns->runnable * 100))))
1752 return node_has_spare;
1754 return node_fully_busy;
1757 #ifdef CONFIG_SCHED_SMT
1758 /* Forward declarations of select_idle_sibling helpers */
1759 static inline bool test_idle_cores(int cpu);
1760 static inline int numa_idle_core(int idle_core, int cpu)
1762 if (!static_branch_likely(&sched_smt_present) ||
1763 idle_core >= 0 || !test_idle_cores(cpu))
1767 * Prefer cores instead of packing HT siblings
1768 * and triggering future load balancing.
1770 if (is_core_idle(cpu))
1776 static inline int numa_idle_core(int idle_core, int cpu)
1783 * Gather all necessary information to make NUMA balancing placement
1784 * decisions that are compatible with standard load balancer. This
1785 * borrows code and logic from update_sg_lb_stats but sharing a
1786 * common implementation is impractical.
1788 static void update_numa_stats(struct task_numa_env *env,
1789 struct numa_stats *ns, int nid,
1792 int cpu, idle_core = -1;
1794 memset(ns, 0, sizeof(*ns));
1798 for_each_cpu(cpu, cpumask_of_node(nid)) {
1799 struct rq *rq = cpu_rq(cpu);
1801 ns->load += cpu_load(rq);
1802 ns->runnable += cpu_runnable(rq);
1803 ns->util += cpu_util_cfs(cpu);
1804 ns->nr_running += rq->cfs.h_nr_running;
1805 ns->compute_capacity += capacity_of(cpu);
1807 if (find_idle && !rq->nr_running && idle_cpu(cpu)) {
1808 if (READ_ONCE(rq->numa_migrate_on) ||
1809 !cpumask_test_cpu(cpu, env->p->cpus_ptr))
1812 if (ns->idle_cpu == -1)
1815 idle_core = numa_idle_core(idle_core, cpu);
1820 ns->weight = cpumask_weight(cpumask_of_node(nid));
1822 ns->node_type = numa_classify(env->imbalance_pct, ns);
1825 ns->idle_cpu = idle_core;
1828 static void task_numa_assign(struct task_numa_env *env,
1829 struct task_struct *p, long imp)
1831 struct rq *rq = cpu_rq(env->dst_cpu);
1833 /* Check if run-queue part of active NUMA balance. */
1834 if (env->best_cpu != env->dst_cpu && xchg(&rq->numa_migrate_on, 1)) {
1836 int start = env->dst_cpu;
1838 /* Find alternative idle CPU. */
1839 for_each_cpu_wrap(cpu, cpumask_of_node(env->dst_nid), start) {
1840 if (cpu == env->best_cpu || !idle_cpu(cpu) ||
1841 !cpumask_test_cpu(cpu, env->p->cpus_ptr)) {
1846 rq = cpu_rq(env->dst_cpu);
1847 if (!xchg(&rq->numa_migrate_on, 1))
1851 /* Failed to find an alternative idle CPU */
1857 * Clear previous best_cpu/rq numa-migrate flag, since task now
1858 * found a better CPU to move/swap.
1860 if (env->best_cpu != -1 && env->best_cpu != env->dst_cpu) {
1861 rq = cpu_rq(env->best_cpu);
1862 WRITE_ONCE(rq->numa_migrate_on, 0);
1866 put_task_struct(env->best_task);
1871 env->best_imp = imp;
1872 env->best_cpu = env->dst_cpu;
1875 static bool load_too_imbalanced(long src_load, long dst_load,
1876 struct task_numa_env *env)
1879 long orig_src_load, orig_dst_load;
1880 long src_capacity, dst_capacity;
1883 * The load is corrected for the CPU capacity available on each node.
1886 * ------------ vs ---------
1887 * src_capacity dst_capacity
1889 src_capacity = env->src_stats.compute_capacity;
1890 dst_capacity = env->dst_stats.compute_capacity;
1892 imb = abs(dst_load * src_capacity - src_load * dst_capacity);
1894 orig_src_load = env->src_stats.load;
1895 orig_dst_load = env->dst_stats.load;
1897 old_imb = abs(orig_dst_load * src_capacity - orig_src_load * dst_capacity);
1899 /* Would this change make things worse? */
1900 return (imb > old_imb);
1904 * Maximum NUMA importance can be 1998 (2*999);
1905 * SMALLIMP @ 30 would be close to 1998/64.
1906 * Used to deter task migration.
1911 * This checks if the overall compute and NUMA accesses of the system would
1912 * be improved if the source tasks was migrated to the target dst_cpu taking
1913 * into account that it might be best if task running on the dst_cpu should
1914 * be exchanged with the source task
1916 static bool task_numa_compare(struct task_numa_env *env,
1917 long taskimp, long groupimp, bool maymove)
1919 struct numa_group *cur_ng, *p_ng = deref_curr_numa_group(env->p);
1920 struct rq *dst_rq = cpu_rq(env->dst_cpu);
1921 long imp = p_ng ? groupimp : taskimp;
1922 struct task_struct *cur;
1923 long src_load, dst_load;
1924 int dist = env->dist;
1927 bool stopsearch = false;
1929 if (READ_ONCE(dst_rq->numa_migrate_on))
1933 cur = rcu_dereference(dst_rq->curr);
1934 if (cur && ((cur->flags & PF_EXITING) || is_idle_task(cur)))
1938 * Because we have preemption enabled we can get migrated around and
1939 * end try selecting ourselves (current == env->p) as a swap candidate.
1941 if (cur == env->p) {
1947 if (maymove && moveimp >= env->best_imp)
1953 /* Skip this swap candidate if cannot move to the source cpu. */
1954 if (!cpumask_test_cpu(env->src_cpu, cur->cpus_ptr))
1958 * Skip this swap candidate if it is not moving to its preferred
1959 * node and the best task is.
1961 if (env->best_task &&
1962 env->best_task->numa_preferred_nid == env->src_nid &&
1963 cur->numa_preferred_nid != env->src_nid) {
1968 * "imp" is the fault differential for the source task between the
1969 * source and destination node. Calculate the total differential for
1970 * the source task and potential destination task. The more negative
1971 * the value is, the more remote accesses that would be expected to
1972 * be incurred if the tasks were swapped.
1974 * If dst and source tasks are in the same NUMA group, or not
1975 * in any group then look only at task weights.
1977 cur_ng = rcu_dereference(cur->numa_group);
1978 if (cur_ng == p_ng) {
1980 * Do not swap within a group or between tasks that have
1981 * no group if there is spare capacity. Swapping does
1982 * not address the load imbalance and helps one task at
1983 * the cost of punishing another.
1985 if (env->dst_stats.node_type == node_has_spare)
1988 imp = taskimp + task_weight(cur, env->src_nid, dist) -
1989 task_weight(cur, env->dst_nid, dist);
1991 * Add some hysteresis to prevent swapping the
1992 * tasks within a group over tiny differences.
1998 * Compare the group weights. If a task is all by itself
1999 * (not part of a group), use the task weight instead.
2002 imp += group_weight(cur, env->src_nid, dist) -
2003 group_weight(cur, env->dst_nid, dist);
2005 imp += task_weight(cur, env->src_nid, dist) -
2006 task_weight(cur, env->dst_nid, dist);
2009 /* Discourage picking a task already on its preferred node */
2010 if (cur->numa_preferred_nid == env->dst_nid)
2014 * Encourage picking a task that moves to its preferred node.
2015 * This potentially makes imp larger than it's maximum of
2016 * 1998 (see SMALLIMP and task_weight for why) but in this
2017 * case, it does not matter.
2019 if (cur->numa_preferred_nid == env->src_nid)
2022 if (maymove && moveimp > imp && moveimp > env->best_imp) {
2029 * Prefer swapping with a task moving to its preferred node over a
2032 if (env->best_task && cur->numa_preferred_nid == env->src_nid &&
2033 env->best_task->numa_preferred_nid != env->src_nid) {
2038 * If the NUMA importance is less than SMALLIMP,
2039 * task migration might only result in ping pong
2040 * of tasks and also hurt performance due to cache
2043 if (imp < SMALLIMP || imp <= env->best_imp + SMALLIMP / 2)
2047 * In the overloaded case, try and keep the load balanced.
2049 load = task_h_load(env->p) - task_h_load(cur);
2053 dst_load = env->dst_stats.load + load;
2054 src_load = env->src_stats.load - load;
2056 if (load_too_imbalanced(src_load, dst_load, env))
2060 /* Evaluate an idle CPU for a task numa move. */
2062 int cpu = env->dst_stats.idle_cpu;
2064 /* Nothing cached so current CPU went idle since the search. */
2069 * If the CPU is no longer truly idle and the previous best CPU
2070 * is, keep using it.
2072 if (!idle_cpu(cpu) && env->best_cpu >= 0 &&
2073 idle_cpu(env->best_cpu)) {
2074 cpu = env->best_cpu;
2080 task_numa_assign(env, cur, imp);
2083 * If a move to idle is allowed because there is capacity or load
2084 * balance improves then stop the search. While a better swap
2085 * candidate may exist, a search is not free.
2087 if (maymove && !cur && env->best_cpu >= 0 && idle_cpu(env->best_cpu))
2091 * If a swap candidate must be identified and the current best task
2092 * moves its preferred node then stop the search.
2094 if (!maymove && env->best_task &&
2095 env->best_task->numa_preferred_nid == env->src_nid) {
2104 static void task_numa_find_cpu(struct task_numa_env *env,
2105 long taskimp, long groupimp)
2107 bool maymove = false;
2111 * If dst node has spare capacity, then check if there is an
2112 * imbalance that would be overruled by the load balancer.
2114 if (env->dst_stats.node_type == node_has_spare) {
2115 unsigned int imbalance;
2116 int src_running, dst_running;
2119 * Would movement cause an imbalance? Note that if src has
2120 * more running tasks that the imbalance is ignored as the
2121 * move improves the imbalance from the perspective of the
2122 * CPU load balancer.
2124 src_running = env->src_stats.nr_running - 1;
2125 dst_running = env->dst_stats.nr_running + 1;
2126 imbalance = max(0, dst_running - src_running);
2127 imbalance = adjust_numa_imbalance(imbalance, dst_running,
2130 /* Use idle CPU if there is no imbalance */
2133 if (env->dst_stats.idle_cpu >= 0) {
2134 env->dst_cpu = env->dst_stats.idle_cpu;
2135 task_numa_assign(env, NULL, 0);
2140 long src_load, dst_load, load;
2142 * If the improvement from just moving env->p direction is better
2143 * than swapping tasks around, check if a move is possible.
2145 load = task_h_load(env->p);
2146 dst_load = env->dst_stats.load + load;
2147 src_load = env->src_stats.load - load;
2148 maymove = !load_too_imbalanced(src_load, dst_load, env);
2151 for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) {
2152 /* Skip this CPU if the source task cannot migrate */
2153 if (!cpumask_test_cpu(cpu, env->p->cpus_ptr))
2157 if (task_numa_compare(env, taskimp, groupimp, maymove))
2162 static int task_numa_migrate(struct task_struct *p)
2164 struct task_numa_env env = {
2167 .src_cpu = task_cpu(p),
2168 .src_nid = task_node(p),
2170 .imbalance_pct = 112,
2176 unsigned long taskweight, groupweight;
2177 struct sched_domain *sd;
2178 long taskimp, groupimp;
2179 struct numa_group *ng;
2184 * Pick the lowest SD_NUMA domain, as that would have the smallest
2185 * imbalance and would be the first to start moving tasks about.
2187 * And we want to avoid any moving of tasks about, as that would create
2188 * random movement of tasks -- counter the numa conditions we're trying
2192 sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu));
2194 env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2;
2195 env.imb_numa_nr = sd->imb_numa_nr;
2200 * Cpusets can break the scheduler domain tree into smaller
2201 * balance domains, some of which do not cross NUMA boundaries.
2202 * Tasks that are "trapped" in such domains cannot be migrated
2203 * elsewhere, so there is no point in (re)trying.
2205 if (unlikely(!sd)) {
2206 sched_setnuma(p, task_node(p));
2210 env.dst_nid = p->numa_preferred_nid;
2211 dist = env.dist = node_distance(env.src_nid, env.dst_nid);
2212 taskweight = task_weight(p, env.src_nid, dist);
2213 groupweight = group_weight(p, env.src_nid, dist);
2214 update_numa_stats(&env, &env.src_stats, env.src_nid, false);
2215 taskimp = task_weight(p, env.dst_nid, dist) - taskweight;
2216 groupimp = group_weight(p, env.dst_nid, dist) - groupweight;
2217 update_numa_stats(&env, &env.dst_stats, env.dst_nid, true);
2219 /* Try to find a spot on the preferred nid. */
2220 task_numa_find_cpu(&env, taskimp, groupimp);
2223 * Look at other nodes in these cases:
2224 * - there is no space available on the preferred_nid
2225 * - the task is part of a numa_group that is interleaved across
2226 * multiple NUMA nodes; in order to better consolidate the group,
2227 * we need to check other locations.
2229 ng = deref_curr_numa_group(p);
2230 if (env.best_cpu == -1 || (ng && ng->active_nodes > 1)) {
2231 for_each_node_state(nid, N_CPU) {
2232 if (nid == env.src_nid || nid == p->numa_preferred_nid)
2235 dist = node_distance(env.src_nid, env.dst_nid);
2236 if (sched_numa_topology_type == NUMA_BACKPLANE &&
2238 taskweight = task_weight(p, env.src_nid, dist);
2239 groupweight = group_weight(p, env.src_nid, dist);
2242 /* Only consider nodes where both task and groups benefit */
2243 taskimp = task_weight(p, nid, dist) - taskweight;
2244 groupimp = group_weight(p, nid, dist) - groupweight;
2245 if (taskimp < 0 && groupimp < 0)
2250 update_numa_stats(&env, &env.dst_stats, env.dst_nid, true);
2251 task_numa_find_cpu(&env, taskimp, groupimp);
2256 * If the task is part of a workload that spans multiple NUMA nodes,
2257 * and is migrating into one of the workload's active nodes, remember
2258 * this node as the task's preferred numa node, so the workload can
2260 * A task that migrated to a second choice node will be better off
2261 * trying for a better one later. Do not set the preferred node here.
2264 if (env.best_cpu == -1)
2267 nid = cpu_to_node(env.best_cpu);
2269 if (nid != p->numa_preferred_nid)
2270 sched_setnuma(p, nid);
2273 /* No better CPU than the current one was found. */
2274 if (env.best_cpu == -1) {
2275 trace_sched_stick_numa(p, env.src_cpu, NULL, -1);
2279 best_rq = cpu_rq(env.best_cpu);
2280 if (env.best_task == NULL) {
2281 ret = migrate_task_to(p, env.best_cpu);
2282 WRITE_ONCE(best_rq->numa_migrate_on, 0);
2284 trace_sched_stick_numa(p, env.src_cpu, NULL, env.best_cpu);
2288 ret = migrate_swap(p, env.best_task, env.best_cpu, env.src_cpu);
2289 WRITE_ONCE(best_rq->numa_migrate_on, 0);
2292 trace_sched_stick_numa(p, env.src_cpu, env.best_task, env.best_cpu);
2293 put_task_struct(env.best_task);
2297 /* Attempt to migrate a task to a CPU on the preferred node. */
2298 static void numa_migrate_preferred(struct task_struct *p)
2300 unsigned long interval = HZ;
2302 /* This task has no NUMA fault statistics yet */
2303 if (unlikely(p->numa_preferred_nid == NUMA_NO_NODE || !p->numa_faults))
2306 /* Periodically retry migrating the task to the preferred node */
2307 interval = min(interval, msecs_to_jiffies(p->numa_scan_period) / 16);
2308 p->numa_migrate_retry = jiffies + interval;
2310 /* Success if task is already running on preferred CPU */
2311 if (task_node(p) == p->numa_preferred_nid)
2314 /* Otherwise, try migrate to a CPU on the preferred node */
2315 task_numa_migrate(p);
2319 * Find out how many nodes the workload is actively running on. Do this by
2320 * tracking the nodes from which NUMA hinting faults are triggered. This can
2321 * be different from the set of nodes where the workload's memory is currently
2324 static void numa_group_count_active_nodes(struct numa_group *numa_group)
2326 unsigned long faults, max_faults = 0;
2327 int nid, active_nodes = 0;
2329 for_each_node_state(nid, N_CPU) {
2330 faults = group_faults_cpu(numa_group, nid);
2331 if (faults > max_faults)
2332 max_faults = faults;
2335 for_each_node_state(nid, N_CPU) {
2336 faults = group_faults_cpu(numa_group, nid);
2337 if (faults * ACTIVE_NODE_FRACTION > max_faults)
2341 numa_group->max_faults_cpu = max_faults;
2342 numa_group->active_nodes = active_nodes;
2346 * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
2347 * increments. The more local the fault statistics are, the higher the scan
2348 * period will be for the next scan window. If local/(local+remote) ratio is
2349 * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS)
2350 * the scan period will decrease. Aim for 70% local accesses.
2352 #define NUMA_PERIOD_SLOTS 10
2353 #define NUMA_PERIOD_THRESHOLD 7
2356 * Increase the scan period (slow down scanning) if the majority of
2357 * our memory is already on our local node, or if the majority of
2358 * the page accesses are shared with other processes.
2359 * Otherwise, decrease the scan period.
2361 static void update_task_scan_period(struct task_struct *p,
2362 unsigned long shared, unsigned long private)
2364 unsigned int period_slot;
2365 int lr_ratio, ps_ratio;
2368 unsigned long remote = p->numa_faults_locality[0];
2369 unsigned long local = p->numa_faults_locality[1];
2372 * If there were no record hinting faults then either the task is
2373 * completely idle or all activity is in areas that are not of interest
2374 * to automatic numa balancing. Related to that, if there were failed
2375 * migration then it implies we are migrating too quickly or the local
2376 * node is overloaded. In either case, scan slower
2378 if (local + shared == 0 || p->numa_faults_locality[2]) {
2379 p->numa_scan_period = min(p->numa_scan_period_max,
2380 p->numa_scan_period << 1);
2382 p->mm->numa_next_scan = jiffies +
2383 msecs_to_jiffies(p->numa_scan_period);
2389 * Prepare to scale scan period relative to the current period.
2390 * == NUMA_PERIOD_THRESHOLD scan period stays the same
2391 * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
2392 * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
2394 period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS);
2395 lr_ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote);
2396 ps_ratio = (private * NUMA_PERIOD_SLOTS) / (private + shared);
2398 if (ps_ratio >= NUMA_PERIOD_THRESHOLD) {
2400 * Most memory accesses are local. There is no need to
2401 * do fast NUMA scanning, since memory is already local.
2403 int slot = ps_ratio - NUMA_PERIOD_THRESHOLD;
2406 diff = slot * period_slot;
2407 } else if (lr_ratio >= NUMA_PERIOD_THRESHOLD) {
2409 * Most memory accesses are shared with other tasks.
2410 * There is no point in continuing fast NUMA scanning,
2411 * since other tasks may just move the memory elsewhere.
2413 int slot = lr_ratio - NUMA_PERIOD_THRESHOLD;
2416 diff = slot * period_slot;
2419 * Private memory faults exceed (SLOTS-THRESHOLD)/SLOTS,
2420 * yet they are not on the local NUMA node. Speed up
2421 * NUMA scanning to get the memory moved over.
2423 int ratio = max(lr_ratio, ps_ratio);
2424 diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot;
2427 p->numa_scan_period = clamp(p->numa_scan_period + diff,
2428 task_scan_min(p), task_scan_max(p));
2429 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
2433 * Get the fraction of time the task has been running since the last
2434 * NUMA placement cycle. The scheduler keeps similar statistics, but
2435 * decays those on a 32ms period, which is orders of magnitude off
2436 * from the dozens-of-seconds NUMA balancing period. Use the scheduler
2437 * stats only if the task is so new there are no NUMA statistics yet.
2439 static u64 numa_get_avg_runtime(struct task_struct *p, u64 *period)
2441 u64 runtime, delta, now;
2442 /* Use the start of this time slice to avoid calculations. */
2443 now = p->se.exec_start;
2444 runtime = p->se.sum_exec_runtime;
2446 if (p->last_task_numa_placement) {
2447 delta = runtime - p->last_sum_exec_runtime;
2448 *period = now - p->last_task_numa_placement;
2450 /* Avoid time going backwards, prevent potential divide error: */
2451 if (unlikely((s64)*period < 0))
2454 delta = p->se.avg.load_sum;
2455 *period = LOAD_AVG_MAX;
2458 p->last_sum_exec_runtime = runtime;
2459 p->last_task_numa_placement = now;
2465 * Determine the preferred nid for a task in a numa_group. This needs to
2466 * be done in a way that produces consistent results with group_weight,
2467 * otherwise workloads might not converge.
2469 static int preferred_group_nid(struct task_struct *p, int nid)
2474 /* Direct connections between all NUMA nodes. */
2475 if (sched_numa_topology_type == NUMA_DIRECT)
2479 * On a system with glueless mesh NUMA topology, group_weight
2480 * scores nodes according to the number of NUMA hinting faults on
2481 * both the node itself, and on nearby nodes.
2483 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
2484 unsigned long score, max_score = 0;
2485 int node, max_node = nid;
2487 dist = sched_max_numa_distance;
2489 for_each_node_state(node, N_CPU) {
2490 score = group_weight(p, node, dist);
2491 if (score > max_score) {
2500 * Finding the preferred nid in a system with NUMA backplane
2501 * interconnect topology is more involved. The goal is to locate
2502 * tasks from numa_groups near each other in the system, and
2503 * untangle workloads from different sides of the system. This requires
2504 * searching down the hierarchy of node groups, recursively searching
2505 * inside the highest scoring group of nodes. The nodemask tricks
2506 * keep the complexity of the search down.
2508 nodes = node_states[N_CPU];
2509 for (dist = sched_max_numa_distance; dist > LOCAL_DISTANCE; dist--) {
2510 unsigned long max_faults = 0;
2511 nodemask_t max_group = NODE_MASK_NONE;
2514 /* Are there nodes at this distance from each other? */
2515 if (!find_numa_distance(dist))
2518 for_each_node_mask(a, nodes) {
2519 unsigned long faults = 0;
2520 nodemask_t this_group;
2521 nodes_clear(this_group);
2523 /* Sum group's NUMA faults; includes a==b case. */
2524 for_each_node_mask(b, nodes) {
2525 if (node_distance(a, b) < dist) {
2526 faults += group_faults(p, b);
2527 node_set(b, this_group);
2528 node_clear(b, nodes);
2532 /* Remember the top group. */
2533 if (faults > max_faults) {
2534 max_faults = faults;
2535 max_group = this_group;
2537 * subtle: at the smallest distance there is
2538 * just one node left in each "group", the
2539 * winner is the preferred nid.
2544 /* Next round, evaluate the nodes within max_group. */
2552 static void task_numa_placement(struct task_struct *p)
2554 int seq, nid, max_nid = NUMA_NO_NODE;
2555 unsigned long max_faults = 0;
2556 unsigned long fault_types[2] = { 0, 0 };
2557 unsigned long total_faults;
2558 u64 runtime, period;
2559 spinlock_t *group_lock = NULL;
2560 struct numa_group *ng;
2563 * The p->mm->numa_scan_seq field gets updated without
2564 * exclusive access. Use READ_ONCE() here to ensure
2565 * that the field is read in a single access:
2567 seq = READ_ONCE(p->mm->numa_scan_seq);
2568 if (p->numa_scan_seq == seq)
2570 p->numa_scan_seq = seq;
2571 p->numa_scan_period_max = task_scan_max(p);
2573 total_faults = p->numa_faults_locality[0] +
2574 p->numa_faults_locality[1];
2575 runtime = numa_get_avg_runtime(p, &period);
2577 /* If the task is part of a group prevent parallel updates to group stats */
2578 ng = deref_curr_numa_group(p);
2580 group_lock = &ng->lock;
2581 spin_lock_irq(group_lock);
2584 /* Find the node with the highest number of faults */
2585 for_each_online_node(nid) {
2586 /* Keep track of the offsets in numa_faults array */
2587 int mem_idx, membuf_idx, cpu_idx, cpubuf_idx;
2588 unsigned long faults = 0, group_faults = 0;
2591 for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) {
2592 long diff, f_diff, f_weight;
2594 mem_idx = task_faults_idx(NUMA_MEM, nid, priv);
2595 membuf_idx = task_faults_idx(NUMA_MEMBUF, nid, priv);
2596 cpu_idx = task_faults_idx(NUMA_CPU, nid, priv);
2597 cpubuf_idx = task_faults_idx(NUMA_CPUBUF, nid, priv);
2599 /* Decay existing window, copy faults since last scan */
2600 diff = p->numa_faults[membuf_idx] - p->numa_faults[mem_idx] / 2;
2601 fault_types[priv] += p->numa_faults[membuf_idx];
2602 p->numa_faults[membuf_idx] = 0;
2605 * Normalize the faults_from, so all tasks in a group
2606 * count according to CPU use, instead of by the raw
2607 * number of faults. Tasks with little runtime have
2608 * little over-all impact on throughput, and thus their
2609 * faults are less important.
2611 f_weight = div64_u64(runtime << 16, period + 1);
2612 f_weight = (f_weight * p->numa_faults[cpubuf_idx]) /
2614 f_diff = f_weight - p->numa_faults[cpu_idx] / 2;
2615 p->numa_faults[cpubuf_idx] = 0;
2617 p->numa_faults[mem_idx] += diff;
2618 p->numa_faults[cpu_idx] += f_diff;
2619 faults += p->numa_faults[mem_idx];
2620 p->total_numa_faults += diff;
2623 * safe because we can only change our own group
2625 * mem_idx represents the offset for a given
2626 * nid and priv in a specific region because it
2627 * is at the beginning of the numa_faults array.
2629 ng->faults[mem_idx] += diff;
2630 ng->faults[cpu_idx] += f_diff;
2631 ng->total_faults += diff;
2632 group_faults += ng->faults[mem_idx];
2637 if (faults > max_faults) {
2638 max_faults = faults;
2641 } else if (group_faults > max_faults) {
2642 max_faults = group_faults;
2647 /* Cannot migrate task to CPU-less node */
2648 if (max_nid != NUMA_NO_NODE && !node_state(max_nid, N_CPU)) {
2649 int near_nid = max_nid;
2650 int distance, near_distance = INT_MAX;
2652 for_each_node_state(nid, N_CPU) {
2653 distance = node_distance(max_nid, nid);
2654 if (distance < near_distance) {
2656 near_distance = distance;
2663 numa_group_count_active_nodes(ng);
2664 spin_unlock_irq(group_lock);
2665 max_nid = preferred_group_nid(p, max_nid);
2669 /* Set the new preferred node */
2670 if (max_nid != p->numa_preferred_nid)
2671 sched_setnuma(p, max_nid);
2674 update_task_scan_period(p, fault_types[0], fault_types[1]);
2677 static inline int get_numa_group(struct numa_group *grp)
2679 return refcount_inc_not_zero(&grp->refcount);
2682 static inline void put_numa_group(struct numa_group *grp)
2684 if (refcount_dec_and_test(&grp->refcount))
2685 kfree_rcu(grp, rcu);
2688 static void task_numa_group(struct task_struct *p, int cpupid, int flags,
2691 struct numa_group *grp, *my_grp;
2692 struct task_struct *tsk;
2694 int cpu = cpupid_to_cpu(cpupid);
2697 if (unlikely(!deref_curr_numa_group(p))) {
2698 unsigned int size = sizeof(struct numa_group) +
2699 NR_NUMA_HINT_FAULT_STATS *
2700 nr_node_ids * sizeof(unsigned long);
2702 grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
2706 refcount_set(&grp->refcount, 1);
2707 grp->active_nodes = 1;
2708 grp->max_faults_cpu = 0;
2709 spin_lock_init(&grp->lock);
2712 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2713 grp->faults[i] = p->numa_faults[i];
2715 grp->total_faults = p->total_numa_faults;
2718 rcu_assign_pointer(p->numa_group, grp);
2722 tsk = READ_ONCE(cpu_rq(cpu)->curr);
2724 if (!cpupid_match_pid(tsk, cpupid))
2727 grp = rcu_dereference(tsk->numa_group);
2731 my_grp = deref_curr_numa_group(p);
2736 * Only join the other group if its bigger; if we're the bigger group,
2737 * the other task will join us.
2739 if (my_grp->nr_tasks > grp->nr_tasks)
2743 * Tie-break on the grp address.
2745 if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp)
2748 /* Always join threads in the same process. */
2749 if (tsk->mm == current->mm)
2752 /* Simple filter to avoid false positives due to PID collisions */
2753 if (flags & TNF_SHARED)
2756 /* Update priv based on whether false sharing was detected */
2759 if (join && !get_numa_group(grp))
2767 WARN_ON_ONCE(irqs_disabled());
2768 double_lock_irq(&my_grp->lock, &grp->lock);
2770 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) {
2771 my_grp->faults[i] -= p->numa_faults[i];
2772 grp->faults[i] += p->numa_faults[i];
2774 my_grp->total_faults -= p->total_numa_faults;
2775 grp->total_faults += p->total_numa_faults;
2780 spin_unlock(&my_grp->lock);
2781 spin_unlock_irq(&grp->lock);
2783 rcu_assign_pointer(p->numa_group, grp);
2785 put_numa_group(my_grp);
2794 * Get rid of NUMA statistics associated with a task (either current or dead).
2795 * If @final is set, the task is dead and has reached refcount zero, so we can
2796 * safely free all relevant data structures. Otherwise, there might be
2797 * concurrent reads from places like load balancing and procfs, and we should
2798 * reset the data back to default state without freeing ->numa_faults.
2800 void task_numa_free(struct task_struct *p, bool final)
2802 /* safe: p either is current or is being freed by current */
2803 struct numa_group *grp = rcu_dereference_raw(p->numa_group);
2804 unsigned long *numa_faults = p->numa_faults;
2805 unsigned long flags;
2812 spin_lock_irqsave(&grp->lock, flags);
2813 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2814 grp->faults[i] -= p->numa_faults[i];
2815 grp->total_faults -= p->total_numa_faults;
2818 spin_unlock_irqrestore(&grp->lock, flags);
2819 RCU_INIT_POINTER(p->numa_group, NULL);
2820 put_numa_group(grp);
2824 p->numa_faults = NULL;
2827 p->total_numa_faults = 0;
2828 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2834 * Got a PROT_NONE fault for a page on @node.
2836 void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags)
2838 struct task_struct *p = current;
2839 bool migrated = flags & TNF_MIGRATED;
2840 int cpu_node = task_node(current);
2841 int local = !!(flags & TNF_FAULT_LOCAL);
2842 struct numa_group *ng;
2845 if (!static_branch_likely(&sched_numa_balancing))
2848 /* for example, ksmd faulting in a user's mm */
2853 * NUMA faults statistics are unnecessary for the slow memory
2854 * node for memory tiering mode.
2856 if (!node_is_toptier(mem_node) &&
2857 (sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING ||
2858 !cpupid_valid(last_cpupid)))
2861 /* Allocate buffer to track faults on a per-node basis */
2862 if (unlikely(!p->numa_faults)) {
2863 int size = sizeof(*p->numa_faults) *
2864 NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids;
2866 p->numa_faults = kzalloc(size, GFP_KERNEL|__GFP_NOWARN);
2867 if (!p->numa_faults)
2870 p->total_numa_faults = 0;
2871 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
2875 * First accesses are treated as private, otherwise consider accesses
2876 * to be private if the accessing pid has not changed
2878 if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) {
2881 priv = cpupid_match_pid(p, last_cpupid);
2882 if (!priv && !(flags & TNF_NO_GROUP))
2883 task_numa_group(p, last_cpupid, flags, &priv);
2887 * If a workload spans multiple NUMA nodes, a shared fault that
2888 * occurs wholly within the set of nodes that the workload is
2889 * actively using should be counted as local. This allows the
2890 * scan rate to slow down when a workload has settled down.
2892 ng = deref_curr_numa_group(p);
2893 if (!priv && !local && ng && ng->active_nodes > 1 &&
2894 numa_is_active_node(cpu_node, ng) &&
2895 numa_is_active_node(mem_node, ng))
2899 * Retry to migrate task to preferred node periodically, in case it
2900 * previously failed, or the scheduler moved us.
2902 if (time_after(jiffies, p->numa_migrate_retry)) {
2903 task_numa_placement(p);
2904 numa_migrate_preferred(p);
2908 p->numa_pages_migrated += pages;
2909 if (flags & TNF_MIGRATE_FAIL)
2910 p->numa_faults_locality[2] += pages;
2912 p->numa_faults[task_faults_idx(NUMA_MEMBUF, mem_node, priv)] += pages;
2913 p->numa_faults[task_faults_idx(NUMA_CPUBUF, cpu_node, priv)] += pages;
2914 p->numa_faults_locality[local] += pages;
2917 static void reset_ptenuma_scan(struct task_struct *p)
2920 * We only did a read acquisition of the mmap sem, so
2921 * p->mm->numa_scan_seq is written to without exclusive access
2922 * and the update is not guaranteed to be atomic. That's not
2923 * much of an issue though, since this is just used for
2924 * statistical sampling. Use READ_ONCE/WRITE_ONCE, which are not
2925 * expensive, to avoid any form of compiler optimizations:
2927 WRITE_ONCE(p->mm->numa_scan_seq, READ_ONCE(p->mm->numa_scan_seq) + 1);
2928 p->mm->numa_scan_offset = 0;
2932 * The expensive part of numa migration is done from task_work context.
2933 * Triggered from task_tick_numa().
2935 static void task_numa_work(struct callback_head *work)
2937 unsigned long migrate, next_scan, now = jiffies;
2938 struct task_struct *p = current;
2939 struct mm_struct *mm = p->mm;
2940 u64 runtime = p->se.sum_exec_runtime;
2941 MA_STATE(mas, &mm->mm_mt, 0, 0);
2942 struct vm_area_struct *vma;
2943 unsigned long start, end;
2944 unsigned long nr_pte_updates = 0;
2945 long pages, virtpages;
2947 SCHED_WARN_ON(p != container_of(work, struct task_struct, numa_work));
2951 * Who cares about NUMA placement when they're dying.
2953 * NOTE: make sure not to dereference p->mm before this check,
2954 * exit_task_work() happens _after_ exit_mm() so we could be called
2955 * without p->mm even though we still had it when we enqueued this
2958 if (p->flags & PF_EXITING)
2961 if (!mm->numa_next_scan) {
2962 mm->numa_next_scan = now +
2963 msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2967 * Enforce maximal scan/migration frequency..
2969 migrate = mm->numa_next_scan;
2970 if (time_before(now, migrate))
2973 if (p->numa_scan_period == 0) {
2974 p->numa_scan_period_max = task_scan_max(p);
2975 p->numa_scan_period = task_scan_start(p);
2978 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
2979 if (!try_cmpxchg(&mm->numa_next_scan, &migrate, next_scan))
2983 * Delay this task enough that another task of this mm will likely win
2984 * the next time around.
2986 p->node_stamp += 2 * TICK_NSEC;
2988 start = mm->numa_scan_offset;
2989 pages = sysctl_numa_balancing_scan_size;
2990 pages <<= 20 - PAGE_SHIFT; /* MB in pages */
2991 virtpages = pages * 8; /* Scan up to this much virtual space */
2996 if (!mmap_read_trylock(mm))
2998 mas_set(&mas, start);
2999 vma = mas_find(&mas, ULONG_MAX);
3001 reset_ptenuma_scan(p);
3003 mas_set(&mas, start);
3004 vma = mas_find(&mas, ULONG_MAX);
3007 for (; vma; vma = mas_find(&mas, ULONG_MAX)) {
3008 if (!vma_migratable(vma) || !vma_policy_mof(vma) ||
3009 is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_MIXEDMAP)) {
3014 * Shared library pages mapped by multiple processes are not
3015 * migrated as it is expected they are cache replicated. Avoid
3016 * hinting faults in read-only file-backed mappings or the vdso
3017 * as migrating the pages will be of marginal benefit.
3020 (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ)))
3024 * Skip inaccessible VMAs to avoid any confusion between
3025 * PROT_NONE and NUMA hinting ptes
3027 if (!vma_is_accessible(vma))
3031 start = max(start, vma->vm_start);
3032 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
3033 end = min(end, vma->vm_end);
3034 nr_pte_updates = change_prot_numa(vma, start, end);
3037 * Try to scan sysctl_numa_balancing_size worth of
3038 * hpages that have at least one present PTE that
3039 * is not already pte-numa. If the VMA contains
3040 * areas that are unused or already full of prot_numa
3041 * PTEs, scan up to virtpages, to skip through those
3045 pages -= (end - start) >> PAGE_SHIFT;
3046 virtpages -= (end - start) >> PAGE_SHIFT;
3049 if (pages <= 0 || virtpages <= 0)
3053 } while (end != vma->vm_end);
3058 * It is possible to reach the end of the VMA list but the last few
3059 * VMAs are not guaranteed to the vma_migratable. If they are not, we
3060 * would find the !migratable VMA on the next scan but not reset the
3061 * scanner to the start so check it now.
3064 mm->numa_scan_offset = start;
3066 reset_ptenuma_scan(p);
3067 mmap_read_unlock(mm);
3070 * Make sure tasks use at least 32x as much time to run other code
3071 * than they used here, to limit NUMA PTE scanning overhead to 3% max.
3072 * Usually update_task_scan_period slows down scanning enough; on an
3073 * overloaded system we need to limit overhead on a per task basis.
3075 if (unlikely(p->se.sum_exec_runtime != runtime)) {
3076 u64 diff = p->se.sum_exec_runtime - runtime;
3077 p->node_stamp += 32 * diff;
3081 void init_numa_balancing(unsigned long clone_flags, struct task_struct *p)
3084 struct mm_struct *mm = p->mm;
3087 mm_users = atomic_read(&mm->mm_users);
3088 if (mm_users == 1) {
3089 mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
3090 mm->numa_scan_seq = 0;
3094 p->numa_scan_seq = mm ? mm->numa_scan_seq : 0;
3095 p->numa_scan_period = sysctl_numa_balancing_scan_delay;
3096 p->numa_migrate_retry = 0;
3097 /* Protect against double add, see task_tick_numa and task_numa_work */
3098 p->numa_work.next = &p->numa_work;
3099 p->numa_faults = NULL;
3100 p->numa_pages_migrated = 0;
3101 p->total_numa_faults = 0;
3102 RCU_INIT_POINTER(p->numa_group, NULL);
3103 p->last_task_numa_placement = 0;
3104 p->last_sum_exec_runtime = 0;
3106 init_task_work(&p->numa_work, task_numa_work);
3108 /* New address space, reset the preferred nid */
3109 if (!(clone_flags & CLONE_VM)) {
3110 p->numa_preferred_nid = NUMA_NO_NODE;
3115 * New thread, keep existing numa_preferred_nid which should be copied
3116 * already by arch_dup_task_struct but stagger when scans start.
3121 delay = min_t(unsigned int, task_scan_max(current),
3122 current->numa_scan_period * mm_users * NSEC_PER_MSEC);
3123 delay += 2 * TICK_NSEC;
3124 p->node_stamp = delay;
3129 * Drive the periodic memory faults..
3131 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
3133 struct callback_head *work = &curr->numa_work;
3137 * We don't care about NUMA placement if we don't have memory.
3139 if (!curr->mm || (curr->flags & (PF_EXITING | PF_KTHREAD)) || work->next != work)
3143 * Using runtime rather than walltime has the dual advantage that
3144 * we (mostly) drive the selection from busy threads and that the
3145 * task needs to have done some actual work before we bother with
3148 now = curr->se.sum_exec_runtime;
3149 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
3151 if (now > curr->node_stamp + period) {
3152 if (!curr->node_stamp)
3153 curr->numa_scan_period = task_scan_start(curr);
3154 curr->node_stamp += period;
3156 if (!time_before(jiffies, curr->mm->numa_next_scan))
3157 task_work_add(curr, work, TWA_RESUME);
3161 static void update_scan_period(struct task_struct *p, int new_cpu)
3163 int src_nid = cpu_to_node(task_cpu(p));
3164 int dst_nid = cpu_to_node(new_cpu);
3166 if (!static_branch_likely(&sched_numa_balancing))
3169 if (!p->mm || !p->numa_faults || (p->flags & PF_EXITING))
3172 if (src_nid == dst_nid)
3176 * Allow resets if faults have been trapped before one scan
3177 * has completed. This is most likely due to a new task that
3178 * is pulled cross-node due to wakeups or load balancing.
3180 if (p->numa_scan_seq) {
3182 * Avoid scan adjustments if moving to the preferred
3183 * node or if the task was not previously running on
3184 * the preferred node.
3186 if (dst_nid == p->numa_preferred_nid ||
3187 (p->numa_preferred_nid != NUMA_NO_NODE &&
3188 src_nid != p->numa_preferred_nid))
3192 p->numa_scan_period = task_scan_start(p);
3196 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
3200 static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p)
3204 static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p)
3208 static inline void update_scan_period(struct task_struct *p, int new_cpu)
3212 #endif /* CONFIG_NUMA_BALANCING */
3215 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
3217 update_load_add(&cfs_rq->load, se->load.weight);
3219 if (entity_is_task(se)) {
3220 struct rq *rq = rq_of(cfs_rq);
3222 account_numa_enqueue(rq, task_of(se));
3223 list_add(&se->group_node, &rq->cfs_tasks);
3226 cfs_rq->nr_running++;
3228 cfs_rq->idle_nr_running++;
3232 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
3234 update_load_sub(&cfs_rq->load, se->load.weight);
3236 if (entity_is_task(se)) {
3237 account_numa_dequeue(rq_of(cfs_rq), task_of(se));
3238 list_del_init(&se->group_node);
3241 cfs_rq->nr_running--;
3243 cfs_rq->idle_nr_running--;
3247 * Signed add and clamp on underflow.
3249 * Explicitly do a load-store to ensure the intermediate value never hits
3250 * memory. This allows lockless observations without ever seeing the negative
3253 #define add_positive(_ptr, _val) do { \
3254 typeof(_ptr) ptr = (_ptr); \
3255 typeof(_val) val = (_val); \
3256 typeof(*ptr) res, var = READ_ONCE(*ptr); \
3260 if (val < 0 && res > var) \
3263 WRITE_ONCE(*ptr, res); \
3267 * Unsigned subtract and clamp on underflow.
3269 * Explicitly do a load-store to ensure the intermediate value never hits
3270 * memory. This allows lockless observations without ever seeing the negative
3273 #define sub_positive(_ptr, _val) do { \
3274 typeof(_ptr) ptr = (_ptr); \
3275 typeof(*ptr) val = (_val); \
3276 typeof(*ptr) res, var = READ_ONCE(*ptr); \
3280 WRITE_ONCE(*ptr, res); \
3284 * Remove and clamp on negative, from a local variable.
3286 * A variant of sub_positive(), which does not use explicit load-store
3287 * and is thus optimized for local variable updates.
3289 #define lsub_positive(_ptr, _val) do { \
3290 typeof(_ptr) ptr = (_ptr); \
3291 *ptr -= min_t(typeof(*ptr), *ptr, _val); \
3296 enqueue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3298 cfs_rq->avg.load_avg += se->avg.load_avg;
3299 cfs_rq->avg.load_sum += se_weight(se) * se->avg.load_sum;
3303 dequeue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3305 sub_positive(&cfs_rq->avg.load_avg, se->avg.load_avg);
3306 sub_positive(&cfs_rq->avg.load_sum, se_weight(se) * se->avg.load_sum);
3307 /* See update_cfs_rq_load_avg() */
3308 cfs_rq->avg.load_sum = max_t(u32, cfs_rq->avg.load_sum,
3309 cfs_rq->avg.load_avg * PELT_MIN_DIVIDER);
3313 enqueue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
3315 dequeue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
3318 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
3319 unsigned long weight)
3322 /* commit outstanding execution time */
3323 if (cfs_rq->curr == se)
3324 update_curr(cfs_rq);
3325 update_load_sub(&cfs_rq->load, se->load.weight);
3327 dequeue_load_avg(cfs_rq, se);
3329 update_load_set(&se->load, weight);
3333 u32 divider = get_pelt_divider(&se->avg);
3335 se->avg.load_avg = div_u64(se_weight(se) * se->avg.load_sum, divider);
3339 enqueue_load_avg(cfs_rq, se);
3341 update_load_add(&cfs_rq->load, se->load.weight);
3345 void reweight_task(struct task_struct *p, int prio)
3347 struct sched_entity *se = &p->se;
3348 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3349 struct load_weight *load = &se->load;
3350 unsigned long weight = scale_load(sched_prio_to_weight[prio]);
3352 reweight_entity(cfs_rq, se, weight);
3353 load->inv_weight = sched_prio_to_wmult[prio];
3356 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
3358 #ifdef CONFIG_FAIR_GROUP_SCHED
3361 * All this does is approximate the hierarchical proportion which includes that
3362 * global sum we all love to hate.
3364 * That is, the weight of a group entity, is the proportional share of the
3365 * group weight based on the group runqueue weights. That is:
3367 * tg->weight * grq->load.weight
3368 * ge->load.weight = ----------------------------- (1)
3369 * \Sum grq->load.weight
3371 * Now, because computing that sum is prohibitively expensive to compute (been
3372 * there, done that) we approximate it with this average stuff. The average
3373 * moves slower and therefore the approximation is cheaper and more stable.
3375 * So instead of the above, we substitute:
3377 * grq->load.weight -> grq->avg.load_avg (2)
3379 * which yields the following:
3381 * tg->weight * grq->avg.load_avg
3382 * ge->load.weight = ------------------------------ (3)
3385 * Where: tg->load_avg ~= \Sum grq->avg.load_avg
3387 * That is shares_avg, and it is right (given the approximation (2)).
3389 * The problem with it is that because the average is slow -- it was designed
3390 * to be exactly that of course -- this leads to transients in boundary
3391 * conditions. In specific, the case where the group was idle and we start the
3392 * one task. It takes time for our CPU's grq->avg.load_avg to build up,
3393 * yielding bad latency etc..
3395 * Now, in that special case (1) reduces to:
3397 * tg->weight * grq->load.weight
3398 * ge->load.weight = ----------------------------- = tg->weight (4)
3401 * That is, the sum collapses because all other CPUs are idle; the UP scenario.
3403 * So what we do is modify our approximation (3) to approach (4) in the (near)
3408 * tg->weight * grq->load.weight
3409 * --------------------------------------------------- (5)
3410 * tg->load_avg - grq->avg.load_avg + grq->load.weight
3412 * But because grq->load.weight can drop to 0, resulting in a divide by zero,
3413 * we need to use grq->avg.load_avg as its lower bound, which then gives:
3416 * tg->weight * grq->load.weight
3417 * ge->load.weight = ----------------------------- (6)
3422 * tg_load_avg' = tg->load_avg - grq->avg.load_avg +
3423 * max(grq->load.weight, grq->avg.load_avg)
3425 * And that is shares_weight and is icky. In the (near) UP case it approaches
3426 * (4) while in the normal case it approaches (3). It consistently
3427 * overestimates the ge->load.weight and therefore:
3429 * \Sum ge->load.weight >= tg->weight
3433 static long calc_group_shares(struct cfs_rq *cfs_rq)
3435 long tg_weight, tg_shares, load, shares;
3436 struct task_group *tg = cfs_rq->tg;
3438 tg_shares = READ_ONCE(tg->shares);
3440 load = max(scale_load_down(cfs_rq->load.weight), cfs_rq->avg.load_avg);
3442 tg_weight = atomic_long_read(&tg->load_avg);
3444 /* Ensure tg_weight >= load */
3445 tg_weight -= cfs_rq->tg_load_avg_contrib;
3448 shares = (tg_shares * load);
3450 shares /= tg_weight;
3453 * MIN_SHARES has to be unscaled here to support per-CPU partitioning
3454 * of a group with small tg->shares value. It is a floor value which is
3455 * assigned as a minimum load.weight to the sched_entity representing
3456 * the group on a CPU.
3458 * E.g. on 64-bit for a group with tg->shares of scale_load(15)=15*1024
3459 * on an 8-core system with 8 tasks each runnable on one CPU shares has
3460 * to be 15*1024*1/8=1920 instead of scale_load(MIN_SHARES)=2*1024. In
3461 * case no task is runnable on a CPU MIN_SHARES=2 should be returned
3464 return clamp_t(long, shares, MIN_SHARES, tg_shares);
3466 #endif /* CONFIG_SMP */
3469 * Recomputes the group entity based on the current state of its group
3472 static void update_cfs_group(struct sched_entity *se)
3474 struct cfs_rq *gcfs_rq = group_cfs_rq(se);
3480 if (throttled_hierarchy(gcfs_rq))
3484 shares = READ_ONCE(gcfs_rq->tg->shares);
3486 if (likely(se->load.weight == shares))
3489 shares = calc_group_shares(gcfs_rq);
3492 reweight_entity(cfs_rq_of(se), se, shares);
3495 #else /* CONFIG_FAIR_GROUP_SCHED */
3496 static inline void update_cfs_group(struct sched_entity *se)
3499 #endif /* CONFIG_FAIR_GROUP_SCHED */
3501 static inline void cfs_rq_util_change(struct cfs_rq *cfs_rq, int flags)
3503 struct rq *rq = rq_of(cfs_rq);
3505 if (&rq->cfs == cfs_rq) {
3507 * There are a few boundary cases this might miss but it should
3508 * get called often enough that that should (hopefully) not be
3511 * It will not get called when we go idle, because the idle
3512 * thread is a different class (!fair), nor will the utilization
3513 * number include things like RT tasks.
3515 * As is, the util number is not freq-invariant (we'd have to
3516 * implement arch_scale_freq_capacity() for that).
3518 * See cpu_util_cfs().
3520 cpufreq_update_util(rq, flags);
3525 static inline bool load_avg_is_decayed(struct sched_avg *sa)
3533 if (sa->runnable_sum)
3537 * _avg must be null when _sum are null because _avg = _sum / divider
3538 * Make sure that rounding and/or propagation of PELT values never
3541 SCHED_WARN_ON(sa->load_avg ||
3548 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
3550 return u64_u32_load_copy(cfs_rq->avg.last_update_time,
3551 cfs_rq->last_update_time_copy);
3553 #ifdef CONFIG_FAIR_GROUP_SCHED
3555 * Because list_add_leaf_cfs_rq always places a child cfs_rq on the list
3556 * immediately before a parent cfs_rq, and cfs_rqs are removed from the list
3557 * bottom-up, we only have to test whether the cfs_rq before us on the list
3559 * If cfs_rq is not on the list, test whether a child needs its to be added to
3560 * connect a branch to the tree * (see list_add_leaf_cfs_rq() for details).
3562 static inline bool child_cfs_rq_on_list(struct cfs_rq *cfs_rq)
3564 struct cfs_rq *prev_cfs_rq;
3565 struct list_head *prev;
3567 if (cfs_rq->on_list) {
3568 prev = cfs_rq->leaf_cfs_rq_list.prev;
3570 struct rq *rq = rq_of(cfs_rq);
3572 prev = rq->tmp_alone_branch;
3575 prev_cfs_rq = container_of(prev, struct cfs_rq, leaf_cfs_rq_list);
3577 return (prev_cfs_rq->tg->parent == cfs_rq->tg);
3580 static inline bool cfs_rq_is_decayed(struct cfs_rq *cfs_rq)
3582 if (cfs_rq->load.weight)
3585 if (!load_avg_is_decayed(&cfs_rq->avg))
3588 if (child_cfs_rq_on_list(cfs_rq))
3595 * update_tg_load_avg - update the tg's load avg
3596 * @cfs_rq: the cfs_rq whose avg changed
3598 * This function 'ensures': tg->load_avg := \Sum tg->cfs_rq[]->avg.load.
3599 * However, because tg->load_avg is a global value there are performance
3602 * In order to avoid having to look at the other cfs_rq's, we use a
3603 * differential update where we store the last value we propagated. This in
3604 * turn allows skipping updates if the differential is 'small'.
3606 * Updating tg's load_avg is necessary before update_cfs_share().
3608 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq)
3610 long delta = cfs_rq->avg.load_avg - cfs_rq->tg_load_avg_contrib;
3613 * No need to update load_avg for root_task_group as it is not used.
3615 if (cfs_rq->tg == &root_task_group)
3618 if (abs(delta) > cfs_rq->tg_load_avg_contrib / 64) {
3619 atomic_long_add(delta, &cfs_rq->tg->load_avg);
3620 cfs_rq->tg_load_avg_contrib = cfs_rq->avg.load_avg;
3625 * Called within set_task_rq() right before setting a task's CPU. The
3626 * caller only guarantees p->pi_lock is held; no other assumptions,
3627 * including the state of rq->lock, should be made.
3629 void set_task_rq_fair(struct sched_entity *se,
3630 struct cfs_rq *prev, struct cfs_rq *next)
3632 u64 p_last_update_time;
3633 u64 n_last_update_time;
3635 if (!sched_feat(ATTACH_AGE_LOAD))
3639 * We are supposed to update the task to "current" time, then its up to
3640 * date and ready to go to new CPU/cfs_rq. But we have difficulty in
3641 * getting what current time is, so simply throw away the out-of-date
3642 * time. This will result in the wakee task is less decayed, but giving
3643 * the wakee more load sounds not bad.
3645 if (!(se->avg.last_update_time && prev))
3648 p_last_update_time = cfs_rq_last_update_time(prev);
3649 n_last_update_time = cfs_rq_last_update_time(next);
3651 __update_load_avg_blocked_se(p_last_update_time, se);
3652 se->avg.last_update_time = n_last_update_time;
3656 * When on migration a sched_entity joins/leaves the PELT hierarchy, we need to
3657 * propagate its contribution. The key to this propagation is the invariant
3658 * that for each group:
3660 * ge->avg == grq->avg (1)
3662 * _IFF_ we look at the pure running and runnable sums. Because they
3663 * represent the very same entity, just at different points in the hierarchy.
3665 * Per the above update_tg_cfs_util() and update_tg_cfs_runnable() are trivial
3666 * and simply copies the running/runnable sum over (but still wrong, because
3667 * the group entity and group rq do not have their PELT windows aligned).
3669 * However, update_tg_cfs_load() is more complex. So we have:
3671 * ge->avg.load_avg = ge->load.weight * ge->avg.runnable_avg (2)
3673 * And since, like util, the runnable part should be directly transferable,
3674 * the following would _appear_ to be the straight forward approach:
3676 * grq->avg.load_avg = grq->load.weight * grq->avg.runnable_avg (3)
3678 * And per (1) we have:
3680 * ge->avg.runnable_avg == grq->avg.runnable_avg
3684 * ge->load.weight * grq->avg.load_avg
3685 * ge->avg.load_avg = ----------------------------------- (4)
3688 * Except that is wrong!
3690 * Because while for entities historical weight is not important and we
3691 * really only care about our future and therefore can consider a pure
3692 * runnable sum, runqueues can NOT do this.
3694 * We specifically want runqueues to have a load_avg that includes
3695 * historical weights. Those represent the blocked load, the load we expect
3696 * to (shortly) return to us. This only works by keeping the weights as
3697 * integral part of the sum. We therefore cannot decompose as per (3).
3699 * Another reason this doesn't work is that runnable isn't a 0-sum entity.
3700 * Imagine a rq with 2 tasks that each are runnable 2/3 of the time. Then the
3701 * rq itself is runnable anywhere between 2/3 and 1 depending on how the
3702 * runnable section of these tasks overlap (or not). If they were to perfectly
3703 * align the rq as a whole would be runnable 2/3 of the time. If however we
3704 * always have at least 1 runnable task, the rq as a whole is always runnable.
3706 * So we'll have to approximate.. :/
3708 * Given the constraint:
3710 * ge->avg.running_sum <= ge->avg.runnable_sum <= LOAD_AVG_MAX
3712 * We can construct a rule that adds runnable to a rq by assuming minimal
3715 * On removal, we'll assume each task is equally runnable; which yields:
3717 * grq->avg.runnable_sum = grq->avg.load_sum / grq->load.weight
3719 * XXX: only do this for the part of runnable > running ?
3723 update_tg_cfs_util(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq)
3725 long delta_sum, delta_avg = gcfs_rq->avg.util_avg - se->avg.util_avg;
3726 u32 new_sum, divider;
3728 /* Nothing to update */
3733 * cfs_rq->avg.period_contrib can be used for both cfs_rq and se.
3734 * See ___update_load_avg() for details.
3736 divider = get_pelt_divider(&cfs_rq->avg);
3739 /* Set new sched_entity's utilization */
3740 se->avg.util_avg = gcfs_rq->avg.util_avg;
3741 new_sum = se->avg.util_avg * divider;
3742 delta_sum = (long)new_sum - (long)se->avg.util_sum;
3743 se->avg.util_sum = new_sum;
3745 /* Update parent cfs_rq utilization */
3746 add_positive(&cfs_rq->avg.util_avg, delta_avg);
3747 add_positive(&cfs_rq->avg.util_sum, delta_sum);
3749 /* See update_cfs_rq_load_avg() */
3750 cfs_rq->avg.util_sum = max_t(u32, cfs_rq->avg.util_sum,
3751 cfs_rq->avg.util_avg * PELT_MIN_DIVIDER);
3755 update_tg_cfs_runnable(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq)
3757 long delta_sum, delta_avg = gcfs_rq->avg.runnable_avg - se->avg.runnable_avg;
3758 u32 new_sum, divider;
3760 /* Nothing to update */
3765 * cfs_rq->avg.period_contrib can be used for both cfs_rq and se.
3766 * See ___update_load_avg() for details.
3768 divider = get_pelt_divider(&cfs_rq->avg);
3770 /* Set new sched_entity's runnable */
3771 se->avg.runnable_avg = gcfs_rq->avg.runnable_avg;
3772 new_sum = se->avg.runnable_avg * divider;
3773 delta_sum = (long)new_sum - (long)se->avg.runnable_sum;
3774 se->avg.runnable_sum = new_sum;
3776 /* Update parent cfs_rq runnable */
3777 add_positive(&cfs_rq->avg.runnable_avg, delta_avg);
3778 add_positive(&cfs_rq->avg.runnable_sum, delta_sum);
3779 /* See update_cfs_rq_load_avg() */
3780 cfs_rq->avg.runnable_sum = max_t(u32, cfs_rq->avg.runnable_sum,
3781 cfs_rq->avg.runnable_avg * PELT_MIN_DIVIDER);
3785 update_tg_cfs_load(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq)
3787 long delta_avg, running_sum, runnable_sum = gcfs_rq->prop_runnable_sum;
3788 unsigned long load_avg;
3796 gcfs_rq->prop_runnable_sum = 0;
3799 * cfs_rq->avg.period_contrib can be used for both cfs_rq and se.
3800 * See ___update_load_avg() for details.
3802 divider = get_pelt_divider(&cfs_rq->avg);
3804 if (runnable_sum >= 0) {
3806 * Add runnable; clip at LOAD_AVG_MAX. Reflects that until
3807 * the CPU is saturated running == runnable.
3809 runnable_sum += se->avg.load_sum;
3810 runnable_sum = min_t(long, runnable_sum, divider);
3813 * Estimate the new unweighted runnable_sum of the gcfs_rq by
3814 * assuming all tasks are equally runnable.
3816 if (scale_load_down(gcfs_rq->load.weight)) {
3817 load_sum = div_u64(gcfs_rq->avg.load_sum,
3818 scale_load_down(gcfs_rq->load.weight));
3821 /* But make sure to not inflate se's runnable */
3822 runnable_sum = min(se->avg.load_sum, load_sum);
3826 * runnable_sum can't be lower than running_sum
3827 * Rescale running sum to be in the same range as runnable sum
3828 * running_sum is in [0 : LOAD_AVG_MAX << SCHED_CAPACITY_SHIFT]
3829 * runnable_sum is in [0 : LOAD_AVG_MAX]
3831 running_sum = se->avg.util_sum >> SCHED_CAPACITY_SHIFT;
3832 runnable_sum = max(runnable_sum, running_sum);
3834 load_sum = se_weight(se) * runnable_sum;
3835 load_avg = div_u64(load_sum, divider);
3837 delta_avg = load_avg - se->avg.load_avg;
3841 delta_sum = load_sum - (s64)se_weight(se) * se->avg.load_sum;
3843 se->avg.load_sum = runnable_sum;
3844 se->avg.load_avg = load_avg;
3845 add_positive(&cfs_rq->avg.load_avg, delta_avg);
3846 add_positive(&cfs_rq->avg.load_sum, delta_sum);
3847 /* See update_cfs_rq_load_avg() */
3848 cfs_rq->avg.load_sum = max_t(u32, cfs_rq->avg.load_sum,
3849 cfs_rq->avg.load_avg * PELT_MIN_DIVIDER);
3852 static inline void add_tg_cfs_propagate(struct cfs_rq *cfs_rq, long runnable_sum)
3854 cfs_rq->propagate = 1;
3855 cfs_rq->prop_runnable_sum += runnable_sum;
3858 /* Update task and its cfs_rq load average */
3859 static inline int propagate_entity_load_avg(struct sched_entity *se)
3861 struct cfs_rq *cfs_rq, *gcfs_rq;
3863 if (entity_is_task(se))
3866 gcfs_rq = group_cfs_rq(se);
3867 if (!gcfs_rq->propagate)
3870 gcfs_rq->propagate = 0;
3872 cfs_rq = cfs_rq_of(se);
3874 add_tg_cfs_propagate(cfs_rq, gcfs_rq->prop_runnable_sum);
3876 update_tg_cfs_util(cfs_rq, se, gcfs_rq);
3877 update_tg_cfs_runnable(cfs_rq, se, gcfs_rq);
3878 update_tg_cfs_load(cfs_rq, se, gcfs_rq);
3880 trace_pelt_cfs_tp(cfs_rq);
3881 trace_pelt_se_tp(se);
3887 * Check if we need to update the load and the utilization of a blocked
3890 static inline bool skip_blocked_update(struct sched_entity *se)
3892 struct cfs_rq *gcfs_rq = group_cfs_rq(se);
3895 * If sched_entity still have not zero load or utilization, we have to
3898 if (se->avg.load_avg || se->avg.util_avg)
3902 * If there is a pending propagation, we have to update the load and
3903 * the utilization of the sched_entity:
3905 if (gcfs_rq->propagate)
3909 * Otherwise, the load and the utilization of the sched_entity is
3910 * already zero and there is no pending propagation, so it will be a
3911 * waste of time to try to decay it:
3916 #else /* CONFIG_FAIR_GROUP_SCHED */
3918 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq) {}
3920 static inline int propagate_entity_load_avg(struct sched_entity *se)
3925 static inline void add_tg_cfs_propagate(struct cfs_rq *cfs_rq, long runnable_sum) {}
3927 #endif /* CONFIG_FAIR_GROUP_SCHED */
3929 #ifdef CONFIG_NO_HZ_COMMON
3930 static inline void migrate_se_pelt_lag(struct sched_entity *se)
3932 u64 throttled = 0, now, lut;
3933 struct cfs_rq *cfs_rq;
3937 if (load_avg_is_decayed(&se->avg))
3940 cfs_rq = cfs_rq_of(se);
3944 is_idle = is_idle_task(rcu_dereference(rq->curr));
3948 * The lag estimation comes with a cost we don't want to pay all the
3949 * time. Hence, limiting to the case where the source CPU is idle and
3950 * we know we are at the greatest risk to have an outdated clock.
3956 * Estimated "now" is: last_update_time + cfs_idle_lag + rq_idle_lag, where:
3958 * last_update_time (the cfs_rq's last_update_time)
3959 * = cfs_rq_clock_pelt()@cfs_rq_idle
3960 * = rq_clock_pelt()@cfs_rq_idle
3961 * - cfs->throttled_clock_pelt_time@cfs_rq_idle
3963 * cfs_idle_lag (delta between rq's update and cfs_rq's update)
3964 * = rq_clock_pelt()@rq_idle - rq_clock_pelt()@cfs_rq_idle
3966 * rq_idle_lag (delta between now and rq's update)
3967 * = sched_clock_cpu() - rq_clock()@rq_idle
3969 * We can then write:
3971 * now = rq_clock_pelt()@rq_idle - cfs->throttled_clock_pelt_time +
3972 * sched_clock_cpu() - rq_clock()@rq_idle
3974 * rq_clock_pelt()@rq_idle is rq->clock_pelt_idle
3975 * rq_clock()@rq_idle is rq->clock_idle
3976 * cfs->throttled_clock_pelt_time@cfs_rq_idle
3977 * is cfs_rq->throttled_pelt_idle
3980 #ifdef CONFIG_CFS_BANDWIDTH
3981 throttled = u64_u32_load(cfs_rq->throttled_pelt_idle);
3982 /* The clock has been stopped for throttling */
3983 if (throttled == U64_MAX)
3986 now = u64_u32_load(rq->clock_pelt_idle);
3988 * Paired with _update_idle_rq_clock_pelt(). It ensures at the worst case
3989 * is observed the old clock_pelt_idle value and the new clock_idle,
3990 * which lead to an underestimation. The opposite would lead to an
3994 lut = cfs_rq_last_update_time(cfs_rq);
3999 * cfs_rq->avg.last_update_time is more recent than our
4000 * estimation, let's use it.
4004 now += sched_clock_cpu(cpu_of(rq)) - u64_u32_load(rq->clock_idle);
4006 __update_load_avg_blocked_se(now, se);
4009 static void migrate_se_pelt_lag(struct sched_entity *se) {}
4013 * update_cfs_rq_load_avg - update the cfs_rq's load/util averages
4014 * @now: current time, as per cfs_rq_clock_pelt()
4015 * @cfs_rq: cfs_rq to update
4017 * The cfs_rq avg is the direct sum of all its entities (blocked and runnable)
4018 * avg. The immediate corollary is that all (fair) tasks must be attached.
4020 * cfs_rq->avg is used for task_h_load() and update_cfs_share() for example.
4022 * Return: true if the load decayed or we removed load.
4024 * Since both these conditions indicate a changed cfs_rq->avg.load we should
4025 * call update_tg_load_avg() when this function returns true.
4028 update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq)
4030 unsigned long removed_load = 0, removed_util = 0, removed_runnable = 0;
4031 struct sched_avg *sa = &cfs_rq->avg;
4034 if (cfs_rq->removed.nr) {
4036 u32 divider = get_pelt_divider(&cfs_rq->avg);
4038 raw_spin_lock(&cfs_rq->removed.lock);
4039 swap(cfs_rq->removed.util_avg, removed_util);
4040 swap(cfs_rq->removed.load_avg, removed_load);
4041 swap(cfs_rq->removed.runnable_avg, removed_runnable);
4042 cfs_rq->removed.nr = 0;
4043 raw_spin_unlock(&cfs_rq->removed.lock);
4046 sub_positive(&sa->load_avg, r);
4047 sub_positive(&sa->load_sum, r * divider);
4048 /* See sa->util_sum below */
4049 sa->load_sum = max_t(u32, sa->load_sum, sa->load_avg * PELT_MIN_DIVIDER);
4052 sub_positive(&sa->util_avg, r);
4053 sub_positive(&sa->util_sum, r * divider);
4055 * Because of rounding, se->util_sum might ends up being +1 more than
4056 * cfs->util_sum. Although this is not a problem by itself, detaching
4057 * a lot of tasks with the rounding problem between 2 updates of
4058 * util_avg (~1ms) can make cfs->util_sum becoming null whereas
4059 * cfs_util_avg is not.
4060 * Check that util_sum is still above its lower bound for the new
4061 * util_avg. Given that period_contrib might have moved since the last
4062 * sync, we are only sure that util_sum must be above or equal to
4063 * util_avg * minimum possible divider
4065 sa->util_sum = max_t(u32, sa->util_sum, sa->util_avg * PELT_MIN_DIVIDER);
4067 r = removed_runnable;
4068 sub_positive(&sa->runnable_avg, r);
4069 sub_positive(&sa->runnable_sum, r * divider);
4070 /* See sa->util_sum above */
4071 sa->runnable_sum = max_t(u32, sa->runnable_sum,
4072 sa->runnable_avg * PELT_MIN_DIVIDER);
4075 * removed_runnable is the unweighted version of removed_load so we
4076 * can use it to estimate removed_load_sum.
4078 add_tg_cfs_propagate(cfs_rq,
4079 -(long)(removed_runnable * divider) >> SCHED_CAPACITY_SHIFT);
4084 decayed |= __update_load_avg_cfs_rq(now, cfs_rq);
4085 u64_u32_store_copy(sa->last_update_time,
4086 cfs_rq->last_update_time_copy,
4087 sa->last_update_time);
4092 * attach_entity_load_avg - attach this entity to its cfs_rq load avg
4093 * @cfs_rq: cfs_rq to attach to
4094 * @se: sched_entity to attach
4096 * Must call update_cfs_rq_load_avg() before this, since we rely on
4097 * cfs_rq->avg.last_update_time being current.
4099 static void attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
4102 * cfs_rq->avg.period_contrib can be used for both cfs_rq and se.
4103 * See ___update_load_avg() for details.
4105 u32 divider = get_pelt_divider(&cfs_rq->avg);
4108 * When we attach the @se to the @cfs_rq, we must align the decay
4109 * window because without that, really weird and wonderful things can
4114 se->avg.last_update_time = cfs_rq->avg.last_update_time;
4115 se->avg.period_contrib = cfs_rq->avg.period_contrib;
4118 * Hell(o) Nasty stuff.. we need to recompute _sum based on the new
4119 * period_contrib. This isn't strictly correct, but since we're
4120 * entirely outside of the PELT hierarchy, nobody cares if we truncate
4123 se->avg.util_sum = se->avg.util_avg * divider;
4125 se->avg.runnable_sum = se->avg.runnable_avg * divider;
4127 se->avg.load_sum = se->avg.load_avg * divider;
4128 if (se_weight(se) < se->avg.load_sum)
4129 se->avg.load_sum = div_u64(se->avg.load_sum, se_weight(se));
4131 se->avg.load_sum = 1;
4133 enqueue_load_avg(cfs_rq, se);
4134 cfs_rq->avg.util_avg += se->avg.util_avg;
4135 cfs_rq->avg.util_sum += se->avg.util_sum;
4136 cfs_rq->avg.runnable_avg += se->avg.runnable_avg;
4137 cfs_rq->avg.runnable_sum += se->avg.runnable_sum;
4139 add_tg_cfs_propagate(cfs_rq, se->avg.load_sum);
4141 cfs_rq_util_change(cfs_rq, 0);
4143 trace_pelt_cfs_tp(cfs_rq);
4147 * detach_entity_load_avg - detach this entity from its cfs_rq load avg
4148 * @cfs_rq: cfs_rq to detach from
4149 * @se: sched_entity to detach
4151 * Must call update_cfs_rq_load_avg() before this, since we rely on
4152 * cfs_rq->avg.last_update_time being current.
4154 static void detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
4156 dequeue_load_avg(cfs_rq, se);
4157 sub_positive(&cfs_rq->avg.util_avg, se->avg.util_avg);
4158 sub_positive(&cfs_rq->avg.util_sum, se->avg.util_sum);
4159 /* See update_cfs_rq_load_avg() */
4160 cfs_rq->avg.util_sum = max_t(u32, cfs_rq->avg.util_sum,
4161 cfs_rq->avg.util_avg * PELT_MIN_DIVIDER);
4163 sub_positive(&cfs_rq->avg.runnable_avg, se->avg.runnable_avg);
4164 sub_positive(&cfs_rq->avg.runnable_sum, se->avg.runnable_sum);
4165 /* See update_cfs_rq_load_avg() */
4166 cfs_rq->avg.runnable_sum = max_t(u32, cfs_rq->avg.runnable_sum,
4167 cfs_rq->avg.runnable_avg * PELT_MIN_DIVIDER);
4169 add_tg_cfs_propagate(cfs_rq, -se->avg.load_sum);
4171 cfs_rq_util_change(cfs_rq, 0);
4173 trace_pelt_cfs_tp(cfs_rq);
4177 * Optional action to be done while updating the load average
4179 #define UPDATE_TG 0x1
4180 #define SKIP_AGE_LOAD 0x2
4181 #define DO_ATTACH 0x4
4182 #define DO_DETACH 0x8
4184 /* Update task and its cfs_rq load average */
4185 static inline void update_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
4187 u64 now = cfs_rq_clock_pelt(cfs_rq);
4191 * Track task load average for carrying it to new CPU after migrated, and
4192 * track group sched_entity load average for task_h_load calc in migration
4194 if (se->avg.last_update_time && !(flags & SKIP_AGE_LOAD))
4195 __update_load_avg_se(now, cfs_rq, se);
4197 decayed = update_cfs_rq_load_avg(now, cfs_rq);
4198 decayed |= propagate_entity_load_avg(se);
4200 if (!se->avg.last_update_time && (flags & DO_ATTACH)) {
4203 * DO_ATTACH means we're here from enqueue_entity().
4204 * !last_update_time means we've passed through
4205 * migrate_task_rq_fair() indicating we migrated.
4207 * IOW we're enqueueing a task on a new CPU.
4209 attach_entity_load_avg(cfs_rq, se);
4210 update_tg_load_avg(cfs_rq);
4212 } else if (flags & DO_DETACH) {
4214 * DO_DETACH means we're here from dequeue_entity()
4215 * and we are migrating task out of the CPU.
4217 detach_entity_load_avg(cfs_rq, se);
4218 update_tg_load_avg(cfs_rq);
4219 } else if (decayed) {
4220 cfs_rq_util_change(cfs_rq, 0);
4222 if (flags & UPDATE_TG)
4223 update_tg_load_avg(cfs_rq);
4228 * Synchronize entity load avg of dequeued entity without locking
4231 static void sync_entity_load_avg(struct sched_entity *se)
4233 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4234 u64 last_update_time;
4236 last_update_time = cfs_rq_last_update_time(cfs_rq);
4237 __update_load_avg_blocked_se(last_update_time, se);
4241 * Task first catches up with cfs_rq, and then subtract
4242 * itself from the cfs_rq (task must be off the queue now).
4244 static void remove_entity_load_avg(struct sched_entity *se)
4246 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4247 unsigned long flags;
4250 * tasks cannot exit without having gone through wake_up_new_task() ->
4251 * enqueue_task_fair() which will have added things to the cfs_rq,
4252 * so we can remove unconditionally.
4255 sync_entity_load_avg(se);
4257 raw_spin_lock_irqsave(&cfs_rq->removed.lock, flags);
4258 ++cfs_rq->removed.nr;
4259 cfs_rq->removed.util_avg += se->avg.util_avg;
4260 cfs_rq->removed.load_avg += se->avg.load_avg;
4261 cfs_rq->removed.runnable_avg += se->avg.runnable_avg;
4262 raw_spin_unlock_irqrestore(&cfs_rq->removed.lock, flags);
4265 static inline unsigned long cfs_rq_runnable_avg(struct cfs_rq *cfs_rq)
4267 return cfs_rq->avg.runnable_avg;
4270 static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq)
4272 return cfs_rq->avg.load_avg;
4275 static int newidle_balance(struct rq *this_rq, struct rq_flags *rf);
4277 static inline unsigned long task_util(struct task_struct *p)
4279 return READ_ONCE(p->se.avg.util_avg);
4282 static inline unsigned long _task_util_est(struct task_struct *p)
4284 struct util_est ue = READ_ONCE(p->se.avg.util_est);
4286 return max(ue.ewma, (ue.enqueued & ~UTIL_AVG_UNCHANGED));
4289 static inline unsigned long task_util_est(struct task_struct *p)
4291 return max(task_util(p), _task_util_est(p));
4294 #ifdef CONFIG_UCLAMP_TASK
4295 static inline unsigned long uclamp_task_util(struct task_struct *p,
4296 unsigned long uclamp_min,
4297 unsigned long uclamp_max)
4299 return clamp(task_util_est(p), uclamp_min, uclamp_max);
4302 static inline unsigned long uclamp_task_util(struct task_struct *p,
4303 unsigned long uclamp_min,
4304 unsigned long uclamp_max)
4306 return task_util_est(p);
4310 static inline void util_est_enqueue(struct cfs_rq *cfs_rq,
4311 struct task_struct *p)
4313 unsigned int enqueued;
4315 if (!sched_feat(UTIL_EST))
4318 /* Update root cfs_rq's estimated utilization */
4319 enqueued = cfs_rq->avg.util_est.enqueued;
4320 enqueued += _task_util_est(p);
4321 WRITE_ONCE(cfs_rq->avg.util_est.enqueued, enqueued);
4323 trace_sched_util_est_cfs_tp(cfs_rq);
4326 static inline void util_est_dequeue(struct cfs_rq *cfs_rq,
4327 struct task_struct *p)
4329 unsigned int enqueued;
4331 if (!sched_feat(UTIL_EST))
4334 /* Update root cfs_rq's estimated utilization */
4335 enqueued = cfs_rq->avg.util_est.enqueued;
4336 enqueued -= min_t(unsigned int, enqueued, _task_util_est(p));
4337 WRITE_ONCE(cfs_rq->avg.util_est.enqueued, enqueued);
4339 trace_sched_util_est_cfs_tp(cfs_rq);
4342 #define UTIL_EST_MARGIN (SCHED_CAPACITY_SCALE / 100)
4345 * Check if a (signed) value is within a specified (unsigned) margin,
4346 * based on the observation that:
4348 * abs(x) < y := (unsigned)(x + y - 1) < (2 * y - 1)
4350 * NOTE: this only works when value + margin < INT_MAX.
4352 static inline bool within_margin(int value, int margin)
4354 return ((unsigned int)(value + margin - 1) < (2 * margin - 1));
4357 static inline void util_est_update(struct cfs_rq *cfs_rq,
4358 struct task_struct *p,
4361 long last_ewma_diff, last_enqueued_diff;
4364 if (!sched_feat(UTIL_EST))
4368 * Skip update of task's estimated utilization when the task has not
4369 * yet completed an activation, e.g. being migrated.
4375 * If the PELT values haven't changed since enqueue time,
4376 * skip the util_est update.
4378 ue = p->se.avg.util_est;
4379 if (ue.enqueued & UTIL_AVG_UNCHANGED)
4382 last_enqueued_diff = ue.enqueued;
4385 * Reset EWMA on utilization increases, the moving average is used only
4386 * to smooth utilization decreases.
4388 ue.enqueued = task_util(p);
4389 if (sched_feat(UTIL_EST_FASTUP)) {
4390 if (ue.ewma < ue.enqueued) {
4391 ue.ewma = ue.enqueued;
4397 * Skip update of task's estimated utilization when its members are
4398 * already ~1% close to its last activation value.
4400 last_ewma_diff = ue.enqueued - ue.ewma;
4401 last_enqueued_diff -= ue.enqueued;
4402 if (within_margin(last_ewma_diff, UTIL_EST_MARGIN)) {
4403 if (!within_margin(last_enqueued_diff, UTIL_EST_MARGIN))
4410 * To avoid overestimation of actual task utilization, skip updates if
4411 * we cannot grant there is idle time in this CPU.
4413 if (task_util(p) > capacity_orig_of(cpu_of(rq_of(cfs_rq))))
4417 * Update Task's estimated utilization
4419 * When *p completes an activation we can consolidate another sample
4420 * of the task size. This is done by storing the current PELT value
4421 * as ue.enqueued and by using this value to update the Exponential
4422 * Weighted Moving Average (EWMA):
4424 * ewma(t) = w * task_util(p) + (1-w) * ewma(t-1)
4425 * = w * task_util(p) + ewma(t-1) - w * ewma(t-1)
4426 * = w * (task_util(p) - ewma(t-1)) + ewma(t-1)
4427 * = w * ( last_ewma_diff ) + ewma(t-1)
4428 * = w * (last_ewma_diff + ewma(t-1) / w)
4430 * Where 'w' is the weight of new samples, which is configured to be
4431 * 0.25, thus making w=1/4 ( >>= UTIL_EST_WEIGHT_SHIFT)
4433 ue.ewma <<= UTIL_EST_WEIGHT_SHIFT;
4434 ue.ewma += last_ewma_diff;
4435 ue.ewma >>= UTIL_EST_WEIGHT_SHIFT;
4437 ue.enqueued |= UTIL_AVG_UNCHANGED;
4438 WRITE_ONCE(p->se.avg.util_est, ue);
4440 trace_sched_util_est_se_tp(&p->se);
4443 static inline int util_fits_cpu(unsigned long util,
4444 unsigned long uclamp_min,
4445 unsigned long uclamp_max,
4448 unsigned long capacity_orig, capacity_orig_thermal;
4449 unsigned long capacity = capacity_of(cpu);
4450 bool fits, uclamp_max_fits;
4453 * Check if the real util fits without any uclamp boost/cap applied.
4455 fits = fits_capacity(util, capacity);
4457 if (!uclamp_is_used())
4461 * We must use capacity_orig_of() for comparing against uclamp_min and
4462 * uclamp_max. We only care about capacity pressure (by using
4463 * capacity_of()) for comparing against the real util.
4465 * If a task is boosted to 1024 for example, we don't want a tiny
4466 * pressure to skew the check whether it fits a CPU or not.
4468 * Similarly if a task is capped to capacity_orig_of(little_cpu), it
4469 * should fit a little cpu even if there's some pressure.
4471 * Only exception is for thermal pressure since it has a direct impact
4472 * on available OPP of the system.
4474 * We honour it for uclamp_min only as a drop in performance level
4475 * could result in not getting the requested minimum performance level.
4477 * For uclamp_max, we can tolerate a drop in performance level as the
4478 * goal is to cap the task. So it's okay if it's getting less.
4480 * In case of capacity inversion we should honour the inverted capacity
4481 * for both uclamp_min and uclamp_max all the time.
4483 capacity_orig = cpu_in_capacity_inversion(cpu);
4484 if (capacity_orig) {
4485 capacity_orig_thermal = capacity_orig;
4487 capacity_orig = capacity_orig_of(cpu);
4488 capacity_orig_thermal = capacity_orig - arch_scale_thermal_pressure(cpu);
4492 * We want to force a task to fit a cpu as implied by uclamp_max.
4493 * But we do have some corner cases to cater for..
4499 * |_ _ _ _ _ _ _ _ _ ___ _ _ _ | _ | _ _ _ _ _ uclamp_max
4502 * | | | | | | | (util somewhere in this region)
4505 * +----------------------------------------
4508 * In the above example if a task is capped to a specific performance
4509 * point, y, then when:
4511 * * util = 80% of x then it does not fit on cpu0 and should migrate
4513 * * util = 80% of y then it is forced to fit on cpu1 to honour
4514 * uclamp_max request.
4516 * which is what we're enforcing here. A task always fits if
4517 * uclamp_max <= capacity_orig. But when uclamp_max > capacity_orig,
4518 * the normal upmigration rules should withhold still.
4520 * Only exception is when we are on max capacity, then we need to be
4521 * careful not to block overutilized state. This is so because:
4523 * 1. There's no concept of capping at max_capacity! We can't go
4524 * beyond this performance level anyway.
4525 * 2. The system is being saturated when we're operating near
4526 * max capacity, it doesn't make sense to block overutilized.
4528 uclamp_max_fits = (capacity_orig == SCHED_CAPACITY_SCALE) && (uclamp_max == SCHED_CAPACITY_SCALE);
4529 uclamp_max_fits = !uclamp_max_fits && (uclamp_max <= capacity_orig);
4530 fits = fits || uclamp_max_fits;
4535 * | ___ (region a, capped, util >= uclamp_max)
4537 * |_ _ _ _ _ _ _ _ _ ___ _ _ _ | _ | _ _ _ _ _ uclamp_max
4539 * | ___ | | | | (region b, uclamp_min <= util <= uclamp_max)
4540 * |_ _ _|_ _|_ _ _ _| _ | _ _ _| _ | _ _ _ _ _ uclamp_min
4542 * | | | | | | | (region c, boosted, util < uclamp_min)
4543 * +----------------------------------------
4546 * a) If util > uclamp_max, then we're capped, we don't care about
4547 * actual fitness value here. We only care if uclamp_max fits
4548 * capacity without taking margin/pressure into account.
4549 * See comment above.
4551 * b) If uclamp_min <= util <= uclamp_max, then the normal
4552 * fits_capacity() rules apply. Except we need to ensure that we
4553 * enforce we remain within uclamp_max, see comment above.
4555 * c) If util < uclamp_min, then we are boosted. Same as (b) but we
4556 * need to take into account the boosted value fits the CPU without
4557 * taking margin/pressure into account.
4559 * Cases (a) and (b) are handled in the 'fits' variable already. We
4560 * just need to consider an extra check for case (c) after ensuring we
4561 * handle the case uclamp_min > uclamp_max.
4563 uclamp_min = min(uclamp_min, uclamp_max);
4564 if (util < uclamp_min && capacity_orig != SCHED_CAPACITY_SCALE)
4565 fits = fits && (uclamp_min <= capacity_orig_thermal);
4570 static inline int task_fits_cpu(struct task_struct *p, int cpu)
4572 unsigned long uclamp_min = uclamp_eff_value(p, UCLAMP_MIN);
4573 unsigned long uclamp_max = uclamp_eff_value(p, UCLAMP_MAX);
4574 unsigned long util = task_util_est(p);
4575 return util_fits_cpu(util, uclamp_min, uclamp_max, cpu);
4578 static inline void update_misfit_status(struct task_struct *p, struct rq *rq)
4580 if (!sched_asym_cpucap_active())
4583 if (!p || p->nr_cpus_allowed == 1) {
4584 rq->misfit_task_load = 0;
4588 if (task_fits_cpu(p, cpu_of(rq))) {
4589 rq->misfit_task_load = 0;
4594 * Make sure that misfit_task_load will not be null even if
4595 * task_h_load() returns 0.
4597 rq->misfit_task_load = max_t(unsigned long, task_h_load(p), 1);
4600 #else /* CONFIG_SMP */
4602 static inline bool cfs_rq_is_decayed(struct cfs_rq *cfs_rq)
4607 #define UPDATE_TG 0x0
4608 #define SKIP_AGE_LOAD 0x0
4609 #define DO_ATTACH 0x0
4610 #define DO_DETACH 0x0
4612 static inline void update_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int not_used1)
4614 cfs_rq_util_change(cfs_rq, 0);
4617 static inline void remove_entity_load_avg(struct sched_entity *se) {}
4620 attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
4622 detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
4624 static inline int newidle_balance(struct rq *rq, struct rq_flags *rf)
4630 util_est_enqueue(struct cfs_rq *cfs_rq, struct task_struct *p) {}
4633 util_est_dequeue(struct cfs_rq *cfs_rq, struct task_struct *p) {}
4636 util_est_update(struct cfs_rq *cfs_rq, struct task_struct *p,
4638 static inline void update_misfit_status(struct task_struct *p, struct rq *rq) {}
4640 #endif /* CONFIG_SMP */
4642 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
4644 #ifdef CONFIG_SCHED_DEBUG
4645 s64 d = se->vruntime - cfs_rq->min_vruntime;
4650 if (d > 3*sysctl_sched_latency)
4651 schedstat_inc(cfs_rq->nr_spread_over);
4656 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
4658 u64 vruntime = cfs_rq->min_vruntime;
4661 * The 'current' period is already promised to the current tasks,
4662 * however the extra weight of the new task will slow them down a
4663 * little, place the new task so that it fits in the slot that
4664 * stays open at the end.
4666 if (initial && sched_feat(START_DEBIT))
4667 vruntime += sched_vslice(cfs_rq, se);
4669 /* sleeps up to a single latency don't count. */
4671 unsigned long thresh;
4674 thresh = sysctl_sched_min_granularity;
4676 thresh = sysctl_sched_latency;
4679 * Halve their sleep time's effect, to allow
4680 * for a gentler effect of sleepers:
4682 if (sched_feat(GENTLE_FAIR_SLEEPERS))
4688 /* ensure we never gain time by being placed backwards. */
4689 se->vruntime = max_vruntime(se->vruntime, vruntime);
4692 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
4694 static inline bool cfs_bandwidth_used(void);
4701 * update_min_vruntime()
4702 * vruntime -= min_vruntime
4706 * update_min_vruntime()
4707 * vruntime += min_vruntime
4709 * this way the vruntime transition between RQs is done when both
4710 * min_vruntime are up-to-date.
4714 * ->migrate_task_rq_fair() (p->state == TASK_WAKING)
4715 * vruntime -= min_vruntime
4719 * update_min_vruntime()
4720 * vruntime += min_vruntime
4722 * this way we don't have the most up-to-date min_vruntime on the originating
4723 * CPU and an up-to-date min_vruntime on the destination CPU.
4727 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
4729 bool renorm = !(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_MIGRATED);
4730 bool curr = cfs_rq->curr == se;
4733 * If we're the current task, we must renormalise before calling
4737 se->vruntime += cfs_rq->min_vruntime;
4739 update_curr(cfs_rq);
4742 * Otherwise, renormalise after, such that we're placed at the current
4743 * moment in time, instead of some random moment in the past. Being
4744 * placed in the past could significantly boost this task to the
4745 * fairness detriment of existing tasks.
4747 if (renorm && !curr)
4748 se->vruntime += cfs_rq->min_vruntime;
4751 * When enqueuing a sched_entity, we must:
4752 * - Update loads to have both entity and cfs_rq synced with now.
4753 * - For group_entity, update its runnable_weight to reflect the new
4754 * h_nr_running of its group cfs_rq.
4755 * - For group_entity, update its weight to reflect the new share of
4757 * - Add its new weight to cfs_rq->load.weight
4759 update_load_avg(cfs_rq, se, UPDATE_TG | DO_ATTACH);
4760 se_update_runnable(se);
4761 update_cfs_group(se);
4762 account_entity_enqueue(cfs_rq, se);
4764 if (flags & ENQUEUE_WAKEUP)
4765 place_entity(cfs_rq, se, 0);
4767 check_schedstat_required();
4768 update_stats_enqueue_fair(cfs_rq, se, flags);
4769 check_spread(cfs_rq, se);
4771 __enqueue_entity(cfs_rq, se);
4774 if (cfs_rq->nr_running == 1) {
4775 check_enqueue_throttle(cfs_rq);
4776 if (!throttled_hierarchy(cfs_rq))
4777 list_add_leaf_cfs_rq(cfs_rq);
4781 static void __clear_buddies_last(struct sched_entity *se)
4783 for_each_sched_entity(se) {
4784 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4785 if (cfs_rq->last != se)
4788 cfs_rq->last = NULL;
4792 static void __clear_buddies_next(struct sched_entity *se)
4794 for_each_sched_entity(se) {
4795 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4796 if (cfs_rq->next != se)
4799 cfs_rq->next = NULL;
4803 static void __clear_buddies_skip(struct sched_entity *se)
4805 for_each_sched_entity(se) {
4806 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4807 if (cfs_rq->skip != se)
4810 cfs_rq->skip = NULL;
4814 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
4816 if (cfs_rq->last == se)
4817 __clear_buddies_last(se);
4819 if (cfs_rq->next == se)
4820 __clear_buddies_next(se);
4822 if (cfs_rq->skip == se)
4823 __clear_buddies_skip(se);
4826 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
4829 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
4831 int action = UPDATE_TG;
4833 if (entity_is_task(se) && task_on_rq_migrating(task_of(se)))
4834 action |= DO_DETACH;
4837 * Update run-time statistics of the 'current'.
4839 update_curr(cfs_rq);
4842 * When dequeuing a sched_entity, we must:
4843 * - Update loads to have both entity and cfs_rq synced with now.
4844 * - For group_entity, update its runnable_weight to reflect the new
4845 * h_nr_running of its group cfs_rq.
4846 * - Subtract its previous weight from cfs_rq->load.weight.
4847 * - For group entity, update its weight to reflect the new share
4848 * of its group cfs_rq.
4850 update_load_avg(cfs_rq, se, action);
4851 se_update_runnable(se);
4853 update_stats_dequeue_fair(cfs_rq, se, flags);
4855 clear_buddies(cfs_rq, se);
4857 if (se != cfs_rq->curr)
4858 __dequeue_entity(cfs_rq, se);
4860 account_entity_dequeue(cfs_rq, se);
4863 * Normalize after update_curr(); which will also have moved
4864 * min_vruntime if @se is the one holding it back. But before doing
4865 * update_min_vruntime() again, which will discount @se's position and
4866 * can move min_vruntime forward still more.
4868 if (!(flags & DEQUEUE_SLEEP))
4869 se->vruntime -= cfs_rq->min_vruntime;
4871 /* return excess runtime on last dequeue */
4872 return_cfs_rq_runtime(cfs_rq);
4874 update_cfs_group(se);
4877 * Now advance min_vruntime if @se was the entity holding it back,
4878 * except when: DEQUEUE_SAVE && !DEQUEUE_MOVE, in this case we'll be
4879 * put back on, and if we advance min_vruntime, we'll be placed back
4880 * further than we started -- ie. we'll be penalized.
4882 if ((flags & (DEQUEUE_SAVE | DEQUEUE_MOVE)) != DEQUEUE_SAVE)
4883 update_min_vruntime(cfs_rq);
4885 if (cfs_rq->nr_running == 0)
4886 update_idle_cfs_rq_clock_pelt(cfs_rq);
4890 * Preempt the current task with a newly woken task if needed:
4893 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
4895 unsigned long ideal_runtime, delta_exec;
4896 struct sched_entity *se;
4899 ideal_runtime = sched_slice(cfs_rq, curr);
4900 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
4901 if (delta_exec > ideal_runtime) {
4902 resched_curr(rq_of(cfs_rq));
4904 * The current task ran long enough, ensure it doesn't get
4905 * re-elected due to buddy favours.
4907 clear_buddies(cfs_rq, curr);
4912 * Ensure that a task that missed wakeup preemption by a
4913 * narrow margin doesn't have to wait for a full slice.
4914 * This also mitigates buddy induced latencies under load.
4916 if (delta_exec < sysctl_sched_min_granularity)
4919 se = __pick_first_entity(cfs_rq);
4920 delta = curr->vruntime - se->vruntime;
4925 if (delta > ideal_runtime)
4926 resched_curr(rq_of(cfs_rq));
4930 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
4932 clear_buddies(cfs_rq, se);
4934 /* 'current' is not kept within the tree. */
4937 * Any task has to be enqueued before it get to execute on
4938 * a CPU. So account for the time it spent waiting on the
4941 update_stats_wait_end_fair(cfs_rq, se);
4942 __dequeue_entity(cfs_rq, se);
4943 update_load_avg(cfs_rq, se, UPDATE_TG);
4946 update_stats_curr_start(cfs_rq, se);
4950 * Track our maximum slice length, if the CPU's load is at
4951 * least twice that of our own weight (i.e. dont track it
4952 * when there are only lesser-weight tasks around):
4954 if (schedstat_enabled() &&
4955 rq_of(cfs_rq)->cfs.load.weight >= 2*se->load.weight) {
4956 struct sched_statistics *stats;
4958 stats = __schedstats_from_se(se);
4959 __schedstat_set(stats->slice_max,
4960 max((u64)stats->slice_max,
4961 se->sum_exec_runtime - se->prev_sum_exec_runtime));
4964 se->prev_sum_exec_runtime = se->sum_exec_runtime;
4968 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
4971 * Pick the next process, keeping these things in mind, in this order:
4972 * 1) keep things fair between processes/task groups
4973 * 2) pick the "next" process, since someone really wants that to run
4974 * 3) pick the "last" process, for cache locality
4975 * 4) do not run the "skip" process, if something else is available
4977 static struct sched_entity *
4978 pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr)
4980 struct sched_entity *left = __pick_first_entity(cfs_rq);
4981 struct sched_entity *se;
4984 * If curr is set we have to see if its left of the leftmost entity
4985 * still in the tree, provided there was anything in the tree at all.
4987 if (!left || (curr && entity_before(curr, left)))
4990 se = left; /* ideally we run the leftmost entity */
4993 * Avoid running the skip buddy, if running something else can
4994 * be done without getting too unfair.
4996 if (cfs_rq->skip && cfs_rq->skip == se) {
4997 struct sched_entity *second;
5000 second = __pick_first_entity(cfs_rq);
5002 second = __pick_next_entity(se);
5003 if (!second || (curr && entity_before(curr, second)))
5007 if (second && wakeup_preempt_entity(second, left) < 1)
5011 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1) {
5013 * Someone really wants this to run. If it's not unfair, run it.
5016 } else if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1) {
5018 * Prefer last buddy, try to return the CPU to a preempted task.
5026 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
5028 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
5031 * If still on the runqueue then deactivate_task()
5032 * was not called and update_curr() has to be done:
5035 update_curr(cfs_rq);
5037 /* throttle cfs_rqs exceeding runtime */
5038 check_cfs_rq_runtime(cfs_rq);
5040 check_spread(cfs_rq, prev);
5043 update_stats_wait_start_fair(cfs_rq, prev);
5044 /* Put 'current' back into the tree. */
5045 __enqueue_entity(cfs_rq, prev);
5046 /* in !on_rq case, update occurred at dequeue */
5047 update_load_avg(cfs_rq, prev, 0);
5049 cfs_rq->curr = NULL;
5053 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
5056 * Update run-time statistics of the 'current'.
5058 update_curr(cfs_rq);
5061 * Ensure that runnable average is periodically updated.
5063 update_load_avg(cfs_rq, curr, UPDATE_TG);
5064 update_cfs_group(curr);
5066 #ifdef CONFIG_SCHED_HRTICK
5068 * queued ticks are scheduled to match the slice, so don't bother
5069 * validating it and just reschedule.
5072 resched_curr(rq_of(cfs_rq));
5076 * don't let the period tick interfere with the hrtick preemption
5078 if (!sched_feat(DOUBLE_TICK) &&
5079 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
5083 if (cfs_rq->nr_running > 1)
5084 check_preempt_tick(cfs_rq, curr);
5088 /**************************************************
5089 * CFS bandwidth control machinery
5092 #ifdef CONFIG_CFS_BANDWIDTH
5094 #ifdef CONFIG_JUMP_LABEL
5095 static struct static_key __cfs_bandwidth_used;
5097 static inline bool cfs_bandwidth_used(void)
5099 return static_key_false(&__cfs_bandwidth_used);
5102 void cfs_bandwidth_usage_inc(void)
5104 static_key_slow_inc_cpuslocked(&__cfs_bandwidth_used);
5107 void cfs_bandwidth_usage_dec(void)
5109 static_key_slow_dec_cpuslocked(&__cfs_bandwidth_used);
5111 #else /* CONFIG_JUMP_LABEL */
5112 static bool cfs_bandwidth_used(void)
5117 void cfs_bandwidth_usage_inc(void) {}
5118 void cfs_bandwidth_usage_dec(void) {}
5119 #endif /* CONFIG_JUMP_LABEL */
5122 * default period for cfs group bandwidth.
5123 * default: 0.1s, units: nanoseconds
5125 static inline u64 default_cfs_period(void)
5127 return 100000000ULL;
5130 static inline u64 sched_cfs_bandwidth_slice(void)
5132 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
5136 * Replenish runtime according to assigned quota. We use sched_clock_cpu
5137 * directly instead of rq->clock to avoid adding additional synchronization
5140 * requires cfs_b->lock
5142 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
5146 if (unlikely(cfs_b->quota == RUNTIME_INF))
5149 cfs_b->runtime += cfs_b->quota;
5150 runtime = cfs_b->runtime_snap - cfs_b->runtime;
5152 cfs_b->burst_time += runtime;
5156 cfs_b->runtime = min(cfs_b->runtime, cfs_b->quota + cfs_b->burst);
5157 cfs_b->runtime_snap = cfs_b->runtime;
5160 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
5162 return &tg->cfs_bandwidth;
5165 /* returns 0 on failure to allocate runtime */
5166 static int __assign_cfs_rq_runtime(struct cfs_bandwidth *cfs_b,
5167 struct cfs_rq *cfs_rq, u64 target_runtime)
5169 u64 min_amount, amount = 0;
5171 lockdep_assert_held(&cfs_b->lock);
5173 /* note: this is a positive sum as runtime_remaining <= 0 */
5174 min_amount = target_runtime - cfs_rq->runtime_remaining;
5176 if (cfs_b->quota == RUNTIME_INF)
5177 amount = min_amount;
5179 start_cfs_bandwidth(cfs_b);
5181 if (cfs_b->runtime > 0) {
5182 amount = min(cfs_b->runtime, min_amount);
5183 cfs_b->runtime -= amount;
5188 cfs_rq->runtime_remaining += amount;
5190 return cfs_rq->runtime_remaining > 0;
5193 /* returns 0 on failure to allocate runtime */
5194 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
5196 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
5199 raw_spin_lock(&cfs_b->lock);
5200 ret = __assign_cfs_rq_runtime(cfs_b, cfs_rq, sched_cfs_bandwidth_slice());
5201 raw_spin_unlock(&cfs_b->lock);
5206 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
5208 /* dock delta_exec before expiring quota (as it could span periods) */
5209 cfs_rq->runtime_remaining -= delta_exec;
5211 if (likely(cfs_rq->runtime_remaining > 0))
5214 if (cfs_rq->throttled)
5217 * if we're unable to extend our runtime we resched so that the active
5218 * hierarchy can be throttled
5220 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
5221 resched_curr(rq_of(cfs_rq));
5224 static __always_inline
5225 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
5227 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
5230 __account_cfs_rq_runtime(cfs_rq, delta_exec);
5233 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
5235 return cfs_bandwidth_used() && cfs_rq->throttled;
5238 /* check whether cfs_rq, or any parent, is throttled */
5239 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
5241 return cfs_bandwidth_used() && cfs_rq->throttle_count;
5245 * Ensure that neither of the group entities corresponding to src_cpu or
5246 * dest_cpu are members of a throttled hierarchy when performing group
5247 * load-balance operations.
5249 static inline int throttled_lb_pair(struct task_group *tg,
5250 int src_cpu, int dest_cpu)
5252 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
5254 src_cfs_rq = tg->cfs_rq[src_cpu];
5255 dest_cfs_rq = tg->cfs_rq[dest_cpu];
5257 return throttled_hierarchy(src_cfs_rq) ||
5258 throttled_hierarchy(dest_cfs_rq);
5261 static int tg_unthrottle_up(struct task_group *tg, void *data)
5263 struct rq *rq = data;
5264 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
5266 cfs_rq->throttle_count--;
5267 if (!cfs_rq->throttle_count) {
5268 cfs_rq->throttled_clock_pelt_time += rq_clock_pelt(rq) -
5269 cfs_rq->throttled_clock_pelt;
5271 /* Add cfs_rq with load or one or more already running entities to the list */
5272 if (!cfs_rq_is_decayed(cfs_rq))
5273 list_add_leaf_cfs_rq(cfs_rq);
5279 static int tg_throttle_down(struct task_group *tg, void *data)
5281 struct rq *rq = data;
5282 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
5284 /* group is entering throttled state, stop time */
5285 if (!cfs_rq->throttle_count) {
5286 cfs_rq->throttled_clock_pelt = rq_clock_pelt(rq);
5287 list_del_leaf_cfs_rq(cfs_rq);
5289 cfs_rq->throttle_count++;
5294 static bool throttle_cfs_rq(struct cfs_rq *cfs_rq)
5296 struct rq *rq = rq_of(cfs_rq);
5297 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
5298 struct sched_entity *se;
5299 long task_delta, idle_task_delta, dequeue = 1;
5301 raw_spin_lock(&cfs_b->lock);
5302 /* This will start the period timer if necessary */
5303 if (__assign_cfs_rq_runtime(cfs_b, cfs_rq, 1)) {
5305 * We have raced with bandwidth becoming available, and if we
5306 * actually throttled the timer might not unthrottle us for an
5307 * entire period. We additionally needed to make sure that any
5308 * subsequent check_cfs_rq_runtime calls agree not to throttle
5309 * us, as we may commit to do cfs put_prev+pick_next, so we ask
5310 * for 1ns of runtime rather than just check cfs_b.
5314 list_add_tail_rcu(&cfs_rq->throttled_list,
5315 &cfs_b->throttled_cfs_rq);
5317 raw_spin_unlock(&cfs_b->lock);
5320 return false; /* Throttle no longer required. */
5322 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
5324 /* freeze hierarchy runnable averages while throttled */
5326 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
5329 task_delta = cfs_rq->h_nr_running;
5330 idle_task_delta = cfs_rq->idle_h_nr_running;
5331 for_each_sched_entity(se) {
5332 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
5333 /* throttled entity or throttle-on-deactivate */
5337 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
5339 if (cfs_rq_is_idle(group_cfs_rq(se)))
5340 idle_task_delta = cfs_rq->h_nr_running;
5342 qcfs_rq->h_nr_running -= task_delta;
5343 qcfs_rq->idle_h_nr_running -= idle_task_delta;
5345 if (qcfs_rq->load.weight) {
5346 /* Avoid re-evaluating load for this entity: */
5347 se = parent_entity(se);
5352 for_each_sched_entity(se) {
5353 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
5354 /* throttled entity or throttle-on-deactivate */
5358 update_load_avg(qcfs_rq, se, 0);
5359 se_update_runnable(se);
5361 if (cfs_rq_is_idle(group_cfs_rq(se)))
5362 idle_task_delta = cfs_rq->h_nr_running;
5364 qcfs_rq->h_nr_running -= task_delta;
5365 qcfs_rq->idle_h_nr_running -= idle_task_delta;
5368 /* At this point se is NULL and we are at root level*/
5369 sub_nr_running(rq, task_delta);
5373 * Note: distribution will already see us throttled via the
5374 * throttled-list. rq->lock protects completion.
5376 cfs_rq->throttled = 1;
5377 cfs_rq->throttled_clock = rq_clock(rq);
5381 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
5383 struct rq *rq = rq_of(cfs_rq);
5384 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
5385 struct sched_entity *se;
5386 long task_delta, idle_task_delta;
5388 se = cfs_rq->tg->se[cpu_of(rq)];
5390 cfs_rq->throttled = 0;
5392 update_rq_clock(rq);
5394 raw_spin_lock(&cfs_b->lock);
5395 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
5396 list_del_rcu(&cfs_rq->throttled_list);
5397 raw_spin_unlock(&cfs_b->lock);
5399 /* update hierarchical throttle state */
5400 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
5402 if (!cfs_rq->load.weight) {
5403 if (!cfs_rq->on_list)
5406 * Nothing to run but something to decay (on_list)?
5407 * Complete the branch.
5409 for_each_sched_entity(se) {
5410 if (list_add_leaf_cfs_rq(cfs_rq_of(se)))
5413 goto unthrottle_throttle;
5416 task_delta = cfs_rq->h_nr_running;
5417 idle_task_delta = cfs_rq->idle_h_nr_running;
5418 for_each_sched_entity(se) {
5419 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
5423 enqueue_entity(qcfs_rq, se, ENQUEUE_WAKEUP);
5425 if (cfs_rq_is_idle(group_cfs_rq(se)))
5426 idle_task_delta = cfs_rq->h_nr_running;
5428 qcfs_rq->h_nr_running += task_delta;
5429 qcfs_rq->idle_h_nr_running += idle_task_delta;
5431 /* end evaluation on encountering a throttled cfs_rq */
5432 if (cfs_rq_throttled(qcfs_rq))
5433 goto unthrottle_throttle;
5436 for_each_sched_entity(se) {
5437 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
5439 update_load_avg(qcfs_rq, se, UPDATE_TG);
5440 se_update_runnable(se);
5442 if (cfs_rq_is_idle(group_cfs_rq(se)))
5443 idle_task_delta = cfs_rq->h_nr_running;
5445 qcfs_rq->h_nr_running += task_delta;
5446 qcfs_rq->idle_h_nr_running += idle_task_delta;
5448 /* end evaluation on encountering a throttled cfs_rq */
5449 if (cfs_rq_throttled(qcfs_rq))
5450 goto unthrottle_throttle;
5453 /* At this point se is NULL and we are at root level*/
5454 add_nr_running(rq, task_delta);
5456 unthrottle_throttle:
5457 assert_list_leaf_cfs_rq(rq);
5459 /* Determine whether we need to wake up potentially idle CPU: */
5460 if (rq->curr == rq->idle && rq->cfs.nr_running)
5464 static void distribute_cfs_runtime(struct cfs_bandwidth *cfs_b)
5466 struct cfs_rq *cfs_rq;
5467 u64 runtime, remaining = 1;
5470 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
5472 struct rq *rq = rq_of(cfs_rq);
5475 rq_lock_irqsave(rq, &rf);
5476 if (!cfs_rq_throttled(cfs_rq))
5479 /* By the above check, this should never be true */
5480 SCHED_WARN_ON(cfs_rq->runtime_remaining > 0);
5482 raw_spin_lock(&cfs_b->lock);
5483 runtime = -cfs_rq->runtime_remaining + 1;
5484 if (runtime > cfs_b->runtime)
5485 runtime = cfs_b->runtime;
5486 cfs_b->runtime -= runtime;
5487 remaining = cfs_b->runtime;
5488 raw_spin_unlock(&cfs_b->lock);
5490 cfs_rq->runtime_remaining += runtime;
5492 /* we check whether we're throttled above */
5493 if (cfs_rq->runtime_remaining > 0)
5494 unthrottle_cfs_rq(cfs_rq);
5497 rq_unlock_irqrestore(rq, &rf);
5506 * Responsible for refilling a task_group's bandwidth and unthrottling its
5507 * cfs_rqs as appropriate. If there has been no activity within the last
5508 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
5509 * used to track this state.
5511 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun, unsigned long flags)
5515 /* no need to continue the timer with no bandwidth constraint */
5516 if (cfs_b->quota == RUNTIME_INF)
5517 goto out_deactivate;
5519 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
5520 cfs_b->nr_periods += overrun;
5522 /* Refill extra burst quota even if cfs_b->idle */
5523 __refill_cfs_bandwidth_runtime(cfs_b);
5526 * idle depends on !throttled (for the case of a large deficit), and if
5527 * we're going inactive then everything else can be deferred
5529 if (cfs_b->idle && !throttled)
5530 goto out_deactivate;
5533 /* mark as potentially idle for the upcoming period */
5538 /* account preceding periods in which throttling occurred */
5539 cfs_b->nr_throttled += overrun;
5542 * This check is repeated as we release cfs_b->lock while we unthrottle.
5544 while (throttled && cfs_b->runtime > 0) {
5545 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
5546 /* we can't nest cfs_b->lock while distributing bandwidth */
5547 distribute_cfs_runtime(cfs_b);
5548 raw_spin_lock_irqsave(&cfs_b->lock, flags);
5550 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
5554 * While we are ensured activity in the period following an
5555 * unthrottle, this also covers the case in which the new bandwidth is
5556 * insufficient to cover the existing bandwidth deficit. (Forcing the
5557 * timer to remain active while there are any throttled entities.)
5567 /* a cfs_rq won't donate quota below this amount */
5568 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
5569 /* minimum remaining period time to redistribute slack quota */
5570 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
5571 /* how long we wait to gather additional slack before distributing */
5572 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
5575 * Are we near the end of the current quota period?
5577 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
5578 * hrtimer base being cleared by hrtimer_start. In the case of
5579 * migrate_hrtimers, base is never cleared, so we are fine.
5581 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
5583 struct hrtimer *refresh_timer = &cfs_b->period_timer;
5586 /* if the call-back is running a quota refresh is already occurring */
5587 if (hrtimer_callback_running(refresh_timer))
5590 /* is a quota refresh about to occur? */
5591 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
5592 if (remaining < (s64)min_expire)
5598 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
5600 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
5602 /* if there's a quota refresh soon don't bother with slack */
5603 if (runtime_refresh_within(cfs_b, min_left))
5606 /* don't push forwards an existing deferred unthrottle */
5607 if (cfs_b->slack_started)
5609 cfs_b->slack_started = true;
5611 hrtimer_start(&cfs_b->slack_timer,
5612 ns_to_ktime(cfs_bandwidth_slack_period),
5616 /* we know any runtime found here is valid as update_curr() precedes return */
5617 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
5619 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
5620 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
5622 if (slack_runtime <= 0)
5625 raw_spin_lock(&cfs_b->lock);
5626 if (cfs_b->quota != RUNTIME_INF) {
5627 cfs_b->runtime += slack_runtime;
5629 /* we are under rq->lock, defer unthrottling using a timer */
5630 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
5631 !list_empty(&cfs_b->throttled_cfs_rq))
5632 start_cfs_slack_bandwidth(cfs_b);
5634 raw_spin_unlock(&cfs_b->lock);
5636 /* even if it's not valid for return we don't want to try again */
5637 cfs_rq->runtime_remaining -= slack_runtime;
5640 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
5642 if (!cfs_bandwidth_used())
5645 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
5648 __return_cfs_rq_runtime(cfs_rq);
5652 * This is done with a timer (instead of inline with bandwidth return) since
5653 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
5655 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
5657 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
5658 unsigned long flags;
5660 /* confirm we're still not at a refresh boundary */
5661 raw_spin_lock_irqsave(&cfs_b->lock, flags);
5662 cfs_b->slack_started = false;
5664 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
5665 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
5669 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice)
5670 runtime = cfs_b->runtime;
5672 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
5677 distribute_cfs_runtime(cfs_b);
5681 * When a group wakes up we want to make sure that its quota is not already
5682 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
5683 * runtime as update_curr() throttling can not trigger until it's on-rq.
5685 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
5687 if (!cfs_bandwidth_used())
5690 /* an active group must be handled by the update_curr()->put() path */
5691 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
5694 /* ensure the group is not already throttled */
5695 if (cfs_rq_throttled(cfs_rq))
5698 /* update runtime allocation */
5699 account_cfs_rq_runtime(cfs_rq, 0);
5700 if (cfs_rq->runtime_remaining <= 0)
5701 throttle_cfs_rq(cfs_rq);
5704 static void sync_throttle(struct task_group *tg, int cpu)
5706 struct cfs_rq *pcfs_rq, *cfs_rq;
5708 if (!cfs_bandwidth_used())
5714 cfs_rq = tg->cfs_rq[cpu];
5715 pcfs_rq = tg->parent->cfs_rq[cpu];
5717 cfs_rq->throttle_count = pcfs_rq->throttle_count;
5718 cfs_rq->throttled_clock_pelt = rq_clock_pelt(cpu_rq(cpu));
5721 /* conditionally throttle active cfs_rq's from put_prev_entity() */
5722 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
5724 if (!cfs_bandwidth_used())
5727 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
5731 * it's possible for a throttled entity to be forced into a running
5732 * state (e.g. set_curr_task), in this case we're finished.
5734 if (cfs_rq_throttled(cfs_rq))
5737 return throttle_cfs_rq(cfs_rq);
5740 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
5742 struct cfs_bandwidth *cfs_b =
5743 container_of(timer, struct cfs_bandwidth, slack_timer);
5745 do_sched_cfs_slack_timer(cfs_b);
5747 return HRTIMER_NORESTART;
5750 extern const u64 max_cfs_quota_period;
5752 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
5754 struct cfs_bandwidth *cfs_b =
5755 container_of(timer, struct cfs_bandwidth, period_timer);
5756 unsigned long flags;
5761 raw_spin_lock_irqsave(&cfs_b->lock, flags);
5763 overrun = hrtimer_forward_now(timer, cfs_b->period);
5767 idle = do_sched_cfs_period_timer(cfs_b, overrun, flags);
5770 u64 new, old = ktime_to_ns(cfs_b->period);
5773 * Grow period by a factor of 2 to avoid losing precision.
5774 * Precision loss in the quota/period ratio can cause __cfs_schedulable
5778 if (new < max_cfs_quota_period) {
5779 cfs_b->period = ns_to_ktime(new);
5783 pr_warn_ratelimited(
5784 "cfs_period_timer[cpu%d]: period too short, scaling up (new cfs_period_us = %lld, cfs_quota_us = %lld)\n",
5786 div_u64(new, NSEC_PER_USEC),
5787 div_u64(cfs_b->quota, NSEC_PER_USEC));
5789 pr_warn_ratelimited(
5790 "cfs_period_timer[cpu%d]: period too short, but cannot scale up without losing precision (cfs_period_us = %lld, cfs_quota_us = %lld)\n",
5792 div_u64(old, NSEC_PER_USEC),
5793 div_u64(cfs_b->quota, NSEC_PER_USEC));
5796 /* reset count so we don't come right back in here */
5801 cfs_b->period_active = 0;
5802 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
5804 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
5807 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
5809 raw_spin_lock_init(&cfs_b->lock);
5811 cfs_b->quota = RUNTIME_INF;
5812 cfs_b->period = ns_to_ktime(default_cfs_period());
5815 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
5816 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED);
5817 cfs_b->period_timer.function = sched_cfs_period_timer;
5818 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
5819 cfs_b->slack_timer.function = sched_cfs_slack_timer;
5820 cfs_b->slack_started = false;
5823 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
5825 cfs_rq->runtime_enabled = 0;
5826 INIT_LIST_HEAD(&cfs_rq->throttled_list);
5829 void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
5831 lockdep_assert_held(&cfs_b->lock);
5833 if (cfs_b->period_active)
5836 cfs_b->period_active = 1;
5837 hrtimer_forward_now(&cfs_b->period_timer, cfs_b->period);
5838 hrtimer_start_expires(&cfs_b->period_timer, HRTIMER_MODE_ABS_PINNED);
5841 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
5843 /* init_cfs_bandwidth() was not called */
5844 if (!cfs_b->throttled_cfs_rq.next)
5847 hrtimer_cancel(&cfs_b->period_timer);
5848 hrtimer_cancel(&cfs_b->slack_timer);
5852 * Both these CPU hotplug callbacks race against unregister_fair_sched_group()
5854 * The race is harmless, since modifying bandwidth settings of unhooked group
5855 * bits doesn't do much.
5858 /* cpu online callback */
5859 static void __maybe_unused update_runtime_enabled(struct rq *rq)
5861 struct task_group *tg;
5863 lockdep_assert_rq_held(rq);
5866 list_for_each_entry_rcu(tg, &task_groups, list) {
5867 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
5868 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
5870 raw_spin_lock(&cfs_b->lock);
5871 cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF;
5872 raw_spin_unlock(&cfs_b->lock);
5877 /* cpu offline callback */
5878 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
5880 struct task_group *tg;
5882 lockdep_assert_rq_held(rq);
5885 list_for_each_entry_rcu(tg, &task_groups, list) {
5886 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
5888 if (!cfs_rq->runtime_enabled)
5892 * clock_task is not advancing so we just need to make sure
5893 * there's some valid quota amount
5895 cfs_rq->runtime_remaining = 1;
5897 * Offline rq is schedulable till CPU is completely disabled
5898 * in take_cpu_down(), so we prevent new cfs throttling here.
5900 cfs_rq->runtime_enabled = 0;
5902 if (cfs_rq_throttled(cfs_rq))
5903 unthrottle_cfs_rq(cfs_rq);
5908 #else /* CONFIG_CFS_BANDWIDTH */
5910 static inline bool cfs_bandwidth_used(void)
5915 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {}
5916 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; }
5917 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
5918 static inline void sync_throttle(struct task_group *tg, int cpu) {}
5919 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
5921 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
5926 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
5931 static inline int throttled_lb_pair(struct task_group *tg,
5932 int src_cpu, int dest_cpu)
5937 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
5939 #ifdef CONFIG_FAIR_GROUP_SCHED
5940 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
5943 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
5947 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
5948 static inline void update_runtime_enabled(struct rq *rq) {}
5949 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
5951 #endif /* CONFIG_CFS_BANDWIDTH */
5953 /**************************************************
5954 * CFS operations on tasks:
5957 #ifdef CONFIG_SCHED_HRTICK
5958 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
5960 struct sched_entity *se = &p->se;
5961 struct cfs_rq *cfs_rq = cfs_rq_of(se);
5963 SCHED_WARN_ON(task_rq(p) != rq);
5965 if (rq->cfs.h_nr_running > 1) {
5966 u64 slice = sched_slice(cfs_rq, se);
5967 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
5968 s64 delta = slice - ran;
5971 if (task_current(rq, p))
5975 hrtick_start(rq, delta);
5980 * called from enqueue/dequeue and updates the hrtick when the
5981 * current task is from our class and nr_running is low enough
5984 static void hrtick_update(struct rq *rq)
5986 struct task_struct *curr = rq->curr;
5988 if (!hrtick_enabled_fair(rq) || curr->sched_class != &fair_sched_class)
5991 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
5992 hrtick_start_fair(rq, curr);
5994 #else /* !CONFIG_SCHED_HRTICK */
5996 hrtick_start_fair(struct rq *rq, struct task_struct *p)
6000 static inline void hrtick_update(struct rq *rq)
6006 static inline bool cpu_overutilized(int cpu)
6008 unsigned long rq_util_min = uclamp_rq_get(cpu_rq(cpu), UCLAMP_MIN);
6009 unsigned long rq_util_max = uclamp_rq_get(cpu_rq(cpu), UCLAMP_MAX);
6011 return !util_fits_cpu(cpu_util_cfs(cpu), rq_util_min, rq_util_max, cpu);
6014 static inline void update_overutilized_status(struct rq *rq)
6016 if (!READ_ONCE(rq->rd->overutilized) && cpu_overutilized(rq->cpu)) {
6017 WRITE_ONCE(rq->rd->overutilized, SG_OVERUTILIZED);
6018 trace_sched_overutilized_tp(rq->rd, SG_OVERUTILIZED);
6022 static inline void update_overutilized_status(struct rq *rq) { }
6025 /* Runqueue only has SCHED_IDLE tasks enqueued */
6026 static int sched_idle_rq(struct rq *rq)
6028 return unlikely(rq->nr_running == rq->cfs.idle_h_nr_running &&
6033 * Returns true if cfs_rq only has SCHED_IDLE entities enqueued. Note the use
6034 * of idle_nr_running, which does not consider idle descendants of normal
6037 static bool sched_idle_cfs_rq(struct cfs_rq *cfs_rq)
6039 return cfs_rq->nr_running &&
6040 cfs_rq->nr_running == cfs_rq->idle_nr_running;
6044 static int sched_idle_cpu(int cpu)
6046 return sched_idle_rq(cpu_rq(cpu));
6051 * The enqueue_task method is called before nr_running is
6052 * increased. Here we update the fair scheduling stats and
6053 * then put the task into the rbtree:
6056 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
6058 struct cfs_rq *cfs_rq;
6059 struct sched_entity *se = &p->se;
6060 int idle_h_nr_running = task_has_idle_policy(p);
6061 int task_new = !(flags & ENQUEUE_WAKEUP);
6064 * The code below (indirectly) updates schedutil which looks at
6065 * the cfs_rq utilization to select a frequency.
6066 * Let's add the task's estimated utilization to the cfs_rq's
6067 * estimated utilization, before we update schedutil.
6069 util_est_enqueue(&rq->cfs, p);
6072 * If in_iowait is set, the code below may not trigger any cpufreq
6073 * utilization updates, so do it here explicitly with the IOWAIT flag
6077 cpufreq_update_util(rq, SCHED_CPUFREQ_IOWAIT);
6079 for_each_sched_entity(se) {
6082 cfs_rq = cfs_rq_of(se);
6083 enqueue_entity(cfs_rq, se, flags);
6085 cfs_rq->h_nr_running++;
6086 cfs_rq->idle_h_nr_running += idle_h_nr_running;
6088 if (cfs_rq_is_idle(cfs_rq))
6089 idle_h_nr_running = 1;
6091 /* end evaluation on encountering a throttled cfs_rq */
6092 if (cfs_rq_throttled(cfs_rq))
6093 goto enqueue_throttle;
6095 flags = ENQUEUE_WAKEUP;
6098 for_each_sched_entity(se) {
6099 cfs_rq = cfs_rq_of(se);
6101 update_load_avg(cfs_rq, se, UPDATE_TG);
6102 se_update_runnable(se);
6103 update_cfs_group(se);
6105 cfs_rq->h_nr_running++;
6106 cfs_rq->idle_h_nr_running += idle_h_nr_running;
6108 if (cfs_rq_is_idle(cfs_rq))
6109 idle_h_nr_running = 1;
6111 /* end evaluation on encountering a throttled cfs_rq */
6112 if (cfs_rq_throttled(cfs_rq))
6113 goto enqueue_throttle;
6116 /* At this point se is NULL and we are at root level*/
6117 add_nr_running(rq, 1);
6120 * Since new tasks are assigned an initial util_avg equal to
6121 * half of the spare capacity of their CPU, tiny tasks have the
6122 * ability to cross the overutilized threshold, which will
6123 * result in the load balancer ruining all the task placement
6124 * done by EAS. As a way to mitigate that effect, do not account
6125 * for the first enqueue operation of new tasks during the
6126 * overutilized flag detection.
6128 * A better way of solving this problem would be to wait for
6129 * the PELT signals of tasks to converge before taking them
6130 * into account, but that is not straightforward to implement,
6131 * and the following generally works well enough in practice.
6134 update_overutilized_status(rq);
6137 assert_list_leaf_cfs_rq(rq);
6142 static void set_next_buddy(struct sched_entity *se);
6145 * The dequeue_task method is called before nr_running is
6146 * decreased. We remove the task from the rbtree and
6147 * update the fair scheduling stats:
6149 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
6151 struct cfs_rq *cfs_rq;
6152 struct sched_entity *se = &p->se;
6153 int task_sleep = flags & DEQUEUE_SLEEP;
6154 int idle_h_nr_running = task_has_idle_policy(p);
6155 bool was_sched_idle = sched_idle_rq(rq);
6157 util_est_dequeue(&rq->cfs, p);
6159 for_each_sched_entity(se) {
6160 cfs_rq = cfs_rq_of(se);
6161 dequeue_entity(cfs_rq, se, flags);
6163 cfs_rq->h_nr_running--;
6164 cfs_rq->idle_h_nr_running -= idle_h_nr_running;
6166 if (cfs_rq_is_idle(cfs_rq))
6167 idle_h_nr_running = 1;
6169 /* end evaluation on encountering a throttled cfs_rq */
6170 if (cfs_rq_throttled(cfs_rq))
6171 goto dequeue_throttle;
6173 /* Don't dequeue parent if it has other entities besides us */
6174 if (cfs_rq->load.weight) {
6175 /* Avoid re-evaluating load for this entity: */
6176 se = parent_entity(se);
6178 * Bias pick_next to pick a task from this cfs_rq, as
6179 * p is sleeping when it is within its sched_slice.
6181 if (task_sleep && se && !throttled_hierarchy(cfs_rq))
6185 flags |= DEQUEUE_SLEEP;
6188 for_each_sched_entity(se) {
6189 cfs_rq = cfs_rq_of(se);
6191 update_load_avg(cfs_rq, se, UPDATE_TG);
6192 se_update_runnable(se);
6193 update_cfs_group(se);
6195 cfs_rq->h_nr_running--;
6196 cfs_rq->idle_h_nr_running -= idle_h_nr_running;
6198 if (cfs_rq_is_idle(cfs_rq))
6199 idle_h_nr_running = 1;
6201 /* end evaluation on encountering a throttled cfs_rq */
6202 if (cfs_rq_throttled(cfs_rq))
6203 goto dequeue_throttle;
6207 /* At this point se is NULL and we are at root level*/
6208 sub_nr_running(rq, 1);
6210 /* balance early to pull high priority tasks */
6211 if (unlikely(!was_sched_idle && sched_idle_rq(rq)))
6212 rq->next_balance = jiffies;
6215 util_est_update(&rq->cfs, p, task_sleep);
6221 /* Working cpumask for: load_balance, load_balance_newidle. */
6222 static DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
6223 static DEFINE_PER_CPU(cpumask_var_t, select_rq_mask);
6225 #ifdef CONFIG_NO_HZ_COMMON
6228 cpumask_var_t idle_cpus_mask;
6230 int has_blocked; /* Idle CPUS has blocked load */
6231 int needs_update; /* Newly idle CPUs need their next_balance collated */
6232 unsigned long next_balance; /* in jiffy units */
6233 unsigned long next_blocked; /* Next update of blocked load in jiffies */
6234 } nohz ____cacheline_aligned;
6236 #endif /* CONFIG_NO_HZ_COMMON */
6238 static unsigned long cpu_load(struct rq *rq)
6240 return cfs_rq_load_avg(&rq->cfs);
6244 * cpu_load_without - compute CPU load without any contributions from *p
6245 * @cpu: the CPU which load is requested
6246 * @p: the task which load should be discounted
6248 * The load of a CPU is defined by the load of tasks currently enqueued on that
6249 * CPU as well as tasks which are currently sleeping after an execution on that
6252 * This method returns the load of the specified CPU by discounting the load of
6253 * the specified task, whenever the task is currently contributing to the CPU
6256 static unsigned long cpu_load_without(struct rq *rq, struct task_struct *p)
6258 struct cfs_rq *cfs_rq;
6261 /* Task has no contribution or is new */
6262 if (cpu_of(rq) != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time))
6263 return cpu_load(rq);
6266 load = READ_ONCE(cfs_rq->avg.load_avg);
6268 /* Discount task's util from CPU's util */
6269 lsub_positive(&load, task_h_load(p));
6274 static unsigned long cpu_runnable(struct rq *rq)
6276 return cfs_rq_runnable_avg(&rq->cfs);
6279 static unsigned long cpu_runnable_without(struct rq *rq, struct task_struct *p)
6281 struct cfs_rq *cfs_rq;
6282 unsigned int runnable;
6284 /* Task has no contribution or is new */
6285 if (cpu_of(rq) != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time))
6286 return cpu_runnable(rq);
6289 runnable = READ_ONCE(cfs_rq->avg.runnable_avg);
6291 /* Discount task's runnable from CPU's runnable */
6292 lsub_positive(&runnable, p->se.avg.runnable_avg);
6297 static unsigned long capacity_of(int cpu)
6299 return cpu_rq(cpu)->cpu_capacity;
6302 static void record_wakee(struct task_struct *p)
6305 * Only decay a single time; tasks that have less then 1 wakeup per
6306 * jiffy will not have built up many flips.
6308 if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) {
6309 current->wakee_flips >>= 1;
6310 current->wakee_flip_decay_ts = jiffies;
6313 if (current->last_wakee != p) {
6314 current->last_wakee = p;
6315 current->wakee_flips++;
6320 * Detect M:N waker/wakee relationships via a switching-frequency heuristic.
6322 * A waker of many should wake a different task than the one last awakened
6323 * at a frequency roughly N times higher than one of its wakees.
6325 * In order to determine whether we should let the load spread vs consolidating
6326 * to shared cache, we look for a minimum 'flip' frequency of llc_size in one
6327 * partner, and a factor of lls_size higher frequency in the other.
6329 * With both conditions met, we can be relatively sure that the relationship is
6330 * non-monogamous, with partner count exceeding socket size.
6332 * Waker/wakee being client/server, worker/dispatcher, interrupt source or
6333 * whatever is irrelevant, spread criteria is apparent partner count exceeds
6336 static int wake_wide(struct task_struct *p)
6338 unsigned int master = current->wakee_flips;
6339 unsigned int slave = p->wakee_flips;
6340 int factor = __this_cpu_read(sd_llc_size);
6343 swap(master, slave);
6344 if (slave < factor || master < slave * factor)
6350 * The purpose of wake_affine() is to quickly determine on which CPU we can run
6351 * soonest. For the purpose of speed we only consider the waking and previous
6354 * wake_affine_idle() - only considers 'now', it check if the waking CPU is
6355 * cache-affine and is (or will be) idle.
6357 * wake_affine_weight() - considers the weight to reflect the average
6358 * scheduling latency of the CPUs. This seems to work
6359 * for the overloaded case.
6362 wake_affine_idle(int this_cpu, int prev_cpu, int sync)
6365 * If this_cpu is idle, it implies the wakeup is from interrupt
6366 * context. Only allow the move if cache is shared. Otherwise an
6367 * interrupt intensive workload could force all tasks onto one
6368 * node depending on the IO topology or IRQ affinity settings.
6370 * If the prev_cpu is idle and cache affine then avoid a migration.
6371 * There is no guarantee that the cache hot data from an interrupt
6372 * is more important than cache hot data on the prev_cpu and from
6373 * a cpufreq perspective, it's better to have higher utilisation
6376 if (available_idle_cpu(this_cpu) && cpus_share_cache(this_cpu, prev_cpu))
6377 return available_idle_cpu(prev_cpu) ? prev_cpu : this_cpu;
6379 if (sync && cpu_rq(this_cpu)->nr_running == 1)
6382 if (available_idle_cpu(prev_cpu))
6385 return nr_cpumask_bits;
6389 wake_affine_weight(struct sched_domain *sd, struct task_struct *p,
6390 int this_cpu, int prev_cpu, int sync)
6392 s64 this_eff_load, prev_eff_load;
6393 unsigned long task_load;
6395 this_eff_load = cpu_load(cpu_rq(this_cpu));
6398 unsigned long current_load = task_h_load(current);
6400 if (current_load > this_eff_load)
6403 this_eff_load -= current_load;
6406 task_load = task_h_load(p);
6408 this_eff_load += task_load;
6409 if (sched_feat(WA_BIAS))
6410 this_eff_load *= 100;
6411 this_eff_load *= capacity_of(prev_cpu);
6413 prev_eff_load = cpu_load(cpu_rq(prev_cpu));
6414 prev_eff_load -= task_load;
6415 if (sched_feat(WA_BIAS))
6416 prev_eff_load *= 100 + (sd->imbalance_pct - 100) / 2;
6417 prev_eff_load *= capacity_of(this_cpu);
6420 * If sync, adjust the weight of prev_eff_load such that if
6421 * prev_eff == this_eff that select_idle_sibling() will consider
6422 * stacking the wakee on top of the waker if no other CPU is
6428 return this_eff_load < prev_eff_load ? this_cpu : nr_cpumask_bits;
6431 static int wake_affine(struct sched_domain *sd, struct task_struct *p,
6432 int this_cpu, int prev_cpu, int sync)
6434 int target = nr_cpumask_bits;
6436 if (sched_feat(WA_IDLE))
6437 target = wake_affine_idle(this_cpu, prev_cpu, sync);
6439 if (sched_feat(WA_WEIGHT) && target == nr_cpumask_bits)
6440 target = wake_affine_weight(sd, p, this_cpu, prev_cpu, sync);
6442 schedstat_inc(p->stats.nr_wakeups_affine_attempts);
6443 if (target == nr_cpumask_bits)
6446 schedstat_inc(sd->ttwu_move_affine);
6447 schedstat_inc(p->stats.nr_wakeups_affine);
6451 static struct sched_group *
6452 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu);
6455 * find_idlest_group_cpu - find the idlest CPU among the CPUs in the group.
6458 find_idlest_group_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
6460 unsigned long load, min_load = ULONG_MAX;
6461 unsigned int min_exit_latency = UINT_MAX;
6462 u64 latest_idle_timestamp = 0;
6463 int least_loaded_cpu = this_cpu;
6464 int shallowest_idle_cpu = -1;
6467 /* Check if we have any choice: */
6468 if (group->group_weight == 1)
6469 return cpumask_first(sched_group_span(group));
6471 /* Traverse only the allowed CPUs */
6472 for_each_cpu_and(i, sched_group_span(group), p->cpus_ptr) {
6473 struct rq *rq = cpu_rq(i);
6475 if (!sched_core_cookie_match(rq, p))
6478 if (sched_idle_cpu(i))
6481 if (available_idle_cpu(i)) {
6482 struct cpuidle_state *idle = idle_get_state(rq);
6483 if (idle && idle->exit_latency < min_exit_latency) {
6485 * We give priority to a CPU whose idle state
6486 * has the smallest exit latency irrespective
6487 * of any idle timestamp.
6489 min_exit_latency = idle->exit_latency;
6490 latest_idle_timestamp = rq->idle_stamp;
6491 shallowest_idle_cpu = i;
6492 } else if ((!idle || idle->exit_latency == min_exit_latency) &&
6493 rq->idle_stamp > latest_idle_timestamp) {
6495 * If equal or no active idle state, then
6496 * the most recently idled CPU might have
6499 latest_idle_timestamp = rq->idle_stamp;
6500 shallowest_idle_cpu = i;
6502 } else if (shallowest_idle_cpu == -1) {
6503 load = cpu_load(cpu_rq(i));
6504 if (load < min_load) {
6506 least_loaded_cpu = i;
6511 return shallowest_idle_cpu != -1 ? shallowest_idle_cpu : least_loaded_cpu;
6514 static inline int find_idlest_cpu(struct sched_domain *sd, struct task_struct *p,
6515 int cpu, int prev_cpu, int sd_flag)
6519 if (!cpumask_intersects(sched_domain_span(sd), p->cpus_ptr))
6523 * We need task's util for cpu_util_without, sync it up to
6524 * prev_cpu's last_update_time.
6526 if (!(sd_flag & SD_BALANCE_FORK))
6527 sync_entity_load_avg(&p->se);
6530 struct sched_group *group;
6531 struct sched_domain *tmp;
6534 if (!(sd->flags & sd_flag)) {
6539 group = find_idlest_group(sd, p, cpu);
6545 new_cpu = find_idlest_group_cpu(group, p, cpu);
6546 if (new_cpu == cpu) {
6547 /* Now try balancing at a lower domain level of 'cpu': */
6552 /* Now try balancing at a lower domain level of 'new_cpu': */
6554 weight = sd->span_weight;
6556 for_each_domain(cpu, tmp) {
6557 if (weight <= tmp->span_weight)
6559 if (tmp->flags & sd_flag)
6567 static inline int __select_idle_cpu(int cpu, struct task_struct *p)
6569 if ((available_idle_cpu(cpu) || sched_idle_cpu(cpu)) &&
6570 sched_cpu_cookie_match(cpu_rq(cpu), p))
6576 #ifdef CONFIG_SCHED_SMT
6577 DEFINE_STATIC_KEY_FALSE(sched_smt_present);
6578 EXPORT_SYMBOL_GPL(sched_smt_present);
6580 static inline void set_idle_cores(int cpu, int val)
6582 struct sched_domain_shared *sds;
6584 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
6586 WRITE_ONCE(sds->has_idle_cores, val);
6589 static inline bool test_idle_cores(int cpu)
6591 struct sched_domain_shared *sds;
6593 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
6595 return READ_ONCE(sds->has_idle_cores);
6601 * Scans the local SMT mask to see if the entire core is idle, and records this
6602 * information in sd_llc_shared->has_idle_cores.
6604 * Since SMT siblings share all cache levels, inspecting this limited remote
6605 * state should be fairly cheap.
6607 void __update_idle_core(struct rq *rq)
6609 int core = cpu_of(rq);
6613 if (test_idle_cores(core))
6616 for_each_cpu(cpu, cpu_smt_mask(core)) {
6620 if (!available_idle_cpu(cpu))
6624 set_idle_cores(core, 1);
6630 * Scan the entire LLC domain for idle cores; this dynamically switches off if
6631 * there are no idle cores left in the system; tracked through
6632 * sd_llc->shared->has_idle_cores and enabled through update_idle_core() above.
6634 static int select_idle_core(struct task_struct *p, int core, struct cpumask *cpus, int *idle_cpu)
6639 for_each_cpu(cpu, cpu_smt_mask(core)) {
6640 if (!available_idle_cpu(cpu)) {
6642 if (*idle_cpu == -1) {
6643 if (sched_idle_cpu(cpu) && cpumask_test_cpu(cpu, p->cpus_ptr)) {
6651 if (*idle_cpu == -1 && cpumask_test_cpu(cpu, p->cpus_ptr))
6658 cpumask_andnot(cpus, cpus, cpu_smt_mask(core));
6663 * Scan the local SMT mask for idle CPUs.
6665 static int select_idle_smt(struct task_struct *p, int target)
6669 for_each_cpu_and(cpu, cpu_smt_mask(target), p->cpus_ptr) {
6672 if (available_idle_cpu(cpu) || sched_idle_cpu(cpu))
6679 #else /* CONFIG_SCHED_SMT */
6681 static inline void set_idle_cores(int cpu, int val)
6685 static inline bool test_idle_cores(int cpu)
6690 static inline int select_idle_core(struct task_struct *p, int core, struct cpumask *cpus, int *idle_cpu)
6692 return __select_idle_cpu(core, p);
6695 static inline int select_idle_smt(struct task_struct *p, int target)
6700 #endif /* CONFIG_SCHED_SMT */
6703 * Scan the LLC domain for idle CPUs; this is dynamically regulated by
6704 * comparing the average scan cost (tracked in sd->avg_scan_cost) against the
6705 * average idle time for this rq (as found in rq->avg_idle).
6707 static int select_idle_cpu(struct task_struct *p, struct sched_domain *sd, bool has_idle_core, int target)
6709 struct cpumask *cpus = this_cpu_cpumask_var_ptr(select_rq_mask);
6710 int i, cpu, idle_cpu = -1, nr = INT_MAX;
6711 struct sched_domain_shared *sd_share;
6712 struct rq *this_rq = this_rq();
6713 int this = smp_processor_id();
6714 struct sched_domain *this_sd = NULL;
6717 cpumask_and(cpus, sched_domain_span(sd), p->cpus_ptr);
6719 if (sched_feat(SIS_PROP) && !has_idle_core) {
6720 u64 avg_cost, avg_idle, span_avg;
6721 unsigned long now = jiffies;
6723 this_sd = rcu_dereference(*this_cpu_ptr(&sd_llc));
6728 * If we're busy, the assumption that the last idle period
6729 * predicts the future is flawed; age away the remaining
6730 * predicted idle time.
6732 if (unlikely(this_rq->wake_stamp < now)) {
6733 while (this_rq->wake_stamp < now && this_rq->wake_avg_idle) {
6734 this_rq->wake_stamp++;
6735 this_rq->wake_avg_idle >>= 1;
6739 avg_idle = this_rq->wake_avg_idle;
6740 avg_cost = this_sd->avg_scan_cost + 1;
6742 span_avg = sd->span_weight * avg_idle;
6743 if (span_avg > 4*avg_cost)
6744 nr = div_u64(span_avg, avg_cost);
6748 time = cpu_clock(this);
6751 if (sched_feat(SIS_UTIL)) {
6752 sd_share = rcu_dereference(per_cpu(sd_llc_shared, target));
6754 /* because !--nr is the condition to stop scan */
6755 nr = READ_ONCE(sd_share->nr_idle_scan) + 1;
6756 /* overloaded LLC is unlikely to have idle cpu/core */
6762 for_each_cpu_wrap(cpu, cpus, target + 1) {
6763 if (has_idle_core) {
6764 i = select_idle_core(p, cpu, cpus, &idle_cpu);
6765 if ((unsigned int)i < nr_cpumask_bits)
6771 idle_cpu = __select_idle_cpu(cpu, p);
6772 if ((unsigned int)idle_cpu < nr_cpumask_bits)
6778 set_idle_cores(target, false);
6780 if (sched_feat(SIS_PROP) && this_sd && !has_idle_core) {
6781 time = cpu_clock(this) - time;
6784 * Account for the scan cost of wakeups against the average
6787 this_rq->wake_avg_idle -= min(this_rq->wake_avg_idle, time);
6789 update_avg(&this_sd->avg_scan_cost, time);
6796 * Scan the asym_capacity domain for idle CPUs; pick the first idle one on which
6797 * the task fits. If no CPU is big enough, but there are idle ones, try to
6798 * maximize capacity.
6801 select_idle_capacity(struct task_struct *p, struct sched_domain *sd, int target)
6803 unsigned long task_util, util_min, util_max, best_cap = 0;
6804 int cpu, best_cpu = -1;
6805 struct cpumask *cpus;
6807 cpus = this_cpu_cpumask_var_ptr(select_rq_mask);
6808 cpumask_and(cpus, sched_domain_span(sd), p->cpus_ptr);
6810 task_util = task_util_est(p);
6811 util_min = uclamp_eff_value(p, UCLAMP_MIN);
6812 util_max = uclamp_eff_value(p, UCLAMP_MAX);
6814 for_each_cpu_wrap(cpu, cpus, target) {
6815 unsigned long cpu_cap = capacity_of(cpu);
6817 if (!available_idle_cpu(cpu) && !sched_idle_cpu(cpu))
6819 if (util_fits_cpu(task_util, util_min, util_max, cpu))
6822 if (cpu_cap > best_cap) {
6831 static inline bool asym_fits_cpu(unsigned long util,
6832 unsigned long util_min,
6833 unsigned long util_max,
6836 if (sched_asym_cpucap_active())
6837 return util_fits_cpu(util, util_min, util_max, cpu);
6843 * Try and locate an idle core/thread in the LLC cache domain.
6845 static int select_idle_sibling(struct task_struct *p, int prev, int target)
6847 bool has_idle_core = false;
6848 struct sched_domain *sd;
6849 unsigned long task_util, util_min, util_max;
6850 int i, recent_used_cpu;
6853 * On asymmetric system, update task utilization because we will check
6854 * that the task fits with cpu's capacity.
6856 if (sched_asym_cpucap_active()) {
6857 sync_entity_load_avg(&p->se);
6858 task_util = task_util_est(p);
6859 util_min = uclamp_eff_value(p, UCLAMP_MIN);
6860 util_max = uclamp_eff_value(p, UCLAMP_MAX);
6864 * per-cpu select_rq_mask usage
6866 lockdep_assert_irqs_disabled();
6868 if ((available_idle_cpu(target) || sched_idle_cpu(target)) &&
6869 asym_fits_cpu(task_util, util_min, util_max, target))
6873 * If the previous CPU is cache affine and idle, don't be stupid:
6875 if (prev != target && cpus_share_cache(prev, target) &&
6876 (available_idle_cpu(prev) || sched_idle_cpu(prev)) &&
6877 asym_fits_cpu(task_util, util_min, util_max, prev))
6881 * Allow a per-cpu kthread to stack with the wakee if the
6882 * kworker thread and the tasks previous CPUs are the same.
6883 * The assumption is that the wakee queued work for the
6884 * per-cpu kthread that is now complete and the wakeup is
6885 * essentially a sync wakeup. An obvious example of this
6886 * pattern is IO completions.
6888 if (is_per_cpu_kthread(current) &&
6890 prev == smp_processor_id() &&
6891 this_rq()->nr_running <= 1 &&
6892 asym_fits_cpu(task_util, util_min, util_max, prev)) {
6896 /* Check a recently used CPU as a potential idle candidate: */
6897 recent_used_cpu = p->recent_used_cpu;
6898 p->recent_used_cpu = prev;
6899 if (recent_used_cpu != prev &&
6900 recent_used_cpu != target &&
6901 cpus_share_cache(recent_used_cpu, target) &&
6902 (available_idle_cpu(recent_used_cpu) || sched_idle_cpu(recent_used_cpu)) &&
6903 cpumask_test_cpu(p->recent_used_cpu, p->cpus_ptr) &&
6904 asym_fits_cpu(task_util, util_min, util_max, recent_used_cpu)) {
6905 return recent_used_cpu;
6909 * For asymmetric CPU capacity systems, our domain of interest is
6910 * sd_asym_cpucapacity rather than sd_llc.
6912 if (sched_asym_cpucap_active()) {
6913 sd = rcu_dereference(per_cpu(sd_asym_cpucapacity, target));
6915 * On an asymmetric CPU capacity system where an exclusive
6916 * cpuset defines a symmetric island (i.e. one unique
6917 * capacity_orig value through the cpuset), the key will be set
6918 * but the CPUs within that cpuset will not have a domain with
6919 * SD_ASYM_CPUCAPACITY. These should follow the usual symmetric
6923 i = select_idle_capacity(p, sd, target);
6924 return ((unsigned)i < nr_cpumask_bits) ? i : target;
6928 sd = rcu_dereference(per_cpu(sd_llc, target));
6932 if (sched_smt_active()) {
6933 has_idle_core = test_idle_cores(target);
6935 if (!has_idle_core && cpus_share_cache(prev, target)) {
6936 i = select_idle_smt(p, prev);
6937 if ((unsigned int)i < nr_cpumask_bits)
6942 i = select_idle_cpu(p, sd, has_idle_core, target);
6943 if ((unsigned)i < nr_cpumask_bits)
6950 * Predicts what cpu_util(@cpu) would return if @p was removed from @cpu
6951 * (@dst_cpu = -1) or migrated to @dst_cpu.
6953 static unsigned long cpu_util_next(int cpu, struct task_struct *p, int dst_cpu)
6955 struct cfs_rq *cfs_rq = &cpu_rq(cpu)->cfs;
6956 unsigned long util = READ_ONCE(cfs_rq->avg.util_avg);
6959 * If @dst_cpu is -1 or @p migrates from @cpu to @dst_cpu remove its
6960 * contribution. If @p migrates from another CPU to @cpu add its
6961 * contribution. In all the other cases @cpu is not impacted by the
6962 * migration so its util_avg is already correct.
6964 if (task_cpu(p) == cpu && dst_cpu != cpu)
6965 lsub_positive(&util, task_util(p));
6966 else if (task_cpu(p) != cpu && dst_cpu == cpu)
6967 util += task_util(p);
6969 if (sched_feat(UTIL_EST)) {
6970 unsigned long util_est;
6972 util_est = READ_ONCE(cfs_rq->avg.util_est.enqueued);
6975 * During wake-up @p isn't enqueued yet and doesn't contribute
6976 * to any cpu_rq(cpu)->cfs.avg.util_est.enqueued.
6977 * If @dst_cpu == @cpu add it to "simulate" cpu_util after @p
6978 * has been enqueued.
6980 * During exec (@dst_cpu = -1) @p is enqueued and does
6981 * contribute to cpu_rq(cpu)->cfs.util_est.enqueued.
6982 * Remove it to "simulate" cpu_util without @p's contribution.
6984 * Despite the task_on_rq_queued(@p) check there is still a
6985 * small window for a possible race when an exec
6986 * select_task_rq_fair() races with LB's detach_task().
6990 * p->on_rq = TASK_ON_RQ_MIGRATING;
6991 * -------------------------------- A
6993 * dequeue_task_fair() + Race Time
6994 * util_est_dequeue() /
6995 * -------------------------------- B
6997 * The additional check "current == p" is required to further
6998 * reduce the race window.
7001 util_est += _task_util_est(p);
7002 else if (unlikely(task_on_rq_queued(p) || current == p))
7003 lsub_positive(&util_est, _task_util_est(p));
7005 util = max(util, util_est);
7008 return min(util, capacity_orig_of(cpu));
7012 * cpu_util_without: compute cpu utilization without any contributions from *p
7013 * @cpu: the CPU which utilization is requested
7014 * @p: the task which utilization should be discounted
7016 * The utilization of a CPU is defined by the utilization of tasks currently
7017 * enqueued on that CPU as well as tasks which are currently sleeping after an
7018 * execution on that CPU.
7020 * This method returns the utilization of the specified CPU by discounting the
7021 * utilization of the specified task, whenever the task is currently
7022 * contributing to the CPU utilization.
7024 static unsigned long cpu_util_without(int cpu, struct task_struct *p)
7026 /* Task has no contribution or is new */
7027 if (cpu != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time))
7028 return cpu_util_cfs(cpu);
7030 return cpu_util_next(cpu, p, -1);
7034 * energy_env - Utilization landscape for energy estimation.
7035 * @task_busy_time: Utilization contribution by the task for which we test the
7036 * placement. Given by eenv_task_busy_time().
7037 * @pd_busy_time: Utilization of the whole perf domain without the task
7038 * contribution. Given by eenv_pd_busy_time().
7039 * @cpu_cap: Maximum CPU capacity for the perf domain.
7040 * @pd_cap: Entire perf domain capacity. (pd->nr_cpus * cpu_cap).
7043 unsigned long task_busy_time;
7044 unsigned long pd_busy_time;
7045 unsigned long cpu_cap;
7046 unsigned long pd_cap;
7050 * Compute the task busy time for compute_energy(). This time cannot be
7051 * injected directly into effective_cpu_util() because of the IRQ scaling.
7052 * The latter only makes sense with the most recent CPUs where the task has
7055 static inline void eenv_task_busy_time(struct energy_env *eenv,
7056 struct task_struct *p, int prev_cpu)
7058 unsigned long busy_time, max_cap = arch_scale_cpu_capacity(prev_cpu);
7059 unsigned long irq = cpu_util_irq(cpu_rq(prev_cpu));
7061 if (unlikely(irq >= max_cap))
7062 busy_time = max_cap;
7064 busy_time = scale_irq_capacity(task_util_est(p), irq, max_cap);
7066 eenv->task_busy_time = busy_time;
7070 * Compute the perf_domain (PD) busy time for compute_energy(). Based on the
7071 * utilization for each @pd_cpus, it however doesn't take into account
7072 * clamping since the ratio (utilization / cpu_capacity) is already enough to
7073 * scale the EM reported power consumption at the (eventually clamped)
7076 * The contribution of the task @p for which we want to estimate the
7077 * energy cost is removed (by cpu_util_next()) and must be calculated
7078 * separately (see eenv_task_busy_time). This ensures:
7080 * - A stable PD utilization, no matter which CPU of that PD we want to place
7083 * - A fair comparison between CPUs as the task contribution (task_util())
7084 * will always be the same no matter which CPU utilization we rely on
7085 * (util_avg or util_est).
7087 * Set @eenv busy time for the PD that spans @pd_cpus. This busy time can't
7088 * exceed @eenv->pd_cap.
7090 static inline void eenv_pd_busy_time(struct energy_env *eenv,
7091 struct cpumask *pd_cpus,
7092 struct task_struct *p)
7094 unsigned long busy_time = 0;
7097 for_each_cpu(cpu, pd_cpus) {
7098 unsigned long util = cpu_util_next(cpu, p, -1);
7100 busy_time += effective_cpu_util(cpu, util, ENERGY_UTIL, NULL);
7103 eenv->pd_busy_time = min(eenv->pd_cap, busy_time);
7107 * Compute the maximum utilization for compute_energy() when the task @p
7108 * is placed on the cpu @dst_cpu.
7110 * Returns the maximum utilization among @eenv->cpus. This utilization can't
7111 * exceed @eenv->cpu_cap.
7113 static inline unsigned long
7114 eenv_pd_max_util(struct energy_env *eenv, struct cpumask *pd_cpus,
7115 struct task_struct *p, int dst_cpu)
7117 unsigned long max_util = 0;
7120 for_each_cpu(cpu, pd_cpus) {
7121 struct task_struct *tsk = (cpu == dst_cpu) ? p : NULL;
7122 unsigned long util = cpu_util_next(cpu, p, dst_cpu);
7123 unsigned long cpu_util;
7126 * Performance domain frequency: utilization clamping
7127 * must be considered since it affects the selection
7128 * of the performance domain frequency.
7129 * NOTE: in case RT tasks are running, by default the
7130 * FREQUENCY_UTIL's utilization can be max OPP.
7132 cpu_util = effective_cpu_util(cpu, util, FREQUENCY_UTIL, tsk);
7133 max_util = max(max_util, cpu_util);
7136 return min(max_util, eenv->cpu_cap);
7140 * compute_energy(): Use the Energy Model to estimate the energy that @pd would
7141 * consume for a given utilization landscape @eenv. When @dst_cpu < 0, the task
7142 * contribution is ignored.
7144 static inline unsigned long
7145 compute_energy(struct energy_env *eenv, struct perf_domain *pd,
7146 struct cpumask *pd_cpus, struct task_struct *p, int dst_cpu)
7148 unsigned long max_util = eenv_pd_max_util(eenv, pd_cpus, p, dst_cpu);
7149 unsigned long busy_time = eenv->pd_busy_time;
7152 busy_time = min(eenv->pd_cap, busy_time + eenv->task_busy_time);
7154 return em_cpu_energy(pd->em_pd, max_util, busy_time, eenv->cpu_cap);
7158 * find_energy_efficient_cpu(): Find most energy-efficient target CPU for the
7159 * waking task. find_energy_efficient_cpu() looks for the CPU with maximum
7160 * spare capacity in each performance domain and uses it as a potential
7161 * candidate to execute the task. Then, it uses the Energy Model to figure
7162 * out which of the CPU candidates is the most energy-efficient.
7164 * The rationale for this heuristic is as follows. In a performance domain,
7165 * all the most energy efficient CPU candidates (according to the Energy
7166 * Model) are those for which we'll request a low frequency. When there are
7167 * several CPUs for which the frequency request will be the same, we don't
7168 * have enough data to break the tie between them, because the Energy Model
7169 * only includes active power costs. With this model, if we assume that
7170 * frequency requests follow utilization (e.g. using schedutil), the CPU with
7171 * the maximum spare capacity in a performance domain is guaranteed to be among
7172 * the best candidates of the performance domain.
7174 * In practice, it could be preferable from an energy standpoint to pack
7175 * small tasks on a CPU in order to let other CPUs go in deeper idle states,
7176 * but that could also hurt our chances to go cluster idle, and we have no
7177 * ways to tell with the current Energy Model if this is actually a good
7178 * idea or not. So, find_energy_efficient_cpu() basically favors
7179 * cluster-packing, and spreading inside a cluster. That should at least be
7180 * a good thing for latency, and this is consistent with the idea that most
7181 * of the energy savings of EAS come from the asymmetry of the system, and
7182 * not so much from breaking the tie between identical CPUs. That's also the
7183 * reason why EAS is enabled in the topology code only for systems where
7184 * SD_ASYM_CPUCAPACITY is set.
7186 * NOTE: Forkees are not accepted in the energy-aware wake-up path because
7187 * they don't have any useful utilization data yet and it's not possible to
7188 * forecast their impact on energy consumption. Consequently, they will be
7189 * placed by find_idlest_cpu() on the least loaded CPU, which might turn out
7190 * to be energy-inefficient in some use-cases. The alternative would be to
7191 * bias new tasks towards specific types of CPUs first, or to try to infer
7192 * their util_avg from the parent task, but those heuristics could hurt
7193 * other use-cases too. So, until someone finds a better way to solve this,
7194 * let's keep things simple by re-using the existing slow path.
7196 static int find_energy_efficient_cpu(struct task_struct *p, int prev_cpu)
7198 struct cpumask *cpus = this_cpu_cpumask_var_ptr(select_rq_mask);
7199 unsigned long prev_delta = ULONG_MAX, best_delta = ULONG_MAX;
7200 unsigned long p_util_min = uclamp_is_used() ? uclamp_eff_value(p, UCLAMP_MIN) : 0;
7201 unsigned long p_util_max = uclamp_is_used() ? uclamp_eff_value(p, UCLAMP_MAX) : 1024;
7202 struct root_domain *rd = this_rq()->rd;
7203 int cpu, best_energy_cpu, target = -1;
7204 struct sched_domain *sd;
7205 struct perf_domain *pd;
7206 struct energy_env eenv;
7209 pd = rcu_dereference(rd->pd);
7210 if (!pd || READ_ONCE(rd->overutilized))
7214 * Energy-aware wake-up happens on the lowest sched_domain starting
7215 * from sd_asym_cpucapacity spanning over this_cpu and prev_cpu.
7217 sd = rcu_dereference(*this_cpu_ptr(&sd_asym_cpucapacity));
7218 while (sd && !cpumask_test_cpu(prev_cpu, sched_domain_span(sd)))
7225 sync_entity_load_avg(&p->se);
7226 if (!uclamp_task_util(p, p_util_min, p_util_max))
7229 eenv_task_busy_time(&eenv, p, prev_cpu);
7231 for (; pd; pd = pd->next) {
7232 unsigned long cpu_cap, cpu_thermal_cap, util;
7233 unsigned long cur_delta, max_spare_cap = 0;
7234 unsigned long rq_util_min, rq_util_max;
7235 unsigned long util_min, util_max;
7236 unsigned long prev_spare_cap = 0;
7237 int max_spare_cap_cpu = -1;
7238 unsigned long base_energy;
7240 cpumask_and(cpus, perf_domain_span(pd), cpu_online_mask);
7242 if (cpumask_empty(cpus))
7245 /* Account thermal pressure for the energy estimation */
7246 cpu = cpumask_first(cpus);
7247 cpu_thermal_cap = arch_scale_cpu_capacity(cpu);
7248 cpu_thermal_cap -= arch_scale_thermal_pressure(cpu);
7250 eenv.cpu_cap = cpu_thermal_cap;
7253 for_each_cpu(cpu, cpus) {
7254 eenv.pd_cap += cpu_thermal_cap;
7256 if (!cpumask_test_cpu(cpu, sched_domain_span(sd)))
7259 if (!cpumask_test_cpu(cpu, p->cpus_ptr))
7262 util = cpu_util_next(cpu, p, cpu);
7263 cpu_cap = capacity_of(cpu);
7266 * Skip CPUs that cannot satisfy the capacity request.
7267 * IOW, placing the task there would make the CPU
7268 * overutilized. Take uclamp into account to see how
7269 * much capacity we can get out of the CPU; this is
7270 * aligned with sched_cpu_util().
7272 if (uclamp_is_used()) {
7273 if (uclamp_rq_is_idle(cpu_rq(cpu))) {
7274 util_min = p_util_min;
7275 util_max = p_util_max;
7278 * Open code uclamp_rq_util_with() except for
7279 * the clamp() part. Ie: apply max aggregation
7280 * only. util_fits_cpu() logic requires to
7281 * operate on non clamped util but must use the
7282 * max-aggregated uclamp_{min, max}.
7284 rq_util_min = uclamp_rq_get(cpu_rq(cpu), UCLAMP_MIN);
7285 rq_util_max = uclamp_rq_get(cpu_rq(cpu), UCLAMP_MAX);
7287 util_min = max(rq_util_min, p_util_min);
7288 util_max = max(rq_util_max, p_util_max);
7291 if (!util_fits_cpu(util, util_min, util_max, cpu))
7294 lsub_positive(&cpu_cap, util);
7296 if (cpu == prev_cpu) {
7297 /* Always use prev_cpu as a candidate. */
7298 prev_spare_cap = cpu_cap;
7299 } else if (cpu_cap > max_spare_cap) {
7301 * Find the CPU with the maximum spare capacity
7302 * among the remaining CPUs in the performance
7305 max_spare_cap = cpu_cap;
7306 max_spare_cap_cpu = cpu;
7310 if (max_spare_cap_cpu < 0 && prev_spare_cap == 0)
7313 eenv_pd_busy_time(&eenv, cpus, p);
7314 /* Compute the 'base' energy of the pd, without @p */
7315 base_energy = compute_energy(&eenv, pd, cpus, p, -1);
7317 /* Evaluate the energy impact of using prev_cpu. */
7318 if (prev_spare_cap > 0) {
7319 prev_delta = compute_energy(&eenv, pd, cpus, p,
7321 /* CPU utilization has changed */
7322 if (prev_delta < base_energy)
7324 prev_delta -= base_energy;
7325 best_delta = min(best_delta, prev_delta);
7328 /* Evaluate the energy impact of using max_spare_cap_cpu. */
7329 if (max_spare_cap_cpu >= 0 && max_spare_cap > prev_spare_cap) {
7330 cur_delta = compute_energy(&eenv, pd, cpus, p,
7332 /* CPU utilization has changed */
7333 if (cur_delta < base_energy)
7335 cur_delta -= base_energy;
7336 if (cur_delta < best_delta) {
7337 best_delta = cur_delta;
7338 best_energy_cpu = max_spare_cap_cpu;
7344 if (best_delta < prev_delta)
7345 target = best_energy_cpu;
7356 * select_task_rq_fair: Select target runqueue for the waking task in domains
7357 * that have the relevant SD flag set. In practice, this is SD_BALANCE_WAKE,
7358 * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
7360 * Balances load by selecting the idlest CPU in the idlest group, or under
7361 * certain conditions an idle sibling CPU if the domain has SD_WAKE_AFFINE set.
7363 * Returns the target CPU number.
7366 select_task_rq_fair(struct task_struct *p, int prev_cpu, int wake_flags)
7368 int sync = (wake_flags & WF_SYNC) && !(current->flags & PF_EXITING);
7369 struct sched_domain *tmp, *sd = NULL;
7370 int cpu = smp_processor_id();
7371 int new_cpu = prev_cpu;
7372 int want_affine = 0;
7373 /* SD_flags and WF_flags share the first nibble */
7374 int sd_flag = wake_flags & 0xF;
7377 * required for stable ->cpus_allowed
7379 lockdep_assert_held(&p->pi_lock);
7380 if (wake_flags & WF_TTWU) {
7383 if (sched_energy_enabled()) {
7384 new_cpu = find_energy_efficient_cpu(p, prev_cpu);
7390 want_affine = !wake_wide(p) && cpumask_test_cpu(cpu, p->cpus_ptr);
7394 for_each_domain(cpu, tmp) {
7396 * If both 'cpu' and 'prev_cpu' are part of this domain,
7397 * cpu is a valid SD_WAKE_AFFINE target.
7399 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
7400 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
7401 if (cpu != prev_cpu)
7402 new_cpu = wake_affine(tmp, p, cpu, prev_cpu, sync);
7404 sd = NULL; /* Prefer wake_affine over balance flags */
7409 * Usually only true for WF_EXEC and WF_FORK, as sched_domains
7410 * usually do not have SD_BALANCE_WAKE set. That means wakeup
7411 * will usually go to the fast path.
7413 if (tmp->flags & sd_flag)
7415 else if (!want_affine)
7421 new_cpu = find_idlest_cpu(sd, p, cpu, prev_cpu, sd_flag);
7422 } else if (wake_flags & WF_TTWU) { /* XXX always ? */
7424 new_cpu = select_idle_sibling(p, prev_cpu, new_cpu);
7432 * Called immediately before a task is migrated to a new CPU; task_cpu(p) and
7433 * cfs_rq_of(p) references at time of call are still valid and identify the
7434 * previous CPU. The caller guarantees p->pi_lock or task_rq(p)->lock is held.
7436 static void migrate_task_rq_fair(struct task_struct *p, int new_cpu)
7438 struct sched_entity *se = &p->se;
7441 * As blocked tasks retain absolute vruntime the migration needs to
7442 * deal with this by subtracting the old and adding the new
7443 * min_vruntime -- the latter is done by enqueue_entity() when placing
7444 * the task on the new runqueue.
7446 if (READ_ONCE(p->__state) == TASK_WAKING) {
7447 struct cfs_rq *cfs_rq = cfs_rq_of(se);
7449 se->vruntime -= u64_u32_load(cfs_rq->min_vruntime);
7452 if (!task_on_rq_migrating(p)) {
7453 remove_entity_load_avg(se);
7456 * Here, the task's PELT values have been updated according to
7457 * the current rq's clock. But if that clock hasn't been
7458 * updated in a while, a substantial idle time will be missed,
7459 * leading to an inflation after wake-up on the new rq.
7461 * Estimate the missing time from the cfs_rq last_update_time
7462 * and update sched_avg to improve the PELT continuity after
7465 migrate_se_pelt_lag(se);
7468 /* Tell new CPU we are migrated */
7469 se->avg.last_update_time = 0;
7471 /* We have migrated, no longer consider this task hot */
7474 update_scan_period(p, new_cpu);
7477 static void task_dead_fair(struct task_struct *p)
7479 remove_entity_load_avg(&p->se);
7483 balance_fair(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
7488 return newidle_balance(rq, rf) != 0;
7490 #endif /* CONFIG_SMP */
7492 static unsigned long wakeup_gran(struct sched_entity *se)
7494 unsigned long gran = sysctl_sched_wakeup_granularity;
7497 * Since its curr running now, convert the gran from real-time
7498 * to virtual-time in his units.
7500 * By using 'se' instead of 'curr' we penalize light tasks, so
7501 * they get preempted easier. That is, if 'se' < 'curr' then
7502 * the resulting gran will be larger, therefore penalizing the
7503 * lighter, if otoh 'se' > 'curr' then the resulting gran will
7504 * be smaller, again penalizing the lighter task.
7506 * This is especially important for buddies when the leftmost
7507 * task is higher priority than the buddy.
7509 return calc_delta_fair(gran, se);
7513 * Should 'se' preempt 'curr'.
7527 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
7529 s64 gran, vdiff = curr->vruntime - se->vruntime;
7534 gran = wakeup_gran(se);
7541 static void set_last_buddy(struct sched_entity *se)
7543 for_each_sched_entity(se) {
7544 if (SCHED_WARN_ON(!se->on_rq))
7548 cfs_rq_of(se)->last = se;
7552 static void set_next_buddy(struct sched_entity *se)
7554 for_each_sched_entity(se) {
7555 if (SCHED_WARN_ON(!se->on_rq))
7559 cfs_rq_of(se)->next = se;
7563 static void set_skip_buddy(struct sched_entity *se)
7565 for_each_sched_entity(se)
7566 cfs_rq_of(se)->skip = se;
7570 * Preempt the current task with a newly woken task if needed:
7572 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
7574 struct task_struct *curr = rq->curr;
7575 struct sched_entity *se = &curr->se, *pse = &p->se;
7576 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
7577 int scale = cfs_rq->nr_running >= sched_nr_latency;
7578 int next_buddy_marked = 0;
7579 int cse_is_idle, pse_is_idle;
7581 if (unlikely(se == pse))
7585 * This is possible from callers such as attach_tasks(), in which we
7586 * unconditionally check_preempt_curr() after an enqueue (which may have
7587 * lead to a throttle). This both saves work and prevents false
7588 * next-buddy nomination below.
7590 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
7593 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
7594 set_next_buddy(pse);
7595 next_buddy_marked = 1;
7599 * We can come here with TIF_NEED_RESCHED already set from new task
7602 * Note: this also catches the edge-case of curr being in a throttled
7603 * group (e.g. via set_curr_task), since update_curr() (in the
7604 * enqueue of curr) will have resulted in resched being set. This
7605 * prevents us from potentially nominating it as a false LAST_BUDDY
7608 if (test_tsk_need_resched(curr))
7611 /* Idle tasks are by definition preempted by non-idle tasks. */
7612 if (unlikely(task_has_idle_policy(curr)) &&
7613 likely(!task_has_idle_policy(p)))
7617 * Batch and idle tasks do not preempt non-idle tasks (their preemption
7618 * is driven by the tick):
7620 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
7623 find_matching_se(&se, &pse);
7626 cse_is_idle = se_is_idle(se);
7627 pse_is_idle = se_is_idle(pse);
7630 * Preempt an idle group in favor of a non-idle group (and don't preempt
7631 * in the inverse case).
7633 if (cse_is_idle && !pse_is_idle)
7635 if (cse_is_idle != pse_is_idle)
7638 update_curr(cfs_rq_of(se));
7639 if (wakeup_preempt_entity(se, pse) == 1) {
7641 * Bias pick_next to pick the sched entity that is
7642 * triggering this preemption.
7644 if (!next_buddy_marked)
7645 set_next_buddy(pse);
7654 * Only set the backward buddy when the current task is still
7655 * on the rq. This can happen when a wakeup gets interleaved
7656 * with schedule on the ->pre_schedule() or idle_balance()
7657 * point, either of which can * drop the rq lock.
7659 * Also, during early boot the idle thread is in the fair class,
7660 * for obvious reasons its a bad idea to schedule back to it.
7662 if (unlikely(!se->on_rq || curr == rq->idle))
7665 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
7670 static struct task_struct *pick_task_fair(struct rq *rq)
7672 struct sched_entity *se;
7673 struct cfs_rq *cfs_rq;
7677 if (!cfs_rq->nr_running)
7681 struct sched_entity *curr = cfs_rq->curr;
7683 /* When we pick for a remote RQ, we'll not have done put_prev_entity() */
7686 update_curr(cfs_rq);
7690 if (unlikely(check_cfs_rq_runtime(cfs_rq)))
7694 se = pick_next_entity(cfs_rq, curr);
7695 cfs_rq = group_cfs_rq(se);
7702 struct task_struct *
7703 pick_next_task_fair(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
7705 struct cfs_rq *cfs_rq = &rq->cfs;
7706 struct sched_entity *se;
7707 struct task_struct *p;
7711 if (!sched_fair_runnable(rq))
7714 #ifdef CONFIG_FAIR_GROUP_SCHED
7715 if (!prev || prev->sched_class != &fair_sched_class)
7719 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
7720 * likely that a next task is from the same cgroup as the current.
7722 * Therefore attempt to avoid putting and setting the entire cgroup
7723 * hierarchy, only change the part that actually changes.
7727 struct sched_entity *curr = cfs_rq->curr;
7730 * Since we got here without doing put_prev_entity() we also
7731 * have to consider cfs_rq->curr. If it is still a runnable
7732 * entity, update_curr() will update its vruntime, otherwise
7733 * forget we've ever seen it.
7737 update_curr(cfs_rq);
7742 * This call to check_cfs_rq_runtime() will do the
7743 * throttle and dequeue its entity in the parent(s).
7744 * Therefore the nr_running test will indeed
7747 if (unlikely(check_cfs_rq_runtime(cfs_rq))) {
7750 if (!cfs_rq->nr_running)
7757 se = pick_next_entity(cfs_rq, curr);
7758 cfs_rq = group_cfs_rq(se);
7764 * Since we haven't yet done put_prev_entity and if the selected task
7765 * is a different task than we started out with, try and touch the
7766 * least amount of cfs_rqs.
7769 struct sched_entity *pse = &prev->se;
7771 while (!(cfs_rq = is_same_group(se, pse))) {
7772 int se_depth = se->depth;
7773 int pse_depth = pse->depth;
7775 if (se_depth <= pse_depth) {
7776 put_prev_entity(cfs_rq_of(pse), pse);
7777 pse = parent_entity(pse);
7779 if (se_depth >= pse_depth) {
7780 set_next_entity(cfs_rq_of(se), se);
7781 se = parent_entity(se);
7785 put_prev_entity(cfs_rq, pse);
7786 set_next_entity(cfs_rq, se);
7793 put_prev_task(rq, prev);
7796 se = pick_next_entity(cfs_rq, NULL);
7797 set_next_entity(cfs_rq, se);
7798 cfs_rq = group_cfs_rq(se);
7803 done: __maybe_unused;
7806 * Move the next running task to the front of
7807 * the list, so our cfs_tasks list becomes MRU
7810 list_move(&p->se.group_node, &rq->cfs_tasks);
7813 if (hrtick_enabled_fair(rq))
7814 hrtick_start_fair(rq, p);
7816 update_misfit_status(p, rq);
7824 new_tasks = newidle_balance(rq, rf);
7827 * Because newidle_balance() releases (and re-acquires) rq->lock, it is
7828 * possible for any higher priority task to appear. In that case we
7829 * must re-start the pick_next_entity() loop.
7838 * rq is about to be idle, check if we need to update the
7839 * lost_idle_time of clock_pelt
7841 update_idle_rq_clock_pelt(rq);
7846 static struct task_struct *__pick_next_task_fair(struct rq *rq)
7848 return pick_next_task_fair(rq, NULL, NULL);
7852 * Account for a descheduled task:
7854 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
7856 struct sched_entity *se = &prev->se;
7857 struct cfs_rq *cfs_rq;
7859 for_each_sched_entity(se) {
7860 cfs_rq = cfs_rq_of(se);
7861 put_prev_entity(cfs_rq, se);
7866 * sched_yield() is very simple
7868 * The magic of dealing with the ->skip buddy is in pick_next_entity.
7870 static void yield_task_fair(struct rq *rq)
7872 struct task_struct *curr = rq->curr;
7873 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
7874 struct sched_entity *se = &curr->se;
7877 * Are we the only task in the tree?
7879 if (unlikely(rq->nr_running == 1))
7882 clear_buddies(cfs_rq, se);
7884 if (curr->policy != SCHED_BATCH) {
7885 update_rq_clock(rq);
7887 * Update run-time statistics of the 'current'.
7889 update_curr(cfs_rq);
7891 * Tell update_rq_clock() that we've just updated,
7892 * so we don't do microscopic update in schedule()
7893 * and double the fastpath cost.
7895 rq_clock_skip_update(rq);
7901 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p)
7903 struct sched_entity *se = &p->se;
7905 /* throttled hierarchies are not runnable */
7906 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
7909 /* Tell the scheduler that we'd really like pse to run next. */
7912 yield_task_fair(rq);
7918 /**************************************************
7919 * Fair scheduling class load-balancing methods.
7923 * The purpose of load-balancing is to achieve the same basic fairness the
7924 * per-CPU scheduler provides, namely provide a proportional amount of compute
7925 * time to each task. This is expressed in the following equation:
7927 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
7929 * Where W_i,n is the n-th weight average for CPU i. The instantaneous weight
7930 * W_i,0 is defined as:
7932 * W_i,0 = \Sum_j w_i,j (2)
7934 * Where w_i,j is the weight of the j-th runnable task on CPU i. This weight
7935 * is derived from the nice value as per sched_prio_to_weight[].
7937 * The weight average is an exponential decay average of the instantaneous
7940 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
7942 * C_i is the compute capacity of CPU i, typically it is the
7943 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
7944 * can also include other factors [XXX].
7946 * To achieve this balance we define a measure of imbalance which follows
7947 * directly from (1):
7949 * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4)
7951 * We them move tasks around to minimize the imbalance. In the continuous
7952 * function space it is obvious this converges, in the discrete case we get
7953 * a few fun cases generally called infeasible weight scenarios.
7956 * - infeasible weights;
7957 * - local vs global optima in the discrete case. ]
7962 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
7963 * for all i,j solution, we create a tree of CPUs that follows the hardware
7964 * topology where each level pairs two lower groups (or better). This results
7965 * in O(log n) layers. Furthermore we reduce the number of CPUs going up the
7966 * tree to only the first of the previous level and we decrease the frequency
7967 * of load-balance at each level inv. proportional to the number of CPUs in
7973 * \Sum { --- * --- * 2^i } = O(n) (5)
7975 * `- size of each group
7976 * | | `- number of CPUs doing load-balance
7978 * `- sum over all levels
7980 * Coupled with a limit on how many tasks we can migrate every balance pass,
7981 * this makes (5) the runtime complexity of the balancer.
7983 * An important property here is that each CPU is still (indirectly) connected
7984 * to every other CPU in at most O(log n) steps:
7986 * The adjacency matrix of the resulting graph is given by:
7989 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
7992 * And you'll find that:
7994 * A^(log_2 n)_i,j != 0 for all i,j (7)
7996 * Showing there's indeed a path between every CPU in at most O(log n) steps.
7997 * The task movement gives a factor of O(m), giving a convergence complexity
8000 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
8005 * In order to avoid CPUs going idle while there's still work to do, new idle
8006 * balancing is more aggressive and has the newly idle CPU iterate up the domain
8007 * tree itself instead of relying on other CPUs to bring it work.
8009 * This adds some complexity to both (5) and (8) but it reduces the total idle
8017 * Cgroups make a horror show out of (2), instead of a simple sum we get:
8020 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
8025 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
8027 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on CPU i.
8029 * The big problem is S_k, its a global sum needed to compute a local (W_i)
8032 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
8033 * rewrite all of this once again.]
8036 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
8038 enum fbq_type { regular, remote, all };
8041 * 'group_type' describes the group of CPUs at the moment of load balancing.
8043 * The enum is ordered by pulling priority, with the group with lowest priority
8044 * first so the group_type can simply be compared when selecting the busiest
8045 * group. See update_sd_pick_busiest().
8048 /* The group has spare capacity that can be used to run more tasks. */
8049 group_has_spare = 0,
8051 * The group is fully used and the tasks don't compete for more CPU
8052 * cycles. Nevertheless, some tasks might wait before running.
8056 * One task doesn't fit with CPU's capacity and must be migrated to a
8057 * more powerful CPU.
8061 * SD_ASYM_PACKING only: One local CPU with higher capacity is available,
8062 * and the task should be migrated to it instead of running on the
8067 * The tasks' affinity constraints previously prevented the scheduler
8068 * from balancing the load across the system.
8072 * The CPU is overloaded and can't provide expected CPU cycles to all
8078 enum migration_type {
8085 #define LBF_ALL_PINNED 0x01
8086 #define LBF_NEED_BREAK 0x02
8087 #define LBF_DST_PINNED 0x04
8088 #define LBF_SOME_PINNED 0x08
8089 #define LBF_ACTIVE_LB 0x10
8092 struct sched_domain *sd;
8100 struct cpumask *dst_grpmask;
8102 enum cpu_idle_type idle;
8104 /* The set of CPUs under consideration for load-balancing */
8105 struct cpumask *cpus;
8110 unsigned int loop_break;
8111 unsigned int loop_max;
8113 enum fbq_type fbq_type;
8114 enum migration_type migration_type;
8115 struct list_head tasks;
8119 * Is this task likely cache-hot:
8121 static int task_hot(struct task_struct *p, struct lb_env *env)
8125 lockdep_assert_rq_held(env->src_rq);
8127 if (p->sched_class != &fair_sched_class)
8130 if (unlikely(task_has_idle_policy(p)))
8133 /* SMT siblings share cache */
8134 if (env->sd->flags & SD_SHARE_CPUCAPACITY)
8138 * Buddy candidates are cache hot:
8140 if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running &&
8141 (&p->se == cfs_rq_of(&p->se)->next ||
8142 &p->se == cfs_rq_of(&p->se)->last))
8145 if (sysctl_sched_migration_cost == -1)
8149 * Don't migrate task if the task's cookie does not match
8150 * with the destination CPU's core cookie.
8152 if (!sched_core_cookie_match(cpu_rq(env->dst_cpu), p))
8155 if (sysctl_sched_migration_cost == 0)
8158 delta = rq_clock_task(env->src_rq) - p->se.exec_start;
8160 return delta < (s64)sysctl_sched_migration_cost;
8163 #ifdef CONFIG_NUMA_BALANCING
8165 * Returns 1, if task migration degrades locality
8166 * Returns 0, if task migration improves locality i.e migration preferred.
8167 * Returns -1, if task migration is not affected by locality.
8169 static int migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
8171 struct numa_group *numa_group = rcu_dereference(p->numa_group);
8172 unsigned long src_weight, dst_weight;
8173 int src_nid, dst_nid, dist;
8175 if (!static_branch_likely(&sched_numa_balancing))
8178 if (!p->numa_faults || !(env->sd->flags & SD_NUMA))
8181 src_nid = cpu_to_node(env->src_cpu);
8182 dst_nid = cpu_to_node(env->dst_cpu);
8184 if (src_nid == dst_nid)
8187 /* Migrating away from the preferred node is always bad. */
8188 if (src_nid == p->numa_preferred_nid) {
8189 if (env->src_rq->nr_running > env->src_rq->nr_preferred_running)
8195 /* Encourage migration to the preferred node. */
8196 if (dst_nid == p->numa_preferred_nid)
8199 /* Leaving a core idle is often worse than degrading locality. */
8200 if (env->idle == CPU_IDLE)
8203 dist = node_distance(src_nid, dst_nid);
8205 src_weight = group_weight(p, src_nid, dist);
8206 dst_weight = group_weight(p, dst_nid, dist);
8208 src_weight = task_weight(p, src_nid, dist);
8209 dst_weight = task_weight(p, dst_nid, dist);
8212 return dst_weight < src_weight;
8216 static inline int migrate_degrades_locality(struct task_struct *p,
8224 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
8227 int can_migrate_task(struct task_struct *p, struct lb_env *env)
8231 lockdep_assert_rq_held(env->src_rq);
8234 * We do not migrate tasks that are:
8235 * 1) throttled_lb_pair, or
8236 * 2) cannot be migrated to this CPU due to cpus_ptr, or
8237 * 3) running (obviously), or
8238 * 4) are cache-hot on their current CPU.
8240 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
8243 /* Disregard pcpu kthreads; they are where they need to be. */
8244 if (kthread_is_per_cpu(p))
8247 if (!cpumask_test_cpu(env->dst_cpu, p->cpus_ptr)) {
8250 schedstat_inc(p->stats.nr_failed_migrations_affine);
8252 env->flags |= LBF_SOME_PINNED;
8255 * Remember if this task can be migrated to any other CPU in
8256 * our sched_group. We may want to revisit it if we couldn't
8257 * meet load balance goals by pulling other tasks on src_cpu.
8259 * Avoid computing new_dst_cpu
8261 * - if we have already computed one in current iteration
8262 * - if it's an active balance
8264 if (env->idle == CPU_NEWLY_IDLE ||
8265 env->flags & (LBF_DST_PINNED | LBF_ACTIVE_LB))
8268 /* Prevent to re-select dst_cpu via env's CPUs: */
8269 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
8270 if (cpumask_test_cpu(cpu, p->cpus_ptr)) {
8271 env->flags |= LBF_DST_PINNED;
8272 env->new_dst_cpu = cpu;
8280 /* Record that we found at least one task that could run on dst_cpu */
8281 env->flags &= ~LBF_ALL_PINNED;
8283 if (task_on_cpu(env->src_rq, p)) {
8284 schedstat_inc(p->stats.nr_failed_migrations_running);
8289 * Aggressive migration if:
8291 * 2) destination numa is preferred
8292 * 3) task is cache cold, or
8293 * 4) too many balance attempts have failed.
8295 if (env->flags & LBF_ACTIVE_LB)
8298 tsk_cache_hot = migrate_degrades_locality(p, env);
8299 if (tsk_cache_hot == -1)
8300 tsk_cache_hot = task_hot(p, env);
8302 if (tsk_cache_hot <= 0 ||
8303 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
8304 if (tsk_cache_hot == 1) {
8305 schedstat_inc(env->sd->lb_hot_gained[env->idle]);
8306 schedstat_inc(p->stats.nr_forced_migrations);
8311 schedstat_inc(p->stats.nr_failed_migrations_hot);
8316 * detach_task() -- detach the task for the migration specified in env
8318 static void detach_task(struct task_struct *p, struct lb_env *env)
8320 lockdep_assert_rq_held(env->src_rq);
8322 deactivate_task(env->src_rq, p, DEQUEUE_NOCLOCK);
8323 set_task_cpu(p, env->dst_cpu);
8327 * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
8328 * part of active balancing operations within "domain".
8330 * Returns a task if successful and NULL otherwise.
8332 static struct task_struct *detach_one_task(struct lb_env *env)
8334 struct task_struct *p;
8336 lockdep_assert_rq_held(env->src_rq);
8338 list_for_each_entry_reverse(p,
8339 &env->src_rq->cfs_tasks, se.group_node) {
8340 if (!can_migrate_task(p, env))
8343 detach_task(p, env);
8346 * Right now, this is only the second place where
8347 * lb_gained[env->idle] is updated (other is detach_tasks)
8348 * so we can safely collect stats here rather than
8349 * inside detach_tasks().
8351 schedstat_inc(env->sd->lb_gained[env->idle]);
8358 * detach_tasks() -- tries to detach up to imbalance load/util/tasks from
8359 * busiest_rq, as part of a balancing operation within domain "sd".
8361 * Returns number of detached tasks if successful and 0 otherwise.
8363 static int detach_tasks(struct lb_env *env)
8365 struct list_head *tasks = &env->src_rq->cfs_tasks;
8366 unsigned long util, load;
8367 struct task_struct *p;
8370 lockdep_assert_rq_held(env->src_rq);
8373 * Source run queue has been emptied by another CPU, clear
8374 * LBF_ALL_PINNED flag as we will not test any task.
8376 if (env->src_rq->nr_running <= 1) {
8377 env->flags &= ~LBF_ALL_PINNED;
8381 if (env->imbalance <= 0)
8384 while (!list_empty(tasks)) {
8386 * We don't want to steal all, otherwise we may be treated likewise,
8387 * which could at worst lead to a livelock crash.
8389 if (env->idle != CPU_NOT_IDLE && env->src_rq->nr_running <= 1)
8394 * We've more or less seen every task there is, call it quits
8395 * unless we haven't found any movable task yet.
8397 if (env->loop > env->loop_max &&
8398 !(env->flags & LBF_ALL_PINNED))
8401 /* take a breather every nr_migrate tasks */
8402 if (env->loop > env->loop_break) {
8403 env->loop_break += SCHED_NR_MIGRATE_BREAK;
8404 env->flags |= LBF_NEED_BREAK;
8408 p = list_last_entry(tasks, struct task_struct, se.group_node);
8410 if (!can_migrate_task(p, env))
8413 switch (env->migration_type) {
8416 * Depending of the number of CPUs and tasks and the
8417 * cgroup hierarchy, task_h_load() can return a null
8418 * value. Make sure that env->imbalance decreases
8419 * otherwise detach_tasks() will stop only after
8420 * detaching up to loop_max tasks.
8422 load = max_t(unsigned long, task_h_load(p), 1);
8424 if (sched_feat(LB_MIN) &&
8425 load < 16 && !env->sd->nr_balance_failed)
8429 * Make sure that we don't migrate too much load.
8430 * Nevertheless, let relax the constraint if
8431 * scheduler fails to find a good waiting task to
8434 if (shr_bound(load, env->sd->nr_balance_failed) > env->imbalance)
8437 env->imbalance -= load;
8441 util = task_util_est(p);
8443 if (util > env->imbalance)
8446 env->imbalance -= util;
8453 case migrate_misfit:
8454 /* This is not a misfit task */
8455 if (task_fits_cpu(p, env->src_cpu))
8462 detach_task(p, env);
8463 list_add(&p->se.group_node, &env->tasks);
8467 #ifdef CONFIG_PREEMPTION
8469 * NEWIDLE balancing is a source of latency, so preemptible
8470 * kernels will stop after the first task is detached to minimize
8471 * the critical section.
8473 if (env->idle == CPU_NEWLY_IDLE)
8478 * We only want to steal up to the prescribed amount of
8481 if (env->imbalance <= 0)
8486 list_move(&p->se.group_node, tasks);
8490 * Right now, this is one of only two places we collect this stat
8491 * so we can safely collect detach_one_task() stats here rather
8492 * than inside detach_one_task().
8494 schedstat_add(env->sd->lb_gained[env->idle], detached);
8500 * attach_task() -- attach the task detached by detach_task() to its new rq.
8502 static void attach_task(struct rq *rq, struct task_struct *p)
8504 lockdep_assert_rq_held(rq);
8506 WARN_ON_ONCE(task_rq(p) != rq);
8507 activate_task(rq, p, ENQUEUE_NOCLOCK);
8508 check_preempt_curr(rq, p, 0);
8512 * attach_one_task() -- attaches the task returned from detach_one_task() to
8515 static void attach_one_task(struct rq *rq, struct task_struct *p)
8520 update_rq_clock(rq);
8526 * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
8529 static void attach_tasks(struct lb_env *env)
8531 struct list_head *tasks = &env->tasks;
8532 struct task_struct *p;
8535 rq_lock(env->dst_rq, &rf);
8536 update_rq_clock(env->dst_rq);
8538 while (!list_empty(tasks)) {
8539 p = list_first_entry(tasks, struct task_struct, se.group_node);
8540 list_del_init(&p->se.group_node);
8542 attach_task(env->dst_rq, p);
8545 rq_unlock(env->dst_rq, &rf);
8548 #ifdef CONFIG_NO_HZ_COMMON
8549 static inline bool cfs_rq_has_blocked(struct cfs_rq *cfs_rq)
8551 if (cfs_rq->avg.load_avg)
8554 if (cfs_rq->avg.util_avg)
8560 static inline bool others_have_blocked(struct rq *rq)
8562 if (READ_ONCE(rq->avg_rt.util_avg))
8565 if (READ_ONCE(rq->avg_dl.util_avg))
8568 if (thermal_load_avg(rq))
8571 #ifdef CONFIG_HAVE_SCHED_AVG_IRQ
8572 if (READ_ONCE(rq->avg_irq.util_avg))
8579 static inline void update_blocked_load_tick(struct rq *rq)
8581 WRITE_ONCE(rq->last_blocked_load_update_tick, jiffies);
8584 static inline void update_blocked_load_status(struct rq *rq, bool has_blocked)
8587 rq->has_blocked_load = 0;
8590 static inline bool cfs_rq_has_blocked(struct cfs_rq *cfs_rq) { return false; }
8591 static inline bool others_have_blocked(struct rq *rq) { return false; }
8592 static inline void update_blocked_load_tick(struct rq *rq) {}
8593 static inline void update_blocked_load_status(struct rq *rq, bool has_blocked) {}
8596 static bool __update_blocked_others(struct rq *rq, bool *done)
8598 const struct sched_class *curr_class;
8599 u64 now = rq_clock_pelt(rq);
8600 unsigned long thermal_pressure;
8604 * update_load_avg() can call cpufreq_update_util(). Make sure that RT,
8605 * DL and IRQ signals have been updated before updating CFS.
8607 curr_class = rq->curr->sched_class;
8609 thermal_pressure = arch_scale_thermal_pressure(cpu_of(rq));
8611 decayed = update_rt_rq_load_avg(now, rq, curr_class == &rt_sched_class) |
8612 update_dl_rq_load_avg(now, rq, curr_class == &dl_sched_class) |
8613 update_thermal_load_avg(rq_clock_thermal(rq), rq, thermal_pressure) |
8614 update_irq_load_avg(rq, 0);
8616 if (others_have_blocked(rq))
8622 #ifdef CONFIG_FAIR_GROUP_SCHED
8624 static bool __update_blocked_fair(struct rq *rq, bool *done)
8626 struct cfs_rq *cfs_rq, *pos;
8627 bool decayed = false;
8628 int cpu = cpu_of(rq);
8631 * Iterates the task_group tree in a bottom up fashion, see
8632 * list_add_leaf_cfs_rq() for details.
8634 for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) {
8635 struct sched_entity *se;
8637 if (update_cfs_rq_load_avg(cfs_rq_clock_pelt(cfs_rq), cfs_rq)) {
8638 update_tg_load_avg(cfs_rq);
8640 if (cfs_rq->nr_running == 0)
8641 update_idle_cfs_rq_clock_pelt(cfs_rq);
8643 if (cfs_rq == &rq->cfs)
8647 /* Propagate pending load changes to the parent, if any: */
8648 se = cfs_rq->tg->se[cpu];
8649 if (se && !skip_blocked_update(se))
8650 update_load_avg(cfs_rq_of(se), se, UPDATE_TG);
8653 * There can be a lot of idle CPU cgroups. Don't let fully
8654 * decayed cfs_rqs linger on the list.
8656 if (cfs_rq_is_decayed(cfs_rq))
8657 list_del_leaf_cfs_rq(cfs_rq);
8659 /* Don't need periodic decay once load/util_avg are null */
8660 if (cfs_rq_has_blocked(cfs_rq))
8668 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
8669 * This needs to be done in a top-down fashion because the load of a child
8670 * group is a fraction of its parents load.
8672 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
8674 struct rq *rq = rq_of(cfs_rq);
8675 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
8676 unsigned long now = jiffies;
8679 if (cfs_rq->last_h_load_update == now)
8682 WRITE_ONCE(cfs_rq->h_load_next, NULL);
8683 for_each_sched_entity(se) {
8684 cfs_rq = cfs_rq_of(se);
8685 WRITE_ONCE(cfs_rq->h_load_next, se);
8686 if (cfs_rq->last_h_load_update == now)
8691 cfs_rq->h_load = cfs_rq_load_avg(cfs_rq);
8692 cfs_rq->last_h_load_update = now;
8695 while ((se = READ_ONCE(cfs_rq->h_load_next)) != NULL) {
8696 load = cfs_rq->h_load;
8697 load = div64_ul(load * se->avg.load_avg,
8698 cfs_rq_load_avg(cfs_rq) + 1);
8699 cfs_rq = group_cfs_rq(se);
8700 cfs_rq->h_load = load;
8701 cfs_rq->last_h_load_update = now;
8705 static unsigned long task_h_load(struct task_struct *p)
8707 struct cfs_rq *cfs_rq = task_cfs_rq(p);
8709 update_cfs_rq_h_load(cfs_rq);
8710 return div64_ul(p->se.avg.load_avg * cfs_rq->h_load,
8711 cfs_rq_load_avg(cfs_rq) + 1);
8714 static bool __update_blocked_fair(struct rq *rq, bool *done)
8716 struct cfs_rq *cfs_rq = &rq->cfs;
8719 decayed = update_cfs_rq_load_avg(cfs_rq_clock_pelt(cfs_rq), cfs_rq);
8720 if (cfs_rq_has_blocked(cfs_rq))
8726 static unsigned long task_h_load(struct task_struct *p)
8728 return p->se.avg.load_avg;
8732 static void update_blocked_averages(int cpu)
8734 bool decayed = false, done = true;
8735 struct rq *rq = cpu_rq(cpu);
8738 rq_lock_irqsave(rq, &rf);
8739 update_blocked_load_tick(rq);
8740 update_rq_clock(rq);
8742 decayed |= __update_blocked_others(rq, &done);
8743 decayed |= __update_blocked_fair(rq, &done);
8745 update_blocked_load_status(rq, !done);
8747 cpufreq_update_util(rq, 0);
8748 rq_unlock_irqrestore(rq, &rf);
8751 /********** Helpers for find_busiest_group ************************/
8754 * sg_lb_stats - stats of a sched_group required for load_balancing
8756 struct sg_lb_stats {
8757 unsigned long avg_load; /*Avg load across the CPUs of the group */
8758 unsigned long group_load; /* Total load over the CPUs of the group */
8759 unsigned long group_capacity;
8760 unsigned long group_util; /* Total utilization over the CPUs of the group */
8761 unsigned long group_runnable; /* Total runnable time over the CPUs of the group */
8762 unsigned int sum_nr_running; /* Nr of tasks running in the group */
8763 unsigned int sum_h_nr_running; /* Nr of CFS tasks running in the group */
8764 unsigned int idle_cpus;
8765 unsigned int group_weight;
8766 enum group_type group_type;
8767 unsigned int group_asym_packing; /* Tasks should be moved to preferred CPU */
8768 unsigned long group_misfit_task_load; /* A CPU has a task too big for its capacity */
8769 #ifdef CONFIG_NUMA_BALANCING
8770 unsigned int nr_numa_running;
8771 unsigned int nr_preferred_running;
8776 * sd_lb_stats - Structure to store the statistics of a sched_domain
8777 * during load balancing.
8779 struct sd_lb_stats {
8780 struct sched_group *busiest; /* Busiest group in this sd */
8781 struct sched_group *local; /* Local group in this sd */
8782 unsigned long total_load; /* Total load of all groups in sd */
8783 unsigned long total_capacity; /* Total capacity of all groups in sd */
8784 unsigned long avg_load; /* Average load across all groups in sd */
8785 unsigned int prefer_sibling; /* tasks should go to sibling first */
8787 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
8788 struct sg_lb_stats local_stat; /* Statistics of the local group */
8791 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
8794 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
8795 * local_stat because update_sg_lb_stats() does a full clear/assignment.
8796 * We must however set busiest_stat::group_type and
8797 * busiest_stat::idle_cpus to the worst busiest group because
8798 * update_sd_pick_busiest() reads these before assignment.
8800 *sds = (struct sd_lb_stats){
8804 .total_capacity = 0UL,
8806 .idle_cpus = UINT_MAX,
8807 .group_type = group_has_spare,
8812 static unsigned long scale_rt_capacity(int cpu)
8814 struct rq *rq = cpu_rq(cpu);
8815 unsigned long max = arch_scale_cpu_capacity(cpu);
8816 unsigned long used, free;
8819 irq = cpu_util_irq(rq);
8821 if (unlikely(irq >= max))
8825 * avg_rt.util_avg and avg_dl.util_avg track binary signals
8826 * (running and not running) with weights 0 and 1024 respectively.
8827 * avg_thermal.load_avg tracks thermal pressure and the weighted
8828 * average uses the actual delta max capacity(load).
8830 used = READ_ONCE(rq->avg_rt.util_avg);
8831 used += READ_ONCE(rq->avg_dl.util_avg);
8832 used += thermal_load_avg(rq);
8834 if (unlikely(used >= max))
8839 return scale_irq_capacity(free, irq, max);
8842 static void update_cpu_capacity(struct sched_domain *sd, int cpu)
8844 unsigned long capacity_orig = arch_scale_cpu_capacity(cpu);
8845 unsigned long capacity = scale_rt_capacity(cpu);
8846 struct sched_group *sdg = sd->groups;
8847 struct rq *rq = cpu_rq(cpu);
8849 rq->cpu_capacity_orig = capacity_orig;
8854 rq->cpu_capacity = capacity;
8857 * Detect if the performance domain is in capacity inversion state.
8859 * Capacity inversion happens when another perf domain with equal or
8860 * lower capacity_orig_of() ends up having higher capacity than this
8861 * domain after subtracting thermal pressure.
8863 * We only take into account thermal pressure in this detection as it's
8864 * the only metric that actually results in *real* reduction of
8865 * capacity due to performance points (OPPs) being dropped/become
8866 * unreachable due to thermal throttling.
8869 * * That all cpus in a perf domain have the same capacity_orig
8871 * * Thermal pressure will impact all cpus in this perf domain
8874 if (static_branch_unlikely(&sched_asym_cpucapacity)) {
8875 unsigned long inv_cap = capacity_orig - thermal_load_avg(rq);
8876 struct perf_domain *pd = rcu_dereference(rq->rd->pd);
8878 rq->cpu_capacity_inverted = 0;
8880 for (; pd; pd = pd->next) {
8881 struct cpumask *pd_span = perf_domain_span(pd);
8882 unsigned long pd_cap_orig, pd_cap;
8884 cpu = cpumask_any(pd_span);
8885 pd_cap_orig = arch_scale_cpu_capacity(cpu);
8887 if (capacity_orig < pd_cap_orig)
8891 * handle the case of multiple perf domains have the
8892 * same capacity_orig but one of them is under higher
8893 * thermal pressure. We record it as capacity
8896 if (capacity_orig == pd_cap_orig) {
8897 pd_cap = pd_cap_orig - thermal_load_avg(cpu_rq(cpu));
8899 if (pd_cap > inv_cap) {
8900 rq->cpu_capacity_inverted = inv_cap;
8903 } else if (pd_cap_orig > inv_cap) {
8904 rq->cpu_capacity_inverted = inv_cap;
8910 trace_sched_cpu_capacity_tp(rq);
8912 sdg->sgc->capacity = capacity;
8913 sdg->sgc->min_capacity = capacity;
8914 sdg->sgc->max_capacity = capacity;
8917 void update_group_capacity(struct sched_domain *sd, int cpu)
8919 struct sched_domain *child = sd->child;
8920 struct sched_group *group, *sdg = sd->groups;
8921 unsigned long capacity, min_capacity, max_capacity;
8922 unsigned long interval;
8924 interval = msecs_to_jiffies(sd->balance_interval);
8925 interval = clamp(interval, 1UL, max_load_balance_interval);
8926 sdg->sgc->next_update = jiffies + interval;
8929 update_cpu_capacity(sd, cpu);
8934 min_capacity = ULONG_MAX;
8937 if (child->flags & SD_OVERLAP) {
8939 * SD_OVERLAP domains cannot assume that child groups
8940 * span the current group.
8943 for_each_cpu(cpu, sched_group_span(sdg)) {
8944 unsigned long cpu_cap = capacity_of(cpu);
8946 capacity += cpu_cap;
8947 min_capacity = min(cpu_cap, min_capacity);
8948 max_capacity = max(cpu_cap, max_capacity);
8952 * !SD_OVERLAP domains can assume that child groups
8953 * span the current group.
8956 group = child->groups;
8958 struct sched_group_capacity *sgc = group->sgc;
8960 capacity += sgc->capacity;
8961 min_capacity = min(sgc->min_capacity, min_capacity);
8962 max_capacity = max(sgc->max_capacity, max_capacity);
8963 group = group->next;
8964 } while (group != child->groups);
8967 sdg->sgc->capacity = capacity;
8968 sdg->sgc->min_capacity = min_capacity;
8969 sdg->sgc->max_capacity = max_capacity;
8973 * Check whether the capacity of the rq has been noticeably reduced by side
8974 * activity. The imbalance_pct is used for the threshold.
8975 * Return true is the capacity is reduced
8978 check_cpu_capacity(struct rq *rq, struct sched_domain *sd)
8980 return ((rq->cpu_capacity * sd->imbalance_pct) <
8981 (rq->cpu_capacity_orig * 100));
8985 * Check whether a rq has a misfit task and if it looks like we can actually
8986 * help that task: we can migrate the task to a CPU of higher capacity, or
8987 * the task's current CPU is heavily pressured.
8989 static inline int check_misfit_status(struct rq *rq, struct sched_domain *sd)
8991 return rq->misfit_task_load &&
8992 (rq->cpu_capacity_orig < rq->rd->max_cpu_capacity ||
8993 check_cpu_capacity(rq, sd));
8997 * Group imbalance indicates (and tries to solve) the problem where balancing
8998 * groups is inadequate due to ->cpus_ptr constraints.
9000 * Imagine a situation of two groups of 4 CPUs each and 4 tasks each with a
9001 * cpumask covering 1 CPU of the first group and 3 CPUs of the second group.
9004 * { 0 1 2 3 } { 4 5 6 7 }
9007 * If we were to balance group-wise we'd place two tasks in the first group and
9008 * two tasks in the second group. Clearly this is undesired as it will overload
9009 * cpu 3 and leave one of the CPUs in the second group unused.
9011 * The current solution to this issue is detecting the skew in the first group
9012 * by noticing the lower domain failed to reach balance and had difficulty
9013 * moving tasks due to affinity constraints.
9015 * When this is so detected; this group becomes a candidate for busiest; see
9016 * update_sd_pick_busiest(). And calculate_imbalance() and
9017 * find_busiest_group() avoid some of the usual balance conditions to allow it
9018 * to create an effective group imbalance.
9020 * This is a somewhat tricky proposition since the next run might not find the
9021 * group imbalance and decide the groups need to be balanced again. A most
9022 * subtle and fragile situation.
9025 static inline int sg_imbalanced(struct sched_group *group)
9027 return group->sgc->imbalance;
9031 * group_has_capacity returns true if the group has spare capacity that could
9032 * be used by some tasks.
9033 * We consider that a group has spare capacity if the number of task is
9034 * smaller than the number of CPUs or if the utilization is lower than the
9035 * available capacity for CFS tasks.
9036 * For the latter, we use a threshold to stabilize the state, to take into
9037 * account the variance of the tasks' load and to return true if the available
9038 * capacity in meaningful for the load balancer.
9039 * As an example, an available capacity of 1% can appear but it doesn't make
9040 * any benefit for the load balance.
9043 group_has_capacity(unsigned int imbalance_pct, struct sg_lb_stats *sgs)
9045 if (sgs->sum_nr_running < sgs->group_weight)
9048 if ((sgs->group_capacity * imbalance_pct) <
9049 (sgs->group_runnable * 100))
9052 if ((sgs->group_capacity * 100) >
9053 (sgs->group_util * imbalance_pct))
9060 * group_is_overloaded returns true if the group has more tasks than it can
9062 * group_is_overloaded is not equals to !group_has_capacity because a group
9063 * with the exact right number of tasks, has no more spare capacity but is not
9064 * overloaded so both group_has_capacity and group_is_overloaded return
9068 group_is_overloaded(unsigned int imbalance_pct, struct sg_lb_stats *sgs)
9070 if (sgs->sum_nr_running <= sgs->group_weight)
9073 if ((sgs->group_capacity * 100) <
9074 (sgs->group_util * imbalance_pct))
9077 if ((sgs->group_capacity * imbalance_pct) <
9078 (sgs->group_runnable * 100))
9085 group_type group_classify(unsigned int imbalance_pct,
9086 struct sched_group *group,
9087 struct sg_lb_stats *sgs)
9089 if (group_is_overloaded(imbalance_pct, sgs))
9090 return group_overloaded;
9092 if (sg_imbalanced(group))
9093 return group_imbalanced;
9095 if (sgs->group_asym_packing)
9096 return group_asym_packing;
9098 if (sgs->group_misfit_task_load)
9099 return group_misfit_task;
9101 if (!group_has_capacity(imbalance_pct, sgs))
9102 return group_fully_busy;
9104 return group_has_spare;
9108 * asym_smt_can_pull_tasks - Check whether the load balancing CPU can pull tasks
9109 * @dst_cpu: Destination CPU of the load balancing
9110 * @sds: Load-balancing data with statistics of the local group
9111 * @sgs: Load-balancing statistics of the candidate busiest group
9112 * @sg: The candidate busiest group
9114 * Check the state of the SMT siblings of both @sds::local and @sg and decide
9115 * if @dst_cpu can pull tasks.
9117 * If @dst_cpu does not have SMT siblings, it can pull tasks if two or more of
9118 * the SMT siblings of @sg are busy. If only one CPU in @sg is busy, pull tasks
9119 * only if @dst_cpu has higher priority.
9121 * If both @dst_cpu and @sg have SMT siblings, and @sg has exactly one more
9122 * busy CPU than @sds::local, let @dst_cpu pull tasks if it has higher priority.
9123 * Bigger imbalances in the number of busy CPUs will be dealt with in
9124 * update_sd_pick_busiest().
9126 * If @sg does not have SMT siblings, only pull tasks if all of the SMT siblings
9127 * of @dst_cpu are idle and @sg has lower priority.
9129 * Return: true if @dst_cpu can pull tasks, false otherwise.
9131 static bool asym_smt_can_pull_tasks(int dst_cpu, struct sd_lb_stats *sds,
9132 struct sg_lb_stats *sgs,
9133 struct sched_group *sg)
9135 #ifdef CONFIG_SCHED_SMT
9136 bool local_is_smt, sg_is_smt;
9139 local_is_smt = sds->local->flags & SD_SHARE_CPUCAPACITY;
9140 sg_is_smt = sg->flags & SD_SHARE_CPUCAPACITY;
9142 sg_busy_cpus = sgs->group_weight - sgs->idle_cpus;
9144 if (!local_is_smt) {
9146 * If we are here, @dst_cpu is idle and does not have SMT
9147 * siblings. Pull tasks if candidate group has two or more
9150 if (sg_busy_cpus >= 2) /* implies sg_is_smt */
9154 * @dst_cpu does not have SMT siblings. @sg may have SMT
9155 * siblings and only one is busy. In such case, @dst_cpu
9156 * can help if it has higher priority and is idle (i.e.,
9157 * it has no running tasks).
9159 return sched_asym_prefer(dst_cpu, sg->asym_prefer_cpu);
9162 /* @dst_cpu has SMT siblings. */
9165 int local_busy_cpus = sds->local->group_weight -
9166 sds->local_stat.idle_cpus;
9167 int busy_cpus_delta = sg_busy_cpus - local_busy_cpus;
9169 if (busy_cpus_delta == 1)
9170 return sched_asym_prefer(dst_cpu, sg->asym_prefer_cpu);
9176 * @sg does not have SMT siblings. Ensure that @sds::local does not end
9177 * up with more than one busy SMT sibling and only pull tasks if there
9178 * are not busy CPUs (i.e., no CPU has running tasks).
9180 if (!sds->local_stat.sum_nr_running)
9181 return sched_asym_prefer(dst_cpu, sg->asym_prefer_cpu);
9185 /* Always return false so that callers deal with non-SMT cases. */
9191 sched_asym(struct lb_env *env, struct sd_lb_stats *sds, struct sg_lb_stats *sgs,
9192 struct sched_group *group)
9194 /* Only do SMT checks if either local or candidate have SMT siblings */
9195 if ((sds->local->flags & SD_SHARE_CPUCAPACITY) ||
9196 (group->flags & SD_SHARE_CPUCAPACITY))
9197 return asym_smt_can_pull_tasks(env->dst_cpu, sds, sgs, group);
9199 return sched_asym_prefer(env->dst_cpu, group->asym_prefer_cpu);
9203 sched_reduced_capacity(struct rq *rq, struct sched_domain *sd)
9206 * When there is more than 1 task, the group_overloaded case already
9207 * takes care of cpu with reduced capacity
9209 if (rq->cfs.h_nr_running != 1)
9212 return check_cpu_capacity(rq, sd);
9216 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
9217 * @env: The load balancing environment.
9218 * @sds: Load-balancing data with statistics of the local group.
9219 * @group: sched_group whose statistics are to be updated.
9220 * @sgs: variable to hold the statistics for this group.
9221 * @sg_status: Holds flag indicating the status of the sched_group
9223 static inline void update_sg_lb_stats(struct lb_env *env,
9224 struct sd_lb_stats *sds,
9225 struct sched_group *group,
9226 struct sg_lb_stats *sgs,
9229 int i, nr_running, local_group;
9231 memset(sgs, 0, sizeof(*sgs));
9233 local_group = group == sds->local;
9235 for_each_cpu_and(i, sched_group_span(group), env->cpus) {
9236 struct rq *rq = cpu_rq(i);
9237 unsigned long load = cpu_load(rq);
9239 sgs->group_load += load;
9240 sgs->group_util += cpu_util_cfs(i);
9241 sgs->group_runnable += cpu_runnable(rq);
9242 sgs->sum_h_nr_running += rq->cfs.h_nr_running;
9244 nr_running = rq->nr_running;
9245 sgs->sum_nr_running += nr_running;
9248 *sg_status |= SG_OVERLOAD;
9250 if (cpu_overutilized(i))
9251 *sg_status |= SG_OVERUTILIZED;
9253 #ifdef CONFIG_NUMA_BALANCING
9254 sgs->nr_numa_running += rq->nr_numa_running;
9255 sgs->nr_preferred_running += rq->nr_preferred_running;
9258 * No need to call idle_cpu() if nr_running is not 0
9260 if (!nr_running && idle_cpu(i)) {
9262 /* Idle cpu can't have misfit task */
9269 if (env->sd->flags & SD_ASYM_CPUCAPACITY) {
9270 /* Check for a misfit task on the cpu */
9271 if (sgs->group_misfit_task_load < rq->misfit_task_load) {
9272 sgs->group_misfit_task_load = rq->misfit_task_load;
9273 *sg_status |= SG_OVERLOAD;
9275 } else if ((env->idle != CPU_NOT_IDLE) &&
9276 sched_reduced_capacity(rq, env->sd)) {
9277 /* Check for a task running on a CPU with reduced capacity */
9278 if (sgs->group_misfit_task_load < load)
9279 sgs->group_misfit_task_load = load;
9283 sgs->group_capacity = group->sgc->capacity;
9285 sgs->group_weight = group->group_weight;
9287 /* Check if dst CPU is idle and preferred to this group */
9288 if (!local_group && env->sd->flags & SD_ASYM_PACKING &&
9289 env->idle != CPU_NOT_IDLE && sgs->sum_h_nr_running &&
9290 sched_asym(env, sds, sgs, group)) {
9291 sgs->group_asym_packing = 1;
9294 sgs->group_type = group_classify(env->sd->imbalance_pct, group, sgs);
9296 /* Computing avg_load makes sense only when group is overloaded */
9297 if (sgs->group_type == group_overloaded)
9298 sgs->avg_load = (sgs->group_load * SCHED_CAPACITY_SCALE) /
9299 sgs->group_capacity;
9303 * update_sd_pick_busiest - return 1 on busiest group
9304 * @env: The load balancing environment.
9305 * @sds: sched_domain statistics
9306 * @sg: sched_group candidate to be checked for being the busiest
9307 * @sgs: sched_group statistics
9309 * Determine if @sg is a busier group than the previously selected
9312 * Return: %true if @sg is a busier group than the previously selected
9313 * busiest group. %false otherwise.
9315 static bool update_sd_pick_busiest(struct lb_env *env,
9316 struct sd_lb_stats *sds,
9317 struct sched_group *sg,
9318 struct sg_lb_stats *sgs)
9320 struct sg_lb_stats *busiest = &sds->busiest_stat;
9322 /* Make sure that there is at least one task to pull */
9323 if (!sgs->sum_h_nr_running)
9327 * Don't try to pull misfit tasks we can't help.
9328 * We can use max_capacity here as reduction in capacity on some
9329 * CPUs in the group should either be possible to resolve
9330 * internally or be covered by avg_load imbalance (eventually).
9332 if ((env->sd->flags & SD_ASYM_CPUCAPACITY) &&
9333 (sgs->group_type == group_misfit_task) &&
9334 (!capacity_greater(capacity_of(env->dst_cpu), sg->sgc->max_capacity) ||
9335 sds->local_stat.group_type != group_has_spare))
9338 if (sgs->group_type > busiest->group_type)
9341 if (sgs->group_type < busiest->group_type)
9345 * The candidate and the current busiest group are the same type of
9346 * group. Let check which one is the busiest according to the type.
9349 switch (sgs->group_type) {
9350 case group_overloaded:
9351 /* Select the overloaded group with highest avg_load. */
9352 if (sgs->avg_load <= busiest->avg_load)
9356 case group_imbalanced:
9358 * Select the 1st imbalanced group as we don't have any way to
9359 * choose one more than another.
9363 case group_asym_packing:
9364 /* Prefer to move from lowest priority CPU's work */
9365 if (sched_asym_prefer(sg->asym_prefer_cpu, sds->busiest->asym_prefer_cpu))
9369 case group_misfit_task:
9371 * If we have more than one misfit sg go with the biggest
9374 if (sgs->group_misfit_task_load < busiest->group_misfit_task_load)
9378 case group_fully_busy:
9380 * Select the fully busy group with highest avg_load. In
9381 * theory, there is no need to pull task from such kind of
9382 * group because tasks have all compute capacity that they need
9383 * but we can still improve the overall throughput by reducing
9384 * contention when accessing shared HW resources.
9386 * XXX for now avg_load is not computed and always 0 so we
9387 * select the 1st one.
9389 if (sgs->avg_load <= busiest->avg_load)
9393 case group_has_spare:
9395 * Select not overloaded group with lowest number of idle cpus
9396 * and highest number of running tasks. We could also compare
9397 * the spare capacity which is more stable but it can end up
9398 * that the group has less spare capacity but finally more idle
9399 * CPUs which means less opportunity to pull tasks.
9401 if (sgs->idle_cpus > busiest->idle_cpus)
9403 else if ((sgs->idle_cpus == busiest->idle_cpus) &&
9404 (sgs->sum_nr_running <= busiest->sum_nr_running))
9411 * Candidate sg has no more than one task per CPU and has higher
9412 * per-CPU capacity. Migrating tasks to less capable CPUs may harm
9413 * throughput. Maximize throughput, power/energy consequences are not
9416 if ((env->sd->flags & SD_ASYM_CPUCAPACITY) &&
9417 (sgs->group_type <= group_fully_busy) &&
9418 (capacity_greater(sg->sgc->min_capacity, capacity_of(env->dst_cpu))))
9424 #ifdef CONFIG_NUMA_BALANCING
9425 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
9427 if (sgs->sum_h_nr_running > sgs->nr_numa_running)
9429 if (sgs->sum_h_nr_running > sgs->nr_preferred_running)
9434 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
9436 if (rq->nr_running > rq->nr_numa_running)
9438 if (rq->nr_running > rq->nr_preferred_running)
9443 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
9448 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
9452 #endif /* CONFIG_NUMA_BALANCING */
9458 * task_running_on_cpu - return 1 if @p is running on @cpu.
9461 static unsigned int task_running_on_cpu(int cpu, struct task_struct *p)
9463 /* Task has no contribution or is new */
9464 if (cpu != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time))
9467 if (task_on_rq_queued(p))
9474 * idle_cpu_without - would a given CPU be idle without p ?
9475 * @cpu: the processor on which idleness is tested.
9476 * @p: task which should be ignored.
9478 * Return: 1 if the CPU would be idle. 0 otherwise.
9480 static int idle_cpu_without(int cpu, struct task_struct *p)
9482 struct rq *rq = cpu_rq(cpu);
9484 if (rq->curr != rq->idle && rq->curr != p)
9488 * rq->nr_running can't be used but an updated version without the
9489 * impact of p on cpu must be used instead. The updated nr_running
9490 * be computed and tested before calling idle_cpu_without().
9494 if (rq->ttwu_pending)
9502 * update_sg_wakeup_stats - Update sched_group's statistics for wakeup.
9503 * @sd: The sched_domain level to look for idlest group.
9504 * @group: sched_group whose statistics are to be updated.
9505 * @sgs: variable to hold the statistics for this group.
9506 * @p: The task for which we look for the idlest group/CPU.
9508 static inline void update_sg_wakeup_stats(struct sched_domain *sd,
9509 struct sched_group *group,
9510 struct sg_lb_stats *sgs,
9511 struct task_struct *p)
9515 memset(sgs, 0, sizeof(*sgs));
9517 /* Assume that task can't fit any CPU of the group */
9518 if (sd->flags & SD_ASYM_CPUCAPACITY)
9519 sgs->group_misfit_task_load = 1;
9521 for_each_cpu(i, sched_group_span(group)) {
9522 struct rq *rq = cpu_rq(i);
9525 sgs->group_load += cpu_load_without(rq, p);
9526 sgs->group_util += cpu_util_without(i, p);
9527 sgs->group_runnable += cpu_runnable_without(rq, p);
9528 local = task_running_on_cpu(i, p);
9529 sgs->sum_h_nr_running += rq->cfs.h_nr_running - local;
9531 nr_running = rq->nr_running - local;
9532 sgs->sum_nr_running += nr_running;
9535 * No need to call idle_cpu_without() if nr_running is not 0
9537 if (!nr_running && idle_cpu_without(i, p))
9540 /* Check if task fits in the CPU */
9541 if (sd->flags & SD_ASYM_CPUCAPACITY &&
9542 sgs->group_misfit_task_load &&
9543 task_fits_cpu(p, i))
9544 sgs->group_misfit_task_load = 0;
9548 sgs->group_capacity = group->sgc->capacity;
9550 sgs->group_weight = group->group_weight;
9552 sgs->group_type = group_classify(sd->imbalance_pct, group, sgs);
9555 * Computing avg_load makes sense only when group is fully busy or
9558 if (sgs->group_type == group_fully_busy ||
9559 sgs->group_type == group_overloaded)
9560 sgs->avg_load = (sgs->group_load * SCHED_CAPACITY_SCALE) /
9561 sgs->group_capacity;
9564 static bool update_pick_idlest(struct sched_group *idlest,
9565 struct sg_lb_stats *idlest_sgs,
9566 struct sched_group *group,
9567 struct sg_lb_stats *sgs)
9569 if (sgs->group_type < idlest_sgs->group_type)
9572 if (sgs->group_type > idlest_sgs->group_type)
9576 * The candidate and the current idlest group are the same type of
9577 * group. Let check which one is the idlest according to the type.
9580 switch (sgs->group_type) {
9581 case group_overloaded:
9582 case group_fully_busy:
9583 /* Select the group with lowest avg_load. */
9584 if (idlest_sgs->avg_load <= sgs->avg_load)
9588 case group_imbalanced:
9589 case group_asym_packing:
9590 /* Those types are not used in the slow wakeup path */
9593 case group_misfit_task:
9594 /* Select group with the highest max capacity */
9595 if (idlest->sgc->max_capacity >= group->sgc->max_capacity)
9599 case group_has_spare:
9600 /* Select group with most idle CPUs */
9601 if (idlest_sgs->idle_cpus > sgs->idle_cpus)
9604 /* Select group with lowest group_util */
9605 if (idlest_sgs->idle_cpus == sgs->idle_cpus &&
9606 idlest_sgs->group_util <= sgs->group_util)
9616 * find_idlest_group() finds and returns the least busy CPU group within the
9619 * Assumes p is allowed on at least one CPU in sd.
9621 static struct sched_group *
9622 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
9624 struct sched_group *idlest = NULL, *local = NULL, *group = sd->groups;
9625 struct sg_lb_stats local_sgs, tmp_sgs;
9626 struct sg_lb_stats *sgs;
9627 unsigned long imbalance;
9628 struct sg_lb_stats idlest_sgs = {
9629 .avg_load = UINT_MAX,
9630 .group_type = group_overloaded,
9636 /* Skip over this group if it has no CPUs allowed */
9637 if (!cpumask_intersects(sched_group_span(group),
9641 /* Skip over this group if no cookie matched */
9642 if (!sched_group_cookie_match(cpu_rq(this_cpu), p, group))
9645 local_group = cpumask_test_cpu(this_cpu,
9646 sched_group_span(group));
9655 update_sg_wakeup_stats(sd, group, sgs, p);
9657 if (!local_group && update_pick_idlest(idlest, &idlest_sgs, group, sgs)) {
9662 } while (group = group->next, group != sd->groups);
9665 /* There is no idlest group to push tasks to */
9669 /* The local group has been skipped because of CPU affinity */
9674 * If the local group is idler than the selected idlest group
9675 * don't try and push the task.
9677 if (local_sgs.group_type < idlest_sgs.group_type)
9681 * If the local group is busier than the selected idlest group
9682 * try and push the task.
9684 if (local_sgs.group_type > idlest_sgs.group_type)
9687 switch (local_sgs.group_type) {
9688 case group_overloaded:
9689 case group_fully_busy:
9691 /* Calculate allowed imbalance based on load */
9692 imbalance = scale_load_down(NICE_0_LOAD) *
9693 (sd->imbalance_pct-100) / 100;
9696 * When comparing groups across NUMA domains, it's possible for
9697 * the local domain to be very lightly loaded relative to the
9698 * remote domains but "imbalance" skews the comparison making
9699 * remote CPUs look much more favourable. When considering
9700 * cross-domain, add imbalance to the load on the remote node
9701 * and consider staying local.
9704 if ((sd->flags & SD_NUMA) &&
9705 ((idlest_sgs.avg_load + imbalance) >= local_sgs.avg_load))
9709 * If the local group is less loaded than the selected
9710 * idlest group don't try and push any tasks.
9712 if (idlest_sgs.avg_load >= (local_sgs.avg_load + imbalance))
9715 if (100 * local_sgs.avg_load <= sd->imbalance_pct * idlest_sgs.avg_load)
9719 case group_imbalanced:
9720 case group_asym_packing:
9721 /* Those type are not used in the slow wakeup path */
9724 case group_misfit_task:
9725 /* Select group with the highest max capacity */
9726 if (local->sgc->max_capacity >= idlest->sgc->max_capacity)
9730 case group_has_spare:
9732 if (sd->flags & SD_NUMA) {
9733 int imb_numa_nr = sd->imb_numa_nr;
9734 #ifdef CONFIG_NUMA_BALANCING
9737 * If there is spare capacity at NUMA, try to select
9738 * the preferred node
9740 if (cpu_to_node(this_cpu) == p->numa_preferred_nid)
9743 idlest_cpu = cpumask_first(sched_group_span(idlest));
9744 if (cpu_to_node(idlest_cpu) == p->numa_preferred_nid)
9746 #endif /* CONFIG_NUMA_BALANCING */
9748 * Otherwise, keep the task close to the wakeup source
9749 * and improve locality if the number of running tasks
9750 * would remain below threshold where an imbalance is
9751 * allowed while accounting for the possibility the
9752 * task is pinned to a subset of CPUs. If there is a
9753 * real need of migration, periodic load balance will
9756 if (p->nr_cpus_allowed != NR_CPUS) {
9757 struct cpumask *cpus = this_cpu_cpumask_var_ptr(select_rq_mask);
9759 cpumask_and(cpus, sched_group_span(local), p->cpus_ptr);
9760 imb_numa_nr = min(cpumask_weight(cpus), sd->imb_numa_nr);
9763 imbalance = abs(local_sgs.idle_cpus - idlest_sgs.idle_cpus);
9764 if (!adjust_numa_imbalance(imbalance,
9765 local_sgs.sum_nr_running + 1,
9770 #endif /* CONFIG_NUMA */
9773 * Select group with highest number of idle CPUs. We could also
9774 * compare the utilization which is more stable but it can end
9775 * up that the group has less spare capacity but finally more
9776 * idle CPUs which means more opportunity to run task.
9778 if (local_sgs.idle_cpus >= idlest_sgs.idle_cpus)
9786 static void update_idle_cpu_scan(struct lb_env *env,
9787 unsigned long sum_util)
9789 struct sched_domain_shared *sd_share;
9790 int llc_weight, pct;
9793 * Update the number of CPUs to scan in LLC domain, which could
9794 * be used as a hint in select_idle_cpu(). The update of sd_share
9795 * could be expensive because it is within a shared cache line.
9796 * So the write of this hint only occurs during periodic load
9797 * balancing, rather than CPU_NEWLY_IDLE, because the latter
9798 * can fire way more frequently than the former.
9800 if (!sched_feat(SIS_UTIL) || env->idle == CPU_NEWLY_IDLE)
9803 llc_weight = per_cpu(sd_llc_size, env->dst_cpu);
9804 if (env->sd->span_weight != llc_weight)
9807 sd_share = rcu_dereference(per_cpu(sd_llc_shared, env->dst_cpu));
9812 * The number of CPUs to search drops as sum_util increases, when
9813 * sum_util hits 85% or above, the scan stops.
9814 * The reason to choose 85% as the threshold is because this is the
9815 * imbalance_pct(117) when a LLC sched group is overloaded.
9817 * let y = SCHED_CAPACITY_SCALE - p * x^2 [1]
9818 * and y'= y / SCHED_CAPACITY_SCALE
9820 * x is the ratio of sum_util compared to the CPU capacity:
9821 * x = sum_util / (llc_weight * SCHED_CAPACITY_SCALE)
9822 * y' is the ratio of CPUs to be scanned in the LLC domain,
9823 * and the number of CPUs to scan is calculated by:
9825 * nr_scan = llc_weight * y' [2]
9827 * When x hits the threshold of overloaded, AKA, when
9828 * x = 100 / pct, y drops to 0. According to [1],
9829 * p should be SCHED_CAPACITY_SCALE * pct^2 / 10000
9831 * Scale x by SCHED_CAPACITY_SCALE:
9832 * x' = sum_util / llc_weight; [3]
9834 * and finally [1] becomes:
9835 * y = SCHED_CAPACITY_SCALE -
9836 * x'^2 * pct^2 / (10000 * SCHED_CAPACITY_SCALE) [4]
9841 do_div(x, llc_weight);
9844 pct = env->sd->imbalance_pct;
9845 tmp = x * x * pct * pct;
9846 do_div(tmp, 10000 * SCHED_CAPACITY_SCALE);
9847 tmp = min_t(long, tmp, SCHED_CAPACITY_SCALE);
9848 y = SCHED_CAPACITY_SCALE - tmp;
9852 do_div(y, SCHED_CAPACITY_SCALE);
9853 if ((int)y != sd_share->nr_idle_scan)
9854 WRITE_ONCE(sd_share->nr_idle_scan, (int)y);
9858 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
9859 * @env: The load balancing environment.
9860 * @sds: variable to hold the statistics for this sched_domain.
9863 static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
9865 struct sched_domain *child = env->sd->child;
9866 struct sched_group *sg = env->sd->groups;
9867 struct sg_lb_stats *local = &sds->local_stat;
9868 struct sg_lb_stats tmp_sgs;
9869 unsigned long sum_util = 0;
9873 struct sg_lb_stats *sgs = &tmp_sgs;
9876 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_span(sg));
9881 if (env->idle != CPU_NEWLY_IDLE ||
9882 time_after_eq(jiffies, sg->sgc->next_update))
9883 update_group_capacity(env->sd, env->dst_cpu);
9886 update_sg_lb_stats(env, sds, sg, sgs, &sg_status);
9892 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
9894 sds->busiest_stat = *sgs;
9898 /* Now, start updating sd_lb_stats */
9899 sds->total_load += sgs->group_load;
9900 sds->total_capacity += sgs->group_capacity;
9902 sum_util += sgs->group_util;
9904 } while (sg != env->sd->groups);
9906 /* Tag domain that child domain prefers tasks go to siblings first */
9907 sds->prefer_sibling = child && child->flags & SD_PREFER_SIBLING;
9910 if (env->sd->flags & SD_NUMA)
9911 env->fbq_type = fbq_classify_group(&sds->busiest_stat);
9913 if (!env->sd->parent) {
9914 struct root_domain *rd = env->dst_rq->rd;
9916 /* update overload indicator if we are at root domain */
9917 WRITE_ONCE(rd->overload, sg_status & SG_OVERLOAD);
9919 /* Update over-utilization (tipping point, U >= 0) indicator */
9920 WRITE_ONCE(rd->overutilized, sg_status & SG_OVERUTILIZED);
9921 trace_sched_overutilized_tp(rd, sg_status & SG_OVERUTILIZED);
9922 } else if (sg_status & SG_OVERUTILIZED) {
9923 struct root_domain *rd = env->dst_rq->rd;
9925 WRITE_ONCE(rd->overutilized, SG_OVERUTILIZED);
9926 trace_sched_overutilized_tp(rd, SG_OVERUTILIZED);
9929 update_idle_cpu_scan(env, sum_util);
9933 * calculate_imbalance - Calculate the amount of imbalance present within the
9934 * groups of a given sched_domain during load balance.
9935 * @env: load balance environment
9936 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
9938 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
9940 struct sg_lb_stats *local, *busiest;
9942 local = &sds->local_stat;
9943 busiest = &sds->busiest_stat;
9945 if (busiest->group_type == group_misfit_task) {
9946 if (env->sd->flags & SD_ASYM_CPUCAPACITY) {
9947 /* Set imbalance to allow misfit tasks to be balanced. */
9948 env->migration_type = migrate_misfit;
9952 * Set load imbalance to allow moving task from cpu
9953 * with reduced capacity.
9955 env->migration_type = migrate_load;
9956 env->imbalance = busiest->group_misfit_task_load;
9961 if (busiest->group_type == group_asym_packing) {
9963 * In case of asym capacity, we will try to migrate all load to
9964 * the preferred CPU.
9966 env->migration_type = migrate_task;
9967 env->imbalance = busiest->sum_h_nr_running;
9971 if (busiest->group_type == group_imbalanced) {
9973 * In the group_imb case we cannot rely on group-wide averages
9974 * to ensure CPU-load equilibrium, try to move any task to fix
9975 * the imbalance. The next load balance will take care of
9976 * balancing back the system.
9978 env->migration_type = migrate_task;
9984 * Try to use spare capacity of local group without overloading it or
9987 if (local->group_type == group_has_spare) {
9988 if ((busiest->group_type > group_fully_busy) &&
9989 !(env->sd->flags & SD_SHARE_PKG_RESOURCES)) {
9991 * If busiest is overloaded, try to fill spare
9992 * capacity. This might end up creating spare capacity
9993 * in busiest or busiest still being overloaded but
9994 * there is no simple way to directly compute the
9995 * amount of load to migrate in order to balance the
9998 env->migration_type = migrate_util;
9999 env->imbalance = max(local->group_capacity, local->group_util) -
10003 * In some cases, the group's utilization is max or even
10004 * higher than capacity because of migrations but the
10005 * local CPU is (newly) idle. There is at least one
10006 * waiting task in this overloaded busiest group. Let's
10009 if (env->idle != CPU_NOT_IDLE && env->imbalance == 0) {
10010 env->migration_type = migrate_task;
10011 env->imbalance = 1;
10017 if (busiest->group_weight == 1 || sds->prefer_sibling) {
10018 unsigned int nr_diff = busiest->sum_nr_running;
10020 * When prefer sibling, evenly spread running tasks on
10023 env->migration_type = migrate_task;
10024 lsub_positive(&nr_diff, local->sum_nr_running);
10025 env->imbalance = nr_diff;
10029 * If there is no overload, we just want to even the number of
10032 env->migration_type = migrate_task;
10033 env->imbalance = max_t(long, 0,
10034 (local->idle_cpus - busiest->idle_cpus));
10038 /* Consider allowing a small imbalance between NUMA groups */
10039 if (env->sd->flags & SD_NUMA) {
10040 env->imbalance = adjust_numa_imbalance(env->imbalance,
10041 local->sum_nr_running + 1,
10042 env->sd->imb_numa_nr);
10046 /* Number of tasks to move to restore balance */
10047 env->imbalance >>= 1;
10053 * Local is fully busy but has to take more load to relieve the
10056 if (local->group_type < group_overloaded) {
10058 * Local will become overloaded so the avg_load metrics are
10062 local->avg_load = (local->group_load * SCHED_CAPACITY_SCALE) /
10063 local->group_capacity;
10066 * If the local group is more loaded than the selected
10067 * busiest group don't try to pull any tasks.
10069 if (local->avg_load >= busiest->avg_load) {
10070 env->imbalance = 0;
10074 sds->avg_load = (sds->total_load * SCHED_CAPACITY_SCALE) /
10075 sds->total_capacity;
10079 * Both group are or will become overloaded and we're trying to get all
10080 * the CPUs to the average_load, so we don't want to push ourselves
10081 * above the average load, nor do we wish to reduce the max loaded CPU
10082 * below the average load. At the same time, we also don't want to
10083 * reduce the group load below the group capacity. Thus we look for
10084 * the minimum possible imbalance.
10086 env->migration_type = migrate_load;
10087 env->imbalance = min(
10088 (busiest->avg_load - sds->avg_load) * busiest->group_capacity,
10089 (sds->avg_load - local->avg_load) * local->group_capacity
10090 ) / SCHED_CAPACITY_SCALE;
10093 /******* find_busiest_group() helpers end here *********************/
10096 * Decision matrix according to the local and busiest group type:
10098 * busiest \ local has_spare fully_busy misfit asym imbalanced overloaded
10099 * has_spare nr_idle balanced N/A N/A balanced balanced
10100 * fully_busy nr_idle nr_idle N/A N/A balanced balanced
10101 * misfit_task force N/A N/A N/A N/A N/A
10102 * asym_packing force force N/A N/A force force
10103 * imbalanced force force N/A N/A force force
10104 * overloaded force force N/A N/A force avg_load
10106 * N/A : Not Applicable because already filtered while updating
10108 * balanced : The system is balanced for these 2 groups.
10109 * force : Calculate the imbalance as load migration is probably needed.
10110 * avg_load : Only if imbalance is significant enough.
10111 * nr_idle : dst_cpu is not busy and the number of idle CPUs is quite
10112 * different in groups.
10116 * find_busiest_group - Returns the busiest group within the sched_domain
10117 * if there is an imbalance.
10118 * @env: The load balancing environment.
10120 * Also calculates the amount of runnable load which should be moved
10121 * to restore balance.
10123 * Return: - The busiest group if imbalance exists.
10125 static struct sched_group *find_busiest_group(struct lb_env *env)
10127 struct sg_lb_stats *local, *busiest;
10128 struct sd_lb_stats sds;
10130 init_sd_lb_stats(&sds);
10133 * Compute the various statistics relevant for load balancing at
10136 update_sd_lb_stats(env, &sds);
10138 if (sched_energy_enabled()) {
10139 struct root_domain *rd = env->dst_rq->rd;
10141 if (rcu_dereference(rd->pd) && !READ_ONCE(rd->overutilized))
10145 local = &sds.local_stat;
10146 busiest = &sds.busiest_stat;
10148 /* There is no busy sibling group to pull tasks from */
10152 /* Misfit tasks should be dealt with regardless of the avg load */
10153 if (busiest->group_type == group_misfit_task)
10154 goto force_balance;
10156 /* ASYM feature bypasses nice load balance check */
10157 if (busiest->group_type == group_asym_packing)
10158 goto force_balance;
10161 * If the busiest group is imbalanced the below checks don't
10162 * work because they assume all things are equal, which typically
10163 * isn't true due to cpus_ptr constraints and the like.
10165 if (busiest->group_type == group_imbalanced)
10166 goto force_balance;
10169 * If the local group is busier than the selected busiest group
10170 * don't try and pull any tasks.
10172 if (local->group_type > busiest->group_type)
10176 * When groups are overloaded, use the avg_load to ensure fairness
10179 if (local->group_type == group_overloaded) {
10181 * If the local group is more loaded than the selected
10182 * busiest group don't try to pull any tasks.
10184 if (local->avg_load >= busiest->avg_load)
10187 /* XXX broken for overlapping NUMA groups */
10188 sds.avg_load = (sds.total_load * SCHED_CAPACITY_SCALE) /
10189 sds.total_capacity;
10192 * Don't pull any tasks if this group is already above the
10193 * domain average load.
10195 if (local->avg_load >= sds.avg_load)
10199 * If the busiest group is more loaded, use imbalance_pct to be
10202 if (100 * busiest->avg_load <=
10203 env->sd->imbalance_pct * local->avg_load)
10207 /* Try to move all excess tasks to child's sibling domain */
10208 if (sds.prefer_sibling && local->group_type == group_has_spare &&
10209 busiest->sum_nr_running > local->sum_nr_running + 1)
10210 goto force_balance;
10212 if (busiest->group_type != group_overloaded) {
10213 if (env->idle == CPU_NOT_IDLE)
10215 * If the busiest group is not overloaded (and as a
10216 * result the local one too) but this CPU is already
10217 * busy, let another idle CPU try to pull task.
10221 if (busiest->group_weight > 1 &&
10222 local->idle_cpus <= (busiest->idle_cpus + 1))
10224 * If the busiest group is not overloaded
10225 * and there is no imbalance between this and busiest
10226 * group wrt idle CPUs, it is balanced. The imbalance
10227 * becomes significant if the diff is greater than 1
10228 * otherwise we might end up to just move the imbalance
10229 * on another group. Of course this applies only if
10230 * there is more than 1 CPU per group.
10234 if (busiest->sum_h_nr_running == 1)
10236 * busiest doesn't have any tasks waiting to run
10242 /* Looks like there is an imbalance. Compute it */
10243 calculate_imbalance(env, &sds);
10244 return env->imbalance ? sds.busiest : NULL;
10247 env->imbalance = 0;
10252 * find_busiest_queue - find the busiest runqueue among the CPUs in the group.
10254 static struct rq *find_busiest_queue(struct lb_env *env,
10255 struct sched_group *group)
10257 struct rq *busiest = NULL, *rq;
10258 unsigned long busiest_util = 0, busiest_load = 0, busiest_capacity = 1;
10259 unsigned int busiest_nr = 0;
10262 for_each_cpu_and(i, sched_group_span(group), env->cpus) {
10263 unsigned long capacity, load, util;
10264 unsigned int nr_running;
10268 rt = fbq_classify_rq(rq);
10271 * We classify groups/runqueues into three groups:
10272 * - regular: there are !numa tasks
10273 * - remote: there are numa tasks that run on the 'wrong' node
10274 * - all: there is no distinction
10276 * In order to avoid migrating ideally placed numa tasks,
10277 * ignore those when there's better options.
10279 * If we ignore the actual busiest queue to migrate another
10280 * task, the next balance pass can still reduce the busiest
10281 * queue by moving tasks around inside the node.
10283 * If we cannot move enough load due to this classification
10284 * the next pass will adjust the group classification and
10285 * allow migration of more tasks.
10287 * Both cases only affect the total convergence complexity.
10289 if (rt > env->fbq_type)
10292 nr_running = rq->cfs.h_nr_running;
10296 capacity = capacity_of(i);
10299 * For ASYM_CPUCAPACITY domains, don't pick a CPU that could
10300 * eventually lead to active_balancing high->low capacity.
10301 * Higher per-CPU capacity is considered better than balancing
10304 if (env->sd->flags & SD_ASYM_CPUCAPACITY &&
10305 !capacity_greater(capacity_of(env->dst_cpu), capacity) &&
10309 /* Make sure we only pull tasks from a CPU of lower priority */
10310 if ((env->sd->flags & SD_ASYM_PACKING) &&
10311 sched_asym_prefer(i, env->dst_cpu) &&
10315 switch (env->migration_type) {
10318 * When comparing with load imbalance, use cpu_load()
10319 * which is not scaled with the CPU capacity.
10321 load = cpu_load(rq);
10323 if (nr_running == 1 && load > env->imbalance &&
10324 !check_cpu_capacity(rq, env->sd))
10328 * For the load comparisons with the other CPUs,
10329 * consider the cpu_load() scaled with the CPU
10330 * capacity, so that the load can be moved away
10331 * from the CPU that is potentially running at a
10334 * Thus we're looking for max(load_i / capacity_i),
10335 * crosswise multiplication to rid ourselves of the
10336 * division works out to:
10337 * load_i * capacity_j > load_j * capacity_i;
10338 * where j is our previous maximum.
10340 if (load * busiest_capacity > busiest_load * capacity) {
10341 busiest_load = load;
10342 busiest_capacity = capacity;
10348 util = cpu_util_cfs(i);
10351 * Don't try to pull utilization from a CPU with one
10352 * running task. Whatever its utilization, we will fail
10355 if (nr_running <= 1)
10358 if (busiest_util < util) {
10359 busiest_util = util;
10365 if (busiest_nr < nr_running) {
10366 busiest_nr = nr_running;
10371 case migrate_misfit:
10373 * For ASYM_CPUCAPACITY domains with misfit tasks we
10374 * simply seek the "biggest" misfit task.
10376 if (rq->misfit_task_load > busiest_load) {
10377 busiest_load = rq->misfit_task_load;
10390 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
10391 * so long as it is large enough.
10393 #define MAX_PINNED_INTERVAL 512
10396 asym_active_balance(struct lb_env *env)
10399 * ASYM_PACKING needs to force migrate tasks from busy but
10400 * lower priority CPUs in order to pack all tasks in the
10401 * highest priority CPUs.
10403 return env->idle != CPU_NOT_IDLE && (env->sd->flags & SD_ASYM_PACKING) &&
10404 sched_asym_prefer(env->dst_cpu, env->src_cpu);
10408 imbalanced_active_balance(struct lb_env *env)
10410 struct sched_domain *sd = env->sd;
10413 * The imbalanced case includes the case of pinned tasks preventing a fair
10414 * distribution of the load on the system but also the even distribution of the
10415 * threads on a system with spare capacity
10417 if ((env->migration_type == migrate_task) &&
10418 (sd->nr_balance_failed > sd->cache_nice_tries+2))
10424 static int need_active_balance(struct lb_env *env)
10426 struct sched_domain *sd = env->sd;
10428 if (asym_active_balance(env))
10431 if (imbalanced_active_balance(env))
10435 * The dst_cpu is idle and the src_cpu CPU has only 1 CFS task.
10436 * It's worth migrating the task if the src_cpu's capacity is reduced
10437 * because of other sched_class or IRQs if more capacity stays
10438 * available on dst_cpu.
10440 if ((env->idle != CPU_NOT_IDLE) &&
10441 (env->src_rq->cfs.h_nr_running == 1)) {
10442 if ((check_cpu_capacity(env->src_rq, sd)) &&
10443 (capacity_of(env->src_cpu)*sd->imbalance_pct < capacity_of(env->dst_cpu)*100))
10447 if (env->migration_type == migrate_misfit)
10453 static int active_load_balance_cpu_stop(void *data);
10455 static int should_we_balance(struct lb_env *env)
10457 struct sched_group *sg = env->sd->groups;
10461 * Ensure the balancing environment is consistent; can happen
10462 * when the softirq triggers 'during' hotplug.
10464 if (!cpumask_test_cpu(env->dst_cpu, env->cpus))
10468 * In the newly idle case, we will allow all the CPUs
10469 * to do the newly idle load balance.
10471 * However, we bail out if we already have tasks or a wakeup pending,
10472 * to optimize wakeup latency.
10474 if (env->idle == CPU_NEWLY_IDLE) {
10475 if (env->dst_rq->nr_running > 0 || env->dst_rq->ttwu_pending)
10480 /* Try to find first idle CPU */
10481 for_each_cpu_and(cpu, group_balance_mask(sg), env->cpus) {
10482 if (!idle_cpu(cpu))
10485 /* Are we the first idle CPU? */
10486 return cpu == env->dst_cpu;
10489 /* Are we the first CPU of this group ? */
10490 return group_balance_cpu(sg) == env->dst_cpu;
10494 * Check this_cpu to ensure it is balanced within domain. Attempt to move
10495 * tasks if there is an imbalance.
10497 static int load_balance(int this_cpu, struct rq *this_rq,
10498 struct sched_domain *sd, enum cpu_idle_type idle,
10499 int *continue_balancing)
10501 int ld_moved, cur_ld_moved, active_balance = 0;
10502 struct sched_domain *sd_parent = sd->parent;
10503 struct sched_group *group;
10504 struct rq *busiest;
10505 struct rq_flags rf;
10506 struct cpumask *cpus = this_cpu_cpumask_var_ptr(load_balance_mask);
10507 struct lb_env env = {
10509 .dst_cpu = this_cpu,
10511 .dst_grpmask = sched_group_span(sd->groups),
10513 .loop_break = SCHED_NR_MIGRATE_BREAK,
10516 .tasks = LIST_HEAD_INIT(env.tasks),
10519 cpumask_and(cpus, sched_domain_span(sd), cpu_active_mask);
10521 schedstat_inc(sd->lb_count[idle]);
10524 if (!should_we_balance(&env)) {
10525 *continue_balancing = 0;
10529 group = find_busiest_group(&env);
10531 schedstat_inc(sd->lb_nobusyg[idle]);
10535 busiest = find_busiest_queue(&env, group);
10537 schedstat_inc(sd->lb_nobusyq[idle]);
10541 WARN_ON_ONCE(busiest == env.dst_rq);
10543 schedstat_add(sd->lb_imbalance[idle], env.imbalance);
10545 env.src_cpu = busiest->cpu;
10546 env.src_rq = busiest;
10549 /* Clear this flag as soon as we find a pullable task */
10550 env.flags |= LBF_ALL_PINNED;
10551 if (busiest->nr_running > 1) {
10553 * Attempt to move tasks. If find_busiest_group has found
10554 * an imbalance but busiest->nr_running <= 1, the group is
10555 * still unbalanced. ld_moved simply stays zero, so it is
10556 * correctly treated as an imbalance.
10558 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
10561 rq_lock_irqsave(busiest, &rf);
10562 update_rq_clock(busiest);
10565 * cur_ld_moved - load moved in current iteration
10566 * ld_moved - cumulative load moved across iterations
10568 cur_ld_moved = detach_tasks(&env);
10571 * We've detached some tasks from busiest_rq. Every
10572 * task is masked "TASK_ON_RQ_MIGRATING", so we can safely
10573 * unlock busiest->lock, and we are able to be sure
10574 * that nobody can manipulate the tasks in parallel.
10575 * See task_rq_lock() family for the details.
10578 rq_unlock(busiest, &rf);
10580 if (cur_ld_moved) {
10581 attach_tasks(&env);
10582 ld_moved += cur_ld_moved;
10585 local_irq_restore(rf.flags);
10587 if (env.flags & LBF_NEED_BREAK) {
10588 env.flags &= ~LBF_NEED_BREAK;
10589 /* Stop if we tried all running tasks */
10590 if (env.loop < busiest->nr_running)
10595 * Revisit (affine) tasks on src_cpu that couldn't be moved to
10596 * us and move them to an alternate dst_cpu in our sched_group
10597 * where they can run. The upper limit on how many times we
10598 * iterate on same src_cpu is dependent on number of CPUs in our
10601 * This changes load balance semantics a bit on who can move
10602 * load to a given_cpu. In addition to the given_cpu itself
10603 * (or a ilb_cpu acting on its behalf where given_cpu is
10604 * nohz-idle), we now have balance_cpu in a position to move
10605 * load to given_cpu. In rare situations, this may cause
10606 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
10607 * _independently_ and at _same_ time to move some load to
10608 * given_cpu) causing excess load to be moved to given_cpu.
10609 * This however should not happen so much in practice and
10610 * moreover subsequent load balance cycles should correct the
10611 * excess load moved.
10613 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
10615 /* Prevent to re-select dst_cpu via env's CPUs */
10616 __cpumask_clear_cpu(env.dst_cpu, env.cpus);
10618 env.dst_rq = cpu_rq(env.new_dst_cpu);
10619 env.dst_cpu = env.new_dst_cpu;
10620 env.flags &= ~LBF_DST_PINNED;
10622 env.loop_break = SCHED_NR_MIGRATE_BREAK;
10625 * Go back to "more_balance" rather than "redo" since we
10626 * need to continue with same src_cpu.
10632 * We failed to reach balance because of affinity.
10635 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
10637 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0)
10638 *group_imbalance = 1;
10641 /* All tasks on this runqueue were pinned by CPU affinity */
10642 if (unlikely(env.flags & LBF_ALL_PINNED)) {
10643 __cpumask_clear_cpu(cpu_of(busiest), cpus);
10645 * Attempting to continue load balancing at the current
10646 * sched_domain level only makes sense if there are
10647 * active CPUs remaining as possible busiest CPUs to
10648 * pull load from which are not contained within the
10649 * destination group that is receiving any migrated
10652 if (!cpumask_subset(cpus, env.dst_grpmask)) {
10654 env.loop_break = SCHED_NR_MIGRATE_BREAK;
10657 goto out_all_pinned;
10662 schedstat_inc(sd->lb_failed[idle]);
10664 * Increment the failure counter only on periodic balance.
10665 * We do not want newidle balance, which can be very
10666 * frequent, pollute the failure counter causing
10667 * excessive cache_hot migrations and active balances.
10669 if (idle != CPU_NEWLY_IDLE)
10670 sd->nr_balance_failed++;
10672 if (need_active_balance(&env)) {
10673 unsigned long flags;
10675 raw_spin_rq_lock_irqsave(busiest, flags);
10678 * Don't kick the active_load_balance_cpu_stop,
10679 * if the curr task on busiest CPU can't be
10680 * moved to this_cpu:
10682 if (!cpumask_test_cpu(this_cpu, busiest->curr->cpus_ptr)) {
10683 raw_spin_rq_unlock_irqrestore(busiest, flags);
10684 goto out_one_pinned;
10687 /* Record that we found at least one task that could run on this_cpu */
10688 env.flags &= ~LBF_ALL_PINNED;
10691 * ->active_balance synchronizes accesses to
10692 * ->active_balance_work. Once set, it's cleared
10693 * only after active load balance is finished.
10695 if (!busiest->active_balance) {
10696 busiest->active_balance = 1;
10697 busiest->push_cpu = this_cpu;
10698 active_balance = 1;
10700 raw_spin_rq_unlock_irqrestore(busiest, flags);
10702 if (active_balance) {
10703 stop_one_cpu_nowait(cpu_of(busiest),
10704 active_load_balance_cpu_stop, busiest,
10705 &busiest->active_balance_work);
10709 sd->nr_balance_failed = 0;
10712 if (likely(!active_balance) || need_active_balance(&env)) {
10713 /* We were unbalanced, so reset the balancing interval */
10714 sd->balance_interval = sd->min_interval;
10721 * We reach balance although we may have faced some affinity
10722 * constraints. Clear the imbalance flag only if other tasks got
10723 * a chance to move and fix the imbalance.
10725 if (sd_parent && !(env.flags & LBF_ALL_PINNED)) {
10726 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
10728 if (*group_imbalance)
10729 *group_imbalance = 0;
10734 * We reach balance because all tasks are pinned at this level so
10735 * we can't migrate them. Let the imbalance flag set so parent level
10736 * can try to migrate them.
10738 schedstat_inc(sd->lb_balanced[idle]);
10740 sd->nr_balance_failed = 0;
10746 * newidle_balance() disregards balance intervals, so we could
10747 * repeatedly reach this code, which would lead to balance_interval
10748 * skyrocketing in a short amount of time. Skip the balance_interval
10749 * increase logic to avoid that.
10751 if (env.idle == CPU_NEWLY_IDLE)
10754 /* tune up the balancing interval */
10755 if ((env.flags & LBF_ALL_PINNED &&
10756 sd->balance_interval < MAX_PINNED_INTERVAL) ||
10757 sd->balance_interval < sd->max_interval)
10758 sd->balance_interval *= 2;
10763 static inline unsigned long
10764 get_sd_balance_interval(struct sched_domain *sd, int cpu_busy)
10766 unsigned long interval = sd->balance_interval;
10769 interval *= sd->busy_factor;
10771 /* scale ms to jiffies */
10772 interval = msecs_to_jiffies(interval);
10775 * Reduce likelihood of busy balancing at higher domains racing with
10776 * balancing at lower domains by preventing their balancing periods
10777 * from being multiples of each other.
10782 interval = clamp(interval, 1UL, max_load_balance_interval);
10788 update_next_balance(struct sched_domain *sd, unsigned long *next_balance)
10790 unsigned long interval, next;
10792 /* used by idle balance, so cpu_busy = 0 */
10793 interval = get_sd_balance_interval(sd, 0);
10794 next = sd->last_balance + interval;
10796 if (time_after(*next_balance, next))
10797 *next_balance = next;
10801 * active_load_balance_cpu_stop is run by the CPU stopper. It pushes
10802 * running tasks off the busiest CPU onto idle CPUs. It requires at
10803 * least 1 task to be running on each physical CPU where possible, and
10804 * avoids physical / logical imbalances.
10806 static int active_load_balance_cpu_stop(void *data)
10808 struct rq *busiest_rq = data;
10809 int busiest_cpu = cpu_of(busiest_rq);
10810 int target_cpu = busiest_rq->push_cpu;
10811 struct rq *target_rq = cpu_rq(target_cpu);
10812 struct sched_domain *sd;
10813 struct task_struct *p = NULL;
10814 struct rq_flags rf;
10816 rq_lock_irq(busiest_rq, &rf);
10818 * Between queueing the stop-work and running it is a hole in which
10819 * CPUs can become inactive. We should not move tasks from or to
10822 if (!cpu_active(busiest_cpu) || !cpu_active(target_cpu))
10825 /* Make sure the requested CPU hasn't gone down in the meantime: */
10826 if (unlikely(busiest_cpu != smp_processor_id() ||
10827 !busiest_rq->active_balance))
10830 /* Is there any task to move? */
10831 if (busiest_rq->nr_running <= 1)
10835 * This condition is "impossible", if it occurs
10836 * we need to fix it. Originally reported by
10837 * Bjorn Helgaas on a 128-CPU setup.
10839 WARN_ON_ONCE(busiest_rq == target_rq);
10841 /* Search for an sd spanning us and the target CPU. */
10843 for_each_domain(target_cpu, sd) {
10844 if (cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
10849 struct lb_env env = {
10851 .dst_cpu = target_cpu,
10852 .dst_rq = target_rq,
10853 .src_cpu = busiest_rq->cpu,
10854 .src_rq = busiest_rq,
10856 .flags = LBF_ACTIVE_LB,
10859 schedstat_inc(sd->alb_count);
10860 update_rq_clock(busiest_rq);
10862 p = detach_one_task(&env);
10864 schedstat_inc(sd->alb_pushed);
10865 /* Active balancing done, reset the failure counter. */
10866 sd->nr_balance_failed = 0;
10868 schedstat_inc(sd->alb_failed);
10873 busiest_rq->active_balance = 0;
10874 rq_unlock(busiest_rq, &rf);
10877 attach_one_task(target_rq, p);
10879 local_irq_enable();
10884 static DEFINE_SPINLOCK(balancing);
10887 * Scale the max load_balance interval with the number of CPUs in the system.
10888 * This trades load-balance latency on larger machines for less cross talk.
10890 void update_max_interval(void)
10892 max_load_balance_interval = HZ*num_online_cpus()/10;
10895 static inline bool update_newidle_cost(struct sched_domain *sd, u64 cost)
10897 if (cost > sd->max_newidle_lb_cost) {
10899 * Track max cost of a domain to make sure to not delay the
10900 * next wakeup on the CPU.
10902 sd->max_newidle_lb_cost = cost;
10903 sd->last_decay_max_lb_cost = jiffies;
10904 } else if (time_after(jiffies, sd->last_decay_max_lb_cost + HZ)) {
10906 * Decay the newidle max times by ~1% per second to ensure that
10907 * it is not outdated and the current max cost is actually
10910 sd->max_newidle_lb_cost = (sd->max_newidle_lb_cost * 253) / 256;
10911 sd->last_decay_max_lb_cost = jiffies;
10920 * It checks each scheduling domain to see if it is due to be balanced,
10921 * and initiates a balancing operation if so.
10923 * Balancing parameters are set up in init_sched_domains.
10925 static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle)
10927 int continue_balancing = 1;
10929 int busy = idle != CPU_IDLE && !sched_idle_cpu(cpu);
10930 unsigned long interval;
10931 struct sched_domain *sd;
10932 /* Earliest time when we have to do rebalance again */
10933 unsigned long next_balance = jiffies + 60*HZ;
10934 int update_next_balance = 0;
10935 int need_serialize, need_decay = 0;
10939 for_each_domain(cpu, sd) {
10941 * Decay the newidle max times here because this is a regular
10942 * visit to all the domains.
10944 need_decay = update_newidle_cost(sd, 0);
10945 max_cost += sd->max_newidle_lb_cost;
10948 * Stop the load balance at this level. There is another
10949 * CPU in our sched group which is doing load balancing more
10952 if (!continue_balancing) {
10958 interval = get_sd_balance_interval(sd, busy);
10960 need_serialize = sd->flags & SD_SERIALIZE;
10961 if (need_serialize) {
10962 if (!spin_trylock(&balancing))
10966 if (time_after_eq(jiffies, sd->last_balance + interval)) {
10967 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
10969 * The LBF_DST_PINNED logic could have changed
10970 * env->dst_cpu, so we can't know our idle
10971 * state even if we migrated tasks. Update it.
10973 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
10974 busy = idle != CPU_IDLE && !sched_idle_cpu(cpu);
10976 sd->last_balance = jiffies;
10977 interval = get_sd_balance_interval(sd, busy);
10979 if (need_serialize)
10980 spin_unlock(&balancing);
10982 if (time_after(next_balance, sd->last_balance + interval)) {
10983 next_balance = sd->last_balance + interval;
10984 update_next_balance = 1;
10989 * Ensure the rq-wide value also decays but keep it at a
10990 * reasonable floor to avoid funnies with rq->avg_idle.
10992 rq->max_idle_balance_cost =
10993 max((u64)sysctl_sched_migration_cost, max_cost);
10998 * next_balance will be updated only when there is a need.
10999 * When the cpu is attached to null domain for ex, it will not be
11002 if (likely(update_next_balance))
11003 rq->next_balance = next_balance;
11007 static inline int on_null_domain(struct rq *rq)
11009 return unlikely(!rcu_dereference_sched(rq->sd));
11012 #ifdef CONFIG_NO_HZ_COMMON
11014 * idle load balancing details
11015 * - When one of the busy CPUs notice that there may be an idle rebalancing
11016 * needed, they will kick the idle load balancer, which then does idle
11017 * load balancing for all the idle CPUs.
11018 * - HK_TYPE_MISC CPUs are used for this task, because HK_TYPE_SCHED not set
11022 static inline int find_new_ilb(void)
11025 const struct cpumask *hk_mask;
11027 hk_mask = housekeeping_cpumask(HK_TYPE_MISC);
11029 for_each_cpu_and(ilb, nohz.idle_cpus_mask, hk_mask) {
11031 if (ilb == smp_processor_id())
11042 * Kick a CPU to do the nohz balancing, if it is time for it. We pick any
11043 * idle CPU in the HK_TYPE_MISC housekeeping set (if there is one).
11045 static void kick_ilb(unsigned int flags)
11050 * Increase nohz.next_balance only when if full ilb is triggered but
11051 * not if we only update stats.
11053 if (flags & NOHZ_BALANCE_KICK)
11054 nohz.next_balance = jiffies+1;
11056 ilb_cpu = find_new_ilb();
11058 if (ilb_cpu >= nr_cpu_ids)
11062 * Access to rq::nohz_csd is serialized by NOHZ_KICK_MASK; he who sets
11063 * the first flag owns it; cleared by nohz_csd_func().
11065 flags = atomic_fetch_or(flags, nohz_flags(ilb_cpu));
11066 if (flags & NOHZ_KICK_MASK)
11070 * This way we generate an IPI on the target CPU which
11071 * is idle. And the softirq performing nohz idle load balance
11072 * will be run before returning from the IPI.
11074 smp_call_function_single_async(ilb_cpu, &cpu_rq(ilb_cpu)->nohz_csd);
11078 * Current decision point for kicking the idle load balancer in the presence
11079 * of idle CPUs in the system.
11081 static void nohz_balancer_kick(struct rq *rq)
11083 unsigned long now = jiffies;
11084 struct sched_domain_shared *sds;
11085 struct sched_domain *sd;
11086 int nr_busy, i, cpu = rq->cpu;
11087 unsigned int flags = 0;
11089 if (unlikely(rq->idle_balance))
11093 * We may be recently in ticked or tickless idle mode. At the first
11094 * busy tick after returning from idle, we will update the busy stats.
11096 nohz_balance_exit_idle(rq);
11099 * None are in tickless mode and hence no need for NOHZ idle load
11102 if (likely(!atomic_read(&nohz.nr_cpus)))
11105 if (READ_ONCE(nohz.has_blocked) &&
11106 time_after(now, READ_ONCE(nohz.next_blocked)))
11107 flags = NOHZ_STATS_KICK;
11109 if (time_before(now, nohz.next_balance))
11112 if (rq->nr_running >= 2) {
11113 flags = NOHZ_STATS_KICK | NOHZ_BALANCE_KICK;
11119 sd = rcu_dereference(rq->sd);
11122 * If there's a CFS task and the current CPU has reduced
11123 * capacity; kick the ILB to see if there's a better CPU to run
11126 if (rq->cfs.h_nr_running >= 1 && check_cpu_capacity(rq, sd)) {
11127 flags = NOHZ_STATS_KICK | NOHZ_BALANCE_KICK;
11132 sd = rcu_dereference(per_cpu(sd_asym_packing, cpu));
11135 * When ASYM_PACKING; see if there's a more preferred CPU
11136 * currently idle; in which case, kick the ILB to move tasks
11139 for_each_cpu_and(i, sched_domain_span(sd), nohz.idle_cpus_mask) {
11140 if (sched_asym_prefer(i, cpu)) {
11141 flags = NOHZ_STATS_KICK | NOHZ_BALANCE_KICK;
11147 sd = rcu_dereference(per_cpu(sd_asym_cpucapacity, cpu));
11150 * When ASYM_CPUCAPACITY; see if there's a higher capacity CPU
11151 * to run the misfit task on.
11153 if (check_misfit_status(rq, sd)) {
11154 flags = NOHZ_STATS_KICK | NOHZ_BALANCE_KICK;
11159 * For asymmetric systems, we do not want to nicely balance
11160 * cache use, instead we want to embrace asymmetry and only
11161 * ensure tasks have enough CPU capacity.
11163 * Skip the LLC logic because it's not relevant in that case.
11168 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
11171 * If there is an imbalance between LLC domains (IOW we could
11172 * increase the overall cache use), we need some less-loaded LLC
11173 * domain to pull some load. Likewise, we may need to spread
11174 * load within the current LLC domain (e.g. packed SMT cores but
11175 * other CPUs are idle). We can't really know from here how busy
11176 * the others are - so just get a nohz balance going if it looks
11177 * like this LLC domain has tasks we could move.
11179 nr_busy = atomic_read(&sds->nr_busy_cpus);
11181 flags = NOHZ_STATS_KICK | NOHZ_BALANCE_KICK;
11188 if (READ_ONCE(nohz.needs_update))
11189 flags |= NOHZ_NEXT_KICK;
11195 static void set_cpu_sd_state_busy(int cpu)
11197 struct sched_domain *sd;
11200 sd = rcu_dereference(per_cpu(sd_llc, cpu));
11202 if (!sd || !sd->nohz_idle)
11206 atomic_inc(&sd->shared->nr_busy_cpus);
11211 void nohz_balance_exit_idle(struct rq *rq)
11213 SCHED_WARN_ON(rq != this_rq());
11215 if (likely(!rq->nohz_tick_stopped))
11218 rq->nohz_tick_stopped = 0;
11219 cpumask_clear_cpu(rq->cpu, nohz.idle_cpus_mask);
11220 atomic_dec(&nohz.nr_cpus);
11222 set_cpu_sd_state_busy(rq->cpu);
11225 static void set_cpu_sd_state_idle(int cpu)
11227 struct sched_domain *sd;
11230 sd = rcu_dereference(per_cpu(sd_llc, cpu));
11232 if (!sd || sd->nohz_idle)
11236 atomic_dec(&sd->shared->nr_busy_cpus);
11242 * This routine will record that the CPU is going idle with tick stopped.
11243 * This info will be used in performing idle load balancing in the future.
11245 void nohz_balance_enter_idle(int cpu)
11247 struct rq *rq = cpu_rq(cpu);
11249 SCHED_WARN_ON(cpu != smp_processor_id());
11251 /* If this CPU is going down, then nothing needs to be done: */
11252 if (!cpu_active(cpu))
11255 /* Spare idle load balancing on CPUs that don't want to be disturbed: */
11256 if (!housekeeping_cpu(cpu, HK_TYPE_SCHED))
11260 * Can be set safely without rq->lock held
11261 * If a clear happens, it will have evaluated last additions because
11262 * rq->lock is held during the check and the clear
11264 rq->has_blocked_load = 1;
11267 * The tick is still stopped but load could have been added in the
11268 * meantime. We set the nohz.has_blocked flag to trig a check of the
11269 * *_avg. The CPU is already part of nohz.idle_cpus_mask so the clear
11270 * of nohz.has_blocked can only happen after checking the new load
11272 if (rq->nohz_tick_stopped)
11275 /* If we're a completely isolated CPU, we don't play: */
11276 if (on_null_domain(rq))
11279 rq->nohz_tick_stopped = 1;
11281 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
11282 atomic_inc(&nohz.nr_cpus);
11285 * Ensures that if nohz_idle_balance() fails to observe our
11286 * @idle_cpus_mask store, it must observe the @has_blocked
11287 * and @needs_update stores.
11289 smp_mb__after_atomic();
11291 set_cpu_sd_state_idle(cpu);
11293 WRITE_ONCE(nohz.needs_update, 1);
11296 * Each time a cpu enter idle, we assume that it has blocked load and
11297 * enable the periodic update of the load of idle cpus
11299 WRITE_ONCE(nohz.has_blocked, 1);
11302 static bool update_nohz_stats(struct rq *rq)
11304 unsigned int cpu = rq->cpu;
11306 if (!rq->has_blocked_load)
11309 if (!cpumask_test_cpu(cpu, nohz.idle_cpus_mask))
11312 if (!time_after(jiffies, READ_ONCE(rq->last_blocked_load_update_tick)))
11315 update_blocked_averages(cpu);
11317 return rq->has_blocked_load;
11321 * Internal function that runs load balance for all idle cpus. The load balance
11322 * can be a simple update of blocked load or a complete load balance with
11323 * tasks movement depending of flags.
11325 static void _nohz_idle_balance(struct rq *this_rq, unsigned int flags)
11327 /* Earliest time when we have to do rebalance again */
11328 unsigned long now = jiffies;
11329 unsigned long next_balance = now + 60*HZ;
11330 bool has_blocked_load = false;
11331 int update_next_balance = 0;
11332 int this_cpu = this_rq->cpu;
11336 SCHED_WARN_ON((flags & NOHZ_KICK_MASK) == NOHZ_BALANCE_KICK);
11339 * We assume there will be no idle load after this update and clear
11340 * the has_blocked flag. If a cpu enters idle in the mean time, it will
11341 * set the has_blocked flag and trigger another update of idle load.
11342 * Because a cpu that becomes idle, is added to idle_cpus_mask before
11343 * setting the flag, we are sure to not clear the state and not
11344 * check the load of an idle cpu.
11346 * Same applies to idle_cpus_mask vs needs_update.
11348 if (flags & NOHZ_STATS_KICK)
11349 WRITE_ONCE(nohz.has_blocked, 0);
11350 if (flags & NOHZ_NEXT_KICK)
11351 WRITE_ONCE(nohz.needs_update, 0);
11354 * Ensures that if we miss the CPU, we must see the has_blocked
11355 * store from nohz_balance_enter_idle().
11360 * Start with the next CPU after this_cpu so we will end with this_cpu and let a
11361 * chance for other idle cpu to pull load.
11363 for_each_cpu_wrap(balance_cpu, nohz.idle_cpus_mask, this_cpu+1) {
11364 if (!idle_cpu(balance_cpu))
11368 * If this CPU gets work to do, stop the load balancing
11369 * work being done for other CPUs. Next load
11370 * balancing owner will pick it up.
11372 if (need_resched()) {
11373 if (flags & NOHZ_STATS_KICK)
11374 has_blocked_load = true;
11375 if (flags & NOHZ_NEXT_KICK)
11376 WRITE_ONCE(nohz.needs_update, 1);
11380 rq = cpu_rq(balance_cpu);
11382 if (flags & NOHZ_STATS_KICK)
11383 has_blocked_load |= update_nohz_stats(rq);
11386 * If time for next balance is due,
11389 if (time_after_eq(jiffies, rq->next_balance)) {
11390 struct rq_flags rf;
11392 rq_lock_irqsave(rq, &rf);
11393 update_rq_clock(rq);
11394 rq_unlock_irqrestore(rq, &rf);
11396 if (flags & NOHZ_BALANCE_KICK)
11397 rebalance_domains(rq, CPU_IDLE);
11400 if (time_after(next_balance, rq->next_balance)) {
11401 next_balance = rq->next_balance;
11402 update_next_balance = 1;
11407 * next_balance will be updated only when there is a need.
11408 * When the CPU is attached to null domain for ex, it will not be
11411 if (likely(update_next_balance))
11412 nohz.next_balance = next_balance;
11414 if (flags & NOHZ_STATS_KICK)
11415 WRITE_ONCE(nohz.next_blocked,
11416 now + msecs_to_jiffies(LOAD_AVG_PERIOD));
11419 /* There is still blocked load, enable periodic update */
11420 if (has_blocked_load)
11421 WRITE_ONCE(nohz.has_blocked, 1);
11425 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
11426 * rebalancing for all the cpus for whom scheduler ticks are stopped.
11428 static bool nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
11430 unsigned int flags = this_rq->nohz_idle_balance;
11435 this_rq->nohz_idle_balance = 0;
11437 if (idle != CPU_IDLE)
11440 _nohz_idle_balance(this_rq, flags);
11446 * Check if we need to run the ILB for updating blocked load before entering
11449 void nohz_run_idle_balance(int cpu)
11451 unsigned int flags;
11453 flags = atomic_fetch_andnot(NOHZ_NEWILB_KICK, nohz_flags(cpu));
11456 * Update the blocked load only if no SCHED_SOFTIRQ is about to happen
11457 * (ie NOHZ_STATS_KICK set) and will do the same.
11459 if ((flags == NOHZ_NEWILB_KICK) && !need_resched())
11460 _nohz_idle_balance(cpu_rq(cpu), NOHZ_STATS_KICK);
11463 static void nohz_newidle_balance(struct rq *this_rq)
11465 int this_cpu = this_rq->cpu;
11468 * This CPU doesn't want to be disturbed by scheduler
11471 if (!housekeeping_cpu(this_cpu, HK_TYPE_SCHED))
11474 /* Will wake up very soon. No time for doing anything else*/
11475 if (this_rq->avg_idle < sysctl_sched_migration_cost)
11478 /* Don't need to update blocked load of idle CPUs*/
11479 if (!READ_ONCE(nohz.has_blocked) ||
11480 time_before(jiffies, READ_ONCE(nohz.next_blocked)))
11484 * Set the need to trigger ILB in order to update blocked load
11485 * before entering idle state.
11487 atomic_or(NOHZ_NEWILB_KICK, nohz_flags(this_cpu));
11490 #else /* !CONFIG_NO_HZ_COMMON */
11491 static inline void nohz_balancer_kick(struct rq *rq) { }
11493 static inline bool nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
11498 static inline void nohz_newidle_balance(struct rq *this_rq) { }
11499 #endif /* CONFIG_NO_HZ_COMMON */
11502 * newidle_balance is called by schedule() if this_cpu is about to become
11503 * idle. Attempts to pull tasks from other CPUs.
11506 * < 0 - we released the lock and there are !fair tasks present
11507 * 0 - failed, no new tasks
11508 * > 0 - success, new (fair) tasks present
11510 static int newidle_balance(struct rq *this_rq, struct rq_flags *rf)
11512 unsigned long next_balance = jiffies + HZ;
11513 int this_cpu = this_rq->cpu;
11514 u64 t0, t1, curr_cost = 0;
11515 struct sched_domain *sd;
11516 int pulled_task = 0;
11518 update_misfit_status(NULL, this_rq);
11521 * There is a task waiting to run. No need to search for one.
11522 * Return 0; the task will be enqueued when switching to idle.
11524 if (this_rq->ttwu_pending)
11528 * We must set idle_stamp _before_ calling idle_balance(), such that we
11529 * measure the duration of idle_balance() as idle time.
11531 this_rq->idle_stamp = rq_clock(this_rq);
11534 * Do not pull tasks towards !active CPUs...
11536 if (!cpu_active(this_cpu))
11540 * This is OK, because current is on_cpu, which avoids it being picked
11541 * for load-balance and preemption/IRQs are still disabled avoiding
11542 * further scheduler activity on it and we're being very careful to
11543 * re-start the picking loop.
11545 rq_unpin_lock(this_rq, rf);
11548 sd = rcu_dereference_check_sched_domain(this_rq->sd);
11550 if (!READ_ONCE(this_rq->rd->overload) ||
11551 (sd && this_rq->avg_idle < sd->max_newidle_lb_cost)) {
11554 update_next_balance(sd, &next_balance);
11561 raw_spin_rq_unlock(this_rq);
11563 t0 = sched_clock_cpu(this_cpu);
11564 update_blocked_averages(this_cpu);
11567 for_each_domain(this_cpu, sd) {
11568 int continue_balancing = 1;
11571 update_next_balance(sd, &next_balance);
11573 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost)
11576 if (sd->flags & SD_BALANCE_NEWIDLE) {
11578 pulled_task = load_balance(this_cpu, this_rq,
11579 sd, CPU_NEWLY_IDLE,
11580 &continue_balancing);
11582 t1 = sched_clock_cpu(this_cpu);
11583 domain_cost = t1 - t0;
11584 update_newidle_cost(sd, domain_cost);
11586 curr_cost += domain_cost;
11591 * Stop searching for tasks to pull if there are
11592 * now runnable tasks on this rq.
11594 if (pulled_task || this_rq->nr_running > 0 ||
11595 this_rq->ttwu_pending)
11600 raw_spin_rq_lock(this_rq);
11602 if (curr_cost > this_rq->max_idle_balance_cost)
11603 this_rq->max_idle_balance_cost = curr_cost;
11606 * While browsing the domains, we released the rq lock, a task could
11607 * have been enqueued in the meantime. Since we're not going idle,
11608 * pretend we pulled a task.
11610 if (this_rq->cfs.h_nr_running && !pulled_task)
11613 /* Is there a task of a high priority class? */
11614 if (this_rq->nr_running != this_rq->cfs.h_nr_running)
11618 /* Move the next balance forward */
11619 if (time_after(this_rq->next_balance, next_balance))
11620 this_rq->next_balance = next_balance;
11623 this_rq->idle_stamp = 0;
11625 nohz_newidle_balance(this_rq);
11627 rq_repin_lock(this_rq, rf);
11629 return pulled_task;
11633 * run_rebalance_domains is triggered when needed from the scheduler tick.
11634 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
11636 static __latent_entropy void run_rebalance_domains(struct softirq_action *h)
11638 struct rq *this_rq = this_rq();
11639 enum cpu_idle_type idle = this_rq->idle_balance ?
11640 CPU_IDLE : CPU_NOT_IDLE;
11643 * If this CPU has a pending nohz_balance_kick, then do the
11644 * balancing on behalf of the other idle CPUs whose ticks are
11645 * stopped. Do nohz_idle_balance *before* rebalance_domains to
11646 * give the idle CPUs a chance to load balance. Else we may
11647 * load balance only within the local sched_domain hierarchy
11648 * and abort nohz_idle_balance altogether if we pull some load.
11650 if (nohz_idle_balance(this_rq, idle))
11653 /* normal load balance */
11654 update_blocked_averages(this_rq->cpu);
11655 rebalance_domains(this_rq, idle);
11659 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
11661 void trigger_load_balance(struct rq *rq)
11664 * Don't need to rebalance while attached to NULL domain or
11665 * runqueue CPU is not active
11667 if (unlikely(on_null_domain(rq) || !cpu_active(cpu_of(rq))))
11670 if (time_after_eq(jiffies, rq->next_balance))
11671 raise_softirq(SCHED_SOFTIRQ);
11673 nohz_balancer_kick(rq);
11676 static void rq_online_fair(struct rq *rq)
11680 update_runtime_enabled(rq);
11683 static void rq_offline_fair(struct rq *rq)
11687 /* Ensure any throttled groups are reachable by pick_next_task */
11688 unthrottle_offline_cfs_rqs(rq);
11691 #endif /* CONFIG_SMP */
11693 #ifdef CONFIG_SCHED_CORE
11695 __entity_slice_used(struct sched_entity *se, int min_nr_tasks)
11697 u64 slice = sched_slice(cfs_rq_of(se), se);
11698 u64 rtime = se->sum_exec_runtime - se->prev_sum_exec_runtime;
11700 return (rtime * min_nr_tasks > slice);
11703 #define MIN_NR_TASKS_DURING_FORCEIDLE 2
11704 static inline void task_tick_core(struct rq *rq, struct task_struct *curr)
11706 if (!sched_core_enabled(rq))
11710 * If runqueue has only one task which used up its slice and
11711 * if the sibling is forced idle, then trigger schedule to
11712 * give forced idle task a chance.
11714 * sched_slice() considers only this active rq and it gets the
11715 * whole slice. But during force idle, we have siblings acting
11716 * like a single runqueue and hence we need to consider runnable
11717 * tasks on this CPU and the forced idle CPU. Ideally, we should
11718 * go through the forced idle rq, but that would be a perf hit.
11719 * We can assume that the forced idle CPU has at least
11720 * MIN_NR_TASKS_DURING_FORCEIDLE - 1 tasks and use that to check
11721 * if we need to give up the CPU.
11723 if (rq->core->core_forceidle_count && rq->cfs.nr_running == 1 &&
11724 __entity_slice_used(&curr->se, MIN_NR_TASKS_DURING_FORCEIDLE))
11729 * se_fi_update - Update the cfs_rq->min_vruntime_fi in a CFS hierarchy if needed.
11731 static void se_fi_update(struct sched_entity *se, unsigned int fi_seq, bool forceidle)
11733 for_each_sched_entity(se) {
11734 struct cfs_rq *cfs_rq = cfs_rq_of(se);
11737 if (cfs_rq->forceidle_seq == fi_seq)
11739 cfs_rq->forceidle_seq = fi_seq;
11742 cfs_rq->min_vruntime_fi = cfs_rq->min_vruntime;
11746 void task_vruntime_update(struct rq *rq, struct task_struct *p, bool in_fi)
11748 struct sched_entity *se = &p->se;
11750 if (p->sched_class != &fair_sched_class)
11753 se_fi_update(se, rq->core->core_forceidle_seq, in_fi);
11756 bool cfs_prio_less(struct task_struct *a, struct task_struct *b, bool in_fi)
11758 struct rq *rq = task_rq(a);
11759 struct sched_entity *sea = &a->se;
11760 struct sched_entity *seb = &b->se;
11761 struct cfs_rq *cfs_rqa;
11762 struct cfs_rq *cfs_rqb;
11765 SCHED_WARN_ON(task_rq(b)->core != rq->core);
11767 #ifdef CONFIG_FAIR_GROUP_SCHED
11769 * Find an se in the hierarchy for tasks a and b, such that the se's
11770 * are immediate siblings.
11772 while (sea->cfs_rq->tg != seb->cfs_rq->tg) {
11773 int sea_depth = sea->depth;
11774 int seb_depth = seb->depth;
11776 if (sea_depth >= seb_depth)
11777 sea = parent_entity(sea);
11778 if (sea_depth <= seb_depth)
11779 seb = parent_entity(seb);
11782 se_fi_update(sea, rq->core->core_forceidle_seq, in_fi);
11783 se_fi_update(seb, rq->core->core_forceidle_seq, in_fi);
11785 cfs_rqa = sea->cfs_rq;
11786 cfs_rqb = seb->cfs_rq;
11788 cfs_rqa = &task_rq(a)->cfs;
11789 cfs_rqb = &task_rq(b)->cfs;
11793 * Find delta after normalizing se's vruntime with its cfs_rq's
11794 * min_vruntime_fi, which would have been updated in prior calls
11795 * to se_fi_update().
11797 delta = (s64)(sea->vruntime - seb->vruntime) +
11798 (s64)(cfs_rqb->min_vruntime_fi - cfs_rqa->min_vruntime_fi);
11803 static inline void task_tick_core(struct rq *rq, struct task_struct *curr) {}
11807 * scheduler tick hitting a task of our scheduling class.
11809 * NOTE: This function can be called remotely by the tick offload that
11810 * goes along full dynticks. Therefore no local assumption can be made
11811 * and everything must be accessed through the @rq and @curr passed in
11814 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
11816 struct cfs_rq *cfs_rq;
11817 struct sched_entity *se = &curr->se;
11819 for_each_sched_entity(se) {
11820 cfs_rq = cfs_rq_of(se);
11821 entity_tick(cfs_rq, se, queued);
11824 if (static_branch_unlikely(&sched_numa_balancing))
11825 task_tick_numa(rq, curr);
11827 update_misfit_status(curr, rq);
11828 update_overutilized_status(task_rq(curr));
11830 task_tick_core(rq, curr);
11834 * called on fork with the child task as argument from the parent's context
11835 * - child not yet on the tasklist
11836 * - preemption disabled
11838 static void task_fork_fair(struct task_struct *p)
11840 struct cfs_rq *cfs_rq;
11841 struct sched_entity *se = &p->se, *curr;
11842 struct rq *rq = this_rq();
11843 struct rq_flags rf;
11846 update_rq_clock(rq);
11848 cfs_rq = task_cfs_rq(current);
11849 curr = cfs_rq->curr;
11851 update_curr(cfs_rq);
11852 se->vruntime = curr->vruntime;
11854 place_entity(cfs_rq, se, 1);
11856 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
11858 * Upon rescheduling, sched_class::put_prev_task() will place
11859 * 'current' within the tree based on its new key value.
11861 swap(curr->vruntime, se->vruntime);
11865 se->vruntime -= cfs_rq->min_vruntime;
11866 rq_unlock(rq, &rf);
11870 * Priority of the task has changed. Check to see if we preempt
11871 * the current task.
11874 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
11876 if (!task_on_rq_queued(p))
11879 if (rq->cfs.nr_running == 1)
11883 * Reschedule if we are currently running on this runqueue and
11884 * our priority decreased, or if we are not currently running on
11885 * this runqueue and our priority is higher than the current's
11887 if (task_current(rq, p)) {
11888 if (p->prio > oldprio)
11891 check_preempt_curr(rq, p, 0);
11894 static inline bool vruntime_normalized(struct task_struct *p)
11896 struct sched_entity *se = &p->se;
11899 * In both the TASK_ON_RQ_QUEUED and TASK_ON_RQ_MIGRATING cases,
11900 * the dequeue_entity(.flags=0) will already have normalized the
11907 * When !on_rq, vruntime of the task has usually NOT been normalized.
11908 * But there are some cases where it has already been normalized:
11910 * - A forked child which is waiting for being woken up by
11911 * wake_up_new_task().
11912 * - A task which has been woken up by try_to_wake_up() and
11913 * waiting for actually being woken up by sched_ttwu_pending().
11915 if (!se->sum_exec_runtime ||
11916 (READ_ONCE(p->__state) == TASK_WAKING && p->sched_remote_wakeup))
11922 #ifdef CONFIG_FAIR_GROUP_SCHED
11924 * Propagate the changes of the sched_entity across the tg tree to make it
11925 * visible to the root
11927 static void propagate_entity_cfs_rq(struct sched_entity *se)
11929 struct cfs_rq *cfs_rq = cfs_rq_of(se);
11931 if (cfs_rq_throttled(cfs_rq))
11934 if (!throttled_hierarchy(cfs_rq))
11935 list_add_leaf_cfs_rq(cfs_rq);
11937 /* Start to propagate at parent */
11940 for_each_sched_entity(se) {
11941 cfs_rq = cfs_rq_of(se);
11943 update_load_avg(cfs_rq, se, UPDATE_TG);
11945 if (cfs_rq_throttled(cfs_rq))
11948 if (!throttled_hierarchy(cfs_rq))
11949 list_add_leaf_cfs_rq(cfs_rq);
11953 static void propagate_entity_cfs_rq(struct sched_entity *se) { }
11956 static void detach_entity_cfs_rq(struct sched_entity *se)
11958 struct cfs_rq *cfs_rq = cfs_rq_of(se);
11962 * In case the task sched_avg hasn't been attached:
11963 * - A forked task which hasn't been woken up by wake_up_new_task().
11964 * - A task which has been woken up by try_to_wake_up() but is
11965 * waiting for actually being woken up by sched_ttwu_pending().
11967 if (!se->avg.last_update_time)
11971 /* Catch up with the cfs_rq and remove our load when we leave */
11972 update_load_avg(cfs_rq, se, 0);
11973 detach_entity_load_avg(cfs_rq, se);
11974 update_tg_load_avg(cfs_rq);
11975 propagate_entity_cfs_rq(se);
11978 static void attach_entity_cfs_rq(struct sched_entity *se)
11980 struct cfs_rq *cfs_rq = cfs_rq_of(se);
11982 /* Synchronize entity with its cfs_rq */
11983 update_load_avg(cfs_rq, se, sched_feat(ATTACH_AGE_LOAD) ? 0 : SKIP_AGE_LOAD);
11984 attach_entity_load_avg(cfs_rq, se);
11985 update_tg_load_avg(cfs_rq);
11986 propagate_entity_cfs_rq(se);
11989 static void detach_task_cfs_rq(struct task_struct *p)
11991 struct sched_entity *se = &p->se;
11992 struct cfs_rq *cfs_rq = cfs_rq_of(se);
11994 if (!vruntime_normalized(p)) {
11996 * Fix up our vruntime so that the current sleep doesn't
11997 * cause 'unlimited' sleep bonus.
11999 place_entity(cfs_rq, se, 0);
12000 se->vruntime -= cfs_rq->min_vruntime;
12003 detach_entity_cfs_rq(se);
12006 static void attach_task_cfs_rq(struct task_struct *p)
12008 struct sched_entity *se = &p->se;
12009 struct cfs_rq *cfs_rq = cfs_rq_of(se);
12011 attach_entity_cfs_rq(se);
12013 if (!vruntime_normalized(p))
12014 se->vruntime += cfs_rq->min_vruntime;
12017 static void switched_from_fair(struct rq *rq, struct task_struct *p)
12019 detach_task_cfs_rq(p);
12022 static void switched_to_fair(struct rq *rq, struct task_struct *p)
12024 attach_task_cfs_rq(p);
12026 if (task_on_rq_queued(p)) {
12028 * We were most likely switched from sched_rt, so
12029 * kick off the schedule if running, otherwise just see
12030 * if we can still preempt the current task.
12032 if (task_current(rq, p))
12035 check_preempt_curr(rq, p, 0);
12039 /* Account for a task changing its policy or group.
12041 * This routine is mostly called to set cfs_rq->curr field when a task
12042 * migrates between groups/classes.
12044 static void set_next_task_fair(struct rq *rq, struct task_struct *p, bool first)
12046 struct sched_entity *se = &p->se;
12049 if (task_on_rq_queued(p)) {
12051 * Move the next running task to the front of the list, so our
12052 * cfs_tasks list becomes MRU one.
12054 list_move(&se->group_node, &rq->cfs_tasks);
12058 for_each_sched_entity(se) {
12059 struct cfs_rq *cfs_rq = cfs_rq_of(se);
12061 set_next_entity(cfs_rq, se);
12062 /* ensure bandwidth has been allocated on our new cfs_rq */
12063 account_cfs_rq_runtime(cfs_rq, 0);
12067 void init_cfs_rq(struct cfs_rq *cfs_rq)
12069 cfs_rq->tasks_timeline = RB_ROOT_CACHED;
12070 u64_u32_store(cfs_rq->min_vruntime, (u64)(-(1LL << 20)));
12072 raw_spin_lock_init(&cfs_rq->removed.lock);
12076 #ifdef CONFIG_FAIR_GROUP_SCHED
12077 static void task_change_group_fair(struct task_struct *p)
12080 * We couldn't detach or attach a forked task which
12081 * hasn't been woken up by wake_up_new_task().
12083 if (READ_ONCE(p->__state) == TASK_NEW)
12086 detach_task_cfs_rq(p);
12089 /* Tell se's cfs_rq has been changed -- migrated */
12090 p->se.avg.last_update_time = 0;
12092 set_task_rq(p, task_cpu(p));
12093 attach_task_cfs_rq(p);
12096 void free_fair_sched_group(struct task_group *tg)
12100 for_each_possible_cpu(i) {
12102 kfree(tg->cfs_rq[i]);
12111 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
12113 struct sched_entity *se;
12114 struct cfs_rq *cfs_rq;
12117 tg->cfs_rq = kcalloc(nr_cpu_ids, sizeof(cfs_rq), GFP_KERNEL);
12120 tg->se = kcalloc(nr_cpu_ids, sizeof(se), GFP_KERNEL);
12124 tg->shares = NICE_0_LOAD;
12126 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
12128 for_each_possible_cpu(i) {
12129 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
12130 GFP_KERNEL, cpu_to_node(i));
12134 se = kzalloc_node(sizeof(struct sched_entity_stats),
12135 GFP_KERNEL, cpu_to_node(i));
12139 init_cfs_rq(cfs_rq);
12140 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
12141 init_entity_runnable_average(se);
12152 void online_fair_sched_group(struct task_group *tg)
12154 struct sched_entity *se;
12155 struct rq_flags rf;
12159 for_each_possible_cpu(i) {
12162 rq_lock_irq(rq, &rf);
12163 update_rq_clock(rq);
12164 attach_entity_cfs_rq(se);
12165 sync_throttle(tg, i);
12166 rq_unlock_irq(rq, &rf);
12170 void unregister_fair_sched_group(struct task_group *tg)
12172 unsigned long flags;
12176 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
12178 for_each_possible_cpu(cpu) {
12180 remove_entity_load_avg(tg->se[cpu]);
12183 * Only empty task groups can be destroyed; so we can speculatively
12184 * check on_list without danger of it being re-added.
12186 if (!tg->cfs_rq[cpu]->on_list)
12191 raw_spin_rq_lock_irqsave(rq, flags);
12192 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
12193 raw_spin_rq_unlock_irqrestore(rq, flags);
12197 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
12198 struct sched_entity *se, int cpu,
12199 struct sched_entity *parent)
12201 struct rq *rq = cpu_rq(cpu);
12205 init_cfs_rq_runtime(cfs_rq);
12207 tg->cfs_rq[cpu] = cfs_rq;
12210 /* se could be NULL for root_task_group */
12215 se->cfs_rq = &rq->cfs;
12218 se->cfs_rq = parent->my_q;
12219 se->depth = parent->depth + 1;
12223 /* guarantee group entities always have weight */
12224 update_load_set(&se->load, NICE_0_LOAD);
12225 se->parent = parent;
12228 static DEFINE_MUTEX(shares_mutex);
12230 static int __sched_group_set_shares(struct task_group *tg, unsigned long shares)
12234 lockdep_assert_held(&shares_mutex);
12237 * We can't change the weight of the root cgroup.
12242 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
12244 if (tg->shares == shares)
12247 tg->shares = shares;
12248 for_each_possible_cpu(i) {
12249 struct rq *rq = cpu_rq(i);
12250 struct sched_entity *se = tg->se[i];
12251 struct rq_flags rf;
12253 /* Propagate contribution to hierarchy */
12254 rq_lock_irqsave(rq, &rf);
12255 update_rq_clock(rq);
12256 for_each_sched_entity(se) {
12257 update_load_avg(cfs_rq_of(se), se, UPDATE_TG);
12258 update_cfs_group(se);
12260 rq_unlock_irqrestore(rq, &rf);
12266 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
12270 mutex_lock(&shares_mutex);
12271 if (tg_is_idle(tg))
12274 ret = __sched_group_set_shares(tg, shares);
12275 mutex_unlock(&shares_mutex);
12280 int sched_group_set_idle(struct task_group *tg, long idle)
12284 if (tg == &root_task_group)
12287 if (idle < 0 || idle > 1)
12290 mutex_lock(&shares_mutex);
12292 if (tg->idle == idle) {
12293 mutex_unlock(&shares_mutex);
12299 for_each_possible_cpu(i) {
12300 struct rq *rq = cpu_rq(i);
12301 struct sched_entity *se = tg->se[i];
12302 struct cfs_rq *parent_cfs_rq, *grp_cfs_rq = tg->cfs_rq[i];
12303 bool was_idle = cfs_rq_is_idle(grp_cfs_rq);
12304 long idle_task_delta;
12305 struct rq_flags rf;
12307 rq_lock_irqsave(rq, &rf);
12309 grp_cfs_rq->idle = idle;
12310 if (WARN_ON_ONCE(was_idle == cfs_rq_is_idle(grp_cfs_rq)))
12314 parent_cfs_rq = cfs_rq_of(se);
12315 if (cfs_rq_is_idle(grp_cfs_rq))
12316 parent_cfs_rq->idle_nr_running++;
12318 parent_cfs_rq->idle_nr_running--;
12321 idle_task_delta = grp_cfs_rq->h_nr_running -
12322 grp_cfs_rq->idle_h_nr_running;
12323 if (!cfs_rq_is_idle(grp_cfs_rq))
12324 idle_task_delta *= -1;
12326 for_each_sched_entity(se) {
12327 struct cfs_rq *cfs_rq = cfs_rq_of(se);
12332 cfs_rq->idle_h_nr_running += idle_task_delta;
12334 /* Already accounted at parent level and above. */
12335 if (cfs_rq_is_idle(cfs_rq))
12340 rq_unlock_irqrestore(rq, &rf);
12343 /* Idle groups have minimum weight. */
12344 if (tg_is_idle(tg))
12345 __sched_group_set_shares(tg, scale_load(WEIGHT_IDLEPRIO));
12347 __sched_group_set_shares(tg, NICE_0_LOAD);
12349 mutex_unlock(&shares_mutex);
12353 #else /* CONFIG_FAIR_GROUP_SCHED */
12355 void free_fair_sched_group(struct task_group *tg) { }
12357 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
12362 void online_fair_sched_group(struct task_group *tg) { }
12364 void unregister_fair_sched_group(struct task_group *tg) { }
12366 #endif /* CONFIG_FAIR_GROUP_SCHED */
12369 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
12371 struct sched_entity *se = &task->se;
12372 unsigned int rr_interval = 0;
12375 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
12378 if (rq->cfs.load.weight)
12379 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
12381 return rr_interval;
12385 * All the scheduling class methods:
12387 DEFINE_SCHED_CLASS(fair) = {
12389 .enqueue_task = enqueue_task_fair,
12390 .dequeue_task = dequeue_task_fair,
12391 .yield_task = yield_task_fair,
12392 .yield_to_task = yield_to_task_fair,
12394 .check_preempt_curr = check_preempt_wakeup,
12396 .pick_next_task = __pick_next_task_fair,
12397 .put_prev_task = put_prev_task_fair,
12398 .set_next_task = set_next_task_fair,
12401 .balance = balance_fair,
12402 .pick_task = pick_task_fair,
12403 .select_task_rq = select_task_rq_fair,
12404 .migrate_task_rq = migrate_task_rq_fair,
12406 .rq_online = rq_online_fair,
12407 .rq_offline = rq_offline_fair,
12409 .task_dead = task_dead_fair,
12410 .set_cpus_allowed = set_cpus_allowed_common,
12413 .task_tick = task_tick_fair,
12414 .task_fork = task_fork_fair,
12416 .prio_changed = prio_changed_fair,
12417 .switched_from = switched_from_fair,
12418 .switched_to = switched_to_fair,
12420 .get_rr_interval = get_rr_interval_fair,
12422 .update_curr = update_curr_fair,
12424 #ifdef CONFIG_FAIR_GROUP_SCHED
12425 .task_change_group = task_change_group_fair,
12428 #ifdef CONFIG_UCLAMP_TASK
12429 .uclamp_enabled = 1,
12433 #ifdef CONFIG_SCHED_DEBUG
12434 void print_cfs_stats(struct seq_file *m, int cpu)
12436 struct cfs_rq *cfs_rq, *pos;
12439 for_each_leaf_cfs_rq_safe(cpu_rq(cpu), cfs_rq, pos)
12440 print_cfs_rq(m, cpu, cfs_rq);
12444 #ifdef CONFIG_NUMA_BALANCING
12445 void show_numa_stats(struct task_struct *p, struct seq_file *m)
12448 unsigned long tsf = 0, tpf = 0, gsf = 0, gpf = 0;
12449 struct numa_group *ng;
12452 ng = rcu_dereference(p->numa_group);
12453 for_each_online_node(node) {
12454 if (p->numa_faults) {
12455 tsf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 0)];
12456 tpf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 1)];
12459 gsf = ng->faults[task_faults_idx(NUMA_MEM, node, 0)],
12460 gpf = ng->faults[task_faults_idx(NUMA_MEM, node, 1)];
12462 print_numa_stats(m, node, tsf, tpf, gsf, gpf);
12466 #endif /* CONFIG_NUMA_BALANCING */
12467 #endif /* CONFIG_SCHED_DEBUG */
12469 __init void init_sched_fair_class(void)
12474 for_each_possible_cpu(i) {
12475 zalloc_cpumask_var_node(&per_cpu(load_balance_mask, i), GFP_KERNEL, cpu_to_node(i));
12476 zalloc_cpumask_var_node(&per_cpu(select_rq_mask, i), GFP_KERNEL, cpu_to_node(i));
12479 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
12481 #ifdef CONFIG_NO_HZ_COMMON
12482 nohz.next_balance = jiffies;
12483 nohz.next_blocked = jiffies;
12484 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);