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
26 * Targeted preemption latency for CPU-bound tasks:
28 * NOTE: this latency value is not the same as the concept of
29 * 'timeslice length' - timeslices in CFS are of variable length
30 * and have no persistent notion like in traditional, time-slice
31 * based scheduling concepts.
33 * (to see the precise effective timeslice length of your workload,
34 * run vmstat and monitor the context-switches (cs) field)
36 * (default: 6ms * (1 + ilog(ncpus)), units: nanoseconds)
38 unsigned int sysctl_sched_latency = 6000000ULL;
39 static unsigned int normalized_sysctl_sched_latency = 6000000ULL;
42 * The initial- and re-scaling of tunables is configurable
46 * SCHED_TUNABLESCALING_NONE - unscaled, always *1
47 * SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus)
48 * SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus
50 * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus))
52 enum sched_tunable_scaling sysctl_sched_tunable_scaling = SCHED_TUNABLESCALING_LOG;
55 * Minimal preemption granularity for CPU-bound tasks:
57 * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds)
59 unsigned int sysctl_sched_min_granularity = 750000ULL;
60 static unsigned int normalized_sysctl_sched_min_granularity = 750000ULL;
63 * This value is kept at sysctl_sched_latency/sysctl_sched_min_granularity
65 static unsigned int sched_nr_latency = 8;
68 * After fork, child runs first. If set to 0 (default) then
69 * parent will (try to) run first.
71 unsigned int sysctl_sched_child_runs_first __read_mostly;
74 * SCHED_OTHER wake-up granularity.
76 * This option delays the preemption effects of decoupled workloads
77 * and reduces their over-scheduling. Synchronous workloads will still
78 * have immediate wakeup/sleep latencies.
80 * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds)
82 unsigned int sysctl_sched_wakeup_granularity = 1000000UL;
83 static unsigned int normalized_sysctl_sched_wakeup_granularity = 1000000UL;
85 const_debug unsigned int sysctl_sched_migration_cost = 500000UL;
87 int sched_thermal_decay_shift;
88 static int __init setup_sched_thermal_decay_shift(char *str)
92 if (kstrtoint(str, 0, &_shift))
93 pr_warn("Unable to set scheduler thermal pressure decay shift parameter\n");
95 sched_thermal_decay_shift = clamp(_shift, 0, 10);
98 __setup("sched_thermal_decay_shift=", setup_sched_thermal_decay_shift);
102 * For asym packing, by default the lower numbered CPU has higher priority.
104 int __weak arch_asym_cpu_priority(int cpu)
110 * The margin used when comparing utilization with CPU capacity.
114 #define fits_capacity(cap, max) ((cap) * 1280 < (max) * 1024)
118 #ifdef CONFIG_CFS_BANDWIDTH
120 * Amount of runtime to allocate from global (tg) to local (per-cfs_rq) pool
121 * each time a cfs_rq requests quota.
123 * Note: in the case that the slice exceeds the runtime remaining (either due
124 * to consumption or the quota being specified to be smaller than the slice)
125 * we will always only issue the remaining available time.
127 * (default: 5 msec, units: microseconds)
129 unsigned int sysctl_sched_cfs_bandwidth_slice = 5000UL;
132 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
138 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
144 static inline void update_load_set(struct load_weight *lw, unsigned long w)
151 * Increase the granularity value when there are more CPUs,
152 * because with more CPUs the 'effective latency' as visible
153 * to users decreases. But the relationship is not linear,
154 * so pick a second-best guess by going with the log2 of the
157 * This idea comes from the SD scheduler of Con Kolivas:
159 static unsigned int get_update_sysctl_factor(void)
161 unsigned int cpus = min_t(unsigned int, num_online_cpus(), 8);
164 switch (sysctl_sched_tunable_scaling) {
165 case SCHED_TUNABLESCALING_NONE:
168 case SCHED_TUNABLESCALING_LINEAR:
171 case SCHED_TUNABLESCALING_LOG:
173 factor = 1 + ilog2(cpus);
180 static void update_sysctl(void)
182 unsigned int factor = get_update_sysctl_factor();
184 #define SET_SYSCTL(name) \
185 (sysctl_##name = (factor) * normalized_sysctl_##name)
186 SET_SYSCTL(sched_min_granularity);
187 SET_SYSCTL(sched_latency);
188 SET_SYSCTL(sched_wakeup_granularity);
192 void __init sched_init_granularity(void)
197 #define WMULT_CONST (~0U)
198 #define WMULT_SHIFT 32
200 static void __update_inv_weight(struct load_weight *lw)
204 if (likely(lw->inv_weight))
207 w = scale_load_down(lw->weight);
209 if (BITS_PER_LONG > 32 && unlikely(w >= WMULT_CONST))
211 else if (unlikely(!w))
212 lw->inv_weight = WMULT_CONST;
214 lw->inv_weight = WMULT_CONST / w;
218 * delta_exec * weight / lw.weight
220 * (delta_exec * (weight * lw->inv_weight)) >> WMULT_SHIFT
222 * Either weight := NICE_0_LOAD and lw \e sched_prio_to_wmult[], in which case
223 * we're guaranteed shift stays positive because inv_weight is guaranteed to
224 * fit 32 bits, and NICE_0_LOAD gives another 10 bits; therefore shift >= 22.
226 * Or, weight =< lw.weight (because lw.weight is the runqueue weight), thus
227 * weight/lw.weight <= 1, and therefore our shift will also be positive.
229 static u64 __calc_delta(u64 delta_exec, unsigned long weight, struct load_weight *lw)
231 u64 fact = scale_load_down(weight);
232 u32 fact_hi = (u32)(fact >> 32);
233 int shift = WMULT_SHIFT;
236 __update_inv_weight(lw);
238 if (unlikely(fact_hi)) {
244 fact = mul_u32_u32(fact, lw->inv_weight);
246 fact_hi = (u32)(fact >> 32);
253 return mul_u64_u32_shr(delta_exec, fact, shift);
257 const struct sched_class fair_sched_class;
259 /**************************************************************
260 * CFS operations on generic schedulable entities:
263 #ifdef CONFIG_FAIR_GROUP_SCHED
264 static inline struct task_struct *task_of(struct sched_entity *se)
266 SCHED_WARN_ON(!entity_is_task(se));
267 return container_of(se, struct task_struct, se);
270 /* Walk up scheduling entities hierarchy */
271 #define for_each_sched_entity(se) \
272 for (; se; se = se->parent)
274 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
279 /* runqueue on which this entity is (to be) queued */
280 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
285 /* runqueue "owned" by this group */
286 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
291 static inline void cfs_rq_tg_path(struct cfs_rq *cfs_rq, char *path, int len)
296 if (cfs_rq && task_group_is_autogroup(cfs_rq->tg))
297 autogroup_path(cfs_rq->tg, path, len);
298 else if (cfs_rq && cfs_rq->tg->css.cgroup)
299 cgroup_path(cfs_rq->tg->css.cgroup, path, len);
301 strlcpy(path, "(null)", len);
304 static inline bool list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
306 struct rq *rq = rq_of(cfs_rq);
307 int cpu = cpu_of(rq);
310 return rq->tmp_alone_branch == &rq->leaf_cfs_rq_list;
315 * Ensure we either appear before our parent (if already
316 * enqueued) or force our parent to appear after us when it is
317 * enqueued. The fact that we always enqueue bottom-up
318 * reduces this to two cases and a special case for the root
319 * cfs_rq. Furthermore, it also means that we will always reset
320 * tmp_alone_branch either when the branch is connected
321 * to a tree or when we reach the top of the tree
323 if (cfs_rq->tg->parent &&
324 cfs_rq->tg->parent->cfs_rq[cpu]->on_list) {
326 * If parent is already on the list, we add the child
327 * just before. Thanks to circular linked property of
328 * the list, this means to put the child at the tail
329 * of the list that starts by parent.
331 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
332 &(cfs_rq->tg->parent->cfs_rq[cpu]->leaf_cfs_rq_list));
334 * The branch is now connected to its tree so we can
335 * reset tmp_alone_branch to the beginning of the
338 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
342 if (!cfs_rq->tg->parent) {
344 * cfs rq without parent should be put
345 * at the tail of the list.
347 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
348 &rq->leaf_cfs_rq_list);
350 * We have reach the top of a tree so we can reset
351 * tmp_alone_branch to the beginning of the list.
353 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
358 * The parent has not already been added so we want to
359 * make sure that it will be put after us.
360 * tmp_alone_branch points to the begin of the branch
361 * where we will add parent.
363 list_add_rcu(&cfs_rq->leaf_cfs_rq_list, rq->tmp_alone_branch);
365 * update tmp_alone_branch to points to the new begin
368 rq->tmp_alone_branch = &cfs_rq->leaf_cfs_rq_list;
372 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
374 if (cfs_rq->on_list) {
375 struct rq *rq = rq_of(cfs_rq);
378 * With cfs_rq being unthrottled/throttled during an enqueue,
379 * it can happen the tmp_alone_branch points the a leaf that
380 * we finally want to del. In this case, tmp_alone_branch moves
381 * to the prev element but it will point to rq->leaf_cfs_rq_list
382 * at the end of the enqueue.
384 if (rq->tmp_alone_branch == &cfs_rq->leaf_cfs_rq_list)
385 rq->tmp_alone_branch = cfs_rq->leaf_cfs_rq_list.prev;
387 list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
392 static inline void assert_list_leaf_cfs_rq(struct rq *rq)
394 SCHED_WARN_ON(rq->tmp_alone_branch != &rq->leaf_cfs_rq_list);
397 /* Iterate thr' all leaf cfs_rq's on a runqueue */
398 #define for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) \
399 list_for_each_entry_safe(cfs_rq, pos, &rq->leaf_cfs_rq_list, \
402 /* Do the two (enqueued) entities belong to the same group ? */
403 static inline struct cfs_rq *
404 is_same_group(struct sched_entity *se, struct sched_entity *pse)
406 if (se->cfs_rq == pse->cfs_rq)
412 static inline struct sched_entity *parent_entity(struct sched_entity *se)
418 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
420 int se_depth, pse_depth;
423 * preemption test can be made between sibling entities who are in the
424 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
425 * both tasks until we find their ancestors who are siblings of common
429 /* First walk up until both entities are at same depth */
430 se_depth = (*se)->depth;
431 pse_depth = (*pse)->depth;
433 while (se_depth > pse_depth) {
435 *se = parent_entity(*se);
438 while (pse_depth > se_depth) {
440 *pse = parent_entity(*pse);
443 while (!is_same_group(*se, *pse)) {
444 *se = parent_entity(*se);
445 *pse = parent_entity(*pse);
449 #else /* !CONFIG_FAIR_GROUP_SCHED */
451 static inline struct task_struct *task_of(struct sched_entity *se)
453 return container_of(se, struct task_struct, se);
456 #define for_each_sched_entity(se) \
457 for (; se; se = NULL)
459 static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
461 return &task_rq(p)->cfs;
464 static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
466 struct task_struct *p = task_of(se);
467 struct rq *rq = task_rq(p);
472 /* runqueue "owned" by this group */
473 static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
478 static inline void cfs_rq_tg_path(struct cfs_rq *cfs_rq, char *path, int len)
481 strlcpy(path, "(null)", len);
484 static inline bool list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
489 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
493 static inline void assert_list_leaf_cfs_rq(struct rq *rq)
497 #define for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) \
498 for (cfs_rq = &rq->cfs, pos = NULL; cfs_rq; cfs_rq = pos)
500 static inline struct sched_entity *parent_entity(struct sched_entity *se)
506 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
510 #endif /* CONFIG_FAIR_GROUP_SCHED */
512 static __always_inline
513 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec);
515 /**************************************************************
516 * Scheduling class tree data structure manipulation methods:
519 static inline u64 max_vruntime(u64 max_vruntime, u64 vruntime)
521 s64 delta = (s64)(vruntime - max_vruntime);
523 max_vruntime = vruntime;
528 static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime)
530 s64 delta = (s64)(vruntime - min_vruntime);
532 min_vruntime = vruntime;
537 static inline bool entity_before(struct sched_entity *a,
538 struct sched_entity *b)
540 return (s64)(a->vruntime - b->vruntime) < 0;
543 #define __node_2_se(node) \
544 rb_entry((node), struct sched_entity, run_node)
546 static void update_min_vruntime(struct cfs_rq *cfs_rq)
548 struct sched_entity *curr = cfs_rq->curr;
549 struct rb_node *leftmost = rb_first_cached(&cfs_rq->tasks_timeline);
551 u64 vruntime = cfs_rq->min_vruntime;
555 vruntime = curr->vruntime;
560 if (leftmost) { /* non-empty tree */
561 struct sched_entity *se = __node_2_se(leftmost);
564 vruntime = se->vruntime;
566 vruntime = min_vruntime(vruntime, se->vruntime);
569 /* ensure we never gain time by being placed backwards. */
570 cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime);
573 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
577 static inline bool __entity_less(struct rb_node *a, const struct rb_node *b)
579 return entity_before(__node_2_se(a), __node_2_se(b));
583 * Enqueue an entity into the rb-tree:
585 static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
587 rb_add_cached(&se->run_node, &cfs_rq->tasks_timeline, __entity_less);
590 static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
592 rb_erase_cached(&se->run_node, &cfs_rq->tasks_timeline);
595 struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq)
597 struct rb_node *left = rb_first_cached(&cfs_rq->tasks_timeline);
602 return __node_2_se(left);
605 static struct sched_entity *__pick_next_entity(struct sched_entity *se)
607 struct rb_node *next = rb_next(&se->run_node);
612 return __node_2_se(next);
615 #ifdef CONFIG_SCHED_DEBUG
616 struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
618 struct rb_node *last = rb_last(&cfs_rq->tasks_timeline.rb_root);
623 return __node_2_se(last);
626 /**************************************************************
627 * Scheduling class statistics methods:
630 int sched_proc_update_handler(struct ctl_table *table, int write,
631 void *buffer, size_t *lenp, loff_t *ppos)
633 int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
634 unsigned int factor = get_update_sysctl_factor();
639 sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency,
640 sysctl_sched_min_granularity);
642 #define WRT_SYSCTL(name) \
643 (normalized_sysctl_##name = sysctl_##name / (factor))
644 WRT_SYSCTL(sched_min_granularity);
645 WRT_SYSCTL(sched_latency);
646 WRT_SYSCTL(sched_wakeup_granularity);
656 static inline u64 calc_delta_fair(u64 delta, struct sched_entity *se)
658 if (unlikely(se->load.weight != NICE_0_LOAD))
659 delta = __calc_delta(delta, NICE_0_LOAD, &se->load);
665 * The idea is to set a period in which each task runs once.
667 * When there are too many tasks (sched_nr_latency) we have to stretch
668 * this period because otherwise the slices get too small.
670 * p = (nr <= nl) ? l : l*nr/nl
672 static u64 __sched_period(unsigned long nr_running)
674 if (unlikely(nr_running > sched_nr_latency))
675 return nr_running * sysctl_sched_min_granularity;
677 return sysctl_sched_latency;
681 * We calculate the wall-time slice from the period by taking a part
682 * proportional to the weight.
686 static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
688 u64 slice = __sched_period(cfs_rq->nr_running + !se->on_rq);
690 for_each_sched_entity(se) {
691 struct load_weight *load;
692 struct load_weight lw;
694 cfs_rq = cfs_rq_of(se);
695 load = &cfs_rq->load;
697 if (unlikely(!se->on_rq)) {
700 update_load_add(&lw, se->load.weight);
703 slice = __calc_delta(slice, se->load.weight, load);
709 * We calculate the vruntime slice of a to-be-inserted task.
713 static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se)
715 return calc_delta_fair(sched_slice(cfs_rq, se), se);
721 static int select_idle_sibling(struct task_struct *p, int prev_cpu, int cpu);
722 static unsigned long task_h_load(struct task_struct *p);
723 static unsigned long capacity_of(int cpu);
725 /* Give new sched_entity start runnable values to heavy its load in infant time */
726 void init_entity_runnable_average(struct sched_entity *se)
728 struct sched_avg *sa = &se->avg;
730 memset(sa, 0, sizeof(*sa));
733 * Tasks are initialized with full load to be seen as heavy tasks until
734 * they get a chance to stabilize to their real load level.
735 * Group entities are initialized with zero load to reflect the fact that
736 * nothing has been attached to the task group yet.
738 if (entity_is_task(se))
739 sa->load_avg = scale_load_down(se->load.weight);
741 /* when this task enqueue'ed, it will contribute to its cfs_rq's load_avg */
744 static void attach_entity_cfs_rq(struct sched_entity *se);
747 * With new tasks being created, their initial util_avgs are extrapolated
748 * based on the cfs_rq's current util_avg:
750 * util_avg = cfs_rq->util_avg / (cfs_rq->load_avg + 1) * se.load.weight
752 * However, in many cases, the above util_avg does not give a desired
753 * value. Moreover, the sum of the util_avgs may be divergent, such
754 * as when the series is a harmonic series.
756 * To solve this problem, we also cap the util_avg of successive tasks to
757 * only 1/2 of the left utilization budget:
759 * util_avg_cap = (cpu_scale - cfs_rq->avg.util_avg) / 2^n
761 * where n denotes the nth task and cpu_scale the CPU capacity.
763 * For example, for a CPU with 1024 of capacity, a simplest series from
764 * the beginning would be like:
766 * task util_avg: 512, 256, 128, 64, 32, 16, 8, ...
767 * cfs_rq util_avg: 512, 768, 896, 960, 992, 1008, 1016, ...
769 * Finally, that extrapolated util_avg is clamped to the cap (util_avg_cap)
770 * if util_avg > util_avg_cap.
772 void post_init_entity_util_avg(struct task_struct *p)
774 struct sched_entity *se = &p->se;
775 struct cfs_rq *cfs_rq = cfs_rq_of(se);
776 struct sched_avg *sa = &se->avg;
777 long cpu_scale = arch_scale_cpu_capacity(cpu_of(rq_of(cfs_rq)));
778 long cap = (long)(cpu_scale - cfs_rq->avg.util_avg) / 2;
781 if (cfs_rq->avg.util_avg != 0) {
782 sa->util_avg = cfs_rq->avg.util_avg * se->load.weight;
783 sa->util_avg /= (cfs_rq->avg.load_avg + 1);
785 if (sa->util_avg > cap)
792 sa->runnable_avg = sa->util_avg;
794 if (p->sched_class != &fair_sched_class) {
796 * For !fair tasks do:
798 update_cfs_rq_load_avg(now, cfs_rq);
799 attach_entity_load_avg(cfs_rq, se);
800 switched_from_fair(rq, p);
802 * such that the next switched_to_fair() has the
805 se->avg.last_update_time = cfs_rq_clock_pelt(cfs_rq);
809 attach_entity_cfs_rq(se);
812 #else /* !CONFIG_SMP */
813 void init_entity_runnable_average(struct sched_entity *se)
816 void post_init_entity_util_avg(struct task_struct *p)
819 static void update_tg_load_avg(struct cfs_rq *cfs_rq)
822 #endif /* CONFIG_SMP */
825 * Update the current task's runtime statistics.
827 static void update_curr(struct cfs_rq *cfs_rq)
829 struct sched_entity *curr = cfs_rq->curr;
830 u64 now = rq_clock_task(rq_of(cfs_rq));
836 delta_exec = now - curr->exec_start;
837 if (unlikely((s64)delta_exec <= 0))
840 curr->exec_start = now;
842 schedstat_set(curr->statistics.exec_max,
843 max(delta_exec, curr->statistics.exec_max));
845 curr->sum_exec_runtime += delta_exec;
846 schedstat_add(cfs_rq->exec_clock, delta_exec);
848 curr->vruntime += calc_delta_fair(delta_exec, curr);
849 update_min_vruntime(cfs_rq);
851 if (entity_is_task(curr)) {
852 struct task_struct *curtask = task_of(curr);
854 trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime);
855 cgroup_account_cputime(curtask, delta_exec);
856 account_group_exec_runtime(curtask, delta_exec);
859 account_cfs_rq_runtime(cfs_rq, delta_exec);
862 static void update_curr_fair(struct rq *rq)
864 update_curr(cfs_rq_of(&rq->curr->se));
868 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
870 u64 wait_start, prev_wait_start;
872 if (!schedstat_enabled())
875 wait_start = rq_clock(rq_of(cfs_rq));
876 prev_wait_start = schedstat_val(se->statistics.wait_start);
878 if (entity_is_task(se) && task_on_rq_migrating(task_of(se)) &&
879 likely(wait_start > prev_wait_start))
880 wait_start -= prev_wait_start;
882 __schedstat_set(se->statistics.wait_start, wait_start);
886 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
888 struct task_struct *p;
891 if (!schedstat_enabled())
895 * When the sched_schedstat changes from 0 to 1, some sched se
896 * maybe already in the runqueue, the se->statistics.wait_start
897 * will be 0.So it will let the delta wrong. We need to avoid this
900 if (unlikely(!schedstat_val(se->statistics.wait_start)))
903 delta = rq_clock(rq_of(cfs_rq)) - schedstat_val(se->statistics.wait_start);
905 if (entity_is_task(se)) {
907 if (task_on_rq_migrating(p)) {
909 * Preserve migrating task's wait time so wait_start
910 * time stamp can be adjusted to accumulate wait time
911 * prior to migration.
913 __schedstat_set(se->statistics.wait_start, delta);
916 trace_sched_stat_wait(p, delta);
919 __schedstat_set(se->statistics.wait_max,
920 max(schedstat_val(se->statistics.wait_max), delta));
921 __schedstat_inc(se->statistics.wait_count);
922 __schedstat_add(se->statistics.wait_sum, delta);
923 __schedstat_set(se->statistics.wait_start, 0);
927 update_stats_enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
929 struct task_struct *tsk = NULL;
930 u64 sleep_start, block_start;
932 if (!schedstat_enabled())
935 sleep_start = schedstat_val(se->statistics.sleep_start);
936 block_start = schedstat_val(se->statistics.block_start);
938 if (entity_is_task(se))
942 u64 delta = rq_clock(rq_of(cfs_rq)) - sleep_start;
947 if (unlikely(delta > schedstat_val(se->statistics.sleep_max)))
948 __schedstat_set(se->statistics.sleep_max, delta);
950 __schedstat_set(se->statistics.sleep_start, 0);
951 __schedstat_add(se->statistics.sum_sleep_runtime, delta);
954 account_scheduler_latency(tsk, delta >> 10, 1);
955 trace_sched_stat_sleep(tsk, delta);
959 u64 delta = rq_clock(rq_of(cfs_rq)) - block_start;
964 if (unlikely(delta > schedstat_val(se->statistics.block_max)))
965 __schedstat_set(se->statistics.block_max, delta);
967 __schedstat_set(se->statistics.block_start, 0);
968 __schedstat_add(se->statistics.sum_sleep_runtime, delta);
971 if (tsk->in_iowait) {
972 __schedstat_add(se->statistics.iowait_sum, delta);
973 __schedstat_inc(se->statistics.iowait_count);
974 trace_sched_stat_iowait(tsk, delta);
977 trace_sched_stat_blocked(tsk, delta);
980 * Blocking time is in units of nanosecs, so shift by
981 * 20 to get a milliseconds-range estimation of the
982 * amount of time that the task spent sleeping:
984 if (unlikely(prof_on == SLEEP_PROFILING)) {
985 profile_hits(SLEEP_PROFILING,
986 (void *)get_wchan(tsk),
989 account_scheduler_latency(tsk, delta >> 10, 0);
995 * Task is being enqueued - update stats:
998 update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
1000 if (!schedstat_enabled())
1004 * Are we enqueueing a waiting task? (for current tasks
1005 * a dequeue/enqueue event is a NOP)
1007 if (se != cfs_rq->curr)
1008 update_stats_wait_start(cfs_rq, se);
1010 if (flags & ENQUEUE_WAKEUP)
1011 update_stats_enqueue_sleeper(cfs_rq, se);
1015 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
1018 if (!schedstat_enabled())
1022 * Mark the end of the wait period if dequeueing a
1025 if (se != cfs_rq->curr)
1026 update_stats_wait_end(cfs_rq, se);
1028 if ((flags & DEQUEUE_SLEEP) && entity_is_task(se)) {
1029 struct task_struct *tsk = task_of(se);
1031 if (tsk->state & TASK_INTERRUPTIBLE)
1032 __schedstat_set(se->statistics.sleep_start,
1033 rq_clock(rq_of(cfs_rq)));
1034 if (tsk->state & TASK_UNINTERRUPTIBLE)
1035 __schedstat_set(se->statistics.block_start,
1036 rq_clock(rq_of(cfs_rq)));
1041 * We are picking a new current task - update its stats:
1044 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
1047 * We are starting a new run period:
1049 se->exec_start = rq_clock_task(rq_of(cfs_rq));
1052 /**************************************************
1053 * Scheduling class queueing methods:
1056 #ifdef CONFIG_NUMA_BALANCING
1058 * Approximate time to scan a full NUMA task in ms. The task scan period is
1059 * calculated based on the tasks virtual memory size and
1060 * numa_balancing_scan_size.
1062 unsigned int sysctl_numa_balancing_scan_period_min = 1000;
1063 unsigned int sysctl_numa_balancing_scan_period_max = 60000;
1065 /* Portion of address space to scan in MB */
1066 unsigned int sysctl_numa_balancing_scan_size = 256;
1068 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
1069 unsigned int sysctl_numa_balancing_scan_delay = 1000;
1072 refcount_t refcount;
1074 spinlock_t lock; /* nr_tasks, tasks */
1079 struct rcu_head rcu;
1080 unsigned long total_faults;
1081 unsigned long max_faults_cpu;
1083 * Faults_cpu is used to decide whether memory should move
1084 * towards the CPU. As a consequence, these stats are weighted
1085 * more by CPU use than by memory faults.
1087 unsigned long *faults_cpu;
1088 unsigned long faults[];
1092 * For functions that can be called in multiple contexts that permit reading
1093 * ->numa_group (see struct task_struct for locking rules).
1095 static struct numa_group *deref_task_numa_group(struct task_struct *p)
1097 return rcu_dereference_check(p->numa_group, p == current ||
1098 (lockdep_is_held(&task_rq(p)->lock) && !READ_ONCE(p->on_cpu)));
1101 static struct numa_group *deref_curr_numa_group(struct task_struct *p)
1103 return rcu_dereference_protected(p->numa_group, p == current);
1106 static inline unsigned long group_faults_priv(struct numa_group *ng);
1107 static inline unsigned long group_faults_shared(struct numa_group *ng);
1109 static unsigned int task_nr_scan_windows(struct task_struct *p)
1111 unsigned long rss = 0;
1112 unsigned long nr_scan_pages;
1115 * Calculations based on RSS as non-present and empty pages are skipped
1116 * by the PTE scanner and NUMA hinting faults should be trapped based
1119 nr_scan_pages = sysctl_numa_balancing_scan_size << (20 - PAGE_SHIFT);
1120 rss = get_mm_rss(p->mm);
1122 rss = nr_scan_pages;
1124 rss = round_up(rss, nr_scan_pages);
1125 return rss / nr_scan_pages;
1128 /* For sanitys sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */
1129 #define MAX_SCAN_WINDOW 2560
1131 static unsigned int task_scan_min(struct task_struct *p)
1133 unsigned int scan_size = READ_ONCE(sysctl_numa_balancing_scan_size);
1134 unsigned int scan, floor;
1135 unsigned int windows = 1;
1137 if (scan_size < MAX_SCAN_WINDOW)
1138 windows = MAX_SCAN_WINDOW / scan_size;
1139 floor = 1000 / windows;
1141 scan = sysctl_numa_balancing_scan_period_min / task_nr_scan_windows(p);
1142 return max_t(unsigned int, floor, scan);
1145 static unsigned int task_scan_start(struct task_struct *p)
1147 unsigned long smin = task_scan_min(p);
1148 unsigned long period = smin;
1149 struct numa_group *ng;
1151 /* Scale the maximum scan period with the amount of shared memory. */
1153 ng = rcu_dereference(p->numa_group);
1155 unsigned long shared = group_faults_shared(ng);
1156 unsigned long private = group_faults_priv(ng);
1158 period *= refcount_read(&ng->refcount);
1159 period *= shared + 1;
1160 period /= private + shared + 1;
1164 return max(smin, period);
1167 static unsigned int task_scan_max(struct task_struct *p)
1169 unsigned long smin = task_scan_min(p);
1171 struct numa_group *ng;
1173 /* Watch for min being lower than max due to floor calculations */
1174 smax = sysctl_numa_balancing_scan_period_max / task_nr_scan_windows(p);
1176 /* Scale the maximum scan period with the amount of shared memory. */
1177 ng = deref_curr_numa_group(p);
1179 unsigned long shared = group_faults_shared(ng);
1180 unsigned long private = group_faults_priv(ng);
1181 unsigned long period = smax;
1183 period *= refcount_read(&ng->refcount);
1184 period *= shared + 1;
1185 period /= private + shared + 1;
1187 smax = max(smax, period);
1190 return max(smin, smax);
1193 static void account_numa_enqueue(struct rq *rq, struct task_struct *p)
1195 rq->nr_numa_running += (p->numa_preferred_nid != NUMA_NO_NODE);
1196 rq->nr_preferred_running += (p->numa_preferred_nid == task_node(p));
1199 static void account_numa_dequeue(struct rq *rq, struct task_struct *p)
1201 rq->nr_numa_running -= (p->numa_preferred_nid != NUMA_NO_NODE);
1202 rq->nr_preferred_running -= (p->numa_preferred_nid == task_node(p));
1205 /* Shared or private faults. */
1206 #define NR_NUMA_HINT_FAULT_TYPES 2
1208 /* Memory and CPU locality */
1209 #define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2)
1211 /* Averaged statistics, and temporary buffers. */
1212 #define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2)
1214 pid_t task_numa_group_id(struct task_struct *p)
1216 struct numa_group *ng;
1220 ng = rcu_dereference(p->numa_group);
1229 * The averaged statistics, shared & private, memory & CPU,
1230 * occupy the first half of the array. The second half of the
1231 * array is for current counters, which are averaged into the
1232 * first set by task_numa_placement.
1234 static inline int task_faults_idx(enum numa_faults_stats s, int nid, int priv)
1236 return NR_NUMA_HINT_FAULT_TYPES * (s * nr_node_ids + nid) + priv;
1239 static inline unsigned long task_faults(struct task_struct *p, int nid)
1241 if (!p->numa_faults)
1244 return p->numa_faults[task_faults_idx(NUMA_MEM, nid, 0)] +
1245 p->numa_faults[task_faults_idx(NUMA_MEM, nid, 1)];
1248 static inline unsigned long group_faults(struct task_struct *p, int nid)
1250 struct numa_group *ng = deref_task_numa_group(p);
1255 return ng->faults[task_faults_idx(NUMA_MEM, nid, 0)] +
1256 ng->faults[task_faults_idx(NUMA_MEM, nid, 1)];
1259 static inline unsigned long group_faults_cpu(struct numa_group *group, int nid)
1261 return group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 0)] +
1262 group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 1)];
1265 static inline unsigned long group_faults_priv(struct numa_group *ng)
1267 unsigned long faults = 0;
1270 for_each_online_node(node) {
1271 faults += ng->faults[task_faults_idx(NUMA_MEM, node, 1)];
1277 static inline unsigned long group_faults_shared(struct numa_group *ng)
1279 unsigned long faults = 0;
1282 for_each_online_node(node) {
1283 faults += ng->faults[task_faults_idx(NUMA_MEM, node, 0)];
1290 * A node triggering more than 1/3 as many NUMA faults as the maximum is
1291 * considered part of a numa group's pseudo-interleaving set. Migrations
1292 * between these nodes are slowed down, to allow things to settle down.
1294 #define ACTIVE_NODE_FRACTION 3
1296 static bool numa_is_active_node(int nid, struct numa_group *ng)
1298 return group_faults_cpu(ng, nid) * ACTIVE_NODE_FRACTION > ng->max_faults_cpu;
1301 /* Handle placement on systems where not all nodes are directly connected. */
1302 static unsigned long score_nearby_nodes(struct task_struct *p, int nid,
1303 int maxdist, bool task)
1305 unsigned long score = 0;
1309 * All nodes are directly connected, and the same distance
1310 * from each other. No need for fancy placement algorithms.
1312 if (sched_numa_topology_type == NUMA_DIRECT)
1316 * This code is called for each node, introducing N^2 complexity,
1317 * which should be ok given the number of nodes rarely exceeds 8.
1319 for_each_online_node(node) {
1320 unsigned long faults;
1321 int dist = node_distance(nid, node);
1324 * The furthest away nodes in the system are not interesting
1325 * for placement; nid was already counted.
1327 if (dist == sched_max_numa_distance || node == nid)
1331 * On systems with a backplane NUMA topology, compare groups
1332 * of nodes, and move tasks towards the group with the most
1333 * memory accesses. When comparing two nodes at distance
1334 * "hoplimit", only nodes closer by than "hoplimit" are part
1335 * of each group. Skip other nodes.
1337 if (sched_numa_topology_type == NUMA_BACKPLANE &&
1341 /* Add up the faults from nearby nodes. */
1343 faults = task_faults(p, node);
1345 faults = group_faults(p, node);
1348 * On systems with a glueless mesh NUMA topology, there are
1349 * no fixed "groups of nodes". Instead, nodes that are not
1350 * directly connected bounce traffic through intermediate
1351 * nodes; a numa_group can occupy any set of nodes.
1352 * The further away a node is, the less the faults count.
1353 * This seems to result in good task placement.
1355 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
1356 faults *= (sched_max_numa_distance - dist);
1357 faults /= (sched_max_numa_distance - LOCAL_DISTANCE);
1367 * These return the fraction of accesses done by a particular task, or
1368 * task group, on a particular numa node. The group weight is given a
1369 * larger multiplier, in order to group tasks together that are almost
1370 * evenly spread out between numa nodes.
1372 static inline unsigned long task_weight(struct task_struct *p, int nid,
1375 unsigned long faults, total_faults;
1377 if (!p->numa_faults)
1380 total_faults = p->total_numa_faults;
1385 faults = task_faults(p, nid);
1386 faults += score_nearby_nodes(p, nid, dist, true);
1388 return 1000 * faults / total_faults;
1391 static inline unsigned long group_weight(struct task_struct *p, int nid,
1394 struct numa_group *ng = deref_task_numa_group(p);
1395 unsigned long faults, total_faults;
1400 total_faults = ng->total_faults;
1405 faults = group_faults(p, nid);
1406 faults += score_nearby_nodes(p, nid, dist, false);
1408 return 1000 * faults / total_faults;
1411 bool should_numa_migrate_memory(struct task_struct *p, struct page * page,
1412 int src_nid, int dst_cpu)
1414 struct numa_group *ng = deref_curr_numa_group(p);
1415 int dst_nid = cpu_to_node(dst_cpu);
1416 int last_cpupid, this_cpupid;
1418 this_cpupid = cpu_pid_to_cpupid(dst_cpu, current->pid);
1419 last_cpupid = page_cpupid_xchg_last(page, this_cpupid);
1422 * Allow first faults or private faults to migrate immediately early in
1423 * the lifetime of a task. The magic number 4 is based on waiting for
1424 * two full passes of the "multi-stage node selection" test that is
1427 if ((p->numa_preferred_nid == NUMA_NO_NODE || p->numa_scan_seq <= 4) &&
1428 (cpupid_pid_unset(last_cpupid) || cpupid_match_pid(p, last_cpupid)))
1432 * Multi-stage node selection is used in conjunction with a periodic
1433 * migration fault to build a temporal task<->page relation. By using
1434 * a two-stage filter we remove short/unlikely relations.
1436 * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate
1437 * a task's usage of a particular page (n_p) per total usage of this
1438 * page (n_t) (in a given time-span) to a probability.
1440 * Our periodic faults will sample this probability and getting the
1441 * same result twice in a row, given these samples are fully
1442 * independent, is then given by P(n)^2, provided our sample period
1443 * is sufficiently short compared to the usage pattern.
1445 * This quadric squishes small probabilities, making it less likely we
1446 * act on an unlikely task<->page relation.
1448 if (!cpupid_pid_unset(last_cpupid) &&
1449 cpupid_to_nid(last_cpupid) != dst_nid)
1452 /* Always allow migrate on private faults */
1453 if (cpupid_match_pid(p, last_cpupid))
1456 /* A shared fault, but p->numa_group has not been set up yet. */
1461 * Destination node is much more heavily used than the source
1462 * node? Allow migration.
1464 if (group_faults_cpu(ng, dst_nid) > group_faults_cpu(ng, src_nid) *
1465 ACTIVE_NODE_FRACTION)
1469 * Distribute memory according to CPU & memory use on each node,
1470 * with 3/4 hysteresis to avoid unnecessary memory migrations:
1472 * faults_cpu(dst) 3 faults_cpu(src)
1473 * --------------- * - > ---------------
1474 * faults_mem(dst) 4 faults_mem(src)
1476 return group_faults_cpu(ng, dst_nid) * group_faults(p, src_nid) * 3 >
1477 group_faults_cpu(ng, src_nid) * group_faults(p, dst_nid) * 4;
1481 * 'numa_type' describes the node at the moment of load balancing.
1484 /* The node has spare capacity that can be used to run more tasks. */
1487 * The node is fully used and the tasks don't compete for more CPU
1488 * cycles. Nevertheless, some tasks might wait before running.
1492 * The node is overloaded and can't provide expected CPU cycles to all
1498 /* Cached statistics for all CPUs within a node */
1501 unsigned long runnable;
1503 /* Total compute capacity of CPUs on a node */
1504 unsigned long compute_capacity;
1505 unsigned int nr_running;
1506 unsigned int weight;
1507 enum numa_type node_type;
1511 static inline bool is_core_idle(int cpu)
1513 #ifdef CONFIG_SCHED_SMT
1516 for_each_cpu(sibling, cpu_smt_mask(cpu)) {
1528 struct task_numa_env {
1529 struct task_struct *p;
1531 int src_cpu, src_nid;
1532 int dst_cpu, dst_nid;
1534 struct numa_stats src_stats, dst_stats;
1539 struct task_struct *best_task;
1544 static unsigned long cpu_load(struct rq *rq);
1545 static unsigned long cpu_runnable(struct rq *rq);
1546 static unsigned long cpu_util(int cpu);
1547 static inline long adjust_numa_imbalance(int imbalance,
1548 int dst_running, int dst_weight);
1551 numa_type numa_classify(unsigned int imbalance_pct,
1552 struct numa_stats *ns)
1554 if ((ns->nr_running > ns->weight) &&
1555 (((ns->compute_capacity * 100) < (ns->util * imbalance_pct)) ||
1556 ((ns->compute_capacity * imbalance_pct) < (ns->runnable * 100))))
1557 return node_overloaded;
1559 if ((ns->nr_running < ns->weight) ||
1560 (((ns->compute_capacity * 100) > (ns->util * imbalance_pct)) &&
1561 ((ns->compute_capacity * imbalance_pct) > (ns->runnable * 100))))
1562 return node_has_spare;
1564 return node_fully_busy;
1567 #ifdef CONFIG_SCHED_SMT
1568 /* Forward declarations of select_idle_sibling helpers */
1569 static inline bool test_idle_cores(int cpu, bool def);
1570 static inline int numa_idle_core(int idle_core, int cpu)
1572 if (!static_branch_likely(&sched_smt_present) ||
1573 idle_core >= 0 || !test_idle_cores(cpu, false))
1577 * Prefer cores instead of packing HT siblings
1578 * and triggering future load balancing.
1580 if (is_core_idle(cpu))
1586 static inline int numa_idle_core(int idle_core, int cpu)
1593 * Gather all necessary information to make NUMA balancing placement
1594 * decisions that are compatible with standard load balancer. This
1595 * borrows code and logic from update_sg_lb_stats but sharing a
1596 * common implementation is impractical.
1598 static void update_numa_stats(struct task_numa_env *env,
1599 struct numa_stats *ns, int nid,
1602 int cpu, idle_core = -1;
1604 memset(ns, 0, sizeof(*ns));
1608 for_each_cpu(cpu, cpumask_of_node(nid)) {
1609 struct rq *rq = cpu_rq(cpu);
1611 ns->load += cpu_load(rq);
1612 ns->runnable += cpu_runnable(rq);
1613 ns->util += cpu_util(cpu);
1614 ns->nr_running += rq->cfs.h_nr_running;
1615 ns->compute_capacity += capacity_of(cpu);
1617 if (find_idle && !rq->nr_running && idle_cpu(cpu)) {
1618 if (READ_ONCE(rq->numa_migrate_on) ||
1619 !cpumask_test_cpu(cpu, env->p->cpus_ptr))
1622 if (ns->idle_cpu == -1)
1625 idle_core = numa_idle_core(idle_core, cpu);
1630 ns->weight = cpumask_weight(cpumask_of_node(nid));
1632 ns->node_type = numa_classify(env->imbalance_pct, ns);
1635 ns->idle_cpu = idle_core;
1638 static void task_numa_assign(struct task_numa_env *env,
1639 struct task_struct *p, long imp)
1641 struct rq *rq = cpu_rq(env->dst_cpu);
1643 /* Check if run-queue part of active NUMA balance. */
1644 if (env->best_cpu != env->dst_cpu && xchg(&rq->numa_migrate_on, 1)) {
1646 int start = env->dst_cpu;
1648 /* Find alternative idle CPU. */
1649 for_each_cpu_wrap(cpu, cpumask_of_node(env->dst_nid), start) {
1650 if (cpu == env->best_cpu || !idle_cpu(cpu) ||
1651 !cpumask_test_cpu(cpu, env->p->cpus_ptr)) {
1656 rq = cpu_rq(env->dst_cpu);
1657 if (!xchg(&rq->numa_migrate_on, 1))
1661 /* Failed to find an alternative idle CPU */
1667 * Clear previous best_cpu/rq numa-migrate flag, since task now
1668 * found a better CPU to move/swap.
1670 if (env->best_cpu != -1 && env->best_cpu != env->dst_cpu) {
1671 rq = cpu_rq(env->best_cpu);
1672 WRITE_ONCE(rq->numa_migrate_on, 0);
1676 put_task_struct(env->best_task);
1681 env->best_imp = imp;
1682 env->best_cpu = env->dst_cpu;
1685 static bool load_too_imbalanced(long src_load, long dst_load,
1686 struct task_numa_env *env)
1689 long orig_src_load, orig_dst_load;
1690 long src_capacity, dst_capacity;
1693 * The load is corrected for the CPU capacity available on each node.
1696 * ------------ vs ---------
1697 * src_capacity dst_capacity
1699 src_capacity = env->src_stats.compute_capacity;
1700 dst_capacity = env->dst_stats.compute_capacity;
1702 imb = abs(dst_load * src_capacity - src_load * dst_capacity);
1704 orig_src_load = env->src_stats.load;
1705 orig_dst_load = env->dst_stats.load;
1707 old_imb = abs(orig_dst_load * src_capacity - orig_src_load * dst_capacity);
1709 /* Would this change make things worse? */
1710 return (imb > old_imb);
1714 * Maximum NUMA importance can be 1998 (2*999);
1715 * SMALLIMP @ 30 would be close to 1998/64.
1716 * Used to deter task migration.
1721 * This checks if the overall compute and NUMA accesses of the system would
1722 * be improved if the source tasks was migrated to the target dst_cpu taking
1723 * into account that it might be best if task running on the dst_cpu should
1724 * be exchanged with the source task
1726 static bool task_numa_compare(struct task_numa_env *env,
1727 long taskimp, long groupimp, bool maymove)
1729 struct numa_group *cur_ng, *p_ng = deref_curr_numa_group(env->p);
1730 struct rq *dst_rq = cpu_rq(env->dst_cpu);
1731 long imp = p_ng ? groupimp : taskimp;
1732 struct task_struct *cur;
1733 long src_load, dst_load;
1734 int dist = env->dist;
1737 bool stopsearch = false;
1739 if (READ_ONCE(dst_rq->numa_migrate_on))
1743 cur = rcu_dereference(dst_rq->curr);
1744 if (cur && ((cur->flags & PF_EXITING) || is_idle_task(cur)))
1748 * Because we have preemption enabled we can get migrated around and
1749 * end try selecting ourselves (current == env->p) as a swap candidate.
1751 if (cur == env->p) {
1757 if (maymove && moveimp >= env->best_imp)
1763 /* Skip this swap candidate if cannot move to the source cpu. */
1764 if (!cpumask_test_cpu(env->src_cpu, cur->cpus_ptr))
1768 * Skip this swap candidate if it is not moving to its preferred
1769 * node and the best task is.
1771 if (env->best_task &&
1772 env->best_task->numa_preferred_nid == env->src_nid &&
1773 cur->numa_preferred_nid != env->src_nid) {
1778 * "imp" is the fault differential for the source task between the
1779 * source and destination node. Calculate the total differential for
1780 * the source task and potential destination task. The more negative
1781 * the value is, the more remote accesses that would be expected to
1782 * be incurred if the tasks were swapped.
1784 * If dst and source tasks are in the same NUMA group, or not
1785 * in any group then look only at task weights.
1787 cur_ng = rcu_dereference(cur->numa_group);
1788 if (cur_ng == p_ng) {
1789 imp = taskimp + task_weight(cur, env->src_nid, dist) -
1790 task_weight(cur, env->dst_nid, dist);
1792 * Add some hysteresis to prevent swapping the
1793 * tasks within a group over tiny differences.
1799 * Compare the group weights. If a task is all by itself
1800 * (not part of a group), use the task weight instead.
1803 imp += group_weight(cur, env->src_nid, dist) -
1804 group_weight(cur, env->dst_nid, dist);
1806 imp += task_weight(cur, env->src_nid, dist) -
1807 task_weight(cur, env->dst_nid, dist);
1810 /* Discourage picking a task already on its preferred node */
1811 if (cur->numa_preferred_nid == env->dst_nid)
1815 * Encourage picking a task that moves to its preferred node.
1816 * This potentially makes imp larger than it's maximum of
1817 * 1998 (see SMALLIMP and task_weight for why) but in this
1818 * case, it does not matter.
1820 if (cur->numa_preferred_nid == env->src_nid)
1823 if (maymove && moveimp > imp && moveimp > env->best_imp) {
1830 * Prefer swapping with a task moving to its preferred node over a
1833 if (env->best_task && cur->numa_preferred_nid == env->src_nid &&
1834 env->best_task->numa_preferred_nid != env->src_nid) {
1839 * If the NUMA importance is less than SMALLIMP,
1840 * task migration might only result in ping pong
1841 * of tasks and also hurt performance due to cache
1844 if (imp < SMALLIMP || imp <= env->best_imp + SMALLIMP / 2)
1848 * In the overloaded case, try and keep the load balanced.
1850 load = task_h_load(env->p) - task_h_load(cur);
1854 dst_load = env->dst_stats.load + load;
1855 src_load = env->src_stats.load - load;
1857 if (load_too_imbalanced(src_load, dst_load, env))
1861 /* Evaluate an idle CPU for a task numa move. */
1863 int cpu = env->dst_stats.idle_cpu;
1865 /* Nothing cached so current CPU went idle since the search. */
1870 * If the CPU is no longer truly idle and the previous best CPU
1871 * is, keep using it.
1873 if (!idle_cpu(cpu) && env->best_cpu >= 0 &&
1874 idle_cpu(env->best_cpu)) {
1875 cpu = env->best_cpu;
1881 task_numa_assign(env, cur, imp);
1884 * If a move to idle is allowed because there is capacity or load
1885 * balance improves then stop the search. While a better swap
1886 * candidate may exist, a search is not free.
1888 if (maymove && !cur && env->best_cpu >= 0 && idle_cpu(env->best_cpu))
1892 * If a swap candidate must be identified and the current best task
1893 * moves its preferred node then stop the search.
1895 if (!maymove && env->best_task &&
1896 env->best_task->numa_preferred_nid == env->src_nid) {
1905 static void task_numa_find_cpu(struct task_numa_env *env,
1906 long taskimp, long groupimp)
1908 bool maymove = false;
1912 * If dst node has spare capacity, then check if there is an
1913 * imbalance that would be overruled by the load balancer.
1915 if (env->dst_stats.node_type == node_has_spare) {
1916 unsigned int imbalance;
1917 int src_running, dst_running;
1920 * Would movement cause an imbalance? Note that if src has
1921 * more running tasks that the imbalance is ignored as the
1922 * move improves the imbalance from the perspective of the
1923 * CPU load balancer.
1925 src_running = env->src_stats.nr_running - 1;
1926 dst_running = env->dst_stats.nr_running + 1;
1927 imbalance = max(0, dst_running - src_running);
1928 imbalance = adjust_numa_imbalance(imbalance, dst_running,
1929 env->dst_stats.weight);
1931 /* Use idle CPU if there is no imbalance */
1934 if (env->dst_stats.idle_cpu >= 0) {
1935 env->dst_cpu = env->dst_stats.idle_cpu;
1936 task_numa_assign(env, NULL, 0);
1941 long src_load, dst_load, load;
1943 * If the improvement from just moving env->p direction is better
1944 * than swapping tasks around, check if a move is possible.
1946 load = task_h_load(env->p);
1947 dst_load = env->dst_stats.load + load;
1948 src_load = env->src_stats.load - load;
1949 maymove = !load_too_imbalanced(src_load, dst_load, env);
1952 for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) {
1953 /* Skip this CPU if the source task cannot migrate */
1954 if (!cpumask_test_cpu(cpu, env->p->cpus_ptr))
1958 if (task_numa_compare(env, taskimp, groupimp, maymove))
1963 static int task_numa_migrate(struct task_struct *p)
1965 struct task_numa_env env = {
1968 .src_cpu = task_cpu(p),
1969 .src_nid = task_node(p),
1971 .imbalance_pct = 112,
1977 unsigned long taskweight, groupweight;
1978 struct sched_domain *sd;
1979 long taskimp, groupimp;
1980 struct numa_group *ng;
1985 * Pick the lowest SD_NUMA domain, as that would have the smallest
1986 * imbalance and would be the first to start moving tasks about.
1988 * And we want to avoid any moving of tasks about, as that would create
1989 * random movement of tasks -- counter the numa conditions we're trying
1993 sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu));
1995 env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2;
1999 * Cpusets can break the scheduler domain tree into smaller
2000 * balance domains, some of which do not cross NUMA boundaries.
2001 * Tasks that are "trapped" in such domains cannot be migrated
2002 * elsewhere, so there is no point in (re)trying.
2004 if (unlikely(!sd)) {
2005 sched_setnuma(p, task_node(p));
2009 env.dst_nid = p->numa_preferred_nid;
2010 dist = env.dist = node_distance(env.src_nid, env.dst_nid);
2011 taskweight = task_weight(p, env.src_nid, dist);
2012 groupweight = group_weight(p, env.src_nid, dist);
2013 update_numa_stats(&env, &env.src_stats, env.src_nid, false);
2014 taskimp = task_weight(p, env.dst_nid, dist) - taskweight;
2015 groupimp = group_weight(p, env.dst_nid, dist) - groupweight;
2016 update_numa_stats(&env, &env.dst_stats, env.dst_nid, true);
2018 /* Try to find a spot on the preferred nid. */
2019 task_numa_find_cpu(&env, taskimp, groupimp);
2022 * Look at other nodes in these cases:
2023 * - there is no space available on the preferred_nid
2024 * - the task is part of a numa_group that is interleaved across
2025 * multiple NUMA nodes; in order to better consolidate the group,
2026 * we need to check other locations.
2028 ng = deref_curr_numa_group(p);
2029 if (env.best_cpu == -1 || (ng && ng->active_nodes > 1)) {
2030 for_each_online_node(nid) {
2031 if (nid == env.src_nid || nid == p->numa_preferred_nid)
2034 dist = node_distance(env.src_nid, env.dst_nid);
2035 if (sched_numa_topology_type == NUMA_BACKPLANE &&
2037 taskweight = task_weight(p, env.src_nid, dist);
2038 groupweight = group_weight(p, env.src_nid, dist);
2041 /* Only consider nodes where both task and groups benefit */
2042 taskimp = task_weight(p, nid, dist) - taskweight;
2043 groupimp = group_weight(p, nid, dist) - groupweight;
2044 if (taskimp < 0 && groupimp < 0)
2049 update_numa_stats(&env, &env.dst_stats, env.dst_nid, true);
2050 task_numa_find_cpu(&env, taskimp, groupimp);
2055 * If the task is part of a workload that spans multiple NUMA nodes,
2056 * and is migrating into one of the workload's active nodes, remember
2057 * this node as the task's preferred numa node, so the workload can
2059 * A task that migrated to a second choice node will be better off
2060 * trying for a better one later. Do not set the preferred node here.
2063 if (env.best_cpu == -1)
2066 nid = cpu_to_node(env.best_cpu);
2068 if (nid != p->numa_preferred_nid)
2069 sched_setnuma(p, nid);
2072 /* No better CPU than the current one was found. */
2073 if (env.best_cpu == -1) {
2074 trace_sched_stick_numa(p, env.src_cpu, NULL, -1);
2078 best_rq = cpu_rq(env.best_cpu);
2079 if (env.best_task == NULL) {
2080 ret = migrate_task_to(p, env.best_cpu);
2081 WRITE_ONCE(best_rq->numa_migrate_on, 0);
2083 trace_sched_stick_numa(p, env.src_cpu, NULL, env.best_cpu);
2087 ret = migrate_swap(p, env.best_task, env.best_cpu, env.src_cpu);
2088 WRITE_ONCE(best_rq->numa_migrate_on, 0);
2091 trace_sched_stick_numa(p, env.src_cpu, env.best_task, env.best_cpu);
2092 put_task_struct(env.best_task);
2096 /* Attempt to migrate a task to a CPU on the preferred node. */
2097 static void numa_migrate_preferred(struct task_struct *p)
2099 unsigned long interval = HZ;
2101 /* This task has no NUMA fault statistics yet */
2102 if (unlikely(p->numa_preferred_nid == NUMA_NO_NODE || !p->numa_faults))
2105 /* Periodically retry migrating the task to the preferred node */
2106 interval = min(interval, msecs_to_jiffies(p->numa_scan_period) / 16);
2107 p->numa_migrate_retry = jiffies + interval;
2109 /* Success if task is already running on preferred CPU */
2110 if (task_node(p) == p->numa_preferred_nid)
2113 /* Otherwise, try migrate to a CPU on the preferred node */
2114 task_numa_migrate(p);
2118 * Find out how many nodes on the workload is actively running on. Do this by
2119 * tracking the nodes from which NUMA hinting faults are triggered. This can
2120 * be different from the set of nodes where the workload's memory is currently
2123 static void numa_group_count_active_nodes(struct numa_group *numa_group)
2125 unsigned long faults, max_faults = 0;
2126 int nid, active_nodes = 0;
2128 for_each_online_node(nid) {
2129 faults = group_faults_cpu(numa_group, nid);
2130 if (faults > max_faults)
2131 max_faults = faults;
2134 for_each_online_node(nid) {
2135 faults = group_faults_cpu(numa_group, nid);
2136 if (faults * ACTIVE_NODE_FRACTION > max_faults)
2140 numa_group->max_faults_cpu = max_faults;
2141 numa_group->active_nodes = active_nodes;
2145 * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
2146 * increments. The more local the fault statistics are, the higher the scan
2147 * period will be for the next scan window. If local/(local+remote) ratio is
2148 * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS)
2149 * the scan period will decrease. Aim for 70% local accesses.
2151 #define NUMA_PERIOD_SLOTS 10
2152 #define NUMA_PERIOD_THRESHOLD 7
2155 * Increase the scan period (slow down scanning) if the majority of
2156 * our memory is already on our local node, or if the majority of
2157 * the page accesses are shared with other processes.
2158 * Otherwise, decrease the scan period.
2160 static void update_task_scan_period(struct task_struct *p,
2161 unsigned long shared, unsigned long private)
2163 unsigned int period_slot;
2164 int lr_ratio, ps_ratio;
2167 unsigned long remote = p->numa_faults_locality[0];
2168 unsigned long local = p->numa_faults_locality[1];
2171 * If there were no record hinting faults then either the task is
2172 * completely idle or all activity is areas that are not of interest
2173 * to automatic numa balancing. Related to that, if there were failed
2174 * migration then it implies we are migrating too quickly or the local
2175 * node is overloaded. In either case, scan slower
2177 if (local + shared == 0 || p->numa_faults_locality[2]) {
2178 p->numa_scan_period = min(p->numa_scan_period_max,
2179 p->numa_scan_period << 1);
2181 p->mm->numa_next_scan = jiffies +
2182 msecs_to_jiffies(p->numa_scan_period);
2188 * Prepare to scale scan period relative to the current period.
2189 * == NUMA_PERIOD_THRESHOLD scan period stays the same
2190 * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
2191 * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
2193 period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS);
2194 lr_ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote);
2195 ps_ratio = (private * NUMA_PERIOD_SLOTS) / (private + shared);
2197 if (ps_ratio >= NUMA_PERIOD_THRESHOLD) {
2199 * Most memory accesses are local. There is no need to
2200 * do fast NUMA scanning, since memory is already local.
2202 int slot = ps_ratio - NUMA_PERIOD_THRESHOLD;
2205 diff = slot * period_slot;
2206 } else if (lr_ratio >= NUMA_PERIOD_THRESHOLD) {
2208 * Most memory accesses are shared with other tasks.
2209 * There is no point in continuing fast NUMA scanning,
2210 * since other tasks may just move the memory elsewhere.
2212 int slot = lr_ratio - NUMA_PERIOD_THRESHOLD;
2215 diff = slot * period_slot;
2218 * Private memory faults exceed (SLOTS-THRESHOLD)/SLOTS,
2219 * yet they are not on the local NUMA node. Speed up
2220 * NUMA scanning to get the memory moved over.
2222 int ratio = max(lr_ratio, ps_ratio);
2223 diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot;
2226 p->numa_scan_period = clamp(p->numa_scan_period + diff,
2227 task_scan_min(p), task_scan_max(p));
2228 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
2232 * Get the fraction of time the task has been running since the last
2233 * NUMA placement cycle. The scheduler keeps similar statistics, but
2234 * decays those on a 32ms period, which is orders of magnitude off
2235 * from the dozens-of-seconds NUMA balancing period. Use the scheduler
2236 * stats only if the task is so new there are no NUMA statistics yet.
2238 static u64 numa_get_avg_runtime(struct task_struct *p, u64 *period)
2240 u64 runtime, delta, now;
2241 /* Use the start of this time slice to avoid calculations. */
2242 now = p->se.exec_start;
2243 runtime = p->se.sum_exec_runtime;
2245 if (p->last_task_numa_placement) {
2246 delta = runtime - p->last_sum_exec_runtime;
2247 *period = now - p->last_task_numa_placement;
2249 /* Avoid time going backwards, prevent potential divide error: */
2250 if (unlikely((s64)*period < 0))
2253 delta = p->se.avg.load_sum;
2254 *period = LOAD_AVG_MAX;
2257 p->last_sum_exec_runtime = runtime;
2258 p->last_task_numa_placement = now;
2264 * Determine the preferred nid for a task in a numa_group. This needs to
2265 * be done in a way that produces consistent results with group_weight,
2266 * otherwise workloads might not converge.
2268 static int preferred_group_nid(struct task_struct *p, int nid)
2273 /* Direct connections between all NUMA nodes. */
2274 if (sched_numa_topology_type == NUMA_DIRECT)
2278 * On a system with glueless mesh NUMA topology, group_weight
2279 * scores nodes according to the number of NUMA hinting faults on
2280 * both the node itself, and on nearby nodes.
2282 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
2283 unsigned long score, max_score = 0;
2284 int node, max_node = nid;
2286 dist = sched_max_numa_distance;
2288 for_each_online_node(node) {
2289 score = group_weight(p, node, dist);
2290 if (score > max_score) {
2299 * Finding the preferred nid in a system with NUMA backplane
2300 * interconnect topology is more involved. The goal is to locate
2301 * tasks from numa_groups near each other in the system, and
2302 * untangle workloads from different sides of the system. This requires
2303 * searching down the hierarchy of node groups, recursively searching
2304 * inside the highest scoring group of nodes. The nodemask tricks
2305 * keep the complexity of the search down.
2307 nodes = node_online_map;
2308 for (dist = sched_max_numa_distance; dist > LOCAL_DISTANCE; dist--) {
2309 unsigned long max_faults = 0;
2310 nodemask_t max_group = NODE_MASK_NONE;
2313 /* Are there nodes at this distance from each other? */
2314 if (!find_numa_distance(dist))
2317 for_each_node_mask(a, nodes) {
2318 unsigned long faults = 0;
2319 nodemask_t this_group;
2320 nodes_clear(this_group);
2322 /* Sum group's NUMA faults; includes a==b case. */
2323 for_each_node_mask(b, nodes) {
2324 if (node_distance(a, b) < dist) {
2325 faults += group_faults(p, b);
2326 node_set(b, this_group);
2327 node_clear(b, nodes);
2331 /* Remember the top group. */
2332 if (faults > max_faults) {
2333 max_faults = faults;
2334 max_group = this_group;
2336 * subtle: at the smallest distance there is
2337 * just one node left in each "group", the
2338 * winner is the preferred nid.
2343 /* Next round, evaluate the nodes within max_group. */
2351 static void task_numa_placement(struct task_struct *p)
2353 int seq, nid, max_nid = NUMA_NO_NODE;
2354 unsigned long max_faults = 0;
2355 unsigned long fault_types[2] = { 0, 0 };
2356 unsigned long total_faults;
2357 u64 runtime, period;
2358 spinlock_t *group_lock = NULL;
2359 struct numa_group *ng;
2362 * The p->mm->numa_scan_seq field gets updated without
2363 * exclusive access. Use READ_ONCE() here to ensure
2364 * that the field is read in a single access:
2366 seq = READ_ONCE(p->mm->numa_scan_seq);
2367 if (p->numa_scan_seq == seq)
2369 p->numa_scan_seq = seq;
2370 p->numa_scan_period_max = task_scan_max(p);
2372 total_faults = p->numa_faults_locality[0] +
2373 p->numa_faults_locality[1];
2374 runtime = numa_get_avg_runtime(p, &period);
2376 /* If the task is part of a group prevent parallel updates to group stats */
2377 ng = deref_curr_numa_group(p);
2379 group_lock = &ng->lock;
2380 spin_lock_irq(group_lock);
2383 /* Find the node with the highest number of faults */
2384 for_each_online_node(nid) {
2385 /* Keep track of the offsets in numa_faults array */
2386 int mem_idx, membuf_idx, cpu_idx, cpubuf_idx;
2387 unsigned long faults = 0, group_faults = 0;
2390 for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) {
2391 long diff, f_diff, f_weight;
2393 mem_idx = task_faults_idx(NUMA_MEM, nid, priv);
2394 membuf_idx = task_faults_idx(NUMA_MEMBUF, nid, priv);
2395 cpu_idx = task_faults_idx(NUMA_CPU, nid, priv);
2396 cpubuf_idx = task_faults_idx(NUMA_CPUBUF, nid, priv);
2398 /* Decay existing window, copy faults since last scan */
2399 diff = p->numa_faults[membuf_idx] - p->numa_faults[mem_idx] / 2;
2400 fault_types[priv] += p->numa_faults[membuf_idx];
2401 p->numa_faults[membuf_idx] = 0;
2404 * Normalize the faults_from, so all tasks in a group
2405 * count according to CPU use, instead of by the raw
2406 * number of faults. Tasks with little runtime have
2407 * little over-all impact on throughput, and thus their
2408 * faults are less important.
2410 f_weight = div64_u64(runtime << 16, period + 1);
2411 f_weight = (f_weight * p->numa_faults[cpubuf_idx]) /
2413 f_diff = f_weight - p->numa_faults[cpu_idx] / 2;
2414 p->numa_faults[cpubuf_idx] = 0;
2416 p->numa_faults[mem_idx] += diff;
2417 p->numa_faults[cpu_idx] += f_diff;
2418 faults += p->numa_faults[mem_idx];
2419 p->total_numa_faults += diff;
2422 * safe because we can only change our own group
2424 * mem_idx represents the offset for a given
2425 * nid and priv in a specific region because it
2426 * is at the beginning of the numa_faults array.
2428 ng->faults[mem_idx] += diff;
2429 ng->faults_cpu[mem_idx] += f_diff;
2430 ng->total_faults += diff;
2431 group_faults += ng->faults[mem_idx];
2436 if (faults > max_faults) {
2437 max_faults = faults;
2440 } else if (group_faults > max_faults) {
2441 max_faults = group_faults;
2447 numa_group_count_active_nodes(ng);
2448 spin_unlock_irq(group_lock);
2449 max_nid = preferred_group_nid(p, max_nid);
2453 /* Set the new preferred node */
2454 if (max_nid != p->numa_preferred_nid)
2455 sched_setnuma(p, max_nid);
2458 update_task_scan_period(p, fault_types[0], fault_types[1]);
2461 static inline int get_numa_group(struct numa_group *grp)
2463 return refcount_inc_not_zero(&grp->refcount);
2466 static inline void put_numa_group(struct numa_group *grp)
2468 if (refcount_dec_and_test(&grp->refcount))
2469 kfree_rcu(grp, rcu);
2472 static void task_numa_group(struct task_struct *p, int cpupid, int flags,
2475 struct numa_group *grp, *my_grp;
2476 struct task_struct *tsk;
2478 int cpu = cpupid_to_cpu(cpupid);
2481 if (unlikely(!deref_curr_numa_group(p))) {
2482 unsigned int size = sizeof(struct numa_group) +
2483 4*nr_node_ids*sizeof(unsigned long);
2485 grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
2489 refcount_set(&grp->refcount, 1);
2490 grp->active_nodes = 1;
2491 grp->max_faults_cpu = 0;
2492 spin_lock_init(&grp->lock);
2494 /* Second half of the array tracks nids where faults happen */
2495 grp->faults_cpu = grp->faults + NR_NUMA_HINT_FAULT_TYPES *
2498 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2499 grp->faults[i] = p->numa_faults[i];
2501 grp->total_faults = p->total_numa_faults;
2504 rcu_assign_pointer(p->numa_group, grp);
2508 tsk = READ_ONCE(cpu_rq(cpu)->curr);
2510 if (!cpupid_match_pid(tsk, cpupid))
2513 grp = rcu_dereference(tsk->numa_group);
2517 my_grp = deref_curr_numa_group(p);
2522 * Only join the other group if its bigger; if we're the bigger group,
2523 * the other task will join us.
2525 if (my_grp->nr_tasks > grp->nr_tasks)
2529 * Tie-break on the grp address.
2531 if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp)
2534 /* Always join threads in the same process. */
2535 if (tsk->mm == current->mm)
2538 /* Simple filter to avoid false positives due to PID collisions */
2539 if (flags & TNF_SHARED)
2542 /* Update priv based on whether false sharing was detected */
2545 if (join && !get_numa_group(grp))
2553 BUG_ON(irqs_disabled());
2554 double_lock_irq(&my_grp->lock, &grp->lock);
2556 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) {
2557 my_grp->faults[i] -= p->numa_faults[i];
2558 grp->faults[i] += p->numa_faults[i];
2560 my_grp->total_faults -= p->total_numa_faults;
2561 grp->total_faults += p->total_numa_faults;
2566 spin_unlock(&my_grp->lock);
2567 spin_unlock_irq(&grp->lock);
2569 rcu_assign_pointer(p->numa_group, grp);
2571 put_numa_group(my_grp);
2580 * Get rid of NUMA staticstics associated with a task (either current or dead).
2581 * If @final is set, the task is dead and has reached refcount zero, so we can
2582 * safely free all relevant data structures. Otherwise, there might be
2583 * concurrent reads from places like load balancing and procfs, and we should
2584 * reset the data back to default state without freeing ->numa_faults.
2586 void task_numa_free(struct task_struct *p, bool final)
2588 /* safe: p either is current or is being freed by current */
2589 struct numa_group *grp = rcu_dereference_raw(p->numa_group);
2590 unsigned long *numa_faults = p->numa_faults;
2591 unsigned long flags;
2598 spin_lock_irqsave(&grp->lock, flags);
2599 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2600 grp->faults[i] -= p->numa_faults[i];
2601 grp->total_faults -= p->total_numa_faults;
2604 spin_unlock_irqrestore(&grp->lock, flags);
2605 RCU_INIT_POINTER(p->numa_group, NULL);
2606 put_numa_group(grp);
2610 p->numa_faults = NULL;
2613 p->total_numa_faults = 0;
2614 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2620 * Got a PROT_NONE fault for a page on @node.
2622 void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags)
2624 struct task_struct *p = current;
2625 bool migrated = flags & TNF_MIGRATED;
2626 int cpu_node = task_node(current);
2627 int local = !!(flags & TNF_FAULT_LOCAL);
2628 struct numa_group *ng;
2631 if (!static_branch_likely(&sched_numa_balancing))
2634 /* for example, ksmd faulting in a user's mm */
2638 /* Allocate buffer to track faults on a per-node basis */
2639 if (unlikely(!p->numa_faults)) {
2640 int size = sizeof(*p->numa_faults) *
2641 NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids;
2643 p->numa_faults = kzalloc(size, GFP_KERNEL|__GFP_NOWARN);
2644 if (!p->numa_faults)
2647 p->total_numa_faults = 0;
2648 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
2652 * First accesses are treated as private, otherwise consider accesses
2653 * to be private if the accessing pid has not changed
2655 if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) {
2658 priv = cpupid_match_pid(p, last_cpupid);
2659 if (!priv && !(flags & TNF_NO_GROUP))
2660 task_numa_group(p, last_cpupid, flags, &priv);
2664 * If a workload spans multiple NUMA nodes, a shared fault that
2665 * occurs wholly within the set of nodes that the workload is
2666 * actively using should be counted as local. This allows the
2667 * scan rate to slow down when a workload has settled down.
2669 ng = deref_curr_numa_group(p);
2670 if (!priv && !local && ng && ng->active_nodes > 1 &&
2671 numa_is_active_node(cpu_node, ng) &&
2672 numa_is_active_node(mem_node, ng))
2676 * Retry to migrate task to preferred node periodically, in case it
2677 * previously failed, or the scheduler moved us.
2679 if (time_after(jiffies, p->numa_migrate_retry)) {
2680 task_numa_placement(p);
2681 numa_migrate_preferred(p);
2685 p->numa_pages_migrated += pages;
2686 if (flags & TNF_MIGRATE_FAIL)
2687 p->numa_faults_locality[2] += pages;
2689 p->numa_faults[task_faults_idx(NUMA_MEMBUF, mem_node, priv)] += pages;
2690 p->numa_faults[task_faults_idx(NUMA_CPUBUF, cpu_node, priv)] += pages;
2691 p->numa_faults_locality[local] += pages;
2694 static void reset_ptenuma_scan(struct task_struct *p)
2697 * We only did a read acquisition of the mmap sem, so
2698 * p->mm->numa_scan_seq is written to without exclusive access
2699 * and the update is not guaranteed to be atomic. That's not
2700 * much of an issue though, since this is just used for
2701 * statistical sampling. Use READ_ONCE/WRITE_ONCE, which are not
2702 * expensive, to avoid any form of compiler optimizations:
2704 WRITE_ONCE(p->mm->numa_scan_seq, READ_ONCE(p->mm->numa_scan_seq) + 1);
2705 p->mm->numa_scan_offset = 0;
2709 * The expensive part of numa migration is done from task_work context.
2710 * Triggered from task_tick_numa().
2712 static void task_numa_work(struct callback_head *work)
2714 unsigned long migrate, next_scan, now = jiffies;
2715 struct task_struct *p = current;
2716 struct mm_struct *mm = p->mm;
2717 u64 runtime = p->se.sum_exec_runtime;
2718 struct vm_area_struct *vma;
2719 unsigned long start, end;
2720 unsigned long nr_pte_updates = 0;
2721 long pages, virtpages;
2723 SCHED_WARN_ON(p != container_of(work, struct task_struct, numa_work));
2727 * Who cares about NUMA placement when they're dying.
2729 * NOTE: make sure not to dereference p->mm before this check,
2730 * exit_task_work() happens _after_ exit_mm() so we could be called
2731 * without p->mm even though we still had it when we enqueued this
2734 if (p->flags & PF_EXITING)
2737 if (!mm->numa_next_scan) {
2738 mm->numa_next_scan = now +
2739 msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2743 * Enforce maximal scan/migration frequency..
2745 migrate = mm->numa_next_scan;
2746 if (time_before(now, migrate))
2749 if (p->numa_scan_period == 0) {
2750 p->numa_scan_period_max = task_scan_max(p);
2751 p->numa_scan_period = task_scan_start(p);
2754 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
2755 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
2759 * Delay this task enough that another task of this mm will likely win
2760 * the next time around.
2762 p->node_stamp += 2 * TICK_NSEC;
2764 start = mm->numa_scan_offset;
2765 pages = sysctl_numa_balancing_scan_size;
2766 pages <<= 20 - PAGE_SHIFT; /* MB in pages */
2767 virtpages = pages * 8; /* Scan up to this much virtual space */
2772 if (!mmap_read_trylock(mm))
2774 vma = find_vma(mm, start);
2776 reset_ptenuma_scan(p);
2780 for (; vma; vma = vma->vm_next) {
2781 if (!vma_migratable(vma) || !vma_policy_mof(vma) ||
2782 is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_MIXEDMAP)) {
2787 * Shared library pages mapped by multiple processes are not
2788 * migrated as it is expected they are cache replicated. Avoid
2789 * hinting faults in read-only file-backed mappings or the vdso
2790 * as migrating the pages will be of marginal benefit.
2793 (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ)))
2797 * Skip inaccessible VMAs to avoid any confusion between
2798 * PROT_NONE and NUMA hinting ptes
2800 if (!vma_is_accessible(vma))
2804 start = max(start, vma->vm_start);
2805 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
2806 end = min(end, vma->vm_end);
2807 nr_pte_updates = change_prot_numa(vma, start, end);
2810 * Try to scan sysctl_numa_balancing_size worth of
2811 * hpages that have at least one present PTE that
2812 * is not already pte-numa. If the VMA contains
2813 * areas that are unused or already full of prot_numa
2814 * PTEs, scan up to virtpages, to skip through those
2818 pages -= (end - start) >> PAGE_SHIFT;
2819 virtpages -= (end - start) >> PAGE_SHIFT;
2822 if (pages <= 0 || virtpages <= 0)
2826 } while (end != vma->vm_end);
2831 * It is possible to reach the end of the VMA list but the last few
2832 * VMAs are not guaranteed to the vma_migratable. If they are not, we
2833 * would find the !migratable VMA on the next scan but not reset the
2834 * scanner to the start so check it now.
2837 mm->numa_scan_offset = start;
2839 reset_ptenuma_scan(p);
2840 mmap_read_unlock(mm);
2843 * Make sure tasks use at least 32x as much time to run other code
2844 * than they used here, to limit NUMA PTE scanning overhead to 3% max.
2845 * Usually update_task_scan_period slows down scanning enough; on an
2846 * overloaded system we need to limit overhead on a per task basis.
2848 if (unlikely(p->se.sum_exec_runtime != runtime)) {
2849 u64 diff = p->se.sum_exec_runtime - runtime;
2850 p->node_stamp += 32 * diff;
2854 void init_numa_balancing(unsigned long clone_flags, struct task_struct *p)
2857 struct mm_struct *mm = p->mm;
2860 mm_users = atomic_read(&mm->mm_users);
2861 if (mm_users == 1) {
2862 mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2863 mm->numa_scan_seq = 0;
2867 p->numa_scan_seq = mm ? mm->numa_scan_seq : 0;
2868 p->numa_scan_period = sysctl_numa_balancing_scan_delay;
2869 /* Protect against double add, see task_tick_numa and task_numa_work */
2870 p->numa_work.next = &p->numa_work;
2871 p->numa_faults = NULL;
2872 RCU_INIT_POINTER(p->numa_group, NULL);
2873 p->last_task_numa_placement = 0;
2874 p->last_sum_exec_runtime = 0;
2876 init_task_work(&p->numa_work, task_numa_work);
2878 /* New address space, reset the preferred nid */
2879 if (!(clone_flags & CLONE_VM)) {
2880 p->numa_preferred_nid = NUMA_NO_NODE;
2885 * New thread, keep existing numa_preferred_nid which should be copied
2886 * already by arch_dup_task_struct but stagger when scans start.
2891 delay = min_t(unsigned int, task_scan_max(current),
2892 current->numa_scan_period * mm_users * NSEC_PER_MSEC);
2893 delay += 2 * TICK_NSEC;
2894 p->node_stamp = delay;
2899 * Drive the periodic memory faults..
2901 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
2903 struct callback_head *work = &curr->numa_work;
2907 * We don't care about NUMA placement if we don't have memory.
2909 if ((curr->flags & (PF_EXITING | PF_KTHREAD)) || work->next != work)
2913 * Using runtime rather than walltime has the dual advantage that
2914 * we (mostly) drive the selection from busy threads and that the
2915 * task needs to have done some actual work before we bother with
2918 now = curr->se.sum_exec_runtime;
2919 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
2921 if (now > curr->node_stamp + period) {
2922 if (!curr->node_stamp)
2923 curr->numa_scan_period = task_scan_start(curr);
2924 curr->node_stamp += period;
2926 if (!time_before(jiffies, curr->mm->numa_next_scan))
2927 task_work_add(curr, work, TWA_RESUME);
2931 static void update_scan_period(struct task_struct *p, int new_cpu)
2933 int src_nid = cpu_to_node(task_cpu(p));
2934 int dst_nid = cpu_to_node(new_cpu);
2936 if (!static_branch_likely(&sched_numa_balancing))
2939 if (!p->mm || !p->numa_faults || (p->flags & PF_EXITING))
2942 if (src_nid == dst_nid)
2946 * Allow resets if faults have been trapped before one scan
2947 * has completed. This is most likely due to a new task that
2948 * is pulled cross-node due to wakeups or load balancing.
2950 if (p->numa_scan_seq) {
2952 * Avoid scan adjustments if moving to the preferred
2953 * node or if the task was not previously running on
2954 * the preferred node.
2956 if (dst_nid == p->numa_preferred_nid ||
2957 (p->numa_preferred_nid != NUMA_NO_NODE &&
2958 src_nid != p->numa_preferred_nid))
2962 p->numa_scan_period = task_scan_start(p);
2966 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
2970 static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p)
2974 static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p)
2978 static inline void update_scan_period(struct task_struct *p, int new_cpu)
2982 #endif /* CONFIG_NUMA_BALANCING */
2985 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2987 update_load_add(&cfs_rq->load, se->load.weight);
2989 if (entity_is_task(se)) {
2990 struct rq *rq = rq_of(cfs_rq);
2992 account_numa_enqueue(rq, task_of(se));
2993 list_add(&se->group_node, &rq->cfs_tasks);
2996 cfs_rq->nr_running++;
3000 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
3002 update_load_sub(&cfs_rq->load, se->load.weight);
3004 if (entity_is_task(se)) {
3005 account_numa_dequeue(rq_of(cfs_rq), task_of(se));
3006 list_del_init(&se->group_node);
3009 cfs_rq->nr_running--;
3013 * Signed add and clamp on underflow.
3015 * Explicitly do a load-store to ensure the intermediate value never hits
3016 * memory. This allows lockless observations without ever seeing the negative
3019 #define add_positive(_ptr, _val) do { \
3020 typeof(_ptr) ptr = (_ptr); \
3021 typeof(_val) val = (_val); \
3022 typeof(*ptr) res, var = READ_ONCE(*ptr); \
3026 if (val < 0 && res > var) \
3029 WRITE_ONCE(*ptr, res); \
3033 * Unsigned subtract and clamp on underflow.
3035 * Explicitly do a load-store to ensure the intermediate value never hits
3036 * memory. This allows lockless observations without ever seeing the negative
3039 #define sub_positive(_ptr, _val) do { \
3040 typeof(_ptr) ptr = (_ptr); \
3041 typeof(*ptr) val = (_val); \
3042 typeof(*ptr) res, var = READ_ONCE(*ptr); \
3046 WRITE_ONCE(*ptr, res); \
3050 * Remove and clamp on negative, from a local variable.
3052 * A variant of sub_positive(), which does not use explicit load-store
3053 * and is thus optimized for local variable updates.
3055 #define lsub_positive(_ptr, _val) do { \
3056 typeof(_ptr) ptr = (_ptr); \
3057 *ptr -= min_t(typeof(*ptr), *ptr, _val); \
3062 enqueue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3064 cfs_rq->avg.load_avg += se->avg.load_avg;
3065 cfs_rq->avg.load_sum += se_weight(se) * se->avg.load_sum;
3069 dequeue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3071 sub_positive(&cfs_rq->avg.load_avg, se->avg.load_avg);
3072 sub_positive(&cfs_rq->avg.load_sum, se_weight(se) * se->avg.load_sum);
3076 enqueue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
3078 dequeue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
3081 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
3082 unsigned long weight)
3085 /* commit outstanding execution time */
3086 if (cfs_rq->curr == se)
3087 update_curr(cfs_rq);
3088 update_load_sub(&cfs_rq->load, se->load.weight);
3090 dequeue_load_avg(cfs_rq, se);
3092 update_load_set(&se->load, weight);
3096 u32 divider = get_pelt_divider(&se->avg);
3098 se->avg.load_avg = div_u64(se_weight(se) * se->avg.load_sum, divider);
3102 enqueue_load_avg(cfs_rq, se);
3104 update_load_add(&cfs_rq->load, se->load.weight);
3108 void reweight_task(struct task_struct *p, int prio)
3110 struct sched_entity *se = &p->se;
3111 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3112 struct load_weight *load = &se->load;
3113 unsigned long weight = scale_load(sched_prio_to_weight[prio]);
3115 reweight_entity(cfs_rq, se, weight);
3116 load->inv_weight = sched_prio_to_wmult[prio];
3119 #ifdef CONFIG_FAIR_GROUP_SCHED
3122 * All this does is approximate the hierarchical proportion which includes that
3123 * global sum we all love to hate.
3125 * That is, the weight of a group entity, is the proportional share of the
3126 * group weight based on the group runqueue weights. That is:
3128 * tg->weight * grq->load.weight
3129 * ge->load.weight = ----------------------------- (1)
3130 * \Sum grq->load.weight
3132 * Now, because computing that sum is prohibitively expensive to compute (been
3133 * there, done that) we approximate it with this average stuff. The average
3134 * moves slower and therefore the approximation is cheaper and more stable.
3136 * So instead of the above, we substitute:
3138 * grq->load.weight -> grq->avg.load_avg (2)
3140 * which yields the following:
3142 * tg->weight * grq->avg.load_avg
3143 * ge->load.weight = ------------------------------ (3)
3146 * Where: tg->load_avg ~= \Sum grq->avg.load_avg
3148 * That is shares_avg, and it is right (given the approximation (2)).
3150 * The problem with it is that because the average is slow -- it was designed
3151 * to be exactly that of course -- this leads to transients in boundary
3152 * conditions. In specific, the case where the group was idle and we start the
3153 * one task. It takes time for our CPU's grq->avg.load_avg to build up,
3154 * yielding bad latency etc..
3156 * Now, in that special case (1) reduces to:
3158 * tg->weight * grq->load.weight
3159 * ge->load.weight = ----------------------------- = tg->weight (4)
3162 * That is, the sum collapses because all other CPUs are idle; the UP scenario.
3164 * So what we do is modify our approximation (3) to approach (4) in the (near)
3169 * tg->weight * grq->load.weight
3170 * --------------------------------------------------- (5)
3171 * tg->load_avg - grq->avg.load_avg + grq->load.weight
3173 * But because grq->load.weight can drop to 0, resulting in a divide by zero,
3174 * we need to use grq->avg.load_avg as its lower bound, which then gives:
3177 * tg->weight * grq->load.weight
3178 * ge->load.weight = ----------------------------- (6)
3183 * tg_load_avg' = tg->load_avg - grq->avg.load_avg +
3184 * max(grq->load.weight, grq->avg.load_avg)
3186 * And that is shares_weight and is icky. In the (near) UP case it approaches
3187 * (4) while in the normal case it approaches (3). It consistently
3188 * overestimates the ge->load.weight and therefore:
3190 * \Sum ge->load.weight >= tg->weight
3194 static long calc_group_shares(struct cfs_rq *cfs_rq)
3196 long tg_weight, tg_shares, load, shares;
3197 struct task_group *tg = cfs_rq->tg;
3199 tg_shares = READ_ONCE(tg->shares);
3201 load = max(scale_load_down(cfs_rq->load.weight), cfs_rq->avg.load_avg);
3203 tg_weight = atomic_long_read(&tg->load_avg);
3205 /* Ensure tg_weight >= load */
3206 tg_weight -= cfs_rq->tg_load_avg_contrib;
3209 shares = (tg_shares * load);
3211 shares /= tg_weight;
3214 * MIN_SHARES has to be unscaled here to support per-CPU partitioning
3215 * of a group with small tg->shares value. It is a floor value which is
3216 * assigned as a minimum load.weight to the sched_entity representing
3217 * the group on a CPU.
3219 * E.g. on 64-bit for a group with tg->shares of scale_load(15)=15*1024
3220 * on an 8-core system with 8 tasks each runnable on one CPU shares has
3221 * to be 15*1024*1/8=1920 instead of scale_load(MIN_SHARES)=2*1024. In
3222 * case no task is runnable on a CPU MIN_SHARES=2 should be returned
3225 return clamp_t(long, shares, MIN_SHARES, tg_shares);
3227 #endif /* CONFIG_SMP */
3229 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
3232 * Recomputes the group entity based on the current state of its group
3235 static void update_cfs_group(struct sched_entity *se)
3237 struct cfs_rq *gcfs_rq = group_cfs_rq(se);
3243 if (throttled_hierarchy(gcfs_rq))
3247 shares = READ_ONCE(gcfs_rq->tg->shares);
3249 if (likely(se->load.weight == shares))
3252 shares = calc_group_shares(gcfs_rq);
3255 reweight_entity(cfs_rq_of(se), se, shares);
3258 #else /* CONFIG_FAIR_GROUP_SCHED */
3259 static inline void update_cfs_group(struct sched_entity *se)
3262 #endif /* CONFIG_FAIR_GROUP_SCHED */
3264 static inline void cfs_rq_util_change(struct cfs_rq *cfs_rq, int flags)
3266 struct rq *rq = rq_of(cfs_rq);
3268 if (&rq->cfs == cfs_rq) {
3270 * There are a few boundary cases this might miss but it should
3271 * get called often enough that that should (hopefully) not be
3274 * It will not get called when we go idle, because the idle
3275 * thread is a different class (!fair), nor will the utilization
3276 * number include things like RT tasks.
3278 * As is, the util number is not freq-invariant (we'd have to
3279 * implement arch_scale_freq_capacity() for that).
3283 cpufreq_update_util(rq, flags);
3288 #ifdef CONFIG_FAIR_GROUP_SCHED
3290 * update_tg_load_avg - update the tg's load avg
3291 * @cfs_rq: the cfs_rq whose avg changed
3293 * This function 'ensures': tg->load_avg := \Sum tg->cfs_rq[]->avg.load.
3294 * However, because tg->load_avg is a global value there are performance
3297 * In order to avoid having to look at the other cfs_rq's, we use a
3298 * differential update where we store the last value we propagated. This in
3299 * turn allows skipping updates if the differential is 'small'.
3301 * Updating tg's load_avg is necessary before update_cfs_share().
3303 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq)
3305 long delta = cfs_rq->avg.load_avg - cfs_rq->tg_load_avg_contrib;
3308 * No need to update load_avg for root_task_group as it is not used.
3310 if (cfs_rq->tg == &root_task_group)
3313 if (abs(delta) > cfs_rq->tg_load_avg_contrib / 64) {
3314 atomic_long_add(delta, &cfs_rq->tg->load_avg);
3315 cfs_rq->tg_load_avg_contrib = cfs_rq->avg.load_avg;
3320 * Called within set_task_rq() right before setting a task's CPU. The
3321 * caller only guarantees p->pi_lock is held; no other assumptions,
3322 * including the state of rq->lock, should be made.
3324 void set_task_rq_fair(struct sched_entity *se,
3325 struct cfs_rq *prev, struct cfs_rq *next)
3327 u64 p_last_update_time;
3328 u64 n_last_update_time;
3330 if (!sched_feat(ATTACH_AGE_LOAD))
3334 * We are supposed to update the task to "current" time, then its up to
3335 * date and ready to go to new CPU/cfs_rq. But we have difficulty in
3336 * getting what current time is, so simply throw away the out-of-date
3337 * time. This will result in the wakee task is less decayed, but giving
3338 * the wakee more load sounds not bad.
3340 if (!(se->avg.last_update_time && prev))
3343 #ifndef CONFIG_64BIT
3345 u64 p_last_update_time_copy;
3346 u64 n_last_update_time_copy;
3349 p_last_update_time_copy = prev->load_last_update_time_copy;
3350 n_last_update_time_copy = next->load_last_update_time_copy;
3354 p_last_update_time = prev->avg.last_update_time;
3355 n_last_update_time = next->avg.last_update_time;
3357 } while (p_last_update_time != p_last_update_time_copy ||
3358 n_last_update_time != n_last_update_time_copy);
3361 p_last_update_time = prev->avg.last_update_time;
3362 n_last_update_time = next->avg.last_update_time;
3364 __update_load_avg_blocked_se(p_last_update_time, se);
3365 se->avg.last_update_time = n_last_update_time;
3370 * When on migration a sched_entity joins/leaves the PELT hierarchy, we need to
3371 * propagate its contribution. The key to this propagation is the invariant
3372 * that for each group:
3374 * ge->avg == grq->avg (1)
3376 * _IFF_ we look at the pure running and runnable sums. Because they
3377 * represent the very same entity, just at different points in the hierarchy.
3379 * Per the above update_tg_cfs_util() and update_tg_cfs_runnable() are trivial
3380 * and simply copies the running/runnable sum over (but still wrong, because
3381 * the group entity and group rq do not have their PELT windows aligned).
3383 * However, update_tg_cfs_load() is more complex. So we have:
3385 * ge->avg.load_avg = ge->load.weight * ge->avg.runnable_avg (2)
3387 * And since, like util, the runnable part should be directly transferable,
3388 * the following would _appear_ to be the straight forward approach:
3390 * grq->avg.load_avg = grq->load.weight * grq->avg.runnable_avg (3)
3392 * And per (1) we have:
3394 * ge->avg.runnable_avg == grq->avg.runnable_avg
3398 * ge->load.weight * grq->avg.load_avg
3399 * ge->avg.load_avg = ----------------------------------- (4)
3402 * Except that is wrong!
3404 * Because while for entities historical weight is not important and we
3405 * really only care about our future and therefore can consider a pure
3406 * runnable sum, runqueues can NOT do this.
3408 * We specifically want runqueues to have a load_avg that includes
3409 * historical weights. Those represent the blocked load, the load we expect
3410 * to (shortly) return to us. This only works by keeping the weights as
3411 * integral part of the sum. We therefore cannot decompose as per (3).
3413 * Another reason this doesn't work is that runnable isn't a 0-sum entity.
3414 * Imagine a rq with 2 tasks that each are runnable 2/3 of the time. Then the
3415 * rq itself is runnable anywhere between 2/3 and 1 depending on how the
3416 * runnable section of these tasks overlap (or not). If they were to perfectly
3417 * align the rq as a whole would be runnable 2/3 of the time. If however we
3418 * always have at least 1 runnable task, the rq as a whole is always runnable.
3420 * So we'll have to approximate.. :/
3422 * Given the constraint:
3424 * ge->avg.running_sum <= ge->avg.runnable_sum <= LOAD_AVG_MAX
3426 * We can construct a rule that adds runnable to a rq by assuming minimal
3429 * On removal, we'll assume each task is equally runnable; which yields:
3431 * grq->avg.runnable_sum = grq->avg.load_sum / grq->load.weight
3433 * XXX: only do this for the part of runnable > running ?
3438 update_tg_cfs_util(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq)
3440 long delta = gcfs_rq->avg.util_avg - se->avg.util_avg;
3443 /* Nothing to update */
3448 * cfs_rq->avg.period_contrib can be used for both cfs_rq and se.
3449 * See ___update_load_avg() for details.
3451 divider = get_pelt_divider(&cfs_rq->avg);
3453 /* Set new sched_entity's utilization */
3454 se->avg.util_avg = gcfs_rq->avg.util_avg;
3455 se->avg.util_sum = se->avg.util_avg * divider;
3457 /* Update parent cfs_rq utilization */
3458 add_positive(&cfs_rq->avg.util_avg, delta);
3459 cfs_rq->avg.util_sum = cfs_rq->avg.util_avg * divider;
3463 update_tg_cfs_runnable(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq)
3465 long delta = gcfs_rq->avg.runnable_avg - se->avg.runnable_avg;
3468 /* Nothing to update */
3473 * cfs_rq->avg.period_contrib can be used for both cfs_rq and se.
3474 * See ___update_load_avg() for details.
3476 divider = get_pelt_divider(&cfs_rq->avg);
3478 /* Set new sched_entity's runnable */
3479 se->avg.runnable_avg = gcfs_rq->avg.runnable_avg;
3480 se->avg.runnable_sum = se->avg.runnable_avg * divider;
3482 /* Update parent cfs_rq runnable */
3483 add_positive(&cfs_rq->avg.runnable_avg, delta);
3484 cfs_rq->avg.runnable_sum = cfs_rq->avg.runnable_avg * divider;
3488 update_tg_cfs_load(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq)
3490 long delta_avg, running_sum, runnable_sum = gcfs_rq->prop_runnable_sum;
3491 unsigned long load_avg;
3499 gcfs_rq->prop_runnable_sum = 0;
3502 * cfs_rq->avg.period_contrib can be used for both cfs_rq and se.
3503 * See ___update_load_avg() for details.
3505 divider = get_pelt_divider(&cfs_rq->avg);
3507 if (runnable_sum >= 0) {
3509 * Add runnable; clip at LOAD_AVG_MAX. Reflects that until
3510 * the CPU is saturated running == runnable.
3512 runnable_sum += se->avg.load_sum;
3513 runnable_sum = min_t(long, runnable_sum, divider);
3516 * Estimate the new unweighted runnable_sum of the gcfs_rq by
3517 * assuming all tasks are equally runnable.
3519 if (scale_load_down(gcfs_rq->load.weight)) {
3520 load_sum = div_s64(gcfs_rq->avg.load_sum,
3521 scale_load_down(gcfs_rq->load.weight));
3524 /* But make sure to not inflate se's runnable */
3525 runnable_sum = min(se->avg.load_sum, load_sum);
3529 * runnable_sum can't be lower than running_sum
3530 * Rescale running sum to be in the same range as runnable sum
3531 * running_sum is in [0 : LOAD_AVG_MAX << SCHED_CAPACITY_SHIFT]
3532 * runnable_sum is in [0 : LOAD_AVG_MAX]
3534 running_sum = se->avg.util_sum >> SCHED_CAPACITY_SHIFT;
3535 runnable_sum = max(runnable_sum, running_sum);
3537 load_sum = (s64)se_weight(se) * runnable_sum;
3538 load_avg = div_s64(load_sum, divider);
3540 delta_sum = load_sum - (s64)se_weight(se) * se->avg.load_sum;
3541 delta_avg = load_avg - se->avg.load_avg;
3543 se->avg.load_sum = runnable_sum;
3544 se->avg.load_avg = load_avg;
3545 add_positive(&cfs_rq->avg.load_avg, delta_avg);
3546 add_positive(&cfs_rq->avg.load_sum, delta_sum);
3549 static inline void add_tg_cfs_propagate(struct cfs_rq *cfs_rq, long runnable_sum)
3551 cfs_rq->propagate = 1;
3552 cfs_rq->prop_runnable_sum += runnable_sum;
3555 /* Update task and its cfs_rq load average */
3556 static inline int propagate_entity_load_avg(struct sched_entity *se)
3558 struct cfs_rq *cfs_rq, *gcfs_rq;
3560 if (entity_is_task(se))
3563 gcfs_rq = group_cfs_rq(se);
3564 if (!gcfs_rq->propagate)
3567 gcfs_rq->propagate = 0;
3569 cfs_rq = cfs_rq_of(se);
3571 add_tg_cfs_propagate(cfs_rq, gcfs_rq->prop_runnable_sum);
3573 update_tg_cfs_util(cfs_rq, se, gcfs_rq);
3574 update_tg_cfs_runnable(cfs_rq, se, gcfs_rq);
3575 update_tg_cfs_load(cfs_rq, se, gcfs_rq);
3577 trace_pelt_cfs_tp(cfs_rq);
3578 trace_pelt_se_tp(se);
3584 * Check if we need to update the load and the utilization of a blocked
3587 static inline bool skip_blocked_update(struct sched_entity *se)
3589 struct cfs_rq *gcfs_rq = group_cfs_rq(se);
3592 * If sched_entity still have not zero load or utilization, we have to
3595 if (se->avg.load_avg || se->avg.util_avg)
3599 * If there is a pending propagation, we have to update the load and
3600 * the utilization of the sched_entity:
3602 if (gcfs_rq->propagate)
3606 * Otherwise, the load and the utilization of the sched_entity is
3607 * already zero and there is no pending propagation, so it will be a
3608 * waste of time to try to decay it:
3613 #else /* CONFIG_FAIR_GROUP_SCHED */
3615 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq) {}
3617 static inline int propagate_entity_load_avg(struct sched_entity *se)
3622 static inline void add_tg_cfs_propagate(struct cfs_rq *cfs_rq, long runnable_sum) {}
3624 #endif /* CONFIG_FAIR_GROUP_SCHED */
3627 * update_cfs_rq_load_avg - update the cfs_rq's load/util averages
3628 * @now: current time, as per cfs_rq_clock_pelt()
3629 * @cfs_rq: cfs_rq to update
3631 * The cfs_rq avg is the direct sum of all its entities (blocked and runnable)
3632 * avg. The immediate corollary is that all (fair) tasks must be attached, see
3633 * post_init_entity_util_avg().
3635 * cfs_rq->avg is used for task_h_load() and update_cfs_share() for example.
3637 * Returns true if the load decayed or we removed load.
3639 * Since both these conditions indicate a changed cfs_rq->avg.load we should
3640 * call update_tg_load_avg() when this function returns true.
3643 update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq)
3645 unsigned long removed_load = 0, removed_util = 0, removed_runnable = 0;
3646 struct sched_avg *sa = &cfs_rq->avg;
3649 if (cfs_rq->removed.nr) {
3651 u32 divider = get_pelt_divider(&cfs_rq->avg);
3653 raw_spin_lock(&cfs_rq->removed.lock);
3654 swap(cfs_rq->removed.util_avg, removed_util);
3655 swap(cfs_rq->removed.load_avg, removed_load);
3656 swap(cfs_rq->removed.runnable_avg, removed_runnable);
3657 cfs_rq->removed.nr = 0;
3658 raw_spin_unlock(&cfs_rq->removed.lock);
3661 sub_positive(&sa->load_avg, r);
3662 sub_positive(&sa->load_sum, r * divider);
3665 sub_positive(&sa->util_avg, r);
3666 sub_positive(&sa->util_sum, r * divider);
3668 r = removed_runnable;
3669 sub_positive(&sa->runnable_avg, r);
3670 sub_positive(&sa->runnable_sum, r * divider);
3673 * removed_runnable is the unweighted version of removed_load so we
3674 * can use it to estimate removed_load_sum.
3676 add_tg_cfs_propagate(cfs_rq,
3677 -(long)(removed_runnable * divider) >> SCHED_CAPACITY_SHIFT);
3682 decayed |= __update_load_avg_cfs_rq(now, cfs_rq);
3684 #ifndef CONFIG_64BIT
3686 cfs_rq->load_last_update_time_copy = sa->last_update_time;
3693 * attach_entity_load_avg - attach this entity to its cfs_rq load avg
3694 * @cfs_rq: cfs_rq to attach to
3695 * @se: sched_entity to attach
3697 * Must call update_cfs_rq_load_avg() before this, since we rely on
3698 * cfs_rq->avg.last_update_time being current.
3700 static void attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3703 * cfs_rq->avg.period_contrib can be used for both cfs_rq and se.
3704 * See ___update_load_avg() for details.
3706 u32 divider = get_pelt_divider(&cfs_rq->avg);
3709 * When we attach the @se to the @cfs_rq, we must align the decay
3710 * window because without that, really weird and wonderful things can
3715 se->avg.last_update_time = cfs_rq->avg.last_update_time;
3716 se->avg.period_contrib = cfs_rq->avg.period_contrib;
3719 * Hell(o) Nasty stuff.. we need to recompute _sum based on the new
3720 * period_contrib. This isn't strictly correct, but since we're
3721 * entirely outside of the PELT hierarchy, nobody cares if we truncate
3724 se->avg.util_sum = se->avg.util_avg * divider;
3726 se->avg.runnable_sum = se->avg.runnable_avg * divider;
3728 se->avg.load_sum = divider;
3729 if (se_weight(se)) {
3731 div_u64(se->avg.load_avg * se->avg.load_sum, se_weight(se));
3734 enqueue_load_avg(cfs_rq, se);
3735 cfs_rq->avg.util_avg += se->avg.util_avg;
3736 cfs_rq->avg.util_sum += se->avg.util_sum;
3737 cfs_rq->avg.runnable_avg += se->avg.runnable_avg;
3738 cfs_rq->avg.runnable_sum += se->avg.runnable_sum;
3740 add_tg_cfs_propagate(cfs_rq, se->avg.load_sum);
3742 cfs_rq_util_change(cfs_rq, 0);
3744 trace_pelt_cfs_tp(cfs_rq);
3748 * detach_entity_load_avg - detach this entity from its cfs_rq load avg
3749 * @cfs_rq: cfs_rq to detach from
3750 * @se: sched_entity to detach
3752 * Must call update_cfs_rq_load_avg() before this, since we rely on
3753 * cfs_rq->avg.last_update_time being current.
3755 static void detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3757 dequeue_load_avg(cfs_rq, se);
3758 sub_positive(&cfs_rq->avg.util_avg, se->avg.util_avg);
3759 sub_positive(&cfs_rq->avg.util_sum, se->avg.util_sum);
3760 sub_positive(&cfs_rq->avg.runnable_avg, se->avg.runnable_avg);
3761 sub_positive(&cfs_rq->avg.runnable_sum, se->avg.runnable_sum);
3763 add_tg_cfs_propagate(cfs_rq, -se->avg.load_sum);
3765 cfs_rq_util_change(cfs_rq, 0);
3767 trace_pelt_cfs_tp(cfs_rq);
3771 * Optional action to be done while updating the load average
3773 #define UPDATE_TG 0x1
3774 #define SKIP_AGE_LOAD 0x2
3775 #define DO_ATTACH 0x4
3777 /* Update task and its cfs_rq load average */
3778 static inline void update_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3780 u64 now = cfs_rq_clock_pelt(cfs_rq);
3784 * Track task load average for carrying it to new CPU after migrated, and
3785 * track group sched_entity load average for task_h_load calc in migration
3787 if (se->avg.last_update_time && !(flags & SKIP_AGE_LOAD))
3788 __update_load_avg_se(now, cfs_rq, se);
3790 decayed = update_cfs_rq_load_avg(now, cfs_rq);
3791 decayed |= propagate_entity_load_avg(se);
3793 if (!se->avg.last_update_time && (flags & DO_ATTACH)) {
3796 * DO_ATTACH means we're here from enqueue_entity().
3797 * !last_update_time means we've passed through
3798 * migrate_task_rq_fair() indicating we migrated.
3800 * IOW we're enqueueing a task on a new CPU.
3802 attach_entity_load_avg(cfs_rq, se);
3803 update_tg_load_avg(cfs_rq);
3805 } else if (decayed) {
3806 cfs_rq_util_change(cfs_rq, 0);
3808 if (flags & UPDATE_TG)
3809 update_tg_load_avg(cfs_rq);
3813 #ifndef CONFIG_64BIT
3814 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
3816 u64 last_update_time_copy;
3817 u64 last_update_time;
3820 last_update_time_copy = cfs_rq->load_last_update_time_copy;
3822 last_update_time = cfs_rq->avg.last_update_time;
3823 } while (last_update_time != last_update_time_copy);
3825 return last_update_time;
3828 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
3830 return cfs_rq->avg.last_update_time;
3835 * Synchronize entity load avg of dequeued entity without locking
3838 static void sync_entity_load_avg(struct sched_entity *se)
3840 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3841 u64 last_update_time;
3843 last_update_time = cfs_rq_last_update_time(cfs_rq);
3844 __update_load_avg_blocked_se(last_update_time, se);
3848 * Task first catches up with cfs_rq, and then subtract
3849 * itself from the cfs_rq (task must be off the queue now).
3851 static void remove_entity_load_avg(struct sched_entity *se)
3853 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3854 unsigned long flags;
3857 * tasks cannot exit without having gone through wake_up_new_task() ->
3858 * post_init_entity_util_avg() which will have added things to the
3859 * cfs_rq, so we can remove unconditionally.
3862 sync_entity_load_avg(se);
3864 raw_spin_lock_irqsave(&cfs_rq->removed.lock, flags);
3865 ++cfs_rq->removed.nr;
3866 cfs_rq->removed.util_avg += se->avg.util_avg;
3867 cfs_rq->removed.load_avg += se->avg.load_avg;
3868 cfs_rq->removed.runnable_avg += se->avg.runnable_avg;
3869 raw_spin_unlock_irqrestore(&cfs_rq->removed.lock, flags);
3872 static inline unsigned long cfs_rq_runnable_avg(struct cfs_rq *cfs_rq)
3874 return cfs_rq->avg.runnable_avg;
3877 static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq)
3879 return cfs_rq->avg.load_avg;
3882 static int newidle_balance(struct rq *this_rq, struct rq_flags *rf);
3884 static inline unsigned long task_util(struct task_struct *p)
3886 return READ_ONCE(p->se.avg.util_avg);
3889 static inline unsigned long _task_util_est(struct task_struct *p)
3891 struct util_est ue = READ_ONCE(p->se.avg.util_est);
3893 return (max(ue.ewma, ue.enqueued) | UTIL_AVG_UNCHANGED);
3896 static inline unsigned long task_util_est(struct task_struct *p)
3898 return max(task_util(p), _task_util_est(p));
3901 #ifdef CONFIG_UCLAMP_TASK
3902 static inline unsigned long uclamp_task_util(struct task_struct *p)
3904 return clamp(task_util_est(p),
3905 uclamp_eff_value(p, UCLAMP_MIN),
3906 uclamp_eff_value(p, UCLAMP_MAX));
3909 static inline unsigned long uclamp_task_util(struct task_struct *p)
3911 return task_util_est(p);
3915 static inline void util_est_enqueue(struct cfs_rq *cfs_rq,
3916 struct task_struct *p)
3918 unsigned int enqueued;
3920 if (!sched_feat(UTIL_EST))
3923 /* Update root cfs_rq's estimated utilization */
3924 enqueued = cfs_rq->avg.util_est.enqueued;
3925 enqueued += _task_util_est(p);
3926 WRITE_ONCE(cfs_rq->avg.util_est.enqueued, enqueued);
3928 trace_sched_util_est_cfs_tp(cfs_rq);
3931 static inline void util_est_dequeue(struct cfs_rq *cfs_rq,
3932 struct task_struct *p)
3934 unsigned int enqueued;
3936 if (!sched_feat(UTIL_EST))
3939 /* Update root cfs_rq's estimated utilization */
3940 enqueued = cfs_rq->avg.util_est.enqueued;
3941 enqueued -= min_t(unsigned int, enqueued, _task_util_est(p));
3942 WRITE_ONCE(cfs_rq->avg.util_est.enqueued, enqueued);
3944 trace_sched_util_est_cfs_tp(cfs_rq);
3947 #define UTIL_EST_MARGIN (SCHED_CAPACITY_SCALE / 100)
3950 * Check if a (signed) value is within a specified (unsigned) margin,
3951 * based on the observation that:
3953 * abs(x) < y := (unsigned)(x + y - 1) < (2 * y - 1)
3955 * NOTE: this only works when value + maring < INT_MAX.
3957 static inline bool within_margin(int value, int margin)
3959 return ((unsigned int)(value + margin - 1) < (2 * margin - 1));
3962 static inline void util_est_update(struct cfs_rq *cfs_rq,
3963 struct task_struct *p,
3966 long last_ewma_diff, last_enqueued_diff;
3969 if (!sched_feat(UTIL_EST))
3973 * Skip update of task's estimated utilization when the task has not
3974 * yet completed an activation, e.g. being migrated.
3980 * If the PELT values haven't changed since enqueue time,
3981 * skip the util_est update.
3983 ue = p->se.avg.util_est;
3984 if (ue.enqueued & UTIL_AVG_UNCHANGED)
3987 last_enqueued_diff = ue.enqueued;
3990 * Reset EWMA on utilization increases, the moving average is used only
3991 * to smooth utilization decreases.
3993 ue.enqueued = (task_util(p) | UTIL_AVG_UNCHANGED);
3994 if (sched_feat(UTIL_EST_FASTUP)) {
3995 if (ue.ewma < ue.enqueued) {
3996 ue.ewma = ue.enqueued;
4002 * Skip update of task's estimated utilization when its members are
4003 * already ~1% close to its last activation value.
4005 last_ewma_diff = ue.enqueued - ue.ewma;
4006 last_enqueued_diff -= ue.enqueued;
4007 if (within_margin(last_ewma_diff, UTIL_EST_MARGIN)) {
4008 if (!within_margin(last_enqueued_diff, UTIL_EST_MARGIN))
4015 * To avoid overestimation of actual task utilization, skip updates if
4016 * we cannot grant there is idle time in this CPU.
4018 if (task_util(p) > capacity_orig_of(cpu_of(rq_of(cfs_rq))))
4022 * Update Task's estimated utilization
4024 * When *p completes an activation we can consolidate another sample
4025 * of the task size. This is done by storing the current PELT value
4026 * as ue.enqueued and by using this value to update the Exponential
4027 * Weighted Moving Average (EWMA):
4029 * ewma(t) = w * task_util(p) + (1-w) * ewma(t-1)
4030 * = w * task_util(p) + ewma(t-1) - w * ewma(t-1)
4031 * = w * (task_util(p) - ewma(t-1)) + ewma(t-1)
4032 * = w * ( last_ewma_diff ) + ewma(t-1)
4033 * = w * (last_ewma_diff + ewma(t-1) / w)
4035 * Where 'w' is the weight of new samples, which is configured to be
4036 * 0.25, thus making w=1/4 ( >>= UTIL_EST_WEIGHT_SHIFT)
4038 ue.ewma <<= UTIL_EST_WEIGHT_SHIFT;
4039 ue.ewma += last_ewma_diff;
4040 ue.ewma >>= UTIL_EST_WEIGHT_SHIFT;
4042 WRITE_ONCE(p->se.avg.util_est, ue);
4044 trace_sched_util_est_se_tp(&p->se);
4047 static inline int task_fits_capacity(struct task_struct *p, long capacity)
4049 return fits_capacity(uclamp_task_util(p), capacity);
4052 static inline void update_misfit_status(struct task_struct *p, struct rq *rq)
4054 if (!static_branch_unlikely(&sched_asym_cpucapacity))
4057 if (!p || p->nr_cpus_allowed == 1) {
4058 rq->misfit_task_load = 0;
4062 if (task_fits_capacity(p, capacity_of(cpu_of(rq)))) {
4063 rq->misfit_task_load = 0;
4068 * Make sure that misfit_task_load will not be null even if
4069 * task_h_load() returns 0.
4071 rq->misfit_task_load = max_t(unsigned long, task_h_load(p), 1);
4074 #else /* CONFIG_SMP */
4076 #define UPDATE_TG 0x0
4077 #define SKIP_AGE_LOAD 0x0
4078 #define DO_ATTACH 0x0
4080 static inline void update_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int not_used1)
4082 cfs_rq_util_change(cfs_rq, 0);
4085 static inline void remove_entity_load_avg(struct sched_entity *se) {}
4088 attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
4090 detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
4092 static inline int newidle_balance(struct rq *rq, struct rq_flags *rf)
4098 util_est_enqueue(struct cfs_rq *cfs_rq, struct task_struct *p) {}
4101 util_est_dequeue(struct cfs_rq *cfs_rq, struct task_struct *p) {}
4104 util_est_update(struct cfs_rq *cfs_rq, struct task_struct *p,
4106 static inline void update_misfit_status(struct task_struct *p, struct rq *rq) {}
4108 #endif /* CONFIG_SMP */
4110 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
4112 #ifdef CONFIG_SCHED_DEBUG
4113 s64 d = se->vruntime - cfs_rq->min_vruntime;
4118 if (d > 3*sysctl_sched_latency)
4119 schedstat_inc(cfs_rq->nr_spread_over);
4124 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
4126 u64 vruntime = cfs_rq->min_vruntime;
4129 * The 'current' period is already promised to the current tasks,
4130 * however the extra weight of the new task will slow them down a
4131 * little, place the new task so that it fits in the slot that
4132 * stays open at the end.
4134 if (initial && sched_feat(START_DEBIT))
4135 vruntime += sched_vslice(cfs_rq, se);
4137 /* sleeps up to a single latency don't count. */
4139 unsigned long thresh = sysctl_sched_latency;
4142 * Halve their sleep time's effect, to allow
4143 * for a gentler effect of sleepers:
4145 if (sched_feat(GENTLE_FAIR_SLEEPERS))
4151 /* ensure we never gain time by being placed backwards. */
4152 se->vruntime = max_vruntime(se->vruntime, vruntime);
4155 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
4157 static inline void check_schedstat_required(void)
4159 #ifdef CONFIG_SCHEDSTATS
4160 if (schedstat_enabled())
4163 /* Force schedstat enabled if a dependent tracepoint is active */
4164 if (trace_sched_stat_wait_enabled() ||
4165 trace_sched_stat_sleep_enabled() ||
4166 trace_sched_stat_iowait_enabled() ||
4167 trace_sched_stat_blocked_enabled() ||
4168 trace_sched_stat_runtime_enabled()) {
4169 printk_deferred_once("Scheduler tracepoints stat_sleep, stat_iowait, "
4170 "stat_blocked and stat_runtime require the "
4171 "kernel parameter schedstats=enable or "
4172 "kernel.sched_schedstats=1\n");
4177 static inline bool cfs_bandwidth_used(void);
4184 * update_min_vruntime()
4185 * vruntime -= min_vruntime
4189 * update_min_vruntime()
4190 * vruntime += min_vruntime
4192 * this way the vruntime transition between RQs is done when both
4193 * min_vruntime are up-to-date.
4197 * ->migrate_task_rq_fair() (p->state == TASK_WAKING)
4198 * vruntime -= min_vruntime
4202 * update_min_vruntime()
4203 * vruntime += min_vruntime
4205 * this way we don't have the most up-to-date min_vruntime on the originating
4206 * CPU and an up-to-date min_vruntime on the destination CPU.
4210 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
4212 bool renorm = !(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_MIGRATED);
4213 bool curr = cfs_rq->curr == se;
4216 * If we're the current task, we must renormalise before calling
4220 se->vruntime += cfs_rq->min_vruntime;
4222 update_curr(cfs_rq);
4225 * Otherwise, renormalise after, such that we're placed at the current
4226 * moment in time, instead of some random moment in the past. Being
4227 * placed in the past could significantly boost this task to the
4228 * fairness detriment of existing tasks.
4230 if (renorm && !curr)
4231 se->vruntime += cfs_rq->min_vruntime;
4234 * When enqueuing a sched_entity, we must:
4235 * - Update loads to have both entity and cfs_rq synced with now.
4236 * - Add its load to cfs_rq->runnable_avg
4237 * - For group_entity, update its weight to reflect the new share of
4239 * - Add its new weight to cfs_rq->load.weight
4241 update_load_avg(cfs_rq, se, UPDATE_TG | DO_ATTACH);
4242 se_update_runnable(se);
4243 update_cfs_group(se);
4244 account_entity_enqueue(cfs_rq, se);
4246 if (flags & ENQUEUE_WAKEUP)
4247 place_entity(cfs_rq, se, 0);
4249 check_schedstat_required();
4250 update_stats_enqueue(cfs_rq, se, flags);
4251 check_spread(cfs_rq, se);
4253 __enqueue_entity(cfs_rq, se);
4257 * When bandwidth control is enabled, cfs might have been removed
4258 * because of a parent been throttled but cfs->nr_running > 1. Try to
4259 * add it unconditionnally.
4261 if (cfs_rq->nr_running == 1 || cfs_bandwidth_used())
4262 list_add_leaf_cfs_rq(cfs_rq);
4264 if (cfs_rq->nr_running == 1)
4265 check_enqueue_throttle(cfs_rq);
4268 static void __clear_buddies_last(struct sched_entity *se)
4270 for_each_sched_entity(se) {
4271 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4272 if (cfs_rq->last != se)
4275 cfs_rq->last = NULL;
4279 static void __clear_buddies_next(struct sched_entity *se)
4281 for_each_sched_entity(se) {
4282 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4283 if (cfs_rq->next != se)
4286 cfs_rq->next = NULL;
4290 static void __clear_buddies_skip(struct sched_entity *se)
4292 for_each_sched_entity(se) {
4293 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4294 if (cfs_rq->skip != se)
4297 cfs_rq->skip = NULL;
4301 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
4303 if (cfs_rq->last == se)
4304 __clear_buddies_last(se);
4306 if (cfs_rq->next == se)
4307 __clear_buddies_next(se);
4309 if (cfs_rq->skip == se)
4310 __clear_buddies_skip(se);
4313 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
4316 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
4319 * Update run-time statistics of the 'current'.
4321 update_curr(cfs_rq);
4324 * When dequeuing a sched_entity, we must:
4325 * - Update loads to have both entity and cfs_rq synced with now.
4326 * - Subtract its load from the cfs_rq->runnable_avg.
4327 * - Subtract its previous weight from cfs_rq->load.weight.
4328 * - For group entity, update its weight to reflect the new share
4329 * of its group cfs_rq.
4331 update_load_avg(cfs_rq, se, UPDATE_TG);
4332 se_update_runnable(se);
4334 update_stats_dequeue(cfs_rq, se, flags);
4336 clear_buddies(cfs_rq, se);
4338 if (se != cfs_rq->curr)
4339 __dequeue_entity(cfs_rq, se);
4341 account_entity_dequeue(cfs_rq, se);
4344 * Normalize after update_curr(); which will also have moved
4345 * min_vruntime if @se is the one holding it back. But before doing
4346 * update_min_vruntime() again, which will discount @se's position and
4347 * can move min_vruntime forward still more.
4349 if (!(flags & DEQUEUE_SLEEP))
4350 se->vruntime -= cfs_rq->min_vruntime;
4352 /* return excess runtime on last dequeue */
4353 return_cfs_rq_runtime(cfs_rq);
4355 update_cfs_group(se);
4358 * Now advance min_vruntime if @se was the entity holding it back,
4359 * except when: DEQUEUE_SAVE && !DEQUEUE_MOVE, in this case we'll be
4360 * put back on, and if we advance min_vruntime, we'll be placed back
4361 * further than we started -- ie. we'll be penalized.
4363 if ((flags & (DEQUEUE_SAVE | DEQUEUE_MOVE)) != DEQUEUE_SAVE)
4364 update_min_vruntime(cfs_rq);
4368 * Preempt the current task with a newly woken task if needed:
4371 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
4373 unsigned long ideal_runtime, delta_exec;
4374 struct sched_entity *se;
4377 ideal_runtime = sched_slice(cfs_rq, curr);
4378 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
4379 if (delta_exec > ideal_runtime) {
4380 resched_curr(rq_of(cfs_rq));
4382 * The current task ran long enough, ensure it doesn't get
4383 * re-elected due to buddy favours.
4385 clear_buddies(cfs_rq, curr);
4390 * Ensure that a task that missed wakeup preemption by a
4391 * narrow margin doesn't have to wait for a full slice.
4392 * This also mitigates buddy induced latencies under load.
4394 if (delta_exec < sysctl_sched_min_granularity)
4397 se = __pick_first_entity(cfs_rq);
4398 delta = curr->vruntime - se->vruntime;
4403 if (delta > ideal_runtime)
4404 resched_curr(rq_of(cfs_rq));
4408 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
4410 /* 'current' is not kept within the tree. */
4413 * Any task has to be enqueued before it get to execute on
4414 * a CPU. So account for the time it spent waiting on the
4417 update_stats_wait_end(cfs_rq, se);
4418 __dequeue_entity(cfs_rq, se);
4419 update_load_avg(cfs_rq, se, UPDATE_TG);
4422 update_stats_curr_start(cfs_rq, se);
4426 * Track our maximum slice length, if the CPU's load is at
4427 * least twice that of our own weight (i.e. dont track it
4428 * when there are only lesser-weight tasks around):
4430 if (schedstat_enabled() &&
4431 rq_of(cfs_rq)->cfs.load.weight >= 2*se->load.weight) {
4432 schedstat_set(se->statistics.slice_max,
4433 max((u64)schedstat_val(se->statistics.slice_max),
4434 se->sum_exec_runtime - se->prev_sum_exec_runtime));
4437 se->prev_sum_exec_runtime = se->sum_exec_runtime;
4441 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
4444 * Pick the next process, keeping these things in mind, in this order:
4445 * 1) keep things fair between processes/task groups
4446 * 2) pick the "next" process, since someone really wants that to run
4447 * 3) pick the "last" process, for cache locality
4448 * 4) do not run the "skip" process, if something else is available
4450 static struct sched_entity *
4451 pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr)
4453 struct sched_entity *left = __pick_first_entity(cfs_rq);
4454 struct sched_entity *se;
4457 * If curr is set we have to see if its left of the leftmost entity
4458 * still in the tree, provided there was anything in the tree at all.
4460 if (!left || (curr && entity_before(curr, left)))
4463 se = left; /* ideally we run the leftmost entity */
4466 * Avoid running the skip buddy, if running something else can
4467 * be done without getting too unfair.
4469 if (cfs_rq->skip == se) {
4470 struct sched_entity *second;
4473 second = __pick_first_entity(cfs_rq);
4475 second = __pick_next_entity(se);
4476 if (!second || (curr && entity_before(curr, second)))
4480 if (second && wakeup_preempt_entity(second, left) < 1)
4484 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1) {
4486 * Someone really wants this to run. If it's not unfair, run it.
4489 } else if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1) {
4491 * Prefer last buddy, try to return the CPU to a preempted task.
4496 clear_buddies(cfs_rq, se);
4501 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
4503 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
4506 * If still on the runqueue then deactivate_task()
4507 * was not called and update_curr() has to be done:
4510 update_curr(cfs_rq);
4512 /* throttle cfs_rqs exceeding runtime */
4513 check_cfs_rq_runtime(cfs_rq);
4515 check_spread(cfs_rq, prev);
4518 update_stats_wait_start(cfs_rq, prev);
4519 /* Put 'current' back into the tree. */
4520 __enqueue_entity(cfs_rq, prev);
4521 /* in !on_rq case, update occurred at dequeue */
4522 update_load_avg(cfs_rq, prev, 0);
4524 cfs_rq->curr = NULL;
4528 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
4531 * Update run-time statistics of the 'current'.
4533 update_curr(cfs_rq);
4536 * Ensure that runnable average is periodically updated.
4538 update_load_avg(cfs_rq, curr, UPDATE_TG);
4539 update_cfs_group(curr);
4541 #ifdef CONFIG_SCHED_HRTICK
4543 * queued ticks are scheduled to match the slice, so don't bother
4544 * validating it and just reschedule.
4547 resched_curr(rq_of(cfs_rq));
4551 * don't let the period tick interfere with the hrtick preemption
4553 if (!sched_feat(DOUBLE_TICK) &&
4554 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
4558 if (cfs_rq->nr_running > 1)
4559 check_preempt_tick(cfs_rq, curr);
4563 /**************************************************
4564 * CFS bandwidth control machinery
4567 #ifdef CONFIG_CFS_BANDWIDTH
4569 #ifdef CONFIG_JUMP_LABEL
4570 static struct static_key __cfs_bandwidth_used;
4572 static inline bool cfs_bandwidth_used(void)
4574 return static_key_false(&__cfs_bandwidth_used);
4577 void cfs_bandwidth_usage_inc(void)
4579 static_key_slow_inc_cpuslocked(&__cfs_bandwidth_used);
4582 void cfs_bandwidth_usage_dec(void)
4584 static_key_slow_dec_cpuslocked(&__cfs_bandwidth_used);
4586 #else /* CONFIG_JUMP_LABEL */
4587 static bool cfs_bandwidth_used(void)
4592 void cfs_bandwidth_usage_inc(void) {}
4593 void cfs_bandwidth_usage_dec(void) {}
4594 #endif /* CONFIG_JUMP_LABEL */
4597 * default period for cfs group bandwidth.
4598 * default: 0.1s, units: nanoseconds
4600 static inline u64 default_cfs_period(void)
4602 return 100000000ULL;
4605 static inline u64 sched_cfs_bandwidth_slice(void)
4607 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
4611 * Replenish runtime according to assigned quota. We use sched_clock_cpu
4612 * directly instead of rq->clock to avoid adding additional synchronization
4615 * requires cfs_b->lock
4617 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
4619 if (cfs_b->quota != RUNTIME_INF)
4620 cfs_b->runtime = cfs_b->quota;
4623 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
4625 return &tg->cfs_bandwidth;
4628 /* returns 0 on failure to allocate runtime */
4629 static int __assign_cfs_rq_runtime(struct cfs_bandwidth *cfs_b,
4630 struct cfs_rq *cfs_rq, u64 target_runtime)
4632 u64 min_amount, amount = 0;
4634 lockdep_assert_held(&cfs_b->lock);
4636 /* note: this is a positive sum as runtime_remaining <= 0 */
4637 min_amount = target_runtime - cfs_rq->runtime_remaining;
4639 if (cfs_b->quota == RUNTIME_INF)
4640 amount = min_amount;
4642 start_cfs_bandwidth(cfs_b);
4644 if (cfs_b->runtime > 0) {
4645 amount = min(cfs_b->runtime, min_amount);
4646 cfs_b->runtime -= amount;
4651 cfs_rq->runtime_remaining += amount;
4653 return cfs_rq->runtime_remaining > 0;
4656 /* returns 0 on failure to allocate runtime */
4657 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4659 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4662 raw_spin_lock(&cfs_b->lock);
4663 ret = __assign_cfs_rq_runtime(cfs_b, cfs_rq, sched_cfs_bandwidth_slice());
4664 raw_spin_unlock(&cfs_b->lock);
4669 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
4671 /* dock delta_exec before expiring quota (as it could span periods) */
4672 cfs_rq->runtime_remaining -= delta_exec;
4674 if (likely(cfs_rq->runtime_remaining > 0))
4677 if (cfs_rq->throttled)
4680 * if we're unable to extend our runtime we resched so that the active
4681 * hierarchy can be throttled
4683 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
4684 resched_curr(rq_of(cfs_rq));
4687 static __always_inline
4688 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
4690 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
4693 __account_cfs_rq_runtime(cfs_rq, delta_exec);
4696 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
4698 return cfs_bandwidth_used() && cfs_rq->throttled;
4701 /* check whether cfs_rq, or any parent, is throttled */
4702 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
4704 return cfs_bandwidth_used() && cfs_rq->throttle_count;
4708 * Ensure that neither of the group entities corresponding to src_cpu or
4709 * dest_cpu are members of a throttled hierarchy when performing group
4710 * load-balance operations.
4712 static inline int throttled_lb_pair(struct task_group *tg,
4713 int src_cpu, int dest_cpu)
4715 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
4717 src_cfs_rq = tg->cfs_rq[src_cpu];
4718 dest_cfs_rq = tg->cfs_rq[dest_cpu];
4720 return throttled_hierarchy(src_cfs_rq) ||
4721 throttled_hierarchy(dest_cfs_rq);
4724 static int tg_unthrottle_up(struct task_group *tg, void *data)
4726 struct rq *rq = data;
4727 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
4729 cfs_rq->throttle_count--;
4730 if (!cfs_rq->throttle_count) {
4731 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
4732 cfs_rq->throttled_clock_task;
4734 /* Add cfs_rq with already running entity in the list */
4735 if (cfs_rq->nr_running >= 1)
4736 list_add_leaf_cfs_rq(cfs_rq);
4742 static int tg_throttle_down(struct task_group *tg, void *data)
4744 struct rq *rq = data;
4745 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
4747 /* group is entering throttled state, stop time */
4748 if (!cfs_rq->throttle_count) {
4749 cfs_rq->throttled_clock_task = rq_clock_task(rq);
4750 list_del_leaf_cfs_rq(cfs_rq);
4752 cfs_rq->throttle_count++;
4757 static bool throttle_cfs_rq(struct cfs_rq *cfs_rq)
4759 struct rq *rq = rq_of(cfs_rq);
4760 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4761 struct sched_entity *se;
4762 long task_delta, idle_task_delta, dequeue = 1;
4764 raw_spin_lock(&cfs_b->lock);
4765 /* This will start the period timer if necessary */
4766 if (__assign_cfs_rq_runtime(cfs_b, cfs_rq, 1)) {
4768 * We have raced with bandwidth becoming available, and if we
4769 * actually throttled the timer might not unthrottle us for an
4770 * entire period. We additionally needed to make sure that any
4771 * subsequent check_cfs_rq_runtime calls agree not to throttle
4772 * us, as we may commit to do cfs put_prev+pick_next, so we ask
4773 * for 1ns of runtime rather than just check cfs_b.
4777 list_add_tail_rcu(&cfs_rq->throttled_list,
4778 &cfs_b->throttled_cfs_rq);
4780 raw_spin_unlock(&cfs_b->lock);
4783 return false; /* Throttle no longer required. */
4785 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
4787 /* freeze hierarchy runnable averages while throttled */
4789 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
4792 task_delta = cfs_rq->h_nr_running;
4793 idle_task_delta = cfs_rq->idle_h_nr_running;
4794 for_each_sched_entity(se) {
4795 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
4796 /* throttled entity or throttle-on-deactivate */
4800 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
4802 qcfs_rq->h_nr_running -= task_delta;
4803 qcfs_rq->idle_h_nr_running -= idle_task_delta;
4805 if (qcfs_rq->load.weight) {
4806 /* Avoid re-evaluating load for this entity: */
4807 se = parent_entity(se);
4812 for_each_sched_entity(se) {
4813 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
4814 /* throttled entity or throttle-on-deactivate */
4818 update_load_avg(qcfs_rq, se, 0);
4819 se_update_runnable(se);
4821 qcfs_rq->h_nr_running -= task_delta;
4822 qcfs_rq->idle_h_nr_running -= idle_task_delta;
4825 /* At this point se is NULL and we are at root level*/
4826 sub_nr_running(rq, task_delta);
4830 * Note: distribution will already see us throttled via the
4831 * throttled-list. rq->lock protects completion.
4833 cfs_rq->throttled = 1;
4834 cfs_rq->throttled_clock = rq_clock(rq);
4838 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
4840 struct rq *rq = rq_of(cfs_rq);
4841 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4842 struct sched_entity *se;
4843 long task_delta, idle_task_delta;
4845 se = cfs_rq->tg->se[cpu_of(rq)];
4847 cfs_rq->throttled = 0;
4849 update_rq_clock(rq);
4851 raw_spin_lock(&cfs_b->lock);
4852 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
4853 list_del_rcu(&cfs_rq->throttled_list);
4854 raw_spin_unlock(&cfs_b->lock);
4856 /* update hierarchical throttle state */
4857 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
4859 if (!cfs_rq->load.weight)
4862 task_delta = cfs_rq->h_nr_running;
4863 idle_task_delta = cfs_rq->idle_h_nr_running;
4864 for_each_sched_entity(se) {
4867 cfs_rq = cfs_rq_of(se);
4868 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
4870 cfs_rq->h_nr_running += task_delta;
4871 cfs_rq->idle_h_nr_running += idle_task_delta;
4873 /* end evaluation on encountering a throttled cfs_rq */
4874 if (cfs_rq_throttled(cfs_rq))
4875 goto unthrottle_throttle;
4878 for_each_sched_entity(se) {
4879 cfs_rq = cfs_rq_of(se);
4881 update_load_avg(cfs_rq, se, UPDATE_TG);
4882 se_update_runnable(se);
4884 cfs_rq->h_nr_running += task_delta;
4885 cfs_rq->idle_h_nr_running += idle_task_delta;
4888 /* end evaluation on encountering a throttled cfs_rq */
4889 if (cfs_rq_throttled(cfs_rq))
4890 goto unthrottle_throttle;
4893 * One parent has been throttled and cfs_rq removed from the
4894 * list. Add it back to not break the leaf list.
4896 if (throttled_hierarchy(cfs_rq))
4897 list_add_leaf_cfs_rq(cfs_rq);
4900 /* At this point se is NULL and we are at root level*/
4901 add_nr_running(rq, task_delta);
4903 unthrottle_throttle:
4905 * The cfs_rq_throttled() breaks in the above iteration can result in
4906 * incomplete leaf list maintenance, resulting in triggering the
4909 for_each_sched_entity(se) {
4910 cfs_rq = cfs_rq_of(se);
4912 if (list_add_leaf_cfs_rq(cfs_rq))
4916 assert_list_leaf_cfs_rq(rq);
4918 /* Determine whether we need to wake up potentially idle CPU: */
4919 if (rq->curr == rq->idle && rq->cfs.nr_running)
4923 static void distribute_cfs_runtime(struct cfs_bandwidth *cfs_b)
4925 struct cfs_rq *cfs_rq;
4926 u64 runtime, remaining = 1;
4929 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
4931 struct rq *rq = rq_of(cfs_rq);
4934 rq_lock_irqsave(rq, &rf);
4935 if (!cfs_rq_throttled(cfs_rq))
4938 /* By the above check, this should never be true */
4939 SCHED_WARN_ON(cfs_rq->runtime_remaining > 0);
4941 raw_spin_lock(&cfs_b->lock);
4942 runtime = -cfs_rq->runtime_remaining + 1;
4943 if (runtime > cfs_b->runtime)
4944 runtime = cfs_b->runtime;
4945 cfs_b->runtime -= runtime;
4946 remaining = cfs_b->runtime;
4947 raw_spin_unlock(&cfs_b->lock);
4949 cfs_rq->runtime_remaining += runtime;
4951 /* we check whether we're throttled above */
4952 if (cfs_rq->runtime_remaining > 0)
4953 unthrottle_cfs_rq(cfs_rq);
4956 rq_unlock_irqrestore(rq, &rf);
4965 * Responsible for refilling a task_group's bandwidth and unthrottling its
4966 * cfs_rqs as appropriate. If there has been no activity within the last
4967 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
4968 * used to track this state.
4970 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun, unsigned long flags)
4974 /* no need to continue the timer with no bandwidth constraint */
4975 if (cfs_b->quota == RUNTIME_INF)
4976 goto out_deactivate;
4978 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
4979 cfs_b->nr_periods += overrun;
4982 * idle depends on !throttled (for the case of a large deficit), and if
4983 * we're going inactive then everything else can be deferred
4985 if (cfs_b->idle && !throttled)
4986 goto out_deactivate;
4988 __refill_cfs_bandwidth_runtime(cfs_b);
4991 /* mark as potentially idle for the upcoming period */
4996 /* account preceding periods in which throttling occurred */
4997 cfs_b->nr_throttled += overrun;
5000 * This check is repeated as we release cfs_b->lock while we unthrottle.
5002 while (throttled && cfs_b->runtime > 0) {
5003 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
5004 /* we can't nest cfs_b->lock while distributing bandwidth */
5005 distribute_cfs_runtime(cfs_b);
5006 raw_spin_lock_irqsave(&cfs_b->lock, flags);
5008 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
5012 * While we are ensured activity in the period following an
5013 * unthrottle, this also covers the case in which the new bandwidth is
5014 * insufficient to cover the existing bandwidth deficit. (Forcing the
5015 * timer to remain active while there are any throttled entities.)
5025 /* a cfs_rq won't donate quota below this amount */
5026 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
5027 /* minimum remaining period time to redistribute slack quota */
5028 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
5029 /* how long we wait to gather additional slack before distributing */
5030 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
5033 * Are we near the end of the current quota period?
5035 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
5036 * hrtimer base being cleared by hrtimer_start. In the case of
5037 * migrate_hrtimers, base is never cleared, so we are fine.
5039 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
5041 struct hrtimer *refresh_timer = &cfs_b->period_timer;
5044 /* if the call-back is running a quota refresh is already occurring */
5045 if (hrtimer_callback_running(refresh_timer))
5048 /* is a quota refresh about to occur? */
5049 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
5050 if (remaining < min_expire)
5056 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
5058 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
5060 /* if there's a quota refresh soon don't bother with slack */
5061 if (runtime_refresh_within(cfs_b, min_left))
5064 /* don't push forwards an existing deferred unthrottle */
5065 if (cfs_b->slack_started)
5067 cfs_b->slack_started = true;
5069 hrtimer_start(&cfs_b->slack_timer,
5070 ns_to_ktime(cfs_bandwidth_slack_period),
5074 /* we know any runtime found here is valid as update_curr() precedes return */
5075 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
5077 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
5078 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
5080 if (slack_runtime <= 0)
5083 raw_spin_lock(&cfs_b->lock);
5084 if (cfs_b->quota != RUNTIME_INF) {
5085 cfs_b->runtime += slack_runtime;
5087 /* we are under rq->lock, defer unthrottling using a timer */
5088 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
5089 !list_empty(&cfs_b->throttled_cfs_rq))
5090 start_cfs_slack_bandwidth(cfs_b);
5092 raw_spin_unlock(&cfs_b->lock);
5094 /* even if it's not valid for return we don't want to try again */
5095 cfs_rq->runtime_remaining -= slack_runtime;
5098 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
5100 if (!cfs_bandwidth_used())
5103 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
5106 __return_cfs_rq_runtime(cfs_rq);
5110 * This is done with a timer (instead of inline with bandwidth return) since
5111 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
5113 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
5115 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
5116 unsigned long flags;
5118 /* confirm we're still not at a refresh boundary */
5119 raw_spin_lock_irqsave(&cfs_b->lock, flags);
5120 cfs_b->slack_started = false;
5122 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
5123 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
5127 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice)
5128 runtime = cfs_b->runtime;
5130 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
5135 distribute_cfs_runtime(cfs_b);
5139 * When a group wakes up we want to make sure that its quota is not already
5140 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
5141 * runtime as update_curr() throttling can not trigger until it's on-rq.
5143 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
5145 if (!cfs_bandwidth_used())
5148 /* an active group must be handled by the update_curr()->put() path */
5149 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
5152 /* ensure the group is not already throttled */
5153 if (cfs_rq_throttled(cfs_rq))
5156 /* update runtime allocation */
5157 account_cfs_rq_runtime(cfs_rq, 0);
5158 if (cfs_rq->runtime_remaining <= 0)
5159 throttle_cfs_rq(cfs_rq);
5162 static void sync_throttle(struct task_group *tg, int cpu)
5164 struct cfs_rq *pcfs_rq, *cfs_rq;
5166 if (!cfs_bandwidth_used())
5172 cfs_rq = tg->cfs_rq[cpu];
5173 pcfs_rq = tg->parent->cfs_rq[cpu];
5175 cfs_rq->throttle_count = pcfs_rq->throttle_count;
5176 cfs_rq->throttled_clock_task = rq_clock_task(cpu_rq(cpu));
5179 /* conditionally throttle active cfs_rq's from put_prev_entity() */
5180 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
5182 if (!cfs_bandwidth_used())
5185 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
5189 * it's possible for a throttled entity to be forced into a running
5190 * state (e.g. set_curr_task), in this case we're finished.
5192 if (cfs_rq_throttled(cfs_rq))
5195 return throttle_cfs_rq(cfs_rq);
5198 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
5200 struct cfs_bandwidth *cfs_b =
5201 container_of(timer, struct cfs_bandwidth, slack_timer);
5203 do_sched_cfs_slack_timer(cfs_b);
5205 return HRTIMER_NORESTART;
5208 extern const u64 max_cfs_quota_period;
5210 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
5212 struct cfs_bandwidth *cfs_b =
5213 container_of(timer, struct cfs_bandwidth, period_timer);
5214 unsigned long flags;
5219 raw_spin_lock_irqsave(&cfs_b->lock, flags);
5221 overrun = hrtimer_forward_now(timer, cfs_b->period);
5225 idle = do_sched_cfs_period_timer(cfs_b, overrun, flags);
5228 u64 new, old = ktime_to_ns(cfs_b->period);
5231 * Grow period by a factor of 2 to avoid losing precision.
5232 * Precision loss in the quota/period ratio can cause __cfs_schedulable
5236 if (new < max_cfs_quota_period) {
5237 cfs_b->period = ns_to_ktime(new);
5240 pr_warn_ratelimited(
5241 "cfs_period_timer[cpu%d]: period too short, scaling up (new cfs_period_us = %lld, cfs_quota_us = %lld)\n",
5243 div_u64(new, NSEC_PER_USEC),
5244 div_u64(cfs_b->quota, NSEC_PER_USEC));
5246 pr_warn_ratelimited(
5247 "cfs_period_timer[cpu%d]: period too short, but cannot scale up without losing precision (cfs_period_us = %lld, cfs_quota_us = %lld)\n",
5249 div_u64(old, NSEC_PER_USEC),
5250 div_u64(cfs_b->quota, NSEC_PER_USEC));
5253 /* reset count so we don't come right back in here */
5258 cfs_b->period_active = 0;
5259 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
5261 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
5264 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
5266 raw_spin_lock_init(&cfs_b->lock);
5268 cfs_b->quota = RUNTIME_INF;
5269 cfs_b->period = ns_to_ktime(default_cfs_period());
5271 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
5272 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED);
5273 cfs_b->period_timer.function = sched_cfs_period_timer;
5274 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
5275 cfs_b->slack_timer.function = sched_cfs_slack_timer;
5276 cfs_b->slack_started = false;
5279 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
5281 cfs_rq->runtime_enabled = 0;
5282 INIT_LIST_HEAD(&cfs_rq->throttled_list);
5285 void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
5287 lockdep_assert_held(&cfs_b->lock);
5289 if (cfs_b->period_active)
5292 cfs_b->period_active = 1;
5293 hrtimer_forward_now(&cfs_b->period_timer, cfs_b->period);
5294 hrtimer_start_expires(&cfs_b->period_timer, HRTIMER_MODE_ABS_PINNED);
5297 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
5299 /* init_cfs_bandwidth() was not called */
5300 if (!cfs_b->throttled_cfs_rq.next)
5303 hrtimer_cancel(&cfs_b->period_timer);
5304 hrtimer_cancel(&cfs_b->slack_timer);
5308 * Both these CPU hotplug callbacks race against unregister_fair_sched_group()
5310 * The race is harmless, since modifying bandwidth settings of unhooked group
5311 * bits doesn't do much.
5314 /* cpu online calback */
5315 static void __maybe_unused update_runtime_enabled(struct rq *rq)
5317 struct task_group *tg;
5319 lockdep_assert_held(&rq->lock);
5322 list_for_each_entry_rcu(tg, &task_groups, list) {
5323 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
5324 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
5326 raw_spin_lock(&cfs_b->lock);
5327 cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF;
5328 raw_spin_unlock(&cfs_b->lock);
5333 /* cpu offline callback */
5334 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
5336 struct task_group *tg;
5338 lockdep_assert_held(&rq->lock);
5341 list_for_each_entry_rcu(tg, &task_groups, list) {
5342 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
5344 if (!cfs_rq->runtime_enabled)
5348 * clock_task is not advancing so we just need to make sure
5349 * there's some valid quota amount
5351 cfs_rq->runtime_remaining = 1;
5353 * Offline rq is schedulable till CPU is completely disabled
5354 * in take_cpu_down(), so we prevent new cfs throttling here.
5356 cfs_rq->runtime_enabled = 0;
5358 if (cfs_rq_throttled(cfs_rq))
5359 unthrottle_cfs_rq(cfs_rq);
5364 #else /* CONFIG_CFS_BANDWIDTH */
5366 static inline bool cfs_bandwidth_used(void)
5371 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {}
5372 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; }
5373 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
5374 static inline void sync_throttle(struct task_group *tg, int cpu) {}
5375 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
5377 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
5382 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
5387 static inline int throttled_lb_pair(struct task_group *tg,
5388 int src_cpu, int dest_cpu)
5393 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
5395 #ifdef CONFIG_FAIR_GROUP_SCHED
5396 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
5399 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
5403 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
5404 static inline void update_runtime_enabled(struct rq *rq) {}
5405 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
5407 #endif /* CONFIG_CFS_BANDWIDTH */
5409 /**************************************************
5410 * CFS operations on tasks:
5413 #ifdef CONFIG_SCHED_HRTICK
5414 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
5416 struct sched_entity *se = &p->se;
5417 struct cfs_rq *cfs_rq = cfs_rq_of(se);
5419 SCHED_WARN_ON(task_rq(p) != rq);
5421 if (rq->cfs.h_nr_running > 1) {
5422 u64 slice = sched_slice(cfs_rq, se);
5423 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
5424 s64 delta = slice - ran;
5427 if (task_current(rq, p))
5431 hrtick_start(rq, delta);
5436 * called from enqueue/dequeue and updates the hrtick when the
5437 * current task is from our class and nr_running is low enough
5440 static void hrtick_update(struct rq *rq)
5442 struct task_struct *curr = rq->curr;
5444 if (!hrtick_enabled_fair(rq) || curr->sched_class != &fair_sched_class)
5447 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
5448 hrtick_start_fair(rq, curr);
5450 #else /* !CONFIG_SCHED_HRTICK */
5452 hrtick_start_fair(struct rq *rq, struct task_struct *p)
5456 static inline void hrtick_update(struct rq *rq)
5462 static inline unsigned long cpu_util(int cpu);
5464 static inline bool cpu_overutilized(int cpu)
5466 return !fits_capacity(cpu_util(cpu), capacity_of(cpu));
5469 static inline void update_overutilized_status(struct rq *rq)
5471 if (!READ_ONCE(rq->rd->overutilized) && cpu_overutilized(rq->cpu)) {
5472 WRITE_ONCE(rq->rd->overutilized, SG_OVERUTILIZED);
5473 trace_sched_overutilized_tp(rq->rd, SG_OVERUTILIZED);
5477 static inline void update_overutilized_status(struct rq *rq) { }
5480 /* Runqueue only has SCHED_IDLE tasks enqueued */
5481 static int sched_idle_rq(struct rq *rq)
5483 return unlikely(rq->nr_running == rq->cfs.idle_h_nr_running &&
5488 static int sched_idle_cpu(int cpu)
5490 return sched_idle_rq(cpu_rq(cpu));
5495 * The enqueue_task method is called before nr_running is
5496 * increased. Here we update the fair scheduling stats and
5497 * then put the task into the rbtree:
5500 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
5502 struct cfs_rq *cfs_rq;
5503 struct sched_entity *se = &p->se;
5504 int idle_h_nr_running = task_has_idle_policy(p);
5505 int task_new = !(flags & ENQUEUE_WAKEUP);
5508 * The code below (indirectly) updates schedutil which looks at
5509 * the cfs_rq utilization to select a frequency.
5510 * Let's add the task's estimated utilization to the cfs_rq's
5511 * estimated utilization, before we update schedutil.
5513 util_est_enqueue(&rq->cfs, p);
5516 * If in_iowait is set, the code below may not trigger any cpufreq
5517 * utilization updates, so do it here explicitly with the IOWAIT flag
5521 cpufreq_update_util(rq, SCHED_CPUFREQ_IOWAIT);
5523 for_each_sched_entity(se) {
5526 cfs_rq = cfs_rq_of(se);
5527 enqueue_entity(cfs_rq, se, flags);
5529 cfs_rq->h_nr_running++;
5530 cfs_rq->idle_h_nr_running += idle_h_nr_running;
5532 /* end evaluation on encountering a throttled cfs_rq */
5533 if (cfs_rq_throttled(cfs_rq))
5534 goto enqueue_throttle;
5536 flags = ENQUEUE_WAKEUP;
5539 for_each_sched_entity(se) {
5540 cfs_rq = cfs_rq_of(se);
5542 update_load_avg(cfs_rq, se, UPDATE_TG);
5543 se_update_runnable(se);
5544 update_cfs_group(se);
5546 cfs_rq->h_nr_running++;
5547 cfs_rq->idle_h_nr_running += idle_h_nr_running;
5549 /* end evaluation on encountering a throttled cfs_rq */
5550 if (cfs_rq_throttled(cfs_rq))
5551 goto enqueue_throttle;
5554 * One parent has been throttled and cfs_rq removed from the
5555 * list. Add it back to not break the leaf list.
5557 if (throttled_hierarchy(cfs_rq))
5558 list_add_leaf_cfs_rq(cfs_rq);
5561 /* At this point se is NULL and we are at root level*/
5562 add_nr_running(rq, 1);
5565 * Since new tasks are assigned an initial util_avg equal to
5566 * half of the spare capacity of their CPU, tiny tasks have the
5567 * ability to cross the overutilized threshold, which will
5568 * result in the load balancer ruining all the task placement
5569 * done by EAS. As a way to mitigate that effect, do not account
5570 * for the first enqueue operation of new tasks during the
5571 * overutilized flag detection.
5573 * A better way of solving this problem would be to wait for
5574 * the PELT signals of tasks to converge before taking them
5575 * into account, but that is not straightforward to implement,
5576 * and the following generally works well enough in practice.
5579 update_overutilized_status(rq);
5582 if (cfs_bandwidth_used()) {
5584 * When bandwidth control is enabled; the cfs_rq_throttled()
5585 * breaks in the above iteration can result in incomplete
5586 * leaf list maintenance, resulting in triggering the assertion
5589 for_each_sched_entity(se) {
5590 cfs_rq = cfs_rq_of(se);
5592 if (list_add_leaf_cfs_rq(cfs_rq))
5597 assert_list_leaf_cfs_rq(rq);
5602 static void set_next_buddy(struct sched_entity *se);
5605 * The dequeue_task method is called before nr_running is
5606 * decreased. We remove the task from the rbtree and
5607 * update the fair scheduling stats:
5609 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
5611 struct cfs_rq *cfs_rq;
5612 struct sched_entity *se = &p->se;
5613 int task_sleep = flags & DEQUEUE_SLEEP;
5614 int idle_h_nr_running = task_has_idle_policy(p);
5615 bool was_sched_idle = sched_idle_rq(rq);
5617 util_est_dequeue(&rq->cfs, p);
5619 for_each_sched_entity(se) {
5620 cfs_rq = cfs_rq_of(se);
5621 dequeue_entity(cfs_rq, se, flags);
5623 cfs_rq->h_nr_running--;
5624 cfs_rq->idle_h_nr_running -= idle_h_nr_running;
5626 /* end evaluation on encountering a throttled cfs_rq */
5627 if (cfs_rq_throttled(cfs_rq))
5628 goto dequeue_throttle;
5630 /* Don't dequeue parent if it has other entities besides us */
5631 if (cfs_rq->load.weight) {
5632 /* Avoid re-evaluating load for this entity: */
5633 se = parent_entity(se);
5635 * Bias pick_next to pick a task from this cfs_rq, as
5636 * p is sleeping when it is within its sched_slice.
5638 if (task_sleep && se && !throttled_hierarchy(cfs_rq))
5642 flags |= DEQUEUE_SLEEP;
5645 for_each_sched_entity(se) {
5646 cfs_rq = cfs_rq_of(se);
5648 update_load_avg(cfs_rq, se, UPDATE_TG);
5649 se_update_runnable(se);
5650 update_cfs_group(se);
5652 cfs_rq->h_nr_running--;
5653 cfs_rq->idle_h_nr_running -= idle_h_nr_running;
5655 /* end evaluation on encountering a throttled cfs_rq */
5656 if (cfs_rq_throttled(cfs_rq))
5657 goto dequeue_throttle;
5661 /* At this point se is NULL and we are at root level*/
5662 sub_nr_running(rq, 1);
5664 /* balance early to pull high priority tasks */
5665 if (unlikely(!was_sched_idle && sched_idle_rq(rq)))
5666 rq->next_balance = jiffies;
5669 util_est_update(&rq->cfs, p, task_sleep);
5675 /* Working cpumask for: load_balance, load_balance_newidle. */
5676 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
5677 DEFINE_PER_CPU(cpumask_var_t, select_idle_mask);
5679 #ifdef CONFIG_NO_HZ_COMMON
5682 cpumask_var_t idle_cpus_mask;
5684 int has_blocked; /* Idle CPUS has blocked load */
5685 unsigned long next_balance; /* in jiffy units */
5686 unsigned long next_blocked; /* Next update of blocked load in jiffies */
5687 } nohz ____cacheline_aligned;
5689 #endif /* CONFIG_NO_HZ_COMMON */
5691 static unsigned long cpu_load(struct rq *rq)
5693 return cfs_rq_load_avg(&rq->cfs);
5697 * cpu_load_without - compute CPU load without any contributions from *p
5698 * @cpu: the CPU which load is requested
5699 * @p: the task which load should be discounted
5701 * The load of a CPU is defined by the load of tasks currently enqueued on that
5702 * CPU as well as tasks which are currently sleeping after an execution on that
5705 * This method returns the load of the specified CPU by discounting the load of
5706 * the specified task, whenever the task is currently contributing to the CPU
5709 static unsigned long cpu_load_without(struct rq *rq, struct task_struct *p)
5711 struct cfs_rq *cfs_rq;
5714 /* Task has no contribution or is new */
5715 if (cpu_of(rq) != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time))
5716 return cpu_load(rq);
5719 load = READ_ONCE(cfs_rq->avg.load_avg);
5721 /* Discount task's util from CPU's util */
5722 lsub_positive(&load, task_h_load(p));
5727 static unsigned long cpu_runnable(struct rq *rq)
5729 return cfs_rq_runnable_avg(&rq->cfs);
5732 static unsigned long cpu_runnable_without(struct rq *rq, struct task_struct *p)
5734 struct cfs_rq *cfs_rq;
5735 unsigned int runnable;
5737 /* Task has no contribution or is new */
5738 if (cpu_of(rq) != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time))
5739 return cpu_runnable(rq);
5742 runnable = READ_ONCE(cfs_rq->avg.runnable_avg);
5744 /* Discount task's runnable from CPU's runnable */
5745 lsub_positive(&runnable, p->se.avg.runnable_avg);
5750 static unsigned long capacity_of(int cpu)
5752 return cpu_rq(cpu)->cpu_capacity;
5755 static void record_wakee(struct task_struct *p)
5758 * Only decay a single time; tasks that have less then 1 wakeup per
5759 * jiffy will not have built up many flips.
5761 if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) {
5762 current->wakee_flips >>= 1;
5763 current->wakee_flip_decay_ts = jiffies;
5766 if (current->last_wakee != p) {
5767 current->last_wakee = p;
5768 current->wakee_flips++;
5773 * Detect M:N waker/wakee relationships via a switching-frequency heuristic.
5775 * A waker of many should wake a different task than the one last awakened
5776 * at a frequency roughly N times higher than one of its wakees.
5778 * In order to determine whether we should let the load spread vs consolidating
5779 * to shared cache, we look for a minimum 'flip' frequency of llc_size in one
5780 * partner, and a factor of lls_size higher frequency in the other.
5782 * With both conditions met, we can be relatively sure that the relationship is
5783 * non-monogamous, with partner count exceeding socket size.
5785 * Waker/wakee being client/server, worker/dispatcher, interrupt source or
5786 * whatever is irrelevant, spread criteria is apparent partner count exceeds
5789 static int wake_wide(struct task_struct *p)
5791 unsigned int master = current->wakee_flips;
5792 unsigned int slave = p->wakee_flips;
5793 int factor = __this_cpu_read(sd_llc_size);
5796 swap(master, slave);
5797 if (slave < factor || master < slave * factor)
5803 * The purpose of wake_affine() is to quickly determine on which CPU we can run
5804 * soonest. For the purpose of speed we only consider the waking and previous
5807 * wake_affine_idle() - only considers 'now', it check if the waking CPU is
5808 * cache-affine and is (or will be) idle.
5810 * wake_affine_weight() - considers the weight to reflect the average
5811 * scheduling latency of the CPUs. This seems to work
5812 * for the overloaded case.
5815 wake_affine_idle(int this_cpu, int prev_cpu, int sync)
5818 * If this_cpu is idle, it implies the wakeup is from interrupt
5819 * context. Only allow the move if cache is shared. Otherwise an
5820 * interrupt intensive workload could force all tasks onto one
5821 * node depending on the IO topology or IRQ affinity settings.
5823 * If the prev_cpu is idle and cache affine then avoid a migration.
5824 * There is no guarantee that the cache hot data from an interrupt
5825 * is more important than cache hot data on the prev_cpu and from
5826 * a cpufreq perspective, it's better to have higher utilisation
5829 if (available_idle_cpu(this_cpu) && cpus_share_cache(this_cpu, prev_cpu))
5830 return available_idle_cpu(prev_cpu) ? prev_cpu : this_cpu;
5832 if (sync && cpu_rq(this_cpu)->nr_running == 1)
5835 if (available_idle_cpu(prev_cpu))
5838 return nr_cpumask_bits;
5842 wake_affine_weight(struct sched_domain *sd, struct task_struct *p,
5843 int this_cpu, int prev_cpu, int sync)
5845 s64 this_eff_load, prev_eff_load;
5846 unsigned long task_load;
5848 this_eff_load = cpu_load(cpu_rq(this_cpu));
5851 unsigned long current_load = task_h_load(current);
5853 if (current_load > this_eff_load)
5856 this_eff_load -= current_load;
5859 task_load = task_h_load(p);
5861 this_eff_load += task_load;
5862 if (sched_feat(WA_BIAS))
5863 this_eff_load *= 100;
5864 this_eff_load *= capacity_of(prev_cpu);
5866 prev_eff_load = cpu_load(cpu_rq(prev_cpu));
5867 prev_eff_load -= task_load;
5868 if (sched_feat(WA_BIAS))
5869 prev_eff_load *= 100 + (sd->imbalance_pct - 100) / 2;
5870 prev_eff_load *= capacity_of(this_cpu);
5873 * If sync, adjust the weight of prev_eff_load such that if
5874 * prev_eff == this_eff that select_idle_sibling() will consider
5875 * stacking the wakee on top of the waker if no other CPU is
5881 return this_eff_load < prev_eff_load ? this_cpu : nr_cpumask_bits;
5884 static int wake_affine(struct sched_domain *sd, struct task_struct *p,
5885 int this_cpu, int prev_cpu, int sync)
5887 int target = nr_cpumask_bits;
5889 if (sched_feat(WA_IDLE))
5890 target = wake_affine_idle(this_cpu, prev_cpu, sync);
5892 if (sched_feat(WA_WEIGHT) && target == nr_cpumask_bits)
5893 target = wake_affine_weight(sd, p, this_cpu, prev_cpu, sync);
5895 schedstat_inc(p->se.statistics.nr_wakeups_affine_attempts);
5896 if (target == nr_cpumask_bits)
5899 schedstat_inc(sd->ttwu_move_affine);
5900 schedstat_inc(p->se.statistics.nr_wakeups_affine);
5904 static struct sched_group *
5905 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu);
5908 * find_idlest_group_cpu - find the idlest CPU among the CPUs in the group.
5911 find_idlest_group_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
5913 unsigned long load, min_load = ULONG_MAX;
5914 unsigned int min_exit_latency = UINT_MAX;
5915 u64 latest_idle_timestamp = 0;
5916 int least_loaded_cpu = this_cpu;
5917 int shallowest_idle_cpu = -1;
5920 /* Check if we have any choice: */
5921 if (group->group_weight == 1)
5922 return cpumask_first(sched_group_span(group));
5924 /* Traverse only the allowed CPUs */
5925 for_each_cpu_and(i, sched_group_span(group), p->cpus_ptr) {
5926 if (sched_idle_cpu(i))
5929 if (available_idle_cpu(i)) {
5930 struct rq *rq = cpu_rq(i);
5931 struct cpuidle_state *idle = idle_get_state(rq);
5932 if (idle && idle->exit_latency < min_exit_latency) {
5934 * We give priority to a CPU whose idle state
5935 * has the smallest exit latency irrespective
5936 * of any idle timestamp.
5938 min_exit_latency = idle->exit_latency;
5939 latest_idle_timestamp = rq->idle_stamp;
5940 shallowest_idle_cpu = i;
5941 } else if ((!idle || idle->exit_latency == min_exit_latency) &&
5942 rq->idle_stamp > latest_idle_timestamp) {
5944 * If equal or no active idle state, then
5945 * the most recently idled CPU might have
5948 latest_idle_timestamp = rq->idle_stamp;
5949 shallowest_idle_cpu = i;
5951 } else if (shallowest_idle_cpu == -1) {
5952 load = cpu_load(cpu_rq(i));
5953 if (load < min_load) {
5955 least_loaded_cpu = i;
5960 return shallowest_idle_cpu != -1 ? shallowest_idle_cpu : least_loaded_cpu;
5963 static inline int find_idlest_cpu(struct sched_domain *sd, struct task_struct *p,
5964 int cpu, int prev_cpu, int sd_flag)
5968 if (!cpumask_intersects(sched_domain_span(sd), p->cpus_ptr))
5972 * We need task's util for cpu_util_without, sync it up to
5973 * prev_cpu's last_update_time.
5975 if (!(sd_flag & SD_BALANCE_FORK))
5976 sync_entity_load_avg(&p->se);
5979 struct sched_group *group;
5980 struct sched_domain *tmp;
5983 if (!(sd->flags & sd_flag)) {
5988 group = find_idlest_group(sd, p, cpu);
5994 new_cpu = find_idlest_group_cpu(group, p, cpu);
5995 if (new_cpu == cpu) {
5996 /* Now try balancing at a lower domain level of 'cpu': */
6001 /* Now try balancing at a lower domain level of 'new_cpu': */
6003 weight = sd->span_weight;
6005 for_each_domain(cpu, tmp) {
6006 if (weight <= tmp->span_weight)
6008 if (tmp->flags & sd_flag)
6016 static inline int __select_idle_cpu(int cpu)
6018 if (available_idle_cpu(cpu) || sched_idle_cpu(cpu))
6024 #ifdef CONFIG_SCHED_SMT
6025 DEFINE_STATIC_KEY_FALSE(sched_smt_present);
6026 EXPORT_SYMBOL_GPL(sched_smt_present);
6028 static inline void set_idle_cores(int cpu, int val)
6030 struct sched_domain_shared *sds;
6032 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
6034 WRITE_ONCE(sds->has_idle_cores, val);
6037 static inline bool test_idle_cores(int cpu, bool def)
6039 struct sched_domain_shared *sds;
6041 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
6043 return READ_ONCE(sds->has_idle_cores);
6049 * Scans the local SMT mask to see if the entire core is idle, and records this
6050 * information in sd_llc_shared->has_idle_cores.
6052 * Since SMT siblings share all cache levels, inspecting this limited remote
6053 * state should be fairly cheap.
6055 void __update_idle_core(struct rq *rq)
6057 int core = cpu_of(rq);
6061 if (test_idle_cores(core, true))
6064 for_each_cpu(cpu, cpu_smt_mask(core)) {
6068 if (!available_idle_cpu(cpu))
6072 set_idle_cores(core, 1);
6078 * Scan the entire LLC domain for idle cores; this dynamically switches off if
6079 * there are no idle cores left in the system; tracked through
6080 * sd_llc->shared->has_idle_cores and enabled through update_idle_core() above.
6082 static int select_idle_core(struct task_struct *p, int core, struct cpumask *cpus, int *idle_cpu)
6087 if (!static_branch_likely(&sched_smt_present))
6088 return __select_idle_cpu(core);
6090 for_each_cpu(cpu, cpu_smt_mask(core)) {
6091 if (!available_idle_cpu(cpu)) {
6093 if (*idle_cpu == -1) {
6094 if (sched_idle_cpu(cpu) && cpumask_test_cpu(cpu, p->cpus_ptr)) {
6102 if (*idle_cpu == -1 && cpumask_test_cpu(cpu, p->cpus_ptr))
6109 cpumask_andnot(cpus, cpus, cpu_smt_mask(core));
6113 #else /* CONFIG_SCHED_SMT */
6115 static inline void set_idle_cores(int cpu, int val)
6119 static inline bool test_idle_cores(int cpu, bool def)
6124 static inline int select_idle_core(struct task_struct *p, int core, struct cpumask *cpus, int *idle_cpu)
6126 return __select_idle_cpu(core);
6129 #endif /* CONFIG_SCHED_SMT */
6132 * Scan the LLC domain for idle CPUs; this is dynamically regulated by
6133 * comparing the average scan cost (tracked in sd->avg_scan_cost) against the
6134 * average idle time for this rq (as found in rq->avg_idle).
6136 static int select_idle_cpu(struct task_struct *p, struct sched_domain *sd, int target)
6138 struct cpumask *cpus = this_cpu_cpumask_var_ptr(select_idle_mask);
6139 int i, cpu, idle_cpu = -1, nr = INT_MAX;
6140 bool smt = test_idle_cores(target, false);
6141 int this = smp_processor_id();
6142 struct sched_domain *this_sd;
6145 this_sd = rcu_dereference(*this_cpu_ptr(&sd_llc));
6149 cpumask_and(cpus, sched_domain_span(sd), p->cpus_ptr);
6151 if (sched_feat(SIS_PROP) && !smt) {
6152 u64 avg_cost, avg_idle, span_avg;
6155 * Due to large variance we need a large fuzz factor;
6156 * hackbench in particularly is sensitive here.
6158 avg_idle = this_rq()->avg_idle / 512;
6159 avg_cost = this_sd->avg_scan_cost + 1;
6161 span_avg = sd->span_weight * avg_idle;
6162 if (span_avg > 4*avg_cost)
6163 nr = div_u64(span_avg, avg_cost);
6167 time = cpu_clock(this);
6170 for_each_cpu_wrap(cpu, cpus, target) {
6172 i = select_idle_core(p, cpu, cpus, &idle_cpu);
6173 if ((unsigned int)i < nr_cpumask_bits)
6179 idle_cpu = __select_idle_cpu(cpu);
6180 if ((unsigned int)idle_cpu < nr_cpumask_bits)
6186 set_idle_cores(this, false);
6188 if (sched_feat(SIS_PROP) && !smt) {
6189 time = cpu_clock(this) - time;
6190 update_avg(&this_sd->avg_scan_cost, time);
6197 * Scan the asym_capacity domain for idle CPUs; pick the first idle one on which
6198 * the task fits. If no CPU is big enough, but there are idle ones, try to
6199 * maximize capacity.
6202 select_idle_capacity(struct task_struct *p, struct sched_domain *sd, int target)
6204 unsigned long task_util, best_cap = 0;
6205 int cpu, best_cpu = -1;
6206 struct cpumask *cpus;
6208 cpus = this_cpu_cpumask_var_ptr(select_idle_mask);
6209 cpumask_and(cpus, sched_domain_span(sd), p->cpus_ptr);
6211 task_util = uclamp_task_util(p);
6213 for_each_cpu_wrap(cpu, cpus, target) {
6214 unsigned long cpu_cap = capacity_of(cpu);
6216 if (!available_idle_cpu(cpu) && !sched_idle_cpu(cpu))
6218 if (fits_capacity(task_util, cpu_cap))
6221 if (cpu_cap > best_cap) {
6230 static inline bool asym_fits_capacity(int task_util, int cpu)
6232 if (static_branch_unlikely(&sched_asym_cpucapacity))
6233 return fits_capacity(task_util, capacity_of(cpu));
6239 * Try and locate an idle core/thread in the LLC cache domain.
6241 static int select_idle_sibling(struct task_struct *p, int prev, int target)
6243 struct sched_domain *sd;
6244 unsigned long task_util;
6245 int i, recent_used_cpu;
6248 * On asymmetric system, update task utilization because we will check
6249 * that the task fits with cpu's capacity.
6251 if (static_branch_unlikely(&sched_asym_cpucapacity)) {
6252 sync_entity_load_avg(&p->se);
6253 task_util = uclamp_task_util(p);
6256 if ((available_idle_cpu(target) || sched_idle_cpu(target)) &&
6257 asym_fits_capacity(task_util, target))
6261 * If the previous CPU is cache affine and idle, don't be stupid:
6263 if (prev != target && cpus_share_cache(prev, target) &&
6264 (available_idle_cpu(prev) || sched_idle_cpu(prev)) &&
6265 asym_fits_capacity(task_util, prev))
6269 * Allow a per-cpu kthread to stack with the wakee if the
6270 * kworker thread and the tasks previous CPUs are the same.
6271 * The assumption is that the wakee queued work for the
6272 * per-cpu kthread that is now complete and the wakeup is
6273 * essentially a sync wakeup. An obvious example of this
6274 * pattern is IO completions.
6276 if (is_per_cpu_kthread(current) &&
6277 prev == smp_processor_id() &&
6278 this_rq()->nr_running <= 1) {
6282 /* Check a recently used CPU as a potential idle candidate: */
6283 recent_used_cpu = p->recent_used_cpu;
6284 if (recent_used_cpu != prev &&
6285 recent_used_cpu != target &&
6286 cpus_share_cache(recent_used_cpu, target) &&
6287 (available_idle_cpu(recent_used_cpu) || sched_idle_cpu(recent_used_cpu)) &&
6288 cpumask_test_cpu(p->recent_used_cpu, p->cpus_ptr) &&
6289 asym_fits_capacity(task_util, recent_used_cpu)) {
6291 * Replace recent_used_cpu with prev as it is a potential
6292 * candidate for the next wake:
6294 p->recent_used_cpu = prev;
6295 return recent_used_cpu;
6299 * For asymmetric CPU capacity systems, our domain of interest is
6300 * sd_asym_cpucapacity rather than sd_llc.
6302 if (static_branch_unlikely(&sched_asym_cpucapacity)) {
6303 sd = rcu_dereference(per_cpu(sd_asym_cpucapacity, target));
6305 * On an asymmetric CPU capacity system where an exclusive
6306 * cpuset defines a symmetric island (i.e. one unique
6307 * capacity_orig value through the cpuset), the key will be set
6308 * but the CPUs within that cpuset will not have a domain with
6309 * SD_ASYM_CPUCAPACITY. These should follow the usual symmetric
6313 i = select_idle_capacity(p, sd, target);
6314 return ((unsigned)i < nr_cpumask_bits) ? i : target;
6318 sd = rcu_dereference(per_cpu(sd_llc, target));
6322 i = select_idle_cpu(p, sd, target);
6323 if ((unsigned)i < nr_cpumask_bits)
6330 * cpu_util - Estimates the amount of capacity of a CPU used by CFS tasks.
6331 * @cpu: the CPU to get the utilization of
6333 * The unit of the return value must be the one of capacity so we can compare
6334 * the utilization with the capacity of the CPU that is available for CFS task
6335 * (ie cpu_capacity).
6337 * cfs_rq.avg.util_avg is the sum of running time of runnable tasks plus the
6338 * recent utilization of currently non-runnable tasks on a CPU. It represents
6339 * the amount of utilization of a CPU in the range [0..capacity_orig] where
6340 * capacity_orig is the cpu_capacity available at the highest frequency
6341 * (arch_scale_freq_capacity()).
6342 * The utilization of a CPU converges towards a sum equal to or less than the
6343 * current capacity (capacity_curr <= capacity_orig) of the CPU because it is
6344 * the running time on this CPU scaled by capacity_curr.
6346 * The estimated utilization of a CPU is defined to be the maximum between its
6347 * cfs_rq.avg.util_avg and the sum of the estimated utilization of the tasks
6348 * currently RUNNABLE on that CPU.
6349 * This allows to properly represent the expected utilization of a CPU which
6350 * has just got a big task running since a long sleep period. At the same time
6351 * however it preserves the benefits of the "blocked utilization" in
6352 * describing the potential for other tasks waking up on the same CPU.
6354 * Nevertheless, cfs_rq.avg.util_avg can be higher than capacity_curr or even
6355 * higher than capacity_orig because of unfortunate rounding in
6356 * cfs.avg.util_avg or just after migrating tasks and new task wakeups until
6357 * the average stabilizes with the new running time. We need to check that the
6358 * utilization stays within the range of [0..capacity_orig] and cap it if
6359 * necessary. Without utilization capping, a group could be seen as overloaded
6360 * (CPU0 utilization at 121% + CPU1 utilization at 80%) whereas CPU1 has 20% of
6361 * available capacity. We allow utilization to overshoot capacity_curr (but not
6362 * capacity_orig) as it useful for predicting the capacity required after task
6363 * migrations (scheduler-driven DVFS).
6365 * Return: the (estimated) utilization for the specified CPU
6367 static inline unsigned long cpu_util(int cpu)
6369 struct cfs_rq *cfs_rq;
6372 cfs_rq = &cpu_rq(cpu)->cfs;
6373 util = READ_ONCE(cfs_rq->avg.util_avg);
6375 if (sched_feat(UTIL_EST))
6376 util = max(util, READ_ONCE(cfs_rq->avg.util_est.enqueued));
6378 return min_t(unsigned long, util, capacity_orig_of(cpu));
6382 * cpu_util_without: compute cpu utilization without any contributions from *p
6383 * @cpu: the CPU which utilization is requested
6384 * @p: the task which utilization should be discounted
6386 * The utilization of a CPU is defined by the utilization of tasks currently
6387 * enqueued on that CPU as well as tasks which are currently sleeping after an
6388 * execution on that CPU.
6390 * This method returns the utilization of the specified CPU by discounting the
6391 * utilization of the specified task, whenever the task is currently
6392 * contributing to the CPU utilization.
6394 static unsigned long cpu_util_without(int cpu, struct task_struct *p)
6396 struct cfs_rq *cfs_rq;
6399 /* Task has no contribution or is new */
6400 if (cpu != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time))
6401 return cpu_util(cpu);
6403 cfs_rq = &cpu_rq(cpu)->cfs;
6404 util = READ_ONCE(cfs_rq->avg.util_avg);
6406 /* Discount task's util from CPU's util */
6407 lsub_positive(&util, task_util(p));
6412 * a) if *p is the only task sleeping on this CPU, then:
6413 * cpu_util (== task_util) > util_est (== 0)
6414 * and thus we return:
6415 * cpu_util_without = (cpu_util - task_util) = 0
6417 * b) if other tasks are SLEEPING on this CPU, which is now exiting
6419 * cpu_util >= task_util
6420 * cpu_util > util_est (== 0)
6421 * and thus we discount *p's blocked utilization to return:
6422 * cpu_util_without = (cpu_util - task_util) >= 0
6424 * c) if other tasks are RUNNABLE on that CPU and
6425 * util_est > cpu_util
6426 * then we use util_est since it returns a more restrictive
6427 * estimation of the spare capacity on that CPU, by just
6428 * considering the expected utilization of tasks already
6429 * runnable on that CPU.
6431 * Cases a) and b) are covered by the above code, while case c) is
6432 * covered by the following code when estimated utilization is
6435 if (sched_feat(UTIL_EST)) {
6436 unsigned int estimated =
6437 READ_ONCE(cfs_rq->avg.util_est.enqueued);
6440 * Despite the following checks we still have a small window
6441 * for a possible race, when an execl's select_task_rq_fair()
6442 * races with LB's detach_task():
6445 * p->on_rq = TASK_ON_RQ_MIGRATING;
6446 * ---------------------------------- A
6447 * deactivate_task() \
6448 * dequeue_task() + RaceTime
6449 * util_est_dequeue() /
6450 * ---------------------------------- B
6452 * The additional check on "current == p" it's required to
6453 * properly fix the execl regression and it helps in further
6454 * reducing the chances for the above race.
6456 if (unlikely(task_on_rq_queued(p) || current == p))
6457 lsub_positive(&estimated, _task_util_est(p));
6459 util = max(util, estimated);
6463 * Utilization (estimated) can exceed the CPU capacity, thus let's
6464 * clamp to the maximum CPU capacity to ensure consistency with
6465 * the cpu_util call.
6467 return min_t(unsigned long, util, capacity_orig_of(cpu));
6471 * Predicts what cpu_util(@cpu) would return if @p was migrated (and enqueued)
6474 static unsigned long cpu_util_next(int cpu, struct task_struct *p, int dst_cpu)
6476 struct cfs_rq *cfs_rq = &cpu_rq(cpu)->cfs;
6477 unsigned long util_est, util = READ_ONCE(cfs_rq->avg.util_avg);
6480 * If @p migrates from @cpu to another, remove its contribution. Or,
6481 * if @p migrates from another CPU to @cpu, add its contribution. In
6482 * the other cases, @cpu is not impacted by the migration, so the
6483 * util_avg should already be correct.
6485 if (task_cpu(p) == cpu && dst_cpu != cpu)
6486 lsub_positive(&util, task_util(p));
6487 else if (task_cpu(p) != cpu && dst_cpu == cpu)
6488 util += task_util(p);
6490 if (sched_feat(UTIL_EST)) {
6491 util_est = READ_ONCE(cfs_rq->avg.util_est.enqueued);
6494 * During wake-up, the task isn't enqueued yet and doesn't
6495 * appear in the cfs_rq->avg.util_est.enqueued of any rq,
6496 * so just add it (if needed) to "simulate" what will be
6497 * cpu_util() after the task has been enqueued.
6500 util_est += _task_util_est(p);
6502 util = max(util, util_est);
6505 return min(util, capacity_orig_of(cpu));
6509 * compute_energy(): Estimates the energy that @pd would consume if @p was
6510 * migrated to @dst_cpu. compute_energy() predicts what will be the utilization
6511 * landscape of @pd's CPUs after the task migration, and uses the Energy Model
6512 * to compute what would be the energy if we decided to actually migrate that
6516 compute_energy(struct task_struct *p, int dst_cpu, struct perf_domain *pd)
6518 struct cpumask *pd_mask = perf_domain_span(pd);
6519 unsigned long cpu_cap = arch_scale_cpu_capacity(cpumask_first(pd_mask));
6520 unsigned long max_util = 0, sum_util = 0;
6524 * The capacity state of CPUs of the current rd can be driven by CPUs
6525 * of another rd if they belong to the same pd. So, account for the
6526 * utilization of these CPUs too by masking pd with cpu_online_mask
6527 * instead of the rd span.
6529 * If an entire pd is outside of the current rd, it will not appear in
6530 * its pd list and will not be accounted by compute_energy().
6532 for_each_cpu_and(cpu, pd_mask, cpu_online_mask) {
6533 unsigned long util_freq = cpu_util_next(cpu, p, dst_cpu);
6534 unsigned long cpu_util, util_running = util_freq;
6535 struct task_struct *tsk = NULL;
6538 * When @p is placed on @cpu:
6540 * util_running = max(cpu_util, cpu_util_est) +
6541 * max(task_util, _task_util_est)
6543 * while cpu_util_next is: max(cpu_util + task_util,
6544 * cpu_util_est + _task_util_est)
6546 if (cpu == dst_cpu) {
6549 cpu_util_next(cpu, p, -1) + task_util_est(p);
6553 * Busy time computation: utilization clamping is not
6554 * required since the ratio (sum_util / cpu_capacity)
6555 * is already enough to scale the EM reported power
6556 * consumption at the (eventually clamped) cpu_capacity.
6558 sum_util += effective_cpu_util(cpu, util_running, cpu_cap,
6562 * Performance domain frequency: utilization clamping
6563 * must be considered since it affects the selection
6564 * of the performance domain frequency.
6565 * NOTE: in case RT tasks are running, by default the
6566 * FREQUENCY_UTIL's utilization can be max OPP.
6568 cpu_util = effective_cpu_util(cpu, util_freq, cpu_cap,
6569 FREQUENCY_UTIL, tsk);
6570 max_util = max(max_util, cpu_util);
6573 return em_cpu_energy(pd->em_pd, max_util, sum_util);
6577 * find_energy_efficient_cpu(): Find most energy-efficient target CPU for the
6578 * waking task. find_energy_efficient_cpu() looks for the CPU with maximum
6579 * spare capacity in each performance domain and uses it as a potential
6580 * candidate to execute the task. Then, it uses the Energy Model to figure
6581 * out which of the CPU candidates is the most energy-efficient.
6583 * The rationale for this heuristic is as follows. In a performance domain,
6584 * all the most energy efficient CPU candidates (according to the Energy
6585 * Model) are those for which we'll request a low frequency. When there are
6586 * several CPUs for which the frequency request will be the same, we don't
6587 * have enough data to break the tie between them, because the Energy Model
6588 * only includes active power costs. With this model, if we assume that
6589 * frequency requests follow utilization (e.g. using schedutil), the CPU with
6590 * the maximum spare capacity in a performance domain is guaranteed to be among
6591 * the best candidates of the performance domain.
6593 * In practice, it could be preferable from an energy standpoint to pack
6594 * small tasks on a CPU in order to let other CPUs go in deeper idle states,
6595 * but that could also hurt our chances to go cluster idle, and we have no
6596 * ways to tell with the current Energy Model if this is actually a good
6597 * idea or not. So, find_energy_efficient_cpu() basically favors
6598 * cluster-packing, and spreading inside a cluster. That should at least be
6599 * a good thing for latency, and this is consistent with the idea that most
6600 * of the energy savings of EAS come from the asymmetry of the system, and
6601 * not so much from breaking the tie between identical CPUs. That's also the
6602 * reason why EAS is enabled in the topology code only for systems where
6603 * SD_ASYM_CPUCAPACITY is set.
6605 * NOTE: Forkees are not accepted in the energy-aware wake-up path because
6606 * they don't have any useful utilization data yet and it's not possible to
6607 * forecast their impact on energy consumption. Consequently, they will be
6608 * placed by find_idlest_cpu() on the least loaded CPU, which might turn out
6609 * to be energy-inefficient in some use-cases. The alternative would be to
6610 * bias new tasks towards specific types of CPUs first, or to try to infer
6611 * their util_avg from the parent task, but those heuristics could hurt
6612 * other use-cases too. So, until someone finds a better way to solve this,
6613 * let's keep things simple by re-using the existing slow path.
6615 static int find_energy_efficient_cpu(struct task_struct *p, int prev_cpu)
6617 unsigned long prev_delta = ULONG_MAX, best_delta = ULONG_MAX;
6618 struct root_domain *rd = cpu_rq(smp_processor_id())->rd;
6619 unsigned long cpu_cap, util, base_energy = 0;
6620 int cpu, best_energy_cpu = prev_cpu;
6621 struct sched_domain *sd;
6622 struct perf_domain *pd;
6625 pd = rcu_dereference(rd->pd);
6626 if (!pd || READ_ONCE(rd->overutilized))
6630 * Energy-aware wake-up happens on the lowest sched_domain starting
6631 * from sd_asym_cpucapacity spanning over this_cpu and prev_cpu.
6633 sd = rcu_dereference(*this_cpu_ptr(&sd_asym_cpucapacity));
6634 while (sd && !cpumask_test_cpu(prev_cpu, sched_domain_span(sd)))
6639 sync_entity_load_avg(&p->se);
6640 if (!task_util_est(p))
6643 for (; pd; pd = pd->next) {
6644 unsigned long cur_delta, spare_cap, max_spare_cap = 0;
6645 unsigned long base_energy_pd;
6646 int max_spare_cap_cpu = -1;
6648 /* Compute the 'base' energy of the pd, without @p */
6649 base_energy_pd = compute_energy(p, -1, pd);
6650 base_energy += base_energy_pd;
6652 for_each_cpu_and(cpu, perf_domain_span(pd), sched_domain_span(sd)) {
6653 if (!cpumask_test_cpu(cpu, p->cpus_ptr))
6656 util = cpu_util_next(cpu, p, cpu);
6657 cpu_cap = capacity_of(cpu);
6658 spare_cap = cpu_cap;
6659 lsub_positive(&spare_cap, util);
6662 * Skip CPUs that cannot satisfy the capacity request.
6663 * IOW, placing the task there would make the CPU
6664 * overutilized. Take uclamp into account to see how
6665 * much capacity we can get out of the CPU; this is
6666 * aligned with sched_cpu_util().
6668 util = uclamp_rq_util_with(cpu_rq(cpu), util, p);
6669 if (!fits_capacity(util, cpu_cap))
6672 /* Always use prev_cpu as a candidate. */
6673 if (cpu == prev_cpu) {
6674 prev_delta = compute_energy(p, prev_cpu, pd);
6675 prev_delta -= base_energy_pd;
6676 best_delta = min(best_delta, prev_delta);
6680 * Find the CPU with the maximum spare capacity in
6681 * the performance domain
6683 if (spare_cap > max_spare_cap) {
6684 max_spare_cap = spare_cap;
6685 max_spare_cap_cpu = cpu;
6689 /* Evaluate the energy impact of using this CPU. */
6690 if (max_spare_cap_cpu >= 0 && max_spare_cap_cpu != prev_cpu) {
6691 cur_delta = compute_energy(p, max_spare_cap_cpu, pd);
6692 cur_delta -= base_energy_pd;
6693 if (cur_delta < best_delta) {
6694 best_delta = cur_delta;
6695 best_energy_cpu = max_spare_cap_cpu;
6703 * Pick the best CPU if prev_cpu cannot be used, or if it saves at
6704 * least 6% of the energy used by prev_cpu.
6706 if (prev_delta == ULONG_MAX)
6707 return best_energy_cpu;
6709 if ((prev_delta - best_delta) > ((prev_delta + base_energy) >> 4))
6710 return best_energy_cpu;
6721 * select_task_rq_fair: Select target runqueue for the waking task in domains
6722 * that have the relevant SD flag set. In practice, this is SD_BALANCE_WAKE,
6723 * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
6725 * Balances load by selecting the idlest CPU in the idlest group, or under
6726 * certain conditions an idle sibling CPU if the domain has SD_WAKE_AFFINE set.
6728 * Returns the target CPU number.
6730 * preempt must be disabled.
6733 select_task_rq_fair(struct task_struct *p, int prev_cpu, int wake_flags)
6735 int sync = (wake_flags & WF_SYNC) && !(current->flags & PF_EXITING);
6736 struct sched_domain *tmp, *sd = NULL;
6737 int cpu = smp_processor_id();
6738 int new_cpu = prev_cpu;
6739 int want_affine = 0;
6740 /* SD_flags and WF_flags share the first nibble */
6741 int sd_flag = wake_flags & 0xF;
6743 if (wake_flags & WF_TTWU) {
6746 if (sched_energy_enabled()) {
6747 new_cpu = find_energy_efficient_cpu(p, prev_cpu);
6753 want_affine = !wake_wide(p) && cpumask_test_cpu(cpu, p->cpus_ptr);
6757 for_each_domain(cpu, tmp) {
6759 * If both 'cpu' and 'prev_cpu' are part of this domain,
6760 * cpu is a valid SD_WAKE_AFFINE target.
6762 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
6763 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
6764 if (cpu != prev_cpu)
6765 new_cpu = wake_affine(tmp, p, cpu, prev_cpu, sync);
6767 sd = NULL; /* Prefer wake_affine over balance flags */
6771 if (tmp->flags & sd_flag)
6773 else if (!want_affine)
6779 new_cpu = find_idlest_cpu(sd, p, cpu, prev_cpu, sd_flag);
6780 } else if (wake_flags & WF_TTWU) { /* XXX always ? */
6782 new_cpu = select_idle_sibling(p, prev_cpu, new_cpu);
6785 current->recent_used_cpu = cpu;
6792 static void detach_entity_cfs_rq(struct sched_entity *se);
6795 * Called immediately before a task is migrated to a new CPU; task_cpu(p) and
6796 * cfs_rq_of(p) references at time of call are still valid and identify the
6797 * previous CPU. The caller guarantees p->pi_lock or task_rq(p)->lock is held.
6799 static void migrate_task_rq_fair(struct task_struct *p, int new_cpu)
6802 * As blocked tasks retain absolute vruntime the migration needs to
6803 * deal with this by subtracting the old and adding the new
6804 * min_vruntime -- the latter is done by enqueue_entity() when placing
6805 * the task on the new runqueue.
6807 if (p->state == TASK_WAKING) {
6808 struct sched_entity *se = &p->se;
6809 struct cfs_rq *cfs_rq = cfs_rq_of(se);
6812 #ifndef CONFIG_64BIT
6813 u64 min_vruntime_copy;
6816 min_vruntime_copy = cfs_rq->min_vruntime_copy;
6818 min_vruntime = cfs_rq->min_vruntime;
6819 } while (min_vruntime != min_vruntime_copy);
6821 min_vruntime = cfs_rq->min_vruntime;
6824 se->vruntime -= min_vruntime;
6827 if (p->on_rq == TASK_ON_RQ_MIGRATING) {
6829 * In case of TASK_ON_RQ_MIGRATING we in fact hold the 'old'
6830 * rq->lock and can modify state directly.
6832 lockdep_assert_held(&task_rq(p)->lock);
6833 detach_entity_cfs_rq(&p->se);
6837 * We are supposed to update the task to "current" time, then
6838 * its up to date and ready to go to new CPU/cfs_rq. But we
6839 * have difficulty in getting what current time is, so simply
6840 * throw away the out-of-date time. This will result in the
6841 * wakee task is less decayed, but giving the wakee more load
6844 remove_entity_load_avg(&p->se);
6847 /* Tell new CPU we are migrated */
6848 p->se.avg.last_update_time = 0;
6850 /* We have migrated, no longer consider this task hot */
6851 p->se.exec_start = 0;
6853 update_scan_period(p, new_cpu);
6856 static void task_dead_fair(struct task_struct *p)
6858 remove_entity_load_avg(&p->se);
6862 balance_fair(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
6867 return newidle_balance(rq, rf) != 0;
6869 #endif /* CONFIG_SMP */
6871 static unsigned long wakeup_gran(struct sched_entity *se)
6873 unsigned long gran = sysctl_sched_wakeup_granularity;
6876 * Since its curr running now, convert the gran from real-time
6877 * to virtual-time in his units.
6879 * By using 'se' instead of 'curr' we penalize light tasks, so
6880 * they get preempted easier. That is, if 'se' < 'curr' then
6881 * the resulting gran will be larger, therefore penalizing the
6882 * lighter, if otoh 'se' > 'curr' then the resulting gran will
6883 * be smaller, again penalizing the lighter task.
6885 * This is especially important for buddies when the leftmost
6886 * task is higher priority than the buddy.
6888 return calc_delta_fair(gran, se);
6892 * Should 'se' preempt 'curr'.
6906 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
6908 s64 gran, vdiff = curr->vruntime - se->vruntime;
6913 gran = wakeup_gran(se);
6920 static void set_last_buddy(struct sched_entity *se)
6922 if (entity_is_task(se) && unlikely(task_has_idle_policy(task_of(se))))
6925 for_each_sched_entity(se) {
6926 if (SCHED_WARN_ON(!se->on_rq))
6928 cfs_rq_of(se)->last = se;
6932 static void set_next_buddy(struct sched_entity *se)
6934 if (entity_is_task(se) && unlikely(task_has_idle_policy(task_of(se))))
6937 for_each_sched_entity(se) {
6938 if (SCHED_WARN_ON(!se->on_rq))
6940 cfs_rq_of(se)->next = se;
6944 static void set_skip_buddy(struct sched_entity *se)
6946 for_each_sched_entity(se)
6947 cfs_rq_of(se)->skip = se;
6951 * Preempt the current task with a newly woken task if needed:
6953 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
6955 struct task_struct *curr = rq->curr;
6956 struct sched_entity *se = &curr->se, *pse = &p->se;
6957 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
6958 int scale = cfs_rq->nr_running >= sched_nr_latency;
6959 int next_buddy_marked = 0;
6961 if (unlikely(se == pse))
6965 * This is possible from callers such as attach_tasks(), in which we
6966 * unconditionally check_prempt_curr() after an enqueue (which may have
6967 * lead to a throttle). This both saves work and prevents false
6968 * next-buddy nomination below.
6970 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
6973 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
6974 set_next_buddy(pse);
6975 next_buddy_marked = 1;
6979 * We can come here with TIF_NEED_RESCHED already set from new task
6982 * Note: this also catches the edge-case of curr being in a throttled
6983 * group (e.g. via set_curr_task), since update_curr() (in the
6984 * enqueue of curr) will have resulted in resched being set. This
6985 * prevents us from potentially nominating it as a false LAST_BUDDY
6988 if (test_tsk_need_resched(curr))
6991 /* Idle tasks are by definition preempted by non-idle tasks. */
6992 if (unlikely(task_has_idle_policy(curr)) &&
6993 likely(!task_has_idle_policy(p)))
6997 * Batch and idle tasks do not preempt non-idle tasks (their preemption
6998 * is driven by the tick):
7000 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
7003 find_matching_se(&se, &pse);
7004 update_curr(cfs_rq_of(se));
7006 if (wakeup_preempt_entity(se, pse) == 1) {
7008 * Bias pick_next to pick the sched entity that is
7009 * triggering this preemption.
7011 if (!next_buddy_marked)
7012 set_next_buddy(pse);
7021 * Only set the backward buddy when the current task is still
7022 * on the rq. This can happen when a wakeup gets interleaved
7023 * with schedule on the ->pre_schedule() or idle_balance()
7024 * point, either of which can * drop the rq lock.
7026 * Also, during early boot the idle thread is in the fair class,
7027 * for obvious reasons its a bad idea to schedule back to it.
7029 if (unlikely(!se->on_rq || curr == rq->idle))
7032 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
7036 struct task_struct *
7037 pick_next_task_fair(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
7039 struct cfs_rq *cfs_rq = &rq->cfs;
7040 struct sched_entity *se;
7041 struct task_struct *p;
7045 if (!sched_fair_runnable(rq))
7048 #ifdef CONFIG_FAIR_GROUP_SCHED
7049 if (!prev || prev->sched_class != &fair_sched_class)
7053 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
7054 * likely that a next task is from the same cgroup as the current.
7056 * Therefore attempt to avoid putting and setting the entire cgroup
7057 * hierarchy, only change the part that actually changes.
7061 struct sched_entity *curr = cfs_rq->curr;
7064 * Since we got here without doing put_prev_entity() we also
7065 * have to consider cfs_rq->curr. If it is still a runnable
7066 * entity, update_curr() will update its vruntime, otherwise
7067 * forget we've ever seen it.
7071 update_curr(cfs_rq);
7076 * This call to check_cfs_rq_runtime() will do the
7077 * throttle and dequeue its entity in the parent(s).
7078 * Therefore the nr_running test will indeed
7081 if (unlikely(check_cfs_rq_runtime(cfs_rq))) {
7084 if (!cfs_rq->nr_running)
7091 se = pick_next_entity(cfs_rq, curr);
7092 cfs_rq = group_cfs_rq(se);
7098 * Since we haven't yet done put_prev_entity and if the selected task
7099 * is a different task than we started out with, try and touch the
7100 * least amount of cfs_rqs.
7103 struct sched_entity *pse = &prev->se;
7105 while (!(cfs_rq = is_same_group(se, pse))) {
7106 int se_depth = se->depth;
7107 int pse_depth = pse->depth;
7109 if (se_depth <= pse_depth) {
7110 put_prev_entity(cfs_rq_of(pse), pse);
7111 pse = parent_entity(pse);
7113 if (se_depth >= pse_depth) {
7114 set_next_entity(cfs_rq_of(se), se);
7115 se = parent_entity(se);
7119 put_prev_entity(cfs_rq, pse);
7120 set_next_entity(cfs_rq, se);
7127 put_prev_task(rq, prev);
7130 se = pick_next_entity(cfs_rq, NULL);
7131 set_next_entity(cfs_rq, se);
7132 cfs_rq = group_cfs_rq(se);
7137 done: __maybe_unused;
7140 * Move the next running task to the front of
7141 * the list, so our cfs_tasks list becomes MRU
7144 list_move(&p->se.group_node, &rq->cfs_tasks);
7147 if (hrtick_enabled_fair(rq))
7148 hrtick_start_fair(rq, p);
7150 update_misfit_status(p, rq);
7158 new_tasks = newidle_balance(rq, rf);
7161 * Because newidle_balance() releases (and re-acquires) rq->lock, it is
7162 * possible for any higher priority task to appear. In that case we
7163 * must re-start the pick_next_entity() loop.
7172 * rq is about to be idle, check if we need to update the
7173 * lost_idle_time of clock_pelt
7175 update_idle_rq_clock_pelt(rq);
7180 static struct task_struct *__pick_next_task_fair(struct rq *rq)
7182 return pick_next_task_fair(rq, NULL, NULL);
7186 * Account for a descheduled task:
7188 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
7190 struct sched_entity *se = &prev->se;
7191 struct cfs_rq *cfs_rq;
7193 for_each_sched_entity(se) {
7194 cfs_rq = cfs_rq_of(se);
7195 put_prev_entity(cfs_rq, se);
7200 * sched_yield() is very simple
7202 * The magic of dealing with the ->skip buddy is in pick_next_entity.
7204 static void yield_task_fair(struct rq *rq)
7206 struct task_struct *curr = rq->curr;
7207 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
7208 struct sched_entity *se = &curr->se;
7211 * Are we the only task in the tree?
7213 if (unlikely(rq->nr_running == 1))
7216 clear_buddies(cfs_rq, se);
7218 if (curr->policy != SCHED_BATCH) {
7219 update_rq_clock(rq);
7221 * Update run-time statistics of the 'current'.
7223 update_curr(cfs_rq);
7225 * Tell update_rq_clock() that we've just updated,
7226 * so we don't do microscopic update in schedule()
7227 * and double the fastpath cost.
7229 rq_clock_skip_update(rq);
7235 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p)
7237 struct sched_entity *se = &p->se;
7239 /* throttled hierarchies are not runnable */
7240 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
7243 /* Tell the scheduler that we'd really like pse to run next. */
7246 yield_task_fair(rq);
7252 /**************************************************
7253 * Fair scheduling class load-balancing methods.
7257 * The purpose of load-balancing is to achieve the same basic fairness the
7258 * per-CPU scheduler provides, namely provide a proportional amount of compute
7259 * time to each task. This is expressed in the following equation:
7261 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
7263 * Where W_i,n is the n-th weight average for CPU i. The instantaneous weight
7264 * W_i,0 is defined as:
7266 * W_i,0 = \Sum_j w_i,j (2)
7268 * Where w_i,j is the weight of the j-th runnable task on CPU i. This weight
7269 * is derived from the nice value as per sched_prio_to_weight[].
7271 * The weight average is an exponential decay average of the instantaneous
7274 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
7276 * C_i is the compute capacity of CPU i, typically it is the
7277 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
7278 * can also include other factors [XXX].
7280 * To achieve this balance we define a measure of imbalance which follows
7281 * directly from (1):
7283 * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4)
7285 * We them move tasks around to minimize the imbalance. In the continuous
7286 * function space it is obvious this converges, in the discrete case we get
7287 * a few fun cases generally called infeasible weight scenarios.
7290 * - infeasible weights;
7291 * - local vs global optima in the discrete case. ]
7296 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
7297 * for all i,j solution, we create a tree of CPUs that follows the hardware
7298 * topology where each level pairs two lower groups (or better). This results
7299 * in O(log n) layers. Furthermore we reduce the number of CPUs going up the
7300 * tree to only the first of the previous level and we decrease the frequency
7301 * of load-balance at each level inv. proportional to the number of CPUs in
7307 * \Sum { --- * --- * 2^i } = O(n) (5)
7309 * `- size of each group
7310 * | | `- number of CPUs doing load-balance
7312 * `- sum over all levels
7314 * Coupled with a limit on how many tasks we can migrate every balance pass,
7315 * this makes (5) the runtime complexity of the balancer.
7317 * An important property here is that each CPU is still (indirectly) connected
7318 * to every other CPU in at most O(log n) steps:
7320 * The adjacency matrix of the resulting graph is given by:
7323 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
7326 * And you'll find that:
7328 * A^(log_2 n)_i,j != 0 for all i,j (7)
7330 * Showing there's indeed a path between every CPU in at most O(log n) steps.
7331 * The task movement gives a factor of O(m), giving a convergence complexity
7334 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
7339 * In order to avoid CPUs going idle while there's still work to do, new idle
7340 * balancing is more aggressive and has the newly idle CPU iterate up the domain
7341 * tree itself instead of relying on other CPUs to bring it work.
7343 * This adds some complexity to both (5) and (8) but it reduces the total idle
7351 * Cgroups make a horror show out of (2), instead of a simple sum we get:
7354 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
7359 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
7361 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on CPU i.
7363 * The big problem is S_k, its a global sum needed to compute a local (W_i)
7366 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
7367 * rewrite all of this once again.]
7370 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
7372 enum fbq_type { regular, remote, all };
7375 * 'group_type' describes the group of CPUs at the moment of load balancing.
7377 * The enum is ordered by pulling priority, with the group with lowest priority
7378 * first so the group_type can simply be compared when selecting the busiest
7379 * group. See update_sd_pick_busiest().
7382 /* The group has spare capacity that can be used to run more tasks. */
7383 group_has_spare = 0,
7385 * The group is fully used and the tasks don't compete for more CPU
7386 * cycles. Nevertheless, some tasks might wait before running.
7390 * SD_ASYM_CPUCAPACITY only: One task doesn't fit with CPU's capacity
7391 * and must be migrated to a more powerful CPU.
7395 * SD_ASYM_PACKING only: One local CPU with higher capacity is available,
7396 * and the task should be migrated to it instead of running on the
7401 * The tasks' affinity constraints previously prevented the scheduler
7402 * from balancing the load across the system.
7406 * The CPU is overloaded and can't provide expected CPU cycles to all
7412 enum migration_type {
7419 #define LBF_ALL_PINNED 0x01
7420 #define LBF_NEED_BREAK 0x02
7421 #define LBF_DST_PINNED 0x04
7422 #define LBF_SOME_PINNED 0x08
7425 struct sched_domain *sd;
7433 struct cpumask *dst_grpmask;
7435 enum cpu_idle_type idle;
7437 /* The set of CPUs under consideration for load-balancing */
7438 struct cpumask *cpus;
7443 unsigned int loop_break;
7444 unsigned int loop_max;
7446 enum fbq_type fbq_type;
7447 enum migration_type migration_type;
7448 struct list_head tasks;
7452 * Is this task likely cache-hot:
7454 static int task_hot(struct task_struct *p, struct lb_env *env)
7458 lockdep_assert_held(&env->src_rq->lock);
7460 if (p->sched_class != &fair_sched_class)
7463 if (unlikely(task_has_idle_policy(p)))
7466 /* SMT siblings share cache */
7467 if (env->sd->flags & SD_SHARE_CPUCAPACITY)
7471 * Buddy candidates are cache hot:
7473 if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running &&
7474 (&p->se == cfs_rq_of(&p->se)->next ||
7475 &p->se == cfs_rq_of(&p->se)->last))
7478 if (sysctl_sched_migration_cost == -1)
7480 if (sysctl_sched_migration_cost == 0)
7483 delta = rq_clock_task(env->src_rq) - p->se.exec_start;
7485 return delta < (s64)sysctl_sched_migration_cost;
7488 #ifdef CONFIG_NUMA_BALANCING
7490 * Returns 1, if task migration degrades locality
7491 * Returns 0, if task migration improves locality i.e migration preferred.
7492 * Returns -1, if task migration is not affected by locality.
7494 static int migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
7496 struct numa_group *numa_group = rcu_dereference(p->numa_group);
7497 unsigned long src_weight, dst_weight;
7498 int src_nid, dst_nid, dist;
7500 if (!static_branch_likely(&sched_numa_balancing))
7503 if (!p->numa_faults || !(env->sd->flags & SD_NUMA))
7506 src_nid = cpu_to_node(env->src_cpu);
7507 dst_nid = cpu_to_node(env->dst_cpu);
7509 if (src_nid == dst_nid)
7512 /* Migrating away from the preferred node is always bad. */
7513 if (src_nid == p->numa_preferred_nid) {
7514 if (env->src_rq->nr_running > env->src_rq->nr_preferred_running)
7520 /* Encourage migration to the preferred node. */
7521 if (dst_nid == p->numa_preferred_nid)
7524 /* Leaving a core idle is often worse than degrading locality. */
7525 if (env->idle == CPU_IDLE)
7528 dist = node_distance(src_nid, dst_nid);
7530 src_weight = group_weight(p, src_nid, dist);
7531 dst_weight = group_weight(p, dst_nid, dist);
7533 src_weight = task_weight(p, src_nid, dist);
7534 dst_weight = task_weight(p, dst_nid, dist);
7537 return dst_weight < src_weight;
7541 static inline int migrate_degrades_locality(struct task_struct *p,
7549 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
7552 int can_migrate_task(struct task_struct *p, struct lb_env *env)
7556 lockdep_assert_held(&env->src_rq->lock);
7559 * We do not migrate tasks that are:
7560 * 1) throttled_lb_pair, or
7561 * 2) cannot be migrated to this CPU due to cpus_ptr, or
7562 * 3) running (obviously), or
7563 * 4) are cache-hot on their current CPU.
7565 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
7568 if (!cpumask_test_cpu(env->dst_cpu, p->cpus_ptr)) {
7571 schedstat_inc(p->se.statistics.nr_failed_migrations_affine);
7573 env->flags |= LBF_SOME_PINNED;
7576 * Remember if this task can be migrated to any other CPU in
7577 * our sched_group. We may want to revisit it if we couldn't
7578 * meet load balance goals by pulling other tasks on src_cpu.
7580 * Avoid computing new_dst_cpu for NEWLY_IDLE or if we have
7581 * already computed one in current iteration.
7583 if (env->idle == CPU_NEWLY_IDLE || (env->flags & LBF_DST_PINNED))
7586 /* Prevent to re-select dst_cpu via env's CPUs: */
7587 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
7588 if (cpumask_test_cpu(cpu, p->cpus_ptr)) {
7589 env->flags |= LBF_DST_PINNED;
7590 env->new_dst_cpu = cpu;
7598 /* Record that we found atleast one task that could run on dst_cpu */
7599 env->flags &= ~LBF_ALL_PINNED;
7601 if (task_running(env->src_rq, p)) {
7602 schedstat_inc(p->se.statistics.nr_failed_migrations_running);
7607 * Aggressive migration if:
7608 * 1) destination numa is preferred
7609 * 2) task is cache cold, or
7610 * 3) too many balance attempts have failed.
7612 tsk_cache_hot = migrate_degrades_locality(p, env);
7613 if (tsk_cache_hot == -1)
7614 tsk_cache_hot = task_hot(p, env);
7616 if (tsk_cache_hot <= 0 ||
7617 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
7618 if (tsk_cache_hot == 1) {
7619 schedstat_inc(env->sd->lb_hot_gained[env->idle]);
7620 schedstat_inc(p->se.statistics.nr_forced_migrations);
7625 schedstat_inc(p->se.statistics.nr_failed_migrations_hot);
7630 * detach_task() -- detach the task for the migration specified in env
7632 static void detach_task(struct task_struct *p, struct lb_env *env)
7634 lockdep_assert_held(&env->src_rq->lock);
7636 deactivate_task(env->src_rq, p, DEQUEUE_NOCLOCK);
7637 set_task_cpu(p, env->dst_cpu);
7641 * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
7642 * part of active balancing operations within "domain".
7644 * Returns a task if successful and NULL otherwise.
7646 static struct task_struct *detach_one_task(struct lb_env *env)
7648 struct task_struct *p;
7650 lockdep_assert_held(&env->src_rq->lock);
7652 list_for_each_entry_reverse(p,
7653 &env->src_rq->cfs_tasks, se.group_node) {
7654 if (!can_migrate_task(p, env))
7657 detach_task(p, env);
7660 * Right now, this is only the second place where
7661 * lb_gained[env->idle] is updated (other is detach_tasks)
7662 * so we can safely collect stats here rather than
7663 * inside detach_tasks().
7665 schedstat_inc(env->sd->lb_gained[env->idle]);
7671 static const unsigned int sched_nr_migrate_break = 32;
7674 * detach_tasks() -- tries to detach up to imbalance load/util/tasks from
7675 * busiest_rq, as part of a balancing operation within domain "sd".
7677 * Returns number of detached tasks if successful and 0 otherwise.
7679 static int detach_tasks(struct lb_env *env)
7681 struct list_head *tasks = &env->src_rq->cfs_tasks;
7682 unsigned long util, load;
7683 struct task_struct *p;
7686 lockdep_assert_held(&env->src_rq->lock);
7688 if (env->imbalance <= 0)
7691 while (!list_empty(tasks)) {
7693 * We don't want to steal all, otherwise we may be treated likewise,
7694 * which could at worst lead to a livelock crash.
7696 if (env->idle != CPU_NOT_IDLE && env->src_rq->nr_running <= 1)
7699 p = list_last_entry(tasks, struct task_struct, se.group_node);
7702 /* We've more or less seen every task there is, call it quits */
7703 if (env->loop > env->loop_max)
7706 /* take a breather every nr_migrate tasks */
7707 if (env->loop > env->loop_break) {
7708 env->loop_break += sched_nr_migrate_break;
7709 env->flags |= LBF_NEED_BREAK;
7713 if (!can_migrate_task(p, env))
7716 switch (env->migration_type) {
7719 * Depending of the number of CPUs and tasks and the
7720 * cgroup hierarchy, task_h_load() can return a null
7721 * value. Make sure that env->imbalance decreases
7722 * otherwise detach_tasks() will stop only after
7723 * detaching up to loop_max tasks.
7725 load = max_t(unsigned long, task_h_load(p), 1);
7727 if (sched_feat(LB_MIN) &&
7728 load < 16 && !env->sd->nr_balance_failed)
7732 * Make sure that we don't migrate too much load.
7733 * Nevertheless, let relax the constraint if
7734 * scheduler fails to find a good waiting task to
7737 if (shr_bound(load, env->sd->nr_balance_failed) > env->imbalance)
7740 env->imbalance -= load;
7744 util = task_util_est(p);
7746 if (util > env->imbalance)
7749 env->imbalance -= util;
7756 case migrate_misfit:
7757 /* This is not a misfit task */
7758 if (task_fits_capacity(p, capacity_of(env->src_cpu)))
7765 detach_task(p, env);
7766 list_add(&p->se.group_node, &env->tasks);
7770 #ifdef CONFIG_PREEMPTION
7772 * NEWIDLE balancing is a source of latency, so preemptible
7773 * kernels will stop after the first task is detached to minimize
7774 * the critical section.
7776 if (env->idle == CPU_NEWLY_IDLE)
7781 * We only want to steal up to the prescribed amount of
7784 if (env->imbalance <= 0)
7789 list_move(&p->se.group_node, tasks);
7793 * Right now, this is one of only two places we collect this stat
7794 * so we can safely collect detach_one_task() stats here rather
7795 * than inside detach_one_task().
7797 schedstat_add(env->sd->lb_gained[env->idle], detached);
7803 * attach_task() -- attach the task detached by detach_task() to its new rq.
7805 static void attach_task(struct rq *rq, struct task_struct *p)
7807 lockdep_assert_held(&rq->lock);
7809 BUG_ON(task_rq(p) != rq);
7810 activate_task(rq, p, ENQUEUE_NOCLOCK);
7811 check_preempt_curr(rq, p, 0);
7815 * attach_one_task() -- attaches the task returned from detach_one_task() to
7818 static void attach_one_task(struct rq *rq, struct task_struct *p)
7823 update_rq_clock(rq);
7829 * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
7832 static void attach_tasks(struct lb_env *env)
7834 struct list_head *tasks = &env->tasks;
7835 struct task_struct *p;
7838 rq_lock(env->dst_rq, &rf);
7839 update_rq_clock(env->dst_rq);
7841 while (!list_empty(tasks)) {
7842 p = list_first_entry(tasks, struct task_struct, se.group_node);
7843 list_del_init(&p->se.group_node);
7845 attach_task(env->dst_rq, p);
7848 rq_unlock(env->dst_rq, &rf);
7851 #ifdef CONFIG_NO_HZ_COMMON
7852 static inline bool cfs_rq_has_blocked(struct cfs_rq *cfs_rq)
7854 if (cfs_rq->avg.load_avg)
7857 if (cfs_rq->avg.util_avg)
7863 static inline bool others_have_blocked(struct rq *rq)
7865 if (READ_ONCE(rq->avg_rt.util_avg))
7868 if (READ_ONCE(rq->avg_dl.util_avg))
7871 if (thermal_load_avg(rq))
7874 #ifdef CONFIG_HAVE_SCHED_AVG_IRQ
7875 if (READ_ONCE(rq->avg_irq.util_avg))
7882 static inline void update_blocked_load_tick(struct rq *rq)
7884 WRITE_ONCE(rq->last_blocked_load_update_tick, jiffies);
7887 static inline void update_blocked_load_status(struct rq *rq, bool has_blocked)
7890 rq->has_blocked_load = 0;
7893 static inline bool cfs_rq_has_blocked(struct cfs_rq *cfs_rq) { return false; }
7894 static inline bool others_have_blocked(struct rq *rq) { return false; }
7895 static inline void update_blocked_load_tick(struct rq *rq) {}
7896 static inline void update_blocked_load_status(struct rq *rq, bool has_blocked) {}
7899 static bool __update_blocked_others(struct rq *rq, bool *done)
7901 const struct sched_class *curr_class;
7902 u64 now = rq_clock_pelt(rq);
7903 unsigned long thermal_pressure;
7907 * update_load_avg() can call cpufreq_update_util(). Make sure that RT,
7908 * DL and IRQ signals have been updated before updating CFS.
7910 curr_class = rq->curr->sched_class;
7912 thermal_pressure = arch_scale_thermal_pressure(cpu_of(rq));
7914 decayed = update_rt_rq_load_avg(now, rq, curr_class == &rt_sched_class) |
7915 update_dl_rq_load_avg(now, rq, curr_class == &dl_sched_class) |
7916 update_thermal_load_avg(rq_clock_thermal(rq), rq, thermal_pressure) |
7917 update_irq_load_avg(rq, 0);
7919 if (others_have_blocked(rq))
7925 #ifdef CONFIG_FAIR_GROUP_SCHED
7927 static inline bool cfs_rq_is_decayed(struct cfs_rq *cfs_rq)
7929 if (cfs_rq->load.weight)
7932 if (cfs_rq->avg.load_sum)
7935 if (cfs_rq->avg.util_sum)
7938 if (cfs_rq->avg.runnable_sum)
7944 static bool __update_blocked_fair(struct rq *rq, bool *done)
7946 struct cfs_rq *cfs_rq, *pos;
7947 bool decayed = false;
7948 int cpu = cpu_of(rq);
7951 * Iterates the task_group tree in a bottom up fashion, see
7952 * list_add_leaf_cfs_rq() for details.
7954 for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) {
7955 struct sched_entity *se;
7957 if (update_cfs_rq_load_avg(cfs_rq_clock_pelt(cfs_rq), cfs_rq)) {
7958 update_tg_load_avg(cfs_rq);
7960 if (cfs_rq == &rq->cfs)
7964 /* Propagate pending load changes to the parent, if any: */
7965 se = cfs_rq->tg->se[cpu];
7966 if (se && !skip_blocked_update(se))
7967 update_load_avg(cfs_rq_of(se), se, 0);
7970 * There can be a lot of idle CPU cgroups. Don't let fully
7971 * decayed cfs_rqs linger on the list.
7973 if (cfs_rq_is_decayed(cfs_rq))
7974 list_del_leaf_cfs_rq(cfs_rq);
7976 /* Don't need periodic decay once load/util_avg are null */
7977 if (cfs_rq_has_blocked(cfs_rq))
7985 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
7986 * This needs to be done in a top-down fashion because the load of a child
7987 * group is a fraction of its parents load.
7989 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
7991 struct rq *rq = rq_of(cfs_rq);
7992 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
7993 unsigned long now = jiffies;
7996 if (cfs_rq->last_h_load_update == now)
7999 WRITE_ONCE(cfs_rq->h_load_next, NULL);
8000 for_each_sched_entity(se) {
8001 cfs_rq = cfs_rq_of(se);
8002 WRITE_ONCE(cfs_rq->h_load_next, se);
8003 if (cfs_rq->last_h_load_update == now)
8008 cfs_rq->h_load = cfs_rq_load_avg(cfs_rq);
8009 cfs_rq->last_h_load_update = now;
8012 while ((se = READ_ONCE(cfs_rq->h_load_next)) != NULL) {
8013 load = cfs_rq->h_load;
8014 load = div64_ul(load * se->avg.load_avg,
8015 cfs_rq_load_avg(cfs_rq) + 1);
8016 cfs_rq = group_cfs_rq(se);
8017 cfs_rq->h_load = load;
8018 cfs_rq->last_h_load_update = now;
8022 static unsigned long task_h_load(struct task_struct *p)
8024 struct cfs_rq *cfs_rq = task_cfs_rq(p);
8026 update_cfs_rq_h_load(cfs_rq);
8027 return div64_ul(p->se.avg.load_avg * cfs_rq->h_load,
8028 cfs_rq_load_avg(cfs_rq) + 1);
8031 static bool __update_blocked_fair(struct rq *rq, bool *done)
8033 struct cfs_rq *cfs_rq = &rq->cfs;
8036 decayed = update_cfs_rq_load_avg(cfs_rq_clock_pelt(cfs_rq), cfs_rq);
8037 if (cfs_rq_has_blocked(cfs_rq))
8043 static unsigned long task_h_load(struct task_struct *p)
8045 return p->se.avg.load_avg;
8049 static void update_blocked_averages(int cpu)
8051 bool decayed = false, done = true;
8052 struct rq *rq = cpu_rq(cpu);
8055 rq_lock_irqsave(rq, &rf);
8056 update_blocked_load_tick(rq);
8057 update_rq_clock(rq);
8059 decayed |= __update_blocked_others(rq, &done);
8060 decayed |= __update_blocked_fair(rq, &done);
8062 update_blocked_load_status(rq, !done);
8064 cpufreq_update_util(rq, 0);
8065 rq_unlock_irqrestore(rq, &rf);
8068 /********** Helpers for find_busiest_group ************************/
8071 * sg_lb_stats - stats of a sched_group required for load_balancing
8073 struct sg_lb_stats {
8074 unsigned long avg_load; /*Avg load across the CPUs of the group */
8075 unsigned long group_load; /* Total load over the CPUs of the group */
8076 unsigned long group_capacity;
8077 unsigned long group_util; /* Total utilization over the CPUs of the group */
8078 unsigned long group_runnable; /* Total runnable time over the CPUs of the group */
8079 unsigned int sum_nr_running; /* Nr of tasks running in the group */
8080 unsigned int sum_h_nr_running; /* Nr of CFS tasks running in the group */
8081 unsigned int idle_cpus;
8082 unsigned int group_weight;
8083 enum group_type group_type;
8084 unsigned int group_asym_packing; /* Tasks should be moved to preferred CPU */
8085 unsigned long group_misfit_task_load; /* A CPU has a task too big for its capacity */
8086 #ifdef CONFIG_NUMA_BALANCING
8087 unsigned int nr_numa_running;
8088 unsigned int nr_preferred_running;
8093 * sd_lb_stats - Structure to store the statistics of a sched_domain
8094 * during load balancing.
8096 struct sd_lb_stats {
8097 struct sched_group *busiest; /* Busiest group in this sd */
8098 struct sched_group *local; /* Local group in this sd */
8099 unsigned long total_load; /* Total load of all groups in sd */
8100 unsigned long total_capacity; /* Total capacity of all groups in sd */
8101 unsigned long avg_load; /* Average load across all groups in sd */
8102 unsigned int prefer_sibling; /* tasks should go to sibling first */
8104 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
8105 struct sg_lb_stats local_stat; /* Statistics of the local group */
8108 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
8111 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
8112 * local_stat because update_sg_lb_stats() does a full clear/assignment.
8113 * We must however set busiest_stat::group_type and
8114 * busiest_stat::idle_cpus to the worst busiest group because
8115 * update_sd_pick_busiest() reads these before assignment.
8117 *sds = (struct sd_lb_stats){
8121 .total_capacity = 0UL,
8123 .idle_cpus = UINT_MAX,
8124 .group_type = group_has_spare,
8129 static unsigned long scale_rt_capacity(int cpu)
8131 struct rq *rq = cpu_rq(cpu);
8132 unsigned long max = arch_scale_cpu_capacity(cpu);
8133 unsigned long used, free;
8136 irq = cpu_util_irq(rq);
8138 if (unlikely(irq >= max))
8142 * avg_rt.util_avg and avg_dl.util_avg track binary signals
8143 * (running and not running) with weights 0 and 1024 respectively.
8144 * avg_thermal.load_avg tracks thermal pressure and the weighted
8145 * average uses the actual delta max capacity(load).
8147 used = READ_ONCE(rq->avg_rt.util_avg);
8148 used += READ_ONCE(rq->avg_dl.util_avg);
8149 used += thermal_load_avg(rq);
8151 if (unlikely(used >= max))
8156 return scale_irq_capacity(free, irq, max);
8159 static void update_cpu_capacity(struct sched_domain *sd, int cpu)
8161 unsigned long capacity = scale_rt_capacity(cpu);
8162 struct sched_group *sdg = sd->groups;
8164 cpu_rq(cpu)->cpu_capacity_orig = arch_scale_cpu_capacity(cpu);
8169 cpu_rq(cpu)->cpu_capacity = capacity;
8170 trace_sched_cpu_capacity_tp(cpu_rq(cpu));
8172 sdg->sgc->capacity = capacity;
8173 sdg->sgc->min_capacity = capacity;
8174 sdg->sgc->max_capacity = capacity;
8177 void update_group_capacity(struct sched_domain *sd, int cpu)
8179 struct sched_domain *child = sd->child;
8180 struct sched_group *group, *sdg = sd->groups;
8181 unsigned long capacity, min_capacity, max_capacity;
8182 unsigned long interval;
8184 interval = msecs_to_jiffies(sd->balance_interval);
8185 interval = clamp(interval, 1UL, max_load_balance_interval);
8186 sdg->sgc->next_update = jiffies + interval;
8189 update_cpu_capacity(sd, cpu);
8194 min_capacity = ULONG_MAX;
8197 if (child->flags & SD_OVERLAP) {
8199 * SD_OVERLAP domains cannot assume that child groups
8200 * span the current group.
8203 for_each_cpu(cpu, sched_group_span(sdg)) {
8204 unsigned long cpu_cap = capacity_of(cpu);
8206 capacity += cpu_cap;
8207 min_capacity = min(cpu_cap, min_capacity);
8208 max_capacity = max(cpu_cap, max_capacity);
8212 * !SD_OVERLAP domains can assume that child groups
8213 * span the current group.
8216 group = child->groups;
8218 struct sched_group_capacity *sgc = group->sgc;
8220 capacity += sgc->capacity;
8221 min_capacity = min(sgc->min_capacity, min_capacity);
8222 max_capacity = max(sgc->max_capacity, max_capacity);
8223 group = group->next;
8224 } while (group != child->groups);
8227 sdg->sgc->capacity = capacity;
8228 sdg->sgc->min_capacity = min_capacity;
8229 sdg->sgc->max_capacity = max_capacity;
8233 * Check whether the capacity of the rq has been noticeably reduced by side
8234 * activity. The imbalance_pct is used for the threshold.
8235 * Return true is the capacity is reduced
8238 check_cpu_capacity(struct rq *rq, struct sched_domain *sd)
8240 return ((rq->cpu_capacity * sd->imbalance_pct) <
8241 (rq->cpu_capacity_orig * 100));
8245 * Check whether a rq has a misfit task and if it looks like we can actually
8246 * help that task: we can migrate the task to a CPU of higher capacity, or
8247 * the task's current CPU is heavily pressured.
8249 static inline int check_misfit_status(struct rq *rq, struct sched_domain *sd)
8251 return rq->misfit_task_load &&
8252 (rq->cpu_capacity_orig < rq->rd->max_cpu_capacity ||
8253 check_cpu_capacity(rq, sd));
8257 * Group imbalance indicates (and tries to solve) the problem where balancing
8258 * groups is inadequate due to ->cpus_ptr constraints.
8260 * Imagine a situation of two groups of 4 CPUs each and 4 tasks each with a
8261 * cpumask covering 1 CPU of the first group and 3 CPUs of the second group.
8264 * { 0 1 2 3 } { 4 5 6 7 }
8267 * If we were to balance group-wise we'd place two tasks in the first group and
8268 * two tasks in the second group. Clearly this is undesired as it will overload
8269 * cpu 3 and leave one of the CPUs in the second group unused.
8271 * The current solution to this issue is detecting the skew in the first group
8272 * by noticing the lower domain failed to reach balance and had difficulty
8273 * moving tasks due to affinity constraints.
8275 * When this is so detected; this group becomes a candidate for busiest; see
8276 * update_sd_pick_busiest(). And calculate_imbalance() and
8277 * find_busiest_group() avoid some of the usual balance conditions to allow it
8278 * to create an effective group imbalance.
8280 * This is a somewhat tricky proposition since the next run might not find the
8281 * group imbalance and decide the groups need to be balanced again. A most
8282 * subtle and fragile situation.
8285 static inline int sg_imbalanced(struct sched_group *group)
8287 return group->sgc->imbalance;
8291 * group_has_capacity returns true if the group has spare capacity that could
8292 * be used by some tasks.
8293 * We consider that a group has spare capacity if the * number of task is
8294 * smaller than the number of CPUs or if the utilization is lower than the
8295 * available capacity for CFS tasks.
8296 * For the latter, we use a threshold to stabilize the state, to take into
8297 * account the variance of the tasks' load and to return true if the available
8298 * capacity in meaningful for the load balancer.
8299 * As an example, an available capacity of 1% can appear but it doesn't make
8300 * any benefit for the load balance.
8303 group_has_capacity(unsigned int imbalance_pct, struct sg_lb_stats *sgs)
8305 if (sgs->sum_nr_running < sgs->group_weight)
8308 if ((sgs->group_capacity * imbalance_pct) <
8309 (sgs->group_runnable * 100))
8312 if ((sgs->group_capacity * 100) >
8313 (sgs->group_util * imbalance_pct))
8320 * group_is_overloaded returns true if the group has more tasks than it can
8322 * group_is_overloaded is not equals to !group_has_capacity because a group
8323 * with the exact right number of tasks, has no more spare capacity but is not
8324 * overloaded so both group_has_capacity and group_is_overloaded return
8328 group_is_overloaded(unsigned int imbalance_pct, struct sg_lb_stats *sgs)
8330 if (sgs->sum_nr_running <= sgs->group_weight)
8333 if ((sgs->group_capacity * 100) <
8334 (sgs->group_util * imbalance_pct))
8337 if ((sgs->group_capacity * imbalance_pct) <
8338 (sgs->group_runnable * 100))
8345 * group_smaller_min_cpu_capacity: Returns true if sched_group sg has smaller
8346 * per-CPU capacity than sched_group ref.
8349 group_smaller_min_cpu_capacity(struct sched_group *sg, struct sched_group *ref)
8351 return fits_capacity(sg->sgc->min_capacity, ref->sgc->min_capacity);
8355 * group_smaller_max_cpu_capacity: Returns true if sched_group sg has smaller
8356 * per-CPU capacity_orig than sched_group ref.
8359 group_smaller_max_cpu_capacity(struct sched_group *sg, struct sched_group *ref)
8361 return fits_capacity(sg->sgc->max_capacity, ref->sgc->max_capacity);
8365 group_type group_classify(unsigned int imbalance_pct,
8366 struct sched_group *group,
8367 struct sg_lb_stats *sgs)
8369 if (group_is_overloaded(imbalance_pct, sgs))
8370 return group_overloaded;
8372 if (sg_imbalanced(group))
8373 return group_imbalanced;
8375 if (sgs->group_asym_packing)
8376 return group_asym_packing;
8378 if (sgs->group_misfit_task_load)
8379 return group_misfit_task;
8381 if (!group_has_capacity(imbalance_pct, sgs))
8382 return group_fully_busy;
8384 return group_has_spare;
8387 static bool update_nohz_stats(struct rq *rq)
8389 #ifdef CONFIG_NO_HZ_COMMON
8390 unsigned int cpu = rq->cpu;
8392 if (!rq->has_blocked_load)
8395 if (!cpumask_test_cpu(cpu, nohz.idle_cpus_mask))
8398 if (!time_after(jiffies, READ_ONCE(rq->last_blocked_load_update_tick)))
8401 update_blocked_averages(cpu);
8403 return rq->has_blocked_load;
8410 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
8411 * @env: The load balancing environment.
8412 * @group: sched_group whose statistics are to be updated.
8413 * @sgs: variable to hold the statistics for this group.
8414 * @sg_status: Holds flag indicating the status of the sched_group
8416 static inline void update_sg_lb_stats(struct lb_env *env,
8417 struct sched_group *group,
8418 struct sg_lb_stats *sgs,
8421 int i, nr_running, local_group;
8423 memset(sgs, 0, sizeof(*sgs));
8425 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_span(group));
8427 for_each_cpu_and(i, sched_group_span(group), env->cpus) {
8428 struct rq *rq = cpu_rq(i);
8430 sgs->group_load += cpu_load(rq);
8431 sgs->group_util += cpu_util(i);
8432 sgs->group_runnable += cpu_runnable(rq);
8433 sgs->sum_h_nr_running += rq->cfs.h_nr_running;
8435 nr_running = rq->nr_running;
8436 sgs->sum_nr_running += nr_running;
8439 *sg_status |= SG_OVERLOAD;
8441 if (cpu_overutilized(i))
8442 *sg_status |= SG_OVERUTILIZED;
8444 #ifdef CONFIG_NUMA_BALANCING
8445 sgs->nr_numa_running += rq->nr_numa_running;
8446 sgs->nr_preferred_running += rq->nr_preferred_running;
8449 * No need to call idle_cpu() if nr_running is not 0
8451 if (!nr_running && idle_cpu(i)) {
8453 /* Idle cpu can't have misfit task */
8460 /* Check for a misfit task on the cpu */
8461 if (env->sd->flags & SD_ASYM_CPUCAPACITY &&
8462 sgs->group_misfit_task_load < rq->misfit_task_load) {
8463 sgs->group_misfit_task_load = rq->misfit_task_load;
8464 *sg_status |= SG_OVERLOAD;
8468 /* Check if dst CPU is idle and preferred to this group */
8469 if (env->sd->flags & SD_ASYM_PACKING &&
8470 env->idle != CPU_NOT_IDLE &&
8471 sgs->sum_h_nr_running &&
8472 sched_asym_prefer(env->dst_cpu, group->asym_prefer_cpu)) {
8473 sgs->group_asym_packing = 1;
8476 sgs->group_capacity = group->sgc->capacity;
8478 sgs->group_weight = group->group_weight;
8480 sgs->group_type = group_classify(env->sd->imbalance_pct, group, sgs);
8482 /* Computing avg_load makes sense only when group is overloaded */
8483 if (sgs->group_type == group_overloaded)
8484 sgs->avg_load = (sgs->group_load * SCHED_CAPACITY_SCALE) /
8485 sgs->group_capacity;
8489 * update_sd_pick_busiest - return 1 on busiest group
8490 * @env: The load balancing environment.
8491 * @sds: sched_domain statistics
8492 * @sg: sched_group candidate to be checked for being the busiest
8493 * @sgs: sched_group statistics
8495 * Determine if @sg is a busier group than the previously selected
8498 * Return: %true if @sg is a busier group than the previously selected
8499 * busiest group. %false otherwise.
8501 static bool update_sd_pick_busiest(struct lb_env *env,
8502 struct sd_lb_stats *sds,
8503 struct sched_group *sg,
8504 struct sg_lb_stats *sgs)
8506 struct sg_lb_stats *busiest = &sds->busiest_stat;
8508 /* Make sure that there is at least one task to pull */
8509 if (!sgs->sum_h_nr_running)
8513 * Don't try to pull misfit tasks we can't help.
8514 * We can use max_capacity here as reduction in capacity on some
8515 * CPUs in the group should either be possible to resolve
8516 * internally or be covered by avg_load imbalance (eventually).
8518 if (sgs->group_type == group_misfit_task &&
8519 (!group_smaller_max_cpu_capacity(sg, sds->local) ||
8520 sds->local_stat.group_type != group_has_spare))
8523 if (sgs->group_type > busiest->group_type)
8526 if (sgs->group_type < busiest->group_type)
8530 * The candidate and the current busiest group are the same type of
8531 * group. Let check which one is the busiest according to the type.
8534 switch (sgs->group_type) {
8535 case group_overloaded:
8536 /* Select the overloaded group with highest avg_load. */
8537 if (sgs->avg_load <= busiest->avg_load)
8541 case group_imbalanced:
8543 * Select the 1st imbalanced group as we don't have any way to
8544 * choose one more than another.
8548 case group_asym_packing:
8549 /* Prefer to move from lowest priority CPU's work */
8550 if (sched_asym_prefer(sg->asym_prefer_cpu, sds->busiest->asym_prefer_cpu))
8554 case group_misfit_task:
8556 * If we have more than one misfit sg go with the biggest
8559 if (sgs->group_misfit_task_load < busiest->group_misfit_task_load)
8563 case group_fully_busy:
8565 * Select the fully busy group with highest avg_load. In
8566 * theory, there is no need to pull task from such kind of
8567 * group because tasks have all compute capacity that they need
8568 * but we can still improve the overall throughput by reducing
8569 * contention when accessing shared HW resources.
8571 * XXX for now avg_load is not computed and always 0 so we
8572 * select the 1st one.
8574 if (sgs->avg_load <= busiest->avg_load)
8578 case group_has_spare:
8580 * Select not overloaded group with lowest number of idle cpus
8581 * and highest number of running tasks. We could also compare
8582 * the spare capacity which is more stable but it can end up
8583 * that the group has less spare capacity but finally more idle
8584 * CPUs which means less opportunity to pull tasks.
8586 if (sgs->idle_cpus > busiest->idle_cpus)
8588 else if ((sgs->idle_cpus == busiest->idle_cpus) &&
8589 (sgs->sum_nr_running <= busiest->sum_nr_running))
8596 * Candidate sg has no more than one task per CPU and has higher
8597 * per-CPU capacity. Migrating tasks to less capable CPUs may harm
8598 * throughput. Maximize throughput, power/energy consequences are not
8601 if ((env->sd->flags & SD_ASYM_CPUCAPACITY) &&
8602 (sgs->group_type <= group_fully_busy) &&
8603 (group_smaller_min_cpu_capacity(sds->local, sg)))
8609 #ifdef CONFIG_NUMA_BALANCING
8610 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
8612 if (sgs->sum_h_nr_running > sgs->nr_numa_running)
8614 if (sgs->sum_h_nr_running > sgs->nr_preferred_running)
8619 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
8621 if (rq->nr_running > rq->nr_numa_running)
8623 if (rq->nr_running > rq->nr_preferred_running)
8628 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
8633 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
8637 #endif /* CONFIG_NUMA_BALANCING */
8643 * task_running_on_cpu - return 1 if @p is running on @cpu.
8646 static unsigned int task_running_on_cpu(int cpu, struct task_struct *p)
8648 /* Task has no contribution or is new */
8649 if (cpu != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time))
8652 if (task_on_rq_queued(p))
8659 * idle_cpu_without - would a given CPU be idle without p ?
8660 * @cpu: the processor on which idleness is tested.
8661 * @p: task which should be ignored.
8663 * Return: 1 if the CPU would be idle. 0 otherwise.
8665 static int idle_cpu_without(int cpu, struct task_struct *p)
8667 struct rq *rq = cpu_rq(cpu);
8669 if (rq->curr != rq->idle && rq->curr != p)
8673 * rq->nr_running can't be used but an updated version without the
8674 * impact of p on cpu must be used instead. The updated nr_running
8675 * be computed and tested before calling idle_cpu_without().
8679 if (rq->ttwu_pending)
8687 * update_sg_wakeup_stats - Update sched_group's statistics for wakeup.
8688 * @sd: The sched_domain level to look for idlest group.
8689 * @group: sched_group whose statistics are to be updated.
8690 * @sgs: variable to hold the statistics for this group.
8691 * @p: The task for which we look for the idlest group/CPU.
8693 static inline void update_sg_wakeup_stats(struct sched_domain *sd,
8694 struct sched_group *group,
8695 struct sg_lb_stats *sgs,
8696 struct task_struct *p)
8700 memset(sgs, 0, sizeof(*sgs));
8702 for_each_cpu(i, sched_group_span(group)) {
8703 struct rq *rq = cpu_rq(i);
8706 sgs->group_load += cpu_load_without(rq, p);
8707 sgs->group_util += cpu_util_without(i, p);
8708 sgs->group_runnable += cpu_runnable_without(rq, p);
8709 local = task_running_on_cpu(i, p);
8710 sgs->sum_h_nr_running += rq->cfs.h_nr_running - local;
8712 nr_running = rq->nr_running - local;
8713 sgs->sum_nr_running += nr_running;
8716 * No need to call idle_cpu_without() if nr_running is not 0
8718 if (!nr_running && idle_cpu_without(i, p))
8723 /* Check if task fits in the group */
8724 if (sd->flags & SD_ASYM_CPUCAPACITY &&
8725 !task_fits_capacity(p, group->sgc->max_capacity)) {
8726 sgs->group_misfit_task_load = 1;
8729 sgs->group_capacity = group->sgc->capacity;
8731 sgs->group_weight = group->group_weight;
8733 sgs->group_type = group_classify(sd->imbalance_pct, group, sgs);
8736 * Computing avg_load makes sense only when group is fully busy or
8739 if (sgs->group_type == group_fully_busy ||
8740 sgs->group_type == group_overloaded)
8741 sgs->avg_load = (sgs->group_load * SCHED_CAPACITY_SCALE) /
8742 sgs->group_capacity;
8745 static bool update_pick_idlest(struct sched_group *idlest,
8746 struct sg_lb_stats *idlest_sgs,
8747 struct sched_group *group,
8748 struct sg_lb_stats *sgs)
8750 if (sgs->group_type < idlest_sgs->group_type)
8753 if (sgs->group_type > idlest_sgs->group_type)
8757 * The candidate and the current idlest group are the same type of
8758 * group. Let check which one is the idlest according to the type.
8761 switch (sgs->group_type) {
8762 case group_overloaded:
8763 case group_fully_busy:
8764 /* Select the group with lowest avg_load. */
8765 if (idlest_sgs->avg_load <= sgs->avg_load)
8769 case group_imbalanced:
8770 case group_asym_packing:
8771 /* Those types are not used in the slow wakeup path */
8774 case group_misfit_task:
8775 /* Select group with the highest max capacity */
8776 if (idlest->sgc->max_capacity >= group->sgc->max_capacity)
8780 case group_has_spare:
8781 /* Select group with most idle CPUs */
8782 if (idlest_sgs->idle_cpus > sgs->idle_cpus)
8785 /* Select group with lowest group_util */
8786 if (idlest_sgs->idle_cpus == sgs->idle_cpus &&
8787 idlest_sgs->group_util <= sgs->group_util)
8797 * Allow a NUMA imbalance if busy CPUs is less than 25% of the domain.
8798 * This is an approximation as the number of running tasks may not be
8799 * related to the number of busy CPUs due to sched_setaffinity.
8801 static inline bool allow_numa_imbalance(int dst_running, int dst_weight)
8803 return (dst_running < (dst_weight >> 2));
8807 * find_idlest_group() finds and returns the least busy CPU group within the
8810 * Assumes p is allowed on at least one CPU in sd.
8812 static struct sched_group *
8813 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
8815 struct sched_group *idlest = NULL, *local = NULL, *group = sd->groups;
8816 struct sg_lb_stats local_sgs, tmp_sgs;
8817 struct sg_lb_stats *sgs;
8818 unsigned long imbalance;
8819 struct sg_lb_stats idlest_sgs = {
8820 .avg_load = UINT_MAX,
8821 .group_type = group_overloaded,
8827 /* Skip over this group if it has no CPUs allowed */
8828 if (!cpumask_intersects(sched_group_span(group),
8832 local_group = cpumask_test_cpu(this_cpu,
8833 sched_group_span(group));
8842 update_sg_wakeup_stats(sd, group, sgs, p);
8844 if (!local_group && update_pick_idlest(idlest, &idlest_sgs, group, sgs)) {
8849 } while (group = group->next, group != sd->groups);
8852 /* There is no idlest group to push tasks to */
8856 /* The local group has been skipped because of CPU affinity */
8861 * If the local group is idler than the selected idlest group
8862 * don't try and push the task.
8864 if (local_sgs.group_type < idlest_sgs.group_type)
8868 * If the local group is busier than the selected idlest group
8869 * try and push the task.
8871 if (local_sgs.group_type > idlest_sgs.group_type)
8874 switch (local_sgs.group_type) {
8875 case group_overloaded:
8876 case group_fully_busy:
8878 /* Calculate allowed imbalance based on load */
8879 imbalance = scale_load_down(NICE_0_LOAD) *
8880 (sd->imbalance_pct-100) / 100;
8883 * When comparing groups across NUMA domains, it's possible for
8884 * the local domain to be very lightly loaded relative to the
8885 * remote domains but "imbalance" skews the comparison making
8886 * remote CPUs look much more favourable. When considering
8887 * cross-domain, add imbalance to the load on the remote node
8888 * and consider staying local.
8891 if ((sd->flags & SD_NUMA) &&
8892 ((idlest_sgs.avg_load + imbalance) >= local_sgs.avg_load))
8896 * If the local group is less loaded than the selected
8897 * idlest group don't try and push any tasks.
8899 if (idlest_sgs.avg_load >= (local_sgs.avg_load + imbalance))
8902 if (100 * local_sgs.avg_load <= sd->imbalance_pct * idlest_sgs.avg_load)
8906 case group_imbalanced:
8907 case group_asym_packing:
8908 /* Those type are not used in the slow wakeup path */
8911 case group_misfit_task:
8912 /* Select group with the highest max capacity */
8913 if (local->sgc->max_capacity >= idlest->sgc->max_capacity)
8917 case group_has_spare:
8918 if (sd->flags & SD_NUMA) {
8919 #ifdef CONFIG_NUMA_BALANCING
8922 * If there is spare capacity at NUMA, try to select
8923 * the preferred node
8925 if (cpu_to_node(this_cpu) == p->numa_preferred_nid)
8928 idlest_cpu = cpumask_first(sched_group_span(idlest));
8929 if (cpu_to_node(idlest_cpu) == p->numa_preferred_nid)
8933 * Otherwise, keep the task on this node to stay close
8934 * its wakeup source and improve locality. If there is
8935 * a real need of migration, periodic load balance will
8938 if (allow_numa_imbalance(local_sgs.sum_nr_running, sd->span_weight))
8943 * Select group with highest number of idle CPUs. We could also
8944 * compare the utilization which is more stable but it can end
8945 * up that the group has less spare capacity but finally more
8946 * idle CPUs which means more opportunity to run task.
8948 if (local_sgs.idle_cpus >= idlest_sgs.idle_cpus)
8957 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
8958 * @env: The load balancing environment.
8959 * @sds: variable to hold the statistics for this sched_domain.
8962 static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
8964 struct sched_domain *child = env->sd->child;
8965 struct sched_group *sg = env->sd->groups;
8966 struct sg_lb_stats *local = &sds->local_stat;
8967 struct sg_lb_stats tmp_sgs;
8971 struct sg_lb_stats *sgs = &tmp_sgs;
8974 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_span(sg));
8979 if (env->idle != CPU_NEWLY_IDLE ||
8980 time_after_eq(jiffies, sg->sgc->next_update))
8981 update_group_capacity(env->sd, env->dst_cpu);
8984 update_sg_lb_stats(env, sg, sgs, &sg_status);
8990 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
8992 sds->busiest_stat = *sgs;
8996 /* Now, start updating sd_lb_stats */
8997 sds->total_load += sgs->group_load;
8998 sds->total_capacity += sgs->group_capacity;
9001 } while (sg != env->sd->groups);
9003 /* Tag domain that child domain prefers tasks go to siblings first */
9004 sds->prefer_sibling = child && child->flags & SD_PREFER_SIBLING;
9007 if (env->sd->flags & SD_NUMA)
9008 env->fbq_type = fbq_classify_group(&sds->busiest_stat);
9010 if (!env->sd->parent) {
9011 struct root_domain *rd = env->dst_rq->rd;
9013 /* update overload indicator if we are at root domain */
9014 WRITE_ONCE(rd->overload, sg_status & SG_OVERLOAD);
9016 /* Update over-utilization (tipping point, U >= 0) indicator */
9017 WRITE_ONCE(rd->overutilized, sg_status & SG_OVERUTILIZED);
9018 trace_sched_overutilized_tp(rd, sg_status & SG_OVERUTILIZED);
9019 } else if (sg_status & SG_OVERUTILIZED) {
9020 struct root_domain *rd = env->dst_rq->rd;
9022 WRITE_ONCE(rd->overutilized, SG_OVERUTILIZED);
9023 trace_sched_overutilized_tp(rd, SG_OVERUTILIZED);
9027 #define NUMA_IMBALANCE_MIN 2
9029 static inline long adjust_numa_imbalance(int imbalance,
9030 int dst_running, int dst_weight)
9032 if (!allow_numa_imbalance(dst_running, dst_weight))
9036 * Allow a small imbalance based on a simple pair of communicating
9037 * tasks that remain local when the destination is lightly loaded.
9039 if (imbalance <= NUMA_IMBALANCE_MIN)
9046 * calculate_imbalance - Calculate the amount of imbalance present within the
9047 * groups of a given sched_domain during load balance.
9048 * @env: load balance environment
9049 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
9051 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
9053 struct sg_lb_stats *local, *busiest;
9055 local = &sds->local_stat;
9056 busiest = &sds->busiest_stat;
9058 if (busiest->group_type == group_misfit_task) {
9059 /* Set imbalance to allow misfit tasks to be balanced. */
9060 env->migration_type = migrate_misfit;
9065 if (busiest->group_type == group_asym_packing) {
9067 * In case of asym capacity, we will try to migrate all load to
9068 * the preferred CPU.
9070 env->migration_type = migrate_task;
9071 env->imbalance = busiest->sum_h_nr_running;
9075 if (busiest->group_type == group_imbalanced) {
9077 * In the group_imb case we cannot rely on group-wide averages
9078 * to ensure CPU-load equilibrium, try to move any task to fix
9079 * the imbalance. The next load balance will take care of
9080 * balancing back the system.
9082 env->migration_type = migrate_task;
9088 * Try to use spare capacity of local group without overloading it or
9091 if (local->group_type == group_has_spare) {
9092 if ((busiest->group_type > group_fully_busy) &&
9093 !(env->sd->flags & SD_SHARE_PKG_RESOURCES)) {
9095 * If busiest is overloaded, try to fill spare
9096 * capacity. This might end up creating spare capacity
9097 * in busiest or busiest still being overloaded but
9098 * there is no simple way to directly compute the
9099 * amount of load to migrate in order to balance the
9102 env->migration_type = migrate_util;
9103 env->imbalance = max(local->group_capacity, local->group_util) -
9107 * In some cases, the group's utilization is max or even
9108 * higher than capacity because of migrations but the
9109 * local CPU is (newly) idle. There is at least one
9110 * waiting task in this overloaded busiest group. Let's
9113 if (env->idle != CPU_NOT_IDLE && env->imbalance == 0) {
9114 env->migration_type = migrate_task;
9121 if (busiest->group_weight == 1 || sds->prefer_sibling) {
9122 unsigned int nr_diff = busiest->sum_nr_running;
9124 * When prefer sibling, evenly spread running tasks on
9127 env->migration_type = migrate_task;
9128 lsub_positive(&nr_diff, local->sum_nr_running);
9129 env->imbalance = nr_diff >> 1;
9133 * If there is no overload, we just want to even the number of
9136 env->migration_type = migrate_task;
9137 env->imbalance = max_t(long, 0, (local->idle_cpus -
9138 busiest->idle_cpus) >> 1);
9141 /* Consider allowing a small imbalance between NUMA groups */
9142 if (env->sd->flags & SD_NUMA) {
9143 env->imbalance = adjust_numa_imbalance(env->imbalance,
9144 busiest->sum_nr_running, busiest->group_weight);
9151 * Local is fully busy but has to take more load to relieve the
9154 if (local->group_type < group_overloaded) {
9156 * Local will become overloaded so the avg_load metrics are
9160 local->avg_load = (local->group_load * SCHED_CAPACITY_SCALE) /
9161 local->group_capacity;
9163 sds->avg_load = (sds->total_load * SCHED_CAPACITY_SCALE) /
9164 sds->total_capacity;
9166 * If the local group is more loaded than the selected
9167 * busiest group don't try to pull any tasks.
9169 if (local->avg_load >= busiest->avg_load) {
9176 * Both group are or will become overloaded and we're trying to get all
9177 * the CPUs to the average_load, so we don't want to push ourselves
9178 * above the average load, nor do we wish to reduce the max loaded CPU
9179 * below the average load. At the same time, we also don't want to
9180 * reduce the group load below the group capacity. Thus we look for
9181 * the minimum possible imbalance.
9183 env->migration_type = migrate_load;
9184 env->imbalance = min(
9185 (busiest->avg_load - sds->avg_load) * busiest->group_capacity,
9186 (sds->avg_load - local->avg_load) * local->group_capacity
9187 ) / SCHED_CAPACITY_SCALE;
9190 /******* find_busiest_group() helpers end here *********************/
9193 * Decision matrix according to the local and busiest group type:
9195 * busiest \ local has_spare fully_busy misfit asym imbalanced overloaded
9196 * has_spare nr_idle balanced N/A N/A balanced balanced
9197 * fully_busy nr_idle nr_idle N/A N/A balanced balanced
9198 * misfit_task force N/A N/A N/A force force
9199 * asym_packing force force N/A N/A force force
9200 * imbalanced force force N/A N/A force force
9201 * overloaded force force N/A N/A force avg_load
9203 * N/A : Not Applicable because already filtered while updating
9205 * balanced : The system is balanced for these 2 groups.
9206 * force : Calculate the imbalance as load migration is probably needed.
9207 * avg_load : Only if imbalance is significant enough.
9208 * nr_idle : dst_cpu is not busy and the number of idle CPUs is quite
9209 * different in groups.
9213 * find_busiest_group - Returns the busiest group within the sched_domain
9214 * if there is an imbalance.
9216 * Also calculates the amount of runnable load which should be moved
9217 * to restore balance.
9219 * @env: The load balancing environment.
9221 * Return: - The busiest group if imbalance exists.
9223 static struct sched_group *find_busiest_group(struct lb_env *env)
9225 struct sg_lb_stats *local, *busiest;
9226 struct sd_lb_stats sds;
9228 init_sd_lb_stats(&sds);
9231 * Compute the various statistics relevant for load balancing at
9234 update_sd_lb_stats(env, &sds);
9236 if (sched_energy_enabled()) {
9237 struct root_domain *rd = env->dst_rq->rd;
9239 if (rcu_dereference(rd->pd) && !READ_ONCE(rd->overutilized))
9243 local = &sds.local_stat;
9244 busiest = &sds.busiest_stat;
9246 /* There is no busy sibling group to pull tasks from */
9250 /* Misfit tasks should be dealt with regardless of the avg load */
9251 if (busiest->group_type == group_misfit_task)
9254 /* ASYM feature bypasses nice load balance check */
9255 if (busiest->group_type == group_asym_packing)
9259 * If the busiest group is imbalanced the below checks don't
9260 * work because they assume all things are equal, which typically
9261 * isn't true due to cpus_ptr constraints and the like.
9263 if (busiest->group_type == group_imbalanced)
9267 * If the local group is busier than the selected busiest group
9268 * don't try and pull any tasks.
9270 if (local->group_type > busiest->group_type)
9274 * When groups are overloaded, use the avg_load to ensure fairness
9277 if (local->group_type == group_overloaded) {
9279 * If the local group is more loaded than the selected
9280 * busiest group don't try to pull any tasks.
9282 if (local->avg_load >= busiest->avg_load)
9285 /* XXX broken for overlapping NUMA groups */
9286 sds.avg_load = (sds.total_load * SCHED_CAPACITY_SCALE) /
9290 * Don't pull any tasks if this group is already above the
9291 * domain average load.
9293 if (local->avg_load >= sds.avg_load)
9297 * If the busiest group is more loaded, use imbalance_pct to be
9300 if (100 * busiest->avg_load <=
9301 env->sd->imbalance_pct * local->avg_load)
9305 /* Try to move all excess tasks to child's sibling domain */
9306 if (sds.prefer_sibling && local->group_type == group_has_spare &&
9307 busiest->sum_nr_running > local->sum_nr_running + 1)
9310 if (busiest->group_type != group_overloaded) {
9311 if (env->idle == CPU_NOT_IDLE)
9313 * If the busiest group is not overloaded (and as a
9314 * result the local one too) but this CPU is already
9315 * busy, let another idle CPU try to pull task.
9319 if (busiest->group_weight > 1 &&
9320 local->idle_cpus <= (busiest->idle_cpus + 1))
9322 * If the busiest group is not overloaded
9323 * and there is no imbalance between this and busiest
9324 * group wrt idle CPUs, it is balanced. The imbalance
9325 * becomes significant if the diff is greater than 1
9326 * otherwise we might end up to just move the imbalance
9327 * on another group. Of course this applies only if
9328 * there is more than 1 CPU per group.
9332 if (busiest->sum_h_nr_running == 1)
9334 * busiest doesn't have any tasks waiting to run
9340 /* Looks like there is an imbalance. Compute it */
9341 calculate_imbalance(env, &sds);
9342 return env->imbalance ? sds.busiest : NULL;
9350 * find_busiest_queue - find the busiest runqueue among the CPUs in the group.
9352 static struct rq *find_busiest_queue(struct lb_env *env,
9353 struct sched_group *group)
9355 struct rq *busiest = NULL, *rq;
9356 unsigned long busiest_util = 0, busiest_load = 0, busiest_capacity = 1;
9357 unsigned int busiest_nr = 0;
9360 for_each_cpu_and(i, sched_group_span(group), env->cpus) {
9361 unsigned long capacity, load, util;
9362 unsigned int nr_running;
9366 rt = fbq_classify_rq(rq);
9369 * We classify groups/runqueues into three groups:
9370 * - regular: there are !numa tasks
9371 * - remote: there are numa tasks that run on the 'wrong' node
9372 * - all: there is no distinction
9374 * In order to avoid migrating ideally placed numa tasks,
9375 * ignore those when there's better options.
9377 * If we ignore the actual busiest queue to migrate another
9378 * task, the next balance pass can still reduce the busiest
9379 * queue by moving tasks around inside the node.
9381 * If we cannot move enough load due to this classification
9382 * the next pass will adjust the group classification and
9383 * allow migration of more tasks.
9385 * Both cases only affect the total convergence complexity.
9387 if (rt > env->fbq_type)
9390 nr_running = rq->cfs.h_nr_running;
9394 capacity = capacity_of(i);
9397 * For ASYM_CPUCAPACITY domains, don't pick a CPU that could
9398 * eventually lead to active_balancing high->low capacity.
9399 * Higher per-CPU capacity is considered better than balancing
9402 if (env->sd->flags & SD_ASYM_CPUCAPACITY &&
9403 capacity_of(env->dst_cpu) < capacity &&
9407 switch (env->migration_type) {
9410 * When comparing with load imbalance, use cpu_load()
9411 * which is not scaled with the CPU capacity.
9413 load = cpu_load(rq);
9415 if (nr_running == 1 && load > env->imbalance &&
9416 !check_cpu_capacity(rq, env->sd))
9420 * For the load comparisons with the other CPUs,
9421 * consider the cpu_load() scaled with the CPU
9422 * capacity, so that the load can be moved away
9423 * from the CPU that is potentially running at a
9426 * Thus we're looking for max(load_i / capacity_i),
9427 * crosswise multiplication to rid ourselves of the
9428 * division works out to:
9429 * load_i * capacity_j > load_j * capacity_i;
9430 * where j is our previous maximum.
9432 if (load * busiest_capacity > busiest_load * capacity) {
9433 busiest_load = load;
9434 busiest_capacity = capacity;
9440 util = cpu_util(cpu_of(rq));
9443 * Don't try to pull utilization from a CPU with one
9444 * running task. Whatever its utilization, we will fail
9447 if (nr_running <= 1)
9450 if (busiest_util < util) {
9451 busiest_util = util;
9457 if (busiest_nr < nr_running) {
9458 busiest_nr = nr_running;
9463 case migrate_misfit:
9465 * For ASYM_CPUCAPACITY domains with misfit tasks we
9466 * simply seek the "biggest" misfit task.
9468 if (rq->misfit_task_load > busiest_load) {
9469 busiest_load = rq->misfit_task_load;
9482 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
9483 * so long as it is large enough.
9485 #define MAX_PINNED_INTERVAL 512
9488 asym_active_balance(struct lb_env *env)
9491 * ASYM_PACKING needs to force migrate tasks from busy but
9492 * lower priority CPUs in order to pack all tasks in the
9493 * highest priority CPUs.
9495 return env->idle != CPU_NOT_IDLE && (env->sd->flags & SD_ASYM_PACKING) &&
9496 sched_asym_prefer(env->dst_cpu, env->src_cpu);
9500 imbalanced_active_balance(struct lb_env *env)
9502 struct sched_domain *sd = env->sd;
9505 * The imbalanced case includes the case of pinned tasks preventing a fair
9506 * distribution of the load on the system but also the even distribution of the
9507 * threads on a system with spare capacity
9509 if ((env->migration_type == migrate_task) &&
9510 (sd->nr_balance_failed > sd->cache_nice_tries+2))
9516 static int need_active_balance(struct lb_env *env)
9518 struct sched_domain *sd = env->sd;
9520 if (asym_active_balance(env))
9523 if (imbalanced_active_balance(env))
9527 * The dst_cpu is idle and the src_cpu CPU has only 1 CFS task.
9528 * It's worth migrating the task if the src_cpu's capacity is reduced
9529 * because of other sched_class or IRQs if more capacity stays
9530 * available on dst_cpu.
9532 if ((env->idle != CPU_NOT_IDLE) &&
9533 (env->src_rq->cfs.h_nr_running == 1)) {
9534 if ((check_cpu_capacity(env->src_rq, sd)) &&
9535 (capacity_of(env->src_cpu)*sd->imbalance_pct < capacity_of(env->dst_cpu)*100))
9539 if (env->migration_type == migrate_misfit)
9545 static int active_load_balance_cpu_stop(void *data);
9547 static int should_we_balance(struct lb_env *env)
9549 struct sched_group *sg = env->sd->groups;
9553 * Ensure the balancing environment is consistent; can happen
9554 * when the softirq triggers 'during' hotplug.
9556 if (!cpumask_test_cpu(env->dst_cpu, env->cpus))
9560 * In the newly idle case, we will allow all the CPUs
9561 * to do the newly idle load balance.
9563 if (env->idle == CPU_NEWLY_IDLE)
9566 /* Try to find first idle CPU */
9567 for_each_cpu_and(cpu, group_balance_mask(sg), env->cpus) {
9571 /* Are we the first idle CPU? */
9572 return cpu == env->dst_cpu;
9575 /* Are we the first CPU of this group ? */
9576 return group_balance_cpu(sg) == env->dst_cpu;
9580 * Check this_cpu to ensure it is balanced within domain. Attempt to move
9581 * tasks if there is an imbalance.
9583 static int load_balance(int this_cpu, struct rq *this_rq,
9584 struct sched_domain *sd, enum cpu_idle_type idle,
9585 int *continue_balancing)
9587 int ld_moved, cur_ld_moved, active_balance = 0;
9588 struct sched_domain *sd_parent = sd->parent;
9589 struct sched_group *group;
9592 struct cpumask *cpus = this_cpu_cpumask_var_ptr(load_balance_mask);
9594 struct lb_env env = {
9596 .dst_cpu = this_cpu,
9598 .dst_grpmask = sched_group_span(sd->groups),
9600 .loop_break = sched_nr_migrate_break,
9603 .tasks = LIST_HEAD_INIT(env.tasks),
9606 cpumask_and(cpus, sched_domain_span(sd), cpu_active_mask);
9608 schedstat_inc(sd->lb_count[idle]);
9611 if (!should_we_balance(&env)) {
9612 *continue_balancing = 0;
9616 group = find_busiest_group(&env);
9618 schedstat_inc(sd->lb_nobusyg[idle]);
9622 busiest = find_busiest_queue(&env, group);
9624 schedstat_inc(sd->lb_nobusyq[idle]);
9628 BUG_ON(busiest == env.dst_rq);
9630 schedstat_add(sd->lb_imbalance[idle], env.imbalance);
9632 env.src_cpu = busiest->cpu;
9633 env.src_rq = busiest;
9636 /* Clear this flag as soon as we find a pullable task */
9637 env.flags |= LBF_ALL_PINNED;
9638 if (busiest->nr_running > 1) {
9640 * Attempt to move tasks. If find_busiest_group has found
9641 * an imbalance but busiest->nr_running <= 1, the group is
9642 * still unbalanced. ld_moved simply stays zero, so it is
9643 * correctly treated as an imbalance.
9645 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
9648 rq_lock_irqsave(busiest, &rf);
9649 update_rq_clock(busiest);
9652 * cur_ld_moved - load moved in current iteration
9653 * ld_moved - cumulative load moved across iterations
9655 cur_ld_moved = detach_tasks(&env);
9658 * We've detached some tasks from busiest_rq. Every
9659 * task is masked "TASK_ON_RQ_MIGRATING", so we can safely
9660 * unlock busiest->lock, and we are able to be sure
9661 * that nobody can manipulate the tasks in parallel.
9662 * See task_rq_lock() family for the details.
9665 rq_unlock(busiest, &rf);
9669 ld_moved += cur_ld_moved;
9672 local_irq_restore(rf.flags);
9674 if (env.flags & LBF_NEED_BREAK) {
9675 env.flags &= ~LBF_NEED_BREAK;
9680 * Revisit (affine) tasks on src_cpu that couldn't be moved to
9681 * us and move them to an alternate dst_cpu in our sched_group
9682 * where they can run. The upper limit on how many times we
9683 * iterate on same src_cpu is dependent on number of CPUs in our
9686 * This changes load balance semantics a bit on who can move
9687 * load to a given_cpu. In addition to the given_cpu itself
9688 * (or a ilb_cpu acting on its behalf where given_cpu is
9689 * nohz-idle), we now have balance_cpu in a position to move
9690 * load to given_cpu. In rare situations, this may cause
9691 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
9692 * _independently_ and at _same_ time to move some load to
9693 * given_cpu) causing exceess load to be moved to given_cpu.
9694 * This however should not happen so much in practice and
9695 * moreover subsequent load balance cycles should correct the
9696 * excess load moved.
9698 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
9700 /* Prevent to re-select dst_cpu via env's CPUs */
9701 __cpumask_clear_cpu(env.dst_cpu, env.cpus);
9703 env.dst_rq = cpu_rq(env.new_dst_cpu);
9704 env.dst_cpu = env.new_dst_cpu;
9705 env.flags &= ~LBF_DST_PINNED;
9707 env.loop_break = sched_nr_migrate_break;
9710 * Go back to "more_balance" rather than "redo" since we
9711 * need to continue with same src_cpu.
9717 * We failed to reach balance because of affinity.
9720 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
9722 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0)
9723 *group_imbalance = 1;
9726 /* All tasks on this runqueue were pinned by CPU affinity */
9727 if (unlikely(env.flags & LBF_ALL_PINNED)) {
9728 __cpumask_clear_cpu(cpu_of(busiest), cpus);
9730 * Attempting to continue load balancing at the current
9731 * sched_domain level only makes sense if there are
9732 * active CPUs remaining as possible busiest CPUs to
9733 * pull load from which are not contained within the
9734 * destination group that is receiving any migrated
9737 if (!cpumask_subset(cpus, env.dst_grpmask)) {
9739 env.loop_break = sched_nr_migrate_break;
9742 goto out_all_pinned;
9747 schedstat_inc(sd->lb_failed[idle]);
9749 * Increment the failure counter only on periodic balance.
9750 * We do not want newidle balance, which can be very
9751 * frequent, pollute the failure counter causing
9752 * excessive cache_hot migrations and active balances.
9754 if (idle != CPU_NEWLY_IDLE)
9755 sd->nr_balance_failed++;
9757 if (need_active_balance(&env)) {
9758 unsigned long flags;
9760 raw_spin_lock_irqsave(&busiest->lock, flags);
9763 * Don't kick the active_load_balance_cpu_stop,
9764 * if the curr task on busiest CPU can't be
9765 * moved to this_cpu:
9767 if (!cpumask_test_cpu(this_cpu, busiest->curr->cpus_ptr)) {
9768 raw_spin_unlock_irqrestore(&busiest->lock,
9770 goto out_one_pinned;
9773 /* Record that we found at least one task that could run on this_cpu */
9774 env.flags &= ~LBF_ALL_PINNED;
9777 * ->active_balance synchronizes accesses to
9778 * ->active_balance_work. Once set, it's cleared
9779 * only after active load balance is finished.
9781 if (!busiest->active_balance) {
9782 busiest->active_balance = 1;
9783 busiest->push_cpu = this_cpu;
9786 raw_spin_unlock_irqrestore(&busiest->lock, flags);
9788 if (active_balance) {
9789 stop_one_cpu_nowait(cpu_of(busiest),
9790 active_load_balance_cpu_stop, busiest,
9791 &busiest->active_balance_work);
9794 /* We've kicked active balancing, force task migration. */
9795 sd->nr_balance_failed = sd->cache_nice_tries+1;
9798 sd->nr_balance_failed = 0;
9801 if (likely(!active_balance) || need_active_balance(&env)) {
9802 /* We were unbalanced, so reset the balancing interval */
9803 sd->balance_interval = sd->min_interval;
9810 * We reach balance although we may have faced some affinity
9811 * constraints. Clear the imbalance flag only if other tasks got
9812 * a chance to move and fix the imbalance.
9814 if (sd_parent && !(env.flags & LBF_ALL_PINNED)) {
9815 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
9817 if (*group_imbalance)
9818 *group_imbalance = 0;
9823 * We reach balance because all tasks are pinned at this level so
9824 * we can't migrate them. Let the imbalance flag set so parent level
9825 * can try to migrate them.
9827 schedstat_inc(sd->lb_balanced[idle]);
9829 sd->nr_balance_failed = 0;
9835 * newidle_balance() disregards balance intervals, so we could
9836 * repeatedly reach this code, which would lead to balance_interval
9837 * skyrocketting in a short amount of time. Skip the balance_interval
9838 * increase logic to avoid that.
9840 if (env.idle == CPU_NEWLY_IDLE)
9843 /* tune up the balancing interval */
9844 if ((env.flags & LBF_ALL_PINNED &&
9845 sd->balance_interval < MAX_PINNED_INTERVAL) ||
9846 sd->balance_interval < sd->max_interval)
9847 sd->balance_interval *= 2;
9852 static inline unsigned long
9853 get_sd_balance_interval(struct sched_domain *sd, int cpu_busy)
9855 unsigned long interval = sd->balance_interval;
9858 interval *= sd->busy_factor;
9860 /* scale ms to jiffies */
9861 interval = msecs_to_jiffies(interval);
9864 * Reduce likelihood of busy balancing at higher domains racing with
9865 * balancing at lower domains by preventing their balancing periods
9866 * from being multiples of each other.
9871 interval = clamp(interval, 1UL, max_load_balance_interval);
9877 update_next_balance(struct sched_domain *sd, unsigned long *next_balance)
9879 unsigned long interval, next;
9881 /* used by idle balance, so cpu_busy = 0 */
9882 interval = get_sd_balance_interval(sd, 0);
9883 next = sd->last_balance + interval;
9885 if (time_after(*next_balance, next))
9886 *next_balance = next;
9890 * active_load_balance_cpu_stop is run by the CPU stopper. It pushes
9891 * running tasks off the busiest CPU onto idle CPUs. It requires at
9892 * least 1 task to be running on each physical CPU where possible, and
9893 * avoids physical / logical imbalances.
9895 static int active_load_balance_cpu_stop(void *data)
9897 struct rq *busiest_rq = data;
9898 int busiest_cpu = cpu_of(busiest_rq);
9899 int target_cpu = busiest_rq->push_cpu;
9900 struct rq *target_rq = cpu_rq(target_cpu);
9901 struct sched_domain *sd;
9902 struct task_struct *p = NULL;
9905 rq_lock_irq(busiest_rq, &rf);
9907 * Between queueing the stop-work and running it is a hole in which
9908 * CPUs can become inactive. We should not move tasks from or to
9911 if (!cpu_active(busiest_cpu) || !cpu_active(target_cpu))
9914 /* Make sure the requested CPU hasn't gone down in the meantime: */
9915 if (unlikely(busiest_cpu != smp_processor_id() ||
9916 !busiest_rq->active_balance))
9919 /* Is there any task to move? */
9920 if (busiest_rq->nr_running <= 1)
9924 * This condition is "impossible", if it occurs
9925 * we need to fix it. Originally reported by
9926 * Bjorn Helgaas on a 128-CPU setup.
9928 BUG_ON(busiest_rq == target_rq);
9930 /* Search for an sd spanning us and the target CPU. */
9932 for_each_domain(target_cpu, sd) {
9933 if (cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
9938 struct lb_env env = {
9940 .dst_cpu = target_cpu,
9941 .dst_rq = target_rq,
9942 .src_cpu = busiest_rq->cpu,
9943 .src_rq = busiest_rq,
9946 * can_migrate_task() doesn't need to compute new_dst_cpu
9947 * for active balancing. Since we have CPU_IDLE, but no
9948 * @dst_grpmask we need to make that test go away with lying
9951 .flags = LBF_DST_PINNED,
9954 schedstat_inc(sd->alb_count);
9955 update_rq_clock(busiest_rq);
9957 p = detach_one_task(&env);
9959 schedstat_inc(sd->alb_pushed);
9960 /* Active balancing done, reset the failure counter. */
9961 sd->nr_balance_failed = 0;
9963 schedstat_inc(sd->alb_failed);
9968 busiest_rq->active_balance = 0;
9969 rq_unlock(busiest_rq, &rf);
9972 attach_one_task(target_rq, p);
9979 static DEFINE_SPINLOCK(balancing);
9982 * Scale the max load_balance interval with the number of CPUs in the system.
9983 * This trades load-balance latency on larger machines for less cross talk.
9985 void update_max_interval(void)
9987 max_load_balance_interval = HZ*num_online_cpus()/10;
9991 * It checks each scheduling domain to see if it is due to be balanced,
9992 * and initiates a balancing operation if so.
9994 * Balancing parameters are set up in init_sched_domains.
9996 static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle)
9998 int continue_balancing = 1;
10000 int busy = idle != CPU_IDLE && !sched_idle_cpu(cpu);
10001 unsigned long interval;
10002 struct sched_domain *sd;
10003 /* Earliest time when we have to do rebalance again */
10004 unsigned long next_balance = jiffies + 60*HZ;
10005 int update_next_balance = 0;
10006 int need_serialize, need_decay = 0;
10010 for_each_domain(cpu, sd) {
10012 * Decay the newidle max times here because this is a regular
10013 * visit to all the domains. Decay ~1% per second.
10015 if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
10016 sd->max_newidle_lb_cost =
10017 (sd->max_newidle_lb_cost * 253) / 256;
10018 sd->next_decay_max_lb_cost = jiffies + HZ;
10021 max_cost += sd->max_newidle_lb_cost;
10024 * Stop the load balance at this level. There is another
10025 * CPU in our sched group which is doing load balancing more
10028 if (!continue_balancing) {
10034 interval = get_sd_balance_interval(sd, busy);
10036 need_serialize = sd->flags & SD_SERIALIZE;
10037 if (need_serialize) {
10038 if (!spin_trylock(&balancing))
10042 if (time_after_eq(jiffies, sd->last_balance + interval)) {
10043 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
10045 * The LBF_DST_PINNED logic could have changed
10046 * env->dst_cpu, so we can't know our idle
10047 * state even if we migrated tasks. Update it.
10049 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
10050 busy = idle != CPU_IDLE && !sched_idle_cpu(cpu);
10052 sd->last_balance = jiffies;
10053 interval = get_sd_balance_interval(sd, busy);
10055 if (need_serialize)
10056 spin_unlock(&balancing);
10058 if (time_after(next_balance, sd->last_balance + interval)) {
10059 next_balance = sd->last_balance + interval;
10060 update_next_balance = 1;
10065 * Ensure the rq-wide value also decays but keep it at a
10066 * reasonable floor to avoid funnies with rq->avg_idle.
10068 rq->max_idle_balance_cost =
10069 max((u64)sysctl_sched_migration_cost, max_cost);
10074 * next_balance will be updated only when there is a need.
10075 * When the cpu is attached to null domain for ex, it will not be
10078 if (likely(update_next_balance))
10079 rq->next_balance = next_balance;
10083 static inline int on_null_domain(struct rq *rq)
10085 return unlikely(!rcu_dereference_sched(rq->sd));
10088 #ifdef CONFIG_NO_HZ_COMMON
10090 * idle load balancing details
10091 * - When one of the busy CPUs notice that there may be an idle rebalancing
10092 * needed, they will kick the idle load balancer, which then does idle
10093 * load balancing for all the idle CPUs.
10094 * - HK_FLAG_MISC CPUs are used for this task, because HK_FLAG_SCHED not set
10098 static inline int find_new_ilb(void)
10102 for_each_cpu_and(ilb, nohz.idle_cpus_mask,
10103 housekeeping_cpumask(HK_FLAG_MISC)) {
10105 if (ilb == smp_processor_id())
10116 * Kick a CPU to do the nohz balancing, if it is time for it. We pick any
10117 * idle CPU in the HK_FLAG_MISC housekeeping set (if there is one).
10119 static void kick_ilb(unsigned int flags)
10124 * Increase nohz.next_balance only when if full ilb is triggered but
10125 * not if we only update stats.
10127 if (flags & NOHZ_BALANCE_KICK)
10128 nohz.next_balance = jiffies+1;
10130 ilb_cpu = find_new_ilb();
10132 if (ilb_cpu >= nr_cpu_ids)
10136 * Access to rq::nohz_csd is serialized by NOHZ_KICK_MASK; he who sets
10137 * the first flag owns it; cleared by nohz_csd_func().
10139 flags = atomic_fetch_or(flags, nohz_flags(ilb_cpu));
10140 if (flags & NOHZ_KICK_MASK)
10144 * This way we generate an IPI on the target CPU which
10145 * is idle. And the softirq performing nohz idle load balance
10146 * will be run before returning from the IPI.
10148 smp_call_function_single_async(ilb_cpu, &cpu_rq(ilb_cpu)->nohz_csd);
10152 * Current decision point for kicking the idle load balancer in the presence
10153 * of idle CPUs in the system.
10155 static void nohz_balancer_kick(struct rq *rq)
10157 unsigned long now = jiffies;
10158 struct sched_domain_shared *sds;
10159 struct sched_domain *sd;
10160 int nr_busy, i, cpu = rq->cpu;
10161 unsigned int flags = 0;
10163 if (unlikely(rq->idle_balance))
10167 * We may be recently in ticked or tickless idle mode. At the first
10168 * busy tick after returning from idle, we will update the busy stats.
10170 nohz_balance_exit_idle(rq);
10173 * None are in tickless mode and hence no need for NOHZ idle load
10176 if (likely(!atomic_read(&nohz.nr_cpus)))
10179 if (READ_ONCE(nohz.has_blocked) &&
10180 time_after(now, READ_ONCE(nohz.next_blocked)))
10181 flags = NOHZ_STATS_KICK;
10183 if (time_before(now, nohz.next_balance))
10186 if (rq->nr_running >= 2) {
10187 flags = NOHZ_KICK_MASK;
10193 sd = rcu_dereference(rq->sd);
10196 * If there's a CFS task and the current CPU has reduced
10197 * capacity; kick the ILB to see if there's a better CPU to run
10200 if (rq->cfs.h_nr_running >= 1 && check_cpu_capacity(rq, sd)) {
10201 flags = NOHZ_KICK_MASK;
10206 sd = rcu_dereference(per_cpu(sd_asym_packing, cpu));
10209 * When ASYM_PACKING; see if there's a more preferred CPU
10210 * currently idle; in which case, kick the ILB to move tasks
10213 for_each_cpu_and(i, sched_domain_span(sd), nohz.idle_cpus_mask) {
10214 if (sched_asym_prefer(i, cpu)) {
10215 flags = NOHZ_KICK_MASK;
10221 sd = rcu_dereference(per_cpu(sd_asym_cpucapacity, cpu));
10224 * When ASYM_CPUCAPACITY; see if there's a higher capacity CPU
10225 * to run the misfit task on.
10227 if (check_misfit_status(rq, sd)) {
10228 flags = NOHZ_KICK_MASK;
10233 * For asymmetric systems, we do not want to nicely balance
10234 * cache use, instead we want to embrace asymmetry and only
10235 * ensure tasks have enough CPU capacity.
10237 * Skip the LLC logic because it's not relevant in that case.
10242 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
10245 * If there is an imbalance between LLC domains (IOW we could
10246 * increase the overall cache use), we need some less-loaded LLC
10247 * domain to pull some load. Likewise, we may need to spread
10248 * load within the current LLC domain (e.g. packed SMT cores but
10249 * other CPUs are idle). We can't really know from here how busy
10250 * the others are - so just get a nohz balance going if it looks
10251 * like this LLC domain has tasks we could move.
10253 nr_busy = atomic_read(&sds->nr_busy_cpus);
10255 flags = NOHZ_KICK_MASK;
10266 static void set_cpu_sd_state_busy(int cpu)
10268 struct sched_domain *sd;
10271 sd = rcu_dereference(per_cpu(sd_llc, cpu));
10273 if (!sd || !sd->nohz_idle)
10277 atomic_inc(&sd->shared->nr_busy_cpus);
10282 void nohz_balance_exit_idle(struct rq *rq)
10284 SCHED_WARN_ON(rq != this_rq());
10286 if (likely(!rq->nohz_tick_stopped))
10289 rq->nohz_tick_stopped = 0;
10290 cpumask_clear_cpu(rq->cpu, nohz.idle_cpus_mask);
10291 atomic_dec(&nohz.nr_cpus);
10293 set_cpu_sd_state_busy(rq->cpu);
10296 static void set_cpu_sd_state_idle(int cpu)
10298 struct sched_domain *sd;
10301 sd = rcu_dereference(per_cpu(sd_llc, cpu));
10303 if (!sd || sd->nohz_idle)
10307 atomic_dec(&sd->shared->nr_busy_cpus);
10313 * This routine will record that the CPU is going idle with tick stopped.
10314 * This info will be used in performing idle load balancing in the future.
10316 void nohz_balance_enter_idle(int cpu)
10318 struct rq *rq = cpu_rq(cpu);
10320 SCHED_WARN_ON(cpu != smp_processor_id());
10322 /* If this CPU is going down, then nothing needs to be done: */
10323 if (!cpu_active(cpu))
10326 /* Spare idle load balancing on CPUs that don't want to be disturbed: */
10327 if (!housekeeping_cpu(cpu, HK_FLAG_SCHED))
10331 * Can be set safely without rq->lock held
10332 * If a clear happens, it will have evaluated last additions because
10333 * rq->lock is held during the check and the clear
10335 rq->has_blocked_load = 1;
10338 * The tick is still stopped but load could have been added in the
10339 * meantime. We set the nohz.has_blocked flag to trig a check of the
10340 * *_avg. The CPU is already part of nohz.idle_cpus_mask so the clear
10341 * of nohz.has_blocked can only happen after checking the new load
10343 if (rq->nohz_tick_stopped)
10346 /* If we're a completely isolated CPU, we don't play: */
10347 if (on_null_domain(rq))
10350 rq->nohz_tick_stopped = 1;
10352 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
10353 atomic_inc(&nohz.nr_cpus);
10356 * Ensures that if nohz_idle_balance() fails to observe our
10357 * @idle_cpus_mask store, it must observe the @has_blocked
10360 smp_mb__after_atomic();
10362 set_cpu_sd_state_idle(cpu);
10366 * Each time a cpu enter idle, we assume that it has blocked load and
10367 * enable the periodic update of the load of idle cpus
10369 WRITE_ONCE(nohz.has_blocked, 1);
10373 * Internal function that runs load balance for all idle cpus. The load balance
10374 * can be a simple update of blocked load or a complete load balance with
10375 * tasks movement depending of flags.
10377 static void _nohz_idle_balance(struct rq *this_rq, unsigned int flags,
10378 enum cpu_idle_type idle)
10380 /* Earliest time when we have to do rebalance again */
10381 unsigned long now = jiffies;
10382 unsigned long next_balance = now + 60*HZ;
10383 bool has_blocked_load = false;
10384 int update_next_balance = 0;
10385 int this_cpu = this_rq->cpu;
10389 SCHED_WARN_ON((flags & NOHZ_KICK_MASK) == NOHZ_BALANCE_KICK);
10392 * We assume there will be no idle load after this update and clear
10393 * the has_blocked flag. If a cpu enters idle in the mean time, it will
10394 * set the has_blocked flag and trig another update of idle load.
10395 * Because a cpu that becomes idle, is added to idle_cpus_mask before
10396 * setting the flag, we are sure to not clear the state and not
10397 * check the load of an idle cpu.
10399 WRITE_ONCE(nohz.has_blocked, 0);
10402 * Ensures that if we miss the CPU, we must see the has_blocked
10403 * store from nohz_balance_enter_idle().
10408 * Start with the next CPU after this_cpu so we will end with this_cpu and let a
10409 * chance for other idle cpu to pull load.
10411 for_each_cpu_wrap(balance_cpu, nohz.idle_cpus_mask, this_cpu+1) {
10412 if (!idle_cpu(balance_cpu))
10416 * If this CPU gets work to do, stop the load balancing
10417 * work being done for other CPUs. Next load
10418 * balancing owner will pick it up.
10420 if (need_resched()) {
10421 has_blocked_load = true;
10425 rq = cpu_rq(balance_cpu);
10427 has_blocked_load |= update_nohz_stats(rq);
10430 * If time for next balance is due,
10433 if (time_after_eq(jiffies, rq->next_balance)) {
10434 struct rq_flags rf;
10436 rq_lock_irqsave(rq, &rf);
10437 update_rq_clock(rq);
10438 rq_unlock_irqrestore(rq, &rf);
10440 if (flags & NOHZ_BALANCE_KICK)
10441 rebalance_domains(rq, CPU_IDLE);
10444 if (time_after(next_balance, rq->next_balance)) {
10445 next_balance = rq->next_balance;
10446 update_next_balance = 1;
10451 * next_balance will be updated only when there is a need.
10452 * When the CPU is attached to null domain for ex, it will not be
10455 if (likely(update_next_balance))
10456 nohz.next_balance = next_balance;
10458 WRITE_ONCE(nohz.next_blocked,
10459 now + msecs_to_jiffies(LOAD_AVG_PERIOD));
10462 /* There is still blocked load, enable periodic update */
10463 if (has_blocked_load)
10464 WRITE_ONCE(nohz.has_blocked, 1);
10468 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
10469 * rebalancing for all the cpus for whom scheduler ticks are stopped.
10471 static bool nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
10473 unsigned int flags = this_rq->nohz_idle_balance;
10478 this_rq->nohz_idle_balance = 0;
10480 if (idle != CPU_IDLE)
10483 _nohz_idle_balance(this_rq, flags, idle);
10489 * Check if we need to run the ILB for updating blocked load before entering
10492 void nohz_run_idle_balance(int cpu)
10494 unsigned int flags;
10496 flags = atomic_fetch_andnot(NOHZ_NEWILB_KICK, nohz_flags(cpu));
10499 * Update the blocked load only if no SCHED_SOFTIRQ is about to happen
10500 * (ie NOHZ_STATS_KICK set) and will do the same.
10502 if ((flags == NOHZ_NEWILB_KICK) && !need_resched())
10503 _nohz_idle_balance(cpu_rq(cpu), NOHZ_STATS_KICK, CPU_IDLE);
10506 static void nohz_newidle_balance(struct rq *this_rq)
10508 int this_cpu = this_rq->cpu;
10511 * This CPU doesn't want to be disturbed by scheduler
10514 if (!housekeeping_cpu(this_cpu, HK_FLAG_SCHED))
10517 /* Will wake up very soon. No time for doing anything else*/
10518 if (this_rq->avg_idle < sysctl_sched_migration_cost)
10521 /* Don't need to update blocked load of idle CPUs*/
10522 if (!READ_ONCE(nohz.has_blocked) ||
10523 time_before(jiffies, READ_ONCE(nohz.next_blocked)))
10527 * Set the need to trigger ILB in order to update blocked load
10528 * before entering idle state.
10530 atomic_or(NOHZ_NEWILB_KICK, nohz_flags(this_cpu));
10533 #else /* !CONFIG_NO_HZ_COMMON */
10534 static inline void nohz_balancer_kick(struct rq *rq) { }
10536 static inline bool nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
10541 static inline void nohz_newidle_balance(struct rq *this_rq) { }
10542 #endif /* CONFIG_NO_HZ_COMMON */
10545 * newidle_balance is called by schedule() if this_cpu is about to become
10546 * idle. Attempts to pull tasks from other CPUs.
10549 * < 0 - we released the lock and there are !fair tasks present
10550 * 0 - failed, no new tasks
10551 * > 0 - success, new (fair) tasks present
10553 static int newidle_balance(struct rq *this_rq, struct rq_flags *rf)
10555 unsigned long next_balance = jiffies + HZ;
10556 int this_cpu = this_rq->cpu;
10557 struct sched_domain *sd;
10558 int pulled_task = 0;
10561 update_misfit_status(NULL, this_rq);
10563 * We must set idle_stamp _before_ calling idle_balance(), such that we
10564 * measure the duration of idle_balance() as idle time.
10566 this_rq->idle_stamp = rq_clock(this_rq);
10569 * Do not pull tasks towards !active CPUs...
10571 if (!cpu_active(this_cpu))
10575 * This is OK, because current is on_cpu, which avoids it being picked
10576 * for load-balance and preemption/IRQs are still disabled avoiding
10577 * further scheduler activity on it and we're being very careful to
10578 * re-start the picking loop.
10580 rq_unpin_lock(this_rq, rf);
10582 if (this_rq->avg_idle < sysctl_sched_migration_cost ||
10583 !READ_ONCE(this_rq->rd->overload)) {
10586 sd = rcu_dereference_check_sched_domain(this_rq->sd);
10588 update_next_balance(sd, &next_balance);
10594 raw_spin_unlock(&this_rq->lock);
10596 update_blocked_averages(this_cpu);
10598 for_each_domain(this_cpu, sd) {
10599 int continue_balancing = 1;
10600 u64 t0, domain_cost;
10602 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost) {
10603 update_next_balance(sd, &next_balance);
10607 if (sd->flags & SD_BALANCE_NEWIDLE) {
10608 t0 = sched_clock_cpu(this_cpu);
10610 pulled_task = load_balance(this_cpu, this_rq,
10611 sd, CPU_NEWLY_IDLE,
10612 &continue_balancing);
10614 domain_cost = sched_clock_cpu(this_cpu) - t0;
10615 if (domain_cost > sd->max_newidle_lb_cost)
10616 sd->max_newidle_lb_cost = domain_cost;
10618 curr_cost += domain_cost;
10621 update_next_balance(sd, &next_balance);
10624 * Stop searching for tasks to pull if there are
10625 * now runnable tasks on this rq.
10627 if (pulled_task || this_rq->nr_running > 0)
10632 raw_spin_lock(&this_rq->lock);
10634 if (curr_cost > this_rq->max_idle_balance_cost)
10635 this_rq->max_idle_balance_cost = curr_cost;
10638 * While browsing the domains, we released the rq lock, a task could
10639 * have been enqueued in the meantime. Since we're not going idle,
10640 * pretend we pulled a task.
10642 if (this_rq->cfs.h_nr_running && !pulled_task)
10645 /* Is there a task of a high priority class? */
10646 if (this_rq->nr_running != this_rq->cfs.h_nr_running)
10650 /* Move the next balance forward */
10651 if (time_after(this_rq->next_balance, next_balance))
10652 this_rq->next_balance = next_balance;
10655 this_rq->idle_stamp = 0;
10657 nohz_newidle_balance(this_rq);
10659 rq_repin_lock(this_rq, rf);
10661 return pulled_task;
10665 * run_rebalance_domains is triggered when needed from the scheduler tick.
10666 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
10668 static __latent_entropy void run_rebalance_domains(struct softirq_action *h)
10670 struct rq *this_rq = this_rq();
10671 enum cpu_idle_type idle = this_rq->idle_balance ?
10672 CPU_IDLE : CPU_NOT_IDLE;
10675 * If this CPU has a pending nohz_balance_kick, then do the
10676 * balancing on behalf of the other idle CPUs whose ticks are
10677 * stopped. Do nohz_idle_balance *before* rebalance_domains to
10678 * give the idle CPUs a chance to load balance. Else we may
10679 * load balance only within the local sched_domain hierarchy
10680 * and abort nohz_idle_balance altogether if we pull some load.
10682 if (nohz_idle_balance(this_rq, idle))
10685 /* normal load balance */
10686 update_blocked_averages(this_rq->cpu);
10687 rebalance_domains(this_rq, idle);
10691 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
10693 void trigger_load_balance(struct rq *rq)
10696 * Don't need to rebalance while attached to NULL domain or
10697 * runqueue CPU is not active
10699 if (unlikely(on_null_domain(rq) || !cpu_active(cpu_of(rq))))
10702 if (time_after_eq(jiffies, rq->next_balance))
10703 raise_softirq(SCHED_SOFTIRQ);
10705 nohz_balancer_kick(rq);
10708 static void rq_online_fair(struct rq *rq)
10712 update_runtime_enabled(rq);
10715 static void rq_offline_fair(struct rq *rq)
10719 /* Ensure any throttled groups are reachable by pick_next_task */
10720 unthrottle_offline_cfs_rqs(rq);
10723 #endif /* CONFIG_SMP */
10726 * scheduler tick hitting a task of our scheduling class.
10728 * NOTE: This function can be called remotely by the tick offload that
10729 * goes along full dynticks. Therefore no local assumption can be made
10730 * and everything must be accessed through the @rq and @curr passed in
10733 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
10735 struct cfs_rq *cfs_rq;
10736 struct sched_entity *se = &curr->se;
10738 for_each_sched_entity(se) {
10739 cfs_rq = cfs_rq_of(se);
10740 entity_tick(cfs_rq, se, queued);
10743 if (static_branch_unlikely(&sched_numa_balancing))
10744 task_tick_numa(rq, curr);
10746 update_misfit_status(curr, rq);
10747 update_overutilized_status(task_rq(curr));
10751 * called on fork with the child task as argument from the parent's context
10752 * - child not yet on the tasklist
10753 * - preemption disabled
10755 static void task_fork_fair(struct task_struct *p)
10757 struct cfs_rq *cfs_rq;
10758 struct sched_entity *se = &p->se, *curr;
10759 struct rq *rq = this_rq();
10760 struct rq_flags rf;
10763 update_rq_clock(rq);
10765 cfs_rq = task_cfs_rq(current);
10766 curr = cfs_rq->curr;
10768 update_curr(cfs_rq);
10769 se->vruntime = curr->vruntime;
10771 place_entity(cfs_rq, se, 1);
10773 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
10775 * Upon rescheduling, sched_class::put_prev_task() will place
10776 * 'current' within the tree based on its new key value.
10778 swap(curr->vruntime, se->vruntime);
10782 se->vruntime -= cfs_rq->min_vruntime;
10783 rq_unlock(rq, &rf);
10787 * Priority of the task has changed. Check to see if we preempt
10788 * the current task.
10791 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
10793 if (!task_on_rq_queued(p))
10796 if (rq->cfs.nr_running == 1)
10800 * Reschedule if we are currently running on this runqueue and
10801 * our priority decreased, or if we are not currently running on
10802 * this runqueue and our priority is higher than the current's
10804 if (task_current(rq, p)) {
10805 if (p->prio > oldprio)
10808 check_preempt_curr(rq, p, 0);
10811 static inline bool vruntime_normalized(struct task_struct *p)
10813 struct sched_entity *se = &p->se;
10816 * In both the TASK_ON_RQ_QUEUED and TASK_ON_RQ_MIGRATING cases,
10817 * the dequeue_entity(.flags=0) will already have normalized the
10824 * When !on_rq, vruntime of the task has usually NOT been normalized.
10825 * But there are some cases where it has already been normalized:
10827 * - A forked child which is waiting for being woken up by
10828 * wake_up_new_task().
10829 * - A task which has been woken up by try_to_wake_up() and
10830 * waiting for actually being woken up by sched_ttwu_pending().
10832 if (!se->sum_exec_runtime ||
10833 (p->state == TASK_WAKING && p->sched_remote_wakeup))
10839 #ifdef CONFIG_FAIR_GROUP_SCHED
10841 * Propagate the changes of the sched_entity across the tg tree to make it
10842 * visible to the root
10844 static void propagate_entity_cfs_rq(struct sched_entity *se)
10846 struct cfs_rq *cfs_rq;
10848 /* Start to propagate at parent */
10851 for_each_sched_entity(se) {
10852 cfs_rq = cfs_rq_of(se);
10854 if (cfs_rq_throttled(cfs_rq))
10857 update_load_avg(cfs_rq, se, UPDATE_TG);
10861 static void propagate_entity_cfs_rq(struct sched_entity *se) { }
10864 static void detach_entity_cfs_rq(struct sched_entity *se)
10866 struct cfs_rq *cfs_rq = cfs_rq_of(se);
10868 /* Catch up with the cfs_rq and remove our load when we leave */
10869 update_load_avg(cfs_rq, se, 0);
10870 detach_entity_load_avg(cfs_rq, se);
10871 update_tg_load_avg(cfs_rq);
10872 propagate_entity_cfs_rq(se);
10875 static void attach_entity_cfs_rq(struct sched_entity *se)
10877 struct cfs_rq *cfs_rq = cfs_rq_of(se);
10879 #ifdef CONFIG_FAIR_GROUP_SCHED
10881 * Since the real-depth could have been changed (only FAIR
10882 * class maintain depth value), reset depth properly.
10884 se->depth = se->parent ? se->parent->depth + 1 : 0;
10887 /* Synchronize entity with its cfs_rq */
10888 update_load_avg(cfs_rq, se, sched_feat(ATTACH_AGE_LOAD) ? 0 : SKIP_AGE_LOAD);
10889 attach_entity_load_avg(cfs_rq, se);
10890 update_tg_load_avg(cfs_rq);
10891 propagate_entity_cfs_rq(se);
10894 static void detach_task_cfs_rq(struct task_struct *p)
10896 struct sched_entity *se = &p->se;
10897 struct cfs_rq *cfs_rq = cfs_rq_of(se);
10899 if (!vruntime_normalized(p)) {
10901 * Fix up our vruntime so that the current sleep doesn't
10902 * cause 'unlimited' sleep bonus.
10904 place_entity(cfs_rq, se, 0);
10905 se->vruntime -= cfs_rq->min_vruntime;
10908 detach_entity_cfs_rq(se);
10911 static void attach_task_cfs_rq(struct task_struct *p)
10913 struct sched_entity *se = &p->se;
10914 struct cfs_rq *cfs_rq = cfs_rq_of(se);
10916 attach_entity_cfs_rq(se);
10918 if (!vruntime_normalized(p))
10919 se->vruntime += cfs_rq->min_vruntime;
10922 static void switched_from_fair(struct rq *rq, struct task_struct *p)
10924 detach_task_cfs_rq(p);
10927 static void switched_to_fair(struct rq *rq, struct task_struct *p)
10929 attach_task_cfs_rq(p);
10931 if (task_on_rq_queued(p)) {
10933 * We were most likely switched from sched_rt, so
10934 * kick off the schedule if running, otherwise just see
10935 * if we can still preempt the current task.
10937 if (task_current(rq, p))
10940 check_preempt_curr(rq, p, 0);
10944 /* Account for a task changing its policy or group.
10946 * This routine is mostly called to set cfs_rq->curr field when a task
10947 * migrates between groups/classes.
10949 static void set_next_task_fair(struct rq *rq, struct task_struct *p, bool first)
10951 struct sched_entity *se = &p->se;
10954 if (task_on_rq_queued(p)) {
10956 * Move the next running task to the front of the list, so our
10957 * cfs_tasks list becomes MRU one.
10959 list_move(&se->group_node, &rq->cfs_tasks);
10963 for_each_sched_entity(se) {
10964 struct cfs_rq *cfs_rq = cfs_rq_of(se);
10966 set_next_entity(cfs_rq, se);
10967 /* ensure bandwidth has been allocated on our new cfs_rq */
10968 account_cfs_rq_runtime(cfs_rq, 0);
10972 void init_cfs_rq(struct cfs_rq *cfs_rq)
10974 cfs_rq->tasks_timeline = RB_ROOT_CACHED;
10975 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
10976 #ifndef CONFIG_64BIT
10977 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
10980 raw_spin_lock_init(&cfs_rq->removed.lock);
10984 #ifdef CONFIG_FAIR_GROUP_SCHED
10985 static void task_set_group_fair(struct task_struct *p)
10987 struct sched_entity *se = &p->se;
10989 set_task_rq(p, task_cpu(p));
10990 se->depth = se->parent ? se->parent->depth + 1 : 0;
10993 static void task_move_group_fair(struct task_struct *p)
10995 detach_task_cfs_rq(p);
10996 set_task_rq(p, task_cpu(p));
10999 /* Tell se's cfs_rq has been changed -- migrated */
11000 p->se.avg.last_update_time = 0;
11002 attach_task_cfs_rq(p);
11005 static void task_change_group_fair(struct task_struct *p, int type)
11008 case TASK_SET_GROUP:
11009 task_set_group_fair(p);
11012 case TASK_MOVE_GROUP:
11013 task_move_group_fair(p);
11018 void free_fair_sched_group(struct task_group *tg)
11022 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
11024 for_each_possible_cpu(i) {
11026 kfree(tg->cfs_rq[i]);
11035 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
11037 struct sched_entity *se;
11038 struct cfs_rq *cfs_rq;
11041 tg->cfs_rq = kcalloc(nr_cpu_ids, sizeof(cfs_rq), GFP_KERNEL);
11044 tg->se = kcalloc(nr_cpu_ids, sizeof(se), GFP_KERNEL);
11048 tg->shares = NICE_0_LOAD;
11050 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
11052 for_each_possible_cpu(i) {
11053 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
11054 GFP_KERNEL, cpu_to_node(i));
11058 se = kzalloc_node(sizeof(struct sched_entity),
11059 GFP_KERNEL, cpu_to_node(i));
11063 init_cfs_rq(cfs_rq);
11064 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
11065 init_entity_runnable_average(se);
11076 void online_fair_sched_group(struct task_group *tg)
11078 struct sched_entity *se;
11079 struct rq_flags rf;
11083 for_each_possible_cpu(i) {
11086 rq_lock_irq(rq, &rf);
11087 update_rq_clock(rq);
11088 attach_entity_cfs_rq(se);
11089 sync_throttle(tg, i);
11090 rq_unlock_irq(rq, &rf);
11094 void unregister_fair_sched_group(struct task_group *tg)
11096 unsigned long flags;
11100 for_each_possible_cpu(cpu) {
11102 remove_entity_load_avg(tg->se[cpu]);
11105 * Only empty task groups can be destroyed; so we can speculatively
11106 * check on_list without danger of it being re-added.
11108 if (!tg->cfs_rq[cpu]->on_list)
11113 raw_spin_lock_irqsave(&rq->lock, flags);
11114 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
11115 raw_spin_unlock_irqrestore(&rq->lock, flags);
11119 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
11120 struct sched_entity *se, int cpu,
11121 struct sched_entity *parent)
11123 struct rq *rq = cpu_rq(cpu);
11127 init_cfs_rq_runtime(cfs_rq);
11129 tg->cfs_rq[cpu] = cfs_rq;
11132 /* se could be NULL for root_task_group */
11137 se->cfs_rq = &rq->cfs;
11140 se->cfs_rq = parent->my_q;
11141 se->depth = parent->depth + 1;
11145 /* guarantee group entities always have weight */
11146 update_load_set(&se->load, NICE_0_LOAD);
11147 se->parent = parent;
11150 static DEFINE_MUTEX(shares_mutex);
11152 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
11157 * We can't change the weight of the root cgroup.
11162 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
11164 mutex_lock(&shares_mutex);
11165 if (tg->shares == shares)
11168 tg->shares = shares;
11169 for_each_possible_cpu(i) {
11170 struct rq *rq = cpu_rq(i);
11171 struct sched_entity *se = tg->se[i];
11172 struct rq_flags rf;
11174 /* Propagate contribution to hierarchy */
11175 rq_lock_irqsave(rq, &rf);
11176 update_rq_clock(rq);
11177 for_each_sched_entity(se) {
11178 update_load_avg(cfs_rq_of(se), se, UPDATE_TG);
11179 update_cfs_group(se);
11181 rq_unlock_irqrestore(rq, &rf);
11185 mutex_unlock(&shares_mutex);
11188 #else /* CONFIG_FAIR_GROUP_SCHED */
11190 void free_fair_sched_group(struct task_group *tg) { }
11192 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
11197 void online_fair_sched_group(struct task_group *tg) { }
11199 void unregister_fair_sched_group(struct task_group *tg) { }
11201 #endif /* CONFIG_FAIR_GROUP_SCHED */
11204 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
11206 struct sched_entity *se = &task->se;
11207 unsigned int rr_interval = 0;
11210 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
11213 if (rq->cfs.load.weight)
11214 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
11216 return rr_interval;
11220 * All the scheduling class methods:
11222 DEFINE_SCHED_CLASS(fair) = {
11224 .enqueue_task = enqueue_task_fair,
11225 .dequeue_task = dequeue_task_fair,
11226 .yield_task = yield_task_fair,
11227 .yield_to_task = yield_to_task_fair,
11229 .check_preempt_curr = check_preempt_wakeup,
11231 .pick_next_task = __pick_next_task_fair,
11232 .put_prev_task = put_prev_task_fair,
11233 .set_next_task = set_next_task_fair,
11236 .balance = balance_fair,
11237 .select_task_rq = select_task_rq_fair,
11238 .migrate_task_rq = migrate_task_rq_fair,
11240 .rq_online = rq_online_fair,
11241 .rq_offline = rq_offline_fair,
11243 .task_dead = task_dead_fair,
11244 .set_cpus_allowed = set_cpus_allowed_common,
11247 .task_tick = task_tick_fair,
11248 .task_fork = task_fork_fair,
11250 .prio_changed = prio_changed_fair,
11251 .switched_from = switched_from_fair,
11252 .switched_to = switched_to_fair,
11254 .get_rr_interval = get_rr_interval_fair,
11256 .update_curr = update_curr_fair,
11258 #ifdef CONFIG_FAIR_GROUP_SCHED
11259 .task_change_group = task_change_group_fair,
11262 #ifdef CONFIG_UCLAMP_TASK
11263 .uclamp_enabled = 1,
11267 #ifdef CONFIG_SCHED_DEBUG
11268 void print_cfs_stats(struct seq_file *m, int cpu)
11270 struct cfs_rq *cfs_rq, *pos;
11273 for_each_leaf_cfs_rq_safe(cpu_rq(cpu), cfs_rq, pos)
11274 print_cfs_rq(m, cpu, cfs_rq);
11278 #ifdef CONFIG_NUMA_BALANCING
11279 void show_numa_stats(struct task_struct *p, struct seq_file *m)
11282 unsigned long tsf = 0, tpf = 0, gsf = 0, gpf = 0;
11283 struct numa_group *ng;
11286 ng = rcu_dereference(p->numa_group);
11287 for_each_online_node(node) {
11288 if (p->numa_faults) {
11289 tsf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 0)];
11290 tpf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 1)];
11293 gsf = ng->faults[task_faults_idx(NUMA_MEM, node, 0)],
11294 gpf = ng->faults[task_faults_idx(NUMA_MEM, node, 1)];
11296 print_numa_stats(m, node, tsf, tpf, gsf, gpf);
11300 #endif /* CONFIG_NUMA_BALANCING */
11301 #endif /* CONFIG_SCHED_DEBUG */
11303 __init void init_sched_fair_class(void)
11306 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
11308 #ifdef CONFIG_NO_HZ_COMMON
11309 nohz.next_balance = jiffies;
11310 nohz.next_blocked = jiffies;
11311 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
11318 * Helper functions to facilitate extracting info from tracepoints.
11321 const struct sched_avg *sched_trace_cfs_rq_avg(struct cfs_rq *cfs_rq)
11324 return cfs_rq ? &cfs_rq->avg : NULL;
11329 EXPORT_SYMBOL_GPL(sched_trace_cfs_rq_avg);
11331 char *sched_trace_cfs_rq_path(struct cfs_rq *cfs_rq, char *str, int len)
11335 strlcpy(str, "(null)", len);
11340 cfs_rq_tg_path(cfs_rq, str, len);
11343 EXPORT_SYMBOL_GPL(sched_trace_cfs_rq_path);
11345 int sched_trace_cfs_rq_cpu(struct cfs_rq *cfs_rq)
11347 return cfs_rq ? cpu_of(rq_of(cfs_rq)) : -1;
11349 EXPORT_SYMBOL_GPL(sched_trace_cfs_rq_cpu);
11351 const struct sched_avg *sched_trace_rq_avg_rt(struct rq *rq)
11354 return rq ? &rq->avg_rt : NULL;
11359 EXPORT_SYMBOL_GPL(sched_trace_rq_avg_rt);
11361 const struct sched_avg *sched_trace_rq_avg_dl(struct rq *rq)
11364 return rq ? &rq->avg_dl : NULL;
11369 EXPORT_SYMBOL_GPL(sched_trace_rq_avg_dl);
11371 const struct sched_avg *sched_trace_rq_avg_irq(struct rq *rq)
11373 #if defined(CONFIG_SMP) && defined(CONFIG_HAVE_SCHED_AVG_IRQ)
11374 return rq ? &rq->avg_irq : NULL;
11379 EXPORT_SYMBOL_GPL(sched_trace_rq_avg_irq);
11381 int sched_trace_rq_cpu(struct rq *rq)
11383 return rq ? cpu_of(rq) : -1;
11385 EXPORT_SYMBOL_GPL(sched_trace_rq_cpu);
11387 int sched_trace_rq_cpu_capacity(struct rq *rq)
11393 SCHED_CAPACITY_SCALE
11397 EXPORT_SYMBOL_GPL(sched_trace_rq_cpu_capacity);
11399 const struct cpumask *sched_trace_rd_span(struct root_domain *rd)
11402 return rd ? rd->span : NULL;
11407 EXPORT_SYMBOL_GPL(sched_trace_rd_span);
11409 int sched_trace_rq_nr_running(struct rq *rq)
11411 return rq ? rq->nr_running : -1;
11413 EXPORT_SYMBOL_GPL(sched_trace_rq_nr_running);