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 unsigned int 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)
117 * The margin used when comparing CPU capacities.
118 * is 'cap1' noticeably greater than 'cap2'
122 #define capacity_greater(cap1, cap2) ((cap1) * 1024 > (cap2) * 1078)
125 #ifdef CONFIG_CFS_BANDWIDTH
127 * Amount of runtime to allocate from global (tg) to local (per-cfs_rq) pool
128 * each time a cfs_rq requests quota.
130 * Note: in the case that the slice exceeds the runtime remaining (either due
131 * to consumption or the quota being specified to be smaller than the slice)
132 * we will always only issue the remaining available time.
134 * (default: 5 msec, units: microseconds)
136 unsigned int sysctl_sched_cfs_bandwidth_slice = 5000UL;
139 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
145 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
151 static inline void update_load_set(struct load_weight *lw, unsigned long w)
158 * Increase the granularity value when there are more CPUs,
159 * because with more CPUs the 'effective latency' as visible
160 * to users decreases. But the relationship is not linear,
161 * so pick a second-best guess by going with the log2 of the
164 * This idea comes from the SD scheduler of Con Kolivas:
166 static unsigned int get_update_sysctl_factor(void)
168 unsigned int cpus = min_t(unsigned int, num_online_cpus(), 8);
171 switch (sysctl_sched_tunable_scaling) {
172 case SCHED_TUNABLESCALING_NONE:
175 case SCHED_TUNABLESCALING_LINEAR:
178 case SCHED_TUNABLESCALING_LOG:
180 factor = 1 + ilog2(cpus);
187 static void update_sysctl(void)
189 unsigned int factor = get_update_sysctl_factor();
191 #define SET_SYSCTL(name) \
192 (sysctl_##name = (factor) * normalized_sysctl_##name)
193 SET_SYSCTL(sched_min_granularity);
194 SET_SYSCTL(sched_latency);
195 SET_SYSCTL(sched_wakeup_granularity);
199 void __init sched_init_granularity(void)
204 #define WMULT_CONST (~0U)
205 #define WMULT_SHIFT 32
207 static void __update_inv_weight(struct load_weight *lw)
211 if (likely(lw->inv_weight))
214 w = scale_load_down(lw->weight);
216 if (BITS_PER_LONG > 32 && unlikely(w >= WMULT_CONST))
218 else if (unlikely(!w))
219 lw->inv_weight = WMULT_CONST;
221 lw->inv_weight = WMULT_CONST / w;
225 * delta_exec * weight / lw.weight
227 * (delta_exec * (weight * lw->inv_weight)) >> WMULT_SHIFT
229 * Either weight := NICE_0_LOAD and lw \e sched_prio_to_wmult[], in which case
230 * we're guaranteed shift stays positive because inv_weight is guaranteed to
231 * fit 32 bits, and NICE_0_LOAD gives another 10 bits; therefore shift >= 22.
233 * Or, weight =< lw.weight (because lw.weight is the runqueue weight), thus
234 * weight/lw.weight <= 1, and therefore our shift will also be positive.
236 static u64 __calc_delta(u64 delta_exec, unsigned long weight, struct load_weight *lw)
238 u64 fact = scale_load_down(weight);
239 u32 fact_hi = (u32)(fact >> 32);
240 int shift = WMULT_SHIFT;
243 __update_inv_weight(lw);
245 if (unlikely(fact_hi)) {
251 fact = mul_u32_u32(fact, lw->inv_weight);
253 fact_hi = (u32)(fact >> 32);
260 return mul_u64_u32_shr(delta_exec, fact, shift);
264 const struct sched_class fair_sched_class;
266 /**************************************************************
267 * CFS operations on generic schedulable entities:
270 #ifdef CONFIG_FAIR_GROUP_SCHED
272 /* Walk up scheduling entities hierarchy */
273 #define for_each_sched_entity(se) \
274 for (; se; se = se->parent)
276 static inline void cfs_rq_tg_path(struct cfs_rq *cfs_rq, char *path, int len)
281 if (cfs_rq && task_group_is_autogroup(cfs_rq->tg))
282 autogroup_path(cfs_rq->tg, path, len);
283 else if (cfs_rq && cfs_rq->tg->css.cgroup)
284 cgroup_path(cfs_rq->tg->css.cgroup, path, len);
286 strlcpy(path, "(null)", len);
289 static inline bool list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
291 struct rq *rq = rq_of(cfs_rq);
292 int cpu = cpu_of(rq);
295 return rq->tmp_alone_branch == &rq->leaf_cfs_rq_list;
300 * Ensure we either appear before our parent (if already
301 * enqueued) or force our parent to appear after us when it is
302 * enqueued. The fact that we always enqueue bottom-up
303 * reduces this to two cases and a special case for the root
304 * cfs_rq. Furthermore, it also means that we will always reset
305 * tmp_alone_branch either when the branch is connected
306 * to a tree or when we reach the top of the tree
308 if (cfs_rq->tg->parent &&
309 cfs_rq->tg->parent->cfs_rq[cpu]->on_list) {
311 * If parent is already on the list, we add the child
312 * just before. Thanks to circular linked property of
313 * the list, this means to put the child at the tail
314 * of the list that starts by parent.
316 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
317 &(cfs_rq->tg->parent->cfs_rq[cpu]->leaf_cfs_rq_list));
319 * The branch is now connected to its tree so we can
320 * reset tmp_alone_branch to the beginning of the
323 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
327 if (!cfs_rq->tg->parent) {
329 * cfs rq without parent should be put
330 * at the tail of the list.
332 list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
333 &rq->leaf_cfs_rq_list);
335 * We have reach the top of a tree so we can reset
336 * tmp_alone_branch to the beginning of the list.
338 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
343 * The parent has not already been added so we want to
344 * make sure that it will be put after us.
345 * tmp_alone_branch points to the begin of the branch
346 * where we will add parent.
348 list_add_rcu(&cfs_rq->leaf_cfs_rq_list, rq->tmp_alone_branch);
350 * update tmp_alone_branch to points to the new begin
353 rq->tmp_alone_branch = &cfs_rq->leaf_cfs_rq_list;
357 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
359 if (cfs_rq->on_list) {
360 struct rq *rq = rq_of(cfs_rq);
363 * With cfs_rq being unthrottled/throttled during an enqueue,
364 * it can happen the tmp_alone_branch points the a leaf that
365 * we finally want to del. In this case, tmp_alone_branch moves
366 * to the prev element but it will point to rq->leaf_cfs_rq_list
367 * at the end of the enqueue.
369 if (rq->tmp_alone_branch == &cfs_rq->leaf_cfs_rq_list)
370 rq->tmp_alone_branch = cfs_rq->leaf_cfs_rq_list.prev;
372 list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
377 static inline void assert_list_leaf_cfs_rq(struct rq *rq)
379 SCHED_WARN_ON(rq->tmp_alone_branch != &rq->leaf_cfs_rq_list);
382 /* Iterate thr' all leaf cfs_rq's on a runqueue */
383 #define for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) \
384 list_for_each_entry_safe(cfs_rq, pos, &rq->leaf_cfs_rq_list, \
387 /* Do the two (enqueued) entities belong to the same group ? */
388 static inline struct cfs_rq *
389 is_same_group(struct sched_entity *se, struct sched_entity *pse)
391 if (se->cfs_rq == pse->cfs_rq)
397 static inline struct sched_entity *parent_entity(struct sched_entity *se)
403 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
405 int se_depth, pse_depth;
408 * preemption test can be made between sibling entities who are in the
409 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
410 * both tasks until we find their ancestors who are siblings of common
414 /* First walk up until both entities are at same depth */
415 se_depth = (*se)->depth;
416 pse_depth = (*pse)->depth;
418 while (se_depth > pse_depth) {
420 *se = parent_entity(*se);
423 while (pse_depth > se_depth) {
425 *pse = parent_entity(*pse);
428 while (!is_same_group(*se, *pse)) {
429 *se = parent_entity(*se);
430 *pse = parent_entity(*pse);
434 #else /* !CONFIG_FAIR_GROUP_SCHED */
436 #define for_each_sched_entity(se) \
437 for (; se; se = NULL)
439 static inline void cfs_rq_tg_path(struct cfs_rq *cfs_rq, char *path, int len)
442 strlcpy(path, "(null)", len);
445 static inline bool list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
450 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
454 static inline void assert_list_leaf_cfs_rq(struct rq *rq)
458 #define for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) \
459 for (cfs_rq = &rq->cfs, pos = NULL; cfs_rq; cfs_rq = pos)
461 static inline struct sched_entity *parent_entity(struct sched_entity *se)
467 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
471 #endif /* CONFIG_FAIR_GROUP_SCHED */
473 static __always_inline
474 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec);
476 /**************************************************************
477 * Scheduling class tree data structure manipulation methods:
480 static inline u64 max_vruntime(u64 max_vruntime, u64 vruntime)
482 s64 delta = (s64)(vruntime - max_vruntime);
484 max_vruntime = vruntime;
489 static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime)
491 s64 delta = (s64)(vruntime - min_vruntime);
493 min_vruntime = vruntime;
498 static inline bool entity_before(struct sched_entity *a,
499 struct sched_entity *b)
501 return (s64)(a->vruntime - b->vruntime) < 0;
504 #define __node_2_se(node) \
505 rb_entry((node), struct sched_entity, run_node)
507 static void update_min_vruntime(struct cfs_rq *cfs_rq)
509 struct sched_entity *curr = cfs_rq->curr;
510 struct rb_node *leftmost = rb_first_cached(&cfs_rq->tasks_timeline);
512 u64 vruntime = cfs_rq->min_vruntime;
516 vruntime = curr->vruntime;
521 if (leftmost) { /* non-empty tree */
522 struct sched_entity *se = __node_2_se(leftmost);
525 vruntime = se->vruntime;
527 vruntime = min_vruntime(vruntime, se->vruntime);
530 /* ensure we never gain time by being placed backwards. */
531 cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime);
534 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
538 static inline bool __entity_less(struct rb_node *a, const struct rb_node *b)
540 return entity_before(__node_2_se(a), __node_2_se(b));
544 * Enqueue an entity into the rb-tree:
546 static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
548 rb_add_cached(&se->run_node, &cfs_rq->tasks_timeline, __entity_less);
551 static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
553 rb_erase_cached(&se->run_node, &cfs_rq->tasks_timeline);
556 struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq)
558 struct rb_node *left = rb_first_cached(&cfs_rq->tasks_timeline);
563 return __node_2_se(left);
566 static struct sched_entity *__pick_next_entity(struct sched_entity *se)
568 struct rb_node *next = rb_next(&se->run_node);
573 return __node_2_se(next);
576 #ifdef CONFIG_SCHED_DEBUG
577 struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
579 struct rb_node *last = rb_last(&cfs_rq->tasks_timeline.rb_root);
584 return __node_2_se(last);
587 /**************************************************************
588 * Scheduling class statistics methods:
591 int sched_update_scaling(void)
593 unsigned int factor = get_update_sysctl_factor();
595 sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency,
596 sysctl_sched_min_granularity);
598 #define WRT_SYSCTL(name) \
599 (normalized_sysctl_##name = sysctl_##name / (factor))
600 WRT_SYSCTL(sched_min_granularity);
601 WRT_SYSCTL(sched_latency);
602 WRT_SYSCTL(sched_wakeup_granularity);
612 static inline u64 calc_delta_fair(u64 delta, struct sched_entity *se)
614 if (unlikely(se->load.weight != NICE_0_LOAD))
615 delta = __calc_delta(delta, NICE_0_LOAD, &se->load);
621 * The idea is to set a period in which each task runs once.
623 * When there are too many tasks (sched_nr_latency) we have to stretch
624 * this period because otherwise the slices get too small.
626 * p = (nr <= nl) ? l : l*nr/nl
628 static u64 __sched_period(unsigned long nr_running)
630 if (unlikely(nr_running > sched_nr_latency))
631 return nr_running * sysctl_sched_min_granularity;
633 return sysctl_sched_latency;
637 * We calculate the wall-time slice from the period by taking a part
638 * proportional to the weight.
642 static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
644 unsigned int nr_running = cfs_rq->nr_running;
647 if (sched_feat(ALT_PERIOD))
648 nr_running = rq_of(cfs_rq)->cfs.h_nr_running;
650 slice = __sched_period(nr_running + !se->on_rq);
652 for_each_sched_entity(se) {
653 struct load_weight *load;
654 struct load_weight lw;
656 cfs_rq = cfs_rq_of(se);
657 load = &cfs_rq->load;
659 if (unlikely(!se->on_rq)) {
662 update_load_add(&lw, se->load.weight);
665 slice = __calc_delta(slice, se->load.weight, load);
668 if (sched_feat(BASE_SLICE))
669 slice = max(slice, (u64)sysctl_sched_min_granularity);
675 * We calculate the vruntime slice of a to-be-inserted task.
679 static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se)
681 return calc_delta_fair(sched_slice(cfs_rq, se), se);
687 static int select_idle_sibling(struct task_struct *p, int prev_cpu, int cpu);
688 static unsigned long task_h_load(struct task_struct *p);
689 static unsigned long capacity_of(int cpu);
691 /* Give new sched_entity start runnable values to heavy its load in infant time */
692 void init_entity_runnable_average(struct sched_entity *se)
694 struct sched_avg *sa = &se->avg;
696 memset(sa, 0, sizeof(*sa));
699 * Tasks are initialized with full load to be seen as heavy tasks until
700 * they get a chance to stabilize to their real load level.
701 * Group entities are initialized with zero load to reflect the fact that
702 * nothing has been attached to the task group yet.
704 if (entity_is_task(se))
705 sa->load_avg = scale_load_down(se->load.weight);
707 /* when this task enqueue'ed, it will contribute to its cfs_rq's load_avg */
710 static void attach_entity_cfs_rq(struct sched_entity *se);
713 * With new tasks being created, their initial util_avgs are extrapolated
714 * based on the cfs_rq's current util_avg:
716 * util_avg = cfs_rq->util_avg / (cfs_rq->load_avg + 1) * se.load.weight
718 * However, in many cases, the above util_avg does not give a desired
719 * value. Moreover, the sum of the util_avgs may be divergent, such
720 * as when the series is a harmonic series.
722 * To solve this problem, we also cap the util_avg of successive tasks to
723 * only 1/2 of the left utilization budget:
725 * util_avg_cap = (cpu_scale - cfs_rq->avg.util_avg) / 2^n
727 * where n denotes the nth task and cpu_scale the CPU capacity.
729 * For example, for a CPU with 1024 of capacity, a simplest series from
730 * the beginning would be like:
732 * task util_avg: 512, 256, 128, 64, 32, 16, 8, ...
733 * cfs_rq util_avg: 512, 768, 896, 960, 992, 1008, 1016, ...
735 * Finally, that extrapolated util_avg is clamped to the cap (util_avg_cap)
736 * if util_avg > util_avg_cap.
738 void post_init_entity_util_avg(struct task_struct *p)
740 struct sched_entity *se = &p->se;
741 struct cfs_rq *cfs_rq = cfs_rq_of(se);
742 struct sched_avg *sa = &se->avg;
743 long cpu_scale = arch_scale_cpu_capacity(cpu_of(rq_of(cfs_rq)));
744 long cap = (long)(cpu_scale - cfs_rq->avg.util_avg) / 2;
747 if (cfs_rq->avg.util_avg != 0) {
748 sa->util_avg = cfs_rq->avg.util_avg * se->load.weight;
749 sa->util_avg /= (cfs_rq->avg.load_avg + 1);
751 if (sa->util_avg > cap)
758 sa->runnable_avg = sa->util_avg;
760 if (p->sched_class != &fair_sched_class) {
762 * For !fair tasks do:
764 update_cfs_rq_load_avg(now, cfs_rq);
765 attach_entity_load_avg(cfs_rq, se);
766 switched_from_fair(rq, p);
768 * such that the next switched_to_fair() has the
771 se->avg.last_update_time = cfs_rq_clock_pelt(cfs_rq);
775 attach_entity_cfs_rq(se);
778 #else /* !CONFIG_SMP */
779 void init_entity_runnable_average(struct sched_entity *se)
782 void post_init_entity_util_avg(struct task_struct *p)
785 static void update_tg_load_avg(struct cfs_rq *cfs_rq)
788 #endif /* CONFIG_SMP */
791 * Update the current task's runtime statistics.
793 static void update_curr(struct cfs_rq *cfs_rq)
795 struct sched_entity *curr = cfs_rq->curr;
796 u64 now = rq_clock_task(rq_of(cfs_rq));
802 delta_exec = now - curr->exec_start;
803 if (unlikely((s64)delta_exec <= 0))
806 curr->exec_start = now;
808 schedstat_set(curr->statistics.exec_max,
809 max(delta_exec, curr->statistics.exec_max));
811 curr->sum_exec_runtime += delta_exec;
812 schedstat_add(cfs_rq->exec_clock, delta_exec);
814 curr->vruntime += calc_delta_fair(delta_exec, curr);
815 update_min_vruntime(cfs_rq);
817 if (entity_is_task(curr)) {
818 struct task_struct *curtask = task_of(curr);
820 trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime);
821 cgroup_account_cputime(curtask, delta_exec);
822 account_group_exec_runtime(curtask, delta_exec);
825 account_cfs_rq_runtime(cfs_rq, delta_exec);
828 static void update_curr_fair(struct rq *rq)
830 update_curr(cfs_rq_of(&rq->curr->se));
834 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
836 u64 wait_start, prev_wait_start;
838 if (!schedstat_enabled())
841 wait_start = rq_clock(rq_of(cfs_rq));
842 prev_wait_start = schedstat_val(se->statistics.wait_start);
844 if (entity_is_task(se) && task_on_rq_migrating(task_of(se)) &&
845 likely(wait_start > prev_wait_start))
846 wait_start -= prev_wait_start;
848 __schedstat_set(se->statistics.wait_start, wait_start);
852 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
854 struct task_struct *p;
857 if (!schedstat_enabled())
861 * When the sched_schedstat changes from 0 to 1, some sched se
862 * maybe already in the runqueue, the se->statistics.wait_start
863 * will be 0.So it will let the delta wrong. We need to avoid this
866 if (unlikely(!schedstat_val(se->statistics.wait_start)))
869 delta = rq_clock(rq_of(cfs_rq)) - schedstat_val(se->statistics.wait_start);
871 if (entity_is_task(se)) {
873 if (task_on_rq_migrating(p)) {
875 * Preserve migrating task's wait time so wait_start
876 * time stamp can be adjusted to accumulate wait time
877 * prior to migration.
879 __schedstat_set(se->statistics.wait_start, delta);
882 trace_sched_stat_wait(p, delta);
885 __schedstat_set(se->statistics.wait_max,
886 max(schedstat_val(se->statistics.wait_max), delta));
887 __schedstat_inc(se->statistics.wait_count);
888 __schedstat_add(se->statistics.wait_sum, delta);
889 __schedstat_set(se->statistics.wait_start, 0);
893 update_stats_enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
895 struct task_struct *tsk = NULL;
896 u64 sleep_start, block_start;
898 if (!schedstat_enabled())
901 sleep_start = schedstat_val(se->statistics.sleep_start);
902 block_start = schedstat_val(se->statistics.block_start);
904 if (entity_is_task(se))
908 u64 delta = rq_clock(rq_of(cfs_rq)) - sleep_start;
913 if (unlikely(delta > schedstat_val(se->statistics.sleep_max)))
914 __schedstat_set(se->statistics.sleep_max, delta);
916 __schedstat_set(se->statistics.sleep_start, 0);
917 __schedstat_add(se->statistics.sum_sleep_runtime, delta);
920 account_scheduler_latency(tsk, delta >> 10, 1);
921 trace_sched_stat_sleep(tsk, delta);
925 u64 delta = rq_clock(rq_of(cfs_rq)) - block_start;
930 if (unlikely(delta > schedstat_val(se->statistics.block_max)))
931 __schedstat_set(se->statistics.block_max, delta);
933 __schedstat_set(se->statistics.block_start, 0);
934 __schedstat_add(se->statistics.sum_sleep_runtime, delta);
937 if (tsk->in_iowait) {
938 __schedstat_add(se->statistics.iowait_sum, delta);
939 __schedstat_inc(se->statistics.iowait_count);
940 trace_sched_stat_iowait(tsk, delta);
943 trace_sched_stat_blocked(tsk, delta);
946 * Blocking time is in units of nanosecs, so shift by
947 * 20 to get a milliseconds-range estimation of the
948 * amount of time that the task spent sleeping:
950 if (unlikely(prof_on == SLEEP_PROFILING)) {
951 profile_hits(SLEEP_PROFILING,
952 (void *)get_wchan(tsk),
955 account_scheduler_latency(tsk, delta >> 10, 0);
961 * Task is being enqueued - update stats:
964 update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
966 if (!schedstat_enabled())
970 * Are we enqueueing a waiting task? (for current tasks
971 * a dequeue/enqueue event is a NOP)
973 if (se != cfs_rq->curr)
974 update_stats_wait_start(cfs_rq, se);
976 if (flags & ENQUEUE_WAKEUP)
977 update_stats_enqueue_sleeper(cfs_rq, se);
981 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
984 if (!schedstat_enabled())
988 * Mark the end of the wait period if dequeueing a
991 if (se != cfs_rq->curr)
992 update_stats_wait_end(cfs_rq, se);
994 if ((flags & DEQUEUE_SLEEP) && entity_is_task(se)) {
995 struct task_struct *tsk = task_of(se);
998 /* XXX racy against TTWU */
999 state = READ_ONCE(tsk->__state);
1000 if (state & TASK_INTERRUPTIBLE)
1001 __schedstat_set(se->statistics.sleep_start,
1002 rq_clock(rq_of(cfs_rq)));
1003 if (state & TASK_UNINTERRUPTIBLE)
1004 __schedstat_set(se->statistics.block_start,
1005 rq_clock(rq_of(cfs_rq)));
1010 * We are picking a new current task - update its stats:
1013 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
1016 * We are starting a new run period:
1018 se->exec_start = rq_clock_task(rq_of(cfs_rq));
1021 /**************************************************
1022 * Scheduling class queueing methods:
1025 #ifdef CONFIG_NUMA_BALANCING
1027 * Approximate time to scan a full NUMA task in ms. The task scan period is
1028 * calculated based on the tasks virtual memory size and
1029 * numa_balancing_scan_size.
1031 unsigned int sysctl_numa_balancing_scan_period_min = 1000;
1032 unsigned int sysctl_numa_balancing_scan_period_max = 60000;
1034 /* Portion of address space to scan in MB */
1035 unsigned int sysctl_numa_balancing_scan_size = 256;
1037 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
1038 unsigned int sysctl_numa_balancing_scan_delay = 1000;
1041 refcount_t refcount;
1043 spinlock_t lock; /* nr_tasks, tasks */
1048 struct rcu_head rcu;
1049 unsigned long total_faults;
1050 unsigned long max_faults_cpu;
1052 * Faults_cpu is used to decide whether memory should move
1053 * towards the CPU. As a consequence, these stats are weighted
1054 * more by CPU use than by memory faults.
1056 unsigned long *faults_cpu;
1057 unsigned long faults[];
1061 * For functions that can be called in multiple contexts that permit reading
1062 * ->numa_group (see struct task_struct for locking rules).
1064 static struct numa_group *deref_task_numa_group(struct task_struct *p)
1066 return rcu_dereference_check(p->numa_group, p == current ||
1067 (lockdep_is_held(__rq_lockp(task_rq(p))) && !READ_ONCE(p->on_cpu)));
1070 static struct numa_group *deref_curr_numa_group(struct task_struct *p)
1072 return rcu_dereference_protected(p->numa_group, p == current);
1075 static inline unsigned long group_faults_priv(struct numa_group *ng);
1076 static inline unsigned long group_faults_shared(struct numa_group *ng);
1078 static unsigned int task_nr_scan_windows(struct task_struct *p)
1080 unsigned long rss = 0;
1081 unsigned long nr_scan_pages;
1084 * Calculations based on RSS as non-present and empty pages are skipped
1085 * by the PTE scanner and NUMA hinting faults should be trapped based
1088 nr_scan_pages = sysctl_numa_balancing_scan_size << (20 - PAGE_SHIFT);
1089 rss = get_mm_rss(p->mm);
1091 rss = nr_scan_pages;
1093 rss = round_up(rss, nr_scan_pages);
1094 return rss / nr_scan_pages;
1097 /* For sanity's sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */
1098 #define MAX_SCAN_WINDOW 2560
1100 static unsigned int task_scan_min(struct task_struct *p)
1102 unsigned int scan_size = READ_ONCE(sysctl_numa_balancing_scan_size);
1103 unsigned int scan, floor;
1104 unsigned int windows = 1;
1106 if (scan_size < MAX_SCAN_WINDOW)
1107 windows = MAX_SCAN_WINDOW / scan_size;
1108 floor = 1000 / windows;
1110 scan = sysctl_numa_balancing_scan_period_min / task_nr_scan_windows(p);
1111 return max_t(unsigned int, floor, scan);
1114 static unsigned int task_scan_start(struct task_struct *p)
1116 unsigned long smin = task_scan_min(p);
1117 unsigned long period = smin;
1118 struct numa_group *ng;
1120 /* Scale the maximum scan period with the amount of shared memory. */
1122 ng = rcu_dereference(p->numa_group);
1124 unsigned long shared = group_faults_shared(ng);
1125 unsigned long private = group_faults_priv(ng);
1127 period *= refcount_read(&ng->refcount);
1128 period *= shared + 1;
1129 period /= private + shared + 1;
1133 return max(smin, period);
1136 static unsigned int task_scan_max(struct task_struct *p)
1138 unsigned long smin = task_scan_min(p);
1140 struct numa_group *ng;
1142 /* Watch for min being lower than max due to floor calculations */
1143 smax = sysctl_numa_balancing_scan_period_max / task_nr_scan_windows(p);
1145 /* Scale the maximum scan period with the amount of shared memory. */
1146 ng = deref_curr_numa_group(p);
1148 unsigned long shared = group_faults_shared(ng);
1149 unsigned long private = group_faults_priv(ng);
1150 unsigned long period = smax;
1152 period *= refcount_read(&ng->refcount);
1153 period *= shared + 1;
1154 period /= private + shared + 1;
1156 smax = max(smax, period);
1159 return max(smin, smax);
1162 static void account_numa_enqueue(struct rq *rq, struct task_struct *p)
1164 rq->nr_numa_running += (p->numa_preferred_nid != NUMA_NO_NODE);
1165 rq->nr_preferred_running += (p->numa_preferred_nid == task_node(p));
1168 static void account_numa_dequeue(struct rq *rq, struct task_struct *p)
1170 rq->nr_numa_running -= (p->numa_preferred_nid != NUMA_NO_NODE);
1171 rq->nr_preferred_running -= (p->numa_preferred_nid == task_node(p));
1174 /* Shared or private faults. */
1175 #define NR_NUMA_HINT_FAULT_TYPES 2
1177 /* Memory and CPU locality */
1178 #define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2)
1180 /* Averaged statistics, and temporary buffers. */
1181 #define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2)
1183 pid_t task_numa_group_id(struct task_struct *p)
1185 struct numa_group *ng;
1189 ng = rcu_dereference(p->numa_group);
1198 * The averaged statistics, shared & private, memory & CPU,
1199 * occupy the first half of the array. The second half of the
1200 * array is for current counters, which are averaged into the
1201 * first set by task_numa_placement.
1203 static inline int task_faults_idx(enum numa_faults_stats s, int nid, int priv)
1205 return NR_NUMA_HINT_FAULT_TYPES * (s * nr_node_ids + nid) + priv;
1208 static inline unsigned long task_faults(struct task_struct *p, int nid)
1210 if (!p->numa_faults)
1213 return p->numa_faults[task_faults_idx(NUMA_MEM, nid, 0)] +
1214 p->numa_faults[task_faults_idx(NUMA_MEM, nid, 1)];
1217 static inline unsigned long group_faults(struct task_struct *p, int nid)
1219 struct numa_group *ng = deref_task_numa_group(p);
1224 return ng->faults[task_faults_idx(NUMA_MEM, nid, 0)] +
1225 ng->faults[task_faults_idx(NUMA_MEM, nid, 1)];
1228 static inline unsigned long group_faults_cpu(struct numa_group *group, int nid)
1230 return group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 0)] +
1231 group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 1)];
1234 static inline unsigned long group_faults_priv(struct numa_group *ng)
1236 unsigned long faults = 0;
1239 for_each_online_node(node) {
1240 faults += ng->faults[task_faults_idx(NUMA_MEM, node, 1)];
1246 static inline unsigned long group_faults_shared(struct numa_group *ng)
1248 unsigned long faults = 0;
1251 for_each_online_node(node) {
1252 faults += ng->faults[task_faults_idx(NUMA_MEM, node, 0)];
1259 * A node triggering more than 1/3 as many NUMA faults as the maximum is
1260 * considered part of a numa group's pseudo-interleaving set. Migrations
1261 * between these nodes are slowed down, to allow things to settle down.
1263 #define ACTIVE_NODE_FRACTION 3
1265 static bool numa_is_active_node(int nid, struct numa_group *ng)
1267 return group_faults_cpu(ng, nid) * ACTIVE_NODE_FRACTION > ng->max_faults_cpu;
1270 /* Handle placement on systems where not all nodes are directly connected. */
1271 static unsigned long score_nearby_nodes(struct task_struct *p, int nid,
1272 int maxdist, bool task)
1274 unsigned long score = 0;
1278 * All nodes are directly connected, and the same distance
1279 * from each other. No need for fancy placement algorithms.
1281 if (sched_numa_topology_type == NUMA_DIRECT)
1285 * This code is called for each node, introducing N^2 complexity,
1286 * which should be ok given the number of nodes rarely exceeds 8.
1288 for_each_online_node(node) {
1289 unsigned long faults;
1290 int dist = node_distance(nid, node);
1293 * The furthest away nodes in the system are not interesting
1294 * for placement; nid was already counted.
1296 if (dist == sched_max_numa_distance || node == nid)
1300 * On systems with a backplane NUMA topology, compare groups
1301 * of nodes, and move tasks towards the group with the most
1302 * memory accesses. When comparing two nodes at distance
1303 * "hoplimit", only nodes closer by than "hoplimit" are part
1304 * of each group. Skip other nodes.
1306 if (sched_numa_topology_type == NUMA_BACKPLANE &&
1310 /* Add up the faults from nearby nodes. */
1312 faults = task_faults(p, node);
1314 faults = group_faults(p, node);
1317 * On systems with a glueless mesh NUMA topology, there are
1318 * no fixed "groups of nodes". Instead, nodes that are not
1319 * directly connected bounce traffic through intermediate
1320 * nodes; a numa_group can occupy any set of nodes.
1321 * The further away a node is, the less the faults count.
1322 * This seems to result in good task placement.
1324 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
1325 faults *= (sched_max_numa_distance - dist);
1326 faults /= (sched_max_numa_distance - LOCAL_DISTANCE);
1336 * These return the fraction of accesses done by a particular task, or
1337 * task group, on a particular numa node. The group weight is given a
1338 * larger multiplier, in order to group tasks together that are almost
1339 * evenly spread out between numa nodes.
1341 static inline unsigned long task_weight(struct task_struct *p, int nid,
1344 unsigned long faults, total_faults;
1346 if (!p->numa_faults)
1349 total_faults = p->total_numa_faults;
1354 faults = task_faults(p, nid);
1355 faults += score_nearby_nodes(p, nid, dist, true);
1357 return 1000 * faults / total_faults;
1360 static inline unsigned long group_weight(struct task_struct *p, int nid,
1363 struct numa_group *ng = deref_task_numa_group(p);
1364 unsigned long faults, total_faults;
1369 total_faults = ng->total_faults;
1374 faults = group_faults(p, nid);
1375 faults += score_nearby_nodes(p, nid, dist, false);
1377 return 1000 * faults / total_faults;
1380 bool should_numa_migrate_memory(struct task_struct *p, struct page * page,
1381 int src_nid, int dst_cpu)
1383 struct numa_group *ng = deref_curr_numa_group(p);
1384 int dst_nid = cpu_to_node(dst_cpu);
1385 int last_cpupid, this_cpupid;
1387 this_cpupid = cpu_pid_to_cpupid(dst_cpu, current->pid);
1388 last_cpupid = page_cpupid_xchg_last(page, this_cpupid);
1391 * Allow first faults or private faults to migrate immediately early in
1392 * the lifetime of a task. The magic number 4 is based on waiting for
1393 * two full passes of the "multi-stage node selection" test that is
1396 if ((p->numa_preferred_nid == NUMA_NO_NODE || p->numa_scan_seq <= 4) &&
1397 (cpupid_pid_unset(last_cpupid) || cpupid_match_pid(p, last_cpupid)))
1401 * Multi-stage node selection is used in conjunction with a periodic
1402 * migration fault to build a temporal task<->page relation. By using
1403 * a two-stage filter we remove short/unlikely relations.
1405 * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate
1406 * a task's usage of a particular page (n_p) per total usage of this
1407 * page (n_t) (in a given time-span) to a probability.
1409 * Our periodic faults will sample this probability and getting the
1410 * same result twice in a row, given these samples are fully
1411 * independent, is then given by P(n)^2, provided our sample period
1412 * is sufficiently short compared to the usage pattern.
1414 * This quadric squishes small probabilities, making it less likely we
1415 * act on an unlikely task<->page relation.
1417 if (!cpupid_pid_unset(last_cpupid) &&
1418 cpupid_to_nid(last_cpupid) != dst_nid)
1421 /* Always allow migrate on private faults */
1422 if (cpupid_match_pid(p, last_cpupid))
1425 /* A shared fault, but p->numa_group has not been set up yet. */
1430 * Destination node is much more heavily used than the source
1431 * node? Allow migration.
1433 if (group_faults_cpu(ng, dst_nid) > group_faults_cpu(ng, src_nid) *
1434 ACTIVE_NODE_FRACTION)
1438 * Distribute memory according to CPU & memory use on each node,
1439 * with 3/4 hysteresis to avoid unnecessary memory migrations:
1441 * faults_cpu(dst) 3 faults_cpu(src)
1442 * --------------- * - > ---------------
1443 * faults_mem(dst) 4 faults_mem(src)
1445 return group_faults_cpu(ng, dst_nid) * group_faults(p, src_nid) * 3 >
1446 group_faults_cpu(ng, src_nid) * group_faults(p, dst_nid) * 4;
1450 * 'numa_type' describes the node at the moment of load balancing.
1453 /* The node has spare capacity that can be used to run more tasks. */
1456 * The node is fully used and the tasks don't compete for more CPU
1457 * cycles. Nevertheless, some tasks might wait before running.
1461 * The node is overloaded and can't provide expected CPU cycles to all
1467 /* Cached statistics for all CPUs within a node */
1470 unsigned long runnable;
1472 /* Total compute capacity of CPUs on a node */
1473 unsigned long compute_capacity;
1474 unsigned int nr_running;
1475 unsigned int weight;
1476 enum numa_type node_type;
1480 static inline bool is_core_idle(int cpu)
1482 #ifdef CONFIG_SCHED_SMT
1485 for_each_cpu(sibling, cpu_smt_mask(cpu)) {
1497 struct task_numa_env {
1498 struct task_struct *p;
1500 int src_cpu, src_nid;
1501 int dst_cpu, dst_nid;
1503 struct numa_stats src_stats, dst_stats;
1508 struct task_struct *best_task;
1513 static unsigned long cpu_load(struct rq *rq);
1514 static unsigned long cpu_runnable(struct rq *rq);
1515 static unsigned long cpu_util(int cpu);
1516 static inline long adjust_numa_imbalance(int imbalance,
1517 int dst_running, int dst_weight);
1520 numa_type numa_classify(unsigned int imbalance_pct,
1521 struct numa_stats *ns)
1523 if ((ns->nr_running > ns->weight) &&
1524 (((ns->compute_capacity * 100) < (ns->util * imbalance_pct)) ||
1525 ((ns->compute_capacity * imbalance_pct) < (ns->runnable * 100))))
1526 return node_overloaded;
1528 if ((ns->nr_running < ns->weight) ||
1529 (((ns->compute_capacity * 100) > (ns->util * imbalance_pct)) &&
1530 ((ns->compute_capacity * imbalance_pct) > (ns->runnable * 100))))
1531 return node_has_spare;
1533 return node_fully_busy;
1536 #ifdef CONFIG_SCHED_SMT
1537 /* Forward declarations of select_idle_sibling helpers */
1538 static inline bool test_idle_cores(int cpu, bool def);
1539 static inline int numa_idle_core(int idle_core, int cpu)
1541 if (!static_branch_likely(&sched_smt_present) ||
1542 idle_core >= 0 || !test_idle_cores(cpu, false))
1546 * Prefer cores instead of packing HT siblings
1547 * and triggering future load balancing.
1549 if (is_core_idle(cpu))
1555 static inline int numa_idle_core(int idle_core, int cpu)
1562 * Gather all necessary information to make NUMA balancing placement
1563 * decisions that are compatible with standard load balancer. This
1564 * borrows code and logic from update_sg_lb_stats but sharing a
1565 * common implementation is impractical.
1567 static void update_numa_stats(struct task_numa_env *env,
1568 struct numa_stats *ns, int nid,
1571 int cpu, idle_core = -1;
1573 memset(ns, 0, sizeof(*ns));
1577 for_each_cpu(cpu, cpumask_of_node(nid)) {
1578 struct rq *rq = cpu_rq(cpu);
1580 ns->load += cpu_load(rq);
1581 ns->runnable += cpu_runnable(rq);
1582 ns->util += cpu_util(cpu);
1583 ns->nr_running += rq->cfs.h_nr_running;
1584 ns->compute_capacity += capacity_of(cpu);
1586 if (find_idle && !rq->nr_running && idle_cpu(cpu)) {
1587 if (READ_ONCE(rq->numa_migrate_on) ||
1588 !cpumask_test_cpu(cpu, env->p->cpus_ptr))
1591 if (ns->idle_cpu == -1)
1594 idle_core = numa_idle_core(idle_core, cpu);
1599 ns->weight = cpumask_weight(cpumask_of_node(nid));
1601 ns->node_type = numa_classify(env->imbalance_pct, ns);
1604 ns->idle_cpu = idle_core;
1607 static void task_numa_assign(struct task_numa_env *env,
1608 struct task_struct *p, long imp)
1610 struct rq *rq = cpu_rq(env->dst_cpu);
1612 /* Check if run-queue part of active NUMA balance. */
1613 if (env->best_cpu != env->dst_cpu && xchg(&rq->numa_migrate_on, 1)) {
1615 int start = env->dst_cpu;
1617 /* Find alternative idle CPU. */
1618 for_each_cpu_wrap(cpu, cpumask_of_node(env->dst_nid), start) {
1619 if (cpu == env->best_cpu || !idle_cpu(cpu) ||
1620 !cpumask_test_cpu(cpu, env->p->cpus_ptr)) {
1625 rq = cpu_rq(env->dst_cpu);
1626 if (!xchg(&rq->numa_migrate_on, 1))
1630 /* Failed to find an alternative idle CPU */
1636 * Clear previous best_cpu/rq numa-migrate flag, since task now
1637 * found a better CPU to move/swap.
1639 if (env->best_cpu != -1 && env->best_cpu != env->dst_cpu) {
1640 rq = cpu_rq(env->best_cpu);
1641 WRITE_ONCE(rq->numa_migrate_on, 0);
1645 put_task_struct(env->best_task);
1650 env->best_imp = imp;
1651 env->best_cpu = env->dst_cpu;
1654 static bool load_too_imbalanced(long src_load, long dst_load,
1655 struct task_numa_env *env)
1658 long orig_src_load, orig_dst_load;
1659 long src_capacity, dst_capacity;
1662 * The load is corrected for the CPU capacity available on each node.
1665 * ------------ vs ---------
1666 * src_capacity dst_capacity
1668 src_capacity = env->src_stats.compute_capacity;
1669 dst_capacity = env->dst_stats.compute_capacity;
1671 imb = abs(dst_load * src_capacity - src_load * dst_capacity);
1673 orig_src_load = env->src_stats.load;
1674 orig_dst_load = env->dst_stats.load;
1676 old_imb = abs(orig_dst_load * src_capacity - orig_src_load * dst_capacity);
1678 /* Would this change make things worse? */
1679 return (imb > old_imb);
1683 * Maximum NUMA importance can be 1998 (2*999);
1684 * SMALLIMP @ 30 would be close to 1998/64.
1685 * Used to deter task migration.
1690 * This checks if the overall compute and NUMA accesses of the system would
1691 * be improved if the source tasks was migrated to the target dst_cpu taking
1692 * into account that it might be best if task running on the dst_cpu should
1693 * be exchanged with the source task
1695 static bool task_numa_compare(struct task_numa_env *env,
1696 long taskimp, long groupimp, bool maymove)
1698 struct numa_group *cur_ng, *p_ng = deref_curr_numa_group(env->p);
1699 struct rq *dst_rq = cpu_rq(env->dst_cpu);
1700 long imp = p_ng ? groupimp : taskimp;
1701 struct task_struct *cur;
1702 long src_load, dst_load;
1703 int dist = env->dist;
1706 bool stopsearch = false;
1708 if (READ_ONCE(dst_rq->numa_migrate_on))
1712 cur = rcu_dereference(dst_rq->curr);
1713 if (cur && ((cur->flags & PF_EXITING) || is_idle_task(cur)))
1717 * Because we have preemption enabled we can get migrated around and
1718 * end try selecting ourselves (current == env->p) as a swap candidate.
1720 if (cur == env->p) {
1726 if (maymove && moveimp >= env->best_imp)
1732 /* Skip this swap candidate if cannot move to the source cpu. */
1733 if (!cpumask_test_cpu(env->src_cpu, cur->cpus_ptr))
1737 * Skip this swap candidate if it is not moving to its preferred
1738 * node and the best task is.
1740 if (env->best_task &&
1741 env->best_task->numa_preferred_nid == env->src_nid &&
1742 cur->numa_preferred_nid != env->src_nid) {
1747 * "imp" is the fault differential for the source task between the
1748 * source and destination node. Calculate the total differential for
1749 * the source task and potential destination task. The more negative
1750 * the value is, the more remote accesses that would be expected to
1751 * be incurred if the tasks were swapped.
1753 * If dst and source tasks are in the same NUMA group, or not
1754 * in any group then look only at task weights.
1756 cur_ng = rcu_dereference(cur->numa_group);
1757 if (cur_ng == p_ng) {
1758 imp = taskimp + task_weight(cur, env->src_nid, dist) -
1759 task_weight(cur, env->dst_nid, dist);
1761 * Add some hysteresis to prevent swapping the
1762 * tasks within a group over tiny differences.
1768 * Compare the group weights. If a task is all by itself
1769 * (not part of a group), use the task weight instead.
1772 imp += group_weight(cur, env->src_nid, dist) -
1773 group_weight(cur, env->dst_nid, dist);
1775 imp += task_weight(cur, env->src_nid, dist) -
1776 task_weight(cur, env->dst_nid, dist);
1779 /* Discourage picking a task already on its preferred node */
1780 if (cur->numa_preferred_nid == env->dst_nid)
1784 * Encourage picking a task that moves to its preferred node.
1785 * This potentially makes imp larger than it's maximum of
1786 * 1998 (see SMALLIMP and task_weight for why) but in this
1787 * case, it does not matter.
1789 if (cur->numa_preferred_nid == env->src_nid)
1792 if (maymove && moveimp > imp && moveimp > env->best_imp) {
1799 * Prefer swapping with a task moving to its preferred node over a
1802 if (env->best_task && cur->numa_preferred_nid == env->src_nid &&
1803 env->best_task->numa_preferred_nid != env->src_nid) {
1808 * If the NUMA importance is less than SMALLIMP,
1809 * task migration might only result in ping pong
1810 * of tasks and also hurt performance due to cache
1813 if (imp < SMALLIMP || imp <= env->best_imp + SMALLIMP / 2)
1817 * In the overloaded case, try and keep the load balanced.
1819 load = task_h_load(env->p) - task_h_load(cur);
1823 dst_load = env->dst_stats.load + load;
1824 src_load = env->src_stats.load - load;
1826 if (load_too_imbalanced(src_load, dst_load, env))
1830 /* Evaluate an idle CPU for a task numa move. */
1832 int cpu = env->dst_stats.idle_cpu;
1834 /* Nothing cached so current CPU went idle since the search. */
1839 * If the CPU is no longer truly idle and the previous best CPU
1840 * is, keep using it.
1842 if (!idle_cpu(cpu) && env->best_cpu >= 0 &&
1843 idle_cpu(env->best_cpu)) {
1844 cpu = env->best_cpu;
1850 task_numa_assign(env, cur, imp);
1853 * If a move to idle is allowed because there is capacity or load
1854 * balance improves then stop the search. While a better swap
1855 * candidate may exist, a search is not free.
1857 if (maymove && !cur && env->best_cpu >= 0 && idle_cpu(env->best_cpu))
1861 * If a swap candidate must be identified and the current best task
1862 * moves its preferred node then stop the search.
1864 if (!maymove && env->best_task &&
1865 env->best_task->numa_preferred_nid == env->src_nid) {
1874 static void task_numa_find_cpu(struct task_numa_env *env,
1875 long taskimp, long groupimp)
1877 bool maymove = false;
1881 * If dst node has spare capacity, then check if there is an
1882 * imbalance that would be overruled by the load balancer.
1884 if (env->dst_stats.node_type == node_has_spare) {
1885 unsigned int imbalance;
1886 int src_running, dst_running;
1889 * Would movement cause an imbalance? Note that if src has
1890 * more running tasks that the imbalance is ignored as the
1891 * move improves the imbalance from the perspective of the
1892 * CPU load balancer.
1894 src_running = env->src_stats.nr_running - 1;
1895 dst_running = env->dst_stats.nr_running + 1;
1896 imbalance = max(0, dst_running - src_running);
1897 imbalance = adjust_numa_imbalance(imbalance, dst_running,
1898 env->dst_stats.weight);
1900 /* Use idle CPU if there is no imbalance */
1903 if (env->dst_stats.idle_cpu >= 0) {
1904 env->dst_cpu = env->dst_stats.idle_cpu;
1905 task_numa_assign(env, NULL, 0);
1910 long src_load, dst_load, load;
1912 * If the improvement from just moving env->p direction is better
1913 * than swapping tasks around, check if a move is possible.
1915 load = task_h_load(env->p);
1916 dst_load = env->dst_stats.load + load;
1917 src_load = env->src_stats.load - load;
1918 maymove = !load_too_imbalanced(src_load, dst_load, env);
1921 for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) {
1922 /* Skip this CPU if the source task cannot migrate */
1923 if (!cpumask_test_cpu(cpu, env->p->cpus_ptr))
1927 if (task_numa_compare(env, taskimp, groupimp, maymove))
1932 static int task_numa_migrate(struct task_struct *p)
1934 struct task_numa_env env = {
1937 .src_cpu = task_cpu(p),
1938 .src_nid = task_node(p),
1940 .imbalance_pct = 112,
1946 unsigned long taskweight, groupweight;
1947 struct sched_domain *sd;
1948 long taskimp, groupimp;
1949 struct numa_group *ng;
1954 * Pick the lowest SD_NUMA domain, as that would have the smallest
1955 * imbalance and would be the first to start moving tasks about.
1957 * And we want to avoid any moving of tasks about, as that would create
1958 * random movement of tasks -- counter the numa conditions we're trying
1962 sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu));
1964 env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2;
1968 * Cpusets can break the scheduler domain tree into smaller
1969 * balance domains, some of which do not cross NUMA boundaries.
1970 * Tasks that are "trapped" in such domains cannot be migrated
1971 * elsewhere, so there is no point in (re)trying.
1973 if (unlikely(!sd)) {
1974 sched_setnuma(p, task_node(p));
1978 env.dst_nid = p->numa_preferred_nid;
1979 dist = env.dist = node_distance(env.src_nid, env.dst_nid);
1980 taskweight = task_weight(p, env.src_nid, dist);
1981 groupweight = group_weight(p, env.src_nid, dist);
1982 update_numa_stats(&env, &env.src_stats, env.src_nid, false);
1983 taskimp = task_weight(p, env.dst_nid, dist) - taskweight;
1984 groupimp = group_weight(p, env.dst_nid, dist) - groupweight;
1985 update_numa_stats(&env, &env.dst_stats, env.dst_nid, true);
1987 /* Try to find a spot on the preferred nid. */
1988 task_numa_find_cpu(&env, taskimp, groupimp);
1991 * Look at other nodes in these cases:
1992 * - there is no space available on the preferred_nid
1993 * - the task is part of a numa_group that is interleaved across
1994 * multiple NUMA nodes; in order to better consolidate the group,
1995 * we need to check other locations.
1997 ng = deref_curr_numa_group(p);
1998 if (env.best_cpu == -1 || (ng && ng->active_nodes > 1)) {
1999 for_each_online_node(nid) {
2000 if (nid == env.src_nid || nid == p->numa_preferred_nid)
2003 dist = node_distance(env.src_nid, env.dst_nid);
2004 if (sched_numa_topology_type == NUMA_BACKPLANE &&
2006 taskweight = task_weight(p, env.src_nid, dist);
2007 groupweight = group_weight(p, env.src_nid, dist);
2010 /* Only consider nodes where both task and groups benefit */
2011 taskimp = task_weight(p, nid, dist) - taskweight;
2012 groupimp = group_weight(p, nid, dist) - groupweight;
2013 if (taskimp < 0 && groupimp < 0)
2018 update_numa_stats(&env, &env.dst_stats, env.dst_nid, true);
2019 task_numa_find_cpu(&env, taskimp, groupimp);
2024 * If the task is part of a workload that spans multiple NUMA nodes,
2025 * and is migrating into one of the workload's active nodes, remember
2026 * this node as the task's preferred numa node, so the workload can
2028 * A task that migrated to a second choice node will be better off
2029 * trying for a better one later. Do not set the preferred node here.
2032 if (env.best_cpu == -1)
2035 nid = cpu_to_node(env.best_cpu);
2037 if (nid != p->numa_preferred_nid)
2038 sched_setnuma(p, nid);
2041 /* No better CPU than the current one was found. */
2042 if (env.best_cpu == -1) {
2043 trace_sched_stick_numa(p, env.src_cpu, NULL, -1);
2047 best_rq = cpu_rq(env.best_cpu);
2048 if (env.best_task == NULL) {
2049 ret = migrate_task_to(p, env.best_cpu);
2050 WRITE_ONCE(best_rq->numa_migrate_on, 0);
2052 trace_sched_stick_numa(p, env.src_cpu, NULL, env.best_cpu);
2056 ret = migrate_swap(p, env.best_task, env.best_cpu, env.src_cpu);
2057 WRITE_ONCE(best_rq->numa_migrate_on, 0);
2060 trace_sched_stick_numa(p, env.src_cpu, env.best_task, env.best_cpu);
2061 put_task_struct(env.best_task);
2065 /* Attempt to migrate a task to a CPU on the preferred node. */
2066 static void numa_migrate_preferred(struct task_struct *p)
2068 unsigned long interval = HZ;
2070 /* This task has no NUMA fault statistics yet */
2071 if (unlikely(p->numa_preferred_nid == NUMA_NO_NODE || !p->numa_faults))
2074 /* Periodically retry migrating the task to the preferred node */
2075 interval = min(interval, msecs_to_jiffies(p->numa_scan_period) / 16);
2076 p->numa_migrate_retry = jiffies + interval;
2078 /* Success if task is already running on preferred CPU */
2079 if (task_node(p) == p->numa_preferred_nid)
2082 /* Otherwise, try migrate to a CPU on the preferred node */
2083 task_numa_migrate(p);
2087 * Find out how many nodes on the workload is actively running on. Do this by
2088 * tracking the nodes from which NUMA hinting faults are triggered. This can
2089 * be different from the set of nodes where the workload's memory is currently
2092 static void numa_group_count_active_nodes(struct numa_group *numa_group)
2094 unsigned long faults, max_faults = 0;
2095 int nid, active_nodes = 0;
2097 for_each_online_node(nid) {
2098 faults = group_faults_cpu(numa_group, nid);
2099 if (faults > max_faults)
2100 max_faults = faults;
2103 for_each_online_node(nid) {
2104 faults = group_faults_cpu(numa_group, nid);
2105 if (faults * ACTIVE_NODE_FRACTION > max_faults)
2109 numa_group->max_faults_cpu = max_faults;
2110 numa_group->active_nodes = active_nodes;
2114 * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
2115 * increments. The more local the fault statistics are, the higher the scan
2116 * period will be for the next scan window. If local/(local+remote) ratio is
2117 * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS)
2118 * the scan period will decrease. Aim for 70% local accesses.
2120 #define NUMA_PERIOD_SLOTS 10
2121 #define NUMA_PERIOD_THRESHOLD 7
2124 * Increase the scan period (slow down scanning) if the majority of
2125 * our memory is already on our local node, or if the majority of
2126 * the page accesses are shared with other processes.
2127 * Otherwise, decrease the scan period.
2129 static void update_task_scan_period(struct task_struct *p,
2130 unsigned long shared, unsigned long private)
2132 unsigned int period_slot;
2133 int lr_ratio, ps_ratio;
2136 unsigned long remote = p->numa_faults_locality[0];
2137 unsigned long local = p->numa_faults_locality[1];
2140 * If there were no record hinting faults then either the task is
2141 * completely idle or all activity is areas that are not of interest
2142 * to automatic numa balancing. Related to that, if there were failed
2143 * migration then it implies we are migrating too quickly or the local
2144 * node is overloaded. In either case, scan slower
2146 if (local + shared == 0 || p->numa_faults_locality[2]) {
2147 p->numa_scan_period = min(p->numa_scan_period_max,
2148 p->numa_scan_period << 1);
2150 p->mm->numa_next_scan = jiffies +
2151 msecs_to_jiffies(p->numa_scan_period);
2157 * Prepare to scale scan period relative to the current period.
2158 * == NUMA_PERIOD_THRESHOLD scan period stays the same
2159 * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
2160 * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
2162 period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS);
2163 lr_ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote);
2164 ps_ratio = (private * NUMA_PERIOD_SLOTS) / (private + shared);
2166 if (ps_ratio >= NUMA_PERIOD_THRESHOLD) {
2168 * Most memory accesses are local. There is no need to
2169 * do fast NUMA scanning, since memory is already local.
2171 int slot = ps_ratio - NUMA_PERIOD_THRESHOLD;
2174 diff = slot * period_slot;
2175 } else if (lr_ratio >= NUMA_PERIOD_THRESHOLD) {
2177 * Most memory accesses are shared with other tasks.
2178 * There is no point in continuing fast NUMA scanning,
2179 * since other tasks may just move the memory elsewhere.
2181 int slot = lr_ratio - NUMA_PERIOD_THRESHOLD;
2184 diff = slot * period_slot;
2187 * Private memory faults exceed (SLOTS-THRESHOLD)/SLOTS,
2188 * yet they are not on the local NUMA node. Speed up
2189 * NUMA scanning to get the memory moved over.
2191 int ratio = max(lr_ratio, ps_ratio);
2192 diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot;
2195 p->numa_scan_period = clamp(p->numa_scan_period + diff,
2196 task_scan_min(p), task_scan_max(p));
2197 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
2201 * Get the fraction of time the task has been running since the last
2202 * NUMA placement cycle. The scheduler keeps similar statistics, but
2203 * decays those on a 32ms period, which is orders of magnitude off
2204 * from the dozens-of-seconds NUMA balancing period. Use the scheduler
2205 * stats only if the task is so new there are no NUMA statistics yet.
2207 static u64 numa_get_avg_runtime(struct task_struct *p, u64 *period)
2209 u64 runtime, delta, now;
2210 /* Use the start of this time slice to avoid calculations. */
2211 now = p->se.exec_start;
2212 runtime = p->se.sum_exec_runtime;
2214 if (p->last_task_numa_placement) {
2215 delta = runtime - p->last_sum_exec_runtime;
2216 *period = now - p->last_task_numa_placement;
2218 /* Avoid time going backwards, prevent potential divide error: */
2219 if (unlikely((s64)*period < 0))
2222 delta = p->se.avg.load_sum;
2223 *period = LOAD_AVG_MAX;
2226 p->last_sum_exec_runtime = runtime;
2227 p->last_task_numa_placement = now;
2233 * Determine the preferred nid for a task in a numa_group. This needs to
2234 * be done in a way that produces consistent results with group_weight,
2235 * otherwise workloads might not converge.
2237 static int preferred_group_nid(struct task_struct *p, int nid)
2242 /* Direct connections between all NUMA nodes. */
2243 if (sched_numa_topology_type == NUMA_DIRECT)
2247 * On a system with glueless mesh NUMA topology, group_weight
2248 * scores nodes according to the number of NUMA hinting faults on
2249 * both the node itself, and on nearby nodes.
2251 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
2252 unsigned long score, max_score = 0;
2253 int node, max_node = nid;
2255 dist = sched_max_numa_distance;
2257 for_each_online_node(node) {
2258 score = group_weight(p, node, dist);
2259 if (score > max_score) {
2268 * Finding the preferred nid in a system with NUMA backplane
2269 * interconnect topology is more involved. The goal is to locate
2270 * tasks from numa_groups near each other in the system, and
2271 * untangle workloads from different sides of the system. This requires
2272 * searching down the hierarchy of node groups, recursively searching
2273 * inside the highest scoring group of nodes. The nodemask tricks
2274 * keep the complexity of the search down.
2276 nodes = node_online_map;
2277 for (dist = sched_max_numa_distance; dist > LOCAL_DISTANCE; dist--) {
2278 unsigned long max_faults = 0;
2279 nodemask_t max_group = NODE_MASK_NONE;
2282 /* Are there nodes at this distance from each other? */
2283 if (!find_numa_distance(dist))
2286 for_each_node_mask(a, nodes) {
2287 unsigned long faults = 0;
2288 nodemask_t this_group;
2289 nodes_clear(this_group);
2291 /* Sum group's NUMA faults; includes a==b case. */
2292 for_each_node_mask(b, nodes) {
2293 if (node_distance(a, b) < dist) {
2294 faults += group_faults(p, b);
2295 node_set(b, this_group);
2296 node_clear(b, nodes);
2300 /* Remember the top group. */
2301 if (faults > max_faults) {
2302 max_faults = faults;
2303 max_group = this_group;
2305 * subtle: at the smallest distance there is
2306 * just one node left in each "group", the
2307 * winner is the preferred nid.
2312 /* Next round, evaluate the nodes within max_group. */
2320 static void task_numa_placement(struct task_struct *p)
2322 int seq, nid, max_nid = NUMA_NO_NODE;
2323 unsigned long max_faults = 0;
2324 unsigned long fault_types[2] = { 0, 0 };
2325 unsigned long total_faults;
2326 u64 runtime, period;
2327 spinlock_t *group_lock = NULL;
2328 struct numa_group *ng;
2331 * The p->mm->numa_scan_seq field gets updated without
2332 * exclusive access. Use READ_ONCE() here to ensure
2333 * that the field is read in a single access:
2335 seq = READ_ONCE(p->mm->numa_scan_seq);
2336 if (p->numa_scan_seq == seq)
2338 p->numa_scan_seq = seq;
2339 p->numa_scan_period_max = task_scan_max(p);
2341 total_faults = p->numa_faults_locality[0] +
2342 p->numa_faults_locality[1];
2343 runtime = numa_get_avg_runtime(p, &period);
2345 /* If the task is part of a group prevent parallel updates to group stats */
2346 ng = deref_curr_numa_group(p);
2348 group_lock = &ng->lock;
2349 spin_lock_irq(group_lock);
2352 /* Find the node with the highest number of faults */
2353 for_each_online_node(nid) {
2354 /* Keep track of the offsets in numa_faults array */
2355 int mem_idx, membuf_idx, cpu_idx, cpubuf_idx;
2356 unsigned long faults = 0, group_faults = 0;
2359 for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) {
2360 long diff, f_diff, f_weight;
2362 mem_idx = task_faults_idx(NUMA_MEM, nid, priv);
2363 membuf_idx = task_faults_idx(NUMA_MEMBUF, nid, priv);
2364 cpu_idx = task_faults_idx(NUMA_CPU, nid, priv);
2365 cpubuf_idx = task_faults_idx(NUMA_CPUBUF, nid, priv);
2367 /* Decay existing window, copy faults since last scan */
2368 diff = p->numa_faults[membuf_idx] - p->numa_faults[mem_idx] / 2;
2369 fault_types[priv] += p->numa_faults[membuf_idx];
2370 p->numa_faults[membuf_idx] = 0;
2373 * Normalize the faults_from, so all tasks in a group
2374 * count according to CPU use, instead of by the raw
2375 * number of faults. Tasks with little runtime have
2376 * little over-all impact on throughput, and thus their
2377 * faults are less important.
2379 f_weight = div64_u64(runtime << 16, period + 1);
2380 f_weight = (f_weight * p->numa_faults[cpubuf_idx]) /
2382 f_diff = f_weight - p->numa_faults[cpu_idx] / 2;
2383 p->numa_faults[cpubuf_idx] = 0;
2385 p->numa_faults[mem_idx] += diff;
2386 p->numa_faults[cpu_idx] += f_diff;
2387 faults += p->numa_faults[mem_idx];
2388 p->total_numa_faults += diff;
2391 * safe because we can only change our own group
2393 * mem_idx represents the offset for a given
2394 * nid and priv in a specific region because it
2395 * is at the beginning of the numa_faults array.
2397 ng->faults[mem_idx] += diff;
2398 ng->faults_cpu[mem_idx] += f_diff;
2399 ng->total_faults += diff;
2400 group_faults += ng->faults[mem_idx];
2405 if (faults > max_faults) {
2406 max_faults = faults;
2409 } else if (group_faults > max_faults) {
2410 max_faults = group_faults;
2416 numa_group_count_active_nodes(ng);
2417 spin_unlock_irq(group_lock);
2418 max_nid = preferred_group_nid(p, max_nid);
2422 /* Set the new preferred node */
2423 if (max_nid != p->numa_preferred_nid)
2424 sched_setnuma(p, max_nid);
2427 update_task_scan_period(p, fault_types[0], fault_types[1]);
2430 static inline int get_numa_group(struct numa_group *grp)
2432 return refcount_inc_not_zero(&grp->refcount);
2435 static inline void put_numa_group(struct numa_group *grp)
2437 if (refcount_dec_and_test(&grp->refcount))
2438 kfree_rcu(grp, rcu);
2441 static void task_numa_group(struct task_struct *p, int cpupid, int flags,
2444 struct numa_group *grp, *my_grp;
2445 struct task_struct *tsk;
2447 int cpu = cpupid_to_cpu(cpupid);
2450 if (unlikely(!deref_curr_numa_group(p))) {
2451 unsigned int size = sizeof(struct numa_group) +
2452 4*nr_node_ids*sizeof(unsigned long);
2454 grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
2458 refcount_set(&grp->refcount, 1);
2459 grp->active_nodes = 1;
2460 grp->max_faults_cpu = 0;
2461 spin_lock_init(&grp->lock);
2463 /* Second half of the array tracks nids where faults happen */
2464 grp->faults_cpu = grp->faults + NR_NUMA_HINT_FAULT_TYPES *
2467 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2468 grp->faults[i] = p->numa_faults[i];
2470 grp->total_faults = p->total_numa_faults;
2473 rcu_assign_pointer(p->numa_group, grp);
2477 tsk = READ_ONCE(cpu_rq(cpu)->curr);
2479 if (!cpupid_match_pid(tsk, cpupid))
2482 grp = rcu_dereference(tsk->numa_group);
2486 my_grp = deref_curr_numa_group(p);
2491 * Only join the other group if its bigger; if we're the bigger group,
2492 * the other task will join us.
2494 if (my_grp->nr_tasks > grp->nr_tasks)
2498 * Tie-break on the grp address.
2500 if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp)
2503 /* Always join threads in the same process. */
2504 if (tsk->mm == current->mm)
2507 /* Simple filter to avoid false positives due to PID collisions */
2508 if (flags & TNF_SHARED)
2511 /* Update priv based on whether false sharing was detected */
2514 if (join && !get_numa_group(grp))
2522 BUG_ON(irqs_disabled());
2523 double_lock_irq(&my_grp->lock, &grp->lock);
2525 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) {
2526 my_grp->faults[i] -= p->numa_faults[i];
2527 grp->faults[i] += p->numa_faults[i];
2529 my_grp->total_faults -= p->total_numa_faults;
2530 grp->total_faults += p->total_numa_faults;
2535 spin_unlock(&my_grp->lock);
2536 spin_unlock_irq(&grp->lock);
2538 rcu_assign_pointer(p->numa_group, grp);
2540 put_numa_group(my_grp);
2549 * Get rid of NUMA statistics associated with a task (either current or dead).
2550 * If @final is set, the task is dead and has reached refcount zero, so we can
2551 * safely free all relevant data structures. Otherwise, there might be
2552 * concurrent reads from places like load balancing and procfs, and we should
2553 * reset the data back to default state without freeing ->numa_faults.
2555 void task_numa_free(struct task_struct *p, bool final)
2557 /* safe: p either is current or is being freed by current */
2558 struct numa_group *grp = rcu_dereference_raw(p->numa_group);
2559 unsigned long *numa_faults = p->numa_faults;
2560 unsigned long flags;
2567 spin_lock_irqsave(&grp->lock, flags);
2568 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2569 grp->faults[i] -= p->numa_faults[i];
2570 grp->total_faults -= p->total_numa_faults;
2573 spin_unlock_irqrestore(&grp->lock, flags);
2574 RCU_INIT_POINTER(p->numa_group, NULL);
2575 put_numa_group(grp);
2579 p->numa_faults = NULL;
2582 p->total_numa_faults = 0;
2583 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2589 * Got a PROT_NONE fault for a page on @node.
2591 void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags)
2593 struct task_struct *p = current;
2594 bool migrated = flags & TNF_MIGRATED;
2595 int cpu_node = task_node(current);
2596 int local = !!(flags & TNF_FAULT_LOCAL);
2597 struct numa_group *ng;
2600 if (!static_branch_likely(&sched_numa_balancing))
2603 /* for example, ksmd faulting in a user's mm */
2607 /* Allocate buffer to track faults on a per-node basis */
2608 if (unlikely(!p->numa_faults)) {
2609 int size = sizeof(*p->numa_faults) *
2610 NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids;
2612 p->numa_faults = kzalloc(size, GFP_KERNEL|__GFP_NOWARN);
2613 if (!p->numa_faults)
2616 p->total_numa_faults = 0;
2617 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
2621 * First accesses are treated as private, otherwise consider accesses
2622 * to be private if the accessing pid has not changed
2624 if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) {
2627 priv = cpupid_match_pid(p, last_cpupid);
2628 if (!priv && !(flags & TNF_NO_GROUP))
2629 task_numa_group(p, last_cpupid, flags, &priv);
2633 * If a workload spans multiple NUMA nodes, a shared fault that
2634 * occurs wholly within the set of nodes that the workload is
2635 * actively using should be counted as local. This allows the
2636 * scan rate to slow down when a workload has settled down.
2638 ng = deref_curr_numa_group(p);
2639 if (!priv && !local && ng && ng->active_nodes > 1 &&
2640 numa_is_active_node(cpu_node, ng) &&
2641 numa_is_active_node(mem_node, ng))
2645 * Retry to migrate task to preferred node periodically, in case it
2646 * previously failed, or the scheduler moved us.
2648 if (time_after(jiffies, p->numa_migrate_retry)) {
2649 task_numa_placement(p);
2650 numa_migrate_preferred(p);
2654 p->numa_pages_migrated += pages;
2655 if (flags & TNF_MIGRATE_FAIL)
2656 p->numa_faults_locality[2] += pages;
2658 p->numa_faults[task_faults_idx(NUMA_MEMBUF, mem_node, priv)] += pages;
2659 p->numa_faults[task_faults_idx(NUMA_CPUBUF, cpu_node, priv)] += pages;
2660 p->numa_faults_locality[local] += pages;
2663 static void reset_ptenuma_scan(struct task_struct *p)
2666 * We only did a read acquisition of the mmap sem, so
2667 * p->mm->numa_scan_seq is written to without exclusive access
2668 * and the update is not guaranteed to be atomic. That's not
2669 * much of an issue though, since this is just used for
2670 * statistical sampling. Use READ_ONCE/WRITE_ONCE, which are not
2671 * expensive, to avoid any form of compiler optimizations:
2673 WRITE_ONCE(p->mm->numa_scan_seq, READ_ONCE(p->mm->numa_scan_seq) + 1);
2674 p->mm->numa_scan_offset = 0;
2678 * The expensive part of numa migration is done from task_work context.
2679 * Triggered from task_tick_numa().
2681 static void task_numa_work(struct callback_head *work)
2683 unsigned long migrate, next_scan, now = jiffies;
2684 struct task_struct *p = current;
2685 struct mm_struct *mm = p->mm;
2686 u64 runtime = p->se.sum_exec_runtime;
2687 struct vm_area_struct *vma;
2688 unsigned long start, end;
2689 unsigned long nr_pte_updates = 0;
2690 long pages, virtpages;
2692 SCHED_WARN_ON(p != container_of(work, struct task_struct, numa_work));
2696 * Who cares about NUMA placement when they're dying.
2698 * NOTE: make sure not to dereference p->mm before this check,
2699 * exit_task_work() happens _after_ exit_mm() so we could be called
2700 * without p->mm even though we still had it when we enqueued this
2703 if (p->flags & PF_EXITING)
2706 if (!mm->numa_next_scan) {
2707 mm->numa_next_scan = now +
2708 msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2712 * Enforce maximal scan/migration frequency..
2714 migrate = mm->numa_next_scan;
2715 if (time_before(now, migrate))
2718 if (p->numa_scan_period == 0) {
2719 p->numa_scan_period_max = task_scan_max(p);
2720 p->numa_scan_period = task_scan_start(p);
2723 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
2724 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
2728 * Delay this task enough that another task of this mm will likely win
2729 * the next time around.
2731 p->node_stamp += 2 * TICK_NSEC;
2733 start = mm->numa_scan_offset;
2734 pages = sysctl_numa_balancing_scan_size;
2735 pages <<= 20 - PAGE_SHIFT; /* MB in pages */
2736 virtpages = pages * 8; /* Scan up to this much virtual space */
2741 if (!mmap_read_trylock(mm))
2743 vma = find_vma(mm, start);
2745 reset_ptenuma_scan(p);
2749 for (; vma; vma = vma->vm_next) {
2750 if (!vma_migratable(vma) || !vma_policy_mof(vma) ||
2751 is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_MIXEDMAP)) {
2756 * Shared library pages mapped by multiple processes are not
2757 * migrated as it is expected they are cache replicated. Avoid
2758 * hinting faults in read-only file-backed mappings or the vdso
2759 * as migrating the pages will be of marginal benefit.
2762 (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ)))
2766 * Skip inaccessible VMAs to avoid any confusion between
2767 * PROT_NONE and NUMA hinting ptes
2769 if (!vma_is_accessible(vma))
2773 start = max(start, vma->vm_start);
2774 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
2775 end = min(end, vma->vm_end);
2776 nr_pte_updates = change_prot_numa(vma, start, end);
2779 * Try to scan sysctl_numa_balancing_size worth of
2780 * hpages that have at least one present PTE that
2781 * is not already pte-numa. If the VMA contains
2782 * areas that are unused or already full of prot_numa
2783 * PTEs, scan up to virtpages, to skip through those
2787 pages -= (end - start) >> PAGE_SHIFT;
2788 virtpages -= (end - start) >> PAGE_SHIFT;
2791 if (pages <= 0 || virtpages <= 0)
2795 } while (end != vma->vm_end);
2800 * It is possible to reach the end of the VMA list but the last few
2801 * VMAs are not guaranteed to the vma_migratable. If they are not, we
2802 * would find the !migratable VMA on the next scan but not reset the
2803 * scanner to the start so check it now.
2806 mm->numa_scan_offset = start;
2808 reset_ptenuma_scan(p);
2809 mmap_read_unlock(mm);
2812 * Make sure tasks use at least 32x as much time to run other code
2813 * than they used here, to limit NUMA PTE scanning overhead to 3% max.
2814 * Usually update_task_scan_period slows down scanning enough; on an
2815 * overloaded system we need to limit overhead on a per task basis.
2817 if (unlikely(p->se.sum_exec_runtime != runtime)) {
2818 u64 diff = p->se.sum_exec_runtime - runtime;
2819 p->node_stamp += 32 * diff;
2823 void init_numa_balancing(unsigned long clone_flags, struct task_struct *p)
2826 struct mm_struct *mm = p->mm;
2829 mm_users = atomic_read(&mm->mm_users);
2830 if (mm_users == 1) {
2831 mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2832 mm->numa_scan_seq = 0;
2836 p->numa_scan_seq = mm ? mm->numa_scan_seq : 0;
2837 p->numa_scan_period = sysctl_numa_balancing_scan_delay;
2838 /* Protect against double add, see task_tick_numa and task_numa_work */
2839 p->numa_work.next = &p->numa_work;
2840 p->numa_faults = NULL;
2841 RCU_INIT_POINTER(p->numa_group, NULL);
2842 p->last_task_numa_placement = 0;
2843 p->last_sum_exec_runtime = 0;
2845 init_task_work(&p->numa_work, task_numa_work);
2847 /* New address space, reset the preferred nid */
2848 if (!(clone_flags & CLONE_VM)) {
2849 p->numa_preferred_nid = NUMA_NO_NODE;
2854 * New thread, keep existing numa_preferred_nid which should be copied
2855 * already by arch_dup_task_struct but stagger when scans start.
2860 delay = min_t(unsigned int, task_scan_max(current),
2861 current->numa_scan_period * mm_users * NSEC_PER_MSEC);
2862 delay += 2 * TICK_NSEC;
2863 p->node_stamp = delay;
2868 * Drive the periodic memory faults..
2870 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
2872 struct callback_head *work = &curr->numa_work;
2876 * We don't care about NUMA placement if we don't have memory.
2878 if ((curr->flags & (PF_EXITING | PF_KTHREAD)) || work->next != work)
2882 * Using runtime rather than walltime has the dual advantage that
2883 * we (mostly) drive the selection from busy threads and that the
2884 * task needs to have done some actual work before we bother with
2887 now = curr->se.sum_exec_runtime;
2888 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
2890 if (now > curr->node_stamp + period) {
2891 if (!curr->node_stamp)
2892 curr->numa_scan_period = task_scan_start(curr);
2893 curr->node_stamp += period;
2895 if (!time_before(jiffies, curr->mm->numa_next_scan))
2896 task_work_add(curr, work, TWA_RESUME);
2900 static void update_scan_period(struct task_struct *p, int new_cpu)
2902 int src_nid = cpu_to_node(task_cpu(p));
2903 int dst_nid = cpu_to_node(new_cpu);
2905 if (!static_branch_likely(&sched_numa_balancing))
2908 if (!p->mm || !p->numa_faults || (p->flags & PF_EXITING))
2911 if (src_nid == dst_nid)
2915 * Allow resets if faults have been trapped before one scan
2916 * has completed. This is most likely due to a new task that
2917 * is pulled cross-node due to wakeups or load balancing.
2919 if (p->numa_scan_seq) {
2921 * Avoid scan adjustments if moving to the preferred
2922 * node or if the task was not previously running on
2923 * the preferred node.
2925 if (dst_nid == p->numa_preferred_nid ||
2926 (p->numa_preferred_nid != NUMA_NO_NODE &&
2927 src_nid != p->numa_preferred_nid))
2931 p->numa_scan_period = task_scan_start(p);
2935 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
2939 static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p)
2943 static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p)
2947 static inline void update_scan_period(struct task_struct *p, int new_cpu)
2951 #endif /* CONFIG_NUMA_BALANCING */
2954 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2956 update_load_add(&cfs_rq->load, se->load.weight);
2958 if (entity_is_task(se)) {
2959 struct rq *rq = rq_of(cfs_rq);
2961 account_numa_enqueue(rq, task_of(se));
2962 list_add(&se->group_node, &rq->cfs_tasks);
2965 cfs_rq->nr_running++;
2969 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2971 update_load_sub(&cfs_rq->load, se->load.weight);
2973 if (entity_is_task(se)) {
2974 account_numa_dequeue(rq_of(cfs_rq), task_of(se));
2975 list_del_init(&se->group_node);
2978 cfs_rq->nr_running--;
2982 * Signed add and clamp on underflow.
2984 * Explicitly do a load-store to ensure the intermediate value never hits
2985 * memory. This allows lockless observations without ever seeing the negative
2988 #define add_positive(_ptr, _val) do { \
2989 typeof(_ptr) ptr = (_ptr); \
2990 typeof(_val) val = (_val); \
2991 typeof(*ptr) res, var = READ_ONCE(*ptr); \
2995 if (val < 0 && res > var) \
2998 WRITE_ONCE(*ptr, res); \
3002 * Unsigned subtract and clamp on underflow.
3004 * Explicitly do a load-store to ensure the intermediate value never hits
3005 * memory. This allows lockless observations without ever seeing the negative
3008 #define sub_positive(_ptr, _val) do { \
3009 typeof(_ptr) ptr = (_ptr); \
3010 typeof(*ptr) val = (_val); \
3011 typeof(*ptr) res, var = READ_ONCE(*ptr); \
3015 WRITE_ONCE(*ptr, res); \
3019 * Remove and clamp on negative, from a local variable.
3021 * A variant of sub_positive(), which does not use explicit load-store
3022 * and is thus optimized for local variable updates.
3024 #define lsub_positive(_ptr, _val) do { \
3025 typeof(_ptr) ptr = (_ptr); \
3026 *ptr -= min_t(typeof(*ptr), *ptr, _val); \
3031 enqueue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3033 cfs_rq->avg.load_avg += se->avg.load_avg;
3034 cfs_rq->avg.load_sum += se_weight(se) * se->avg.load_sum;
3038 dequeue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3040 sub_positive(&cfs_rq->avg.load_avg, se->avg.load_avg);
3041 sub_positive(&cfs_rq->avg.load_sum, se_weight(se) * se->avg.load_sum);
3045 enqueue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
3047 dequeue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
3050 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
3051 unsigned long weight)
3054 /* commit outstanding execution time */
3055 if (cfs_rq->curr == se)
3056 update_curr(cfs_rq);
3057 update_load_sub(&cfs_rq->load, se->load.weight);
3059 dequeue_load_avg(cfs_rq, se);
3061 update_load_set(&se->load, weight);
3065 u32 divider = get_pelt_divider(&se->avg);
3067 se->avg.load_avg = div_u64(se_weight(se) * se->avg.load_sum, divider);
3071 enqueue_load_avg(cfs_rq, se);
3073 update_load_add(&cfs_rq->load, se->load.weight);
3077 void reweight_task(struct task_struct *p, int prio)
3079 struct sched_entity *se = &p->se;
3080 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3081 struct load_weight *load = &se->load;
3082 unsigned long weight = scale_load(sched_prio_to_weight[prio]);
3084 reweight_entity(cfs_rq, se, weight);
3085 load->inv_weight = sched_prio_to_wmult[prio];
3088 #ifdef CONFIG_FAIR_GROUP_SCHED
3091 * All this does is approximate the hierarchical proportion which includes that
3092 * global sum we all love to hate.
3094 * That is, the weight of a group entity, is the proportional share of the
3095 * group weight based on the group runqueue weights. That is:
3097 * tg->weight * grq->load.weight
3098 * ge->load.weight = ----------------------------- (1)
3099 * \Sum grq->load.weight
3101 * Now, because computing that sum is prohibitively expensive to compute (been
3102 * there, done that) we approximate it with this average stuff. The average
3103 * moves slower and therefore the approximation is cheaper and more stable.
3105 * So instead of the above, we substitute:
3107 * grq->load.weight -> grq->avg.load_avg (2)
3109 * which yields the following:
3111 * tg->weight * grq->avg.load_avg
3112 * ge->load.weight = ------------------------------ (3)
3115 * Where: tg->load_avg ~= \Sum grq->avg.load_avg
3117 * That is shares_avg, and it is right (given the approximation (2)).
3119 * The problem with it is that because the average is slow -- it was designed
3120 * to be exactly that of course -- this leads to transients in boundary
3121 * conditions. In specific, the case where the group was idle and we start the
3122 * one task. It takes time for our CPU's grq->avg.load_avg to build up,
3123 * yielding bad latency etc..
3125 * Now, in that special case (1) reduces to:
3127 * tg->weight * grq->load.weight
3128 * ge->load.weight = ----------------------------- = tg->weight (4)
3131 * That is, the sum collapses because all other CPUs are idle; the UP scenario.
3133 * So what we do is modify our approximation (3) to approach (4) in the (near)
3138 * tg->weight * grq->load.weight
3139 * --------------------------------------------------- (5)
3140 * tg->load_avg - grq->avg.load_avg + grq->load.weight
3142 * But because grq->load.weight can drop to 0, resulting in a divide by zero,
3143 * we need to use grq->avg.load_avg as its lower bound, which then gives:
3146 * tg->weight * grq->load.weight
3147 * ge->load.weight = ----------------------------- (6)
3152 * tg_load_avg' = tg->load_avg - grq->avg.load_avg +
3153 * max(grq->load.weight, grq->avg.load_avg)
3155 * And that is shares_weight and is icky. In the (near) UP case it approaches
3156 * (4) while in the normal case it approaches (3). It consistently
3157 * overestimates the ge->load.weight and therefore:
3159 * \Sum ge->load.weight >= tg->weight
3163 static long calc_group_shares(struct cfs_rq *cfs_rq)
3165 long tg_weight, tg_shares, load, shares;
3166 struct task_group *tg = cfs_rq->tg;
3168 tg_shares = READ_ONCE(tg->shares);
3170 load = max(scale_load_down(cfs_rq->load.weight), cfs_rq->avg.load_avg);
3172 tg_weight = atomic_long_read(&tg->load_avg);
3174 /* Ensure tg_weight >= load */
3175 tg_weight -= cfs_rq->tg_load_avg_contrib;
3178 shares = (tg_shares * load);
3180 shares /= tg_weight;
3183 * MIN_SHARES has to be unscaled here to support per-CPU partitioning
3184 * of a group with small tg->shares value. It is a floor value which is
3185 * assigned as a minimum load.weight to the sched_entity representing
3186 * the group on a CPU.
3188 * E.g. on 64-bit for a group with tg->shares of scale_load(15)=15*1024
3189 * on an 8-core system with 8 tasks each runnable on one CPU shares has
3190 * to be 15*1024*1/8=1920 instead of scale_load(MIN_SHARES)=2*1024. In
3191 * case no task is runnable on a CPU MIN_SHARES=2 should be returned
3194 return clamp_t(long, shares, MIN_SHARES, tg_shares);
3196 #endif /* CONFIG_SMP */
3198 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
3201 * Recomputes the group entity based on the current state of its group
3204 static void update_cfs_group(struct sched_entity *se)
3206 struct cfs_rq *gcfs_rq = group_cfs_rq(se);
3212 if (throttled_hierarchy(gcfs_rq))
3216 shares = READ_ONCE(gcfs_rq->tg->shares);
3218 if (likely(se->load.weight == shares))
3221 shares = calc_group_shares(gcfs_rq);
3224 reweight_entity(cfs_rq_of(se), se, shares);
3227 #else /* CONFIG_FAIR_GROUP_SCHED */
3228 static inline void update_cfs_group(struct sched_entity *se)
3231 #endif /* CONFIG_FAIR_GROUP_SCHED */
3233 static inline void cfs_rq_util_change(struct cfs_rq *cfs_rq, int flags)
3235 struct rq *rq = rq_of(cfs_rq);
3237 if (&rq->cfs == cfs_rq) {
3239 * There are a few boundary cases this might miss but it should
3240 * get called often enough that that should (hopefully) not be
3243 * It will not get called when we go idle, because the idle
3244 * thread is a different class (!fair), nor will the utilization
3245 * number include things like RT tasks.
3247 * As is, the util number is not freq-invariant (we'd have to
3248 * implement arch_scale_freq_capacity() for that).
3252 cpufreq_update_util(rq, flags);
3257 #ifdef CONFIG_FAIR_GROUP_SCHED
3259 * Because list_add_leaf_cfs_rq always places a child cfs_rq on the list
3260 * immediately before a parent cfs_rq, and cfs_rqs are removed from the list
3261 * bottom-up, we only have to test whether the cfs_rq before us on the list
3263 * If cfs_rq is not on the list, test whether a child needs its to be added to
3264 * connect a branch to the tree * (see list_add_leaf_cfs_rq() for details).
3266 static inline bool child_cfs_rq_on_list(struct cfs_rq *cfs_rq)
3268 struct cfs_rq *prev_cfs_rq;
3269 struct list_head *prev;
3271 if (cfs_rq->on_list) {
3272 prev = cfs_rq->leaf_cfs_rq_list.prev;
3274 struct rq *rq = rq_of(cfs_rq);
3276 prev = rq->tmp_alone_branch;
3279 prev_cfs_rq = container_of(prev, struct cfs_rq, leaf_cfs_rq_list);
3281 return (prev_cfs_rq->tg->parent == cfs_rq->tg);
3284 static inline bool cfs_rq_is_decayed(struct cfs_rq *cfs_rq)
3286 if (cfs_rq->load.weight)
3289 if (cfs_rq->avg.load_sum)
3292 if (cfs_rq->avg.util_sum)
3295 if (cfs_rq->avg.runnable_sum)
3298 if (child_cfs_rq_on_list(cfs_rq))
3302 * _avg must be null when _sum are null because _avg = _sum / divider
3303 * Make sure that rounding and/or propagation of PELT values never
3306 SCHED_WARN_ON(cfs_rq->avg.load_avg ||
3307 cfs_rq->avg.util_avg ||
3308 cfs_rq->avg.runnable_avg);
3314 * update_tg_load_avg - update the tg's load avg
3315 * @cfs_rq: the cfs_rq whose avg changed
3317 * This function 'ensures': tg->load_avg := \Sum tg->cfs_rq[]->avg.load.
3318 * However, because tg->load_avg is a global value there are performance
3321 * In order to avoid having to look at the other cfs_rq's, we use a
3322 * differential update where we store the last value we propagated. This in
3323 * turn allows skipping updates if the differential is 'small'.
3325 * Updating tg's load_avg is necessary before update_cfs_share().
3327 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq)
3329 long delta = cfs_rq->avg.load_avg - cfs_rq->tg_load_avg_contrib;
3332 * No need to update load_avg for root_task_group as it is not used.
3334 if (cfs_rq->tg == &root_task_group)
3337 if (abs(delta) > cfs_rq->tg_load_avg_contrib / 64) {
3338 atomic_long_add(delta, &cfs_rq->tg->load_avg);
3339 cfs_rq->tg_load_avg_contrib = cfs_rq->avg.load_avg;
3344 * Called within set_task_rq() right before setting a task's CPU. The
3345 * caller only guarantees p->pi_lock is held; no other assumptions,
3346 * including the state of rq->lock, should be made.
3348 void set_task_rq_fair(struct sched_entity *se,
3349 struct cfs_rq *prev, struct cfs_rq *next)
3351 u64 p_last_update_time;
3352 u64 n_last_update_time;
3354 if (!sched_feat(ATTACH_AGE_LOAD))
3358 * We are supposed to update the task to "current" time, then its up to
3359 * date and ready to go to new CPU/cfs_rq. But we have difficulty in
3360 * getting what current time is, so simply throw away the out-of-date
3361 * time. This will result in the wakee task is less decayed, but giving
3362 * the wakee more load sounds not bad.
3364 if (!(se->avg.last_update_time && prev))
3367 #ifndef CONFIG_64BIT
3369 u64 p_last_update_time_copy;
3370 u64 n_last_update_time_copy;
3373 p_last_update_time_copy = prev->load_last_update_time_copy;
3374 n_last_update_time_copy = next->load_last_update_time_copy;
3378 p_last_update_time = prev->avg.last_update_time;
3379 n_last_update_time = next->avg.last_update_time;
3381 } while (p_last_update_time != p_last_update_time_copy ||
3382 n_last_update_time != n_last_update_time_copy);
3385 p_last_update_time = prev->avg.last_update_time;
3386 n_last_update_time = next->avg.last_update_time;
3388 __update_load_avg_blocked_se(p_last_update_time, se);
3389 se->avg.last_update_time = n_last_update_time;
3394 * When on migration a sched_entity joins/leaves the PELT hierarchy, we need to
3395 * propagate its contribution. The key to this propagation is the invariant
3396 * that for each group:
3398 * ge->avg == grq->avg (1)
3400 * _IFF_ we look at the pure running and runnable sums. Because they
3401 * represent the very same entity, just at different points in the hierarchy.
3403 * Per the above update_tg_cfs_util() and update_tg_cfs_runnable() are trivial
3404 * and simply copies the running/runnable sum over (but still wrong, because
3405 * the group entity and group rq do not have their PELT windows aligned).
3407 * However, update_tg_cfs_load() is more complex. So we have:
3409 * ge->avg.load_avg = ge->load.weight * ge->avg.runnable_avg (2)
3411 * And since, like util, the runnable part should be directly transferable,
3412 * the following would _appear_ to be the straight forward approach:
3414 * grq->avg.load_avg = grq->load.weight * grq->avg.runnable_avg (3)
3416 * And per (1) we have:
3418 * ge->avg.runnable_avg == grq->avg.runnable_avg
3422 * ge->load.weight * grq->avg.load_avg
3423 * ge->avg.load_avg = ----------------------------------- (4)
3426 * Except that is wrong!
3428 * Because while for entities historical weight is not important and we
3429 * really only care about our future and therefore can consider a pure
3430 * runnable sum, runqueues can NOT do this.
3432 * We specifically want runqueues to have a load_avg that includes
3433 * historical weights. Those represent the blocked load, the load we expect
3434 * to (shortly) return to us. This only works by keeping the weights as
3435 * integral part of the sum. We therefore cannot decompose as per (3).
3437 * Another reason this doesn't work is that runnable isn't a 0-sum entity.
3438 * Imagine a rq with 2 tasks that each are runnable 2/3 of the time. Then the
3439 * rq itself is runnable anywhere between 2/3 and 1 depending on how the
3440 * runnable section of these tasks overlap (or not). If they were to perfectly
3441 * align the rq as a whole would be runnable 2/3 of the time. If however we
3442 * always have at least 1 runnable task, the rq as a whole is always runnable.
3444 * So we'll have to approximate.. :/
3446 * Given the constraint:
3448 * ge->avg.running_sum <= ge->avg.runnable_sum <= LOAD_AVG_MAX
3450 * We can construct a rule that adds runnable to a rq by assuming minimal
3453 * On removal, we'll assume each task is equally runnable; which yields:
3455 * grq->avg.runnable_sum = grq->avg.load_sum / grq->load.weight
3457 * XXX: only do this for the part of runnable > running ?
3462 update_tg_cfs_util(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq)
3464 long delta = gcfs_rq->avg.util_avg - se->avg.util_avg;
3467 /* Nothing to update */
3472 * cfs_rq->avg.period_contrib can be used for both cfs_rq and se.
3473 * See ___update_load_avg() for details.
3475 divider = get_pelt_divider(&cfs_rq->avg);
3477 /* Set new sched_entity's utilization */
3478 se->avg.util_avg = gcfs_rq->avg.util_avg;
3479 se->avg.util_sum = se->avg.util_avg * divider;
3481 /* Update parent cfs_rq utilization */
3482 add_positive(&cfs_rq->avg.util_avg, delta);
3483 cfs_rq->avg.util_sum = cfs_rq->avg.util_avg * divider;
3487 update_tg_cfs_runnable(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq)
3489 long delta = gcfs_rq->avg.runnable_avg - se->avg.runnable_avg;
3492 /* Nothing to update */
3497 * cfs_rq->avg.period_contrib can be used for both cfs_rq and se.
3498 * See ___update_load_avg() for details.
3500 divider = get_pelt_divider(&cfs_rq->avg);
3502 /* Set new sched_entity's runnable */
3503 se->avg.runnable_avg = gcfs_rq->avg.runnable_avg;
3504 se->avg.runnable_sum = se->avg.runnable_avg * divider;
3506 /* Update parent cfs_rq runnable */
3507 add_positive(&cfs_rq->avg.runnable_avg, delta);
3508 cfs_rq->avg.runnable_sum = cfs_rq->avg.runnable_avg * divider;
3512 update_tg_cfs_load(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq)
3514 long delta, running_sum, runnable_sum = gcfs_rq->prop_runnable_sum;
3515 unsigned long load_avg;
3522 gcfs_rq->prop_runnable_sum = 0;
3525 * cfs_rq->avg.period_contrib can be used for both cfs_rq and se.
3526 * See ___update_load_avg() for details.
3528 divider = get_pelt_divider(&cfs_rq->avg);
3530 if (runnable_sum >= 0) {
3532 * Add runnable; clip at LOAD_AVG_MAX. Reflects that until
3533 * the CPU is saturated running == runnable.
3535 runnable_sum += se->avg.load_sum;
3536 runnable_sum = min_t(long, runnable_sum, divider);
3539 * Estimate the new unweighted runnable_sum of the gcfs_rq by
3540 * assuming all tasks are equally runnable.
3542 if (scale_load_down(gcfs_rq->load.weight)) {
3543 load_sum = div_s64(gcfs_rq->avg.load_sum,
3544 scale_load_down(gcfs_rq->load.weight));
3547 /* But make sure to not inflate se's runnable */
3548 runnable_sum = min(se->avg.load_sum, load_sum);
3552 * runnable_sum can't be lower than running_sum
3553 * Rescale running sum to be in the same range as runnable sum
3554 * running_sum is in [0 : LOAD_AVG_MAX << SCHED_CAPACITY_SHIFT]
3555 * runnable_sum is in [0 : LOAD_AVG_MAX]
3557 running_sum = se->avg.util_sum >> SCHED_CAPACITY_SHIFT;
3558 runnable_sum = max(runnable_sum, running_sum);
3560 load_sum = (s64)se_weight(se) * runnable_sum;
3561 load_avg = div_s64(load_sum, divider);
3563 se->avg.load_sum = runnable_sum;
3565 delta = load_avg - se->avg.load_avg;
3569 se->avg.load_avg = load_avg;
3571 add_positive(&cfs_rq->avg.load_avg, delta);
3572 cfs_rq->avg.load_sum = cfs_rq->avg.load_avg * divider;
3575 static inline void add_tg_cfs_propagate(struct cfs_rq *cfs_rq, long runnable_sum)
3577 cfs_rq->propagate = 1;
3578 cfs_rq->prop_runnable_sum += runnable_sum;
3581 /* Update task and its cfs_rq load average */
3582 static inline int propagate_entity_load_avg(struct sched_entity *se)
3584 struct cfs_rq *cfs_rq, *gcfs_rq;
3586 if (entity_is_task(se))
3589 gcfs_rq = group_cfs_rq(se);
3590 if (!gcfs_rq->propagate)
3593 gcfs_rq->propagate = 0;
3595 cfs_rq = cfs_rq_of(se);
3597 add_tg_cfs_propagate(cfs_rq, gcfs_rq->prop_runnable_sum);
3599 update_tg_cfs_util(cfs_rq, se, gcfs_rq);
3600 update_tg_cfs_runnable(cfs_rq, se, gcfs_rq);
3601 update_tg_cfs_load(cfs_rq, se, gcfs_rq);
3603 trace_pelt_cfs_tp(cfs_rq);
3604 trace_pelt_se_tp(se);
3610 * Check if we need to update the load and the utilization of a blocked
3613 static inline bool skip_blocked_update(struct sched_entity *se)
3615 struct cfs_rq *gcfs_rq = group_cfs_rq(se);
3618 * If sched_entity still have not zero load or utilization, we have to
3621 if (se->avg.load_avg || se->avg.util_avg)
3625 * If there is a pending propagation, we have to update the load and
3626 * the utilization of the sched_entity:
3628 if (gcfs_rq->propagate)
3632 * Otherwise, the load and the utilization of the sched_entity is
3633 * already zero and there is no pending propagation, so it will be a
3634 * waste of time to try to decay it:
3639 #else /* CONFIG_FAIR_GROUP_SCHED */
3641 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq) {}
3643 static inline int propagate_entity_load_avg(struct sched_entity *se)
3648 static inline void add_tg_cfs_propagate(struct cfs_rq *cfs_rq, long runnable_sum) {}
3650 #endif /* CONFIG_FAIR_GROUP_SCHED */
3653 * update_cfs_rq_load_avg - update the cfs_rq's load/util averages
3654 * @now: current time, as per cfs_rq_clock_pelt()
3655 * @cfs_rq: cfs_rq to update
3657 * The cfs_rq avg is the direct sum of all its entities (blocked and runnable)
3658 * avg. The immediate corollary is that all (fair) tasks must be attached, see
3659 * post_init_entity_util_avg().
3661 * cfs_rq->avg is used for task_h_load() and update_cfs_share() for example.
3663 * Returns true if the load decayed or we removed load.
3665 * Since both these conditions indicate a changed cfs_rq->avg.load we should
3666 * call update_tg_load_avg() when this function returns true.
3669 update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq)
3671 unsigned long removed_load = 0, removed_util = 0, removed_runnable = 0;
3672 struct sched_avg *sa = &cfs_rq->avg;
3675 if (cfs_rq->removed.nr) {
3677 u32 divider = get_pelt_divider(&cfs_rq->avg);
3679 raw_spin_lock(&cfs_rq->removed.lock);
3680 swap(cfs_rq->removed.util_avg, removed_util);
3681 swap(cfs_rq->removed.load_avg, removed_load);
3682 swap(cfs_rq->removed.runnable_avg, removed_runnable);
3683 cfs_rq->removed.nr = 0;
3684 raw_spin_unlock(&cfs_rq->removed.lock);
3687 sub_positive(&sa->load_avg, r);
3688 sa->load_sum = sa->load_avg * divider;
3691 sub_positive(&sa->util_avg, r);
3692 sa->util_sum = sa->util_avg * divider;
3694 r = removed_runnable;
3695 sub_positive(&sa->runnable_avg, r);
3696 sa->runnable_sum = sa->runnable_avg * divider;
3699 * removed_runnable is the unweighted version of removed_load so we
3700 * can use it to estimate removed_load_sum.
3702 add_tg_cfs_propagate(cfs_rq,
3703 -(long)(removed_runnable * divider) >> SCHED_CAPACITY_SHIFT);
3708 decayed |= __update_load_avg_cfs_rq(now, cfs_rq);
3710 #ifndef CONFIG_64BIT
3712 cfs_rq->load_last_update_time_copy = sa->last_update_time;
3719 * attach_entity_load_avg - attach this entity to its cfs_rq load avg
3720 * @cfs_rq: cfs_rq to attach to
3721 * @se: sched_entity to attach
3723 * Must call update_cfs_rq_load_avg() before this, since we rely on
3724 * cfs_rq->avg.last_update_time being current.
3726 static void attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3729 * cfs_rq->avg.period_contrib can be used for both cfs_rq and se.
3730 * See ___update_load_avg() for details.
3732 u32 divider = get_pelt_divider(&cfs_rq->avg);
3735 * When we attach the @se to the @cfs_rq, we must align the decay
3736 * window because without that, really weird and wonderful things can
3741 se->avg.last_update_time = cfs_rq->avg.last_update_time;
3742 se->avg.period_contrib = cfs_rq->avg.period_contrib;
3745 * Hell(o) Nasty stuff.. we need to recompute _sum based on the new
3746 * period_contrib. This isn't strictly correct, but since we're
3747 * entirely outside of the PELT hierarchy, nobody cares if we truncate
3750 se->avg.util_sum = se->avg.util_avg * divider;
3752 se->avg.runnable_sum = se->avg.runnable_avg * divider;
3754 se->avg.load_sum = divider;
3755 if (se_weight(se)) {
3757 div_u64(se->avg.load_avg * se->avg.load_sum, se_weight(se));
3760 enqueue_load_avg(cfs_rq, se);
3761 cfs_rq->avg.util_avg += se->avg.util_avg;
3762 cfs_rq->avg.util_sum += se->avg.util_sum;
3763 cfs_rq->avg.runnable_avg += se->avg.runnable_avg;
3764 cfs_rq->avg.runnable_sum += se->avg.runnable_sum;
3766 add_tg_cfs_propagate(cfs_rq, se->avg.load_sum);
3768 cfs_rq_util_change(cfs_rq, 0);
3770 trace_pelt_cfs_tp(cfs_rq);
3774 * detach_entity_load_avg - detach this entity from its cfs_rq load avg
3775 * @cfs_rq: cfs_rq to detach from
3776 * @se: sched_entity to detach
3778 * Must call update_cfs_rq_load_avg() before this, since we rely on
3779 * cfs_rq->avg.last_update_time being current.
3781 static void detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3784 * cfs_rq->avg.period_contrib can be used for both cfs_rq and se.
3785 * See ___update_load_avg() for details.
3787 u32 divider = get_pelt_divider(&cfs_rq->avg);
3789 dequeue_load_avg(cfs_rq, se);
3790 sub_positive(&cfs_rq->avg.util_avg, se->avg.util_avg);
3791 cfs_rq->avg.util_sum = cfs_rq->avg.util_avg * divider;
3792 sub_positive(&cfs_rq->avg.runnable_avg, se->avg.runnable_avg);
3793 cfs_rq->avg.runnable_sum = cfs_rq->avg.runnable_avg * divider;
3795 add_tg_cfs_propagate(cfs_rq, -se->avg.load_sum);
3797 cfs_rq_util_change(cfs_rq, 0);
3799 trace_pelt_cfs_tp(cfs_rq);
3803 * Optional action to be done while updating the load average
3805 #define UPDATE_TG 0x1
3806 #define SKIP_AGE_LOAD 0x2
3807 #define DO_ATTACH 0x4
3809 /* Update task and its cfs_rq load average */
3810 static inline void update_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3812 u64 now = cfs_rq_clock_pelt(cfs_rq);
3816 * Track task load average for carrying it to new CPU after migrated, and
3817 * track group sched_entity load average for task_h_load calc in migration
3819 if (se->avg.last_update_time && !(flags & SKIP_AGE_LOAD))
3820 __update_load_avg_se(now, cfs_rq, se);
3822 decayed = update_cfs_rq_load_avg(now, cfs_rq);
3823 decayed |= propagate_entity_load_avg(se);
3825 if (!se->avg.last_update_time && (flags & DO_ATTACH)) {
3828 * DO_ATTACH means we're here from enqueue_entity().
3829 * !last_update_time means we've passed through
3830 * migrate_task_rq_fair() indicating we migrated.
3832 * IOW we're enqueueing a task on a new CPU.
3834 attach_entity_load_avg(cfs_rq, se);
3835 update_tg_load_avg(cfs_rq);
3837 } else if (decayed) {
3838 cfs_rq_util_change(cfs_rq, 0);
3840 if (flags & UPDATE_TG)
3841 update_tg_load_avg(cfs_rq);
3845 #ifndef CONFIG_64BIT
3846 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
3848 u64 last_update_time_copy;
3849 u64 last_update_time;
3852 last_update_time_copy = cfs_rq->load_last_update_time_copy;
3854 last_update_time = cfs_rq->avg.last_update_time;
3855 } while (last_update_time != last_update_time_copy);
3857 return last_update_time;
3860 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
3862 return cfs_rq->avg.last_update_time;
3867 * Synchronize entity load avg of dequeued entity without locking
3870 static void sync_entity_load_avg(struct sched_entity *se)
3872 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3873 u64 last_update_time;
3875 last_update_time = cfs_rq_last_update_time(cfs_rq);
3876 __update_load_avg_blocked_se(last_update_time, se);
3880 * Task first catches up with cfs_rq, and then subtract
3881 * itself from the cfs_rq (task must be off the queue now).
3883 static void remove_entity_load_avg(struct sched_entity *se)
3885 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3886 unsigned long flags;
3889 * tasks cannot exit without having gone through wake_up_new_task() ->
3890 * post_init_entity_util_avg() which will have added things to the
3891 * cfs_rq, so we can remove unconditionally.
3894 sync_entity_load_avg(se);
3896 raw_spin_lock_irqsave(&cfs_rq->removed.lock, flags);
3897 ++cfs_rq->removed.nr;
3898 cfs_rq->removed.util_avg += se->avg.util_avg;
3899 cfs_rq->removed.load_avg += se->avg.load_avg;
3900 cfs_rq->removed.runnable_avg += se->avg.runnable_avg;
3901 raw_spin_unlock_irqrestore(&cfs_rq->removed.lock, flags);
3904 static inline unsigned long cfs_rq_runnable_avg(struct cfs_rq *cfs_rq)
3906 return cfs_rq->avg.runnable_avg;
3909 static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq)
3911 return cfs_rq->avg.load_avg;
3914 static int newidle_balance(struct rq *this_rq, struct rq_flags *rf);
3916 static inline unsigned long task_util(struct task_struct *p)
3918 return READ_ONCE(p->se.avg.util_avg);
3921 static inline unsigned long _task_util_est(struct task_struct *p)
3923 struct util_est ue = READ_ONCE(p->se.avg.util_est);
3925 return max(ue.ewma, (ue.enqueued & ~UTIL_AVG_UNCHANGED));
3928 static inline unsigned long task_util_est(struct task_struct *p)
3930 return max(task_util(p), _task_util_est(p));
3933 #ifdef CONFIG_UCLAMP_TASK
3934 static inline unsigned long uclamp_task_util(struct task_struct *p)
3936 return clamp(task_util_est(p),
3937 uclamp_eff_value(p, UCLAMP_MIN),
3938 uclamp_eff_value(p, UCLAMP_MAX));
3941 static inline unsigned long uclamp_task_util(struct task_struct *p)
3943 return task_util_est(p);
3947 static inline void util_est_enqueue(struct cfs_rq *cfs_rq,
3948 struct task_struct *p)
3950 unsigned int enqueued;
3952 if (!sched_feat(UTIL_EST))
3955 /* Update root cfs_rq's estimated utilization */
3956 enqueued = cfs_rq->avg.util_est.enqueued;
3957 enqueued += _task_util_est(p);
3958 WRITE_ONCE(cfs_rq->avg.util_est.enqueued, enqueued);
3960 trace_sched_util_est_cfs_tp(cfs_rq);
3963 static inline void util_est_dequeue(struct cfs_rq *cfs_rq,
3964 struct task_struct *p)
3966 unsigned int enqueued;
3968 if (!sched_feat(UTIL_EST))
3971 /* Update root cfs_rq's estimated utilization */
3972 enqueued = cfs_rq->avg.util_est.enqueued;
3973 enqueued -= min_t(unsigned int, enqueued, _task_util_est(p));
3974 WRITE_ONCE(cfs_rq->avg.util_est.enqueued, enqueued);
3976 trace_sched_util_est_cfs_tp(cfs_rq);
3979 #define UTIL_EST_MARGIN (SCHED_CAPACITY_SCALE / 100)
3982 * Check if a (signed) value is within a specified (unsigned) margin,
3983 * based on the observation that:
3985 * abs(x) < y := (unsigned)(x + y - 1) < (2 * y - 1)
3987 * NOTE: this only works when value + margin < INT_MAX.
3989 static inline bool within_margin(int value, int margin)
3991 return ((unsigned int)(value + margin - 1) < (2 * margin - 1));
3994 static inline void util_est_update(struct cfs_rq *cfs_rq,
3995 struct task_struct *p,
3998 long last_ewma_diff, last_enqueued_diff;
4001 if (!sched_feat(UTIL_EST))
4005 * Skip update of task's estimated utilization when the task has not
4006 * yet completed an activation, e.g. being migrated.
4012 * If the PELT values haven't changed since enqueue time,
4013 * skip the util_est update.
4015 ue = p->se.avg.util_est;
4016 if (ue.enqueued & UTIL_AVG_UNCHANGED)
4019 last_enqueued_diff = ue.enqueued;
4022 * Reset EWMA on utilization increases, the moving average is used only
4023 * to smooth utilization decreases.
4025 ue.enqueued = task_util(p);
4026 if (sched_feat(UTIL_EST_FASTUP)) {
4027 if (ue.ewma < ue.enqueued) {
4028 ue.ewma = ue.enqueued;
4034 * Skip update of task's estimated utilization when its members are
4035 * already ~1% close to its last activation value.
4037 last_ewma_diff = ue.enqueued - ue.ewma;
4038 last_enqueued_diff -= ue.enqueued;
4039 if (within_margin(last_ewma_diff, UTIL_EST_MARGIN)) {
4040 if (!within_margin(last_enqueued_diff, UTIL_EST_MARGIN))
4047 * To avoid overestimation of actual task utilization, skip updates if
4048 * we cannot grant there is idle time in this CPU.
4050 if (task_util(p) > capacity_orig_of(cpu_of(rq_of(cfs_rq))))
4054 * Update Task's estimated utilization
4056 * When *p completes an activation we can consolidate another sample
4057 * of the task size. This is done by storing the current PELT value
4058 * as ue.enqueued and by using this value to update the Exponential
4059 * Weighted Moving Average (EWMA):
4061 * ewma(t) = w * task_util(p) + (1-w) * ewma(t-1)
4062 * = w * task_util(p) + ewma(t-1) - w * ewma(t-1)
4063 * = w * (task_util(p) - ewma(t-1)) + ewma(t-1)
4064 * = w * ( last_ewma_diff ) + ewma(t-1)
4065 * = w * (last_ewma_diff + ewma(t-1) / w)
4067 * Where 'w' is the weight of new samples, which is configured to be
4068 * 0.25, thus making w=1/4 ( >>= UTIL_EST_WEIGHT_SHIFT)
4070 ue.ewma <<= UTIL_EST_WEIGHT_SHIFT;
4071 ue.ewma += last_ewma_diff;
4072 ue.ewma >>= UTIL_EST_WEIGHT_SHIFT;
4074 ue.enqueued |= UTIL_AVG_UNCHANGED;
4075 WRITE_ONCE(p->se.avg.util_est, ue);
4077 trace_sched_util_est_se_tp(&p->se);
4080 static inline int task_fits_capacity(struct task_struct *p, long capacity)
4082 return fits_capacity(uclamp_task_util(p), capacity);
4085 static inline void update_misfit_status(struct task_struct *p, struct rq *rq)
4087 if (!static_branch_unlikely(&sched_asym_cpucapacity))
4090 if (!p || p->nr_cpus_allowed == 1) {
4091 rq->misfit_task_load = 0;
4095 if (task_fits_capacity(p, capacity_of(cpu_of(rq)))) {
4096 rq->misfit_task_load = 0;
4101 * Make sure that misfit_task_load will not be null even if
4102 * task_h_load() returns 0.
4104 rq->misfit_task_load = max_t(unsigned long, task_h_load(p), 1);
4107 #else /* CONFIG_SMP */
4109 static inline bool cfs_rq_is_decayed(struct cfs_rq *cfs_rq)
4114 #define UPDATE_TG 0x0
4115 #define SKIP_AGE_LOAD 0x0
4116 #define DO_ATTACH 0x0
4118 static inline void update_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int not_used1)
4120 cfs_rq_util_change(cfs_rq, 0);
4123 static inline void remove_entity_load_avg(struct sched_entity *se) {}
4126 attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
4128 detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
4130 static inline int newidle_balance(struct rq *rq, struct rq_flags *rf)
4136 util_est_enqueue(struct cfs_rq *cfs_rq, struct task_struct *p) {}
4139 util_est_dequeue(struct cfs_rq *cfs_rq, struct task_struct *p) {}
4142 util_est_update(struct cfs_rq *cfs_rq, struct task_struct *p,
4144 static inline void update_misfit_status(struct task_struct *p, struct rq *rq) {}
4146 #endif /* CONFIG_SMP */
4148 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
4150 #ifdef CONFIG_SCHED_DEBUG
4151 s64 d = se->vruntime - cfs_rq->min_vruntime;
4156 if (d > 3*sysctl_sched_latency)
4157 schedstat_inc(cfs_rq->nr_spread_over);
4162 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
4164 u64 vruntime = cfs_rq->min_vruntime;
4167 * The 'current' period is already promised to the current tasks,
4168 * however the extra weight of the new task will slow them down a
4169 * little, place the new task so that it fits in the slot that
4170 * stays open at the end.
4172 if (initial && sched_feat(START_DEBIT))
4173 vruntime += sched_vslice(cfs_rq, se);
4175 /* sleeps up to a single latency don't count. */
4177 unsigned long thresh = sysctl_sched_latency;
4180 * Halve their sleep time's effect, to allow
4181 * for a gentler effect of sleepers:
4183 if (sched_feat(GENTLE_FAIR_SLEEPERS))
4189 /* ensure we never gain time by being placed backwards. */
4190 se->vruntime = max_vruntime(se->vruntime, vruntime);
4193 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
4195 static inline void check_schedstat_required(void)
4197 #ifdef CONFIG_SCHEDSTATS
4198 if (schedstat_enabled())
4201 /* Force schedstat enabled if a dependent tracepoint is active */
4202 if (trace_sched_stat_wait_enabled() ||
4203 trace_sched_stat_sleep_enabled() ||
4204 trace_sched_stat_iowait_enabled() ||
4205 trace_sched_stat_blocked_enabled() ||
4206 trace_sched_stat_runtime_enabled()) {
4207 printk_deferred_once("Scheduler tracepoints stat_sleep, stat_iowait, "
4208 "stat_blocked and stat_runtime require the "
4209 "kernel parameter schedstats=enable or "
4210 "kernel.sched_schedstats=1\n");
4215 static inline bool cfs_bandwidth_used(void);
4222 * update_min_vruntime()
4223 * vruntime -= min_vruntime
4227 * update_min_vruntime()
4228 * vruntime += min_vruntime
4230 * this way the vruntime transition between RQs is done when both
4231 * min_vruntime are up-to-date.
4235 * ->migrate_task_rq_fair() (p->state == TASK_WAKING)
4236 * vruntime -= min_vruntime
4240 * update_min_vruntime()
4241 * vruntime += min_vruntime
4243 * this way we don't have the most up-to-date min_vruntime on the originating
4244 * CPU and an up-to-date min_vruntime on the destination CPU.
4248 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
4250 bool renorm = !(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_MIGRATED);
4251 bool curr = cfs_rq->curr == se;
4254 * If we're the current task, we must renormalise before calling
4258 se->vruntime += cfs_rq->min_vruntime;
4260 update_curr(cfs_rq);
4263 * Otherwise, renormalise after, such that we're placed at the current
4264 * moment in time, instead of some random moment in the past. Being
4265 * placed in the past could significantly boost this task to the
4266 * fairness detriment of existing tasks.
4268 if (renorm && !curr)
4269 se->vruntime += cfs_rq->min_vruntime;
4272 * When enqueuing a sched_entity, we must:
4273 * - Update loads to have both entity and cfs_rq synced with now.
4274 * - Add its load to cfs_rq->runnable_avg
4275 * - For group_entity, update its weight to reflect the new share of
4277 * - Add its new weight to cfs_rq->load.weight
4279 update_load_avg(cfs_rq, se, UPDATE_TG | DO_ATTACH);
4280 se_update_runnable(se);
4281 update_cfs_group(se);
4282 account_entity_enqueue(cfs_rq, se);
4284 if (flags & ENQUEUE_WAKEUP)
4285 place_entity(cfs_rq, se, 0);
4287 check_schedstat_required();
4288 update_stats_enqueue(cfs_rq, se, flags);
4289 check_spread(cfs_rq, se);
4291 __enqueue_entity(cfs_rq, se);
4295 * When bandwidth control is enabled, cfs might have been removed
4296 * because of a parent been throttled but cfs->nr_running > 1. Try to
4297 * add it unconditionally.
4299 if (cfs_rq->nr_running == 1 || cfs_bandwidth_used())
4300 list_add_leaf_cfs_rq(cfs_rq);
4302 if (cfs_rq->nr_running == 1)
4303 check_enqueue_throttle(cfs_rq);
4306 static void __clear_buddies_last(struct sched_entity *se)
4308 for_each_sched_entity(se) {
4309 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4310 if (cfs_rq->last != se)
4313 cfs_rq->last = NULL;
4317 static void __clear_buddies_next(struct sched_entity *se)
4319 for_each_sched_entity(se) {
4320 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4321 if (cfs_rq->next != se)
4324 cfs_rq->next = NULL;
4328 static void __clear_buddies_skip(struct sched_entity *se)
4330 for_each_sched_entity(se) {
4331 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4332 if (cfs_rq->skip != se)
4335 cfs_rq->skip = NULL;
4339 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
4341 if (cfs_rq->last == se)
4342 __clear_buddies_last(se);
4344 if (cfs_rq->next == se)
4345 __clear_buddies_next(se);
4347 if (cfs_rq->skip == se)
4348 __clear_buddies_skip(se);
4351 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
4354 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
4357 * Update run-time statistics of the 'current'.
4359 update_curr(cfs_rq);
4362 * When dequeuing a sched_entity, we must:
4363 * - Update loads to have both entity and cfs_rq synced with now.
4364 * - Subtract its load from the cfs_rq->runnable_avg.
4365 * - Subtract its previous weight from cfs_rq->load.weight.
4366 * - For group entity, update its weight to reflect the new share
4367 * of its group cfs_rq.
4369 update_load_avg(cfs_rq, se, UPDATE_TG);
4370 se_update_runnable(se);
4372 update_stats_dequeue(cfs_rq, se, flags);
4374 clear_buddies(cfs_rq, se);
4376 if (se != cfs_rq->curr)
4377 __dequeue_entity(cfs_rq, se);
4379 account_entity_dequeue(cfs_rq, se);
4382 * Normalize after update_curr(); which will also have moved
4383 * min_vruntime if @se is the one holding it back. But before doing
4384 * update_min_vruntime() again, which will discount @se's position and
4385 * can move min_vruntime forward still more.
4387 if (!(flags & DEQUEUE_SLEEP))
4388 se->vruntime -= cfs_rq->min_vruntime;
4390 /* return excess runtime on last dequeue */
4391 return_cfs_rq_runtime(cfs_rq);
4393 update_cfs_group(se);
4396 * Now advance min_vruntime if @se was the entity holding it back,
4397 * except when: DEQUEUE_SAVE && !DEQUEUE_MOVE, in this case we'll be
4398 * put back on, and if we advance min_vruntime, we'll be placed back
4399 * further than we started -- ie. we'll be penalized.
4401 if ((flags & (DEQUEUE_SAVE | DEQUEUE_MOVE)) != DEQUEUE_SAVE)
4402 update_min_vruntime(cfs_rq);
4406 * Preempt the current task with a newly woken task if needed:
4409 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
4411 unsigned long ideal_runtime, delta_exec;
4412 struct sched_entity *se;
4415 ideal_runtime = sched_slice(cfs_rq, curr);
4416 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
4417 if (delta_exec > ideal_runtime) {
4418 resched_curr(rq_of(cfs_rq));
4420 * The current task ran long enough, ensure it doesn't get
4421 * re-elected due to buddy favours.
4423 clear_buddies(cfs_rq, curr);
4428 * Ensure that a task that missed wakeup preemption by a
4429 * narrow margin doesn't have to wait for a full slice.
4430 * This also mitigates buddy induced latencies under load.
4432 if (delta_exec < sysctl_sched_min_granularity)
4435 se = __pick_first_entity(cfs_rq);
4436 delta = curr->vruntime - se->vruntime;
4441 if (delta > ideal_runtime)
4442 resched_curr(rq_of(cfs_rq));
4446 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
4448 clear_buddies(cfs_rq, se);
4450 /* 'current' is not kept within the tree. */
4453 * Any task has to be enqueued before it get to execute on
4454 * a CPU. So account for the time it spent waiting on the
4457 update_stats_wait_end(cfs_rq, se);
4458 __dequeue_entity(cfs_rq, se);
4459 update_load_avg(cfs_rq, se, UPDATE_TG);
4462 update_stats_curr_start(cfs_rq, se);
4466 * Track our maximum slice length, if the CPU's load is at
4467 * least twice that of our own weight (i.e. dont track it
4468 * when there are only lesser-weight tasks around):
4470 if (schedstat_enabled() &&
4471 rq_of(cfs_rq)->cfs.load.weight >= 2*se->load.weight) {
4472 schedstat_set(se->statistics.slice_max,
4473 max((u64)schedstat_val(se->statistics.slice_max),
4474 se->sum_exec_runtime - se->prev_sum_exec_runtime));
4477 se->prev_sum_exec_runtime = se->sum_exec_runtime;
4481 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
4484 * Pick the next process, keeping these things in mind, in this order:
4485 * 1) keep things fair between processes/task groups
4486 * 2) pick the "next" process, since someone really wants that to run
4487 * 3) pick the "last" process, for cache locality
4488 * 4) do not run the "skip" process, if something else is available
4490 static struct sched_entity *
4491 pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr)
4493 struct sched_entity *left = __pick_first_entity(cfs_rq);
4494 struct sched_entity *se;
4497 * If curr is set we have to see if its left of the leftmost entity
4498 * still in the tree, provided there was anything in the tree at all.
4500 if (!left || (curr && entity_before(curr, left)))
4503 se = left; /* ideally we run the leftmost entity */
4506 * Avoid running the skip buddy, if running something else can
4507 * be done without getting too unfair.
4509 if (cfs_rq->skip && cfs_rq->skip == se) {
4510 struct sched_entity *second;
4513 second = __pick_first_entity(cfs_rq);
4515 second = __pick_next_entity(se);
4516 if (!second || (curr && entity_before(curr, second)))
4520 if (second && wakeup_preempt_entity(second, left) < 1)
4524 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1) {
4526 * Someone really wants this to run. If it's not unfair, run it.
4529 } else if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1) {
4531 * Prefer last buddy, try to return the CPU to a preempted task.
4539 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
4541 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
4544 * If still on the runqueue then deactivate_task()
4545 * was not called and update_curr() has to be done:
4548 update_curr(cfs_rq);
4550 /* throttle cfs_rqs exceeding runtime */
4551 check_cfs_rq_runtime(cfs_rq);
4553 check_spread(cfs_rq, prev);
4556 update_stats_wait_start(cfs_rq, prev);
4557 /* Put 'current' back into the tree. */
4558 __enqueue_entity(cfs_rq, prev);
4559 /* in !on_rq case, update occurred at dequeue */
4560 update_load_avg(cfs_rq, prev, 0);
4562 cfs_rq->curr = NULL;
4566 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
4569 * Update run-time statistics of the 'current'.
4571 update_curr(cfs_rq);
4574 * Ensure that runnable average is periodically updated.
4576 update_load_avg(cfs_rq, curr, UPDATE_TG);
4577 update_cfs_group(curr);
4579 #ifdef CONFIG_SCHED_HRTICK
4581 * queued ticks are scheduled to match the slice, so don't bother
4582 * validating it and just reschedule.
4585 resched_curr(rq_of(cfs_rq));
4589 * don't let the period tick interfere with the hrtick preemption
4591 if (!sched_feat(DOUBLE_TICK) &&
4592 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
4596 if (cfs_rq->nr_running > 1)
4597 check_preempt_tick(cfs_rq, curr);
4601 /**************************************************
4602 * CFS bandwidth control machinery
4605 #ifdef CONFIG_CFS_BANDWIDTH
4607 #ifdef CONFIG_JUMP_LABEL
4608 static struct static_key __cfs_bandwidth_used;
4610 static inline bool cfs_bandwidth_used(void)
4612 return static_key_false(&__cfs_bandwidth_used);
4615 void cfs_bandwidth_usage_inc(void)
4617 static_key_slow_inc_cpuslocked(&__cfs_bandwidth_used);
4620 void cfs_bandwidth_usage_dec(void)
4622 static_key_slow_dec_cpuslocked(&__cfs_bandwidth_used);
4624 #else /* CONFIG_JUMP_LABEL */
4625 static bool cfs_bandwidth_used(void)
4630 void cfs_bandwidth_usage_inc(void) {}
4631 void cfs_bandwidth_usage_dec(void) {}
4632 #endif /* CONFIG_JUMP_LABEL */
4635 * default period for cfs group bandwidth.
4636 * default: 0.1s, units: nanoseconds
4638 static inline u64 default_cfs_period(void)
4640 return 100000000ULL;
4643 static inline u64 sched_cfs_bandwidth_slice(void)
4645 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
4649 * Replenish runtime according to assigned quota. We use sched_clock_cpu
4650 * directly instead of rq->clock to avoid adding additional synchronization
4653 * requires cfs_b->lock
4655 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
4657 if (unlikely(cfs_b->quota == RUNTIME_INF))
4660 cfs_b->runtime += cfs_b->quota;
4661 cfs_b->runtime = min(cfs_b->runtime, cfs_b->quota + cfs_b->burst);
4664 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
4666 return &tg->cfs_bandwidth;
4669 /* returns 0 on failure to allocate runtime */
4670 static int __assign_cfs_rq_runtime(struct cfs_bandwidth *cfs_b,
4671 struct cfs_rq *cfs_rq, u64 target_runtime)
4673 u64 min_amount, amount = 0;
4675 lockdep_assert_held(&cfs_b->lock);
4677 /* note: this is a positive sum as runtime_remaining <= 0 */
4678 min_amount = target_runtime - cfs_rq->runtime_remaining;
4680 if (cfs_b->quota == RUNTIME_INF)
4681 amount = min_amount;
4683 start_cfs_bandwidth(cfs_b);
4685 if (cfs_b->runtime > 0) {
4686 amount = min(cfs_b->runtime, min_amount);
4687 cfs_b->runtime -= amount;
4692 cfs_rq->runtime_remaining += amount;
4694 return cfs_rq->runtime_remaining > 0;
4697 /* returns 0 on failure to allocate runtime */
4698 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4700 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4703 raw_spin_lock(&cfs_b->lock);
4704 ret = __assign_cfs_rq_runtime(cfs_b, cfs_rq, sched_cfs_bandwidth_slice());
4705 raw_spin_unlock(&cfs_b->lock);
4710 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
4712 /* dock delta_exec before expiring quota (as it could span periods) */
4713 cfs_rq->runtime_remaining -= delta_exec;
4715 if (likely(cfs_rq->runtime_remaining > 0))
4718 if (cfs_rq->throttled)
4721 * if we're unable to extend our runtime we resched so that the active
4722 * hierarchy can be throttled
4724 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
4725 resched_curr(rq_of(cfs_rq));
4728 static __always_inline
4729 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
4731 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
4734 __account_cfs_rq_runtime(cfs_rq, delta_exec);
4737 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
4739 return cfs_bandwidth_used() && cfs_rq->throttled;
4742 /* check whether cfs_rq, or any parent, is throttled */
4743 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
4745 return cfs_bandwidth_used() && cfs_rq->throttle_count;
4749 * Ensure that neither of the group entities corresponding to src_cpu or
4750 * dest_cpu are members of a throttled hierarchy when performing group
4751 * load-balance operations.
4753 static inline int throttled_lb_pair(struct task_group *tg,
4754 int src_cpu, int dest_cpu)
4756 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
4758 src_cfs_rq = tg->cfs_rq[src_cpu];
4759 dest_cfs_rq = tg->cfs_rq[dest_cpu];
4761 return throttled_hierarchy(src_cfs_rq) ||
4762 throttled_hierarchy(dest_cfs_rq);
4765 static int tg_unthrottle_up(struct task_group *tg, void *data)
4767 struct rq *rq = data;
4768 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
4770 cfs_rq->throttle_count--;
4771 if (!cfs_rq->throttle_count) {
4772 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
4773 cfs_rq->throttled_clock_task;
4775 /* Add cfs_rq with load or one or more already running entities to the list */
4776 if (!cfs_rq_is_decayed(cfs_rq) || cfs_rq->nr_running)
4777 list_add_leaf_cfs_rq(cfs_rq);
4783 static int tg_throttle_down(struct task_group *tg, void *data)
4785 struct rq *rq = data;
4786 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
4788 /* group is entering throttled state, stop time */
4789 if (!cfs_rq->throttle_count) {
4790 cfs_rq->throttled_clock_task = rq_clock_task(rq);
4791 list_del_leaf_cfs_rq(cfs_rq);
4793 cfs_rq->throttle_count++;
4798 static bool throttle_cfs_rq(struct cfs_rq *cfs_rq)
4800 struct rq *rq = rq_of(cfs_rq);
4801 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4802 struct sched_entity *se;
4803 long task_delta, idle_task_delta, dequeue = 1;
4805 raw_spin_lock(&cfs_b->lock);
4806 /* This will start the period timer if necessary */
4807 if (__assign_cfs_rq_runtime(cfs_b, cfs_rq, 1)) {
4809 * We have raced with bandwidth becoming available, and if we
4810 * actually throttled the timer might not unthrottle us for an
4811 * entire period. We additionally needed to make sure that any
4812 * subsequent check_cfs_rq_runtime calls agree not to throttle
4813 * us, as we may commit to do cfs put_prev+pick_next, so we ask
4814 * for 1ns of runtime rather than just check cfs_b.
4818 list_add_tail_rcu(&cfs_rq->throttled_list,
4819 &cfs_b->throttled_cfs_rq);
4821 raw_spin_unlock(&cfs_b->lock);
4824 return false; /* Throttle no longer required. */
4826 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
4828 /* freeze hierarchy runnable averages while throttled */
4830 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
4833 task_delta = cfs_rq->h_nr_running;
4834 idle_task_delta = cfs_rq->idle_h_nr_running;
4835 for_each_sched_entity(se) {
4836 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
4837 /* throttled entity or throttle-on-deactivate */
4841 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
4843 qcfs_rq->h_nr_running -= task_delta;
4844 qcfs_rq->idle_h_nr_running -= idle_task_delta;
4846 if (qcfs_rq->load.weight) {
4847 /* Avoid re-evaluating load for this entity: */
4848 se = parent_entity(se);
4853 for_each_sched_entity(se) {
4854 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
4855 /* throttled entity or throttle-on-deactivate */
4859 update_load_avg(qcfs_rq, se, 0);
4860 se_update_runnable(se);
4862 qcfs_rq->h_nr_running -= task_delta;
4863 qcfs_rq->idle_h_nr_running -= idle_task_delta;
4866 /* At this point se is NULL and we are at root level*/
4867 sub_nr_running(rq, task_delta);
4871 * Note: distribution will already see us throttled via the
4872 * throttled-list. rq->lock protects completion.
4874 cfs_rq->throttled = 1;
4875 cfs_rq->throttled_clock = rq_clock(rq);
4879 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
4881 struct rq *rq = rq_of(cfs_rq);
4882 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4883 struct sched_entity *se;
4884 long task_delta, idle_task_delta;
4886 se = cfs_rq->tg->se[cpu_of(rq)];
4888 cfs_rq->throttled = 0;
4890 update_rq_clock(rq);
4892 raw_spin_lock(&cfs_b->lock);
4893 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
4894 list_del_rcu(&cfs_rq->throttled_list);
4895 raw_spin_unlock(&cfs_b->lock);
4897 /* update hierarchical throttle state */
4898 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
4900 if (!cfs_rq->load.weight)
4903 task_delta = cfs_rq->h_nr_running;
4904 idle_task_delta = cfs_rq->idle_h_nr_running;
4905 for_each_sched_entity(se) {
4908 cfs_rq = cfs_rq_of(se);
4909 enqueue_entity(cfs_rq, se, ENQUEUE_WAKEUP);
4911 cfs_rq->h_nr_running += task_delta;
4912 cfs_rq->idle_h_nr_running += idle_task_delta;
4914 /* end evaluation on encountering a throttled cfs_rq */
4915 if (cfs_rq_throttled(cfs_rq))
4916 goto unthrottle_throttle;
4919 for_each_sched_entity(se) {
4920 cfs_rq = cfs_rq_of(se);
4922 update_load_avg(cfs_rq, se, UPDATE_TG);
4923 se_update_runnable(se);
4925 cfs_rq->h_nr_running += task_delta;
4926 cfs_rq->idle_h_nr_running += idle_task_delta;
4929 /* end evaluation on encountering a throttled cfs_rq */
4930 if (cfs_rq_throttled(cfs_rq))
4931 goto unthrottle_throttle;
4934 * One parent has been throttled and cfs_rq removed from the
4935 * list. Add it back to not break the leaf list.
4937 if (throttled_hierarchy(cfs_rq))
4938 list_add_leaf_cfs_rq(cfs_rq);
4941 /* At this point se is NULL and we are at root level*/
4942 add_nr_running(rq, task_delta);
4944 unthrottle_throttle:
4946 * The cfs_rq_throttled() breaks in the above iteration can result in
4947 * incomplete leaf list maintenance, resulting in triggering the
4950 for_each_sched_entity(se) {
4951 cfs_rq = cfs_rq_of(se);
4953 if (list_add_leaf_cfs_rq(cfs_rq))
4957 assert_list_leaf_cfs_rq(rq);
4959 /* Determine whether we need to wake up potentially idle CPU: */
4960 if (rq->curr == rq->idle && rq->cfs.nr_running)
4964 static void distribute_cfs_runtime(struct cfs_bandwidth *cfs_b)
4966 struct cfs_rq *cfs_rq;
4967 u64 runtime, remaining = 1;
4970 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
4972 struct rq *rq = rq_of(cfs_rq);
4975 rq_lock_irqsave(rq, &rf);
4976 if (!cfs_rq_throttled(cfs_rq))
4979 /* By the above check, this should never be true */
4980 SCHED_WARN_ON(cfs_rq->runtime_remaining > 0);
4982 raw_spin_lock(&cfs_b->lock);
4983 runtime = -cfs_rq->runtime_remaining + 1;
4984 if (runtime > cfs_b->runtime)
4985 runtime = cfs_b->runtime;
4986 cfs_b->runtime -= runtime;
4987 remaining = cfs_b->runtime;
4988 raw_spin_unlock(&cfs_b->lock);
4990 cfs_rq->runtime_remaining += runtime;
4992 /* we check whether we're throttled above */
4993 if (cfs_rq->runtime_remaining > 0)
4994 unthrottle_cfs_rq(cfs_rq);
4997 rq_unlock_irqrestore(rq, &rf);
5006 * Responsible for refilling a task_group's bandwidth and unthrottling its
5007 * cfs_rqs as appropriate. If there has been no activity within the last
5008 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
5009 * used to track this state.
5011 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun, unsigned long flags)
5015 /* no need to continue the timer with no bandwidth constraint */
5016 if (cfs_b->quota == RUNTIME_INF)
5017 goto out_deactivate;
5019 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
5020 cfs_b->nr_periods += overrun;
5022 /* Refill extra burst quota even if cfs_b->idle */
5023 __refill_cfs_bandwidth_runtime(cfs_b);
5026 * idle depends on !throttled (for the case of a large deficit), and if
5027 * we're going inactive then everything else can be deferred
5029 if (cfs_b->idle && !throttled)
5030 goto out_deactivate;
5033 /* mark as potentially idle for the upcoming period */
5038 /* account preceding periods in which throttling occurred */
5039 cfs_b->nr_throttled += overrun;
5042 * This check is repeated as we release cfs_b->lock while we unthrottle.
5044 while (throttled && cfs_b->runtime > 0) {
5045 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
5046 /* we can't nest cfs_b->lock while distributing bandwidth */
5047 distribute_cfs_runtime(cfs_b);
5048 raw_spin_lock_irqsave(&cfs_b->lock, flags);
5050 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
5054 * While we are ensured activity in the period following an
5055 * unthrottle, this also covers the case in which the new bandwidth is
5056 * insufficient to cover the existing bandwidth deficit. (Forcing the
5057 * timer to remain active while there are any throttled entities.)
5067 /* a cfs_rq won't donate quota below this amount */
5068 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
5069 /* minimum remaining period time to redistribute slack quota */
5070 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
5071 /* how long we wait to gather additional slack before distributing */
5072 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
5075 * Are we near the end of the current quota period?
5077 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
5078 * hrtimer base being cleared by hrtimer_start. In the case of
5079 * migrate_hrtimers, base is never cleared, so we are fine.
5081 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
5083 struct hrtimer *refresh_timer = &cfs_b->period_timer;
5086 /* if the call-back is running a quota refresh is already occurring */
5087 if (hrtimer_callback_running(refresh_timer))
5090 /* is a quota refresh about to occur? */
5091 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
5092 if (remaining < min_expire)
5098 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
5100 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
5102 /* if there's a quota refresh soon don't bother with slack */
5103 if (runtime_refresh_within(cfs_b, min_left))
5106 /* don't push forwards an existing deferred unthrottle */
5107 if (cfs_b->slack_started)
5109 cfs_b->slack_started = true;
5111 hrtimer_start(&cfs_b->slack_timer,
5112 ns_to_ktime(cfs_bandwidth_slack_period),
5116 /* we know any runtime found here is valid as update_curr() precedes return */
5117 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
5119 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
5120 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
5122 if (slack_runtime <= 0)
5125 raw_spin_lock(&cfs_b->lock);
5126 if (cfs_b->quota != RUNTIME_INF) {
5127 cfs_b->runtime += slack_runtime;
5129 /* we are under rq->lock, defer unthrottling using a timer */
5130 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
5131 !list_empty(&cfs_b->throttled_cfs_rq))
5132 start_cfs_slack_bandwidth(cfs_b);
5134 raw_spin_unlock(&cfs_b->lock);
5136 /* even if it's not valid for return we don't want to try again */
5137 cfs_rq->runtime_remaining -= slack_runtime;
5140 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
5142 if (!cfs_bandwidth_used())
5145 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
5148 __return_cfs_rq_runtime(cfs_rq);
5152 * This is done with a timer (instead of inline with bandwidth return) since
5153 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
5155 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
5157 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
5158 unsigned long flags;
5160 /* confirm we're still not at a refresh boundary */
5161 raw_spin_lock_irqsave(&cfs_b->lock, flags);
5162 cfs_b->slack_started = false;
5164 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
5165 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
5169 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice)
5170 runtime = cfs_b->runtime;
5172 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
5177 distribute_cfs_runtime(cfs_b);
5181 * When a group wakes up we want to make sure that its quota is not already
5182 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
5183 * runtime as update_curr() throttling can not trigger until it's on-rq.
5185 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
5187 if (!cfs_bandwidth_used())
5190 /* an active group must be handled by the update_curr()->put() path */
5191 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
5194 /* ensure the group is not already throttled */
5195 if (cfs_rq_throttled(cfs_rq))
5198 /* update runtime allocation */
5199 account_cfs_rq_runtime(cfs_rq, 0);
5200 if (cfs_rq->runtime_remaining <= 0)
5201 throttle_cfs_rq(cfs_rq);
5204 static void sync_throttle(struct task_group *tg, int cpu)
5206 struct cfs_rq *pcfs_rq, *cfs_rq;
5208 if (!cfs_bandwidth_used())
5214 cfs_rq = tg->cfs_rq[cpu];
5215 pcfs_rq = tg->parent->cfs_rq[cpu];
5217 cfs_rq->throttle_count = pcfs_rq->throttle_count;
5218 cfs_rq->throttled_clock_task = rq_clock_task(cpu_rq(cpu));
5221 /* conditionally throttle active cfs_rq's from put_prev_entity() */
5222 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
5224 if (!cfs_bandwidth_used())
5227 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
5231 * it's possible for a throttled entity to be forced into a running
5232 * state (e.g. set_curr_task), in this case we're finished.
5234 if (cfs_rq_throttled(cfs_rq))
5237 return throttle_cfs_rq(cfs_rq);
5240 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
5242 struct cfs_bandwidth *cfs_b =
5243 container_of(timer, struct cfs_bandwidth, slack_timer);
5245 do_sched_cfs_slack_timer(cfs_b);
5247 return HRTIMER_NORESTART;
5250 extern const u64 max_cfs_quota_period;
5252 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
5254 struct cfs_bandwidth *cfs_b =
5255 container_of(timer, struct cfs_bandwidth, period_timer);
5256 unsigned long flags;
5261 raw_spin_lock_irqsave(&cfs_b->lock, flags);
5263 overrun = hrtimer_forward_now(timer, cfs_b->period);
5267 idle = do_sched_cfs_period_timer(cfs_b, overrun, flags);
5270 u64 new, old = ktime_to_ns(cfs_b->period);
5273 * Grow period by a factor of 2 to avoid losing precision.
5274 * Precision loss in the quota/period ratio can cause __cfs_schedulable
5278 if (new < max_cfs_quota_period) {
5279 cfs_b->period = ns_to_ktime(new);
5283 pr_warn_ratelimited(
5284 "cfs_period_timer[cpu%d]: period too short, scaling up (new cfs_period_us = %lld, cfs_quota_us = %lld)\n",
5286 div_u64(new, NSEC_PER_USEC),
5287 div_u64(cfs_b->quota, NSEC_PER_USEC));
5289 pr_warn_ratelimited(
5290 "cfs_period_timer[cpu%d]: period too short, but cannot scale up without losing precision (cfs_period_us = %lld, cfs_quota_us = %lld)\n",
5292 div_u64(old, NSEC_PER_USEC),
5293 div_u64(cfs_b->quota, NSEC_PER_USEC));
5296 /* reset count so we don't come right back in here */
5301 cfs_b->period_active = 0;
5302 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
5304 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
5307 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
5309 raw_spin_lock_init(&cfs_b->lock);
5311 cfs_b->quota = RUNTIME_INF;
5312 cfs_b->period = ns_to_ktime(default_cfs_period());
5315 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
5316 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED);
5317 cfs_b->period_timer.function = sched_cfs_period_timer;
5318 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
5319 cfs_b->slack_timer.function = sched_cfs_slack_timer;
5320 cfs_b->slack_started = false;
5323 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
5325 cfs_rq->runtime_enabled = 0;
5326 INIT_LIST_HEAD(&cfs_rq->throttled_list);
5329 void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
5331 lockdep_assert_held(&cfs_b->lock);
5333 if (cfs_b->period_active)
5336 cfs_b->period_active = 1;
5337 hrtimer_forward_now(&cfs_b->period_timer, cfs_b->period);
5338 hrtimer_start_expires(&cfs_b->period_timer, HRTIMER_MODE_ABS_PINNED);
5341 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
5343 /* init_cfs_bandwidth() was not called */
5344 if (!cfs_b->throttled_cfs_rq.next)
5347 hrtimer_cancel(&cfs_b->period_timer);
5348 hrtimer_cancel(&cfs_b->slack_timer);
5352 * Both these CPU hotplug callbacks race against unregister_fair_sched_group()
5354 * The race is harmless, since modifying bandwidth settings of unhooked group
5355 * bits doesn't do much.
5358 /* cpu online callback */
5359 static void __maybe_unused update_runtime_enabled(struct rq *rq)
5361 struct task_group *tg;
5363 lockdep_assert_rq_held(rq);
5366 list_for_each_entry_rcu(tg, &task_groups, list) {
5367 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
5368 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
5370 raw_spin_lock(&cfs_b->lock);
5371 cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF;
5372 raw_spin_unlock(&cfs_b->lock);
5377 /* cpu offline callback */
5378 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
5380 struct task_group *tg;
5382 lockdep_assert_rq_held(rq);
5385 list_for_each_entry_rcu(tg, &task_groups, list) {
5386 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
5388 if (!cfs_rq->runtime_enabled)
5392 * clock_task is not advancing so we just need to make sure
5393 * there's some valid quota amount
5395 cfs_rq->runtime_remaining = 1;
5397 * Offline rq is schedulable till CPU is completely disabled
5398 * in take_cpu_down(), so we prevent new cfs throttling here.
5400 cfs_rq->runtime_enabled = 0;
5402 if (cfs_rq_throttled(cfs_rq))
5403 unthrottle_cfs_rq(cfs_rq);
5408 #else /* CONFIG_CFS_BANDWIDTH */
5410 static inline bool cfs_bandwidth_used(void)
5415 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {}
5416 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; }
5417 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
5418 static inline void sync_throttle(struct task_group *tg, int cpu) {}
5419 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
5421 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
5426 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
5431 static inline int throttled_lb_pair(struct task_group *tg,
5432 int src_cpu, int dest_cpu)
5437 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
5439 #ifdef CONFIG_FAIR_GROUP_SCHED
5440 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
5443 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
5447 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
5448 static inline void update_runtime_enabled(struct rq *rq) {}
5449 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
5451 #endif /* CONFIG_CFS_BANDWIDTH */
5453 /**************************************************
5454 * CFS operations on tasks:
5457 #ifdef CONFIG_SCHED_HRTICK
5458 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
5460 struct sched_entity *se = &p->se;
5461 struct cfs_rq *cfs_rq = cfs_rq_of(se);
5463 SCHED_WARN_ON(task_rq(p) != rq);
5465 if (rq->cfs.h_nr_running > 1) {
5466 u64 slice = sched_slice(cfs_rq, se);
5467 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
5468 s64 delta = slice - ran;
5471 if (task_current(rq, p))
5475 hrtick_start(rq, delta);
5480 * called from enqueue/dequeue and updates the hrtick when the
5481 * current task is from our class and nr_running is low enough
5484 static void hrtick_update(struct rq *rq)
5486 struct task_struct *curr = rq->curr;
5488 if (!hrtick_enabled_fair(rq) || curr->sched_class != &fair_sched_class)
5491 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
5492 hrtick_start_fair(rq, curr);
5494 #else /* !CONFIG_SCHED_HRTICK */
5496 hrtick_start_fair(struct rq *rq, struct task_struct *p)
5500 static inline void hrtick_update(struct rq *rq)
5506 static inline unsigned long cpu_util(int cpu);
5508 static inline bool cpu_overutilized(int cpu)
5510 return !fits_capacity(cpu_util(cpu), capacity_of(cpu));
5513 static inline void update_overutilized_status(struct rq *rq)
5515 if (!READ_ONCE(rq->rd->overutilized) && cpu_overutilized(rq->cpu)) {
5516 WRITE_ONCE(rq->rd->overutilized, SG_OVERUTILIZED);
5517 trace_sched_overutilized_tp(rq->rd, SG_OVERUTILIZED);
5521 static inline void update_overutilized_status(struct rq *rq) { }
5524 /* Runqueue only has SCHED_IDLE tasks enqueued */
5525 static int sched_idle_rq(struct rq *rq)
5527 return unlikely(rq->nr_running == rq->cfs.idle_h_nr_running &&
5532 static int sched_idle_cpu(int cpu)
5534 return sched_idle_rq(cpu_rq(cpu));
5539 * The enqueue_task method is called before nr_running is
5540 * increased. Here we update the fair scheduling stats and
5541 * then put the task into the rbtree:
5544 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
5546 struct cfs_rq *cfs_rq;
5547 struct sched_entity *se = &p->se;
5548 int idle_h_nr_running = task_has_idle_policy(p);
5549 int task_new = !(flags & ENQUEUE_WAKEUP);
5552 * The code below (indirectly) updates schedutil which looks at
5553 * the cfs_rq utilization to select a frequency.
5554 * Let's add the task's estimated utilization to the cfs_rq's
5555 * estimated utilization, before we update schedutil.
5557 util_est_enqueue(&rq->cfs, p);
5560 * If in_iowait is set, the code below may not trigger any cpufreq
5561 * utilization updates, so do it here explicitly with the IOWAIT flag
5565 cpufreq_update_util(rq, SCHED_CPUFREQ_IOWAIT);
5567 for_each_sched_entity(se) {
5570 cfs_rq = cfs_rq_of(se);
5571 enqueue_entity(cfs_rq, se, flags);
5573 cfs_rq->h_nr_running++;
5574 cfs_rq->idle_h_nr_running += idle_h_nr_running;
5576 /* end evaluation on encountering a throttled cfs_rq */
5577 if (cfs_rq_throttled(cfs_rq))
5578 goto enqueue_throttle;
5580 flags = ENQUEUE_WAKEUP;
5583 for_each_sched_entity(se) {
5584 cfs_rq = cfs_rq_of(se);
5586 update_load_avg(cfs_rq, se, UPDATE_TG);
5587 se_update_runnable(se);
5588 update_cfs_group(se);
5590 cfs_rq->h_nr_running++;
5591 cfs_rq->idle_h_nr_running += idle_h_nr_running;
5593 /* end evaluation on encountering a throttled cfs_rq */
5594 if (cfs_rq_throttled(cfs_rq))
5595 goto enqueue_throttle;
5598 * One parent has been throttled and cfs_rq removed from the
5599 * list. Add it back to not break the leaf list.
5601 if (throttled_hierarchy(cfs_rq))
5602 list_add_leaf_cfs_rq(cfs_rq);
5605 /* At this point se is NULL and we are at root level*/
5606 add_nr_running(rq, 1);
5609 * Since new tasks are assigned an initial util_avg equal to
5610 * half of the spare capacity of their CPU, tiny tasks have the
5611 * ability to cross the overutilized threshold, which will
5612 * result in the load balancer ruining all the task placement
5613 * done by EAS. As a way to mitigate that effect, do not account
5614 * for the first enqueue operation of new tasks during the
5615 * overutilized flag detection.
5617 * A better way of solving this problem would be to wait for
5618 * the PELT signals of tasks to converge before taking them
5619 * into account, but that is not straightforward to implement,
5620 * and the following generally works well enough in practice.
5623 update_overutilized_status(rq);
5626 if (cfs_bandwidth_used()) {
5628 * When bandwidth control is enabled; the cfs_rq_throttled()
5629 * breaks in the above iteration can result in incomplete
5630 * leaf list maintenance, resulting in triggering the assertion
5633 for_each_sched_entity(se) {
5634 cfs_rq = cfs_rq_of(se);
5636 if (list_add_leaf_cfs_rq(cfs_rq))
5641 assert_list_leaf_cfs_rq(rq);
5646 static void set_next_buddy(struct sched_entity *se);
5649 * The dequeue_task method is called before nr_running is
5650 * decreased. We remove the task from the rbtree and
5651 * update the fair scheduling stats:
5653 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
5655 struct cfs_rq *cfs_rq;
5656 struct sched_entity *se = &p->se;
5657 int task_sleep = flags & DEQUEUE_SLEEP;
5658 int idle_h_nr_running = task_has_idle_policy(p);
5659 bool was_sched_idle = sched_idle_rq(rq);
5661 util_est_dequeue(&rq->cfs, p);
5663 for_each_sched_entity(se) {
5664 cfs_rq = cfs_rq_of(se);
5665 dequeue_entity(cfs_rq, se, flags);
5667 cfs_rq->h_nr_running--;
5668 cfs_rq->idle_h_nr_running -= idle_h_nr_running;
5670 /* end evaluation on encountering a throttled cfs_rq */
5671 if (cfs_rq_throttled(cfs_rq))
5672 goto dequeue_throttle;
5674 /* Don't dequeue parent if it has other entities besides us */
5675 if (cfs_rq->load.weight) {
5676 /* Avoid re-evaluating load for this entity: */
5677 se = parent_entity(se);
5679 * Bias pick_next to pick a task from this cfs_rq, as
5680 * p is sleeping when it is within its sched_slice.
5682 if (task_sleep && se && !throttled_hierarchy(cfs_rq))
5686 flags |= DEQUEUE_SLEEP;
5689 for_each_sched_entity(se) {
5690 cfs_rq = cfs_rq_of(se);
5692 update_load_avg(cfs_rq, se, UPDATE_TG);
5693 se_update_runnable(se);
5694 update_cfs_group(se);
5696 cfs_rq->h_nr_running--;
5697 cfs_rq->idle_h_nr_running -= idle_h_nr_running;
5699 /* end evaluation on encountering a throttled cfs_rq */
5700 if (cfs_rq_throttled(cfs_rq))
5701 goto dequeue_throttle;
5705 /* At this point se is NULL and we are at root level*/
5706 sub_nr_running(rq, 1);
5708 /* balance early to pull high priority tasks */
5709 if (unlikely(!was_sched_idle && sched_idle_rq(rq)))
5710 rq->next_balance = jiffies;
5713 util_est_update(&rq->cfs, p, task_sleep);
5719 /* Working cpumask for: load_balance, load_balance_newidle. */
5720 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
5721 DEFINE_PER_CPU(cpumask_var_t, select_idle_mask);
5723 #ifdef CONFIG_NO_HZ_COMMON
5726 cpumask_var_t idle_cpus_mask;
5728 int has_blocked; /* Idle CPUS has blocked load */
5729 unsigned long next_balance; /* in jiffy units */
5730 unsigned long next_blocked; /* Next update of blocked load in jiffies */
5731 } nohz ____cacheline_aligned;
5733 #endif /* CONFIG_NO_HZ_COMMON */
5735 static unsigned long cpu_load(struct rq *rq)
5737 return cfs_rq_load_avg(&rq->cfs);
5741 * cpu_load_without - compute CPU load without any contributions from *p
5742 * @cpu: the CPU which load is requested
5743 * @p: the task which load should be discounted
5745 * The load of a CPU is defined by the load of tasks currently enqueued on that
5746 * CPU as well as tasks which are currently sleeping after an execution on that
5749 * This method returns the load of the specified CPU by discounting the load of
5750 * the specified task, whenever the task is currently contributing to the CPU
5753 static unsigned long cpu_load_without(struct rq *rq, struct task_struct *p)
5755 struct cfs_rq *cfs_rq;
5758 /* Task has no contribution or is new */
5759 if (cpu_of(rq) != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time))
5760 return cpu_load(rq);
5763 load = READ_ONCE(cfs_rq->avg.load_avg);
5765 /* Discount task's util from CPU's util */
5766 lsub_positive(&load, task_h_load(p));
5771 static unsigned long cpu_runnable(struct rq *rq)
5773 return cfs_rq_runnable_avg(&rq->cfs);
5776 static unsigned long cpu_runnable_without(struct rq *rq, struct task_struct *p)
5778 struct cfs_rq *cfs_rq;
5779 unsigned int runnable;
5781 /* Task has no contribution or is new */
5782 if (cpu_of(rq) != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time))
5783 return cpu_runnable(rq);
5786 runnable = READ_ONCE(cfs_rq->avg.runnable_avg);
5788 /* Discount task's runnable from CPU's runnable */
5789 lsub_positive(&runnable, p->se.avg.runnable_avg);
5794 static unsigned long capacity_of(int cpu)
5796 return cpu_rq(cpu)->cpu_capacity;
5799 static void record_wakee(struct task_struct *p)
5802 * Only decay a single time; tasks that have less then 1 wakeup per
5803 * jiffy will not have built up many flips.
5805 if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) {
5806 current->wakee_flips >>= 1;
5807 current->wakee_flip_decay_ts = jiffies;
5810 if (current->last_wakee != p) {
5811 current->last_wakee = p;
5812 current->wakee_flips++;
5817 * Detect M:N waker/wakee relationships via a switching-frequency heuristic.
5819 * A waker of many should wake a different task than the one last awakened
5820 * at a frequency roughly N times higher than one of its wakees.
5822 * In order to determine whether we should let the load spread vs consolidating
5823 * to shared cache, we look for a minimum 'flip' frequency of llc_size in one
5824 * partner, and a factor of lls_size higher frequency in the other.
5826 * With both conditions met, we can be relatively sure that the relationship is
5827 * non-monogamous, with partner count exceeding socket size.
5829 * Waker/wakee being client/server, worker/dispatcher, interrupt source or
5830 * whatever is irrelevant, spread criteria is apparent partner count exceeds
5833 static int wake_wide(struct task_struct *p)
5835 unsigned int master = current->wakee_flips;
5836 unsigned int slave = p->wakee_flips;
5837 int factor = __this_cpu_read(sd_llc_size);
5840 swap(master, slave);
5841 if (slave < factor || master < slave * factor)
5847 * The purpose of wake_affine() is to quickly determine on which CPU we can run
5848 * soonest. For the purpose of speed we only consider the waking and previous
5851 * wake_affine_idle() - only considers 'now', it check if the waking CPU is
5852 * cache-affine and is (or will be) idle.
5854 * wake_affine_weight() - considers the weight to reflect the average
5855 * scheduling latency of the CPUs. This seems to work
5856 * for the overloaded case.
5859 wake_affine_idle(int this_cpu, int prev_cpu, int sync)
5862 * If this_cpu is idle, it implies the wakeup is from interrupt
5863 * context. Only allow the move if cache is shared. Otherwise an
5864 * interrupt intensive workload could force all tasks onto one
5865 * node depending on the IO topology or IRQ affinity settings.
5867 * If the prev_cpu is idle and cache affine then avoid a migration.
5868 * There is no guarantee that the cache hot data from an interrupt
5869 * is more important than cache hot data on the prev_cpu and from
5870 * a cpufreq perspective, it's better to have higher utilisation
5873 if (available_idle_cpu(this_cpu) && cpus_share_cache(this_cpu, prev_cpu))
5874 return available_idle_cpu(prev_cpu) ? prev_cpu : this_cpu;
5876 if (sync && cpu_rq(this_cpu)->nr_running == 1)
5879 if (available_idle_cpu(prev_cpu))
5882 return nr_cpumask_bits;
5886 wake_affine_weight(struct sched_domain *sd, struct task_struct *p,
5887 int this_cpu, int prev_cpu, int sync)
5889 s64 this_eff_load, prev_eff_load;
5890 unsigned long task_load;
5892 this_eff_load = cpu_load(cpu_rq(this_cpu));
5895 unsigned long current_load = task_h_load(current);
5897 if (current_load > this_eff_load)
5900 this_eff_load -= current_load;
5903 task_load = task_h_load(p);
5905 this_eff_load += task_load;
5906 if (sched_feat(WA_BIAS))
5907 this_eff_load *= 100;
5908 this_eff_load *= capacity_of(prev_cpu);
5910 prev_eff_load = cpu_load(cpu_rq(prev_cpu));
5911 prev_eff_load -= task_load;
5912 if (sched_feat(WA_BIAS))
5913 prev_eff_load *= 100 + (sd->imbalance_pct - 100) / 2;
5914 prev_eff_load *= capacity_of(this_cpu);
5917 * If sync, adjust the weight of prev_eff_load such that if
5918 * prev_eff == this_eff that select_idle_sibling() will consider
5919 * stacking the wakee on top of the waker if no other CPU is
5925 return this_eff_load < prev_eff_load ? this_cpu : nr_cpumask_bits;
5928 static int wake_affine(struct sched_domain *sd, struct task_struct *p,
5929 int this_cpu, int prev_cpu, int sync)
5931 int target = nr_cpumask_bits;
5933 if (sched_feat(WA_IDLE))
5934 target = wake_affine_idle(this_cpu, prev_cpu, sync);
5936 if (sched_feat(WA_WEIGHT) && target == nr_cpumask_bits)
5937 target = wake_affine_weight(sd, p, this_cpu, prev_cpu, sync);
5939 schedstat_inc(p->se.statistics.nr_wakeups_affine_attempts);
5940 if (target == nr_cpumask_bits)
5943 schedstat_inc(sd->ttwu_move_affine);
5944 schedstat_inc(p->se.statistics.nr_wakeups_affine);
5948 static struct sched_group *
5949 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu);
5952 * find_idlest_group_cpu - find the idlest CPU among the CPUs in the group.
5955 find_idlest_group_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
5957 unsigned long load, min_load = ULONG_MAX;
5958 unsigned int min_exit_latency = UINT_MAX;
5959 u64 latest_idle_timestamp = 0;
5960 int least_loaded_cpu = this_cpu;
5961 int shallowest_idle_cpu = -1;
5964 /* Check if we have any choice: */
5965 if (group->group_weight == 1)
5966 return cpumask_first(sched_group_span(group));
5968 /* Traverse only the allowed CPUs */
5969 for_each_cpu_and(i, sched_group_span(group), p->cpus_ptr) {
5970 struct rq *rq = cpu_rq(i);
5972 if (!sched_core_cookie_match(rq, p))
5975 if (sched_idle_cpu(i))
5978 if (available_idle_cpu(i)) {
5979 struct cpuidle_state *idle = idle_get_state(rq);
5980 if (idle && idle->exit_latency < min_exit_latency) {
5982 * We give priority to a CPU whose idle state
5983 * has the smallest exit latency irrespective
5984 * of any idle timestamp.
5986 min_exit_latency = idle->exit_latency;
5987 latest_idle_timestamp = rq->idle_stamp;
5988 shallowest_idle_cpu = i;
5989 } else if ((!idle || idle->exit_latency == min_exit_latency) &&
5990 rq->idle_stamp > latest_idle_timestamp) {
5992 * If equal or no active idle state, then
5993 * the most recently idled CPU might have
5996 latest_idle_timestamp = rq->idle_stamp;
5997 shallowest_idle_cpu = i;
5999 } else if (shallowest_idle_cpu == -1) {
6000 load = cpu_load(cpu_rq(i));
6001 if (load < min_load) {
6003 least_loaded_cpu = i;
6008 return shallowest_idle_cpu != -1 ? shallowest_idle_cpu : least_loaded_cpu;
6011 static inline int find_idlest_cpu(struct sched_domain *sd, struct task_struct *p,
6012 int cpu, int prev_cpu, int sd_flag)
6016 if (!cpumask_intersects(sched_domain_span(sd), p->cpus_ptr))
6020 * We need task's util for cpu_util_without, sync it up to
6021 * prev_cpu's last_update_time.
6023 if (!(sd_flag & SD_BALANCE_FORK))
6024 sync_entity_load_avg(&p->se);
6027 struct sched_group *group;
6028 struct sched_domain *tmp;
6031 if (!(sd->flags & sd_flag)) {
6036 group = find_idlest_group(sd, p, cpu);
6042 new_cpu = find_idlest_group_cpu(group, p, cpu);
6043 if (new_cpu == cpu) {
6044 /* Now try balancing at a lower domain level of 'cpu': */
6049 /* Now try balancing at a lower domain level of 'new_cpu': */
6051 weight = sd->span_weight;
6053 for_each_domain(cpu, tmp) {
6054 if (weight <= tmp->span_weight)
6056 if (tmp->flags & sd_flag)
6064 static inline int __select_idle_cpu(int cpu, struct task_struct *p)
6066 if ((available_idle_cpu(cpu) || sched_idle_cpu(cpu)) &&
6067 sched_cpu_cookie_match(cpu_rq(cpu), p))
6073 #ifdef CONFIG_SCHED_SMT
6074 DEFINE_STATIC_KEY_FALSE(sched_smt_present);
6075 EXPORT_SYMBOL_GPL(sched_smt_present);
6077 static inline void set_idle_cores(int cpu, int val)
6079 struct sched_domain_shared *sds;
6081 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
6083 WRITE_ONCE(sds->has_idle_cores, val);
6086 static inline bool test_idle_cores(int cpu, bool def)
6088 struct sched_domain_shared *sds;
6090 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
6092 return READ_ONCE(sds->has_idle_cores);
6098 * Scans the local SMT mask to see if the entire core is idle, and records this
6099 * information in sd_llc_shared->has_idle_cores.
6101 * Since SMT siblings share all cache levels, inspecting this limited remote
6102 * state should be fairly cheap.
6104 void __update_idle_core(struct rq *rq)
6106 int core = cpu_of(rq);
6110 if (test_idle_cores(core, true))
6113 for_each_cpu(cpu, cpu_smt_mask(core)) {
6117 if (!available_idle_cpu(cpu))
6121 set_idle_cores(core, 1);
6127 * Scan the entire LLC domain for idle cores; this dynamically switches off if
6128 * there are no idle cores left in the system; tracked through
6129 * sd_llc->shared->has_idle_cores and enabled through update_idle_core() above.
6131 static int select_idle_core(struct task_struct *p, int core, struct cpumask *cpus, int *idle_cpu)
6136 if (!static_branch_likely(&sched_smt_present))
6137 return __select_idle_cpu(core, p);
6139 for_each_cpu(cpu, cpu_smt_mask(core)) {
6140 if (!available_idle_cpu(cpu)) {
6142 if (*idle_cpu == -1) {
6143 if (sched_idle_cpu(cpu) && cpumask_test_cpu(cpu, p->cpus_ptr)) {
6151 if (*idle_cpu == -1 && cpumask_test_cpu(cpu, p->cpus_ptr))
6158 cpumask_andnot(cpus, cpus, cpu_smt_mask(core));
6163 * Scan the local SMT mask for idle CPUs.
6165 static int select_idle_smt(struct task_struct *p, struct sched_domain *sd, int target)
6169 for_each_cpu(cpu, cpu_smt_mask(target)) {
6170 if (!cpumask_test_cpu(cpu, p->cpus_ptr) ||
6171 !cpumask_test_cpu(cpu, sched_domain_span(sd)))
6173 if (available_idle_cpu(cpu) || sched_idle_cpu(cpu))
6180 #else /* CONFIG_SCHED_SMT */
6182 static inline void set_idle_cores(int cpu, int val)
6186 static inline bool test_idle_cores(int cpu, bool def)
6191 static inline int select_idle_core(struct task_struct *p, int core, struct cpumask *cpus, int *idle_cpu)
6193 return __select_idle_cpu(core, p);
6196 static inline int select_idle_smt(struct task_struct *p, struct sched_domain *sd, int target)
6201 #endif /* CONFIG_SCHED_SMT */
6204 * Scan the LLC domain for idle CPUs; this is dynamically regulated by
6205 * comparing the average scan cost (tracked in sd->avg_scan_cost) against the
6206 * average idle time for this rq (as found in rq->avg_idle).
6208 static int select_idle_cpu(struct task_struct *p, struct sched_domain *sd, bool has_idle_core, int target)
6210 struct cpumask *cpus = this_cpu_cpumask_var_ptr(select_idle_mask);
6211 int i, cpu, idle_cpu = -1, nr = INT_MAX;
6212 struct rq *this_rq = this_rq();
6213 int this = smp_processor_id();
6214 struct sched_domain *this_sd;
6217 this_sd = rcu_dereference(*this_cpu_ptr(&sd_llc));
6221 cpumask_and(cpus, sched_domain_span(sd), p->cpus_ptr);
6223 if (sched_feat(SIS_PROP) && !has_idle_core) {
6224 u64 avg_cost, avg_idle, span_avg;
6225 unsigned long now = jiffies;
6228 * If we're busy, the assumption that the last idle period
6229 * predicts the future is flawed; age away the remaining
6230 * predicted idle time.
6232 if (unlikely(this_rq->wake_stamp < now)) {
6233 while (this_rq->wake_stamp < now && this_rq->wake_avg_idle) {
6234 this_rq->wake_stamp++;
6235 this_rq->wake_avg_idle >>= 1;
6239 avg_idle = this_rq->wake_avg_idle;
6240 avg_cost = this_sd->avg_scan_cost + 1;
6242 span_avg = sd->span_weight * avg_idle;
6243 if (span_avg > 4*avg_cost)
6244 nr = div_u64(span_avg, avg_cost);
6248 time = cpu_clock(this);
6251 for_each_cpu_wrap(cpu, cpus, target) {
6252 if (has_idle_core) {
6253 i = select_idle_core(p, cpu, cpus, &idle_cpu);
6254 if ((unsigned int)i < nr_cpumask_bits)
6260 idle_cpu = __select_idle_cpu(cpu, p);
6261 if ((unsigned int)idle_cpu < nr_cpumask_bits)
6267 set_idle_cores(target, false);
6269 if (sched_feat(SIS_PROP) && !has_idle_core) {
6270 time = cpu_clock(this) - time;
6273 * Account for the scan cost of wakeups against the average
6276 this_rq->wake_avg_idle -= min(this_rq->wake_avg_idle, time);
6278 update_avg(&this_sd->avg_scan_cost, time);
6285 * Scan the asym_capacity domain for idle CPUs; pick the first idle one on which
6286 * the task fits. If no CPU is big enough, but there are idle ones, try to
6287 * maximize capacity.
6290 select_idle_capacity(struct task_struct *p, struct sched_domain *sd, int target)
6292 unsigned long task_util, best_cap = 0;
6293 int cpu, best_cpu = -1;
6294 struct cpumask *cpus;
6296 cpus = this_cpu_cpumask_var_ptr(select_idle_mask);
6297 cpumask_and(cpus, sched_domain_span(sd), p->cpus_ptr);
6299 task_util = uclamp_task_util(p);
6301 for_each_cpu_wrap(cpu, cpus, target) {
6302 unsigned long cpu_cap = capacity_of(cpu);
6304 if (!available_idle_cpu(cpu) && !sched_idle_cpu(cpu))
6306 if (fits_capacity(task_util, cpu_cap))
6309 if (cpu_cap > best_cap) {
6318 static inline bool asym_fits_capacity(int task_util, int cpu)
6320 if (static_branch_unlikely(&sched_asym_cpucapacity))
6321 return fits_capacity(task_util, capacity_of(cpu));
6327 * Try and locate an idle core/thread in the LLC cache domain.
6329 static int select_idle_sibling(struct task_struct *p, int prev, int target)
6331 bool has_idle_core = false;
6332 struct sched_domain *sd;
6333 unsigned long task_util;
6334 int i, recent_used_cpu;
6337 * On asymmetric system, update task utilization because we will check
6338 * that the task fits with cpu's capacity.
6340 if (static_branch_unlikely(&sched_asym_cpucapacity)) {
6341 sync_entity_load_avg(&p->se);
6342 task_util = uclamp_task_util(p);
6346 * per-cpu select_idle_mask usage
6348 lockdep_assert_irqs_disabled();
6350 if ((available_idle_cpu(target) || sched_idle_cpu(target)) &&
6351 asym_fits_capacity(task_util, target))
6355 * If the previous CPU is cache affine and idle, don't be stupid:
6357 if (prev != target && cpus_share_cache(prev, target) &&
6358 (available_idle_cpu(prev) || sched_idle_cpu(prev)) &&
6359 asym_fits_capacity(task_util, prev))
6363 * Allow a per-cpu kthread to stack with the wakee if the
6364 * kworker thread and the tasks previous CPUs are the same.
6365 * The assumption is that the wakee queued work for the
6366 * per-cpu kthread that is now complete and the wakeup is
6367 * essentially a sync wakeup. An obvious example of this
6368 * pattern is IO completions.
6370 if (is_per_cpu_kthread(current) &&
6371 prev == smp_processor_id() &&
6372 this_rq()->nr_running <= 1) {
6376 /* Check a recently used CPU as a potential idle candidate: */
6377 recent_used_cpu = p->recent_used_cpu;
6378 if (recent_used_cpu != prev &&
6379 recent_used_cpu != target &&
6380 cpus_share_cache(recent_used_cpu, target) &&
6381 (available_idle_cpu(recent_used_cpu) || sched_idle_cpu(recent_used_cpu)) &&
6382 cpumask_test_cpu(p->recent_used_cpu, p->cpus_ptr) &&
6383 asym_fits_capacity(task_util, recent_used_cpu)) {
6385 * Replace recent_used_cpu with prev as it is a potential
6386 * candidate for the next wake:
6388 p->recent_used_cpu = prev;
6389 return recent_used_cpu;
6393 * For asymmetric CPU capacity systems, our domain of interest is
6394 * sd_asym_cpucapacity rather than sd_llc.
6396 if (static_branch_unlikely(&sched_asym_cpucapacity)) {
6397 sd = rcu_dereference(per_cpu(sd_asym_cpucapacity, target));
6399 * On an asymmetric CPU capacity system where an exclusive
6400 * cpuset defines a symmetric island (i.e. one unique
6401 * capacity_orig value through the cpuset), the key will be set
6402 * but the CPUs within that cpuset will not have a domain with
6403 * SD_ASYM_CPUCAPACITY. These should follow the usual symmetric
6407 i = select_idle_capacity(p, sd, target);
6408 return ((unsigned)i < nr_cpumask_bits) ? i : target;
6412 sd = rcu_dereference(per_cpu(sd_llc, target));
6416 if (sched_smt_active()) {
6417 has_idle_core = test_idle_cores(target, false);
6419 if (!has_idle_core && cpus_share_cache(prev, target)) {
6420 i = select_idle_smt(p, sd, prev);
6421 if ((unsigned int)i < nr_cpumask_bits)
6426 i = select_idle_cpu(p, sd, has_idle_core, target);
6427 if ((unsigned)i < nr_cpumask_bits)
6434 * cpu_util - Estimates the amount of capacity of a CPU used by CFS tasks.
6435 * @cpu: the CPU to get the utilization of
6437 * The unit of the return value must be the one of capacity so we can compare
6438 * the utilization with the capacity of the CPU that is available for CFS task
6439 * (ie cpu_capacity).
6441 * cfs_rq.avg.util_avg is the sum of running time of runnable tasks plus the
6442 * recent utilization of currently non-runnable tasks on a CPU. It represents
6443 * the amount of utilization of a CPU in the range [0..capacity_orig] where
6444 * capacity_orig is the cpu_capacity available at the highest frequency
6445 * (arch_scale_freq_capacity()).
6446 * The utilization of a CPU converges towards a sum equal to or less than the
6447 * current capacity (capacity_curr <= capacity_orig) of the CPU because it is
6448 * the running time on this CPU scaled by capacity_curr.
6450 * The estimated utilization of a CPU is defined to be the maximum between its
6451 * cfs_rq.avg.util_avg and the sum of the estimated utilization of the tasks
6452 * currently RUNNABLE on that CPU.
6453 * This allows to properly represent the expected utilization of a CPU which
6454 * has just got a big task running since a long sleep period. At the same time
6455 * however it preserves the benefits of the "blocked utilization" in
6456 * describing the potential for other tasks waking up on the same CPU.
6458 * Nevertheless, cfs_rq.avg.util_avg can be higher than capacity_curr or even
6459 * higher than capacity_orig because of unfortunate rounding in
6460 * cfs.avg.util_avg or just after migrating tasks and new task wakeups until
6461 * the average stabilizes with the new running time. We need to check that the
6462 * utilization stays within the range of [0..capacity_orig] and cap it if
6463 * necessary. Without utilization capping, a group could be seen as overloaded
6464 * (CPU0 utilization at 121% + CPU1 utilization at 80%) whereas CPU1 has 20% of
6465 * available capacity. We allow utilization to overshoot capacity_curr (but not
6466 * capacity_orig) as it useful for predicting the capacity required after task
6467 * migrations (scheduler-driven DVFS).
6469 * Return: the (estimated) utilization for the specified CPU
6471 static inline unsigned long cpu_util(int cpu)
6473 struct cfs_rq *cfs_rq;
6476 cfs_rq = &cpu_rq(cpu)->cfs;
6477 util = READ_ONCE(cfs_rq->avg.util_avg);
6479 if (sched_feat(UTIL_EST))
6480 util = max(util, READ_ONCE(cfs_rq->avg.util_est.enqueued));
6482 return min_t(unsigned long, util, capacity_orig_of(cpu));
6486 * cpu_util_without: compute cpu utilization without any contributions from *p
6487 * @cpu: the CPU which utilization is requested
6488 * @p: the task which utilization should be discounted
6490 * The utilization of a CPU is defined by the utilization of tasks currently
6491 * enqueued on that CPU as well as tasks which are currently sleeping after an
6492 * execution on that CPU.
6494 * This method returns the utilization of the specified CPU by discounting the
6495 * utilization of the specified task, whenever the task is currently
6496 * contributing to the CPU utilization.
6498 static unsigned long cpu_util_without(int cpu, struct task_struct *p)
6500 struct cfs_rq *cfs_rq;
6503 /* Task has no contribution or is new */
6504 if (cpu != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time))
6505 return cpu_util(cpu);
6507 cfs_rq = &cpu_rq(cpu)->cfs;
6508 util = READ_ONCE(cfs_rq->avg.util_avg);
6510 /* Discount task's util from CPU's util */
6511 lsub_positive(&util, task_util(p));
6516 * a) if *p is the only task sleeping on this CPU, then:
6517 * cpu_util (== task_util) > util_est (== 0)
6518 * and thus we return:
6519 * cpu_util_without = (cpu_util - task_util) = 0
6521 * b) if other tasks are SLEEPING on this CPU, which is now exiting
6523 * cpu_util >= task_util
6524 * cpu_util > util_est (== 0)
6525 * and thus we discount *p's blocked utilization to return:
6526 * cpu_util_without = (cpu_util - task_util) >= 0
6528 * c) if other tasks are RUNNABLE on that CPU and
6529 * util_est > cpu_util
6530 * then we use util_est since it returns a more restrictive
6531 * estimation of the spare capacity on that CPU, by just
6532 * considering the expected utilization of tasks already
6533 * runnable on that CPU.
6535 * Cases a) and b) are covered by the above code, while case c) is
6536 * covered by the following code when estimated utilization is
6539 if (sched_feat(UTIL_EST)) {
6540 unsigned int estimated =
6541 READ_ONCE(cfs_rq->avg.util_est.enqueued);
6544 * Despite the following checks we still have a small window
6545 * for a possible race, when an execl's select_task_rq_fair()
6546 * races with LB's detach_task():
6549 * p->on_rq = TASK_ON_RQ_MIGRATING;
6550 * ---------------------------------- A
6551 * deactivate_task() \
6552 * dequeue_task() + RaceTime
6553 * util_est_dequeue() /
6554 * ---------------------------------- B
6556 * The additional check on "current == p" it's required to
6557 * properly fix the execl regression and it helps in further
6558 * reducing the chances for the above race.
6560 if (unlikely(task_on_rq_queued(p) || current == p))
6561 lsub_positive(&estimated, _task_util_est(p));
6563 util = max(util, estimated);
6567 * Utilization (estimated) can exceed the CPU capacity, thus let's
6568 * clamp to the maximum CPU capacity to ensure consistency with
6569 * the cpu_util call.
6571 return min_t(unsigned long, util, capacity_orig_of(cpu));
6575 * Predicts what cpu_util(@cpu) would return if @p was migrated (and enqueued)
6578 static unsigned long cpu_util_next(int cpu, struct task_struct *p, int dst_cpu)
6580 struct cfs_rq *cfs_rq = &cpu_rq(cpu)->cfs;
6581 unsigned long util_est, util = READ_ONCE(cfs_rq->avg.util_avg);
6584 * If @p migrates from @cpu to another, remove its contribution. Or,
6585 * if @p migrates from another CPU to @cpu, add its contribution. In
6586 * the other cases, @cpu is not impacted by the migration, so the
6587 * util_avg should already be correct.
6589 if (task_cpu(p) == cpu && dst_cpu != cpu)
6590 lsub_positive(&util, task_util(p));
6591 else if (task_cpu(p) != cpu && dst_cpu == cpu)
6592 util += task_util(p);
6594 if (sched_feat(UTIL_EST)) {
6595 util_est = READ_ONCE(cfs_rq->avg.util_est.enqueued);
6598 * During wake-up, the task isn't enqueued yet and doesn't
6599 * appear in the cfs_rq->avg.util_est.enqueued of any rq,
6600 * so just add it (if needed) to "simulate" what will be
6601 * cpu_util() after the task has been enqueued.
6604 util_est += _task_util_est(p);
6606 util = max(util, util_est);
6609 return min(util, capacity_orig_of(cpu));
6613 * compute_energy(): Estimates the energy that @pd would consume if @p was
6614 * migrated to @dst_cpu. compute_energy() predicts what will be the utilization
6615 * landscape of @pd's CPUs after the task migration, and uses the Energy Model
6616 * to compute what would be the energy if we decided to actually migrate that
6620 compute_energy(struct task_struct *p, int dst_cpu, struct perf_domain *pd)
6622 struct cpumask *pd_mask = perf_domain_span(pd);
6623 unsigned long cpu_cap = arch_scale_cpu_capacity(cpumask_first(pd_mask));
6624 unsigned long max_util = 0, sum_util = 0;
6625 unsigned long _cpu_cap = cpu_cap;
6628 _cpu_cap -= arch_scale_thermal_pressure(cpumask_first(pd_mask));
6631 * The capacity state of CPUs of the current rd can be driven by CPUs
6632 * of another rd if they belong to the same pd. So, account for the
6633 * utilization of these CPUs too by masking pd with cpu_online_mask
6634 * instead of the rd span.
6636 * If an entire pd is outside of the current rd, it will not appear in
6637 * its pd list and will not be accounted by compute_energy().
6639 for_each_cpu_and(cpu, pd_mask, cpu_online_mask) {
6640 unsigned long util_freq = cpu_util_next(cpu, p, dst_cpu);
6641 unsigned long cpu_util, util_running = util_freq;
6642 struct task_struct *tsk = NULL;
6645 * When @p is placed on @cpu:
6647 * util_running = max(cpu_util, cpu_util_est) +
6648 * max(task_util, _task_util_est)
6650 * while cpu_util_next is: max(cpu_util + task_util,
6651 * cpu_util_est + _task_util_est)
6653 if (cpu == dst_cpu) {
6656 cpu_util_next(cpu, p, -1) + task_util_est(p);
6660 * Busy time computation: utilization clamping is not
6661 * required since the ratio (sum_util / cpu_capacity)
6662 * is already enough to scale the EM reported power
6663 * consumption at the (eventually clamped) cpu_capacity.
6665 cpu_util = effective_cpu_util(cpu, util_running, cpu_cap,
6668 sum_util += min(cpu_util, _cpu_cap);
6671 * Performance domain frequency: utilization clamping
6672 * must be considered since it affects the selection
6673 * of the performance domain frequency.
6674 * NOTE: in case RT tasks are running, by default the
6675 * FREQUENCY_UTIL's utilization can be max OPP.
6677 cpu_util = effective_cpu_util(cpu, util_freq, cpu_cap,
6678 FREQUENCY_UTIL, tsk);
6679 max_util = max(max_util, min(cpu_util, _cpu_cap));
6682 return em_cpu_energy(pd->em_pd, max_util, sum_util, _cpu_cap);
6686 * find_energy_efficient_cpu(): Find most energy-efficient target CPU for the
6687 * waking task. find_energy_efficient_cpu() looks for the CPU with maximum
6688 * spare capacity in each performance domain and uses it as a potential
6689 * candidate to execute the task. Then, it uses the Energy Model to figure
6690 * out which of the CPU candidates is the most energy-efficient.
6692 * The rationale for this heuristic is as follows. In a performance domain,
6693 * all the most energy efficient CPU candidates (according to the Energy
6694 * Model) are those for which we'll request a low frequency. When there are
6695 * several CPUs for which the frequency request will be the same, we don't
6696 * have enough data to break the tie between them, because the Energy Model
6697 * only includes active power costs. With this model, if we assume that
6698 * frequency requests follow utilization (e.g. using schedutil), the CPU with
6699 * the maximum spare capacity in a performance domain is guaranteed to be among
6700 * the best candidates of the performance domain.
6702 * In practice, it could be preferable from an energy standpoint to pack
6703 * small tasks on a CPU in order to let other CPUs go in deeper idle states,
6704 * but that could also hurt our chances to go cluster idle, and we have no
6705 * ways to tell with the current Energy Model if this is actually a good
6706 * idea or not. So, find_energy_efficient_cpu() basically favors
6707 * cluster-packing, and spreading inside a cluster. That should at least be
6708 * a good thing for latency, and this is consistent with the idea that most
6709 * of the energy savings of EAS come from the asymmetry of the system, and
6710 * not so much from breaking the tie between identical CPUs. That's also the
6711 * reason why EAS is enabled in the topology code only for systems where
6712 * SD_ASYM_CPUCAPACITY is set.
6714 * NOTE: Forkees are not accepted in the energy-aware wake-up path because
6715 * they don't have any useful utilization data yet and it's not possible to
6716 * forecast their impact on energy consumption. Consequently, they will be
6717 * placed by find_idlest_cpu() on the least loaded CPU, which might turn out
6718 * to be energy-inefficient in some use-cases. The alternative would be to
6719 * bias new tasks towards specific types of CPUs first, or to try to infer
6720 * their util_avg from the parent task, but those heuristics could hurt
6721 * other use-cases too. So, until someone finds a better way to solve this,
6722 * let's keep things simple by re-using the existing slow path.
6724 static int find_energy_efficient_cpu(struct task_struct *p, int prev_cpu)
6726 unsigned long prev_delta = ULONG_MAX, best_delta = ULONG_MAX;
6727 struct root_domain *rd = cpu_rq(smp_processor_id())->rd;
6728 int cpu, best_energy_cpu = prev_cpu, target = -1;
6729 unsigned long cpu_cap, util, base_energy = 0;
6730 struct sched_domain *sd;
6731 struct perf_domain *pd;
6734 pd = rcu_dereference(rd->pd);
6735 if (!pd || READ_ONCE(rd->overutilized))
6739 * Energy-aware wake-up happens on the lowest sched_domain starting
6740 * from sd_asym_cpucapacity spanning over this_cpu and prev_cpu.
6742 sd = rcu_dereference(*this_cpu_ptr(&sd_asym_cpucapacity));
6743 while (sd && !cpumask_test_cpu(prev_cpu, sched_domain_span(sd)))
6750 sync_entity_load_avg(&p->se);
6751 if (!task_util_est(p))
6754 for (; pd; pd = pd->next) {
6755 unsigned long cur_delta, spare_cap, max_spare_cap = 0;
6756 bool compute_prev_delta = false;
6757 unsigned long base_energy_pd;
6758 int max_spare_cap_cpu = -1;
6760 for_each_cpu_and(cpu, perf_domain_span(pd), sched_domain_span(sd)) {
6761 if (!cpumask_test_cpu(cpu, p->cpus_ptr))
6764 util = cpu_util_next(cpu, p, cpu);
6765 cpu_cap = capacity_of(cpu);
6766 spare_cap = cpu_cap;
6767 lsub_positive(&spare_cap, util);
6770 * Skip CPUs that cannot satisfy the capacity request.
6771 * IOW, placing the task there would make the CPU
6772 * overutilized. Take uclamp into account to see how
6773 * much capacity we can get out of the CPU; this is
6774 * aligned with sched_cpu_util().
6776 util = uclamp_rq_util_with(cpu_rq(cpu), util, p);
6777 if (!fits_capacity(util, cpu_cap))
6780 if (cpu == prev_cpu) {
6781 /* Always use prev_cpu as a candidate. */
6782 compute_prev_delta = true;
6783 } else if (spare_cap > max_spare_cap) {
6785 * Find the CPU with the maximum spare capacity
6786 * in the performance domain.
6788 max_spare_cap = spare_cap;
6789 max_spare_cap_cpu = cpu;
6793 if (max_spare_cap_cpu < 0 && !compute_prev_delta)
6796 /* Compute the 'base' energy of the pd, without @p */
6797 base_energy_pd = compute_energy(p, -1, pd);
6798 base_energy += base_energy_pd;
6800 /* Evaluate the energy impact of using prev_cpu. */
6801 if (compute_prev_delta) {
6802 prev_delta = compute_energy(p, prev_cpu, pd);
6803 if (prev_delta < base_energy_pd)
6805 prev_delta -= base_energy_pd;
6806 best_delta = min(best_delta, prev_delta);
6809 /* Evaluate the energy impact of using max_spare_cap_cpu. */
6810 if (max_spare_cap_cpu >= 0) {
6811 cur_delta = compute_energy(p, max_spare_cap_cpu, pd);
6812 if (cur_delta < base_energy_pd)
6814 cur_delta -= base_energy_pd;
6815 if (cur_delta < best_delta) {
6816 best_delta = cur_delta;
6817 best_energy_cpu = max_spare_cap_cpu;
6824 * Pick the best CPU if prev_cpu cannot be used, or if it saves at
6825 * least 6% of the energy used by prev_cpu.
6827 if ((prev_delta == ULONG_MAX) ||
6828 (prev_delta - best_delta) > ((prev_delta + base_energy) >> 4))
6829 target = best_energy_cpu;
6840 * select_task_rq_fair: Select target runqueue for the waking task in domains
6841 * that have the relevant SD flag set. In practice, this is SD_BALANCE_WAKE,
6842 * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
6844 * Balances load by selecting the idlest CPU in the idlest group, or under
6845 * certain conditions an idle sibling CPU if the domain has SD_WAKE_AFFINE set.
6847 * Returns the target CPU number.
6850 select_task_rq_fair(struct task_struct *p, int prev_cpu, int wake_flags)
6852 int sync = (wake_flags & WF_SYNC) && !(current->flags & PF_EXITING);
6853 struct sched_domain *tmp, *sd = NULL;
6854 int cpu = smp_processor_id();
6855 int new_cpu = prev_cpu;
6856 int want_affine = 0;
6857 /* SD_flags and WF_flags share the first nibble */
6858 int sd_flag = wake_flags & 0xF;
6861 * required for stable ->cpus_allowed
6863 lockdep_assert_held(&p->pi_lock);
6864 if (wake_flags & WF_TTWU) {
6867 if (sched_energy_enabled()) {
6868 new_cpu = find_energy_efficient_cpu(p, prev_cpu);
6874 want_affine = !wake_wide(p) && cpumask_test_cpu(cpu, p->cpus_ptr);
6878 for_each_domain(cpu, tmp) {
6880 * If both 'cpu' and 'prev_cpu' are part of this domain,
6881 * cpu is a valid SD_WAKE_AFFINE target.
6883 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
6884 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
6885 if (cpu != prev_cpu)
6886 new_cpu = wake_affine(tmp, p, cpu, prev_cpu, sync);
6888 sd = NULL; /* Prefer wake_affine over balance flags */
6892 if (tmp->flags & sd_flag)
6894 else if (!want_affine)
6900 new_cpu = find_idlest_cpu(sd, p, cpu, prev_cpu, sd_flag);
6901 } else if (wake_flags & WF_TTWU) { /* XXX always ? */
6903 new_cpu = select_idle_sibling(p, prev_cpu, new_cpu);
6906 current->recent_used_cpu = cpu;
6913 static void detach_entity_cfs_rq(struct sched_entity *se);
6916 * Called immediately before a task is migrated to a new CPU; task_cpu(p) and
6917 * cfs_rq_of(p) references at time of call are still valid and identify the
6918 * previous CPU. The caller guarantees p->pi_lock or task_rq(p)->lock is held.
6920 static void migrate_task_rq_fair(struct task_struct *p, int new_cpu)
6923 * As blocked tasks retain absolute vruntime the migration needs to
6924 * deal with this by subtracting the old and adding the new
6925 * min_vruntime -- the latter is done by enqueue_entity() when placing
6926 * the task on the new runqueue.
6928 if (READ_ONCE(p->__state) == TASK_WAKING) {
6929 struct sched_entity *se = &p->se;
6930 struct cfs_rq *cfs_rq = cfs_rq_of(se);
6933 #ifndef CONFIG_64BIT
6934 u64 min_vruntime_copy;
6937 min_vruntime_copy = cfs_rq->min_vruntime_copy;
6939 min_vruntime = cfs_rq->min_vruntime;
6940 } while (min_vruntime != min_vruntime_copy);
6942 min_vruntime = cfs_rq->min_vruntime;
6945 se->vruntime -= min_vruntime;
6948 if (p->on_rq == TASK_ON_RQ_MIGRATING) {
6950 * In case of TASK_ON_RQ_MIGRATING we in fact hold the 'old'
6951 * rq->lock and can modify state directly.
6953 lockdep_assert_rq_held(task_rq(p));
6954 detach_entity_cfs_rq(&p->se);
6958 * We are supposed to update the task to "current" time, then
6959 * its up to date and ready to go to new CPU/cfs_rq. But we
6960 * have difficulty in getting what current time is, so simply
6961 * throw away the out-of-date time. This will result in the
6962 * wakee task is less decayed, but giving the wakee more load
6965 remove_entity_load_avg(&p->se);
6968 /* Tell new CPU we are migrated */
6969 p->se.avg.last_update_time = 0;
6971 /* We have migrated, no longer consider this task hot */
6972 p->se.exec_start = 0;
6974 update_scan_period(p, new_cpu);
6977 static void task_dead_fair(struct task_struct *p)
6979 remove_entity_load_avg(&p->se);
6983 balance_fair(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
6988 return newidle_balance(rq, rf) != 0;
6990 #endif /* CONFIG_SMP */
6992 static unsigned long wakeup_gran(struct sched_entity *se)
6994 unsigned long gran = sysctl_sched_wakeup_granularity;
6997 * Since its curr running now, convert the gran from real-time
6998 * to virtual-time in his units.
7000 * By using 'se' instead of 'curr' we penalize light tasks, so
7001 * they get preempted easier. That is, if 'se' < 'curr' then
7002 * the resulting gran will be larger, therefore penalizing the
7003 * lighter, if otoh 'se' > 'curr' then the resulting gran will
7004 * be smaller, again penalizing the lighter task.
7006 * This is especially important for buddies when the leftmost
7007 * task is higher priority than the buddy.
7009 return calc_delta_fair(gran, se);
7013 * Should 'se' preempt 'curr'.
7027 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
7029 s64 gran, vdiff = curr->vruntime - se->vruntime;
7034 gran = wakeup_gran(se);
7041 static void set_last_buddy(struct sched_entity *se)
7043 if (entity_is_task(se) && unlikely(task_has_idle_policy(task_of(se))))
7046 for_each_sched_entity(se) {
7047 if (SCHED_WARN_ON(!se->on_rq))
7049 cfs_rq_of(se)->last = se;
7053 static void set_next_buddy(struct sched_entity *se)
7055 if (entity_is_task(se) && unlikely(task_has_idle_policy(task_of(se))))
7058 for_each_sched_entity(se) {
7059 if (SCHED_WARN_ON(!se->on_rq))
7061 cfs_rq_of(se)->next = se;
7065 static void set_skip_buddy(struct sched_entity *se)
7067 for_each_sched_entity(se)
7068 cfs_rq_of(se)->skip = se;
7072 * Preempt the current task with a newly woken task if needed:
7074 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
7076 struct task_struct *curr = rq->curr;
7077 struct sched_entity *se = &curr->se, *pse = &p->se;
7078 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
7079 int scale = cfs_rq->nr_running >= sched_nr_latency;
7080 int next_buddy_marked = 0;
7082 if (unlikely(se == pse))
7086 * This is possible from callers such as attach_tasks(), in which we
7087 * unconditionally check_preempt_curr() after an enqueue (which may have
7088 * lead to a throttle). This both saves work and prevents false
7089 * next-buddy nomination below.
7091 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
7094 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
7095 set_next_buddy(pse);
7096 next_buddy_marked = 1;
7100 * We can come here with TIF_NEED_RESCHED already set from new task
7103 * Note: this also catches the edge-case of curr being in a throttled
7104 * group (e.g. via set_curr_task), since update_curr() (in the
7105 * enqueue of curr) will have resulted in resched being set. This
7106 * prevents us from potentially nominating it as a false LAST_BUDDY
7109 if (test_tsk_need_resched(curr))
7112 /* Idle tasks are by definition preempted by non-idle tasks. */
7113 if (unlikely(task_has_idle_policy(curr)) &&
7114 likely(!task_has_idle_policy(p)))
7118 * Batch and idle tasks do not preempt non-idle tasks (their preemption
7119 * is driven by the tick):
7121 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
7124 find_matching_se(&se, &pse);
7125 update_curr(cfs_rq_of(se));
7127 if (wakeup_preempt_entity(se, pse) == 1) {
7129 * Bias pick_next to pick the sched entity that is
7130 * triggering this preemption.
7132 if (!next_buddy_marked)
7133 set_next_buddy(pse);
7142 * Only set the backward buddy when the current task is still
7143 * on the rq. This can happen when a wakeup gets interleaved
7144 * with schedule on the ->pre_schedule() or idle_balance()
7145 * point, either of which can * drop the rq lock.
7147 * Also, during early boot the idle thread is in the fair class,
7148 * for obvious reasons its a bad idea to schedule back to it.
7150 if (unlikely(!se->on_rq || curr == rq->idle))
7153 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
7158 static struct task_struct *pick_task_fair(struct rq *rq)
7160 struct sched_entity *se;
7161 struct cfs_rq *cfs_rq;
7165 if (!cfs_rq->nr_running)
7169 struct sched_entity *curr = cfs_rq->curr;
7171 /* When we pick for a remote RQ, we'll not have done put_prev_entity() */
7174 update_curr(cfs_rq);
7178 if (unlikely(check_cfs_rq_runtime(cfs_rq)))
7182 se = pick_next_entity(cfs_rq, curr);
7183 cfs_rq = group_cfs_rq(se);
7190 struct task_struct *
7191 pick_next_task_fair(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
7193 struct cfs_rq *cfs_rq = &rq->cfs;
7194 struct sched_entity *se;
7195 struct task_struct *p;
7199 if (!sched_fair_runnable(rq))
7202 #ifdef CONFIG_FAIR_GROUP_SCHED
7203 if (!prev || prev->sched_class != &fair_sched_class)
7207 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
7208 * likely that a next task is from the same cgroup as the current.
7210 * Therefore attempt to avoid putting and setting the entire cgroup
7211 * hierarchy, only change the part that actually changes.
7215 struct sched_entity *curr = cfs_rq->curr;
7218 * Since we got here without doing put_prev_entity() we also
7219 * have to consider cfs_rq->curr. If it is still a runnable
7220 * entity, update_curr() will update its vruntime, otherwise
7221 * forget we've ever seen it.
7225 update_curr(cfs_rq);
7230 * This call to check_cfs_rq_runtime() will do the
7231 * throttle and dequeue its entity in the parent(s).
7232 * Therefore the nr_running test will indeed
7235 if (unlikely(check_cfs_rq_runtime(cfs_rq))) {
7238 if (!cfs_rq->nr_running)
7245 se = pick_next_entity(cfs_rq, curr);
7246 cfs_rq = group_cfs_rq(se);
7252 * Since we haven't yet done put_prev_entity and if the selected task
7253 * is a different task than we started out with, try and touch the
7254 * least amount of cfs_rqs.
7257 struct sched_entity *pse = &prev->se;
7259 while (!(cfs_rq = is_same_group(se, pse))) {
7260 int se_depth = se->depth;
7261 int pse_depth = pse->depth;
7263 if (se_depth <= pse_depth) {
7264 put_prev_entity(cfs_rq_of(pse), pse);
7265 pse = parent_entity(pse);
7267 if (se_depth >= pse_depth) {
7268 set_next_entity(cfs_rq_of(se), se);
7269 se = parent_entity(se);
7273 put_prev_entity(cfs_rq, pse);
7274 set_next_entity(cfs_rq, se);
7281 put_prev_task(rq, prev);
7284 se = pick_next_entity(cfs_rq, NULL);
7285 set_next_entity(cfs_rq, se);
7286 cfs_rq = group_cfs_rq(se);
7291 done: __maybe_unused;
7294 * Move the next running task to the front of
7295 * the list, so our cfs_tasks list becomes MRU
7298 list_move(&p->se.group_node, &rq->cfs_tasks);
7301 if (hrtick_enabled_fair(rq))
7302 hrtick_start_fair(rq, p);
7304 update_misfit_status(p, rq);
7312 new_tasks = newidle_balance(rq, rf);
7315 * Because newidle_balance() releases (and re-acquires) rq->lock, it is
7316 * possible for any higher priority task to appear. In that case we
7317 * must re-start the pick_next_entity() loop.
7326 * rq is about to be idle, check if we need to update the
7327 * lost_idle_time of clock_pelt
7329 update_idle_rq_clock_pelt(rq);
7334 static struct task_struct *__pick_next_task_fair(struct rq *rq)
7336 return pick_next_task_fair(rq, NULL, NULL);
7340 * Account for a descheduled task:
7342 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
7344 struct sched_entity *se = &prev->se;
7345 struct cfs_rq *cfs_rq;
7347 for_each_sched_entity(se) {
7348 cfs_rq = cfs_rq_of(se);
7349 put_prev_entity(cfs_rq, se);
7354 * sched_yield() is very simple
7356 * The magic of dealing with the ->skip buddy is in pick_next_entity.
7358 static void yield_task_fair(struct rq *rq)
7360 struct task_struct *curr = rq->curr;
7361 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
7362 struct sched_entity *se = &curr->se;
7365 * Are we the only task in the tree?
7367 if (unlikely(rq->nr_running == 1))
7370 clear_buddies(cfs_rq, se);
7372 if (curr->policy != SCHED_BATCH) {
7373 update_rq_clock(rq);
7375 * Update run-time statistics of the 'current'.
7377 update_curr(cfs_rq);
7379 * Tell update_rq_clock() that we've just updated,
7380 * so we don't do microscopic update in schedule()
7381 * and double the fastpath cost.
7383 rq_clock_skip_update(rq);
7389 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p)
7391 struct sched_entity *se = &p->se;
7393 /* throttled hierarchies are not runnable */
7394 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
7397 /* Tell the scheduler that we'd really like pse to run next. */
7400 yield_task_fair(rq);
7406 /**************************************************
7407 * Fair scheduling class load-balancing methods.
7411 * The purpose of load-balancing is to achieve the same basic fairness the
7412 * per-CPU scheduler provides, namely provide a proportional amount of compute
7413 * time to each task. This is expressed in the following equation:
7415 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
7417 * Where W_i,n is the n-th weight average for CPU i. The instantaneous weight
7418 * W_i,0 is defined as:
7420 * W_i,0 = \Sum_j w_i,j (2)
7422 * Where w_i,j is the weight of the j-th runnable task on CPU i. This weight
7423 * is derived from the nice value as per sched_prio_to_weight[].
7425 * The weight average is an exponential decay average of the instantaneous
7428 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
7430 * C_i is the compute capacity of CPU i, typically it is the
7431 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
7432 * can also include other factors [XXX].
7434 * To achieve this balance we define a measure of imbalance which follows
7435 * directly from (1):
7437 * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4)
7439 * We them move tasks around to minimize the imbalance. In the continuous
7440 * function space it is obvious this converges, in the discrete case we get
7441 * a few fun cases generally called infeasible weight scenarios.
7444 * - infeasible weights;
7445 * - local vs global optima in the discrete case. ]
7450 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
7451 * for all i,j solution, we create a tree of CPUs that follows the hardware
7452 * topology where each level pairs two lower groups (or better). This results
7453 * in O(log n) layers. Furthermore we reduce the number of CPUs going up the
7454 * tree to only the first of the previous level and we decrease the frequency
7455 * of load-balance at each level inv. proportional to the number of CPUs in
7461 * \Sum { --- * --- * 2^i } = O(n) (5)
7463 * `- size of each group
7464 * | | `- number of CPUs doing load-balance
7466 * `- sum over all levels
7468 * Coupled with a limit on how many tasks we can migrate every balance pass,
7469 * this makes (5) the runtime complexity of the balancer.
7471 * An important property here is that each CPU is still (indirectly) connected
7472 * to every other CPU in at most O(log n) steps:
7474 * The adjacency matrix of the resulting graph is given by:
7477 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
7480 * And you'll find that:
7482 * A^(log_2 n)_i,j != 0 for all i,j (7)
7484 * Showing there's indeed a path between every CPU in at most O(log n) steps.
7485 * The task movement gives a factor of O(m), giving a convergence complexity
7488 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
7493 * In order to avoid CPUs going idle while there's still work to do, new idle
7494 * balancing is more aggressive and has the newly idle CPU iterate up the domain
7495 * tree itself instead of relying on other CPUs to bring it work.
7497 * This adds some complexity to both (5) and (8) but it reduces the total idle
7505 * Cgroups make a horror show out of (2), instead of a simple sum we get:
7508 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
7513 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
7515 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on CPU i.
7517 * The big problem is S_k, its a global sum needed to compute a local (W_i)
7520 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
7521 * rewrite all of this once again.]
7524 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
7526 enum fbq_type { regular, remote, all };
7529 * 'group_type' describes the group of CPUs at the moment of load balancing.
7531 * The enum is ordered by pulling priority, with the group with lowest priority
7532 * first so the group_type can simply be compared when selecting the busiest
7533 * group. See update_sd_pick_busiest().
7536 /* The group has spare capacity that can be used to run more tasks. */
7537 group_has_spare = 0,
7539 * The group is fully used and the tasks don't compete for more CPU
7540 * cycles. Nevertheless, some tasks might wait before running.
7544 * SD_ASYM_CPUCAPACITY only: One task doesn't fit with CPU's capacity
7545 * and must be migrated to a more powerful CPU.
7549 * SD_ASYM_PACKING only: One local CPU with higher capacity is available,
7550 * and the task should be migrated to it instead of running on the
7555 * The tasks' affinity constraints previously prevented the scheduler
7556 * from balancing the load across the system.
7560 * The CPU is overloaded and can't provide expected CPU cycles to all
7566 enum migration_type {
7573 #define LBF_ALL_PINNED 0x01
7574 #define LBF_NEED_BREAK 0x02
7575 #define LBF_DST_PINNED 0x04
7576 #define LBF_SOME_PINNED 0x08
7577 #define LBF_ACTIVE_LB 0x10
7580 struct sched_domain *sd;
7588 struct cpumask *dst_grpmask;
7590 enum cpu_idle_type idle;
7592 /* The set of CPUs under consideration for load-balancing */
7593 struct cpumask *cpus;
7598 unsigned int loop_break;
7599 unsigned int loop_max;
7601 enum fbq_type fbq_type;
7602 enum migration_type migration_type;
7603 struct list_head tasks;
7607 * Is this task likely cache-hot:
7609 static int task_hot(struct task_struct *p, struct lb_env *env)
7613 lockdep_assert_rq_held(env->src_rq);
7615 if (p->sched_class != &fair_sched_class)
7618 if (unlikely(task_has_idle_policy(p)))
7621 /* SMT siblings share cache */
7622 if (env->sd->flags & SD_SHARE_CPUCAPACITY)
7626 * Buddy candidates are cache hot:
7628 if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running &&
7629 (&p->se == cfs_rq_of(&p->se)->next ||
7630 &p->se == cfs_rq_of(&p->se)->last))
7633 if (sysctl_sched_migration_cost == -1)
7637 * Don't migrate task if the task's cookie does not match
7638 * with the destination CPU's core cookie.
7640 if (!sched_core_cookie_match(cpu_rq(env->dst_cpu), p))
7643 if (sysctl_sched_migration_cost == 0)
7646 delta = rq_clock_task(env->src_rq) - p->se.exec_start;
7648 return delta < (s64)sysctl_sched_migration_cost;
7651 #ifdef CONFIG_NUMA_BALANCING
7653 * Returns 1, if task migration degrades locality
7654 * Returns 0, if task migration improves locality i.e migration preferred.
7655 * Returns -1, if task migration is not affected by locality.
7657 static int migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
7659 struct numa_group *numa_group = rcu_dereference(p->numa_group);
7660 unsigned long src_weight, dst_weight;
7661 int src_nid, dst_nid, dist;
7663 if (!static_branch_likely(&sched_numa_balancing))
7666 if (!p->numa_faults || !(env->sd->flags & SD_NUMA))
7669 src_nid = cpu_to_node(env->src_cpu);
7670 dst_nid = cpu_to_node(env->dst_cpu);
7672 if (src_nid == dst_nid)
7675 /* Migrating away from the preferred node is always bad. */
7676 if (src_nid == p->numa_preferred_nid) {
7677 if (env->src_rq->nr_running > env->src_rq->nr_preferred_running)
7683 /* Encourage migration to the preferred node. */
7684 if (dst_nid == p->numa_preferred_nid)
7687 /* Leaving a core idle is often worse than degrading locality. */
7688 if (env->idle == CPU_IDLE)
7691 dist = node_distance(src_nid, dst_nid);
7693 src_weight = group_weight(p, src_nid, dist);
7694 dst_weight = group_weight(p, dst_nid, dist);
7696 src_weight = task_weight(p, src_nid, dist);
7697 dst_weight = task_weight(p, dst_nid, dist);
7700 return dst_weight < src_weight;
7704 static inline int migrate_degrades_locality(struct task_struct *p,
7712 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
7715 int can_migrate_task(struct task_struct *p, struct lb_env *env)
7719 lockdep_assert_rq_held(env->src_rq);
7722 * We do not migrate tasks that are:
7723 * 1) throttled_lb_pair, or
7724 * 2) cannot be migrated to this CPU due to cpus_ptr, or
7725 * 3) running (obviously), or
7726 * 4) are cache-hot on their current CPU.
7728 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
7731 /* Disregard pcpu kthreads; they are where they need to be. */
7732 if (kthread_is_per_cpu(p))
7735 if (!cpumask_test_cpu(env->dst_cpu, p->cpus_ptr)) {
7738 schedstat_inc(p->se.statistics.nr_failed_migrations_affine);
7740 env->flags |= LBF_SOME_PINNED;
7743 * Remember if this task can be migrated to any other CPU in
7744 * our sched_group. We may want to revisit it if we couldn't
7745 * meet load balance goals by pulling other tasks on src_cpu.
7747 * Avoid computing new_dst_cpu
7749 * - if we have already computed one in current iteration
7750 * - if it's an active balance
7752 if (env->idle == CPU_NEWLY_IDLE ||
7753 env->flags & (LBF_DST_PINNED | LBF_ACTIVE_LB))
7756 /* Prevent to re-select dst_cpu via env's CPUs: */
7757 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
7758 if (cpumask_test_cpu(cpu, p->cpus_ptr)) {
7759 env->flags |= LBF_DST_PINNED;
7760 env->new_dst_cpu = cpu;
7768 /* Record that we found at least one task that could run on dst_cpu */
7769 env->flags &= ~LBF_ALL_PINNED;
7771 if (task_running(env->src_rq, p)) {
7772 schedstat_inc(p->se.statistics.nr_failed_migrations_running);
7777 * Aggressive migration if:
7779 * 2) destination numa is preferred
7780 * 3) task is cache cold, or
7781 * 4) too many balance attempts have failed.
7783 if (env->flags & LBF_ACTIVE_LB)
7786 tsk_cache_hot = migrate_degrades_locality(p, env);
7787 if (tsk_cache_hot == -1)
7788 tsk_cache_hot = task_hot(p, env);
7790 if (tsk_cache_hot <= 0 ||
7791 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
7792 if (tsk_cache_hot == 1) {
7793 schedstat_inc(env->sd->lb_hot_gained[env->idle]);
7794 schedstat_inc(p->se.statistics.nr_forced_migrations);
7799 schedstat_inc(p->se.statistics.nr_failed_migrations_hot);
7804 * detach_task() -- detach the task for the migration specified in env
7806 static void detach_task(struct task_struct *p, struct lb_env *env)
7808 lockdep_assert_rq_held(env->src_rq);
7810 deactivate_task(env->src_rq, p, DEQUEUE_NOCLOCK);
7811 set_task_cpu(p, env->dst_cpu);
7815 * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
7816 * part of active balancing operations within "domain".
7818 * Returns a task if successful and NULL otherwise.
7820 static struct task_struct *detach_one_task(struct lb_env *env)
7822 struct task_struct *p;
7824 lockdep_assert_rq_held(env->src_rq);
7826 list_for_each_entry_reverse(p,
7827 &env->src_rq->cfs_tasks, se.group_node) {
7828 if (!can_migrate_task(p, env))
7831 detach_task(p, env);
7834 * Right now, this is only the second place where
7835 * lb_gained[env->idle] is updated (other is detach_tasks)
7836 * so we can safely collect stats here rather than
7837 * inside detach_tasks().
7839 schedstat_inc(env->sd->lb_gained[env->idle]);
7845 static const unsigned int sched_nr_migrate_break = 32;
7848 * detach_tasks() -- tries to detach up to imbalance load/util/tasks from
7849 * busiest_rq, as part of a balancing operation within domain "sd".
7851 * Returns number of detached tasks if successful and 0 otherwise.
7853 static int detach_tasks(struct lb_env *env)
7855 struct list_head *tasks = &env->src_rq->cfs_tasks;
7856 unsigned long util, load;
7857 struct task_struct *p;
7860 lockdep_assert_rq_held(env->src_rq);
7863 * Source run queue has been emptied by another CPU, clear
7864 * LBF_ALL_PINNED flag as we will not test any task.
7866 if (env->src_rq->nr_running <= 1) {
7867 env->flags &= ~LBF_ALL_PINNED;
7871 if (env->imbalance <= 0)
7874 while (!list_empty(tasks)) {
7876 * We don't want to steal all, otherwise we may be treated likewise,
7877 * which could at worst lead to a livelock crash.
7879 if (env->idle != CPU_NOT_IDLE && env->src_rq->nr_running <= 1)
7882 p = list_last_entry(tasks, struct task_struct, se.group_node);
7885 /* We've more or less seen every task there is, call it quits */
7886 if (env->loop > env->loop_max)
7889 /* take a breather every nr_migrate tasks */
7890 if (env->loop > env->loop_break) {
7891 env->loop_break += sched_nr_migrate_break;
7892 env->flags |= LBF_NEED_BREAK;
7896 if (!can_migrate_task(p, env))
7899 switch (env->migration_type) {
7902 * Depending of the number of CPUs and tasks and the
7903 * cgroup hierarchy, task_h_load() can return a null
7904 * value. Make sure that env->imbalance decreases
7905 * otherwise detach_tasks() will stop only after
7906 * detaching up to loop_max tasks.
7908 load = max_t(unsigned long, task_h_load(p), 1);
7910 if (sched_feat(LB_MIN) &&
7911 load < 16 && !env->sd->nr_balance_failed)
7915 * Make sure that we don't migrate too much load.
7916 * Nevertheless, let relax the constraint if
7917 * scheduler fails to find a good waiting task to
7920 if (shr_bound(load, env->sd->nr_balance_failed) > env->imbalance)
7923 env->imbalance -= load;
7927 util = task_util_est(p);
7929 if (util > env->imbalance)
7932 env->imbalance -= util;
7939 case migrate_misfit:
7940 /* This is not a misfit task */
7941 if (task_fits_capacity(p, capacity_of(env->src_cpu)))
7948 detach_task(p, env);
7949 list_add(&p->se.group_node, &env->tasks);
7953 #ifdef CONFIG_PREEMPTION
7955 * NEWIDLE balancing is a source of latency, so preemptible
7956 * kernels will stop after the first task is detached to minimize
7957 * the critical section.
7959 if (env->idle == CPU_NEWLY_IDLE)
7964 * We only want to steal up to the prescribed amount of
7967 if (env->imbalance <= 0)
7972 list_move(&p->se.group_node, tasks);
7976 * Right now, this is one of only two places we collect this stat
7977 * so we can safely collect detach_one_task() stats here rather
7978 * than inside detach_one_task().
7980 schedstat_add(env->sd->lb_gained[env->idle], detached);
7986 * attach_task() -- attach the task detached by detach_task() to its new rq.
7988 static void attach_task(struct rq *rq, struct task_struct *p)
7990 lockdep_assert_rq_held(rq);
7992 BUG_ON(task_rq(p) != rq);
7993 activate_task(rq, p, ENQUEUE_NOCLOCK);
7994 check_preempt_curr(rq, p, 0);
7998 * attach_one_task() -- attaches the task returned from detach_one_task() to
8001 static void attach_one_task(struct rq *rq, struct task_struct *p)
8006 update_rq_clock(rq);
8012 * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
8015 static void attach_tasks(struct lb_env *env)
8017 struct list_head *tasks = &env->tasks;
8018 struct task_struct *p;
8021 rq_lock(env->dst_rq, &rf);
8022 update_rq_clock(env->dst_rq);
8024 while (!list_empty(tasks)) {
8025 p = list_first_entry(tasks, struct task_struct, se.group_node);
8026 list_del_init(&p->se.group_node);
8028 attach_task(env->dst_rq, p);
8031 rq_unlock(env->dst_rq, &rf);
8034 #ifdef CONFIG_NO_HZ_COMMON
8035 static inline bool cfs_rq_has_blocked(struct cfs_rq *cfs_rq)
8037 if (cfs_rq->avg.load_avg)
8040 if (cfs_rq->avg.util_avg)
8046 static inline bool others_have_blocked(struct rq *rq)
8048 if (READ_ONCE(rq->avg_rt.util_avg))
8051 if (READ_ONCE(rq->avg_dl.util_avg))
8054 if (thermal_load_avg(rq))
8057 #ifdef CONFIG_HAVE_SCHED_AVG_IRQ
8058 if (READ_ONCE(rq->avg_irq.util_avg))
8065 static inline void update_blocked_load_tick(struct rq *rq)
8067 WRITE_ONCE(rq->last_blocked_load_update_tick, jiffies);
8070 static inline void update_blocked_load_status(struct rq *rq, bool has_blocked)
8073 rq->has_blocked_load = 0;
8076 static inline bool cfs_rq_has_blocked(struct cfs_rq *cfs_rq) { return false; }
8077 static inline bool others_have_blocked(struct rq *rq) { return false; }
8078 static inline void update_blocked_load_tick(struct rq *rq) {}
8079 static inline void update_blocked_load_status(struct rq *rq, bool has_blocked) {}
8082 static bool __update_blocked_others(struct rq *rq, bool *done)
8084 const struct sched_class *curr_class;
8085 u64 now = rq_clock_pelt(rq);
8086 unsigned long thermal_pressure;
8090 * update_load_avg() can call cpufreq_update_util(). Make sure that RT,
8091 * DL and IRQ signals have been updated before updating CFS.
8093 curr_class = rq->curr->sched_class;
8095 thermal_pressure = arch_scale_thermal_pressure(cpu_of(rq));
8097 decayed = update_rt_rq_load_avg(now, rq, curr_class == &rt_sched_class) |
8098 update_dl_rq_load_avg(now, rq, curr_class == &dl_sched_class) |
8099 update_thermal_load_avg(rq_clock_thermal(rq), rq, thermal_pressure) |
8100 update_irq_load_avg(rq, 0);
8102 if (others_have_blocked(rq))
8108 #ifdef CONFIG_FAIR_GROUP_SCHED
8110 static bool __update_blocked_fair(struct rq *rq, bool *done)
8112 struct cfs_rq *cfs_rq, *pos;
8113 bool decayed = false;
8114 int cpu = cpu_of(rq);
8117 * Iterates the task_group tree in a bottom up fashion, see
8118 * list_add_leaf_cfs_rq() for details.
8120 for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) {
8121 struct sched_entity *se;
8123 if (update_cfs_rq_load_avg(cfs_rq_clock_pelt(cfs_rq), cfs_rq)) {
8124 update_tg_load_avg(cfs_rq);
8126 if (cfs_rq == &rq->cfs)
8130 /* Propagate pending load changes to the parent, if any: */
8131 se = cfs_rq->tg->se[cpu];
8132 if (se && !skip_blocked_update(se))
8133 update_load_avg(cfs_rq_of(se), se, UPDATE_TG);
8136 * There can be a lot of idle CPU cgroups. Don't let fully
8137 * decayed cfs_rqs linger on the list.
8139 if (cfs_rq_is_decayed(cfs_rq))
8140 list_del_leaf_cfs_rq(cfs_rq);
8142 /* Don't need periodic decay once load/util_avg are null */
8143 if (cfs_rq_has_blocked(cfs_rq))
8151 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
8152 * This needs to be done in a top-down fashion because the load of a child
8153 * group is a fraction of its parents load.
8155 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
8157 struct rq *rq = rq_of(cfs_rq);
8158 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
8159 unsigned long now = jiffies;
8162 if (cfs_rq->last_h_load_update == now)
8165 WRITE_ONCE(cfs_rq->h_load_next, NULL);
8166 for_each_sched_entity(se) {
8167 cfs_rq = cfs_rq_of(se);
8168 WRITE_ONCE(cfs_rq->h_load_next, se);
8169 if (cfs_rq->last_h_load_update == now)
8174 cfs_rq->h_load = cfs_rq_load_avg(cfs_rq);
8175 cfs_rq->last_h_load_update = now;
8178 while ((se = READ_ONCE(cfs_rq->h_load_next)) != NULL) {
8179 load = cfs_rq->h_load;
8180 load = div64_ul(load * se->avg.load_avg,
8181 cfs_rq_load_avg(cfs_rq) + 1);
8182 cfs_rq = group_cfs_rq(se);
8183 cfs_rq->h_load = load;
8184 cfs_rq->last_h_load_update = now;
8188 static unsigned long task_h_load(struct task_struct *p)
8190 struct cfs_rq *cfs_rq = task_cfs_rq(p);
8192 update_cfs_rq_h_load(cfs_rq);
8193 return div64_ul(p->se.avg.load_avg * cfs_rq->h_load,
8194 cfs_rq_load_avg(cfs_rq) + 1);
8197 static bool __update_blocked_fair(struct rq *rq, bool *done)
8199 struct cfs_rq *cfs_rq = &rq->cfs;
8202 decayed = update_cfs_rq_load_avg(cfs_rq_clock_pelt(cfs_rq), cfs_rq);
8203 if (cfs_rq_has_blocked(cfs_rq))
8209 static unsigned long task_h_load(struct task_struct *p)
8211 return p->se.avg.load_avg;
8215 static void update_blocked_averages(int cpu)
8217 bool decayed = false, done = true;
8218 struct rq *rq = cpu_rq(cpu);
8221 rq_lock_irqsave(rq, &rf);
8222 update_blocked_load_tick(rq);
8223 update_rq_clock(rq);
8225 decayed |= __update_blocked_others(rq, &done);
8226 decayed |= __update_blocked_fair(rq, &done);
8228 update_blocked_load_status(rq, !done);
8230 cpufreq_update_util(rq, 0);
8231 rq_unlock_irqrestore(rq, &rf);
8234 /********** Helpers for find_busiest_group ************************/
8237 * sg_lb_stats - stats of a sched_group required for load_balancing
8239 struct sg_lb_stats {
8240 unsigned long avg_load; /*Avg load across the CPUs of the group */
8241 unsigned long group_load; /* Total load over the CPUs of the group */
8242 unsigned long group_capacity;
8243 unsigned long group_util; /* Total utilization over the CPUs of the group */
8244 unsigned long group_runnable; /* Total runnable time over the CPUs of the group */
8245 unsigned int sum_nr_running; /* Nr of tasks running in the group */
8246 unsigned int sum_h_nr_running; /* Nr of CFS tasks running in the group */
8247 unsigned int idle_cpus;
8248 unsigned int group_weight;
8249 enum group_type group_type;
8250 unsigned int group_asym_packing; /* Tasks should be moved to preferred CPU */
8251 unsigned long group_misfit_task_load; /* A CPU has a task too big for its capacity */
8252 #ifdef CONFIG_NUMA_BALANCING
8253 unsigned int nr_numa_running;
8254 unsigned int nr_preferred_running;
8259 * sd_lb_stats - Structure to store the statistics of a sched_domain
8260 * during load balancing.
8262 struct sd_lb_stats {
8263 struct sched_group *busiest; /* Busiest group in this sd */
8264 struct sched_group *local; /* Local group in this sd */
8265 unsigned long total_load; /* Total load of all groups in sd */
8266 unsigned long total_capacity; /* Total capacity of all groups in sd */
8267 unsigned long avg_load; /* Average load across all groups in sd */
8268 unsigned int prefer_sibling; /* tasks should go to sibling first */
8270 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
8271 struct sg_lb_stats local_stat; /* Statistics of the local group */
8274 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
8277 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
8278 * local_stat because update_sg_lb_stats() does a full clear/assignment.
8279 * We must however set busiest_stat::group_type and
8280 * busiest_stat::idle_cpus to the worst busiest group because
8281 * update_sd_pick_busiest() reads these before assignment.
8283 *sds = (struct sd_lb_stats){
8287 .total_capacity = 0UL,
8289 .idle_cpus = UINT_MAX,
8290 .group_type = group_has_spare,
8295 static unsigned long scale_rt_capacity(int cpu)
8297 struct rq *rq = cpu_rq(cpu);
8298 unsigned long max = arch_scale_cpu_capacity(cpu);
8299 unsigned long used, free;
8302 irq = cpu_util_irq(rq);
8304 if (unlikely(irq >= max))
8308 * avg_rt.util_avg and avg_dl.util_avg track binary signals
8309 * (running and not running) with weights 0 and 1024 respectively.
8310 * avg_thermal.load_avg tracks thermal pressure and the weighted
8311 * average uses the actual delta max capacity(load).
8313 used = READ_ONCE(rq->avg_rt.util_avg);
8314 used += READ_ONCE(rq->avg_dl.util_avg);
8315 used += thermal_load_avg(rq);
8317 if (unlikely(used >= max))
8322 return scale_irq_capacity(free, irq, max);
8325 static void update_cpu_capacity(struct sched_domain *sd, int cpu)
8327 unsigned long capacity = scale_rt_capacity(cpu);
8328 struct sched_group *sdg = sd->groups;
8330 cpu_rq(cpu)->cpu_capacity_orig = arch_scale_cpu_capacity(cpu);
8335 cpu_rq(cpu)->cpu_capacity = capacity;
8336 trace_sched_cpu_capacity_tp(cpu_rq(cpu));
8338 sdg->sgc->capacity = capacity;
8339 sdg->sgc->min_capacity = capacity;
8340 sdg->sgc->max_capacity = capacity;
8343 void update_group_capacity(struct sched_domain *sd, int cpu)
8345 struct sched_domain *child = sd->child;
8346 struct sched_group *group, *sdg = sd->groups;
8347 unsigned long capacity, min_capacity, max_capacity;
8348 unsigned long interval;
8350 interval = msecs_to_jiffies(sd->balance_interval);
8351 interval = clamp(interval, 1UL, max_load_balance_interval);
8352 sdg->sgc->next_update = jiffies + interval;
8355 update_cpu_capacity(sd, cpu);
8360 min_capacity = ULONG_MAX;
8363 if (child->flags & SD_OVERLAP) {
8365 * SD_OVERLAP domains cannot assume that child groups
8366 * span the current group.
8369 for_each_cpu(cpu, sched_group_span(sdg)) {
8370 unsigned long cpu_cap = capacity_of(cpu);
8372 capacity += cpu_cap;
8373 min_capacity = min(cpu_cap, min_capacity);
8374 max_capacity = max(cpu_cap, max_capacity);
8378 * !SD_OVERLAP domains can assume that child groups
8379 * span the current group.
8382 group = child->groups;
8384 struct sched_group_capacity *sgc = group->sgc;
8386 capacity += sgc->capacity;
8387 min_capacity = min(sgc->min_capacity, min_capacity);
8388 max_capacity = max(sgc->max_capacity, max_capacity);
8389 group = group->next;
8390 } while (group != child->groups);
8393 sdg->sgc->capacity = capacity;
8394 sdg->sgc->min_capacity = min_capacity;
8395 sdg->sgc->max_capacity = max_capacity;
8399 * Check whether the capacity of the rq has been noticeably reduced by side
8400 * activity. The imbalance_pct is used for the threshold.
8401 * Return true is the capacity is reduced
8404 check_cpu_capacity(struct rq *rq, struct sched_domain *sd)
8406 return ((rq->cpu_capacity * sd->imbalance_pct) <
8407 (rq->cpu_capacity_orig * 100));
8411 * Check whether a rq has a misfit task and if it looks like we can actually
8412 * help that task: we can migrate the task to a CPU of higher capacity, or
8413 * the task's current CPU is heavily pressured.
8415 static inline int check_misfit_status(struct rq *rq, struct sched_domain *sd)
8417 return rq->misfit_task_load &&
8418 (rq->cpu_capacity_orig < rq->rd->max_cpu_capacity ||
8419 check_cpu_capacity(rq, sd));
8423 * Group imbalance indicates (and tries to solve) the problem where balancing
8424 * groups is inadequate due to ->cpus_ptr constraints.
8426 * Imagine a situation of two groups of 4 CPUs each and 4 tasks each with a
8427 * cpumask covering 1 CPU of the first group and 3 CPUs of the second group.
8430 * { 0 1 2 3 } { 4 5 6 7 }
8433 * If we were to balance group-wise we'd place two tasks in the first group and
8434 * two tasks in the second group. Clearly this is undesired as it will overload
8435 * cpu 3 and leave one of the CPUs in the second group unused.
8437 * The current solution to this issue is detecting the skew in the first group
8438 * by noticing the lower domain failed to reach balance and had difficulty
8439 * moving tasks due to affinity constraints.
8441 * When this is so detected; this group becomes a candidate for busiest; see
8442 * update_sd_pick_busiest(). And calculate_imbalance() and
8443 * find_busiest_group() avoid some of the usual balance conditions to allow it
8444 * to create an effective group imbalance.
8446 * This is a somewhat tricky proposition since the next run might not find the
8447 * group imbalance and decide the groups need to be balanced again. A most
8448 * subtle and fragile situation.
8451 static inline int sg_imbalanced(struct sched_group *group)
8453 return group->sgc->imbalance;
8457 * group_has_capacity returns true if the group has spare capacity that could
8458 * be used by some tasks.
8459 * We consider that a group has spare capacity if the * number of task is
8460 * smaller than the number of CPUs or if the utilization is lower than the
8461 * available capacity for CFS tasks.
8462 * For the latter, we use a threshold to stabilize the state, to take into
8463 * account the variance of the tasks' load and to return true if the available
8464 * capacity in meaningful for the load balancer.
8465 * As an example, an available capacity of 1% can appear but it doesn't make
8466 * any benefit for the load balance.
8469 group_has_capacity(unsigned int imbalance_pct, struct sg_lb_stats *sgs)
8471 if (sgs->sum_nr_running < sgs->group_weight)
8474 if ((sgs->group_capacity * imbalance_pct) <
8475 (sgs->group_runnable * 100))
8478 if ((sgs->group_capacity * 100) >
8479 (sgs->group_util * imbalance_pct))
8486 * group_is_overloaded returns true if the group has more tasks than it can
8488 * group_is_overloaded is not equals to !group_has_capacity because a group
8489 * with the exact right number of tasks, has no more spare capacity but is not
8490 * overloaded so both group_has_capacity and group_is_overloaded return
8494 group_is_overloaded(unsigned int imbalance_pct, struct sg_lb_stats *sgs)
8496 if (sgs->sum_nr_running <= sgs->group_weight)
8499 if ((sgs->group_capacity * 100) <
8500 (sgs->group_util * imbalance_pct))
8503 if ((sgs->group_capacity * imbalance_pct) <
8504 (sgs->group_runnable * 100))
8511 group_type group_classify(unsigned int imbalance_pct,
8512 struct sched_group *group,
8513 struct sg_lb_stats *sgs)
8515 if (group_is_overloaded(imbalance_pct, sgs))
8516 return group_overloaded;
8518 if (sg_imbalanced(group))
8519 return group_imbalanced;
8521 if (sgs->group_asym_packing)
8522 return group_asym_packing;
8524 if (sgs->group_misfit_task_load)
8525 return group_misfit_task;
8527 if (!group_has_capacity(imbalance_pct, sgs))
8528 return group_fully_busy;
8530 return group_has_spare;
8534 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
8535 * @env: The load balancing environment.
8536 * @group: sched_group whose statistics are to be updated.
8537 * @sgs: variable to hold the statistics for this group.
8538 * @sg_status: Holds flag indicating the status of the sched_group
8540 static inline void update_sg_lb_stats(struct lb_env *env,
8541 struct sched_group *group,
8542 struct sg_lb_stats *sgs,
8545 int i, nr_running, local_group;
8547 memset(sgs, 0, sizeof(*sgs));
8549 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_span(group));
8551 for_each_cpu_and(i, sched_group_span(group), env->cpus) {
8552 struct rq *rq = cpu_rq(i);
8554 sgs->group_load += cpu_load(rq);
8555 sgs->group_util += cpu_util(i);
8556 sgs->group_runnable += cpu_runnable(rq);
8557 sgs->sum_h_nr_running += rq->cfs.h_nr_running;
8559 nr_running = rq->nr_running;
8560 sgs->sum_nr_running += nr_running;
8563 *sg_status |= SG_OVERLOAD;
8565 if (cpu_overutilized(i))
8566 *sg_status |= SG_OVERUTILIZED;
8568 #ifdef CONFIG_NUMA_BALANCING
8569 sgs->nr_numa_running += rq->nr_numa_running;
8570 sgs->nr_preferred_running += rq->nr_preferred_running;
8573 * No need to call idle_cpu() if nr_running is not 0
8575 if (!nr_running && idle_cpu(i)) {
8577 /* Idle cpu can't have misfit task */
8584 /* Check for a misfit task on the cpu */
8585 if (env->sd->flags & SD_ASYM_CPUCAPACITY &&
8586 sgs->group_misfit_task_load < rq->misfit_task_load) {
8587 sgs->group_misfit_task_load = rq->misfit_task_load;
8588 *sg_status |= SG_OVERLOAD;
8592 /* Check if dst CPU is idle and preferred to this group */
8593 if (env->sd->flags & SD_ASYM_PACKING &&
8594 env->idle != CPU_NOT_IDLE &&
8595 sgs->sum_h_nr_running &&
8596 sched_asym_prefer(env->dst_cpu, group->asym_prefer_cpu)) {
8597 sgs->group_asym_packing = 1;
8600 sgs->group_capacity = group->sgc->capacity;
8602 sgs->group_weight = group->group_weight;
8604 sgs->group_type = group_classify(env->sd->imbalance_pct, group, sgs);
8606 /* Computing avg_load makes sense only when group is overloaded */
8607 if (sgs->group_type == group_overloaded)
8608 sgs->avg_load = (sgs->group_load * SCHED_CAPACITY_SCALE) /
8609 sgs->group_capacity;
8613 * update_sd_pick_busiest - return 1 on busiest group
8614 * @env: The load balancing environment.
8615 * @sds: sched_domain statistics
8616 * @sg: sched_group candidate to be checked for being the busiest
8617 * @sgs: sched_group statistics
8619 * Determine if @sg is a busier group than the previously selected
8622 * Return: %true if @sg is a busier group than the previously selected
8623 * busiest group. %false otherwise.
8625 static bool update_sd_pick_busiest(struct lb_env *env,
8626 struct sd_lb_stats *sds,
8627 struct sched_group *sg,
8628 struct sg_lb_stats *sgs)
8630 struct sg_lb_stats *busiest = &sds->busiest_stat;
8632 /* Make sure that there is at least one task to pull */
8633 if (!sgs->sum_h_nr_running)
8637 * Don't try to pull misfit tasks we can't help.
8638 * We can use max_capacity here as reduction in capacity on some
8639 * CPUs in the group should either be possible to resolve
8640 * internally or be covered by avg_load imbalance (eventually).
8642 if (sgs->group_type == group_misfit_task &&
8643 (!capacity_greater(capacity_of(env->dst_cpu), sg->sgc->max_capacity) ||
8644 sds->local_stat.group_type != group_has_spare))
8647 if (sgs->group_type > busiest->group_type)
8650 if (sgs->group_type < busiest->group_type)
8654 * The candidate and the current busiest group are the same type of
8655 * group. Let check which one is the busiest according to the type.
8658 switch (sgs->group_type) {
8659 case group_overloaded:
8660 /* Select the overloaded group with highest avg_load. */
8661 if (sgs->avg_load <= busiest->avg_load)
8665 case group_imbalanced:
8667 * Select the 1st imbalanced group as we don't have any way to
8668 * choose one more than another.
8672 case group_asym_packing:
8673 /* Prefer to move from lowest priority CPU's work */
8674 if (sched_asym_prefer(sg->asym_prefer_cpu, sds->busiest->asym_prefer_cpu))
8678 case group_misfit_task:
8680 * If we have more than one misfit sg go with the biggest
8683 if (sgs->group_misfit_task_load < busiest->group_misfit_task_load)
8687 case group_fully_busy:
8689 * Select the fully busy group with highest avg_load. In
8690 * theory, there is no need to pull task from such kind of
8691 * group because tasks have all compute capacity that they need
8692 * but we can still improve the overall throughput by reducing
8693 * contention when accessing shared HW resources.
8695 * XXX for now avg_load is not computed and always 0 so we
8696 * select the 1st one.
8698 if (sgs->avg_load <= busiest->avg_load)
8702 case group_has_spare:
8704 * Select not overloaded group with lowest number of idle cpus
8705 * and highest number of running tasks. We could also compare
8706 * the spare capacity which is more stable but it can end up
8707 * that the group has less spare capacity but finally more idle
8708 * CPUs which means less opportunity to pull tasks.
8710 if (sgs->idle_cpus > busiest->idle_cpus)
8712 else if ((sgs->idle_cpus == busiest->idle_cpus) &&
8713 (sgs->sum_nr_running <= busiest->sum_nr_running))
8720 * Candidate sg has no more than one task per CPU and has higher
8721 * per-CPU capacity. Migrating tasks to less capable CPUs may harm
8722 * throughput. Maximize throughput, power/energy consequences are not
8725 if ((env->sd->flags & SD_ASYM_CPUCAPACITY) &&
8726 (sgs->group_type <= group_fully_busy) &&
8727 (capacity_greater(sg->sgc->min_capacity, capacity_of(env->dst_cpu))))
8733 #ifdef CONFIG_NUMA_BALANCING
8734 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
8736 if (sgs->sum_h_nr_running > sgs->nr_numa_running)
8738 if (sgs->sum_h_nr_running > sgs->nr_preferred_running)
8743 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
8745 if (rq->nr_running > rq->nr_numa_running)
8747 if (rq->nr_running > rq->nr_preferred_running)
8752 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
8757 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
8761 #endif /* CONFIG_NUMA_BALANCING */
8767 * task_running_on_cpu - return 1 if @p is running on @cpu.
8770 static unsigned int task_running_on_cpu(int cpu, struct task_struct *p)
8772 /* Task has no contribution or is new */
8773 if (cpu != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time))
8776 if (task_on_rq_queued(p))
8783 * idle_cpu_without - would a given CPU be idle without p ?
8784 * @cpu: the processor on which idleness is tested.
8785 * @p: task which should be ignored.
8787 * Return: 1 if the CPU would be idle. 0 otherwise.
8789 static int idle_cpu_without(int cpu, struct task_struct *p)
8791 struct rq *rq = cpu_rq(cpu);
8793 if (rq->curr != rq->idle && rq->curr != p)
8797 * rq->nr_running can't be used but an updated version without the
8798 * impact of p on cpu must be used instead. The updated nr_running
8799 * be computed and tested before calling idle_cpu_without().
8803 if (rq->ttwu_pending)
8811 * update_sg_wakeup_stats - Update sched_group's statistics for wakeup.
8812 * @sd: The sched_domain level to look for idlest group.
8813 * @group: sched_group whose statistics are to be updated.
8814 * @sgs: variable to hold the statistics for this group.
8815 * @p: The task for which we look for the idlest group/CPU.
8817 static inline void update_sg_wakeup_stats(struct sched_domain *sd,
8818 struct sched_group *group,
8819 struct sg_lb_stats *sgs,
8820 struct task_struct *p)
8824 memset(sgs, 0, sizeof(*sgs));
8826 for_each_cpu(i, sched_group_span(group)) {
8827 struct rq *rq = cpu_rq(i);
8830 sgs->group_load += cpu_load_without(rq, p);
8831 sgs->group_util += cpu_util_without(i, p);
8832 sgs->group_runnable += cpu_runnable_without(rq, p);
8833 local = task_running_on_cpu(i, p);
8834 sgs->sum_h_nr_running += rq->cfs.h_nr_running - local;
8836 nr_running = rq->nr_running - local;
8837 sgs->sum_nr_running += nr_running;
8840 * No need to call idle_cpu_without() if nr_running is not 0
8842 if (!nr_running && idle_cpu_without(i, p))
8847 /* Check if task fits in the group */
8848 if (sd->flags & SD_ASYM_CPUCAPACITY &&
8849 !task_fits_capacity(p, group->sgc->max_capacity)) {
8850 sgs->group_misfit_task_load = 1;
8853 sgs->group_capacity = group->sgc->capacity;
8855 sgs->group_weight = group->group_weight;
8857 sgs->group_type = group_classify(sd->imbalance_pct, group, sgs);
8860 * Computing avg_load makes sense only when group is fully busy or
8863 if (sgs->group_type == group_fully_busy ||
8864 sgs->group_type == group_overloaded)
8865 sgs->avg_load = (sgs->group_load * SCHED_CAPACITY_SCALE) /
8866 sgs->group_capacity;
8869 static bool update_pick_idlest(struct sched_group *idlest,
8870 struct sg_lb_stats *idlest_sgs,
8871 struct sched_group *group,
8872 struct sg_lb_stats *sgs)
8874 if (sgs->group_type < idlest_sgs->group_type)
8877 if (sgs->group_type > idlest_sgs->group_type)
8881 * The candidate and the current idlest group are the same type of
8882 * group. Let check which one is the idlest according to the type.
8885 switch (sgs->group_type) {
8886 case group_overloaded:
8887 case group_fully_busy:
8888 /* Select the group with lowest avg_load. */
8889 if (idlest_sgs->avg_load <= sgs->avg_load)
8893 case group_imbalanced:
8894 case group_asym_packing:
8895 /* Those types are not used in the slow wakeup path */
8898 case group_misfit_task:
8899 /* Select group with the highest max capacity */
8900 if (idlest->sgc->max_capacity >= group->sgc->max_capacity)
8904 case group_has_spare:
8905 /* Select group with most idle CPUs */
8906 if (idlest_sgs->idle_cpus > sgs->idle_cpus)
8909 /* Select group with lowest group_util */
8910 if (idlest_sgs->idle_cpus == sgs->idle_cpus &&
8911 idlest_sgs->group_util <= sgs->group_util)
8921 * Allow a NUMA imbalance if busy CPUs is less than 25% of the domain.
8922 * This is an approximation as the number of running tasks may not be
8923 * related to the number of busy CPUs due to sched_setaffinity.
8925 static inline bool allow_numa_imbalance(int dst_running, int dst_weight)
8927 return (dst_running < (dst_weight >> 2));
8931 * find_idlest_group() finds and returns the least busy CPU group within the
8934 * Assumes p is allowed on at least one CPU in sd.
8936 static struct sched_group *
8937 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
8939 struct sched_group *idlest = NULL, *local = NULL, *group = sd->groups;
8940 struct sg_lb_stats local_sgs, tmp_sgs;
8941 struct sg_lb_stats *sgs;
8942 unsigned long imbalance;
8943 struct sg_lb_stats idlest_sgs = {
8944 .avg_load = UINT_MAX,
8945 .group_type = group_overloaded,
8951 /* Skip over this group if it has no CPUs allowed */
8952 if (!cpumask_intersects(sched_group_span(group),
8956 /* Skip over this group if no cookie matched */
8957 if (!sched_group_cookie_match(cpu_rq(this_cpu), p, group))
8960 local_group = cpumask_test_cpu(this_cpu,
8961 sched_group_span(group));
8970 update_sg_wakeup_stats(sd, group, sgs, p);
8972 if (!local_group && update_pick_idlest(idlest, &idlest_sgs, group, sgs)) {
8977 } while (group = group->next, group != sd->groups);
8980 /* There is no idlest group to push tasks to */
8984 /* The local group has been skipped because of CPU affinity */
8989 * If the local group is idler than the selected idlest group
8990 * don't try and push the task.
8992 if (local_sgs.group_type < idlest_sgs.group_type)
8996 * If the local group is busier than the selected idlest group
8997 * try and push the task.
8999 if (local_sgs.group_type > idlest_sgs.group_type)
9002 switch (local_sgs.group_type) {
9003 case group_overloaded:
9004 case group_fully_busy:
9006 /* Calculate allowed imbalance based on load */
9007 imbalance = scale_load_down(NICE_0_LOAD) *
9008 (sd->imbalance_pct-100) / 100;
9011 * When comparing groups across NUMA domains, it's possible for
9012 * the local domain to be very lightly loaded relative to the
9013 * remote domains but "imbalance" skews the comparison making
9014 * remote CPUs look much more favourable. When considering
9015 * cross-domain, add imbalance to the load on the remote node
9016 * and consider staying local.
9019 if ((sd->flags & SD_NUMA) &&
9020 ((idlest_sgs.avg_load + imbalance) >= local_sgs.avg_load))
9024 * If the local group is less loaded than the selected
9025 * idlest group don't try and push any tasks.
9027 if (idlest_sgs.avg_load >= (local_sgs.avg_load + imbalance))
9030 if (100 * local_sgs.avg_load <= sd->imbalance_pct * idlest_sgs.avg_load)
9034 case group_imbalanced:
9035 case group_asym_packing:
9036 /* Those type are not used in the slow wakeup path */
9039 case group_misfit_task:
9040 /* Select group with the highest max capacity */
9041 if (local->sgc->max_capacity >= idlest->sgc->max_capacity)
9045 case group_has_spare:
9046 if (sd->flags & SD_NUMA) {
9047 #ifdef CONFIG_NUMA_BALANCING
9050 * If there is spare capacity at NUMA, try to select
9051 * the preferred node
9053 if (cpu_to_node(this_cpu) == p->numa_preferred_nid)
9056 idlest_cpu = cpumask_first(sched_group_span(idlest));
9057 if (cpu_to_node(idlest_cpu) == p->numa_preferred_nid)
9061 * Otherwise, keep the task on this node to stay close
9062 * its wakeup source and improve locality. If there is
9063 * a real need of migration, periodic load balance will
9066 if (allow_numa_imbalance(local_sgs.sum_nr_running, sd->span_weight))
9071 * Select group with highest number of idle CPUs. We could also
9072 * compare the utilization which is more stable but it can end
9073 * up that the group has less spare capacity but finally more
9074 * idle CPUs which means more opportunity to run task.
9076 if (local_sgs.idle_cpus >= idlest_sgs.idle_cpus)
9085 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
9086 * @env: The load balancing environment.
9087 * @sds: variable to hold the statistics for this sched_domain.
9090 static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
9092 struct sched_domain *child = env->sd->child;
9093 struct sched_group *sg = env->sd->groups;
9094 struct sg_lb_stats *local = &sds->local_stat;
9095 struct sg_lb_stats tmp_sgs;
9099 struct sg_lb_stats *sgs = &tmp_sgs;
9102 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_span(sg));
9107 if (env->idle != CPU_NEWLY_IDLE ||
9108 time_after_eq(jiffies, sg->sgc->next_update))
9109 update_group_capacity(env->sd, env->dst_cpu);
9112 update_sg_lb_stats(env, sg, sgs, &sg_status);
9118 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
9120 sds->busiest_stat = *sgs;
9124 /* Now, start updating sd_lb_stats */
9125 sds->total_load += sgs->group_load;
9126 sds->total_capacity += sgs->group_capacity;
9129 } while (sg != env->sd->groups);
9131 /* Tag domain that child domain prefers tasks go to siblings first */
9132 sds->prefer_sibling = child && child->flags & SD_PREFER_SIBLING;
9135 if (env->sd->flags & SD_NUMA)
9136 env->fbq_type = fbq_classify_group(&sds->busiest_stat);
9138 if (!env->sd->parent) {
9139 struct root_domain *rd = env->dst_rq->rd;
9141 /* update overload indicator if we are at root domain */
9142 WRITE_ONCE(rd->overload, sg_status & SG_OVERLOAD);
9144 /* Update over-utilization (tipping point, U >= 0) indicator */
9145 WRITE_ONCE(rd->overutilized, sg_status & SG_OVERUTILIZED);
9146 trace_sched_overutilized_tp(rd, sg_status & SG_OVERUTILIZED);
9147 } else if (sg_status & SG_OVERUTILIZED) {
9148 struct root_domain *rd = env->dst_rq->rd;
9150 WRITE_ONCE(rd->overutilized, SG_OVERUTILIZED);
9151 trace_sched_overutilized_tp(rd, SG_OVERUTILIZED);
9155 #define NUMA_IMBALANCE_MIN 2
9157 static inline long adjust_numa_imbalance(int imbalance,
9158 int dst_running, int dst_weight)
9160 if (!allow_numa_imbalance(dst_running, dst_weight))
9164 * Allow a small imbalance based on a simple pair of communicating
9165 * tasks that remain local when the destination is lightly loaded.
9167 if (imbalance <= NUMA_IMBALANCE_MIN)
9174 * calculate_imbalance - Calculate the amount of imbalance present within the
9175 * groups of a given sched_domain during load balance.
9176 * @env: load balance environment
9177 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
9179 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
9181 struct sg_lb_stats *local, *busiest;
9183 local = &sds->local_stat;
9184 busiest = &sds->busiest_stat;
9186 if (busiest->group_type == group_misfit_task) {
9187 /* Set imbalance to allow misfit tasks to be balanced. */
9188 env->migration_type = migrate_misfit;
9193 if (busiest->group_type == group_asym_packing) {
9195 * In case of asym capacity, we will try to migrate all load to
9196 * the preferred CPU.
9198 env->migration_type = migrate_task;
9199 env->imbalance = busiest->sum_h_nr_running;
9203 if (busiest->group_type == group_imbalanced) {
9205 * In the group_imb case we cannot rely on group-wide averages
9206 * to ensure CPU-load equilibrium, try to move any task to fix
9207 * the imbalance. The next load balance will take care of
9208 * balancing back the system.
9210 env->migration_type = migrate_task;
9216 * Try to use spare capacity of local group without overloading it or
9219 if (local->group_type == group_has_spare) {
9220 if ((busiest->group_type > group_fully_busy) &&
9221 !(env->sd->flags & SD_SHARE_PKG_RESOURCES)) {
9223 * If busiest is overloaded, try to fill spare
9224 * capacity. This might end up creating spare capacity
9225 * in busiest or busiest still being overloaded but
9226 * there is no simple way to directly compute the
9227 * amount of load to migrate in order to balance the
9230 env->migration_type = migrate_util;
9231 env->imbalance = max(local->group_capacity, local->group_util) -
9235 * In some cases, the group's utilization is max or even
9236 * higher than capacity because of migrations but the
9237 * local CPU is (newly) idle. There is at least one
9238 * waiting task in this overloaded busiest group. Let's
9241 if (env->idle != CPU_NOT_IDLE && env->imbalance == 0) {
9242 env->migration_type = migrate_task;
9249 if (busiest->group_weight == 1 || sds->prefer_sibling) {
9250 unsigned int nr_diff = busiest->sum_nr_running;
9252 * When prefer sibling, evenly spread running tasks on
9255 env->migration_type = migrate_task;
9256 lsub_positive(&nr_diff, local->sum_nr_running);
9257 env->imbalance = nr_diff >> 1;
9261 * If there is no overload, we just want to even the number of
9264 env->migration_type = migrate_task;
9265 env->imbalance = max_t(long, 0, (local->idle_cpus -
9266 busiest->idle_cpus) >> 1);
9269 /* Consider allowing a small imbalance between NUMA groups */
9270 if (env->sd->flags & SD_NUMA) {
9271 env->imbalance = adjust_numa_imbalance(env->imbalance,
9272 busiest->sum_nr_running, busiest->group_weight);
9279 * Local is fully busy but has to take more load to relieve the
9282 if (local->group_type < group_overloaded) {
9284 * Local will become overloaded so the avg_load metrics are
9288 local->avg_load = (local->group_load * SCHED_CAPACITY_SCALE) /
9289 local->group_capacity;
9291 sds->avg_load = (sds->total_load * SCHED_CAPACITY_SCALE) /
9292 sds->total_capacity;
9294 * If the local group is more loaded than the selected
9295 * busiest group don't try to pull any tasks.
9297 if (local->avg_load >= busiest->avg_load) {
9304 * Both group are or will become overloaded and we're trying to get all
9305 * the CPUs to the average_load, so we don't want to push ourselves
9306 * above the average load, nor do we wish to reduce the max loaded CPU
9307 * below the average load. At the same time, we also don't want to
9308 * reduce the group load below the group capacity. Thus we look for
9309 * the minimum possible imbalance.
9311 env->migration_type = migrate_load;
9312 env->imbalance = min(
9313 (busiest->avg_load - sds->avg_load) * busiest->group_capacity,
9314 (sds->avg_load - local->avg_load) * local->group_capacity
9315 ) / SCHED_CAPACITY_SCALE;
9318 /******* find_busiest_group() helpers end here *********************/
9321 * Decision matrix according to the local and busiest group type:
9323 * busiest \ local has_spare fully_busy misfit asym imbalanced overloaded
9324 * has_spare nr_idle balanced N/A N/A balanced balanced
9325 * fully_busy nr_idle nr_idle N/A N/A balanced balanced
9326 * misfit_task force N/A N/A N/A force force
9327 * asym_packing force force N/A N/A force force
9328 * imbalanced force force N/A N/A force force
9329 * overloaded force force N/A N/A force avg_load
9331 * N/A : Not Applicable because already filtered while updating
9333 * balanced : The system is balanced for these 2 groups.
9334 * force : Calculate the imbalance as load migration is probably needed.
9335 * avg_load : Only if imbalance is significant enough.
9336 * nr_idle : dst_cpu is not busy and the number of idle CPUs is quite
9337 * different in groups.
9341 * find_busiest_group - Returns the busiest group within the sched_domain
9342 * if there is an imbalance.
9344 * Also calculates the amount of runnable load which should be moved
9345 * to restore balance.
9347 * @env: The load balancing environment.
9349 * Return: - The busiest group if imbalance exists.
9351 static struct sched_group *find_busiest_group(struct lb_env *env)
9353 struct sg_lb_stats *local, *busiest;
9354 struct sd_lb_stats sds;
9356 init_sd_lb_stats(&sds);
9359 * Compute the various statistics relevant for load balancing at
9362 update_sd_lb_stats(env, &sds);
9364 if (sched_energy_enabled()) {
9365 struct root_domain *rd = env->dst_rq->rd;
9367 if (rcu_dereference(rd->pd) && !READ_ONCE(rd->overutilized))
9371 local = &sds.local_stat;
9372 busiest = &sds.busiest_stat;
9374 /* There is no busy sibling group to pull tasks from */
9378 /* Misfit tasks should be dealt with regardless of the avg load */
9379 if (busiest->group_type == group_misfit_task)
9382 /* ASYM feature bypasses nice load balance check */
9383 if (busiest->group_type == group_asym_packing)
9387 * If the busiest group is imbalanced the below checks don't
9388 * work because they assume all things are equal, which typically
9389 * isn't true due to cpus_ptr constraints and the like.
9391 if (busiest->group_type == group_imbalanced)
9395 * If the local group is busier than the selected busiest group
9396 * don't try and pull any tasks.
9398 if (local->group_type > busiest->group_type)
9402 * When groups are overloaded, use the avg_load to ensure fairness
9405 if (local->group_type == group_overloaded) {
9407 * If the local group is more loaded than the selected
9408 * busiest group don't try to pull any tasks.
9410 if (local->avg_load >= busiest->avg_load)
9413 /* XXX broken for overlapping NUMA groups */
9414 sds.avg_load = (sds.total_load * SCHED_CAPACITY_SCALE) /
9418 * Don't pull any tasks if this group is already above the
9419 * domain average load.
9421 if (local->avg_load >= sds.avg_load)
9425 * If the busiest group is more loaded, use imbalance_pct to be
9428 if (100 * busiest->avg_load <=
9429 env->sd->imbalance_pct * local->avg_load)
9433 /* Try to move all excess tasks to child's sibling domain */
9434 if (sds.prefer_sibling && local->group_type == group_has_spare &&
9435 busiest->sum_nr_running > local->sum_nr_running + 1)
9438 if (busiest->group_type != group_overloaded) {
9439 if (env->idle == CPU_NOT_IDLE)
9441 * If the busiest group is not overloaded (and as a
9442 * result the local one too) but this CPU is already
9443 * busy, let another idle CPU try to pull task.
9447 if (busiest->group_weight > 1 &&
9448 local->idle_cpus <= (busiest->idle_cpus + 1))
9450 * If the busiest group is not overloaded
9451 * and there is no imbalance between this and busiest
9452 * group wrt idle CPUs, it is balanced. The imbalance
9453 * becomes significant if the diff is greater than 1
9454 * otherwise we might end up to just move the imbalance
9455 * on another group. Of course this applies only if
9456 * there is more than 1 CPU per group.
9460 if (busiest->sum_h_nr_running == 1)
9462 * busiest doesn't have any tasks waiting to run
9468 /* Looks like there is an imbalance. Compute it */
9469 calculate_imbalance(env, &sds);
9470 return env->imbalance ? sds.busiest : NULL;
9478 * find_busiest_queue - find the busiest runqueue among the CPUs in the group.
9480 static struct rq *find_busiest_queue(struct lb_env *env,
9481 struct sched_group *group)
9483 struct rq *busiest = NULL, *rq;
9484 unsigned long busiest_util = 0, busiest_load = 0, busiest_capacity = 1;
9485 unsigned int busiest_nr = 0;
9488 for_each_cpu_and(i, sched_group_span(group), env->cpus) {
9489 unsigned long capacity, load, util;
9490 unsigned int nr_running;
9494 rt = fbq_classify_rq(rq);
9497 * We classify groups/runqueues into three groups:
9498 * - regular: there are !numa tasks
9499 * - remote: there are numa tasks that run on the 'wrong' node
9500 * - all: there is no distinction
9502 * In order to avoid migrating ideally placed numa tasks,
9503 * ignore those when there's better options.
9505 * If we ignore the actual busiest queue to migrate another
9506 * task, the next balance pass can still reduce the busiest
9507 * queue by moving tasks around inside the node.
9509 * If we cannot move enough load due to this classification
9510 * the next pass will adjust the group classification and
9511 * allow migration of more tasks.
9513 * Both cases only affect the total convergence complexity.
9515 if (rt > env->fbq_type)
9518 nr_running = rq->cfs.h_nr_running;
9522 capacity = capacity_of(i);
9525 * For ASYM_CPUCAPACITY domains, don't pick a CPU that could
9526 * eventually lead to active_balancing high->low capacity.
9527 * Higher per-CPU capacity is considered better than balancing
9530 if (env->sd->flags & SD_ASYM_CPUCAPACITY &&
9531 !capacity_greater(capacity_of(env->dst_cpu), capacity) &&
9535 switch (env->migration_type) {
9538 * When comparing with load imbalance, use cpu_load()
9539 * which is not scaled with the CPU capacity.
9541 load = cpu_load(rq);
9543 if (nr_running == 1 && load > env->imbalance &&
9544 !check_cpu_capacity(rq, env->sd))
9548 * For the load comparisons with the other CPUs,
9549 * consider the cpu_load() scaled with the CPU
9550 * capacity, so that the load can be moved away
9551 * from the CPU that is potentially running at a
9554 * Thus we're looking for max(load_i / capacity_i),
9555 * crosswise multiplication to rid ourselves of the
9556 * division works out to:
9557 * load_i * capacity_j > load_j * capacity_i;
9558 * where j is our previous maximum.
9560 if (load * busiest_capacity > busiest_load * capacity) {
9561 busiest_load = load;
9562 busiest_capacity = capacity;
9568 util = cpu_util(cpu_of(rq));
9571 * Don't try to pull utilization from a CPU with one
9572 * running task. Whatever its utilization, we will fail
9575 if (nr_running <= 1)
9578 if (busiest_util < util) {
9579 busiest_util = util;
9585 if (busiest_nr < nr_running) {
9586 busiest_nr = nr_running;
9591 case migrate_misfit:
9593 * For ASYM_CPUCAPACITY domains with misfit tasks we
9594 * simply seek the "biggest" misfit task.
9596 if (rq->misfit_task_load > busiest_load) {
9597 busiest_load = rq->misfit_task_load;
9610 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
9611 * so long as it is large enough.
9613 #define MAX_PINNED_INTERVAL 512
9616 asym_active_balance(struct lb_env *env)
9619 * ASYM_PACKING needs to force migrate tasks from busy but
9620 * lower priority CPUs in order to pack all tasks in the
9621 * highest priority CPUs.
9623 return env->idle != CPU_NOT_IDLE && (env->sd->flags & SD_ASYM_PACKING) &&
9624 sched_asym_prefer(env->dst_cpu, env->src_cpu);
9628 imbalanced_active_balance(struct lb_env *env)
9630 struct sched_domain *sd = env->sd;
9633 * The imbalanced case includes the case of pinned tasks preventing a fair
9634 * distribution of the load on the system but also the even distribution of the
9635 * threads on a system with spare capacity
9637 if ((env->migration_type == migrate_task) &&
9638 (sd->nr_balance_failed > sd->cache_nice_tries+2))
9644 static int need_active_balance(struct lb_env *env)
9646 struct sched_domain *sd = env->sd;
9648 if (asym_active_balance(env))
9651 if (imbalanced_active_balance(env))
9655 * The dst_cpu is idle and the src_cpu CPU has only 1 CFS task.
9656 * It's worth migrating the task if the src_cpu's capacity is reduced
9657 * because of other sched_class or IRQs if more capacity stays
9658 * available on dst_cpu.
9660 if ((env->idle != CPU_NOT_IDLE) &&
9661 (env->src_rq->cfs.h_nr_running == 1)) {
9662 if ((check_cpu_capacity(env->src_rq, sd)) &&
9663 (capacity_of(env->src_cpu)*sd->imbalance_pct < capacity_of(env->dst_cpu)*100))
9667 if (env->migration_type == migrate_misfit)
9673 static int active_load_balance_cpu_stop(void *data);
9675 static int should_we_balance(struct lb_env *env)
9677 struct sched_group *sg = env->sd->groups;
9681 * Ensure the balancing environment is consistent; can happen
9682 * when the softirq triggers 'during' hotplug.
9684 if (!cpumask_test_cpu(env->dst_cpu, env->cpus))
9688 * In the newly idle case, we will allow all the CPUs
9689 * to do the newly idle load balance.
9691 if (env->idle == CPU_NEWLY_IDLE)
9694 /* Try to find first idle CPU */
9695 for_each_cpu_and(cpu, group_balance_mask(sg), env->cpus) {
9699 /* Are we the first idle CPU? */
9700 return cpu == env->dst_cpu;
9703 /* Are we the first CPU of this group ? */
9704 return group_balance_cpu(sg) == env->dst_cpu;
9708 * Check this_cpu to ensure it is balanced within domain. Attempt to move
9709 * tasks if there is an imbalance.
9711 static int load_balance(int this_cpu, struct rq *this_rq,
9712 struct sched_domain *sd, enum cpu_idle_type idle,
9713 int *continue_balancing)
9715 int ld_moved, cur_ld_moved, active_balance = 0;
9716 struct sched_domain *sd_parent = sd->parent;
9717 struct sched_group *group;
9720 struct cpumask *cpus = this_cpu_cpumask_var_ptr(load_balance_mask);
9722 struct lb_env env = {
9724 .dst_cpu = this_cpu,
9726 .dst_grpmask = sched_group_span(sd->groups),
9728 .loop_break = sched_nr_migrate_break,
9731 .tasks = LIST_HEAD_INIT(env.tasks),
9734 cpumask_and(cpus, sched_domain_span(sd), cpu_active_mask);
9736 schedstat_inc(sd->lb_count[idle]);
9739 if (!should_we_balance(&env)) {
9740 *continue_balancing = 0;
9744 group = find_busiest_group(&env);
9746 schedstat_inc(sd->lb_nobusyg[idle]);
9750 busiest = find_busiest_queue(&env, group);
9752 schedstat_inc(sd->lb_nobusyq[idle]);
9756 BUG_ON(busiest == env.dst_rq);
9758 schedstat_add(sd->lb_imbalance[idle], env.imbalance);
9760 env.src_cpu = busiest->cpu;
9761 env.src_rq = busiest;
9764 /* Clear this flag as soon as we find a pullable task */
9765 env.flags |= LBF_ALL_PINNED;
9766 if (busiest->nr_running > 1) {
9768 * Attempt to move tasks. If find_busiest_group has found
9769 * an imbalance but busiest->nr_running <= 1, the group is
9770 * still unbalanced. ld_moved simply stays zero, so it is
9771 * correctly treated as an imbalance.
9773 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
9776 rq_lock_irqsave(busiest, &rf);
9777 update_rq_clock(busiest);
9780 * cur_ld_moved - load moved in current iteration
9781 * ld_moved - cumulative load moved across iterations
9783 cur_ld_moved = detach_tasks(&env);
9786 * We've detached some tasks from busiest_rq. Every
9787 * task is masked "TASK_ON_RQ_MIGRATING", so we can safely
9788 * unlock busiest->lock, and we are able to be sure
9789 * that nobody can manipulate the tasks in parallel.
9790 * See task_rq_lock() family for the details.
9793 rq_unlock(busiest, &rf);
9797 ld_moved += cur_ld_moved;
9800 local_irq_restore(rf.flags);
9802 if (env.flags & LBF_NEED_BREAK) {
9803 env.flags &= ~LBF_NEED_BREAK;
9808 * Revisit (affine) tasks on src_cpu that couldn't be moved to
9809 * us and move them to an alternate dst_cpu in our sched_group
9810 * where they can run. The upper limit on how many times we
9811 * iterate on same src_cpu is dependent on number of CPUs in our
9814 * This changes load balance semantics a bit on who can move
9815 * load to a given_cpu. In addition to the given_cpu itself
9816 * (or a ilb_cpu acting on its behalf where given_cpu is
9817 * nohz-idle), we now have balance_cpu in a position to move
9818 * load to given_cpu. In rare situations, this may cause
9819 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
9820 * _independently_ and at _same_ time to move some load to
9821 * given_cpu) causing excess load to be moved to given_cpu.
9822 * This however should not happen so much in practice and
9823 * moreover subsequent load balance cycles should correct the
9824 * excess load moved.
9826 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
9828 /* Prevent to re-select dst_cpu via env's CPUs */
9829 __cpumask_clear_cpu(env.dst_cpu, env.cpus);
9831 env.dst_rq = cpu_rq(env.new_dst_cpu);
9832 env.dst_cpu = env.new_dst_cpu;
9833 env.flags &= ~LBF_DST_PINNED;
9835 env.loop_break = sched_nr_migrate_break;
9838 * Go back to "more_balance" rather than "redo" since we
9839 * need to continue with same src_cpu.
9845 * We failed to reach balance because of affinity.
9848 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
9850 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0)
9851 *group_imbalance = 1;
9854 /* All tasks on this runqueue were pinned by CPU affinity */
9855 if (unlikely(env.flags & LBF_ALL_PINNED)) {
9856 __cpumask_clear_cpu(cpu_of(busiest), cpus);
9858 * Attempting to continue load balancing at the current
9859 * sched_domain level only makes sense if there are
9860 * active CPUs remaining as possible busiest CPUs to
9861 * pull load from which are not contained within the
9862 * destination group that is receiving any migrated
9865 if (!cpumask_subset(cpus, env.dst_grpmask)) {
9867 env.loop_break = sched_nr_migrate_break;
9870 goto out_all_pinned;
9875 schedstat_inc(sd->lb_failed[idle]);
9877 * Increment the failure counter only on periodic balance.
9878 * We do not want newidle balance, which can be very
9879 * frequent, pollute the failure counter causing
9880 * excessive cache_hot migrations and active balances.
9882 if (idle != CPU_NEWLY_IDLE)
9883 sd->nr_balance_failed++;
9885 if (need_active_balance(&env)) {
9886 unsigned long flags;
9888 raw_spin_rq_lock_irqsave(busiest, flags);
9891 * Don't kick the active_load_balance_cpu_stop,
9892 * if the curr task on busiest CPU can't be
9893 * moved to this_cpu:
9895 if (!cpumask_test_cpu(this_cpu, busiest->curr->cpus_ptr)) {
9896 raw_spin_rq_unlock_irqrestore(busiest, flags);
9897 goto out_one_pinned;
9900 /* Record that we found at least one task that could run on this_cpu */
9901 env.flags &= ~LBF_ALL_PINNED;
9904 * ->active_balance synchronizes accesses to
9905 * ->active_balance_work. Once set, it's cleared
9906 * only after active load balance is finished.
9908 if (!busiest->active_balance) {
9909 busiest->active_balance = 1;
9910 busiest->push_cpu = this_cpu;
9913 raw_spin_rq_unlock_irqrestore(busiest, flags);
9915 if (active_balance) {
9916 stop_one_cpu_nowait(cpu_of(busiest),
9917 active_load_balance_cpu_stop, busiest,
9918 &busiest->active_balance_work);
9922 sd->nr_balance_failed = 0;
9925 if (likely(!active_balance) || need_active_balance(&env)) {
9926 /* We were unbalanced, so reset the balancing interval */
9927 sd->balance_interval = sd->min_interval;
9934 * We reach balance although we may have faced some affinity
9935 * constraints. Clear the imbalance flag only if other tasks got
9936 * a chance to move and fix the imbalance.
9938 if (sd_parent && !(env.flags & LBF_ALL_PINNED)) {
9939 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
9941 if (*group_imbalance)
9942 *group_imbalance = 0;
9947 * We reach balance because all tasks are pinned at this level so
9948 * we can't migrate them. Let the imbalance flag set so parent level
9949 * can try to migrate them.
9951 schedstat_inc(sd->lb_balanced[idle]);
9953 sd->nr_balance_failed = 0;
9959 * newidle_balance() disregards balance intervals, so we could
9960 * repeatedly reach this code, which would lead to balance_interval
9961 * skyrocketing in a short amount of time. Skip the balance_interval
9962 * increase logic to avoid that.
9964 if (env.idle == CPU_NEWLY_IDLE)
9967 /* tune up the balancing interval */
9968 if ((env.flags & LBF_ALL_PINNED &&
9969 sd->balance_interval < MAX_PINNED_INTERVAL) ||
9970 sd->balance_interval < sd->max_interval)
9971 sd->balance_interval *= 2;
9976 static inline unsigned long
9977 get_sd_balance_interval(struct sched_domain *sd, int cpu_busy)
9979 unsigned long interval = sd->balance_interval;
9982 interval *= sd->busy_factor;
9984 /* scale ms to jiffies */
9985 interval = msecs_to_jiffies(interval);
9988 * Reduce likelihood of busy balancing at higher domains racing with
9989 * balancing at lower domains by preventing their balancing periods
9990 * from being multiples of each other.
9995 interval = clamp(interval, 1UL, max_load_balance_interval);
10001 update_next_balance(struct sched_domain *sd, unsigned long *next_balance)
10003 unsigned long interval, next;
10005 /* used by idle balance, so cpu_busy = 0 */
10006 interval = get_sd_balance_interval(sd, 0);
10007 next = sd->last_balance + interval;
10009 if (time_after(*next_balance, next))
10010 *next_balance = next;
10014 * active_load_balance_cpu_stop is run by the CPU stopper. It pushes
10015 * running tasks off the busiest CPU onto idle CPUs. It requires at
10016 * least 1 task to be running on each physical CPU where possible, and
10017 * avoids physical / logical imbalances.
10019 static int active_load_balance_cpu_stop(void *data)
10021 struct rq *busiest_rq = data;
10022 int busiest_cpu = cpu_of(busiest_rq);
10023 int target_cpu = busiest_rq->push_cpu;
10024 struct rq *target_rq = cpu_rq(target_cpu);
10025 struct sched_domain *sd;
10026 struct task_struct *p = NULL;
10027 struct rq_flags rf;
10029 rq_lock_irq(busiest_rq, &rf);
10031 * Between queueing the stop-work and running it is a hole in which
10032 * CPUs can become inactive. We should not move tasks from or to
10035 if (!cpu_active(busiest_cpu) || !cpu_active(target_cpu))
10038 /* Make sure the requested CPU hasn't gone down in the meantime: */
10039 if (unlikely(busiest_cpu != smp_processor_id() ||
10040 !busiest_rq->active_balance))
10043 /* Is there any task to move? */
10044 if (busiest_rq->nr_running <= 1)
10048 * This condition is "impossible", if it occurs
10049 * we need to fix it. Originally reported by
10050 * Bjorn Helgaas on a 128-CPU setup.
10052 BUG_ON(busiest_rq == target_rq);
10054 /* Search for an sd spanning us and the target CPU. */
10056 for_each_domain(target_cpu, sd) {
10057 if (cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
10062 struct lb_env env = {
10064 .dst_cpu = target_cpu,
10065 .dst_rq = target_rq,
10066 .src_cpu = busiest_rq->cpu,
10067 .src_rq = busiest_rq,
10069 .flags = LBF_ACTIVE_LB,
10072 schedstat_inc(sd->alb_count);
10073 update_rq_clock(busiest_rq);
10075 p = detach_one_task(&env);
10077 schedstat_inc(sd->alb_pushed);
10078 /* Active balancing done, reset the failure counter. */
10079 sd->nr_balance_failed = 0;
10081 schedstat_inc(sd->alb_failed);
10086 busiest_rq->active_balance = 0;
10087 rq_unlock(busiest_rq, &rf);
10090 attach_one_task(target_rq, p);
10092 local_irq_enable();
10097 static DEFINE_SPINLOCK(balancing);
10100 * Scale the max load_balance interval with the number of CPUs in the system.
10101 * This trades load-balance latency on larger machines for less cross talk.
10103 void update_max_interval(void)
10105 max_load_balance_interval = HZ*num_online_cpus()/10;
10109 * It checks each scheduling domain to see if it is due to be balanced,
10110 * and initiates a balancing operation if so.
10112 * Balancing parameters are set up in init_sched_domains.
10114 static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle)
10116 int continue_balancing = 1;
10118 int busy = idle != CPU_IDLE && !sched_idle_cpu(cpu);
10119 unsigned long interval;
10120 struct sched_domain *sd;
10121 /* Earliest time when we have to do rebalance again */
10122 unsigned long next_balance = jiffies + 60*HZ;
10123 int update_next_balance = 0;
10124 int need_serialize, need_decay = 0;
10128 for_each_domain(cpu, sd) {
10130 * Decay the newidle max times here because this is a regular
10131 * visit to all the domains. Decay ~1% per second.
10133 if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
10134 sd->max_newidle_lb_cost =
10135 (sd->max_newidle_lb_cost * 253) / 256;
10136 sd->next_decay_max_lb_cost = jiffies + HZ;
10139 max_cost += sd->max_newidle_lb_cost;
10142 * Stop the load balance at this level. There is another
10143 * CPU in our sched group which is doing load balancing more
10146 if (!continue_balancing) {
10152 interval = get_sd_balance_interval(sd, busy);
10154 need_serialize = sd->flags & SD_SERIALIZE;
10155 if (need_serialize) {
10156 if (!spin_trylock(&balancing))
10160 if (time_after_eq(jiffies, sd->last_balance + interval)) {
10161 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
10163 * The LBF_DST_PINNED logic could have changed
10164 * env->dst_cpu, so we can't know our idle
10165 * state even if we migrated tasks. Update it.
10167 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
10168 busy = idle != CPU_IDLE && !sched_idle_cpu(cpu);
10170 sd->last_balance = jiffies;
10171 interval = get_sd_balance_interval(sd, busy);
10173 if (need_serialize)
10174 spin_unlock(&balancing);
10176 if (time_after(next_balance, sd->last_balance + interval)) {
10177 next_balance = sd->last_balance + interval;
10178 update_next_balance = 1;
10183 * Ensure the rq-wide value also decays but keep it at a
10184 * reasonable floor to avoid funnies with rq->avg_idle.
10186 rq->max_idle_balance_cost =
10187 max((u64)sysctl_sched_migration_cost, max_cost);
10192 * next_balance will be updated only when there is a need.
10193 * When the cpu is attached to null domain for ex, it will not be
10196 if (likely(update_next_balance))
10197 rq->next_balance = next_balance;
10201 static inline int on_null_domain(struct rq *rq)
10203 return unlikely(!rcu_dereference_sched(rq->sd));
10206 #ifdef CONFIG_NO_HZ_COMMON
10208 * idle load balancing details
10209 * - When one of the busy CPUs notice that there may be an idle rebalancing
10210 * needed, they will kick the idle load balancer, which then does idle
10211 * load balancing for all the idle CPUs.
10212 * - HK_FLAG_MISC CPUs are used for this task, because HK_FLAG_SCHED not set
10216 static inline int find_new_ilb(void)
10220 for_each_cpu_and(ilb, nohz.idle_cpus_mask,
10221 housekeeping_cpumask(HK_FLAG_MISC)) {
10223 if (ilb == smp_processor_id())
10234 * Kick a CPU to do the nohz balancing, if it is time for it. We pick any
10235 * idle CPU in the HK_FLAG_MISC housekeeping set (if there is one).
10237 static void kick_ilb(unsigned int flags)
10242 * Increase nohz.next_balance only when if full ilb is triggered but
10243 * not if we only update stats.
10245 if (flags & NOHZ_BALANCE_KICK)
10246 nohz.next_balance = jiffies+1;
10248 ilb_cpu = find_new_ilb();
10250 if (ilb_cpu >= nr_cpu_ids)
10254 * Access to rq::nohz_csd is serialized by NOHZ_KICK_MASK; he who sets
10255 * the first flag owns it; cleared by nohz_csd_func().
10257 flags = atomic_fetch_or(flags, nohz_flags(ilb_cpu));
10258 if (flags & NOHZ_KICK_MASK)
10262 * This way we generate an IPI on the target CPU which
10263 * is idle. And the softirq performing nohz idle load balance
10264 * will be run before returning from the IPI.
10266 smp_call_function_single_async(ilb_cpu, &cpu_rq(ilb_cpu)->nohz_csd);
10270 * Current decision point for kicking the idle load balancer in the presence
10271 * of idle CPUs in the system.
10273 static void nohz_balancer_kick(struct rq *rq)
10275 unsigned long now = jiffies;
10276 struct sched_domain_shared *sds;
10277 struct sched_domain *sd;
10278 int nr_busy, i, cpu = rq->cpu;
10279 unsigned int flags = 0;
10281 if (unlikely(rq->idle_balance))
10285 * We may be recently in ticked or tickless idle mode. At the first
10286 * busy tick after returning from idle, we will update the busy stats.
10288 nohz_balance_exit_idle(rq);
10291 * None are in tickless mode and hence no need for NOHZ idle load
10294 if (likely(!atomic_read(&nohz.nr_cpus)))
10297 if (READ_ONCE(nohz.has_blocked) &&
10298 time_after(now, READ_ONCE(nohz.next_blocked)))
10299 flags = NOHZ_STATS_KICK;
10301 if (time_before(now, nohz.next_balance))
10304 if (rq->nr_running >= 2) {
10305 flags = NOHZ_KICK_MASK;
10311 sd = rcu_dereference(rq->sd);
10314 * If there's a CFS task and the current CPU has reduced
10315 * capacity; kick the ILB to see if there's a better CPU to run
10318 if (rq->cfs.h_nr_running >= 1 && check_cpu_capacity(rq, sd)) {
10319 flags = NOHZ_KICK_MASK;
10324 sd = rcu_dereference(per_cpu(sd_asym_packing, cpu));
10327 * When ASYM_PACKING; see if there's a more preferred CPU
10328 * currently idle; in which case, kick the ILB to move tasks
10331 for_each_cpu_and(i, sched_domain_span(sd), nohz.idle_cpus_mask) {
10332 if (sched_asym_prefer(i, cpu)) {
10333 flags = NOHZ_KICK_MASK;
10339 sd = rcu_dereference(per_cpu(sd_asym_cpucapacity, cpu));
10342 * When ASYM_CPUCAPACITY; see if there's a higher capacity CPU
10343 * to run the misfit task on.
10345 if (check_misfit_status(rq, sd)) {
10346 flags = NOHZ_KICK_MASK;
10351 * For asymmetric systems, we do not want to nicely balance
10352 * cache use, instead we want to embrace asymmetry and only
10353 * ensure tasks have enough CPU capacity.
10355 * Skip the LLC logic because it's not relevant in that case.
10360 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
10363 * If there is an imbalance between LLC domains (IOW we could
10364 * increase the overall cache use), we need some less-loaded LLC
10365 * domain to pull some load. Likewise, we may need to spread
10366 * load within the current LLC domain (e.g. packed SMT cores but
10367 * other CPUs are idle). We can't really know from here how busy
10368 * the others are - so just get a nohz balance going if it looks
10369 * like this LLC domain has tasks we could move.
10371 nr_busy = atomic_read(&sds->nr_busy_cpus);
10373 flags = NOHZ_KICK_MASK;
10384 static void set_cpu_sd_state_busy(int cpu)
10386 struct sched_domain *sd;
10389 sd = rcu_dereference(per_cpu(sd_llc, cpu));
10391 if (!sd || !sd->nohz_idle)
10395 atomic_inc(&sd->shared->nr_busy_cpus);
10400 void nohz_balance_exit_idle(struct rq *rq)
10402 SCHED_WARN_ON(rq != this_rq());
10404 if (likely(!rq->nohz_tick_stopped))
10407 rq->nohz_tick_stopped = 0;
10408 cpumask_clear_cpu(rq->cpu, nohz.idle_cpus_mask);
10409 atomic_dec(&nohz.nr_cpus);
10411 set_cpu_sd_state_busy(rq->cpu);
10414 static void set_cpu_sd_state_idle(int cpu)
10416 struct sched_domain *sd;
10419 sd = rcu_dereference(per_cpu(sd_llc, cpu));
10421 if (!sd || sd->nohz_idle)
10425 atomic_dec(&sd->shared->nr_busy_cpus);
10431 * This routine will record that the CPU is going idle with tick stopped.
10432 * This info will be used in performing idle load balancing in the future.
10434 void nohz_balance_enter_idle(int cpu)
10436 struct rq *rq = cpu_rq(cpu);
10438 SCHED_WARN_ON(cpu != smp_processor_id());
10440 /* If this CPU is going down, then nothing needs to be done: */
10441 if (!cpu_active(cpu))
10444 /* Spare idle load balancing on CPUs that don't want to be disturbed: */
10445 if (!housekeeping_cpu(cpu, HK_FLAG_SCHED))
10449 * Can be set safely without rq->lock held
10450 * If a clear happens, it will have evaluated last additions because
10451 * rq->lock is held during the check and the clear
10453 rq->has_blocked_load = 1;
10456 * The tick is still stopped but load could have been added in the
10457 * meantime. We set the nohz.has_blocked flag to trig a check of the
10458 * *_avg. The CPU is already part of nohz.idle_cpus_mask so the clear
10459 * of nohz.has_blocked can only happen after checking the new load
10461 if (rq->nohz_tick_stopped)
10464 /* If we're a completely isolated CPU, we don't play: */
10465 if (on_null_domain(rq))
10468 rq->nohz_tick_stopped = 1;
10470 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
10471 atomic_inc(&nohz.nr_cpus);
10474 * Ensures that if nohz_idle_balance() fails to observe our
10475 * @idle_cpus_mask store, it must observe the @has_blocked
10478 smp_mb__after_atomic();
10480 set_cpu_sd_state_idle(cpu);
10484 * Each time a cpu enter idle, we assume that it has blocked load and
10485 * enable the periodic update of the load of idle cpus
10487 WRITE_ONCE(nohz.has_blocked, 1);
10490 static bool update_nohz_stats(struct rq *rq)
10492 unsigned int cpu = rq->cpu;
10494 if (!rq->has_blocked_load)
10497 if (!cpumask_test_cpu(cpu, nohz.idle_cpus_mask))
10500 if (!time_after(jiffies, READ_ONCE(rq->last_blocked_load_update_tick)))
10503 update_blocked_averages(cpu);
10505 return rq->has_blocked_load;
10509 * Internal function that runs load balance for all idle cpus. The load balance
10510 * can be a simple update of blocked load or a complete load balance with
10511 * tasks movement depending of flags.
10513 static void _nohz_idle_balance(struct rq *this_rq, unsigned int flags,
10514 enum cpu_idle_type idle)
10516 /* Earliest time when we have to do rebalance again */
10517 unsigned long now = jiffies;
10518 unsigned long next_balance = now + 60*HZ;
10519 bool has_blocked_load = false;
10520 int update_next_balance = 0;
10521 int this_cpu = this_rq->cpu;
10525 SCHED_WARN_ON((flags & NOHZ_KICK_MASK) == NOHZ_BALANCE_KICK);
10528 * We assume there will be no idle load after this update and clear
10529 * the has_blocked flag. If a cpu enters idle in the mean time, it will
10530 * set the has_blocked flag and trig another update of idle load.
10531 * Because a cpu that becomes idle, is added to idle_cpus_mask before
10532 * setting the flag, we are sure to not clear the state and not
10533 * check the load of an idle cpu.
10535 WRITE_ONCE(nohz.has_blocked, 0);
10538 * Ensures that if we miss the CPU, we must see the has_blocked
10539 * store from nohz_balance_enter_idle().
10544 * Start with the next CPU after this_cpu so we will end with this_cpu and let a
10545 * chance for other idle cpu to pull load.
10547 for_each_cpu_wrap(balance_cpu, nohz.idle_cpus_mask, this_cpu+1) {
10548 if (!idle_cpu(balance_cpu))
10552 * If this CPU gets work to do, stop the load balancing
10553 * work being done for other CPUs. Next load
10554 * balancing owner will pick it up.
10556 if (need_resched()) {
10557 has_blocked_load = true;
10561 rq = cpu_rq(balance_cpu);
10563 has_blocked_load |= update_nohz_stats(rq);
10566 * If time for next balance is due,
10569 if (time_after_eq(jiffies, rq->next_balance)) {
10570 struct rq_flags rf;
10572 rq_lock_irqsave(rq, &rf);
10573 update_rq_clock(rq);
10574 rq_unlock_irqrestore(rq, &rf);
10576 if (flags & NOHZ_BALANCE_KICK)
10577 rebalance_domains(rq, CPU_IDLE);
10580 if (time_after(next_balance, rq->next_balance)) {
10581 next_balance = rq->next_balance;
10582 update_next_balance = 1;
10587 * next_balance will be updated only when there is a need.
10588 * When the CPU is attached to null domain for ex, it will not be
10591 if (likely(update_next_balance))
10592 nohz.next_balance = next_balance;
10594 WRITE_ONCE(nohz.next_blocked,
10595 now + msecs_to_jiffies(LOAD_AVG_PERIOD));
10598 /* There is still blocked load, enable periodic update */
10599 if (has_blocked_load)
10600 WRITE_ONCE(nohz.has_blocked, 1);
10604 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
10605 * rebalancing for all the cpus for whom scheduler ticks are stopped.
10607 static bool nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
10609 unsigned int flags = this_rq->nohz_idle_balance;
10614 this_rq->nohz_idle_balance = 0;
10616 if (idle != CPU_IDLE)
10619 _nohz_idle_balance(this_rq, flags, idle);
10625 * Check if we need to run the ILB for updating blocked load before entering
10628 void nohz_run_idle_balance(int cpu)
10630 unsigned int flags;
10632 flags = atomic_fetch_andnot(NOHZ_NEWILB_KICK, nohz_flags(cpu));
10635 * Update the blocked load only if no SCHED_SOFTIRQ is about to happen
10636 * (ie NOHZ_STATS_KICK set) and will do the same.
10638 if ((flags == NOHZ_NEWILB_KICK) && !need_resched())
10639 _nohz_idle_balance(cpu_rq(cpu), NOHZ_STATS_KICK, CPU_IDLE);
10642 static void nohz_newidle_balance(struct rq *this_rq)
10644 int this_cpu = this_rq->cpu;
10647 * This CPU doesn't want to be disturbed by scheduler
10650 if (!housekeeping_cpu(this_cpu, HK_FLAG_SCHED))
10653 /* Will wake up very soon. No time for doing anything else*/
10654 if (this_rq->avg_idle < sysctl_sched_migration_cost)
10657 /* Don't need to update blocked load of idle CPUs*/
10658 if (!READ_ONCE(nohz.has_blocked) ||
10659 time_before(jiffies, READ_ONCE(nohz.next_blocked)))
10663 * Set the need to trigger ILB in order to update blocked load
10664 * before entering idle state.
10666 atomic_or(NOHZ_NEWILB_KICK, nohz_flags(this_cpu));
10669 #else /* !CONFIG_NO_HZ_COMMON */
10670 static inline void nohz_balancer_kick(struct rq *rq) { }
10672 static inline bool nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
10677 static inline void nohz_newidle_balance(struct rq *this_rq) { }
10678 #endif /* CONFIG_NO_HZ_COMMON */
10681 * newidle_balance is called by schedule() if this_cpu is about to become
10682 * idle. Attempts to pull tasks from other CPUs.
10685 * < 0 - we released the lock and there are !fair tasks present
10686 * 0 - failed, no new tasks
10687 * > 0 - success, new (fair) tasks present
10689 static int newidle_balance(struct rq *this_rq, struct rq_flags *rf)
10691 unsigned long next_balance = jiffies + HZ;
10692 int this_cpu = this_rq->cpu;
10693 struct sched_domain *sd;
10694 int pulled_task = 0;
10697 update_misfit_status(NULL, this_rq);
10700 * There is a task waiting to run. No need to search for one.
10701 * Return 0; the task will be enqueued when switching to idle.
10703 if (this_rq->ttwu_pending)
10707 * We must set idle_stamp _before_ calling idle_balance(), such that we
10708 * measure the duration of idle_balance() as idle time.
10710 this_rq->idle_stamp = rq_clock(this_rq);
10713 * Do not pull tasks towards !active CPUs...
10715 if (!cpu_active(this_cpu))
10719 * This is OK, because current is on_cpu, which avoids it being picked
10720 * for load-balance and preemption/IRQs are still disabled avoiding
10721 * further scheduler activity on it and we're being very careful to
10722 * re-start the picking loop.
10724 rq_unpin_lock(this_rq, rf);
10726 if (this_rq->avg_idle < sysctl_sched_migration_cost ||
10727 !READ_ONCE(this_rq->rd->overload)) {
10730 sd = rcu_dereference_check_sched_domain(this_rq->sd);
10732 update_next_balance(sd, &next_balance);
10738 raw_spin_rq_unlock(this_rq);
10740 update_blocked_averages(this_cpu);
10742 for_each_domain(this_cpu, sd) {
10743 int continue_balancing = 1;
10744 u64 t0, domain_cost;
10746 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost) {
10747 update_next_balance(sd, &next_balance);
10751 if (sd->flags & SD_BALANCE_NEWIDLE) {
10752 t0 = sched_clock_cpu(this_cpu);
10754 pulled_task = load_balance(this_cpu, this_rq,
10755 sd, CPU_NEWLY_IDLE,
10756 &continue_balancing);
10758 domain_cost = sched_clock_cpu(this_cpu) - t0;
10759 if (domain_cost > sd->max_newidle_lb_cost)
10760 sd->max_newidle_lb_cost = domain_cost;
10762 curr_cost += domain_cost;
10765 update_next_balance(sd, &next_balance);
10768 * Stop searching for tasks to pull if there are
10769 * now runnable tasks on this rq.
10771 if (pulled_task || this_rq->nr_running > 0 ||
10772 this_rq->ttwu_pending)
10777 raw_spin_rq_lock(this_rq);
10779 if (curr_cost > this_rq->max_idle_balance_cost)
10780 this_rq->max_idle_balance_cost = curr_cost;
10783 * While browsing the domains, we released the rq lock, a task could
10784 * have been enqueued in the meantime. Since we're not going idle,
10785 * pretend we pulled a task.
10787 if (this_rq->cfs.h_nr_running && !pulled_task)
10790 /* Is there a task of a high priority class? */
10791 if (this_rq->nr_running != this_rq->cfs.h_nr_running)
10795 /* Move the next balance forward */
10796 if (time_after(this_rq->next_balance, next_balance))
10797 this_rq->next_balance = next_balance;
10800 this_rq->idle_stamp = 0;
10802 nohz_newidle_balance(this_rq);
10804 rq_repin_lock(this_rq, rf);
10806 return pulled_task;
10810 * run_rebalance_domains is triggered when needed from the scheduler tick.
10811 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
10813 static __latent_entropy void run_rebalance_domains(struct softirq_action *h)
10815 struct rq *this_rq = this_rq();
10816 enum cpu_idle_type idle = this_rq->idle_balance ?
10817 CPU_IDLE : CPU_NOT_IDLE;
10820 * If this CPU has a pending nohz_balance_kick, then do the
10821 * balancing on behalf of the other idle CPUs whose ticks are
10822 * stopped. Do nohz_idle_balance *before* rebalance_domains to
10823 * give the idle CPUs a chance to load balance. Else we may
10824 * load balance only within the local sched_domain hierarchy
10825 * and abort nohz_idle_balance altogether if we pull some load.
10827 if (nohz_idle_balance(this_rq, idle))
10830 /* normal load balance */
10831 update_blocked_averages(this_rq->cpu);
10832 rebalance_domains(this_rq, idle);
10836 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
10838 void trigger_load_balance(struct rq *rq)
10841 * Don't need to rebalance while attached to NULL domain or
10842 * runqueue CPU is not active
10844 if (unlikely(on_null_domain(rq) || !cpu_active(cpu_of(rq))))
10847 if (time_after_eq(jiffies, rq->next_balance))
10848 raise_softirq(SCHED_SOFTIRQ);
10850 nohz_balancer_kick(rq);
10853 static void rq_online_fair(struct rq *rq)
10857 update_runtime_enabled(rq);
10860 static void rq_offline_fair(struct rq *rq)
10864 /* Ensure any throttled groups are reachable by pick_next_task */
10865 unthrottle_offline_cfs_rqs(rq);
10868 #endif /* CONFIG_SMP */
10870 #ifdef CONFIG_SCHED_CORE
10872 __entity_slice_used(struct sched_entity *se, int min_nr_tasks)
10874 u64 slice = sched_slice(cfs_rq_of(se), se);
10875 u64 rtime = se->sum_exec_runtime - se->prev_sum_exec_runtime;
10877 return (rtime * min_nr_tasks > slice);
10880 #define MIN_NR_TASKS_DURING_FORCEIDLE 2
10881 static inline void task_tick_core(struct rq *rq, struct task_struct *curr)
10883 if (!sched_core_enabled(rq))
10887 * If runqueue has only one task which used up its slice and
10888 * if the sibling is forced idle, then trigger schedule to
10889 * give forced idle task a chance.
10891 * sched_slice() considers only this active rq and it gets the
10892 * whole slice. But during force idle, we have siblings acting
10893 * like a single runqueue and hence we need to consider runnable
10894 * tasks on this CPU and the forced idle CPU. Ideally, we should
10895 * go through the forced idle rq, but that would be a perf hit.
10896 * We can assume that the forced idle CPU has at least
10897 * MIN_NR_TASKS_DURING_FORCEIDLE - 1 tasks and use that to check
10898 * if we need to give up the CPU.
10900 if (rq->core->core_forceidle && rq->cfs.nr_running == 1 &&
10901 __entity_slice_used(&curr->se, MIN_NR_TASKS_DURING_FORCEIDLE))
10906 * se_fi_update - Update the cfs_rq->min_vruntime_fi in a CFS hierarchy if needed.
10908 static void se_fi_update(struct sched_entity *se, unsigned int fi_seq, bool forceidle)
10910 for_each_sched_entity(se) {
10911 struct cfs_rq *cfs_rq = cfs_rq_of(se);
10914 if (cfs_rq->forceidle_seq == fi_seq)
10916 cfs_rq->forceidle_seq = fi_seq;
10919 cfs_rq->min_vruntime_fi = cfs_rq->min_vruntime;
10923 void task_vruntime_update(struct rq *rq, struct task_struct *p, bool in_fi)
10925 struct sched_entity *se = &p->se;
10927 if (p->sched_class != &fair_sched_class)
10930 se_fi_update(se, rq->core->core_forceidle_seq, in_fi);
10933 bool cfs_prio_less(struct task_struct *a, struct task_struct *b, bool in_fi)
10935 struct rq *rq = task_rq(a);
10936 struct sched_entity *sea = &a->se;
10937 struct sched_entity *seb = &b->se;
10938 struct cfs_rq *cfs_rqa;
10939 struct cfs_rq *cfs_rqb;
10942 SCHED_WARN_ON(task_rq(b)->core != rq->core);
10944 #ifdef CONFIG_FAIR_GROUP_SCHED
10946 * Find an se in the hierarchy for tasks a and b, such that the se's
10947 * are immediate siblings.
10949 while (sea->cfs_rq->tg != seb->cfs_rq->tg) {
10950 int sea_depth = sea->depth;
10951 int seb_depth = seb->depth;
10953 if (sea_depth >= seb_depth)
10954 sea = parent_entity(sea);
10955 if (sea_depth <= seb_depth)
10956 seb = parent_entity(seb);
10959 se_fi_update(sea, rq->core->core_forceidle_seq, in_fi);
10960 se_fi_update(seb, rq->core->core_forceidle_seq, in_fi);
10962 cfs_rqa = sea->cfs_rq;
10963 cfs_rqb = seb->cfs_rq;
10965 cfs_rqa = &task_rq(a)->cfs;
10966 cfs_rqb = &task_rq(b)->cfs;
10970 * Find delta after normalizing se's vruntime with its cfs_rq's
10971 * min_vruntime_fi, which would have been updated in prior calls
10972 * to se_fi_update().
10974 delta = (s64)(sea->vruntime - seb->vruntime) +
10975 (s64)(cfs_rqb->min_vruntime_fi - cfs_rqa->min_vruntime_fi);
10980 static inline void task_tick_core(struct rq *rq, struct task_struct *curr) {}
10984 * scheduler tick hitting a task of our scheduling class.
10986 * NOTE: This function can be called remotely by the tick offload that
10987 * goes along full dynticks. Therefore no local assumption can be made
10988 * and everything must be accessed through the @rq and @curr passed in
10991 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
10993 struct cfs_rq *cfs_rq;
10994 struct sched_entity *se = &curr->se;
10996 for_each_sched_entity(se) {
10997 cfs_rq = cfs_rq_of(se);
10998 entity_tick(cfs_rq, se, queued);
11001 if (static_branch_unlikely(&sched_numa_balancing))
11002 task_tick_numa(rq, curr);
11004 update_misfit_status(curr, rq);
11005 update_overutilized_status(task_rq(curr));
11007 task_tick_core(rq, curr);
11011 * called on fork with the child task as argument from the parent's context
11012 * - child not yet on the tasklist
11013 * - preemption disabled
11015 static void task_fork_fair(struct task_struct *p)
11017 struct cfs_rq *cfs_rq;
11018 struct sched_entity *se = &p->se, *curr;
11019 struct rq *rq = this_rq();
11020 struct rq_flags rf;
11023 update_rq_clock(rq);
11025 cfs_rq = task_cfs_rq(current);
11026 curr = cfs_rq->curr;
11028 update_curr(cfs_rq);
11029 se->vruntime = curr->vruntime;
11031 place_entity(cfs_rq, se, 1);
11033 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
11035 * Upon rescheduling, sched_class::put_prev_task() will place
11036 * 'current' within the tree based on its new key value.
11038 swap(curr->vruntime, se->vruntime);
11042 se->vruntime -= cfs_rq->min_vruntime;
11043 rq_unlock(rq, &rf);
11047 * Priority of the task has changed. Check to see if we preempt
11048 * the current task.
11051 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
11053 if (!task_on_rq_queued(p))
11056 if (rq->cfs.nr_running == 1)
11060 * Reschedule if we are currently running on this runqueue and
11061 * our priority decreased, or if we are not currently running on
11062 * this runqueue and our priority is higher than the current's
11064 if (task_current(rq, p)) {
11065 if (p->prio > oldprio)
11068 check_preempt_curr(rq, p, 0);
11071 static inline bool vruntime_normalized(struct task_struct *p)
11073 struct sched_entity *se = &p->se;
11076 * In both the TASK_ON_RQ_QUEUED and TASK_ON_RQ_MIGRATING cases,
11077 * the dequeue_entity(.flags=0) will already have normalized the
11084 * When !on_rq, vruntime of the task has usually NOT been normalized.
11085 * But there are some cases where it has already been normalized:
11087 * - A forked child which is waiting for being woken up by
11088 * wake_up_new_task().
11089 * - A task which has been woken up by try_to_wake_up() and
11090 * waiting for actually being woken up by sched_ttwu_pending().
11092 if (!se->sum_exec_runtime ||
11093 (READ_ONCE(p->__state) == TASK_WAKING && p->sched_remote_wakeup))
11099 #ifdef CONFIG_FAIR_GROUP_SCHED
11101 * Propagate the changes of the sched_entity across the tg tree to make it
11102 * visible to the root
11104 static void propagate_entity_cfs_rq(struct sched_entity *se)
11106 struct cfs_rq *cfs_rq;
11108 list_add_leaf_cfs_rq(cfs_rq_of(se));
11110 /* Start to propagate at parent */
11113 for_each_sched_entity(se) {
11114 cfs_rq = cfs_rq_of(se);
11116 if (!cfs_rq_throttled(cfs_rq)){
11117 update_load_avg(cfs_rq, se, UPDATE_TG);
11118 list_add_leaf_cfs_rq(cfs_rq);
11122 if (list_add_leaf_cfs_rq(cfs_rq))
11127 static void propagate_entity_cfs_rq(struct sched_entity *se) { }
11130 static void detach_entity_cfs_rq(struct sched_entity *se)
11132 struct cfs_rq *cfs_rq = cfs_rq_of(se);
11134 /* Catch up with the cfs_rq and remove our load when we leave */
11135 update_load_avg(cfs_rq, se, 0);
11136 detach_entity_load_avg(cfs_rq, se);
11137 update_tg_load_avg(cfs_rq);
11138 propagate_entity_cfs_rq(se);
11141 static void attach_entity_cfs_rq(struct sched_entity *se)
11143 struct cfs_rq *cfs_rq = cfs_rq_of(se);
11145 #ifdef CONFIG_FAIR_GROUP_SCHED
11147 * Since the real-depth could have been changed (only FAIR
11148 * class maintain depth value), reset depth properly.
11150 se->depth = se->parent ? se->parent->depth + 1 : 0;
11153 /* Synchronize entity with its cfs_rq */
11154 update_load_avg(cfs_rq, se, sched_feat(ATTACH_AGE_LOAD) ? 0 : SKIP_AGE_LOAD);
11155 attach_entity_load_avg(cfs_rq, se);
11156 update_tg_load_avg(cfs_rq);
11157 propagate_entity_cfs_rq(se);
11160 static void detach_task_cfs_rq(struct task_struct *p)
11162 struct sched_entity *se = &p->se;
11163 struct cfs_rq *cfs_rq = cfs_rq_of(se);
11165 if (!vruntime_normalized(p)) {
11167 * Fix up our vruntime so that the current sleep doesn't
11168 * cause 'unlimited' sleep bonus.
11170 place_entity(cfs_rq, se, 0);
11171 se->vruntime -= cfs_rq->min_vruntime;
11174 detach_entity_cfs_rq(se);
11177 static void attach_task_cfs_rq(struct task_struct *p)
11179 struct sched_entity *se = &p->se;
11180 struct cfs_rq *cfs_rq = cfs_rq_of(se);
11182 attach_entity_cfs_rq(se);
11184 if (!vruntime_normalized(p))
11185 se->vruntime += cfs_rq->min_vruntime;
11188 static void switched_from_fair(struct rq *rq, struct task_struct *p)
11190 detach_task_cfs_rq(p);
11193 static void switched_to_fair(struct rq *rq, struct task_struct *p)
11195 attach_task_cfs_rq(p);
11197 if (task_on_rq_queued(p)) {
11199 * We were most likely switched from sched_rt, so
11200 * kick off the schedule if running, otherwise just see
11201 * if we can still preempt the current task.
11203 if (task_current(rq, p))
11206 check_preempt_curr(rq, p, 0);
11210 /* Account for a task changing its policy or group.
11212 * This routine is mostly called to set cfs_rq->curr field when a task
11213 * migrates between groups/classes.
11215 static void set_next_task_fair(struct rq *rq, struct task_struct *p, bool first)
11217 struct sched_entity *se = &p->se;
11220 if (task_on_rq_queued(p)) {
11222 * Move the next running task to the front of the list, so our
11223 * cfs_tasks list becomes MRU one.
11225 list_move(&se->group_node, &rq->cfs_tasks);
11229 for_each_sched_entity(se) {
11230 struct cfs_rq *cfs_rq = cfs_rq_of(se);
11232 set_next_entity(cfs_rq, se);
11233 /* ensure bandwidth has been allocated on our new cfs_rq */
11234 account_cfs_rq_runtime(cfs_rq, 0);
11238 void init_cfs_rq(struct cfs_rq *cfs_rq)
11240 cfs_rq->tasks_timeline = RB_ROOT_CACHED;
11241 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
11242 #ifndef CONFIG_64BIT
11243 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
11246 raw_spin_lock_init(&cfs_rq->removed.lock);
11250 #ifdef CONFIG_FAIR_GROUP_SCHED
11251 static void task_set_group_fair(struct task_struct *p)
11253 struct sched_entity *se = &p->se;
11255 set_task_rq(p, task_cpu(p));
11256 se->depth = se->parent ? se->parent->depth + 1 : 0;
11259 static void task_move_group_fair(struct task_struct *p)
11261 detach_task_cfs_rq(p);
11262 set_task_rq(p, task_cpu(p));
11265 /* Tell se's cfs_rq has been changed -- migrated */
11266 p->se.avg.last_update_time = 0;
11268 attach_task_cfs_rq(p);
11271 static void task_change_group_fair(struct task_struct *p, int type)
11274 case TASK_SET_GROUP:
11275 task_set_group_fair(p);
11278 case TASK_MOVE_GROUP:
11279 task_move_group_fair(p);
11284 void free_fair_sched_group(struct task_group *tg)
11288 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
11290 for_each_possible_cpu(i) {
11292 kfree(tg->cfs_rq[i]);
11301 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
11303 struct sched_entity *se;
11304 struct cfs_rq *cfs_rq;
11307 tg->cfs_rq = kcalloc(nr_cpu_ids, sizeof(cfs_rq), GFP_KERNEL);
11310 tg->se = kcalloc(nr_cpu_ids, sizeof(se), GFP_KERNEL);
11314 tg->shares = NICE_0_LOAD;
11316 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
11318 for_each_possible_cpu(i) {
11319 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
11320 GFP_KERNEL, cpu_to_node(i));
11324 se = kzalloc_node(sizeof(struct sched_entity),
11325 GFP_KERNEL, cpu_to_node(i));
11329 init_cfs_rq(cfs_rq);
11330 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
11331 init_entity_runnable_average(se);
11342 void online_fair_sched_group(struct task_group *tg)
11344 struct sched_entity *se;
11345 struct rq_flags rf;
11349 for_each_possible_cpu(i) {
11352 rq_lock_irq(rq, &rf);
11353 update_rq_clock(rq);
11354 attach_entity_cfs_rq(se);
11355 sync_throttle(tg, i);
11356 rq_unlock_irq(rq, &rf);
11360 void unregister_fair_sched_group(struct task_group *tg)
11362 unsigned long flags;
11366 for_each_possible_cpu(cpu) {
11368 remove_entity_load_avg(tg->se[cpu]);
11371 * Only empty task groups can be destroyed; so we can speculatively
11372 * check on_list without danger of it being re-added.
11374 if (!tg->cfs_rq[cpu]->on_list)
11379 raw_spin_rq_lock_irqsave(rq, flags);
11380 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
11381 raw_spin_rq_unlock_irqrestore(rq, flags);
11385 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
11386 struct sched_entity *se, int cpu,
11387 struct sched_entity *parent)
11389 struct rq *rq = cpu_rq(cpu);
11393 init_cfs_rq_runtime(cfs_rq);
11395 tg->cfs_rq[cpu] = cfs_rq;
11398 /* se could be NULL for root_task_group */
11403 se->cfs_rq = &rq->cfs;
11406 se->cfs_rq = parent->my_q;
11407 se->depth = parent->depth + 1;
11411 /* guarantee group entities always have weight */
11412 update_load_set(&se->load, NICE_0_LOAD);
11413 se->parent = parent;
11416 static DEFINE_MUTEX(shares_mutex);
11418 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
11423 * We can't change the weight of the root cgroup.
11428 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
11430 mutex_lock(&shares_mutex);
11431 if (tg->shares == shares)
11434 tg->shares = shares;
11435 for_each_possible_cpu(i) {
11436 struct rq *rq = cpu_rq(i);
11437 struct sched_entity *se = tg->se[i];
11438 struct rq_flags rf;
11440 /* Propagate contribution to hierarchy */
11441 rq_lock_irqsave(rq, &rf);
11442 update_rq_clock(rq);
11443 for_each_sched_entity(se) {
11444 update_load_avg(cfs_rq_of(se), se, UPDATE_TG);
11445 update_cfs_group(se);
11447 rq_unlock_irqrestore(rq, &rf);
11451 mutex_unlock(&shares_mutex);
11454 #else /* CONFIG_FAIR_GROUP_SCHED */
11456 void free_fair_sched_group(struct task_group *tg) { }
11458 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
11463 void online_fair_sched_group(struct task_group *tg) { }
11465 void unregister_fair_sched_group(struct task_group *tg) { }
11467 #endif /* CONFIG_FAIR_GROUP_SCHED */
11470 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
11472 struct sched_entity *se = &task->se;
11473 unsigned int rr_interval = 0;
11476 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
11479 if (rq->cfs.load.weight)
11480 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
11482 return rr_interval;
11486 * All the scheduling class methods:
11488 DEFINE_SCHED_CLASS(fair) = {
11490 .enqueue_task = enqueue_task_fair,
11491 .dequeue_task = dequeue_task_fair,
11492 .yield_task = yield_task_fair,
11493 .yield_to_task = yield_to_task_fair,
11495 .check_preempt_curr = check_preempt_wakeup,
11497 .pick_next_task = __pick_next_task_fair,
11498 .put_prev_task = put_prev_task_fair,
11499 .set_next_task = set_next_task_fair,
11502 .balance = balance_fair,
11503 .pick_task = pick_task_fair,
11504 .select_task_rq = select_task_rq_fair,
11505 .migrate_task_rq = migrate_task_rq_fair,
11507 .rq_online = rq_online_fair,
11508 .rq_offline = rq_offline_fair,
11510 .task_dead = task_dead_fair,
11511 .set_cpus_allowed = set_cpus_allowed_common,
11514 .task_tick = task_tick_fair,
11515 .task_fork = task_fork_fair,
11517 .prio_changed = prio_changed_fair,
11518 .switched_from = switched_from_fair,
11519 .switched_to = switched_to_fair,
11521 .get_rr_interval = get_rr_interval_fair,
11523 .update_curr = update_curr_fair,
11525 #ifdef CONFIG_FAIR_GROUP_SCHED
11526 .task_change_group = task_change_group_fair,
11529 #ifdef CONFIG_UCLAMP_TASK
11530 .uclamp_enabled = 1,
11534 #ifdef CONFIG_SCHED_DEBUG
11535 void print_cfs_stats(struct seq_file *m, int cpu)
11537 struct cfs_rq *cfs_rq, *pos;
11540 for_each_leaf_cfs_rq_safe(cpu_rq(cpu), cfs_rq, pos)
11541 print_cfs_rq(m, cpu, cfs_rq);
11545 #ifdef CONFIG_NUMA_BALANCING
11546 void show_numa_stats(struct task_struct *p, struct seq_file *m)
11549 unsigned long tsf = 0, tpf = 0, gsf = 0, gpf = 0;
11550 struct numa_group *ng;
11553 ng = rcu_dereference(p->numa_group);
11554 for_each_online_node(node) {
11555 if (p->numa_faults) {
11556 tsf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 0)];
11557 tpf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 1)];
11560 gsf = ng->faults[task_faults_idx(NUMA_MEM, node, 0)],
11561 gpf = ng->faults[task_faults_idx(NUMA_MEM, node, 1)];
11563 print_numa_stats(m, node, tsf, tpf, gsf, gpf);
11567 #endif /* CONFIG_NUMA_BALANCING */
11568 #endif /* CONFIG_SCHED_DEBUG */
11570 __init void init_sched_fair_class(void)
11573 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
11575 #ifdef CONFIG_NO_HZ_COMMON
11576 nohz.next_balance = jiffies;
11577 nohz.next_blocked = jiffies;
11578 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
11585 * Helper functions to facilitate extracting info from tracepoints.
11588 const struct sched_avg *sched_trace_cfs_rq_avg(struct cfs_rq *cfs_rq)
11591 return cfs_rq ? &cfs_rq->avg : NULL;
11596 EXPORT_SYMBOL_GPL(sched_trace_cfs_rq_avg);
11598 char *sched_trace_cfs_rq_path(struct cfs_rq *cfs_rq, char *str, int len)
11602 strlcpy(str, "(null)", len);
11607 cfs_rq_tg_path(cfs_rq, str, len);
11610 EXPORT_SYMBOL_GPL(sched_trace_cfs_rq_path);
11612 int sched_trace_cfs_rq_cpu(struct cfs_rq *cfs_rq)
11614 return cfs_rq ? cpu_of(rq_of(cfs_rq)) : -1;
11616 EXPORT_SYMBOL_GPL(sched_trace_cfs_rq_cpu);
11618 const struct sched_avg *sched_trace_rq_avg_rt(struct rq *rq)
11621 return rq ? &rq->avg_rt : NULL;
11626 EXPORT_SYMBOL_GPL(sched_trace_rq_avg_rt);
11628 const struct sched_avg *sched_trace_rq_avg_dl(struct rq *rq)
11631 return rq ? &rq->avg_dl : NULL;
11636 EXPORT_SYMBOL_GPL(sched_trace_rq_avg_dl);
11638 const struct sched_avg *sched_trace_rq_avg_irq(struct rq *rq)
11640 #if defined(CONFIG_SMP) && defined(CONFIG_HAVE_SCHED_AVG_IRQ)
11641 return rq ? &rq->avg_irq : NULL;
11646 EXPORT_SYMBOL_GPL(sched_trace_rq_avg_irq);
11648 int sched_trace_rq_cpu(struct rq *rq)
11650 return rq ? cpu_of(rq) : -1;
11652 EXPORT_SYMBOL_GPL(sched_trace_rq_cpu);
11654 int sched_trace_rq_cpu_capacity(struct rq *rq)
11660 SCHED_CAPACITY_SCALE
11664 EXPORT_SYMBOL_GPL(sched_trace_rq_cpu_capacity);
11666 const struct cpumask *sched_trace_rd_span(struct root_domain *rd)
11669 return rd ? rd->span : NULL;
11674 EXPORT_SYMBOL_GPL(sched_trace_rd_span);
11676 int sched_trace_rq_nr_running(struct rq *rq)
11678 return rq ? rq->nr_running : -1;
11680 EXPORT_SYMBOL_GPL(sched_trace_rq_nr_running);