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 static int tg_is_idle(struct task_group *tg)
439 static int cfs_rq_is_idle(struct cfs_rq *cfs_rq)
441 return cfs_rq->idle > 0;
444 static int se_is_idle(struct sched_entity *se)
446 if (entity_is_task(se))
447 return task_has_idle_policy(task_of(se));
448 return cfs_rq_is_idle(group_cfs_rq(se));
451 #else /* !CONFIG_FAIR_GROUP_SCHED */
453 #define for_each_sched_entity(se) \
454 for (; se; se = NULL)
456 static inline void cfs_rq_tg_path(struct cfs_rq *cfs_rq, char *path, int len)
459 strlcpy(path, "(null)", len);
462 static inline bool list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
467 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
471 static inline void assert_list_leaf_cfs_rq(struct rq *rq)
475 #define for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) \
476 for (cfs_rq = &rq->cfs, pos = NULL; cfs_rq; cfs_rq = pos)
478 static inline struct sched_entity *parent_entity(struct sched_entity *se)
484 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
488 static int tg_is_idle(struct task_group *tg)
493 static int cfs_rq_is_idle(struct cfs_rq *cfs_rq)
498 static int se_is_idle(struct sched_entity *se)
503 #endif /* CONFIG_FAIR_GROUP_SCHED */
505 static __always_inline
506 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec);
508 /**************************************************************
509 * Scheduling class tree data structure manipulation methods:
512 static inline u64 max_vruntime(u64 max_vruntime, u64 vruntime)
514 s64 delta = (s64)(vruntime - max_vruntime);
516 max_vruntime = vruntime;
521 static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime)
523 s64 delta = (s64)(vruntime - min_vruntime);
525 min_vruntime = vruntime;
530 static inline bool entity_before(struct sched_entity *a,
531 struct sched_entity *b)
533 return (s64)(a->vruntime - b->vruntime) < 0;
536 #define __node_2_se(node) \
537 rb_entry((node), struct sched_entity, run_node)
539 static void update_min_vruntime(struct cfs_rq *cfs_rq)
541 struct sched_entity *curr = cfs_rq->curr;
542 struct rb_node *leftmost = rb_first_cached(&cfs_rq->tasks_timeline);
544 u64 vruntime = cfs_rq->min_vruntime;
548 vruntime = curr->vruntime;
553 if (leftmost) { /* non-empty tree */
554 struct sched_entity *se = __node_2_se(leftmost);
557 vruntime = se->vruntime;
559 vruntime = min_vruntime(vruntime, se->vruntime);
562 /* ensure we never gain time by being placed backwards. */
563 cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime);
566 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
570 static inline bool __entity_less(struct rb_node *a, const struct rb_node *b)
572 return entity_before(__node_2_se(a), __node_2_se(b));
576 * Enqueue an entity into the rb-tree:
578 static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
580 rb_add_cached(&se->run_node, &cfs_rq->tasks_timeline, __entity_less);
583 static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
585 rb_erase_cached(&se->run_node, &cfs_rq->tasks_timeline);
588 struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq)
590 struct rb_node *left = rb_first_cached(&cfs_rq->tasks_timeline);
595 return __node_2_se(left);
598 static struct sched_entity *__pick_next_entity(struct sched_entity *se)
600 struct rb_node *next = rb_next(&se->run_node);
605 return __node_2_se(next);
608 #ifdef CONFIG_SCHED_DEBUG
609 struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
611 struct rb_node *last = rb_last(&cfs_rq->tasks_timeline.rb_root);
616 return __node_2_se(last);
619 /**************************************************************
620 * Scheduling class statistics methods:
623 int sched_update_scaling(void)
625 unsigned int factor = get_update_sysctl_factor();
627 sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency,
628 sysctl_sched_min_granularity);
630 #define WRT_SYSCTL(name) \
631 (normalized_sysctl_##name = sysctl_##name / (factor))
632 WRT_SYSCTL(sched_min_granularity);
633 WRT_SYSCTL(sched_latency);
634 WRT_SYSCTL(sched_wakeup_granularity);
644 static inline u64 calc_delta_fair(u64 delta, struct sched_entity *se)
646 if (unlikely(se->load.weight != NICE_0_LOAD))
647 delta = __calc_delta(delta, NICE_0_LOAD, &se->load);
653 * The idea is to set a period in which each task runs once.
655 * When there are too many tasks (sched_nr_latency) we have to stretch
656 * this period because otherwise the slices get too small.
658 * p = (nr <= nl) ? l : l*nr/nl
660 static u64 __sched_period(unsigned long nr_running)
662 if (unlikely(nr_running > sched_nr_latency))
663 return nr_running * sysctl_sched_min_granularity;
665 return sysctl_sched_latency;
669 * We calculate the wall-time slice from the period by taking a part
670 * proportional to the weight.
674 static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
676 unsigned int nr_running = cfs_rq->nr_running;
679 if (sched_feat(ALT_PERIOD))
680 nr_running = rq_of(cfs_rq)->cfs.h_nr_running;
682 slice = __sched_period(nr_running + !se->on_rq);
684 for_each_sched_entity(se) {
685 struct load_weight *load;
686 struct load_weight lw;
688 cfs_rq = cfs_rq_of(se);
689 load = &cfs_rq->load;
691 if (unlikely(!se->on_rq)) {
694 update_load_add(&lw, se->load.weight);
697 slice = __calc_delta(slice, se->load.weight, load);
700 if (sched_feat(BASE_SLICE))
701 slice = max(slice, (u64)sysctl_sched_min_granularity);
707 * We calculate the vruntime slice of a to-be-inserted task.
711 static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se)
713 return calc_delta_fair(sched_slice(cfs_rq, se), se);
719 static int select_idle_sibling(struct task_struct *p, int prev_cpu, int cpu);
720 static unsigned long task_h_load(struct task_struct *p);
721 static unsigned long capacity_of(int cpu);
723 /* Give new sched_entity start runnable values to heavy its load in infant time */
724 void init_entity_runnable_average(struct sched_entity *se)
726 struct sched_avg *sa = &se->avg;
728 memset(sa, 0, sizeof(*sa));
731 * Tasks are initialized with full load to be seen as heavy tasks until
732 * they get a chance to stabilize to their real load level.
733 * Group entities are initialized with zero load to reflect the fact that
734 * nothing has been attached to the task group yet.
736 if (entity_is_task(se))
737 sa->load_avg = scale_load_down(se->load.weight);
739 /* when this task enqueue'ed, it will contribute to its cfs_rq's load_avg */
742 static void attach_entity_cfs_rq(struct sched_entity *se);
745 * With new tasks being created, their initial util_avgs are extrapolated
746 * based on the cfs_rq's current util_avg:
748 * util_avg = cfs_rq->util_avg / (cfs_rq->load_avg + 1) * se.load.weight
750 * However, in many cases, the above util_avg does not give a desired
751 * value. Moreover, the sum of the util_avgs may be divergent, such
752 * as when the series is a harmonic series.
754 * To solve this problem, we also cap the util_avg of successive tasks to
755 * only 1/2 of the left utilization budget:
757 * util_avg_cap = (cpu_scale - cfs_rq->avg.util_avg) / 2^n
759 * where n denotes the nth task and cpu_scale the CPU capacity.
761 * For example, for a CPU with 1024 of capacity, a simplest series from
762 * the beginning would be like:
764 * task util_avg: 512, 256, 128, 64, 32, 16, 8, ...
765 * cfs_rq util_avg: 512, 768, 896, 960, 992, 1008, 1016, ...
767 * Finally, that extrapolated util_avg is clamped to the cap (util_avg_cap)
768 * if util_avg > util_avg_cap.
770 void post_init_entity_util_avg(struct task_struct *p)
772 struct sched_entity *se = &p->se;
773 struct cfs_rq *cfs_rq = cfs_rq_of(se);
774 struct sched_avg *sa = &se->avg;
775 long cpu_scale = arch_scale_cpu_capacity(cpu_of(rq_of(cfs_rq)));
776 long cap = (long)(cpu_scale - cfs_rq->avg.util_avg) / 2;
779 if (cfs_rq->avg.util_avg != 0) {
780 sa->util_avg = cfs_rq->avg.util_avg * se->load.weight;
781 sa->util_avg /= (cfs_rq->avg.load_avg + 1);
783 if (sa->util_avg > cap)
790 sa->runnable_avg = sa->util_avg;
792 if (p->sched_class != &fair_sched_class) {
794 * For !fair tasks do:
796 update_cfs_rq_load_avg(now, cfs_rq);
797 attach_entity_load_avg(cfs_rq, se);
798 switched_from_fair(rq, p);
800 * such that the next switched_to_fair() has the
803 se->avg.last_update_time = cfs_rq_clock_pelt(cfs_rq);
807 attach_entity_cfs_rq(se);
810 #else /* !CONFIG_SMP */
811 void init_entity_runnable_average(struct sched_entity *se)
814 void post_init_entity_util_avg(struct task_struct *p)
817 static void update_tg_load_avg(struct cfs_rq *cfs_rq)
820 #endif /* CONFIG_SMP */
823 * Update the current task's runtime statistics.
825 static void update_curr(struct cfs_rq *cfs_rq)
827 struct sched_entity *curr = cfs_rq->curr;
828 u64 now = rq_clock_task(rq_of(cfs_rq));
834 delta_exec = now - curr->exec_start;
835 if (unlikely((s64)delta_exec <= 0))
838 curr->exec_start = now;
840 schedstat_set(curr->statistics.exec_max,
841 max(delta_exec, curr->statistics.exec_max));
843 curr->sum_exec_runtime += delta_exec;
844 schedstat_add(cfs_rq->exec_clock, delta_exec);
846 curr->vruntime += calc_delta_fair(delta_exec, curr);
847 update_min_vruntime(cfs_rq);
849 if (entity_is_task(curr)) {
850 struct task_struct *curtask = task_of(curr);
852 trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime);
853 cgroup_account_cputime(curtask, delta_exec);
854 account_group_exec_runtime(curtask, delta_exec);
857 account_cfs_rq_runtime(cfs_rq, delta_exec);
860 static void update_curr_fair(struct rq *rq)
862 update_curr(cfs_rq_of(&rq->curr->se));
866 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
868 u64 wait_start, prev_wait_start;
870 if (!schedstat_enabled())
873 wait_start = rq_clock(rq_of(cfs_rq));
874 prev_wait_start = schedstat_val(se->statistics.wait_start);
876 if (entity_is_task(se) && task_on_rq_migrating(task_of(se)) &&
877 likely(wait_start > prev_wait_start))
878 wait_start -= prev_wait_start;
880 __schedstat_set(se->statistics.wait_start, wait_start);
884 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
886 struct task_struct *p;
889 if (!schedstat_enabled())
893 * When the sched_schedstat changes from 0 to 1, some sched se
894 * maybe already in the runqueue, the se->statistics.wait_start
895 * will be 0.So it will let the delta wrong. We need to avoid this
898 if (unlikely(!schedstat_val(se->statistics.wait_start)))
901 delta = rq_clock(rq_of(cfs_rq)) - schedstat_val(se->statistics.wait_start);
903 if (entity_is_task(se)) {
905 if (task_on_rq_migrating(p)) {
907 * Preserve migrating task's wait time so wait_start
908 * time stamp can be adjusted to accumulate wait time
909 * prior to migration.
911 __schedstat_set(se->statistics.wait_start, delta);
914 trace_sched_stat_wait(p, delta);
917 __schedstat_set(se->statistics.wait_max,
918 max(schedstat_val(se->statistics.wait_max), delta));
919 __schedstat_inc(se->statistics.wait_count);
920 __schedstat_add(se->statistics.wait_sum, delta);
921 __schedstat_set(se->statistics.wait_start, 0);
925 update_stats_enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
927 struct task_struct *tsk = NULL;
928 u64 sleep_start, block_start;
930 if (!schedstat_enabled())
933 sleep_start = schedstat_val(se->statistics.sleep_start);
934 block_start = schedstat_val(se->statistics.block_start);
936 if (entity_is_task(se))
940 u64 delta = rq_clock(rq_of(cfs_rq)) - sleep_start;
945 if (unlikely(delta > schedstat_val(se->statistics.sleep_max)))
946 __schedstat_set(se->statistics.sleep_max, delta);
948 __schedstat_set(se->statistics.sleep_start, 0);
949 __schedstat_add(se->statistics.sum_sleep_runtime, delta);
952 account_scheduler_latency(tsk, delta >> 10, 1);
953 trace_sched_stat_sleep(tsk, delta);
957 u64 delta = rq_clock(rq_of(cfs_rq)) - block_start;
962 if (unlikely(delta > schedstat_val(se->statistics.block_max)))
963 __schedstat_set(se->statistics.block_max, delta);
965 __schedstat_set(se->statistics.block_start, 0);
966 __schedstat_add(se->statistics.sum_sleep_runtime, delta);
969 if (tsk->in_iowait) {
970 __schedstat_add(se->statistics.iowait_sum, delta);
971 __schedstat_inc(se->statistics.iowait_count);
972 trace_sched_stat_iowait(tsk, delta);
975 trace_sched_stat_blocked(tsk, delta);
978 * Blocking time is in units of nanosecs, so shift by
979 * 20 to get a milliseconds-range estimation of the
980 * amount of time that the task spent sleeping:
982 if (unlikely(prof_on == SLEEP_PROFILING)) {
983 profile_hits(SLEEP_PROFILING,
984 (void *)get_wchan(tsk),
987 account_scheduler_latency(tsk, delta >> 10, 0);
993 * Task is being enqueued - update stats:
996 update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
998 if (!schedstat_enabled())
1002 * Are we enqueueing a waiting task? (for current tasks
1003 * a dequeue/enqueue event is a NOP)
1005 if (se != cfs_rq->curr)
1006 update_stats_wait_start(cfs_rq, se);
1008 if (flags & ENQUEUE_WAKEUP)
1009 update_stats_enqueue_sleeper(cfs_rq, se);
1013 update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
1016 if (!schedstat_enabled())
1020 * Mark the end of the wait period if dequeueing a
1023 if (se != cfs_rq->curr)
1024 update_stats_wait_end(cfs_rq, se);
1026 if ((flags & DEQUEUE_SLEEP) && entity_is_task(se)) {
1027 struct task_struct *tsk = task_of(se);
1030 /* XXX racy against TTWU */
1031 state = READ_ONCE(tsk->__state);
1032 if (state & TASK_INTERRUPTIBLE)
1033 __schedstat_set(se->statistics.sleep_start,
1034 rq_clock(rq_of(cfs_rq)));
1035 if (state & TASK_UNINTERRUPTIBLE)
1036 __schedstat_set(se->statistics.block_start,
1037 rq_clock(rq_of(cfs_rq)));
1042 * We are picking a new current task - update its stats:
1045 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
1048 * We are starting a new run period:
1050 se->exec_start = rq_clock_task(rq_of(cfs_rq));
1053 /**************************************************
1054 * Scheduling class queueing methods:
1057 #ifdef CONFIG_NUMA_BALANCING
1059 * Approximate time to scan a full NUMA task in ms. The task scan period is
1060 * calculated based on the tasks virtual memory size and
1061 * numa_balancing_scan_size.
1063 unsigned int sysctl_numa_balancing_scan_period_min = 1000;
1064 unsigned int sysctl_numa_balancing_scan_period_max = 60000;
1066 /* Portion of address space to scan in MB */
1067 unsigned int sysctl_numa_balancing_scan_size = 256;
1069 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
1070 unsigned int sysctl_numa_balancing_scan_delay = 1000;
1073 refcount_t refcount;
1075 spinlock_t lock; /* nr_tasks, tasks */
1080 struct rcu_head rcu;
1081 unsigned long total_faults;
1082 unsigned long max_faults_cpu;
1084 * Faults_cpu is used to decide whether memory should move
1085 * towards the CPU. As a consequence, these stats are weighted
1086 * more by CPU use than by memory faults.
1088 unsigned long *faults_cpu;
1089 unsigned long faults[];
1093 * For functions that can be called in multiple contexts that permit reading
1094 * ->numa_group (see struct task_struct for locking rules).
1096 static struct numa_group *deref_task_numa_group(struct task_struct *p)
1098 return rcu_dereference_check(p->numa_group, p == current ||
1099 (lockdep_is_held(__rq_lockp(task_rq(p))) && !READ_ONCE(p->on_cpu)));
1102 static struct numa_group *deref_curr_numa_group(struct task_struct *p)
1104 return rcu_dereference_protected(p->numa_group, p == current);
1107 static inline unsigned long group_faults_priv(struct numa_group *ng);
1108 static inline unsigned long group_faults_shared(struct numa_group *ng);
1110 static unsigned int task_nr_scan_windows(struct task_struct *p)
1112 unsigned long rss = 0;
1113 unsigned long nr_scan_pages;
1116 * Calculations based on RSS as non-present and empty pages are skipped
1117 * by the PTE scanner and NUMA hinting faults should be trapped based
1120 nr_scan_pages = sysctl_numa_balancing_scan_size << (20 - PAGE_SHIFT);
1121 rss = get_mm_rss(p->mm);
1123 rss = nr_scan_pages;
1125 rss = round_up(rss, nr_scan_pages);
1126 return rss / nr_scan_pages;
1129 /* For sanity's sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */
1130 #define MAX_SCAN_WINDOW 2560
1132 static unsigned int task_scan_min(struct task_struct *p)
1134 unsigned int scan_size = READ_ONCE(sysctl_numa_balancing_scan_size);
1135 unsigned int scan, floor;
1136 unsigned int windows = 1;
1138 if (scan_size < MAX_SCAN_WINDOW)
1139 windows = MAX_SCAN_WINDOW / scan_size;
1140 floor = 1000 / windows;
1142 scan = sysctl_numa_balancing_scan_period_min / task_nr_scan_windows(p);
1143 return max_t(unsigned int, floor, scan);
1146 static unsigned int task_scan_start(struct task_struct *p)
1148 unsigned long smin = task_scan_min(p);
1149 unsigned long period = smin;
1150 struct numa_group *ng;
1152 /* Scale the maximum scan period with the amount of shared memory. */
1154 ng = rcu_dereference(p->numa_group);
1156 unsigned long shared = group_faults_shared(ng);
1157 unsigned long private = group_faults_priv(ng);
1159 period *= refcount_read(&ng->refcount);
1160 period *= shared + 1;
1161 period /= private + shared + 1;
1165 return max(smin, period);
1168 static unsigned int task_scan_max(struct task_struct *p)
1170 unsigned long smin = task_scan_min(p);
1172 struct numa_group *ng;
1174 /* Watch for min being lower than max due to floor calculations */
1175 smax = sysctl_numa_balancing_scan_period_max / task_nr_scan_windows(p);
1177 /* Scale the maximum scan period with the amount of shared memory. */
1178 ng = deref_curr_numa_group(p);
1180 unsigned long shared = group_faults_shared(ng);
1181 unsigned long private = group_faults_priv(ng);
1182 unsigned long period = smax;
1184 period *= refcount_read(&ng->refcount);
1185 period *= shared + 1;
1186 period /= private + shared + 1;
1188 smax = max(smax, period);
1191 return max(smin, smax);
1194 static void account_numa_enqueue(struct rq *rq, struct task_struct *p)
1196 rq->nr_numa_running += (p->numa_preferred_nid != NUMA_NO_NODE);
1197 rq->nr_preferred_running += (p->numa_preferred_nid == task_node(p));
1200 static void account_numa_dequeue(struct rq *rq, struct task_struct *p)
1202 rq->nr_numa_running -= (p->numa_preferred_nid != NUMA_NO_NODE);
1203 rq->nr_preferred_running -= (p->numa_preferred_nid == task_node(p));
1206 /* Shared or private faults. */
1207 #define NR_NUMA_HINT_FAULT_TYPES 2
1209 /* Memory and CPU locality */
1210 #define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2)
1212 /* Averaged statistics, and temporary buffers. */
1213 #define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2)
1215 pid_t task_numa_group_id(struct task_struct *p)
1217 struct numa_group *ng;
1221 ng = rcu_dereference(p->numa_group);
1230 * The averaged statistics, shared & private, memory & CPU,
1231 * occupy the first half of the array. The second half of the
1232 * array is for current counters, which are averaged into the
1233 * first set by task_numa_placement.
1235 static inline int task_faults_idx(enum numa_faults_stats s, int nid, int priv)
1237 return NR_NUMA_HINT_FAULT_TYPES * (s * nr_node_ids + nid) + priv;
1240 static inline unsigned long task_faults(struct task_struct *p, int nid)
1242 if (!p->numa_faults)
1245 return p->numa_faults[task_faults_idx(NUMA_MEM, nid, 0)] +
1246 p->numa_faults[task_faults_idx(NUMA_MEM, nid, 1)];
1249 static inline unsigned long group_faults(struct task_struct *p, int nid)
1251 struct numa_group *ng = deref_task_numa_group(p);
1256 return ng->faults[task_faults_idx(NUMA_MEM, nid, 0)] +
1257 ng->faults[task_faults_idx(NUMA_MEM, nid, 1)];
1260 static inline unsigned long group_faults_cpu(struct numa_group *group, int nid)
1262 return group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 0)] +
1263 group->faults_cpu[task_faults_idx(NUMA_MEM, nid, 1)];
1266 static inline unsigned long group_faults_priv(struct numa_group *ng)
1268 unsigned long faults = 0;
1271 for_each_online_node(node) {
1272 faults += ng->faults[task_faults_idx(NUMA_MEM, node, 1)];
1278 static inline unsigned long group_faults_shared(struct numa_group *ng)
1280 unsigned long faults = 0;
1283 for_each_online_node(node) {
1284 faults += ng->faults[task_faults_idx(NUMA_MEM, node, 0)];
1291 * A node triggering more than 1/3 as many NUMA faults as the maximum is
1292 * considered part of a numa group's pseudo-interleaving set. Migrations
1293 * between these nodes are slowed down, to allow things to settle down.
1295 #define ACTIVE_NODE_FRACTION 3
1297 static bool numa_is_active_node(int nid, struct numa_group *ng)
1299 return group_faults_cpu(ng, nid) * ACTIVE_NODE_FRACTION > ng->max_faults_cpu;
1302 /* Handle placement on systems where not all nodes are directly connected. */
1303 static unsigned long score_nearby_nodes(struct task_struct *p, int nid,
1304 int maxdist, bool task)
1306 unsigned long score = 0;
1310 * All nodes are directly connected, and the same distance
1311 * from each other. No need for fancy placement algorithms.
1313 if (sched_numa_topology_type == NUMA_DIRECT)
1317 * This code is called for each node, introducing N^2 complexity,
1318 * which should be ok given the number of nodes rarely exceeds 8.
1320 for_each_online_node(node) {
1321 unsigned long faults;
1322 int dist = node_distance(nid, node);
1325 * The furthest away nodes in the system are not interesting
1326 * for placement; nid was already counted.
1328 if (dist == sched_max_numa_distance || node == nid)
1332 * On systems with a backplane NUMA topology, compare groups
1333 * of nodes, and move tasks towards the group with the most
1334 * memory accesses. When comparing two nodes at distance
1335 * "hoplimit", only nodes closer by than "hoplimit" are part
1336 * of each group. Skip other nodes.
1338 if (sched_numa_topology_type == NUMA_BACKPLANE &&
1342 /* Add up the faults from nearby nodes. */
1344 faults = task_faults(p, node);
1346 faults = group_faults(p, node);
1349 * On systems with a glueless mesh NUMA topology, there are
1350 * no fixed "groups of nodes". Instead, nodes that are not
1351 * directly connected bounce traffic through intermediate
1352 * nodes; a numa_group can occupy any set of nodes.
1353 * The further away a node is, the less the faults count.
1354 * This seems to result in good task placement.
1356 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
1357 faults *= (sched_max_numa_distance - dist);
1358 faults /= (sched_max_numa_distance - LOCAL_DISTANCE);
1368 * These return the fraction of accesses done by a particular task, or
1369 * task group, on a particular numa node. The group weight is given a
1370 * larger multiplier, in order to group tasks together that are almost
1371 * evenly spread out between numa nodes.
1373 static inline unsigned long task_weight(struct task_struct *p, int nid,
1376 unsigned long faults, total_faults;
1378 if (!p->numa_faults)
1381 total_faults = p->total_numa_faults;
1386 faults = task_faults(p, nid);
1387 faults += score_nearby_nodes(p, nid, dist, true);
1389 return 1000 * faults / total_faults;
1392 static inline unsigned long group_weight(struct task_struct *p, int nid,
1395 struct numa_group *ng = deref_task_numa_group(p);
1396 unsigned long faults, total_faults;
1401 total_faults = ng->total_faults;
1406 faults = group_faults(p, nid);
1407 faults += score_nearby_nodes(p, nid, dist, false);
1409 return 1000 * faults / total_faults;
1412 bool should_numa_migrate_memory(struct task_struct *p, struct page * page,
1413 int src_nid, int dst_cpu)
1415 struct numa_group *ng = deref_curr_numa_group(p);
1416 int dst_nid = cpu_to_node(dst_cpu);
1417 int last_cpupid, this_cpupid;
1419 this_cpupid = cpu_pid_to_cpupid(dst_cpu, current->pid);
1420 last_cpupid = page_cpupid_xchg_last(page, this_cpupid);
1423 * Allow first faults or private faults to migrate immediately early in
1424 * the lifetime of a task. The magic number 4 is based on waiting for
1425 * two full passes of the "multi-stage node selection" test that is
1428 if ((p->numa_preferred_nid == NUMA_NO_NODE || p->numa_scan_seq <= 4) &&
1429 (cpupid_pid_unset(last_cpupid) || cpupid_match_pid(p, last_cpupid)))
1433 * Multi-stage node selection is used in conjunction with a periodic
1434 * migration fault to build a temporal task<->page relation. By using
1435 * a two-stage filter we remove short/unlikely relations.
1437 * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate
1438 * a task's usage of a particular page (n_p) per total usage of this
1439 * page (n_t) (in a given time-span) to a probability.
1441 * Our periodic faults will sample this probability and getting the
1442 * same result twice in a row, given these samples are fully
1443 * independent, is then given by P(n)^2, provided our sample period
1444 * is sufficiently short compared to the usage pattern.
1446 * This quadric squishes small probabilities, making it less likely we
1447 * act on an unlikely task<->page relation.
1449 if (!cpupid_pid_unset(last_cpupid) &&
1450 cpupid_to_nid(last_cpupid) != dst_nid)
1453 /* Always allow migrate on private faults */
1454 if (cpupid_match_pid(p, last_cpupid))
1457 /* A shared fault, but p->numa_group has not been set up yet. */
1462 * Destination node is much more heavily used than the source
1463 * node? Allow migration.
1465 if (group_faults_cpu(ng, dst_nid) > group_faults_cpu(ng, src_nid) *
1466 ACTIVE_NODE_FRACTION)
1470 * Distribute memory according to CPU & memory use on each node,
1471 * with 3/4 hysteresis to avoid unnecessary memory migrations:
1473 * faults_cpu(dst) 3 faults_cpu(src)
1474 * --------------- * - > ---------------
1475 * faults_mem(dst) 4 faults_mem(src)
1477 return group_faults_cpu(ng, dst_nid) * group_faults(p, src_nid) * 3 >
1478 group_faults_cpu(ng, src_nid) * group_faults(p, dst_nid) * 4;
1482 * 'numa_type' describes the node at the moment of load balancing.
1485 /* The node has spare capacity that can be used to run more tasks. */
1488 * The node is fully used and the tasks don't compete for more CPU
1489 * cycles. Nevertheless, some tasks might wait before running.
1493 * The node is overloaded and can't provide expected CPU cycles to all
1499 /* Cached statistics for all CPUs within a node */
1502 unsigned long runnable;
1504 /* Total compute capacity of CPUs on a node */
1505 unsigned long compute_capacity;
1506 unsigned int nr_running;
1507 unsigned int weight;
1508 enum numa_type node_type;
1512 static inline bool is_core_idle(int cpu)
1514 #ifdef CONFIG_SCHED_SMT
1517 for_each_cpu(sibling, cpu_smt_mask(cpu)) {
1521 if (!idle_cpu(sibling))
1529 struct task_numa_env {
1530 struct task_struct *p;
1532 int src_cpu, src_nid;
1533 int dst_cpu, dst_nid;
1535 struct numa_stats src_stats, dst_stats;
1540 struct task_struct *best_task;
1545 static unsigned long cpu_load(struct rq *rq);
1546 static unsigned long cpu_runnable(struct rq *rq);
1547 static unsigned long cpu_util(int cpu);
1548 static inline long adjust_numa_imbalance(int imbalance,
1549 int dst_running, int dst_weight);
1552 numa_type numa_classify(unsigned int imbalance_pct,
1553 struct numa_stats *ns)
1555 if ((ns->nr_running > ns->weight) &&
1556 (((ns->compute_capacity * 100) < (ns->util * imbalance_pct)) ||
1557 ((ns->compute_capacity * imbalance_pct) < (ns->runnable * 100))))
1558 return node_overloaded;
1560 if ((ns->nr_running < ns->weight) ||
1561 (((ns->compute_capacity * 100) > (ns->util * imbalance_pct)) &&
1562 ((ns->compute_capacity * imbalance_pct) > (ns->runnable * 100))))
1563 return node_has_spare;
1565 return node_fully_busy;
1568 #ifdef CONFIG_SCHED_SMT
1569 /* Forward declarations of select_idle_sibling helpers */
1570 static inline bool test_idle_cores(int cpu, bool def);
1571 static inline int numa_idle_core(int idle_core, int cpu)
1573 if (!static_branch_likely(&sched_smt_present) ||
1574 idle_core >= 0 || !test_idle_cores(cpu, false))
1578 * Prefer cores instead of packing HT siblings
1579 * and triggering future load balancing.
1581 if (is_core_idle(cpu))
1587 static inline int numa_idle_core(int idle_core, int cpu)
1594 * Gather all necessary information to make NUMA balancing placement
1595 * decisions that are compatible with standard load balancer. This
1596 * borrows code and logic from update_sg_lb_stats but sharing a
1597 * common implementation is impractical.
1599 static void update_numa_stats(struct task_numa_env *env,
1600 struct numa_stats *ns, int nid,
1603 int cpu, idle_core = -1;
1605 memset(ns, 0, sizeof(*ns));
1609 for_each_cpu(cpu, cpumask_of_node(nid)) {
1610 struct rq *rq = cpu_rq(cpu);
1612 ns->load += cpu_load(rq);
1613 ns->runnable += cpu_runnable(rq);
1614 ns->util += cpu_util(cpu);
1615 ns->nr_running += rq->cfs.h_nr_running;
1616 ns->compute_capacity += capacity_of(cpu);
1618 if (find_idle && !rq->nr_running && idle_cpu(cpu)) {
1619 if (READ_ONCE(rq->numa_migrate_on) ||
1620 !cpumask_test_cpu(cpu, env->p->cpus_ptr))
1623 if (ns->idle_cpu == -1)
1626 idle_core = numa_idle_core(idle_core, cpu);
1631 ns->weight = cpumask_weight(cpumask_of_node(nid));
1633 ns->node_type = numa_classify(env->imbalance_pct, ns);
1636 ns->idle_cpu = idle_core;
1639 static void task_numa_assign(struct task_numa_env *env,
1640 struct task_struct *p, long imp)
1642 struct rq *rq = cpu_rq(env->dst_cpu);
1644 /* Check if run-queue part of active NUMA balance. */
1645 if (env->best_cpu != env->dst_cpu && xchg(&rq->numa_migrate_on, 1)) {
1647 int start = env->dst_cpu;
1649 /* Find alternative idle CPU. */
1650 for_each_cpu_wrap(cpu, cpumask_of_node(env->dst_nid), start) {
1651 if (cpu == env->best_cpu || !idle_cpu(cpu) ||
1652 !cpumask_test_cpu(cpu, env->p->cpus_ptr)) {
1657 rq = cpu_rq(env->dst_cpu);
1658 if (!xchg(&rq->numa_migrate_on, 1))
1662 /* Failed to find an alternative idle CPU */
1668 * Clear previous best_cpu/rq numa-migrate flag, since task now
1669 * found a better CPU to move/swap.
1671 if (env->best_cpu != -1 && env->best_cpu != env->dst_cpu) {
1672 rq = cpu_rq(env->best_cpu);
1673 WRITE_ONCE(rq->numa_migrate_on, 0);
1677 put_task_struct(env->best_task);
1682 env->best_imp = imp;
1683 env->best_cpu = env->dst_cpu;
1686 static bool load_too_imbalanced(long src_load, long dst_load,
1687 struct task_numa_env *env)
1690 long orig_src_load, orig_dst_load;
1691 long src_capacity, dst_capacity;
1694 * The load is corrected for the CPU capacity available on each node.
1697 * ------------ vs ---------
1698 * src_capacity dst_capacity
1700 src_capacity = env->src_stats.compute_capacity;
1701 dst_capacity = env->dst_stats.compute_capacity;
1703 imb = abs(dst_load * src_capacity - src_load * dst_capacity);
1705 orig_src_load = env->src_stats.load;
1706 orig_dst_load = env->dst_stats.load;
1708 old_imb = abs(orig_dst_load * src_capacity - orig_src_load * dst_capacity);
1710 /* Would this change make things worse? */
1711 return (imb > old_imb);
1715 * Maximum NUMA importance can be 1998 (2*999);
1716 * SMALLIMP @ 30 would be close to 1998/64.
1717 * Used to deter task migration.
1722 * This checks if the overall compute and NUMA accesses of the system would
1723 * be improved if the source tasks was migrated to the target dst_cpu taking
1724 * into account that it might be best if task running on the dst_cpu should
1725 * be exchanged with the source task
1727 static bool task_numa_compare(struct task_numa_env *env,
1728 long taskimp, long groupimp, bool maymove)
1730 struct numa_group *cur_ng, *p_ng = deref_curr_numa_group(env->p);
1731 struct rq *dst_rq = cpu_rq(env->dst_cpu);
1732 long imp = p_ng ? groupimp : taskimp;
1733 struct task_struct *cur;
1734 long src_load, dst_load;
1735 int dist = env->dist;
1738 bool stopsearch = false;
1740 if (READ_ONCE(dst_rq->numa_migrate_on))
1744 cur = rcu_dereference(dst_rq->curr);
1745 if (cur && ((cur->flags & PF_EXITING) || is_idle_task(cur)))
1749 * Because we have preemption enabled we can get migrated around and
1750 * end try selecting ourselves (current == env->p) as a swap candidate.
1752 if (cur == env->p) {
1758 if (maymove && moveimp >= env->best_imp)
1764 /* Skip this swap candidate if cannot move to the source cpu. */
1765 if (!cpumask_test_cpu(env->src_cpu, cur->cpus_ptr))
1769 * Skip this swap candidate if it is not moving to its preferred
1770 * node and the best task is.
1772 if (env->best_task &&
1773 env->best_task->numa_preferred_nid == env->src_nid &&
1774 cur->numa_preferred_nid != env->src_nid) {
1779 * "imp" is the fault differential for the source task between the
1780 * source and destination node. Calculate the total differential for
1781 * the source task and potential destination task. The more negative
1782 * the value is, the more remote accesses that would be expected to
1783 * be incurred if the tasks were swapped.
1785 * If dst and source tasks are in the same NUMA group, or not
1786 * in any group then look only at task weights.
1788 cur_ng = rcu_dereference(cur->numa_group);
1789 if (cur_ng == p_ng) {
1790 imp = taskimp + task_weight(cur, env->src_nid, dist) -
1791 task_weight(cur, env->dst_nid, dist);
1793 * Add some hysteresis to prevent swapping the
1794 * tasks within a group over tiny differences.
1800 * Compare the group weights. If a task is all by itself
1801 * (not part of a group), use the task weight instead.
1804 imp += group_weight(cur, env->src_nid, dist) -
1805 group_weight(cur, env->dst_nid, dist);
1807 imp += task_weight(cur, env->src_nid, dist) -
1808 task_weight(cur, env->dst_nid, dist);
1811 /* Discourage picking a task already on its preferred node */
1812 if (cur->numa_preferred_nid == env->dst_nid)
1816 * Encourage picking a task that moves to its preferred node.
1817 * This potentially makes imp larger than it's maximum of
1818 * 1998 (see SMALLIMP and task_weight for why) but in this
1819 * case, it does not matter.
1821 if (cur->numa_preferred_nid == env->src_nid)
1824 if (maymove && moveimp > imp && moveimp > env->best_imp) {
1831 * Prefer swapping with a task moving to its preferred node over a
1834 if (env->best_task && cur->numa_preferred_nid == env->src_nid &&
1835 env->best_task->numa_preferred_nid != env->src_nid) {
1840 * If the NUMA importance is less than SMALLIMP,
1841 * task migration might only result in ping pong
1842 * of tasks and also hurt performance due to cache
1845 if (imp < SMALLIMP || imp <= env->best_imp + SMALLIMP / 2)
1849 * In the overloaded case, try and keep the load balanced.
1851 load = task_h_load(env->p) - task_h_load(cur);
1855 dst_load = env->dst_stats.load + load;
1856 src_load = env->src_stats.load - load;
1858 if (load_too_imbalanced(src_load, dst_load, env))
1862 /* Evaluate an idle CPU for a task numa move. */
1864 int cpu = env->dst_stats.idle_cpu;
1866 /* Nothing cached so current CPU went idle since the search. */
1871 * If the CPU is no longer truly idle and the previous best CPU
1872 * is, keep using it.
1874 if (!idle_cpu(cpu) && env->best_cpu >= 0 &&
1875 idle_cpu(env->best_cpu)) {
1876 cpu = env->best_cpu;
1882 task_numa_assign(env, cur, imp);
1885 * If a move to idle is allowed because there is capacity or load
1886 * balance improves then stop the search. While a better swap
1887 * candidate may exist, a search is not free.
1889 if (maymove && !cur && env->best_cpu >= 0 && idle_cpu(env->best_cpu))
1893 * If a swap candidate must be identified and the current best task
1894 * moves its preferred node then stop the search.
1896 if (!maymove && env->best_task &&
1897 env->best_task->numa_preferred_nid == env->src_nid) {
1906 static void task_numa_find_cpu(struct task_numa_env *env,
1907 long taskimp, long groupimp)
1909 bool maymove = false;
1913 * If dst node has spare capacity, then check if there is an
1914 * imbalance that would be overruled by the load balancer.
1916 if (env->dst_stats.node_type == node_has_spare) {
1917 unsigned int imbalance;
1918 int src_running, dst_running;
1921 * Would movement cause an imbalance? Note that if src has
1922 * more running tasks that the imbalance is ignored as the
1923 * move improves the imbalance from the perspective of the
1924 * CPU load balancer.
1926 src_running = env->src_stats.nr_running - 1;
1927 dst_running = env->dst_stats.nr_running + 1;
1928 imbalance = max(0, dst_running - src_running);
1929 imbalance = adjust_numa_imbalance(imbalance, dst_running,
1930 env->dst_stats.weight);
1932 /* Use idle CPU if there is no imbalance */
1935 if (env->dst_stats.idle_cpu >= 0) {
1936 env->dst_cpu = env->dst_stats.idle_cpu;
1937 task_numa_assign(env, NULL, 0);
1942 long src_load, dst_load, load;
1944 * If the improvement from just moving env->p direction is better
1945 * than swapping tasks around, check if a move is possible.
1947 load = task_h_load(env->p);
1948 dst_load = env->dst_stats.load + load;
1949 src_load = env->src_stats.load - load;
1950 maymove = !load_too_imbalanced(src_load, dst_load, env);
1953 for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) {
1954 /* Skip this CPU if the source task cannot migrate */
1955 if (!cpumask_test_cpu(cpu, env->p->cpus_ptr))
1959 if (task_numa_compare(env, taskimp, groupimp, maymove))
1964 static int task_numa_migrate(struct task_struct *p)
1966 struct task_numa_env env = {
1969 .src_cpu = task_cpu(p),
1970 .src_nid = task_node(p),
1972 .imbalance_pct = 112,
1978 unsigned long taskweight, groupweight;
1979 struct sched_domain *sd;
1980 long taskimp, groupimp;
1981 struct numa_group *ng;
1986 * Pick the lowest SD_NUMA domain, as that would have the smallest
1987 * imbalance and would be the first to start moving tasks about.
1989 * And we want to avoid any moving of tasks about, as that would create
1990 * random movement of tasks -- counter the numa conditions we're trying
1994 sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu));
1996 env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2;
2000 * Cpusets can break the scheduler domain tree into smaller
2001 * balance domains, some of which do not cross NUMA boundaries.
2002 * Tasks that are "trapped" in such domains cannot be migrated
2003 * elsewhere, so there is no point in (re)trying.
2005 if (unlikely(!sd)) {
2006 sched_setnuma(p, task_node(p));
2010 env.dst_nid = p->numa_preferred_nid;
2011 dist = env.dist = node_distance(env.src_nid, env.dst_nid);
2012 taskweight = task_weight(p, env.src_nid, dist);
2013 groupweight = group_weight(p, env.src_nid, dist);
2014 update_numa_stats(&env, &env.src_stats, env.src_nid, false);
2015 taskimp = task_weight(p, env.dst_nid, dist) - taskweight;
2016 groupimp = group_weight(p, env.dst_nid, dist) - groupweight;
2017 update_numa_stats(&env, &env.dst_stats, env.dst_nid, true);
2019 /* Try to find a spot on the preferred nid. */
2020 task_numa_find_cpu(&env, taskimp, groupimp);
2023 * Look at other nodes in these cases:
2024 * - there is no space available on the preferred_nid
2025 * - the task is part of a numa_group that is interleaved across
2026 * multiple NUMA nodes; in order to better consolidate the group,
2027 * we need to check other locations.
2029 ng = deref_curr_numa_group(p);
2030 if (env.best_cpu == -1 || (ng && ng->active_nodes > 1)) {
2031 for_each_online_node(nid) {
2032 if (nid == env.src_nid || nid == p->numa_preferred_nid)
2035 dist = node_distance(env.src_nid, env.dst_nid);
2036 if (sched_numa_topology_type == NUMA_BACKPLANE &&
2038 taskweight = task_weight(p, env.src_nid, dist);
2039 groupweight = group_weight(p, env.src_nid, dist);
2042 /* Only consider nodes where both task and groups benefit */
2043 taskimp = task_weight(p, nid, dist) - taskweight;
2044 groupimp = group_weight(p, nid, dist) - groupweight;
2045 if (taskimp < 0 && groupimp < 0)
2050 update_numa_stats(&env, &env.dst_stats, env.dst_nid, true);
2051 task_numa_find_cpu(&env, taskimp, groupimp);
2056 * If the task is part of a workload that spans multiple NUMA nodes,
2057 * and is migrating into one of the workload's active nodes, remember
2058 * this node as the task's preferred numa node, so the workload can
2060 * A task that migrated to a second choice node will be better off
2061 * trying for a better one later. Do not set the preferred node here.
2064 if (env.best_cpu == -1)
2067 nid = cpu_to_node(env.best_cpu);
2069 if (nid != p->numa_preferred_nid)
2070 sched_setnuma(p, nid);
2073 /* No better CPU than the current one was found. */
2074 if (env.best_cpu == -1) {
2075 trace_sched_stick_numa(p, env.src_cpu, NULL, -1);
2079 best_rq = cpu_rq(env.best_cpu);
2080 if (env.best_task == NULL) {
2081 ret = migrate_task_to(p, env.best_cpu);
2082 WRITE_ONCE(best_rq->numa_migrate_on, 0);
2084 trace_sched_stick_numa(p, env.src_cpu, NULL, env.best_cpu);
2088 ret = migrate_swap(p, env.best_task, env.best_cpu, env.src_cpu);
2089 WRITE_ONCE(best_rq->numa_migrate_on, 0);
2092 trace_sched_stick_numa(p, env.src_cpu, env.best_task, env.best_cpu);
2093 put_task_struct(env.best_task);
2097 /* Attempt to migrate a task to a CPU on the preferred node. */
2098 static void numa_migrate_preferred(struct task_struct *p)
2100 unsigned long interval = HZ;
2102 /* This task has no NUMA fault statistics yet */
2103 if (unlikely(p->numa_preferred_nid == NUMA_NO_NODE || !p->numa_faults))
2106 /* Periodically retry migrating the task to the preferred node */
2107 interval = min(interval, msecs_to_jiffies(p->numa_scan_period) / 16);
2108 p->numa_migrate_retry = jiffies + interval;
2110 /* Success if task is already running on preferred CPU */
2111 if (task_node(p) == p->numa_preferred_nid)
2114 /* Otherwise, try migrate to a CPU on the preferred node */
2115 task_numa_migrate(p);
2119 * Find out how many nodes on the workload is actively running on. Do this by
2120 * tracking the nodes from which NUMA hinting faults are triggered. This can
2121 * be different from the set of nodes where the workload's memory is currently
2124 static void numa_group_count_active_nodes(struct numa_group *numa_group)
2126 unsigned long faults, max_faults = 0;
2127 int nid, active_nodes = 0;
2129 for_each_online_node(nid) {
2130 faults = group_faults_cpu(numa_group, nid);
2131 if (faults > max_faults)
2132 max_faults = faults;
2135 for_each_online_node(nid) {
2136 faults = group_faults_cpu(numa_group, nid);
2137 if (faults * ACTIVE_NODE_FRACTION > max_faults)
2141 numa_group->max_faults_cpu = max_faults;
2142 numa_group->active_nodes = active_nodes;
2146 * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
2147 * increments. The more local the fault statistics are, the higher the scan
2148 * period will be for the next scan window. If local/(local+remote) ratio is
2149 * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS)
2150 * the scan period will decrease. Aim for 70% local accesses.
2152 #define NUMA_PERIOD_SLOTS 10
2153 #define NUMA_PERIOD_THRESHOLD 7
2156 * Increase the scan period (slow down scanning) if the majority of
2157 * our memory is already on our local node, or if the majority of
2158 * the page accesses are shared with other processes.
2159 * Otherwise, decrease the scan period.
2161 static void update_task_scan_period(struct task_struct *p,
2162 unsigned long shared, unsigned long private)
2164 unsigned int period_slot;
2165 int lr_ratio, ps_ratio;
2168 unsigned long remote = p->numa_faults_locality[0];
2169 unsigned long local = p->numa_faults_locality[1];
2172 * If there were no record hinting faults then either the task is
2173 * completely idle or all activity is areas that are not of interest
2174 * to automatic numa balancing. Related to that, if there were failed
2175 * migration then it implies we are migrating too quickly or the local
2176 * node is overloaded. In either case, scan slower
2178 if (local + shared == 0 || p->numa_faults_locality[2]) {
2179 p->numa_scan_period = min(p->numa_scan_period_max,
2180 p->numa_scan_period << 1);
2182 p->mm->numa_next_scan = jiffies +
2183 msecs_to_jiffies(p->numa_scan_period);
2189 * Prepare to scale scan period relative to the current period.
2190 * == NUMA_PERIOD_THRESHOLD scan period stays the same
2191 * < NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
2192 * >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
2194 period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS);
2195 lr_ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote);
2196 ps_ratio = (private * NUMA_PERIOD_SLOTS) / (private + shared);
2198 if (ps_ratio >= NUMA_PERIOD_THRESHOLD) {
2200 * Most memory accesses are local. There is no need to
2201 * do fast NUMA scanning, since memory is already local.
2203 int slot = ps_ratio - NUMA_PERIOD_THRESHOLD;
2206 diff = slot * period_slot;
2207 } else if (lr_ratio >= NUMA_PERIOD_THRESHOLD) {
2209 * Most memory accesses are shared with other tasks.
2210 * There is no point in continuing fast NUMA scanning,
2211 * since other tasks may just move the memory elsewhere.
2213 int slot = lr_ratio - NUMA_PERIOD_THRESHOLD;
2216 diff = slot * period_slot;
2219 * Private memory faults exceed (SLOTS-THRESHOLD)/SLOTS,
2220 * yet they are not on the local NUMA node. Speed up
2221 * NUMA scanning to get the memory moved over.
2223 int ratio = max(lr_ratio, ps_ratio);
2224 diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot;
2227 p->numa_scan_period = clamp(p->numa_scan_period + diff,
2228 task_scan_min(p), task_scan_max(p));
2229 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
2233 * Get the fraction of time the task has been running since the last
2234 * NUMA placement cycle. The scheduler keeps similar statistics, but
2235 * decays those on a 32ms period, which is orders of magnitude off
2236 * from the dozens-of-seconds NUMA balancing period. Use the scheduler
2237 * stats only if the task is so new there are no NUMA statistics yet.
2239 static u64 numa_get_avg_runtime(struct task_struct *p, u64 *period)
2241 u64 runtime, delta, now;
2242 /* Use the start of this time slice to avoid calculations. */
2243 now = p->se.exec_start;
2244 runtime = p->se.sum_exec_runtime;
2246 if (p->last_task_numa_placement) {
2247 delta = runtime - p->last_sum_exec_runtime;
2248 *period = now - p->last_task_numa_placement;
2250 /* Avoid time going backwards, prevent potential divide error: */
2251 if (unlikely((s64)*period < 0))
2254 delta = p->se.avg.load_sum;
2255 *period = LOAD_AVG_MAX;
2258 p->last_sum_exec_runtime = runtime;
2259 p->last_task_numa_placement = now;
2265 * Determine the preferred nid for a task in a numa_group. This needs to
2266 * be done in a way that produces consistent results with group_weight,
2267 * otherwise workloads might not converge.
2269 static int preferred_group_nid(struct task_struct *p, int nid)
2274 /* Direct connections between all NUMA nodes. */
2275 if (sched_numa_topology_type == NUMA_DIRECT)
2279 * On a system with glueless mesh NUMA topology, group_weight
2280 * scores nodes according to the number of NUMA hinting faults on
2281 * both the node itself, and on nearby nodes.
2283 if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
2284 unsigned long score, max_score = 0;
2285 int node, max_node = nid;
2287 dist = sched_max_numa_distance;
2289 for_each_online_node(node) {
2290 score = group_weight(p, node, dist);
2291 if (score > max_score) {
2300 * Finding the preferred nid in a system with NUMA backplane
2301 * interconnect topology is more involved. The goal is to locate
2302 * tasks from numa_groups near each other in the system, and
2303 * untangle workloads from different sides of the system. This requires
2304 * searching down the hierarchy of node groups, recursively searching
2305 * inside the highest scoring group of nodes. The nodemask tricks
2306 * keep the complexity of the search down.
2308 nodes = node_online_map;
2309 for (dist = sched_max_numa_distance; dist > LOCAL_DISTANCE; dist--) {
2310 unsigned long max_faults = 0;
2311 nodemask_t max_group = NODE_MASK_NONE;
2314 /* Are there nodes at this distance from each other? */
2315 if (!find_numa_distance(dist))
2318 for_each_node_mask(a, nodes) {
2319 unsigned long faults = 0;
2320 nodemask_t this_group;
2321 nodes_clear(this_group);
2323 /* Sum group's NUMA faults; includes a==b case. */
2324 for_each_node_mask(b, nodes) {
2325 if (node_distance(a, b) < dist) {
2326 faults += group_faults(p, b);
2327 node_set(b, this_group);
2328 node_clear(b, nodes);
2332 /* Remember the top group. */
2333 if (faults > max_faults) {
2334 max_faults = faults;
2335 max_group = this_group;
2337 * subtle: at the smallest distance there is
2338 * just one node left in each "group", the
2339 * winner is the preferred nid.
2344 /* Next round, evaluate the nodes within max_group. */
2352 static void task_numa_placement(struct task_struct *p)
2354 int seq, nid, max_nid = NUMA_NO_NODE;
2355 unsigned long max_faults = 0;
2356 unsigned long fault_types[2] = { 0, 0 };
2357 unsigned long total_faults;
2358 u64 runtime, period;
2359 spinlock_t *group_lock = NULL;
2360 struct numa_group *ng;
2363 * The p->mm->numa_scan_seq field gets updated without
2364 * exclusive access. Use READ_ONCE() here to ensure
2365 * that the field is read in a single access:
2367 seq = READ_ONCE(p->mm->numa_scan_seq);
2368 if (p->numa_scan_seq == seq)
2370 p->numa_scan_seq = seq;
2371 p->numa_scan_period_max = task_scan_max(p);
2373 total_faults = p->numa_faults_locality[0] +
2374 p->numa_faults_locality[1];
2375 runtime = numa_get_avg_runtime(p, &period);
2377 /* If the task is part of a group prevent parallel updates to group stats */
2378 ng = deref_curr_numa_group(p);
2380 group_lock = &ng->lock;
2381 spin_lock_irq(group_lock);
2384 /* Find the node with the highest number of faults */
2385 for_each_online_node(nid) {
2386 /* Keep track of the offsets in numa_faults array */
2387 int mem_idx, membuf_idx, cpu_idx, cpubuf_idx;
2388 unsigned long faults = 0, group_faults = 0;
2391 for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) {
2392 long diff, f_diff, f_weight;
2394 mem_idx = task_faults_idx(NUMA_MEM, nid, priv);
2395 membuf_idx = task_faults_idx(NUMA_MEMBUF, nid, priv);
2396 cpu_idx = task_faults_idx(NUMA_CPU, nid, priv);
2397 cpubuf_idx = task_faults_idx(NUMA_CPUBUF, nid, priv);
2399 /* Decay existing window, copy faults since last scan */
2400 diff = p->numa_faults[membuf_idx] - p->numa_faults[mem_idx] / 2;
2401 fault_types[priv] += p->numa_faults[membuf_idx];
2402 p->numa_faults[membuf_idx] = 0;
2405 * Normalize the faults_from, so all tasks in a group
2406 * count according to CPU use, instead of by the raw
2407 * number of faults. Tasks with little runtime have
2408 * little over-all impact on throughput, and thus their
2409 * faults are less important.
2411 f_weight = div64_u64(runtime << 16, period + 1);
2412 f_weight = (f_weight * p->numa_faults[cpubuf_idx]) /
2414 f_diff = f_weight - p->numa_faults[cpu_idx] / 2;
2415 p->numa_faults[cpubuf_idx] = 0;
2417 p->numa_faults[mem_idx] += diff;
2418 p->numa_faults[cpu_idx] += f_diff;
2419 faults += p->numa_faults[mem_idx];
2420 p->total_numa_faults += diff;
2423 * safe because we can only change our own group
2425 * mem_idx represents the offset for a given
2426 * nid and priv in a specific region because it
2427 * is at the beginning of the numa_faults array.
2429 ng->faults[mem_idx] += diff;
2430 ng->faults_cpu[mem_idx] += f_diff;
2431 ng->total_faults += diff;
2432 group_faults += ng->faults[mem_idx];
2437 if (faults > max_faults) {
2438 max_faults = faults;
2441 } else if (group_faults > max_faults) {
2442 max_faults = group_faults;
2448 numa_group_count_active_nodes(ng);
2449 spin_unlock_irq(group_lock);
2450 max_nid = preferred_group_nid(p, max_nid);
2454 /* Set the new preferred node */
2455 if (max_nid != p->numa_preferred_nid)
2456 sched_setnuma(p, max_nid);
2459 update_task_scan_period(p, fault_types[0], fault_types[1]);
2462 static inline int get_numa_group(struct numa_group *grp)
2464 return refcount_inc_not_zero(&grp->refcount);
2467 static inline void put_numa_group(struct numa_group *grp)
2469 if (refcount_dec_and_test(&grp->refcount))
2470 kfree_rcu(grp, rcu);
2473 static void task_numa_group(struct task_struct *p, int cpupid, int flags,
2476 struct numa_group *grp, *my_grp;
2477 struct task_struct *tsk;
2479 int cpu = cpupid_to_cpu(cpupid);
2482 if (unlikely(!deref_curr_numa_group(p))) {
2483 unsigned int size = sizeof(struct numa_group) +
2484 4*nr_node_ids*sizeof(unsigned long);
2486 grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
2490 refcount_set(&grp->refcount, 1);
2491 grp->active_nodes = 1;
2492 grp->max_faults_cpu = 0;
2493 spin_lock_init(&grp->lock);
2495 /* Second half of the array tracks nids where faults happen */
2496 grp->faults_cpu = grp->faults + NR_NUMA_HINT_FAULT_TYPES *
2499 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2500 grp->faults[i] = p->numa_faults[i];
2502 grp->total_faults = p->total_numa_faults;
2505 rcu_assign_pointer(p->numa_group, grp);
2509 tsk = READ_ONCE(cpu_rq(cpu)->curr);
2511 if (!cpupid_match_pid(tsk, cpupid))
2514 grp = rcu_dereference(tsk->numa_group);
2518 my_grp = deref_curr_numa_group(p);
2523 * Only join the other group if its bigger; if we're the bigger group,
2524 * the other task will join us.
2526 if (my_grp->nr_tasks > grp->nr_tasks)
2530 * Tie-break on the grp address.
2532 if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp)
2535 /* Always join threads in the same process. */
2536 if (tsk->mm == current->mm)
2539 /* Simple filter to avoid false positives due to PID collisions */
2540 if (flags & TNF_SHARED)
2543 /* Update priv based on whether false sharing was detected */
2546 if (join && !get_numa_group(grp))
2554 BUG_ON(irqs_disabled());
2555 double_lock_irq(&my_grp->lock, &grp->lock);
2557 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) {
2558 my_grp->faults[i] -= p->numa_faults[i];
2559 grp->faults[i] += p->numa_faults[i];
2561 my_grp->total_faults -= p->total_numa_faults;
2562 grp->total_faults += p->total_numa_faults;
2567 spin_unlock(&my_grp->lock);
2568 spin_unlock_irq(&grp->lock);
2570 rcu_assign_pointer(p->numa_group, grp);
2572 put_numa_group(my_grp);
2581 * Get rid of NUMA statistics associated with a task (either current or dead).
2582 * If @final is set, the task is dead and has reached refcount zero, so we can
2583 * safely free all relevant data structures. Otherwise, there might be
2584 * concurrent reads from places like load balancing and procfs, and we should
2585 * reset the data back to default state without freeing ->numa_faults.
2587 void task_numa_free(struct task_struct *p, bool final)
2589 /* safe: p either is current or is being freed by current */
2590 struct numa_group *grp = rcu_dereference_raw(p->numa_group);
2591 unsigned long *numa_faults = p->numa_faults;
2592 unsigned long flags;
2599 spin_lock_irqsave(&grp->lock, flags);
2600 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2601 grp->faults[i] -= p->numa_faults[i];
2602 grp->total_faults -= p->total_numa_faults;
2605 spin_unlock_irqrestore(&grp->lock, flags);
2606 RCU_INIT_POINTER(p->numa_group, NULL);
2607 put_numa_group(grp);
2611 p->numa_faults = NULL;
2614 p->total_numa_faults = 0;
2615 for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2621 * Got a PROT_NONE fault for a page on @node.
2623 void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags)
2625 struct task_struct *p = current;
2626 bool migrated = flags & TNF_MIGRATED;
2627 int cpu_node = task_node(current);
2628 int local = !!(flags & TNF_FAULT_LOCAL);
2629 struct numa_group *ng;
2632 if (!static_branch_likely(&sched_numa_balancing))
2635 /* for example, ksmd faulting in a user's mm */
2639 /* Allocate buffer to track faults on a per-node basis */
2640 if (unlikely(!p->numa_faults)) {
2641 int size = sizeof(*p->numa_faults) *
2642 NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids;
2644 p->numa_faults = kzalloc(size, GFP_KERNEL|__GFP_NOWARN);
2645 if (!p->numa_faults)
2648 p->total_numa_faults = 0;
2649 memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
2653 * First accesses are treated as private, otherwise consider accesses
2654 * to be private if the accessing pid has not changed
2656 if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) {
2659 priv = cpupid_match_pid(p, last_cpupid);
2660 if (!priv && !(flags & TNF_NO_GROUP))
2661 task_numa_group(p, last_cpupid, flags, &priv);
2665 * If a workload spans multiple NUMA nodes, a shared fault that
2666 * occurs wholly within the set of nodes that the workload is
2667 * actively using should be counted as local. This allows the
2668 * scan rate to slow down when a workload has settled down.
2670 ng = deref_curr_numa_group(p);
2671 if (!priv && !local && ng && ng->active_nodes > 1 &&
2672 numa_is_active_node(cpu_node, ng) &&
2673 numa_is_active_node(mem_node, ng))
2677 * Retry to migrate task to preferred node periodically, in case it
2678 * previously failed, or the scheduler moved us.
2680 if (time_after(jiffies, p->numa_migrate_retry)) {
2681 task_numa_placement(p);
2682 numa_migrate_preferred(p);
2686 p->numa_pages_migrated += pages;
2687 if (flags & TNF_MIGRATE_FAIL)
2688 p->numa_faults_locality[2] += pages;
2690 p->numa_faults[task_faults_idx(NUMA_MEMBUF, mem_node, priv)] += pages;
2691 p->numa_faults[task_faults_idx(NUMA_CPUBUF, cpu_node, priv)] += pages;
2692 p->numa_faults_locality[local] += pages;
2695 static void reset_ptenuma_scan(struct task_struct *p)
2698 * We only did a read acquisition of the mmap sem, so
2699 * p->mm->numa_scan_seq is written to without exclusive access
2700 * and the update is not guaranteed to be atomic. That's not
2701 * much of an issue though, since this is just used for
2702 * statistical sampling. Use READ_ONCE/WRITE_ONCE, which are not
2703 * expensive, to avoid any form of compiler optimizations:
2705 WRITE_ONCE(p->mm->numa_scan_seq, READ_ONCE(p->mm->numa_scan_seq) + 1);
2706 p->mm->numa_scan_offset = 0;
2710 * The expensive part of numa migration is done from task_work context.
2711 * Triggered from task_tick_numa().
2713 static void task_numa_work(struct callback_head *work)
2715 unsigned long migrate, next_scan, now = jiffies;
2716 struct task_struct *p = current;
2717 struct mm_struct *mm = p->mm;
2718 u64 runtime = p->se.sum_exec_runtime;
2719 struct vm_area_struct *vma;
2720 unsigned long start, end;
2721 unsigned long nr_pte_updates = 0;
2722 long pages, virtpages;
2724 SCHED_WARN_ON(p != container_of(work, struct task_struct, numa_work));
2728 * Who cares about NUMA placement when they're dying.
2730 * NOTE: make sure not to dereference p->mm before this check,
2731 * exit_task_work() happens _after_ exit_mm() so we could be called
2732 * without p->mm even though we still had it when we enqueued this
2735 if (p->flags & PF_EXITING)
2738 if (!mm->numa_next_scan) {
2739 mm->numa_next_scan = now +
2740 msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2744 * Enforce maximal scan/migration frequency..
2746 migrate = mm->numa_next_scan;
2747 if (time_before(now, migrate))
2750 if (p->numa_scan_period == 0) {
2751 p->numa_scan_period_max = task_scan_max(p);
2752 p->numa_scan_period = task_scan_start(p);
2755 next_scan = now + msecs_to_jiffies(p->numa_scan_period);
2756 if (cmpxchg(&mm->numa_next_scan, migrate, next_scan) != migrate)
2760 * Delay this task enough that another task of this mm will likely win
2761 * the next time around.
2763 p->node_stamp += 2 * TICK_NSEC;
2765 start = mm->numa_scan_offset;
2766 pages = sysctl_numa_balancing_scan_size;
2767 pages <<= 20 - PAGE_SHIFT; /* MB in pages */
2768 virtpages = pages * 8; /* Scan up to this much virtual space */
2773 if (!mmap_read_trylock(mm))
2775 vma = find_vma(mm, start);
2777 reset_ptenuma_scan(p);
2781 for (; vma; vma = vma->vm_next) {
2782 if (!vma_migratable(vma) || !vma_policy_mof(vma) ||
2783 is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_MIXEDMAP)) {
2788 * Shared library pages mapped by multiple processes are not
2789 * migrated as it is expected they are cache replicated. Avoid
2790 * hinting faults in read-only file-backed mappings or the vdso
2791 * as migrating the pages will be of marginal benefit.
2794 (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ)))
2798 * Skip inaccessible VMAs to avoid any confusion between
2799 * PROT_NONE and NUMA hinting ptes
2801 if (!vma_is_accessible(vma))
2805 start = max(start, vma->vm_start);
2806 end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
2807 end = min(end, vma->vm_end);
2808 nr_pte_updates = change_prot_numa(vma, start, end);
2811 * Try to scan sysctl_numa_balancing_size worth of
2812 * hpages that have at least one present PTE that
2813 * is not already pte-numa. If the VMA contains
2814 * areas that are unused or already full of prot_numa
2815 * PTEs, scan up to virtpages, to skip through those
2819 pages -= (end - start) >> PAGE_SHIFT;
2820 virtpages -= (end - start) >> PAGE_SHIFT;
2823 if (pages <= 0 || virtpages <= 0)
2827 } while (end != vma->vm_end);
2832 * It is possible to reach the end of the VMA list but the last few
2833 * VMAs are not guaranteed to the vma_migratable. If they are not, we
2834 * would find the !migratable VMA on the next scan but not reset the
2835 * scanner to the start so check it now.
2838 mm->numa_scan_offset = start;
2840 reset_ptenuma_scan(p);
2841 mmap_read_unlock(mm);
2844 * Make sure tasks use at least 32x as much time to run other code
2845 * than they used here, to limit NUMA PTE scanning overhead to 3% max.
2846 * Usually update_task_scan_period slows down scanning enough; on an
2847 * overloaded system we need to limit overhead on a per task basis.
2849 if (unlikely(p->se.sum_exec_runtime != runtime)) {
2850 u64 diff = p->se.sum_exec_runtime - runtime;
2851 p->node_stamp += 32 * diff;
2855 void init_numa_balancing(unsigned long clone_flags, struct task_struct *p)
2858 struct mm_struct *mm = p->mm;
2861 mm_users = atomic_read(&mm->mm_users);
2862 if (mm_users == 1) {
2863 mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2864 mm->numa_scan_seq = 0;
2868 p->numa_scan_seq = mm ? mm->numa_scan_seq : 0;
2869 p->numa_scan_period = sysctl_numa_balancing_scan_delay;
2870 /* Protect against double add, see task_tick_numa and task_numa_work */
2871 p->numa_work.next = &p->numa_work;
2872 p->numa_faults = NULL;
2873 RCU_INIT_POINTER(p->numa_group, NULL);
2874 p->last_task_numa_placement = 0;
2875 p->last_sum_exec_runtime = 0;
2877 init_task_work(&p->numa_work, task_numa_work);
2879 /* New address space, reset the preferred nid */
2880 if (!(clone_flags & CLONE_VM)) {
2881 p->numa_preferred_nid = NUMA_NO_NODE;
2886 * New thread, keep existing numa_preferred_nid which should be copied
2887 * already by arch_dup_task_struct but stagger when scans start.
2892 delay = min_t(unsigned int, task_scan_max(current),
2893 current->numa_scan_period * mm_users * NSEC_PER_MSEC);
2894 delay += 2 * TICK_NSEC;
2895 p->node_stamp = delay;
2900 * Drive the periodic memory faults..
2902 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
2904 struct callback_head *work = &curr->numa_work;
2908 * We don't care about NUMA placement if we don't have memory.
2910 if ((curr->flags & (PF_EXITING | PF_KTHREAD)) || work->next != work)
2914 * Using runtime rather than walltime has the dual advantage that
2915 * we (mostly) drive the selection from busy threads and that the
2916 * task needs to have done some actual work before we bother with
2919 now = curr->se.sum_exec_runtime;
2920 period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
2922 if (now > curr->node_stamp + period) {
2923 if (!curr->node_stamp)
2924 curr->numa_scan_period = task_scan_start(curr);
2925 curr->node_stamp += period;
2927 if (!time_before(jiffies, curr->mm->numa_next_scan))
2928 task_work_add(curr, work, TWA_RESUME);
2932 static void update_scan_period(struct task_struct *p, int new_cpu)
2934 int src_nid = cpu_to_node(task_cpu(p));
2935 int dst_nid = cpu_to_node(new_cpu);
2937 if (!static_branch_likely(&sched_numa_balancing))
2940 if (!p->mm || !p->numa_faults || (p->flags & PF_EXITING))
2943 if (src_nid == dst_nid)
2947 * Allow resets if faults have been trapped before one scan
2948 * has completed. This is most likely due to a new task that
2949 * is pulled cross-node due to wakeups or load balancing.
2951 if (p->numa_scan_seq) {
2953 * Avoid scan adjustments if moving to the preferred
2954 * node or if the task was not previously running on
2955 * the preferred node.
2957 if (dst_nid == p->numa_preferred_nid ||
2958 (p->numa_preferred_nid != NUMA_NO_NODE &&
2959 src_nid != p->numa_preferred_nid))
2963 p->numa_scan_period = task_scan_start(p);
2967 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
2971 static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p)
2975 static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p)
2979 static inline void update_scan_period(struct task_struct *p, int new_cpu)
2983 #endif /* CONFIG_NUMA_BALANCING */
2986 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
2988 update_load_add(&cfs_rq->load, se->load.weight);
2990 if (entity_is_task(se)) {
2991 struct rq *rq = rq_of(cfs_rq);
2993 account_numa_enqueue(rq, task_of(se));
2994 list_add(&se->group_node, &rq->cfs_tasks);
2997 cfs_rq->nr_running++;
3001 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
3003 update_load_sub(&cfs_rq->load, se->load.weight);
3005 if (entity_is_task(se)) {
3006 account_numa_dequeue(rq_of(cfs_rq), task_of(se));
3007 list_del_init(&se->group_node);
3010 cfs_rq->nr_running--;
3014 * Signed add and clamp on underflow.
3016 * Explicitly do a load-store to ensure the intermediate value never hits
3017 * memory. This allows lockless observations without ever seeing the negative
3020 #define add_positive(_ptr, _val) do { \
3021 typeof(_ptr) ptr = (_ptr); \
3022 typeof(_val) val = (_val); \
3023 typeof(*ptr) res, var = READ_ONCE(*ptr); \
3027 if (val < 0 && res > var) \
3030 WRITE_ONCE(*ptr, res); \
3034 * Unsigned subtract and clamp on underflow.
3036 * Explicitly do a load-store to ensure the intermediate value never hits
3037 * memory. This allows lockless observations without ever seeing the negative
3040 #define sub_positive(_ptr, _val) do { \
3041 typeof(_ptr) ptr = (_ptr); \
3042 typeof(*ptr) val = (_val); \
3043 typeof(*ptr) res, var = READ_ONCE(*ptr); \
3047 WRITE_ONCE(*ptr, res); \
3051 * Remove and clamp on negative, from a local variable.
3053 * A variant of sub_positive(), which does not use explicit load-store
3054 * and is thus optimized for local variable updates.
3056 #define lsub_positive(_ptr, _val) do { \
3057 typeof(_ptr) ptr = (_ptr); \
3058 *ptr -= min_t(typeof(*ptr), *ptr, _val); \
3063 enqueue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3065 cfs_rq->avg.load_avg += se->avg.load_avg;
3066 cfs_rq->avg.load_sum += se_weight(se) * se->avg.load_sum;
3070 dequeue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3072 sub_positive(&cfs_rq->avg.load_avg, se->avg.load_avg);
3073 sub_positive(&cfs_rq->avg.load_sum, se_weight(se) * se->avg.load_sum);
3077 enqueue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
3079 dequeue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
3082 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
3083 unsigned long weight)
3086 /* commit outstanding execution time */
3087 if (cfs_rq->curr == se)
3088 update_curr(cfs_rq);
3089 update_load_sub(&cfs_rq->load, se->load.weight);
3091 dequeue_load_avg(cfs_rq, se);
3093 update_load_set(&se->load, weight);
3097 u32 divider = get_pelt_divider(&se->avg);
3099 se->avg.load_avg = div_u64(se_weight(se) * se->avg.load_sum, divider);
3103 enqueue_load_avg(cfs_rq, se);
3105 update_load_add(&cfs_rq->load, se->load.weight);
3109 void reweight_task(struct task_struct *p, int prio)
3111 struct sched_entity *se = &p->se;
3112 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3113 struct load_weight *load = &se->load;
3114 unsigned long weight = scale_load(sched_prio_to_weight[prio]);
3116 reweight_entity(cfs_rq, se, weight);
3117 load->inv_weight = sched_prio_to_wmult[prio];
3120 #ifdef CONFIG_FAIR_GROUP_SCHED
3123 * All this does is approximate the hierarchical proportion which includes that
3124 * global sum we all love to hate.
3126 * That is, the weight of a group entity, is the proportional share of the
3127 * group weight based on the group runqueue weights. That is:
3129 * tg->weight * grq->load.weight
3130 * ge->load.weight = ----------------------------- (1)
3131 * \Sum grq->load.weight
3133 * Now, because computing that sum is prohibitively expensive to compute (been
3134 * there, done that) we approximate it with this average stuff. The average
3135 * moves slower and therefore the approximation is cheaper and more stable.
3137 * So instead of the above, we substitute:
3139 * grq->load.weight -> grq->avg.load_avg (2)
3141 * which yields the following:
3143 * tg->weight * grq->avg.load_avg
3144 * ge->load.weight = ------------------------------ (3)
3147 * Where: tg->load_avg ~= \Sum grq->avg.load_avg
3149 * That is shares_avg, and it is right (given the approximation (2)).
3151 * The problem with it is that because the average is slow -- it was designed
3152 * to be exactly that of course -- this leads to transients in boundary
3153 * conditions. In specific, the case where the group was idle and we start the
3154 * one task. It takes time for our CPU's grq->avg.load_avg to build up,
3155 * yielding bad latency etc..
3157 * Now, in that special case (1) reduces to:
3159 * tg->weight * grq->load.weight
3160 * ge->load.weight = ----------------------------- = tg->weight (4)
3163 * That is, the sum collapses because all other CPUs are idle; the UP scenario.
3165 * So what we do is modify our approximation (3) to approach (4) in the (near)
3170 * tg->weight * grq->load.weight
3171 * --------------------------------------------------- (5)
3172 * tg->load_avg - grq->avg.load_avg + grq->load.weight
3174 * But because grq->load.weight can drop to 0, resulting in a divide by zero,
3175 * we need to use grq->avg.load_avg as its lower bound, which then gives:
3178 * tg->weight * grq->load.weight
3179 * ge->load.weight = ----------------------------- (6)
3184 * tg_load_avg' = tg->load_avg - grq->avg.load_avg +
3185 * max(grq->load.weight, grq->avg.load_avg)
3187 * And that is shares_weight and is icky. In the (near) UP case it approaches
3188 * (4) while in the normal case it approaches (3). It consistently
3189 * overestimates the ge->load.weight and therefore:
3191 * \Sum ge->load.weight >= tg->weight
3195 static long calc_group_shares(struct cfs_rq *cfs_rq)
3197 long tg_weight, tg_shares, load, shares;
3198 struct task_group *tg = cfs_rq->tg;
3200 tg_shares = READ_ONCE(tg->shares);
3202 load = max(scale_load_down(cfs_rq->load.weight), cfs_rq->avg.load_avg);
3204 tg_weight = atomic_long_read(&tg->load_avg);
3206 /* Ensure tg_weight >= load */
3207 tg_weight -= cfs_rq->tg_load_avg_contrib;
3210 shares = (tg_shares * load);
3212 shares /= tg_weight;
3215 * MIN_SHARES has to be unscaled here to support per-CPU partitioning
3216 * of a group with small tg->shares value. It is a floor value which is
3217 * assigned as a minimum load.weight to the sched_entity representing
3218 * the group on a CPU.
3220 * E.g. on 64-bit for a group with tg->shares of scale_load(15)=15*1024
3221 * on an 8-core system with 8 tasks each runnable on one CPU shares has
3222 * to be 15*1024*1/8=1920 instead of scale_load(MIN_SHARES)=2*1024. In
3223 * case no task is runnable on a CPU MIN_SHARES=2 should be returned
3226 return clamp_t(long, shares, MIN_SHARES, tg_shares);
3228 #endif /* CONFIG_SMP */
3230 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
3233 * Recomputes the group entity based on the current state of its group
3236 static void update_cfs_group(struct sched_entity *se)
3238 struct cfs_rq *gcfs_rq = group_cfs_rq(se);
3244 if (throttled_hierarchy(gcfs_rq))
3248 shares = READ_ONCE(gcfs_rq->tg->shares);
3250 if (likely(se->load.weight == shares))
3253 shares = calc_group_shares(gcfs_rq);
3256 reweight_entity(cfs_rq_of(se), se, shares);
3259 #else /* CONFIG_FAIR_GROUP_SCHED */
3260 static inline void update_cfs_group(struct sched_entity *se)
3263 #endif /* CONFIG_FAIR_GROUP_SCHED */
3265 static inline void cfs_rq_util_change(struct cfs_rq *cfs_rq, int flags)
3267 struct rq *rq = rq_of(cfs_rq);
3269 if (&rq->cfs == cfs_rq) {
3271 * There are a few boundary cases this might miss but it should
3272 * get called often enough that that should (hopefully) not be
3275 * It will not get called when we go idle, because the idle
3276 * thread is a different class (!fair), nor will the utilization
3277 * number include things like RT tasks.
3279 * As is, the util number is not freq-invariant (we'd have to
3280 * implement arch_scale_freq_capacity() for that).
3284 cpufreq_update_util(rq, flags);
3289 #ifdef CONFIG_FAIR_GROUP_SCHED
3291 static inline bool cfs_rq_is_decayed(struct cfs_rq *cfs_rq)
3293 if (cfs_rq->load.weight)
3296 if (cfs_rq->avg.load_sum)
3299 if (cfs_rq->avg.util_sum)
3302 if (cfs_rq->avg.runnable_sum)
3306 * _avg must be null when _sum are null because _avg = _sum / divider
3307 * Make sure that rounding and/or propagation of PELT values never
3310 SCHED_WARN_ON(cfs_rq->avg.load_avg ||
3311 cfs_rq->avg.util_avg ||
3312 cfs_rq->avg.runnable_avg);
3318 * update_tg_load_avg - update the tg's load avg
3319 * @cfs_rq: the cfs_rq whose avg changed
3321 * This function 'ensures': tg->load_avg := \Sum tg->cfs_rq[]->avg.load.
3322 * However, because tg->load_avg is a global value there are performance
3325 * In order to avoid having to look at the other cfs_rq's, we use a
3326 * differential update where we store the last value we propagated. This in
3327 * turn allows skipping updates if the differential is 'small'.
3329 * Updating tg's load_avg is necessary before update_cfs_share().
3331 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq)
3333 long delta = cfs_rq->avg.load_avg - cfs_rq->tg_load_avg_contrib;
3336 * No need to update load_avg for root_task_group as it is not used.
3338 if (cfs_rq->tg == &root_task_group)
3341 if (abs(delta) > cfs_rq->tg_load_avg_contrib / 64) {
3342 atomic_long_add(delta, &cfs_rq->tg->load_avg);
3343 cfs_rq->tg_load_avg_contrib = cfs_rq->avg.load_avg;
3348 * Called within set_task_rq() right before setting a task's CPU. The
3349 * caller only guarantees p->pi_lock is held; no other assumptions,
3350 * including the state of rq->lock, should be made.
3352 void set_task_rq_fair(struct sched_entity *se,
3353 struct cfs_rq *prev, struct cfs_rq *next)
3355 u64 p_last_update_time;
3356 u64 n_last_update_time;
3358 if (!sched_feat(ATTACH_AGE_LOAD))
3362 * We are supposed to update the task to "current" time, then its up to
3363 * date and ready to go to new CPU/cfs_rq. But we have difficulty in
3364 * getting what current time is, so simply throw away the out-of-date
3365 * time. This will result in the wakee task is less decayed, but giving
3366 * the wakee more load sounds not bad.
3368 if (!(se->avg.last_update_time && prev))
3371 #ifndef CONFIG_64BIT
3373 u64 p_last_update_time_copy;
3374 u64 n_last_update_time_copy;
3377 p_last_update_time_copy = prev->load_last_update_time_copy;
3378 n_last_update_time_copy = next->load_last_update_time_copy;
3382 p_last_update_time = prev->avg.last_update_time;
3383 n_last_update_time = next->avg.last_update_time;
3385 } while (p_last_update_time != p_last_update_time_copy ||
3386 n_last_update_time != n_last_update_time_copy);
3389 p_last_update_time = prev->avg.last_update_time;
3390 n_last_update_time = next->avg.last_update_time;
3392 __update_load_avg_blocked_se(p_last_update_time, se);
3393 se->avg.last_update_time = n_last_update_time;
3398 * When on migration a sched_entity joins/leaves the PELT hierarchy, we need to
3399 * propagate its contribution. The key to this propagation is the invariant
3400 * that for each group:
3402 * ge->avg == grq->avg (1)
3404 * _IFF_ we look at the pure running and runnable sums. Because they
3405 * represent the very same entity, just at different points in the hierarchy.
3407 * Per the above update_tg_cfs_util() and update_tg_cfs_runnable() are trivial
3408 * and simply copies the running/runnable sum over (but still wrong, because
3409 * the group entity and group rq do not have their PELT windows aligned).
3411 * However, update_tg_cfs_load() is more complex. So we have:
3413 * ge->avg.load_avg = ge->load.weight * ge->avg.runnable_avg (2)
3415 * And since, like util, the runnable part should be directly transferable,
3416 * the following would _appear_ to be the straight forward approach:
3418 * grq->avg.load_avg = grq->load.weight * grq->avg.runnable_avg (3)
3420 * And per (1) we have:
3422 * ge->avg.runnable_avg == grq->avg.runnable_avg
3426 * ge->load.weight * grq->avg.load_avg
3427 * ge->avg.load_avg = ----------------------------------- (4)
3430 * Except that is wrong!
3432 * Because while for entities historical weight is not important and we
3433 * really only care about our future and therefore can consider a pure
3434 * runnable sum, runqueues can NOT do this.
3436 * We specifically want runqueues to have a load_avg that includes
3437 * historical weights. Those represent the blocked load, the load we expect
3438 * to (shortly) return to us. This only works by keeping the weights as
3439 * integral part of the sum. We therefore cannot decompose as per (3).
3441 * Another reason this doesn't work is that runnable isn't a 0-sum entity.
3442 * Imagine a rq with 2 tasks that each are runnable 2/3 of the time. Then the
3443 * rq itself is runnable anywhere between 2/3 and 1 depending on how the
3444 * runnable section of these tasks overlap (or not). If they were to perfectly
3445 * align the rq as a whole would be runnable 2/3 of the time. If however we
3446 * always have at least 1 runnable task, the rq as a whole is always runnable.
3448 * So we'll have to approximate.. :/
3450 * Given the constraint:
3452 * ge->avg.running_sum <= ge->avg.runnable_sum <= LOAD_AVG_MAX
3454 * We can construct a rule that adds runnable to a rq by assuming minimal
3457 * On removal, we'll assume each task is equally runnable; which yields:
3459 * grq->avg.runnable_sum = grq->avg.load_sum / grq->load.weight
3461 * XXX: only do this for the part of runnable > running ?
3466 update_tg_cfs_util(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq)
3468 long delta = gcfs_rq->avg.util_avg - se->avg.util_avg;
3471 /* Nothing to update */
3476 * cfs_rq->avg.period_contrib can be used for both cfs_rq and se.
3477 * See ___update_load_avg() for details.
3479 divider = get_pelt_divider(&cfs_rq->avg);
3481 /* Set new sched_entity's utilization */
3482 se->avg.util_avg = gcfs_rq->avg.util_avg;
3483 se->avg.util_sum = se->avg.util_avg * divider;
3485 /* Update parent cfs_rq utilization */
3486 add_positive(&cfs_rq->avg.util_avg, delta);
3487 cfs_rq->avg.util_sum = cfs_rq->avg.util_avg * divider;
3491 update_tg_cfs_runnable(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq)
3493 long delta = gcfs_rq->avg.runnable_avg - se->avg.runnable_avg;
3496 /* Nothing to update */
3501 * cfs_rq->avg.period_contrib can be used for both cfs_rq and se.
3502 * See ___update_load_avg() for details.
3504 divider = get_pelt_divider(&cfs_rq->avg);
3506 /* Set new sched_entity's runnable */
3507 se->avg.runnable_avg = gcfs_rq->avg.runnable_avg;
3508 se->avg.runnable_sum = se->avg.runnable_avg * divider;
3510 /* Update parent cfs_rq runnable */
3511 add_positive(&cfs_rq->avg.runnable_avg, delta);
3512 cfs_rq->avg.runnable_sum = cfs_rq->avg.runnable_avg * divider;
3516 update_tg_cfs_load(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq)
3518 long delta, running_sum, runnable_sum = gcfs_rq->prop_runnable_sum;
3519 unsigned long load_avg;
3526 gcfs_rq->prop_runnable_sum = 0;
3529 * cfs_rq->avg.period_contrib can be used for both cfs_rq and se.
3530 * See ___update_load_avg() for details.
3532 divider = get_pelt_divider(&cfs_rq->avg);
3534 if (runnable_sum >= 0) {
3536 * Add runnable; clip at LOAD_AVG_MAX. Reflects that until
3537 * the CPU is saturated running == runnable.
3539 runnable_sum += se->avg.load_sum;
3540 runnable_sum = min_t(long, runnable_sum, divider);
3543 * Estimate the new unweighted runnable_sum of the gcfs_rq by
3544 * assuming all tasks are equally runnable.
3546 if (scale_load_down(gcfs_rq->load.weight)) {
3547 load_sum = div_s64(gcfs_rq->avg.load_sum,
3548 scale_load_down(gcfs_rq->load.weight));
3551 /* But make sure to not inflate se's runnable */
3552 runnable_sum = min(se->avg.load_sum, load_sum);
3556 * runnable_sum can't be lower than running_sum
3557 * Rescale running sum to be in the same range as runnable sum
3558 * running_sum is in [0 : LOAD_AVG_MAX << SCHED_CAPACITY_SHIFT]
3559 * runnable_sum is in [0 : LOAD_AVG_MAX]
3561 running_sum = se->avg.util_sum >> SCHED_CAPACITY_SHIFT;
3562 runnable_sum = max(runnable_sum, running_sum);
3564 load_sum = (s64)se_weight(se) * runnable_sum;
3565 load_avg = div_s64(load_sum, divider);
3567 se->avg.load_sum = runnable_sum;
3569 delta = load_avg - se->avg.load_avg;
3573 se->avg.load_avg = load_avg;
3575 add_positive(&cfs_rq->avg.load_avg, delta);
3576 cfs_rq->avg.load_sum = cfs_rq->avg.load_avg * divider;
3579 static inline void add_tg_cfs_propagate(struct cfs_rq *cfs_rq, long runnable_sum)
3581 cfs_rq->propagate = 1;
3582 cfs_rq->prop_runnable_sum += runnable_sum;
3585 /* Update task and its cfs_rq load average */
3586 static inline int propagate_entity_load_avg(struct sched_entity *se)
3588 struct cfs_rq *cfs_rq, *gcfs_rq;
3590 if (entity_is_task(se))
3593 gcfs_rq = group_cfs_rq(se);
3594 if (!gcfs_rq->propagate)
3597 gcfs_rq->propagate = 0;
3599 cfs_rq = cfs_rq_of(se);
3601 add_tg_cfs_propagate(cfs_rq, gcfs_rq->prop_runnable_sum);
3603 update_tg_cfs_util(cfs_rq, se, gcfs_rq);
3604 update_tg_cfs_runnable(cfs_rq, se, gcfs_rq);
3605 update_tg_cfs_load(cfs_rq, se, gcfs_rq);
3607 trace_pelt_cfs_tp(cfs_rq);
3608 trace_pelt_se_tp(se);
3614 * Check if we need to update the load and the utilization of a blocked
3617 static inline bool skip_blocked_update(struct sched_entity *se)
3619 struct cfs_rq *gcfs_rq = group_cfs_rq(se);
3622 * If sched_entity still have not zero load or utilization, we have to
3625 if (se->avg.load_avg || se->avg.util_avg)
3629 * If there is a pending propagation, we have to update the load and
3630 * the utilization of the sched_entity:
3632 if (gcfs_rq->propagate)
3636 * Otherwise, the load and the utilization of the sched_entity is
3637 * already zero and there is no pending propagation, so it will be a
3638 * waste of time to try to decay it:
3643 #else /* CONFIG_FAIR_GROUP_SCHED */
3645 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq) {}
3647 static inline int propagate_entity_load_avg(struct sched_entity *se)
3652 static inline void add_tg_cfs_propagate(struct cfs_rq *cfs_rq, long runnable_sum) {}
3654 #endif /* CONFIG_FAIR_GROUP_SCHED */
3657 * update_cfs_rq_load_avg - update the cfs_rq's load/util averages
3658 * @now: current time, as per cfs_rq_clock_pelt()
3659 * @cfs_rq: cfs_rq to update
3661 * The cfs_rq avg is the direct sum of all its entities (blocked and runnable)
3662 * avg. The immediate corollary is that all (fair) tasks must be attached, see
3663 * post_init_entity_util_avg().
3665 * cfs_rq->avg is used for task_h_load() and update_cfs_share() for example.
3667 * Returns true if the load decayed or we removed load.
3669 * Since both these conditions indicate a changed cfs_rq->avg.load we should
3670 * call update_tg_load_avg() when this function returns true.
3673 update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq)
3675 unsigned long removed_load = 0, removed_util = 0, removed_runnable = 0;
3676 struct sched_avg *sa = &cfs_rq->avg;
3679 if (cfs_rq->removed.nr) {
3681 u32 divider = get_pelt_divider(&cfs_rq->avg);
3683 raw_spin_lock(&cfs_rq->removed.lock);
3684 swap(cfs_rq->removed.util_avg, removed_util);
3685 swap(cfs_rq->removed.load_avg, removed_load);
3686 swap(cfs_rq->removed.runnable_avg, removed_runnable);
3687 cfs_rq->removed.nr = 0;
3688 raw_spin_unlock(&cfs_rq->removed.lock);
3691 sub_positive(&sa->load_avg, r);
3692 sa->load_sum = sa->load_avg * divider;
3695 sub_positive(&sa->util_avg, r);
3696 sa->util_sum = sa->util_avg * divider;
3698 r = removed_runnable;
3699 sub_positive(&sa->runnable_avg, r);
3700 sa->runnable_sum = sa->runnable_avg * divider;
3703 * removed_runnable is the unweighted version of removed_load so we
3704 * can use it to estimate removed_load_sum.
3706 add_tg_cfs_propagate(cfs_rq,
3707 -(long)(removed_runnable * divider) >> SCHED_CAPACITY_SHIFT);
3712 decayed |= __update_load_avg_cfs_rq(now, cfs_rq);
3714 #ifndef CONFIG_64BIT
3716 cfs_rq->load_last_update_time_copy = sa->last_update_time;
3723 * attach_entity_load_avg - attach this entity to its cfs_rq load avg
3724 * @cfs_rq: cfs_rq to attach to
3725 * @se: sched_entity to attach
3727 * Must call update_cfs_rq_load_avg() before this, since we rely on
3728 * cfs_rq->avg.last_update_time being current.
3730 static void attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3733 * cfs_rq->avg.period_contrib can be used for both cfs_rq and se.
3734 * See ___update_load_avg() for details.
3736 u32 divider = get_pelt_divider(&cfs_rq->avg);
3739 * When we attach the @se to the @cfs_rq, we must align the decay
3740 * window because without that, really weird and wonderful things can
3745 se->avg.last_update_time = cfs_rq->avg.last_update_time;
3746 se->avg.period_contrib = cfs_rq->avg.period_contrib;
3749 * Hell(o) Nasty stuff.. we need to recompute _sum based on the new
3750 * period_contrib. This isn't strictly correct, but since we're
3751 * entirely outside of the PELT hierarchy, nobody cares if we truncate
3754 se->avg.util_sum = se->avg.util_avg * divider;
3756 se->avg.runnable_sum = se->avg.runnable_avg * divider;
3758 se->avg.load_sum = divider;
3759 if (se_weight(se)) {
3761 div_u64(se->avg.load_avg * se->avg.load_sum, se_weight(se));
3764 enqueue_load_avg(cfs_rq, se);
3765 cfs_rq->avg.util_avg += se->avg.util_avg;
3766 cfs_rq->avg.util_sum += se->avg.util_sum;
3767 cfs_rq->avg.runnable_avg += se->avg.runnable_avg;
3768 cfs_rq->avg.runnable_sum += se->avg.runnable_sum;
3770 add_tg_cfs_propagate(cfs_rq, se->avg.load_sum);
3772 cfs_rq_util_change(cfs_rq, 0);
3774 trace_pelt_cfs_tp(cfs_rq);
3778 * detach_entity_load_avg - detach this entity from its cfs_rq load avg
3779 * @cfs_rq: cfs_rq to detach from
3780 * @se: sched_entity to detach
3782 * Must call update_cfs_rq_load_avg() before this, since we rely on
3783 * cfs_rq->avg.last_update_time being current.
3785 static void detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3788 * cfs_rq->avg.period_contrib can be used for both cfs_rq and se.
3789 * See ___update_load_avg() for details.
3791 u32 divider = get_pelt_divider(&cfs_rq->avg);
3793 dequeue_load_avg(cfs_rq, se);
3794 sub_positive(&cfs_rq->avg.util_avg, se->avg.util_avg);
3795 cfs_rq->avg.util_sum = cfs_rq->avg.util_avg * divider;
3796 sub_positive(&cfs_rq->avg.runnable_avg, se->avg.runnable_avg);
3797 cfs_rq->avg.runnable_sum = cfs_rq->avg.runnable_avg * divider;
3799 add_tg_cfs_propagate(cfs_rq, -se->avg.load_sum);
3801 cfs_rq_util_change(cfs_rq, 0);
3803 trace_pelt_cfs_tp(cfs_rq);
3807 * Optional action to be done while updating the load average
3809 #define UPDATE_TG 0x1
3810 #define SKIP_AGE_LOAD 0x2
3811 #define DO_ATTACH 0x4
3813 /* Update task and its cfs_rq load average */
3814 static inline void update_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
3816 u64 now = cfs_rq_clock_pelt(cfs_rq);
3820 * Track task load average for carrying it to new CPU after migrated, and
3821 * track group sched_entity load average for task_h_load calc in migration
3823 if (se->avg.last_update_time && !(flags & SKIP_AGE_LOAD))
3824 __update_load_avg_se(now, cfs_rq, se);
3826 decayed = update_cfs_rq_load_avg(now, cfs_rq);
3827 decayed |= propagate_entity_load_avg(se);
3829 if (!se->avg.last_update_time && (flags & DO_ATTACH)) {
3832 * DO_ATTACH means we're here from enqueue_entity().
3833 * !last_update_time means we've passed through
3834 * migrate_task_rq_fair() indicating we migrated.
3836 * IOW we're enqueueing a task on a new CPU.
3838 attach_entity_load_avg(cfs_rq, se);
3839 update_tg_load_avg(cfs_rq);
3841 } else if (decayed) {
3842 cfs_rq_util_change(cfs_rq, 0);
3844 if (flags & UPDATE_TG)
3845 update_tg_load_avg(cfs_rq);
3849 #ifndef CONFIG_64BIT
3850 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
3852 u64 last_update_time_copy;
3853 u64 last_update_time;
3856 last_update_time_copy = cfs_rq->load_last_update_time_copy;
3858 last_update_time = cfs_rq->avg.last_update_time;
3859 } while (last_update_time != last_update_time_copy);
3861 return last_update_time;
3864 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
3866 return cfs_rq->avg.last_update_time;
3871 * Synchronize entity load avg of dequeued entity without locking
3874 static void sync_entity_load_avg(struct sched_entity *se)
3876 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3877 u64 last_update_time;
3879 last_update_time = cfs_rq_last_update_time(cfs_rq);
3880 __update_load_avg_blocked_se(last_update_time, se);
3884 * Task first catches up with cfs_rq, and then subtract
3885 * itself from the cfs_rq (task must be off the queue now).
3887 static void remove_entity_load_avg(struct sched_entity *se)
3889 struct cfs_rq *cfs_rq = cfs_rq_of(se);
3890 unsigned long flags;
3893 * tasks cannot exit without having gone through wake_up_new_task() ->
3894 * post_init_entity_util_avg() which will have added things to the
3895 * cfs_rq, so we can remove unconditionally.
3898 sync_entity_load_avg(se);
3900 raw_spin_lock_irqsave(&cfs_rq->removed.lock, flags);
3901 ++cfs_rq->removed.nr;
3902 cfs_rq->removed.util_avg += se->avg.util_avg;
3903 cfs_rq->removed.load_avg += se->avg.load_avg;
3904 cfs_rq->removed.runnable_avg += se->avg.runnable_avg;
3905 raw_spin_unlock_irqrestore(&cfs_rq->removed.lock, flags);
3908 static inline unsigned long cfs_rq_runnable_avg(struct cfs_rq *cfs_rq)
3910 return cfs_rq->avg.runnable_avg;
3913 static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq)
3915 return cfs_rq->avg.load_avg;
3918 static int newidle_balance(struct rq *this_rq, struct rq_flags *rf);
3920 static inline unsigned long task_util(struct task_struct *p)
3922 return READ_ONCE(p->se.avg.util_avg);
3925 static inline unsigned long _task_util_est(struct task_struct *p)
3927 struct util_est ue = READ_ONCE(p->se.avg.util_est);
3929 return max(ue.ewma, (ue.enqueued & ~UTIL_AVG_UNCHANGED));
3932 static inline unsigned long task_util_est(struct task_struct *p)
3934 return max(task_util(p), _task_util_est(p));
3937 #ifdef CONFIG_UCLAMP_TASK
3938 static inline unsigned long uclamp_task_util(struct task_struct *p)
3940 return clamp(task_util_est(p),
3941 uclamp_eff_value(p, UCLAMP_MIN),
3942 uclamp_eff_value(p, UCLAMP_MAX));
3945 static inline unsigned long uclamp_task_util(struct task_struct *p)
3947 return task_util_est(p);
3951 static inline void util_est_enqueue(struct cfs_rq *cfs_rq,
3952 struct task_struct *p)
3954 unsigned int enqueued;
3956 if (!sched_feat(UTIL_EST))
3959 /* Update root cfs_rq's estimated utilization */
3960 enqueued = cfs_rq->avg.util_est.enqueued;
3961 enqueued += _task_util_est(p);
3962 WRITE_ONCE(cfs_rq->avg.util_est.enqueued, enqueued);
3964 trace_sched_util_est_cfs_tp(cfs_rq);
3967 static inline void util_est_dequeue(struct cfs_rq *cfs_rq,
3968 struct task_struct *p)
3970 unsigned int enqueued;
3972 if (!sched_feat(UTIL_EST))
3975 /* Update root cfs_rq's estimated utilization */
3976 enqueued = cfs_rq->avg.util_est.enqueued;
3977 enqueued -= min_t(unsigned int, enqueued, _task_util_est(p));
3978 WRITE_ONCE(cfs_rq->avg.util_est.enqueued, enqueued);
3980 trace_sched_util_est_cfs_tp(cfs_rq);
3983 #define UTIL_EST_MARGIN (SCHED_CAPACITY_SCALE / 100)
3986 * Check if a (signed) value is within a specified (unsigned) margin,
3987 * based on the observation that:
3989 * abs(x) < y := (unsigned)(x + y - 1) < (2 * y - 1)
3991 * NOTE: this only works when value + margin < INT_MAX.
3993 static inline bool within_margin(int value, int margin)
3995 return ((unsigned int)(value + margin - 1) < (2 * margin - 1));
3998 static inline void util_est_update(struct cfs_rq *cfs_rq,
3999 struct task_struct *p,
4002 long last_ewma_diff, last_enqueued_diff;
4005 if (!sched_feat(UTIL_EST))
4009 * Skip update of task's estimated utilization when the task has not
4010 * yet completed an activation, e.g. being migrated.
4016 * If the PELT values haven't changed since enqueue time,
4017 * skip the util_est update.
4019 ue = p->se.avg.util_est;
4020 if (ue.enqueued & UTIL_AVG_UNCHANGED)
4023 last_enqueued_diff = ue.enqueued;
4026 * Reset EWMA on utilization increases, the moving average is used only
4027 * to smooth utilization decreases.
4029 ue.enqueued = task_util(p);
4030 if (sched_feat(UTIL_EST_FASTUP)) {
4031 if (ue.ewma < ue.enqueued) {
4032 ue.ewma = ue.enqueued;
4038 * Skip update of task's estimated utilization when its members are
4039 * already ~1% close to its last activation value.
4041 last_ewma_diff = ue.enqueued - ue.ewma;
4042 last_enqueued_diff -= ue.enqueued;
4043 if (within_margin(last_ewma_diff, UTIL_EST_MARGIN)) {
4044 if (!within_margin(last_enqueued_diff, UTIL_EST_MARGIN))
4051 * To avoid overestimation of actual task utilization, skip updates if
4052 * we cannot grant there is idle time in this CPU.
4054 if (task_util(p) > capacity_orig_of(cpu_of(rq_of(cfs_rq))))
4058 * Update Task's estimated utilization
4060 * When *p completes an activation we can consolidate another sample
4061 * of the task size. This is done by storing the current PELT value
4062 * as ue.enqueued and by using this value to update the Exponential
4063 * Weighted Moving Average (EWMA):
4065 * ewma(t) = w * task_util(p) + (1-w) * ewma(t-1)
4066 * = w * task_util(p) + ewma(t-1) - w * ewma(t-1)
4067 * = w * (task_util(p) - ewma(t-1)) + ewma(t-1)
4068 * = w * ( last_ewma_diff ) + ewma(t-1)
4069 * = w * (last_ewma_diff + ewma(t-1) / w)
4071 * Where 'w' is the weight of new samples, which is configured to be
4072 * 0.25, thus making w=1/4 ( >>= UTIL_EST_WEIGHT_SHIFT)
4074 ue.ewma <<= UTIL_EST_WEIGHT_SHIFT;
4075 ue.ewma += last_ewma_diff;
4076 ue.ewma >>= UTIL_EST_WEIGHT_SHIFT;
4078 ue.enqueued |= UTIL_AVG_UNCHANGED;
4079 WRITE_ONCE(p->se.avg.util_est, ue);
4081 trace_sched_util_est_se_tp(&p->se);
4084 static inline int task_fits_capacity(struct task_struct *p, long capacity)
4086 return fits_capacity(uclamp_task_util(p), capacity);
4089 static inline void update_misfit_status(struct task_struct *p, struct rq *rq)
4091 if (!static_branch_unlikely(&sched_asym_cpucapacity))
4094 if (!p || p->nr_cpus_allowed == 1) {
4095 rq->misfit_task_load = 0;
4099 if (task_fits_capacity(p, capacity_of(cpu_of(rq)))) {
4100 rq->misfit_task_load = 0;
4105 * Make sure that misfit_task_load will not be null even if
4106 * task_h_load() returns 0.
4108 rq->misfit_task_load = max_t(unsigned long, task_h_load(p), 1);
4111 #else /* CONFIG_SMP */
4113 static inline bool cfs_rq_is_decayed(struct cfs_rq *cfs_rq)
4118 #define UPDATE_TG 0x0
4119 #define SKIP_AGE_LOAD 0x0
4120 #define DO_ATTACH 0x0
4122 static inline void update_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int not_used1)
4124 cfs_rq_util_change(cfs_rq, 0);
4127 static inline void remove_entity_load_avg(struct sched_entity *se) {}
4130 attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
4132 detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
4134 static inline int newidle_balance(struct rq *rq, struct rq_flags *rf)
4140 util_est_enqueue(struct cfs_rq *cfs_rq, struct task_struct *p) {}
4143 util_est_dequeue(struct cfs_rq *cfs_rq, struct task_struct *p) {}
4146 util_est_update(struct cfs_rq *cfs_rq, struct task_struct *p,
4148 static inline void update_misfit_status(struct task_struct *p, struct rq *rq) {}
4150 #endif /* CONFIG_SMP */
4152 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
4154 #ifdef CONFIG_SCHED_DEBUG
4155 s64 d = se->vruntime - cfs_rq->min_vruntime;
4160 if (d > 3*sysctl_sched_latency)
4161 schedstat_inc(cfs_rq->nr_spread_over);
4166 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
4168 u64 vruntime = cfs_rq->min_vruntime;
4171 * The 'current' period is already promised to the current tasks,
4172 * however the extra weight of the new task will slow them down a
4173 * little, place the new task so that it fits in the slot that
4174 * stays open at the end.
4176 if (initial && sched_feat(START_DEBIT))
4177 vruntime += sched_vslice(cfs_rq, se);
4179 /* sleeps up to a single latency don't count. */
4181 unsigned long thresh = sysctl_sched_latency;
4184 * Halve their sleep time's effect, to allow
4185 * for a gentler effect of sleepers:
4187 if (sched_feat(GENTLE_FAIR_SLEEPERS))
4193 /* ensure we never gain time by being placed backwards. */
4194 se->vruntime = max_vruntime(se->vruntime, vruntime);
4197 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
4199 static inline void check_schedstat_required(void)
4201 #ifdef CONFIG_SCHEDSTATS
4202 if (schedstat_enabled())
4205 /* Force schedstat enabled if a dependent tracepoint is active */
4206 if (trace_sched_stat_wait_enabled() ||
4207 trace_sched_stat_sleep_enabled() ||
4208 trace_sched_stat_iowait_enabled() ||
4209 trace_sched_stat_blocked_enabled() ||
4210 trace_sched_stat_runtime_enabled()) {
4211 printk_deferred_once("Scheduler tracepoints stat_sleep, stat_iowait, "
4212 "stat_blocked and stat_runtime require the "
4213 "kernel parameter schedstats=enable or "
4214 "kernel.sched_schedstats=1\n");
4219 static inline bool cfs_bandwidth_used(void);
4226 * update_min_vruntime()
4227 * vruntime -= min_vruntime
4231 * update_min_vruntime()
4232 * vruntime += min_vruntime
4234 * this way the vruntime transition between RQs is done when both
4235 * min_vruntime are up-to-date.
4239 * ->migrate_task_rq_fair() (p->state == TASK_WAKING)
4240 * vruntime -= min_vruntime
4244 * update_min_vruntime()
4245 * vruntime += min_vruntime
4247 * this way we don't have the most up-to-date min_vruntime on the originating
4248 * CPU and an up-to-date min_vruntime on the destination CPU.
4252 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
4254 bool renorm = !(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_MIGRATED);
4255 bool curr = cfs_rq->curr == se;
4258 * If we're the current task, we must renormalise before calling
4262 se->vruntime += cfs_rq->min_vruntime;
4264 update_curr(cfs_rq);
4267 * Otherwise, renormalise after, such that we're placed at the current
4268 * moment in time, instead of some random moment in the past. Being
4269 * placed in the past could significantly boost this task to the
4270 * fairness detriment of existing tasks.
4272 if (renorm && !curr)
4273 se->vruntime += cfs_rq->min_vruntime;
4276 * When enqueuing a sched_entity, we must:
4277 * - Update loads to have both entity and cfs_rq synced with now.
4278 * - Add its load to cfs_rq->runnable_avg
4279 * - For group_entity, update its weight to reflect the new share of
4281 * - Add its new weight to cfs_rq->load.weight
4283 update_load_avg(cfs_rq, se, UPDATE_TG | DO_ATTACH);
4284 se_update_runnable(se);
4285 update_cfs_group(se);
4286 account_entity_enqueue(cfs_rq, se);
4288 if (flags & ENQUEUE_WAKEUP)
4289 place_entity(cfs_rq, se, 0);
4291 check_schedstat_required();
4292 update_stats_enqueue(cfs_rq, se, flags);
4293 check_spread(cfs_rq, se);
4295 __enqueue_entity(cfs_rq, se);
4299 * When bandwidth control is enabled, cfs might have been removed
4300 * because of a parent been throttled but cfs->nr_running > 1. Try to
4301 * add it unconditionally.
4303 if (cfs_rq->nr_running == 1 || cfs_bandwidth_used())
4304 list_add_leaf_cfs_rq(cfs_rq);
4306 if (cfs_rq->nr_running == 1)
4307 check_enqueue_throttle(cfs_rq);
4310 static void __clear_buddies_last(struct sched_entity *se)
4312 for_each_sched_entity(se) {
4313 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4314 if (cfs_rq->last != se)
4317 cfs_rq->last = NULL;
4321 static void __clear_buddies_next(struct sched_entity *se)
4323 for_each_sched_entity(se) {
4324 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4325 if (cfs_rq->next != se)
4328 cfs_rq->next = NULL;
4332 static void __clear_buddies_skip(struct sched_entity *se)
4334 for_each_sched_entity(se) {
4335 struct cfs_rq *cfs_rq = cfs_rq_of(se);
4336 if (cfs_rq->skip != se)
4339 cfs_rq->skip = NULL;
4343 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
4345 if (cfs_rq->last == se)
4346 __clear_buddies_last(se);
4348 if (cfs_rq->next == se)
4349 __clear_buddies_next(se);
4351 if (cfs_rq->skip == se)
4352 __clear_buddies_skip(se);
4355 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
4358 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
4361 * Update run-time statistics of the 'current'.
4363 update_curr(cfs_rq);
4366 * When dequeuing a sched_entity, we must:
4367 * - Update loads to have both entity and cfs_rq synced with now.
4368 * - Subtract its load from the cfs_rq->runnable_avg.
4369 * - Subtract its previous weight from cfs_rq->load.weight.
4370 * - For group entity, update its weight to reflect the new share
4371 * of its group cfs_rq.
4373 update_load_avg(cfs_rq, se, UPDATE_TG);
4374 se_update_runnable(se);
4376 update_stats_dequeue(cfs_rq, se, flags);
4378 clear_buddies(cfs_rq, se);
4380 if (se != cfs_rq->curr)
4381 __dequeue_entity(cfs_rq, se);
4383 account_entity_dequeue(cfs_rq, se);
4386 * Normalize after update_curr(); which will also have moved
4387 * min_vruntime if @se is the one holding it back. But before doing
4388 * update_min_vruntime() again, which will discount @se's position and
4389 * can move min_vruntime forward still more.
4391 if (!(flags & DEQUEUE_SLEEP))
4392 se->vruntime -= cfs_rq->min_vruntime;
4394 /* return excess runtime on last dequeue */
4395 return_cfs_rq_runtime(cfs_rq);
4397 update_cfs_group(se);
4400 * Now advance min_vruntime if @se was the entity holding it back,
4401 * except when: DEQUEUE_SAVE && !DEQUEUE_MOVE, in this case we'll be
4402 * put back on, and if we advance min_vruntime, we'll be placed back
4403 * further than we started -- ie. we'll be penalized.
4405 if ((flags & (DEQUEUE_SAVE | DEQUEUE_MOVE)) != DEQUEUE_SAVE)
4406 update_min_vruntime(cfs_rq);
4410 * Preempt the current task with a newly woken task if needed:
4413 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
4415 unsigned long ideal_runtime, delta_exec;
4416 struct sched_entity *se;
4419 ideal_runtime = sched_slice(cfs_rq, curr);
4420 delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
4421 if (delta_exec > ideal_runtime) {
4422 resched_curr(rq_of(cfs_rq));
4424 * The current task ran long enough, ensure it doesn't get
4425 * re-elected due to buddy favours.
4427 clear_buddies(cfs_rq, curr);
4432 * Ensure that a task that missed wakeup preemption by a
4433 * narrow margin doesn't have to wait for a full slice.
4434 * This also mitigates buddy induced latencies under load.
4436 if (delta_exec < sysctl_sched_min_granularity)
4439 se = __pick_first_entity(cfs_rq);
4440 delta = curr->vruntime - se->vruntime;
4445 if (delta > ideal_runtime)
4446 resched_curr(rq_of(cfs_rq));
4450 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
4452 clear_buddies(cfs_rq, se);
4454 /* 'current' is not kept within the tree. */
4457 * Any task has to be enqueued before it get to execute on
4458 * a CPU. So account for the time it spent waiting on the
4461 update_stats_wait_end(cfs_rq, se);
4462 __dequeue_entity(cfs_rq, se);
4463 update_load_avg(cfs_rq, se, UPDATE_TG);
4466 update_stats_curr_start(cfs_rq, se);
4470 * Track our maximum slice length, if the CPU's load is at
4471 * least twice that of our own weight (i.e. dont track it
4472 * when there are only lesser-weight tasks around):
4474 if (schedstat_enabled() &&
4475 rq_of(cfs_rq)->cfs.load.weight >= 2*se->load.weight) {
4476 schedstat_set(se->statistics.slice_max,
4477 max((u64)schedstat_val(se->statistics.slice_max),
4478 se->sum_exec_runtime - se->prev_sum_exec_runtime));
4481 se->prev_sum_exec_runtime = se->sum_exec_runtime;
4485 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
4488 * Pick the next process, keeping these things in mind, in this order:
4489 * 1) keep things fair between processes/task groups
4490 * 2) pick the "next" process, since someone really wants that to run
4491 * 3) pick the "last" process, for cache locality
4492 * 4) do not run the "skip" process, if something else is available
4494 static struct sched_entity *
4495 pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr)
4497 struct sched_entity *left = __pick_first_entity(cfs_rq);
4498 struct sched_entity *se;
4501 * If curr is set we have to see if its left of the leftmost entity
4502 * still in the tree, provided there was anything in the tree at all.
4504 if (!left || (curr && entity_before(curr, left)))
4507 se = left; /* ideally we run the leftmost entity */
4510 * Avoid running the skip buddy, if running something else can
4511 * be done without getting too unfair.
4513 if (cfs_rq->skip && cfs_rq->skip == se) {
4514 struct sched_entity *second;
4517 second = __pick_first_entity(cfs_rq);
4519 second = __pick_next_entity(se);
4520 if (!second || (curr && entity_before(curr, second)))
4524 if (second && wakeup_preempt_entity(second, left) < 1)
4528 if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1) {
4530 * Someone really wants this to run. If it's not unfair, run it.
4533 } else if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1) {
4535 * Prefer last buddy, try to return the CPU to a preempted task.
4543 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
4545 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
4548 * If still on the runqueue then deactivate_task()
4549 * was not called and update_curr() has to be done:
4552 update_curr(cfs_rq);
4554 /* throttle cfs_rqs exceeding runtime */
4555 check_cfs_rq_runtime(cfs_rq);
4557 check_spread(cfs_rq, prev);
4560 update_stats_wait_start(cfs_rq, prev);
4561 /* Put 'current' back into the tree. */
4562 __enqueue_entity(cfs_rq, prev);
4563 /* in !on_rq case, update occurred at dequeue */
4564 update_load_avg(cfs_rq, prev, 0);
4566 cfs_rq->curr = NULL;
4570 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
4573 * Update run-time statistics of the 'current'.
4575 update_curr(cfs_rq);
4578 * Ensure that runnable average is periodically updated.
4580 update_load_avg(cfs_rq, curr, UPDATE_TG);
4581 update_cfs_group(curr);
4583 #ifdef CONFIG_SCHED_HRTICK
4585 * queued ticks are scheduled to match the slice, so don't bother
4586 * validating it and just reschedule.
4589 resched_curr(rq_of(cfs_rq));
4593 * don't let the period tick interfere with the hrtick preemption
4595 if (!sched_feat(DOUBLE_TICK) &&
4596 hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
4600 if (cfs_rq->nr_running > 1)
4601 check_preempt_tick(cfs_rq, curr);
4605 /**************************************************
4606 * CFS bandwidth control machinery
4609 #ifdef CONFIG_CFS_BANDWIDTH
4611 #ifdef CONFIG_JUMP_LABEL
4612 static struct static_key __cfs_bandwidth_used;
4614 static inline bool cfs_bandwidth_used(void)
4616 return static_key_false(&__cfs_bandwidth_used);
4619 void cfs_bandwidth_usage_inc(void)
4621 static_key_slow_inc_cpuslocked(&__cfs_bandwidth_used);
4624 void cfs_bandwidth_usage_dec(void)
4626 static_key_slow_dec_cpuslocked(&__cfs_bandwidth_used);
4628 #else /* CONFIG_JUMP_LABEL */
4629 static bool cfs_bandwidth_used(void)
4634 void cfs_bandwidth_usage_inc(void) {}
4635 void cfs_bandwidth_usage_dec(void) {}
4636 #endif /* CONFIG_JUMP_LABEL */
4639 * default period for cfs group bandwidth.
4640 * default: 0.1s, units: nanoseconds
4642 static inline u64 default_cfs_period(void)
4644 return 100000000ULL;
4647 static inline u64 sched_cfs_bandwidth_slice(void)
4649 return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
4653 * Replenish runtime according to assigned quota. We use sched_clock_cpu
4654 * directly instead of rq->clock to avoid adding additional synchronization
4657 * requires cfs_b->lock
4659 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
4661 if (unlikely(cfs_b->quota == RUNTIME_INF))
4664 cfs_b->runtime += cfs_b->quota;
4665 cfs_b->runtime = min(cfs_b->runtime, cfs_b->quota + cfs_b->burst);
4668 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
4670 return &tg->cfs_bandwidth;
4673 /* returns 0 on failure to allocate runtime */
4674 static int __assign_cfs_rq_runtime(struct cfs_bandwidth *cfs_b,
4675 struct cfs_rq *cfs_rq, u64 target_runtime)
4677 u64 min_amount, amount = 0;
4679 lockdep_assert_held(&cfs_b->lock);
4681 /* note: this is a positive sum as runtime_remaining <= 0 */
4682 min_amount = target_runtime - cfs_rq->runtime_remaining;
4684 if (cfs_b->quota == RUNTIME_INF)
4685 amount = min_amount;
4687 start_cfs_bandwidth(cfs_b);
4689 if (cfs_b->runtime > 0) {
4690 amount = min(cfs_b->runtime, min_amount);
4691 cfs_b->runtime -= amount;
4696 cfs_rq->runtime_remaining += amount;
4698 return cfs_rq->runtime_remaining > 0;
4701 /* returns 0 on failure to allocate runtime */
4702 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
4704 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4707 raw_spin_lock(&cfs_b->lock);
4708 ret = __assign_cfs_rq_runtime(cfs_b, cfs_rq, sched_cfs_bandwidth_slice());
4709 raw_spin_unlock(&cfs_b->lock);
4714 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
4716 /* dock delta_exec before expiring quota (as it could span periods) */
4717 cfs_rq->runtime_remaining -= delta_exec;
4719 if (likely(cfs_rq->runtime_remaining > 0))
4722 if (cfs_rq->throttled)
4725 * if we're unable to extend our runtime we resched so that the active
4726 * hierarchy can be throttled
4728 if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
4729 resched_curr(rq_of(cfs_rq));
4732 static __always_inline
4733 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
4735 if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
4738 __account_cfs_rq_runtime(cfs_rq, delta_exec);
4741 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
4743 return cfs_bandwidth_used() && cfs_rq->throttled;
4746 /* check whether cfs_rq, or any parent, is throttled */
4747 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
4749 return cfs_bandwidth_used() && cfs_rq->throttle_count;
4753 * Ensure that neither of the group entities corresponding to src_cpu or
4754 * dest_cpu are members of a throttled hierarchy when performing group
4755 * load-balance operations.
4757 static inline int throttled_lb_pair(struct task_group *tg,
4758 int src_cpu, int dest_cpu)
4760 struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
4762 src_cfs_rq = tg->cfs_rq[src_cpu];
4763 dest_cfs_rq = tg->cfs_rq[dest_cpu];
4765 return throttled_hierarchy(src_cfs_rq) ||
4766 throttled_hierarchy(dest_cfs_rq);
4769 static int tg_unthrottle_up(struct task_group *tg, void *data)
4771 struct rq *rq = data;
4772 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
4774 cfs_rq->throttle_count--;
4775 if (!cfs_rq->throttle_count) {
4776 cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
4777 cfs_rq->throttled_clock_task;
4779 /* Add cfs_rq with load or one or more already running entities to the list */
4780 if (!cfs_rq_is_decayed(cfs_rq) || cfs_rq->nr_running)
4781 list_add_leaf_cfs_rq(cfs_rq);
4787 static int tg_throttle_down(struct task_group *tg, void *data)
4789 struct rq *rq = data;
4790 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
4792 /* group is entering throttled state, stop time */
4793 if (!cfs_rq->throttle_count) {
4794 cfs_rq->throttled_clock_task = rq_clock_task(rq);
4795 list_del_leaf_cfs_rq(cfs_rq);
4797 cfs_rq->throttle_count++;
4802 static bool throttle_cfs_rq(struct cfs_rq *cfs_rq)
4804 struct rq *rq = rq_of(cfs_rq);
4805 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4806 struct sched_entity *se;
4807 long task_delta, idle_task_delta, dequeue = 1;
4809 raw_spin_lock(&cfs_b->lock);
4810 /* This will start the period timer if necessary */
4811 if (__assign_cfs_rq_runtime(cfs_b, cfs_rq, 1)) {
4813 * We have raced with bandwidth becoming available, and if we
4814 * actually throttled the timer might not unthrottle us for an
4815 * entire period. We additionally needed to make sure that any
4816 * subsequent check_cfs_rq_runtime calls agree not to throttle
4817 * us, as we may commit to do cfs put_prev+pick_next, so we ask
4818 * for 1ns of runtime rather than just check cfs_b.
4822 list_add_tail_rcu(&cfs_rq->throttled_list,
4823 &cfs_b->throttled_cfs_rq);
4825 raw_spin_unlock(&cfs_b->lock);
4828 return false; /* Throttle no longer required. */
4830 se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
4832 /* freeze hierarchy runnable averages while throttled */
4834 walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
4837 task_delta = cfs_rq->h_nr_running;
4838 idle_task_delta = cfs_rq->idle_h_nr_running;
4839 for_each_sched_entity(se) {
4840 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
4841 /* throttled entity or throttle-on-deactivate */
4845 dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
4847 if (cfs_rq_is_idle(group_cfs_rq(se)))
4848 idle_task_delta = cfs_rq->h_nr_running;
4850 qcfs_rq->h_nr_running -= task_delta;
4851 qcfs_rq->idle_h_nr_running -= idle_task_delta;
4853 if (qcfs_rq->load.weight) {
4854 /* Avoid re-evaluating load for this entity: */
4855 se = parent_entity(se);
4860 for_each_sched_entity(se) {
4861 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
4862 /* throttled entity or throttle-on-deactivate */
4866 update_load_avg(qcfs_rq, se, 0);
4867 se_update_runnable(se);
4869 if (cfs_rq_is_idle(group_cfs_rq(se)))
4870 idle_task_delta = cfs_rq->h_nr_running;
4872 qcfs_rq->h_nr_running -= task_delta;
4873 qcfs_rq->idle_h_nr_running -= idle_task_delta;
4876 /* At this point se is NULL and we are at root level*/
4877 sub_nr_running(rq, task_delta);
4881 * Note: distribution will already see us throttled via the
4882 * throttled-list. rq->lock protects completion.
4884 cfs_rq->throttled = 1;
4885 cfs_rq->throttled_clock = rq_clock(rq);
4889 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
4891 struct rq *rq = rq_of(cfs_rq);
4892 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
4893 struct sched_entity *se;
4894 long task_delta, idle_task_delta;
4896 se = cfs_rq->tg->se[cpu_of(rq)];
4898 cfs_rq->throttled = 0;
4900 update_rq_clock(rq);
4902 raw_spin_lock(&cfs_b->lock);
4903 cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
4904 list_del_rcu(&cfs_rq->throttled_list);
4905 raw_spin_unlock(&cfs_b->lock);
4907 /* update hierarchical throttle state */
4908 walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
4910 if (!cfs_rq->load.weight)
4913 task_delta = cfs_rq->h_nr_running;
4914 idle_task_delta = cfs_rq->idle_h_nr_running;
4915 for_each_sched_entity(se) {
4916 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
4920 enqueue_entity(qcfs_rq, se, ENQUEUE_WAKEUP);
4922 if (cfs_rq_is_idle(group_cfs_rq(se)))
4923 idle_task_delta = cfs_rq->h_nr_running;
4925 qcfs_rq->h_nr_running += task_delta;
4926 qcfs_rq->idle_h_nr_running += idle_task_delta;
4928 /* end evaluation on encountering a throttled cfs_rq */
4929 if (cfs_rq_throttled(qcfs_rq))
4930 goto unthrottle_throttle;
4933 for_each_sched_entity(se) {
4934 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
4936 update_load_avg(qcfs_rq, se, UPDATE_TG);
4937 se_update_runnable(se);
4939 if (cfs_rq_is_idle(group_cfs_rq(se)))
4940 idle_task_delta = cfs_rq->h_nr_running;
4942 qcfs_rq->h_nr_running += task_delta;
4943 qcfs_rq->idle_h_nr_running += idle_task_delta;
4945 /* end evaluation on encountering a throttled cfs_rq */
4946 if (cfs_rq_throttled(qcfs_rq))
4947 goto unthrottle_throttle;
4950 * One parent has been throttled and cfs_rq removed from the
4951 * list. Add it back to not break the leaf list.
4953 if (throttled_hierarchy(qcfs_rq))
4954 list_add_leaf_cfs_rq(qcfs_rq);
4957 /* At this point se is NULL and we are at root level*/
4958 add_nr_running(rq, task_delta);
4960 unthrottle_throttle:
4962 * The cfs_rq_throttled() breaks in the above iteration can result in
4963 * incomplete leaf list maintenance, resulting in triggering the
4966 for_each_sched_entity(se) {
4967 struct cfs_rq *qcfs_rq = cfs_rq_of(se);
4969 if (list_add_leaf_cfs_rq(qcfs_rq))
4973 assert_list_leaf_cfs_rq(rq);
4975 /* Determine whether we need to wake up potentially idle CPU: */
4976 if (rq->curr == rq->idle && rq->cfs.nr_running)
4980 static void distribute_cfs_runtime(struct cfs_bandwidth *cfs_b)
4982 struct cfs_rq *cfs_rq;
4983 u64 runtime, remaining = 1;
4986 list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
4988 struct rq *rq = rq_of(cfs_rq);
4991 rq_lock_irqsave(rq, &rf);
4992 if (!cfs_rq_throttled(cfs_rq))
4995 /* By the above check, this should never be true */
4996 SCHED_WARN_ON(cfs_rq->runtime_remaining > 0);
4998 raw_spin_lock(&cfs_b->lock);
4999 runtime = -cfs_rq->runtime_remaining + 1;
5000 if (runtime > cfs_b->runtime)
5001 runtime = cfs_b->runtime;
5002 cfs_b->runtime -= runtime;
5003 remaining = cfs_b->runtime;
5004 raw_spin_unlock(&cfs_b->lock);
5006 cfs_rq->runtime_remaining += runtime;
5008 /* we check whether we're throttled above */
5009 if (cfs_rq->runtime_remaining > 0)
5010 unthrottle_cfs_rq(cfs_rq);
5013 rq_unlock_irqrestore(rq, &rf);
5022 * Responsible for refilling a task_group's bandwidth and unthrottling its
5023 * cfs_rqs as appropriate. If there has been no activity within the last
5024 * period the timer is deactivated until scheduling resumes; cfs_b->idle is
5025 * used to track this state.
5027 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun, unsigned long flags)
5031 /* no need to continue the timer with no bandwidth constraint */
5032 if (cfs_b->quota == RUNTIME_INF)
5033 goto out_deactivate;
5035 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
5036 cfs_b->nr_periods += overrun;
5038 /* Refill extra burst quota even if cfs_b->idle */
5039 __refill_cfs_bandwidth_runtime(cfs_b);
5042 * idle depends on !throttled (for the case of a large deficit), and if
5043 * we're going inactive then everything else can be deferred
5045 if (cfs_b->idle && !throttled)
5046 goto out_deactivate;
5049 /* mark as potentially idle for the upcoming period */
5054 /* account preceding periods in which throttling occurred */
5055 cfs_b->nr_throttled += overrun;
5058 * This check is repeated as we release cfs_b->lock while we unthrottle.
5060 while (throttled && cfs_b->runtime > 0) {
5061 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
5062 /* we can't nest cfs_b->lock while distributing bandwidth */
5063 distribute_cfs_runtime(cfs_b);
5064 raw_spin_lock_irqsave(&cfs_b->lock, flags);
5066 throttled = !list_empty(&cfs_b->throttled_cfs_rq);
5070 * While we are ensured activity in the period following an
5071 * unthrottle, this also covers the case in which the new bandwidth is
5072 * insufficient to cover the existing bandwidth deficit. (Forcing the
5073 * timer to remain active while there are any throttled entities.)
5083 /* a cfs_rq won't donate quota below this amount */
5084 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
5085 /* minimum remaining period time to redistribute slack quota */
5086 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
5087 /* how long we wait to gather additional slack before distributing */
5088 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
5091 * Are we near the end of the current quota period?
5093 * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
5094 * hrtimer base being cleared by hrtimer_start. In the case of
5095 * migrate_hrtimers, base is never cleared, so we are fine.
5097 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
5099 struct hrtimer *refresh_timer = &cfs_b->period_timer;
5102 /* if the call-back is running a quota refresh is already occurring */
5103 if (hrtimer_callback_running(refresh_timer))
5106 /* is a quota refresh about to occur? */
5107 remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
5108 if (remaining < min_expire)
5114 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
5116 u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
5118 /* if there's a quota refresh soon don't bother with slack */
5119 if (runtime_refresh_within(cfs_b, min_left))
5122 /* don't push forwards an existing deferred unthrottle */
5123 if (cfs_b->slack_started)
5125 cfs_b->slack_started = true;
5127 hrtimer_start(&cfs_b->slack_timer,
5128 ns_to_ktime(cfs_bandwidth_slack_period),
5132 /* we know any runtime found here is valid as update_curr() precedes return */
5133 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
5135 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
5136 s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
5138 if (slack_runtime <= 0)
5141 raw_spin_lock(&cfs_b->lock);
5142 if (cfs_b->quota != RUNTIME_INF) {
5143 cfs_b->runtime += slack_runtime;
5145 /* we are under rq->lock, defer unthrottling using a timer */
5146 if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
5147 !list_empty(&cfs_b->throttled_cfs_rq))
5148 start_cfs_slack_bandwidth(cfs_b);
5150 raw_spin_unlock(&cfs_b->lock);
5152 /* even if it's not valid for return we don't want to try again */
5153 cfs_rq->runtime_remaining -= slack_runtime;
5156 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
5158 if (!cfs_bandwidth_used())
5161 if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
5164 __return_cfs_rq_runtime(cfs_rq);
5168 * This is done with a timer (instead of inline with bandwidth return) since
5169 * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
5171 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
5173 u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
5174 unsigned long flags;
5176 /* confirm we're still not at a refresh boundary */
5177 raw_spin_lock_irqsave(&cfs_b->lock, flags);
5178 cfs_b->slack_started = false;
5180 if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
5181 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
5185 if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice)
5186 runtime = cfs_b->runtime;
5188 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
5193 distribute_cfs_runtime(cfs_b);
5197 * When a group wakes up we want to make sure that its quota is not already
5198 * expired/exceeded, otherwise it may be allowed to steal additional ticks of
5199 * runtime as update_curr() throttling can not trigger until it's on-rq.
5201 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
5203 if (!cfs_bandwidth_used())
5206 /* an active group must be handled by the update_curr()->put() path */
5207 if (!cfs_rq->runtime_enabled || cfs_rq->curr)
5210 /* ensure the group is not already throttled */
5211 if (cfs_rq_throttled(cfs_rq))
5214 /* update runtime allocation */
5215 account_cfs_rq_runtime(cfs_rq, 0);
5216 if (cfs_rq->runtime_remaining <= 0)
5217 throttle_cfs_rq(cfs_rq);
5220 static void sync_throttle(struct task_group *tg, int cpu)
5222 struct cfs_rq *pcfs_rq, *cfs_rq;
5224 if (!cfs_bandwidth_used())
5230 cfs_rq = tg->cfs_rq[cpu];
5231 pcfs_rq = tg->parent->cfs_rq[cpu];
5233 cfs_rq->throttle_count = pcfs_rq->throttle_count;
5234 cfs_rq->throttled_clock_task = rq_clock_task(cpu_rq(cpu));
5237 /* conditionally throttle active cfs_rq's from put_prev_entity() */
5238 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
5240 if (!cfs_bandwidth_used())
5243 if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
5247 * it's possible for a throttled entity to be forced into a running
5248 * state (e.g. set_curr_task), in this case we're finished.
5250 if (cfs_rq_throttled(cfs_rq))
5253 return throttle_cfs_rq(cfs_rq);
5256 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
5258 struct cfs_bandwidth *cfs_b =
5259 container_of(timer, struct cfs_bandwidth, slack_timer);
5261 do_sched_cfs_slack_timer(cfs_b);
5263 return HRTIMER_NORESTART;
5266 extern const u64 max_cfs_quota_period;
5268 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
5270 struct cfs_bandwidth *cfs_b =
5271 container_of(timer, struct cfs_bandwidth, period_timer);
5272 unsigned long flags;
5277 raw_spin_lock_irqsave(&cfs_b->lock, flags);
5279 overrun = hrtimer_forward_now(timer, cfs_b->period);
5283 idle = do_sched_cfs_period_timer(cfs_b, overrun, flags);
5286 u64 new, old = ktime_to_ns(cfs_b->period);
5289 * Grow period by a factor of 2 to avoid losing precision.
5290 * Precision loss in the quota/period ratio can cause __cfs_schedulable
5294 if (new < max_cfs_quota_period) {
5295 cfs_b->period = ns_to_ktime(new);
5299 pr_warn_ratelimited(
5300 "cfs_period_timer[cpu%d]: period too short, scaling up (new cfs_period_us = %lld, cfs_quota_us = %lld)\n",
5302 div_u64(new, NSEC_PER_USEC),
5303 div_u64(cfs_b->quota, NSEC_PER_USEC));
5305 pr_warn_ratelimited(
5306 "cfs_period_timer[cpu%d]: period too short, but cannot scale up without losing precision (cfs_period_us = %lld, cfs_quota_us = %lld)\n",
5308 div_u64(old, NSEC_PER_USEC),
5309 div_u64(cfs_b->quota, NSEC_PER_USEC));
5312 /* reset count so we don't come right back in here */
5317 cfs_b->period_active = 0;
5318 raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
5320 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
5323 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
5325 raw_spin_lock_init(&cfs_b->lock);
5327 cfs_b->quota = RUNTIME_INF;
5328 cfs_b->period = ns_to_ktime(default_cfs_period());
5331 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
5332 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED);
5333 cfs_b->period_timer.function = sched_cfs_period_timer;
5334 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
5335 cfs_b->slack_timer.function = sched_cfs_slack_timer;
5336 cfs_b->slack_started = false;
5339 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
5341 cfs_rq->runtime_enabled = 0;
5342 INIT_LIST_HEAD(&cfs_rq->throttled_list);
5345 void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
5347 lockdep_assert_held(&cfs_b->lock);
5349 if (cfs_b->period_active)
5352 cfs_b->period_active = 1;
5353 hrtimer_forward_now(&cfs_b->period_timer, cfs_b->period);
5354 hrtimer_start_expires(&cfs_b->period_timer, HRTIMER_MODE_ABS_PINNED);
5357 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
5359 /* init_cfs_bandwidth() was not called */
5360 if (!cfs_b->throttled_cfs_rq.next)
5363 hrtimer_cancel(&cfs_b->period_timer);
5364 hrtimer_cancel(&cfs_b->slack_timer);
5368 * Both these CPU hotplug callbacks race against unregister_fair_sched_group()
5370 * The race is harmless, since modifying bandwidth settings of unhooked group
5371 * bits doesn't do much.
5374 /* cpu online callback */
5375 static void __maybe_unused update_runtime_enabled(struct rq *rq)
5377 struct task_group *tg;
5379 lockdep_assert_rq_held(rq);
5382 list_for_each_entry_rcu(tg, &task_groups, list) {
5383 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
5384 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
5386 raw_spin_lock(&cfs_b->lock);
5387 cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF;
5388 raw_spin_unlock(&cfs_b->lock);
5393 /* cpu offline callback */
5394 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
5396 struct task_group *tg;
5398 lockdep_assert_rq_held(rq);
5401 list_for_each_entry_rcu(tg, &task_groups, list) {
5402 struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
5404 if (!cfs_rq->runtime_enabled)
5408 * clock_task is not advancing so we just need to make sure
5409 * there's some valid quota amount
5411 cfs_rq->runtime_remaining = 1;
5413 * Offline rq is schedulable till CPU is completely disabled
5414 * in take_cpu_down(), so we prevent new cfs throttling here.
5416 cfs_rq->runtime_enabled = 0;
5418 if (cfs_rq_throttled(cfs_rq))
5419 unthrottle_cfs_rq(cfs_rq);
5424 #else /* CONFIG_CFS_BANDWIDTH */
5426 static inline bool cfs_bandwidth_used(void)
5431 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {}
5432 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; }
5433 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
5434 static inline void sync_throttle(struct task_group *tg, int cpu) {}
5435 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
5437 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
5442 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
5447 static inline int throttled_lb_pair(struct task_group *tg,
5448 int src_cpu, int dest_cpu)
5453 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
5455 #ifdef CONFIG_FAIR_GROUP_SCHED
5456 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
5459 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
5463 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
5464 static inline void update_runtime_enabled(struct rq *rq) {}
5465 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
5467 #endif /* CONFIG_CFS_BANDWIDTH */
5469 /**************************************************
5470 * CFS operations on tasks:
5473 #ifdef CONFIG_SCHED_HRTICK
5474 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
5476 struct sched_entity *se = &p->se;
5477 struct cfs_rq *cfs_rq = cfs_rq_of(se);
5479 SCHED_WARN_ON(task_rq(p) != rq);
5481 if (rq->cfs.h_nr_running > 1) {
5482 u64 slice = sched_slice(cfs_rq, se);
5483 u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
5484 s64 delta = slice - ran;
5487 if (task_current(rq, p))
5491 hrtick_start(rq, delta);
5496 * called from enqueue/dequeue and updates the hrtick when the
5497 * current task is from our class and nr_running is low enough
5500 static void hrtick_update(struct rq *rq)
5502 struct task_struct *curr = rq->curr;
5504 if (!hrtick_enabled_fair(rq) || curr->sched_class != &fair_sched_class)
5507 if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
5508 hrtick_start_fair(rq, curr);
5510 #else /* !CONFIG_SCHED_HRTICK */
5512 hrtick_start_fair(struct rq *rq, struct task_struct *p)
5516 static inline void hrtick_update(struct rq *rq)
5522 static inline unsigned long cpu_util(int cpu);
5524 static inline bool cpu_overutilized(int cpu)
5526 return !fits_capacity(cpu_util(cpu), capacity_of(cpu));
5529 static inline void update_overutilized_status(struct rq *rq)
5531 if (!READ_ONCE(rq->rd->overutilized) && cpu_overutilized(rq->cpu)) {
5532 WRITE_ONCE(rq->rd->overutilized, SG_OVERUTILIZED);
5533 trace_sched_overutilized_tp(rq->rd, SG_OVERUTILIZED);
5537 static inline void update_overutilized_status(struct rq *rq) { }
5540 /* Runqueue only has SCHED_IDLE tasks enqueued */
5541 static int sched_idle_rq(struct rq *rq)
5543 return unlikely(rq->nr_running == rq->cfs.idle_h_nr_running &&
5548 static int sched_idle_cpu(int cpu)
5550 return sched_idle_rq(cpu_rq(cpu));
5555 * The enqueue_task method is called before nr_running is
5556 * increased. Here we update the fair scheduling stats and
5557 * then put the task into the rbtree:
5560 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
5562 struct cfs_rq *cfs_rq;
5563 struct sched_entity *se = &p->se;
5564 int idle_h_nr_running = task_has_idle_policy(p);
5565 int task_new = !(flags & ENQUEUE_WAKEUP);
5568 * The code below (indirectly) updates schedutil which looks at
5569 * the cfs_rq utilization to select a frequency.
5570 * Let's add the task's estimated utilization to the cfs_rq's
5571 * estimated utilization, before we update schedutil.
5573 util_est_enqueue(&rq->cfs, p);
5576 * If in_iowait is set, the code below may not trigger any cpufreq
5577 * utilization updates, so do it here explicitly with the IOWAIT flag
5581 cpufreq_update_util(rq, SCHED_CPUFREQ_IOWAIT);
5583 for_each_sched_entity(se) {
5586 cfs_rq = cfs_rq_of(se);
5587 enqueue_entity(cfs_rq, se, flags);
5589 cfs_rq->h_nr_running++;
5590 cfs_rq->idle_h_nr_running += idle_h_nr_running;
5592 if (cfs_rq_is_idle(cfs_rq))
5593 idle_h_nr_running = 1;
5595 /* end evaluation on encountering a throttled cfs_rq */
5596 if (cfs_rq_throttled(cfs_rq))
5597 goto enqueue_throttle;
5599 flags = ENQUEUE_WAKEUP;
5602 for_each_sched_entity(se) {
5603 cfs_rq = cfs_rq_of(se);
5605 update_load_avg(cfs_rq, se, UPDATE_TG);
5606 se_update_runnable(se);
5607 update_cfs_group(se);
5609 cfs_rq->h_nr_running++;
5610 cfs_rq->idle_h_nr_running += idle_h_nr_running;
5612 if (cfs_rq_is_idle(cfs_rq))
5613 idle_h_nr_running = 1;
5615 /* end evaluation on encountering a throttled cfs_rq */
5616 if (cfs_rq_throttled(cfs_rq))
5617 goto enqueue_throttle;
5620 * One parent has been throttled and cfs_rq removed from the
5621 * list. Add it back to not break the leaf list.
5623 if (throttled_hierarchy(cfs_rq))
5624 list_add_leaf_cfs_rq(cfs_rq);
5627 /* At this point se is NULL and we are at root level*/
5628 add_nr_running(rq, 1);
5631 * Since new tasks are assigned an initial util_avg equal to
5632 * half of the spare capacity of their CPU, tiny tasks have the
5633 * ability to cross the overutilized threshold, which will
5634 * result in the load balancer ruining all the task placement
5635 * done by EAS. As a way to mitigate that effect, do not account
5636 * for the first enqueue operation of new tasks during the
5637 * overutilized flag detection.
5639 * A better way of solving this problem would be to wait for
5640 * the PELT signals of tasks to converge before taking them
5641 * into account, but that is not straightforward to implement,
5642 * and the following generally works well enough in practice.
5645 update_overutilized_status(rq);
5648 if (cfs_bandwidth_used()) {
5650 * When bandwidth control is enabled; the cfs_rq_throttled()
5651 * breaks in the above iteration can result in incomplete
5652 * leaf list maintenance, resulting in triggering the assertion
5655 for_each_sched_entity(se) {
5656 cfs_rq = cfs_rq_of(se);
5658 if (list_add_leaf_cfs_rq(cfs_rq))
5663 assert_list_leaf_cfs_rq(rq);
5668 static void set_next_buddy(struct sched_entity *se);
5671 * The dequeue_task method is called before nr_running is
5672 * decreased. We remove the task from the rbtree and
5673 * update the fair scheduling stats:
5675 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
5677 struct cfs_rq *cfs_rq;
5678 struct sched_entity *se = &p->se;
5679 int task_sleep = flags & DEQUEUE_SLEEP;
5680 int idle_h_nr_running = task_has_idle_policy(p);
5681 bool was_sched_idle = sched_idle_rq(rq);
5683 util_est_dequeue(&rq->cfs, p);
5685 for_each_sched_entity(se) {
5686 cfs_rq = cfs_rq_of(se);
5687 dequeue_entity(cfs_rq, se, flags);
5689 cfs_rq->h_nr_running--;
5690 cfs_rq->idle_h_nr_running -= idle_h_nr_running;
5692 if (cfs_rq_is_idle(cfs_rq))
5693 idle_h_nr_running = 1;
5695 /* end evaluation on encountering a throttled cfs_rq */
5696 if (cfs_rq_throttled(cfs_rq))
5697 goto dequeue_throttle;
5699 /* Don't dequeue parent if it has other entities besides us */
5700 if (cfs_rq->load.weight) {
5701 /* Avoid re-evaluating load for this entity: */
5702 se = parent_entity(se);
5704 * Bias pick_next to pick a task from this cfs_rq, as
5705 * p is sleeping when it is within its sched_slice.
5707 if (task_sleep && se && !throttled_hierarchy(cfs_rq))
5711 flags |= DEQUEUE_SLEEP;
5714 for_each_sched_entity(se) {
5715 cfs_rq = cfs_rq_of(se);
5717 update_load_avg(cfs_rq, se, UPDATE_TG);
5718 se_update_runnable(se);
5719 update_cfs_group(se);
5721 cfs_rq->h_nr_running--;
5722 cfs_rq->idle_h_nr_running -= idle_h_nr_running;
5724 if (cfs_rq_is_idle(cfs_rq))
5725 idle_h_nr_running = 1;
5727 /* end evaluation on encountering a throttled cfs_rq */
5728 if (cfs_rq_throttled(cfs_rq))
5729 goto dequeue_throttle;
5733 /* At this point se is NULL and we are at root level*/
5734 sub_nr_running(rq, 1);
5736 /* balance early to pull high priority tasks */
5737 if (unlikely(!was_sched_idle && sched_idle_rq(rq)))
5738 rq->next_balance = jiffies;
5741 util_est_update(&rq->cfs, p, task_sleep);
5747 /* Working cpumask for: load_balance, load_balance_newidle. */
5748 DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
5749 DEFINE_PER_CPU(cpumask_var_t, select_idle_mask);
5751 #ifdef CONFIG_NO_HZ_COMMON
5754 cpumask_var_t idle_cpus_mask;
5756 int has_blocked; /* Idle CPUS has blocked load */
5757 unsigned long next_balance; /* in jiffy units */
5758 unsigned long next_blocked; /* Next update of blocked load in jiffies */
5759 } nohz ____cacheline_aligned;
5761 #endif /* CONFIG_NO_HZ_COMMON */
5763 static unsigned long cpu_load(struct rq *rq)
5765 return cfs_rq_load_avg(&rq->cfs);
5769 * cpu_load_without - compute CPU load without any contributions from *p
5770 * @cpu: the CPU which load is requested
5771 * @p: the task which load should be discounted
5773 * The load of a CPU is defined by the load of tasks currently enqueued on that
5774 * CPU as well as tasks which are currently sleeping after an execution on that
5777 * This method returns the load of the specified CPU by discounting the load of
5778 * the specified task, whenever the task is currently contributing to the CPU
5781 static unsigned long cpu_load_without(struct rq *rq, struct task_struct *p)
5783 struct cfs_rq *cfs_rq;
5786 /* Task has no contribution or is new */
5787 if (cpu_of(rq) != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time))
5788 return cpu_load(rq);
5791 load = READ_ONCE(cfs_rq->avg.load_avg);
5793 /* Discount task's util from CPU's util */
5794 lsub_positive(&load, task_h_load(p));
5799 static unsigned long cpu_runnable(struct rq *rq)
5801 return cfs_rq_runnable_avg(&rq->cfs);
5804 static unsigned long cpu_runnable_without(struct rq *rq, struct task_struct *p)
5806 struct cfs_rq *cfs_rq;
5807 unsigned int runnable;
5809 /* Task has no contribution or is new */
5810 if (cpu_of(rq) != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time))
5811 return cpu_runnable(rq);
5814 runnable = READ_ONCE(cfs_rq->avg.runnable_avg);
5816 /* Discount task's runnable from CPU's runnable */
5817 lsub_positive(&runnable, p->se.avg.runnable_avg);
5822 static unsigned long capacity_of(int cpu)
5824 return cpu_rq(cpu)->cpu_capacity;
5827 static void record_wakee(struct task_struct *p)
5830 * Only decay a single time; tasks that have less then 1 wakeup per
5831 * jiffy will not have built up many flips.
5833 if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) {
5834 current->wakee_flips >>= 1;
5835 current->wakee_flip_decay_ts = jiffies;
5838 if (current->last_wakee != p) {
5839 current->last_wakee = p;
5840 current->wakee_flips++;
5845 * Detect M:N waker/wakee relationships via a switching-frequency heuristic.
5847 * A waker of many should wake a different task than the one last awakened
5848 * at a frequency roughly N times higher than one of its wakees.
5850 * In order to determine whether we should let the load spread vs consolidating
5851 * to shared cache, we look for a minimum 'flip' frequency of llc_size in one
5852 * partner, and a factor of lls_size higher frequency in the other.
5854 * With both conditions met, we can be relatively sure that the relationship is
5855 * non-monogamous, with partner count exceeding socket size.
5857 * Waker/wakee being client/server, worker/dispatcher, interrupt source or
5858 * whatever is irrelevant, spread criteria is apparent partner count exceeds
5861 static int wake_wide(struct task_struct *p)
5863 unsigned int master = current->wakee_flips;
5864 unsigned int slave = p->wakee_flips;
5865 int factor = __this_cpu_read(sd_llc_size);
5868 swap(master, slave);
5869 if (slave < factor || master < slave * factor)
5875 * The purpose of wake_affine() is to quickly determine on which CPU we can run
5876 * soonest. For the purpose of speed we only consider the waking and previous
5879 * wake_affine_idle() - only considers 'now', it check if the waking CPU is
5880 * cache-affine and is (or will be) idle.
5882 * wake_affine_weight() - considers the weight to reflect the average
5883 * scheduling latency of the CPUs. This seems to work
5884 * for the overloaded case.
5887 wake_affine_idle(int this_cpu, int prev_cpu, int sync)
5890 * If this_cpu is idle, it implies the wakeup is from interrupt
5891 * context. Only allow the move if cache is shared. Otherwise an
5892 * interrupt intensive workload could force all tasks onto one
5893 * node depending on the IO topology or IRQ affinity settings.
5895 * If the prev_cpu is idle and cache affine then avoid a migration.
5896 * There is no guarantee that the cache hot data from an interrupt
5897 * is more important than cache hot data on the prev_cpu and from
5898 * a cpufreq perspective, it's better to have higher utilisation
5901 if (available_idle_cpu(this_cpu) && cpus_share_cache(this_cpu, prev_cpu))
5902 return available_idle_cpu(prev_cpu) ? prev_cpu : this_cpu;
5904 if (sync && cpu_rq(this_cpu)->nr_running == 1)
5907 if (available_idle_cpu(prev_cpu))
5910 return nr_cpumask_bits;
5914 wake_affine_weight(struct sched_domain *sd, struct task_struct *p,
5915 int this_cpu, int prev_cpu, int sync)
5917 s64 this_eff_load, prev_eff_load;
5918 unsigned long task_load;
5920 this_eff_load = cpu_load(cpu_rq(this_cpu));
5923 unsigned long current_load = task_h_load(current);
5925 if (current_load > this_eff_load)
5928 this_eff_load -= current_load;
5931 task_load = task_h_load(p);
5933 this_eff_load += task_load;
5934 if (sched_feat(WA_BIAS))
5935 this_eff_load *= 100;
5936 this_eff_load *= capacity_of(prev_cpu);
5938 prev_eff_load = cpu_load(cpu_rq(prev_cpu));
5939 prev_eff_load -= task_load;
5940 if (sched_feat(WA_BIAS))
5941 prev_eff_load *= 100 + (sd->imbalance_pct - 100) / 2;
5942 prev_eff_load *= capacity_of(this_cpu);
5945 * If sync, adjust the weight of prev_eff_load such that if
5946 * prev_eff == this_eff that select_idle_sibling() will consider
5947 * stacking the wakee on top of the waker if no other CPU is
5953 return this_eff_load < prev_eff_load ? this_cpu : nr_cpumask_bits;
5956 static int wake_affine(struct sched_domain *sd, struct task_struct *p,
5957 int this_cpu, int prev_cpu, int sync)
5959 int target = nr_cpumask_bits;
5961 if (sched_feat(WA_IDLE))
5962 target = wake_affine_idle(this_cpu, prev_cpu, sync);
5964 if (sched_feat(WA_WEIGHT) && target == nr_cpumask_bits)
5965 target = wake_affine_weight(sd, p, this_cpu, prev_cpu, sync);
5967 schedstat_inc(p->se.statistics.nr_wakeups_affine_attempts);
5968 if (target == nr_cpumask_bits)
5971 schedstat_inc(sd->ttwu_move_affine);
5972 schedstat_inc(p->se.statistics.nr_wakeups_affine);
5976 static struct sched_group *
5977 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu);
5980 * find_idlest_group_cpu - find the idlest CPU among the CPUs in the group.
5983 find_idlest_group_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
5985 unsigned long load, min_load = ULONG_MAX;
5986 unsigned int min_exit_latency = UINT_MAX;
5987 u64 latest_idle_timestamp = 0;
5988 int least_loaded_cpu = this_cpu;
5989 int shallowest_idle_cpu = -1;
5992 /* Check if we have any choice: */
5993 if (group->group_weight == 1)
5994 return cpumask_first(sched_group_span(group));
5996 /* Traverse only the allowed CPUs */
5997 for_each_cpu_and(i, sched_group_span(group), p->cpus_ptr) {
5998 struct rq *rq = cpu_rq(i);
6000 if (!sched_core_cookie_match(rq, p))
6003 if (sched_idle_cpu(i))
6006 if (available_idle_cpu(i)) {
6007 struct cpuidle_state *idle = idle_get_state(rq);
6008 if (idle && idle->exit_latency < min_exit_latency) {
6010 * We give priority to a CPU whose idle state
6011 * has the smallest exit latency irrespective
6012 * of any idle timestamp.
6014 min_exit_latency = idle->exit_latency;
6015 latest_idle_timestamp = rq->idle_stamp;
6016 shallowest_idle_cpu = i;
6017 } else if ((!idle || idle->exit_latency == min_exit_latency) &&
6018 rq->idle_stamp > latest_idle_timestamp) {
6020 * If equal or no active idle state, then
6021 * the most recently idled CPU might have
6024 latest_idle_timestamp = rq->idle_stamp;
6025 shallowest_idle_cpu = i;
6027 } else if (shallowest_idle_cpu == -1) {
6028 load = cpu_load(cpu_rq(i));
6029 if (load < min_load) {
6031 least_loaded_cpu = i;
6036 return shallowest_idle_cpu != -1 ? shallowest_idle_cpu : least_loaded_cpu;
6039 static inline int find_idlest_cpu(struct sched_domain *sd, struct task_struct *p,
6040 int cpu, int prev_cpu, int sd_flag)
6044 if (!cpumask_intersects(sched_domain_span(sd), p->cpus_ptr))
6048 * We need task's util for cpu_util_without, sync it up to
6049 * prev_cpu's last_update_time.
6051 if (!(sd_flag & SD_BALANCE_FORK))
6052 sync_entity_load_avg(&p->se);
6055 struct sched_group *group;
6056 struct sched_domain *tmp;
6059 if (!(sd->flags & sd_flag)) {
6064 group = find_idlest_group(sd, p, cpu);
6070 new_cpu = find_idlest_group_cpu(group, p, cpu);
6071 if (new_cpu == cpu) {
6072 /* Now try balancing at a lower domain level of 'cpu': */
6077 /* Now try balancing at a lower domain level of 'new_cpu': */
6079 weight = sd->span_weight;
6081 for_each_domain(cpu, tmp) {
6082 if (weight <= tmp->span_weight)
6084 if (tmp->flags & sd_flag)
6092 static inline int __select_idle_cpu(int cpu, struct task_struct *p)
6094 if ((available_idle_cpu(cpu) || sched_idle_cpu(cpu)) &&
6095 sched_cpu_cookie_match(cpu_rq(cpu), p))
6101 #ifdef CONFIG_SCHED_SMT
6102 DEFINE_STATIC_KEY_FALSE(sched_smt_present);
6103 EXPORT_SYMBOL_GPL(sched_smt_present);
6105 static inline void set_idle_cores(int cpu, int val)
6107 struct sched_domain_shared *sds;
6109 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
6111 WRITE_ONCE(sds->has_idle_cores, val);
6114 static inline bool test_idle_cores(int cpu, bool def)
6116 struct sched_domain_shared *sds;
6118 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
6120 return READ_ONCE(sds->has_idle_cores);
6126 * Scans the local SMT mask to see if the entire core is idle, and records this
6127 * information in sd_llc_shared->has_idle_cores.
6129 * Since SMT siblings share all cache levels, inspecting this limited remote
6130 * state should be fairly cheap.
6132 void __update_idle_core(struct rq *rq)
6134 int core = cpu_of(rq);
6138 if (test_idle_cores(core, true))
6141 for_each_cpu(cpu, cpu_smt_mask(core)) {
6145 if (!available_idle_cpu(cpu))
6149 set_idle_cores(core, 1);
6155 * Scan the entire LLC domain for idle cores; this dynamically switches off if
6156 * there are no idle cores left in the system; tracked through
6157 * sd_llc->shared->has_idle_cores and enabled through update_idle_core() above.
6159 static int select_idle_core(struct task_struct *p, int core, struct cpumask *cpus, int *idle_cpu)
6164 if (!static_branch_likely(&sched_smt_present))
6165 return __select_idle_cpu(core, p);
6167 for_each_cpu(cpu, cpu_smt_mask(core)) {
6168 if (!available_idle_cpu(cpu)) {
6170 if (*idle_cpu == -1) {
6171 if (sched_idle_cpu(cpu) && cpumask_test_cpu(cpu, p->cpus_ptr)) {
6179 if (*idle_cpu == -1 && cpumask_test_cpu(cpu, p->cpus_ptr))
6186 cpumask_andnot(cpus, cpus, cpu_smt_mask(core));
6191 * Scan the local SMT mask for idle CPUs.
6193 static int select_idle_smt(struct task_struct *p, struct sched_domain *sd, int target)
6197 for_each_cpu(cpu, cpu_smt_mask(target)) {
6198 if (!cpumask_test_cpu(cpu, p->cpus_ptr) ||
6199 !cpumask_test_cpu(cpu, sched_domain_span(sd)))
6201 if (available_idle_cpu(cpu) || sched_idle_cpu(cpu))
6208 #else /* CONFIG_SCHED_SMT */
6210 static inline void set_idle_cores(int cpu, int val)
6214 static inline bool test_idle_cores(int cpu, bool def)
6219 static inline int select_idle_core(struct task_struct *p, int core, struct cpumask *cpus, int *idle_cpu)
6221 return __select_idle_cpu(core, p);
6224 static inline int select_idle_smt(struct task_struct *p, struct sched_domain *sd, int target)
6229 #endif /* CONFIG_SCHED_SMT */
6232 * Scan the LLC domain for idle CPUs; this is dynamically regulated by
6233 * comparing the average scan cost (tracked in sd->avg_scan_cost) against the
6234 * average idle time for this rq (as found in rq->avg_idle).
6236 static int select_idle_cpu(struct task_struct *p, struct sched_domain *sd, bool has_idle_core, int target)
6238 struct cpumask *cpus = this_cpu_cpumask_var_ptr(select_idle_mask);
6239 int i, cpu, idle_cpu = -1, nr = INT_MAX;
6240 struct rq *this_rq = this_rq();
6241 int this = smp_processor_id();
6242 struct sched_domain *this_sd;
6245 this_sd = rcu_dereference(*this_cpu_ptr(&sd_llc));
6249 cpumask_and(cpus, sched_domain_span(sd), p->cpus_ptr);
6251 if (sched_feat(SIS_PROP) && !has_idle_core) {
6252 u64 avg_cost, avg_idle, span_avg;
6253 unsigned long now = jiffies;
6256 * If we're busy, the assumption that the last idle period
6257 * predicts the future is flawed; age away the remaining
6258 * predicted idle time.
6260 if (unlikely(this_rq->wake_stamp < now)) {
6261 while (this_rq->wake_stamp < now && this_rq->wake_avg_idle) {
6262 this_rq->wake_stamp++;
6263 this_rq->wake_avg_idle >>= 1;
6267 avg_idle = this_rq->wake_avg_idle;
6268 avg_cost = this_sd->avg_scan_cost + 1;
6270 span_avg = sd->span_weight * avg_idle;
6271 if (span_avg > 4*avg_cost)
6272 nr = div_u64(span_avg, avg_cost);
6276 time = cpu_clock(this);
6279 for_each_cpu_wrap(cpu, cpus, target + 1) {
6280 if (has_idle_core) {
6281 i = select_idle_core(p, cpu, cpus, &idle_cpu);
6282 if ((unsigned int)i < nr_cpumask_bits)
6288 idle_cpu = __select_idle_cpu(cpu, p);
6289 if ((unsigned int)idle_cpu < nr_cpumask_bits)
6295 set_idle_cores(target, false);
6297 if (sched_feat(SIS_PROP) && !has_idle_core) {
6298 time = cpu_clock(this) - time;
6301 * Account for the scan cost of wakeups against the average
6304 this_rq->wake_avg_idle -= min(this_rq->wake_avg_idle, time);
6306 update_avg(&this_sd->avg_scan_cost, time);
6313 * Scan the asym_capacity domain for idle CPUs; pick the first idle one on which
6314 * the task fits. If no CPU is big enough, but there are idle ones, try to
6315 * maximize capacity.
6318 select_idle_capacity(struct task_struct *p, struct sched_domain *sd, int target)
6320 unsigned long task_util, best_cap = 0;
6321 int cpu, best_cpu = -1;
6322 struct cpumask *cpus;
6324 cpus = this_cpu_cpumask_var_ptr(select_idle_mask);
6325 cpumask_and(cpus, sched_domain_span(sd), p->cpus_ptr);
6327 task_util = uclamp_task_util(p);
6329 for_each_cpu_wrap(cpu, cpus, target) {
6330 unsigned long cpu_cap = capacity_of(cpu);
6332 if (!available_idle_cpu(cpu) && !sched_idle_cpu(cpu))
6334 if (fits_capacity(task_util, cpu_cap))
6337 if (cpu_cap > best_cap) {
6346 static inline bool asym_fits_capacity(int task_util, int cpu)
6348 if (static_branch_unlikely(&sched_asym_cpucapacity))
6349 return fits_capacity(task_util, capacity_of(cpu));
6355 * Try and locate an idle core/thread in the LLC cache domain.
6357 static int select_idle_sibling(struct task_struct *p, int prev, int target)
6359 bool has_idle_core = false;
6360 struct sched_domain *sd;
6361 unsigned long task_util;
6362 int i, recent_used_cpu;
6365 * On asymmetric system, update task utilization because we will check
6366 * that the task fits with cpu's capacity.
6368 if (static_branch_unlikely(&sched_asym_cpucapacity)) {
6369 sync_entity_load_avg(&p->se);
6370 task_util = uclamp_task_util(p);
6374 * per-cpu select_idle_mask usage
6376 lockdep_assert_irqs_disabled();
6378 if ((available_idle_cpu(target) || sched_idle_cpu(target)) &&
6379 asym_fits_capacity(task_util, target))
6383 * If the previous CPU is cache affine and idle, don't be stupid:
6385 if (prev != target && cpus_share_cache(prev, target) &&
6386 (available_idle_cpu(prev) || sched_idle_cpu(prev)) &&
6387 asym_fits_capacity(task_util, prev))
6391 * Allow a per-cpu kthread to stack with the wakee if the
6392 * kworker thread and the tasks previous CPUs are the same.
6393 * The assumption is that the wakee queued work for the
6394 * per-cpu kthread that is now complete and the wakeup is
6395 * essentially a sync wakeup. An obvious example of this
6396 * pattern is IO completions.
6398 if (is_per_cpu_kthread(current) &&
6399 prev == smp_processor_id() &&
6400 this_rq()->nr_running <= 1) {
6404 /* Check a recently used CPU as a potential idle candidate: */
6405 recent_used_cpu = p->recent_used_cpu;
6406 p->recent_used_cpu = prev;
6407 if (recent_used_cpu != prev &&
6408 recent_used_cpu != target &&
6409 cpus_share_cache(recent_used_cpu, target) &&
6410 (available_idle_cpu(recent_used_cpu) || sched_idle_cpu(recent_used_cpu)) &&
6411 cpumask_test_cpu(p->recent_used_cpu, p->cpus_ptr) &&
6412 asym_fits_capacity(task_util, recent_used_cpu)) {
6414 * Replace recent_used_cpu with prev as it is a potential
6415 * candidate for the next wake:
6417 p->recent_used_cpu = prev;
6418 return recent_used_cpu;
6422 * For asymmetric CPU capacity systems, our domain of interest is
6423 * sd_asym_cpucapacity rather than sd_llc.
6425 if (static_branch_unlikely(&sched_asym_cpucapacity)) {
6426 sd = rcu_dereference(per_cpu(sd_asym_cpucapacity, target));
6428 * On an asymmetric CPU capacity system where an exclusive
6429 * cpuset defines a symmetric island (i.e. one unique
6430 * capacity_orig value through the cpuset), the key will be set
6431 * but the CPUs within that cpuset will not have a domain with
6432 * SD_ASYM_CPUCAPACITY. These should follow the usual symmetric
6436 i = select_idle_capacity(p, sd, target);
6437 return ((unsigned)i < nr_cpumask_bits) ? i : target;
6441 sd = rcu_dereference(per_cpu(sd_llc, target));
6445 if (sched_smt_active()) {
6446 has_idle_core = test_idle_cores(target, false);
6448 if (!has_idle_core && cpus_share_cache(prev, target)) {
6449 i = select_idle_smt(p, sd, prev);
6450 if ((unsigned int)i < nr_cpumask_bits)
6455 i = select_idle_cpu(p, sd, has_idle_core, target);
6456 if ((unsigned)i < nr_cpumask_bits)
6463 * cpu_util - Estimates the amount of capacity of a CPU used by CFS tasks.
6464 * @cpu: the CPU to get the utilization of
6466 * The unit of the return value must be the one of capacity so we can compare
6467 * the utilization with the capacity of the CPU that is available for CFS task
6468 * (ie cpu_capacity).
6470 * cfs_rq.avg.util_avg is the sum of running time of runnable tasks plus the
6471 * recent utilization of currently non-runnable tasks on a CPU. It represents
6472 * the amount of utilization of a CPU in the range [0..capacity_orig] where
6473 * capacity_orig is the cpu_capacity available at the highest frequency
6474 * (arch_scale_freq_capacity()).
6475 * The utilization of a CPU converges towards a sum equal to or less than the
6476 * current capacity (capacity_curr <= capacity_orig) of the CPU because it is
6477 * the running time on this CPU scaled by capacity_curr.
6479 * The estimated utilization of a CPU is defined to be the maximum between its
6480 * cfs_rq.avg.util_avg and the sum of the estimated utilization of the tasks
6481 * currently RUNNABLE on that CPU.
6482 * This allows to properly represent the expected utilization of a CPU which
6483 * has just got a big task running since a long sleep period. At the same time
6484 * however it preserves the benefits of the "blocked utilization" in
6485 * describing the potential for other tasks waking up on the same CPU.
6487 * Nevertheless, cfs_rq.avg.util_avg can be higher than capacity_curr or even
6488 * higher than capacity_orig because of unfortunate rounding in
6489 * cfs.avg.util_avg or just after migrating tasks and new task wakeups until
6490 * the average stabilizes with the new running time. We need to check that the
6491 * utilization stays within the range of [0..capacity_orig] and cap it if
6492 * necessary. Without utilization capping, a group could be seen as overloaded
6493 * (CPU0 utilization at 121% + CPU1 utilization at 80%) whereas CPU1 has 20% of
6494 * available capacity. We allow utilization to overshoot capacity_curr (but not
6495 * capacity_orig) as it useful for predicting the capacity required after task
6496 * migrations (scheduler-driven DVFS).
6498 * Return: the (estimated) utilization for the specified CPU
6500 static inline unsigned long cpu_util(int cpu)
6502 struct cfs_rq *cfs_rq;
6505 cfs_rq = &cpu_rq(cpu)->cfs;
6506 util = READ_ONCE(cfs_rq->avg.util_avg);
6508 if (sched_feat(UTIL_EST))
6509 util = max(util, READ_ONCE(cfs_rq->avg.util_est.enqueued));
6511 return min_t(unsigned long, util, capacity_orig_of(cpu));
6515 * cpu_util_without: compute cpu utilization without any contributions from *p
6516 * @cpu: the CPU which utilization is requested
6517 * @p: the task which utilization should be discounted
6519 * The utilization of a CPU is defined by the utilization of tasks currently
6520 * enqueued on that CPU as well as tasks which are currently sleeping after an
6521 * execution on that CPU.
6523 * This method returns the utilization of the specified CPU by discounting the
6524 * utilization of the specified task, whenever the task is currently
6525 * contributing to the CPU utilization.
6527 static unsigned long cpu_util_without(int cpu, struct task_struct *p)
6529 struct cfs_rq *cfs_rq;
6532 /* Task has no contribution or is new */
6533 if (cpu != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time))
6534 return cpu_util(cpu);
6536 cfs_rq = &cpu_rq(cpu)->cfs;
6537 util = READ_ONCE(cfs_rq->avg.util_avg);
6539 /* Discount task's util from CPU's util */
6540 lsub_positive(&util, task_util(p));
6545 * a) if *p is the only task sleeping on this CPU, then:
6546 * cpu_util (== task_util) > util_est (== 0)
6547 * and thus we return:
6548 * cpu_util_without = (cpu_util - task_util) = 0
6550 * b) if other tasks are SLEEPING on this CPU, which is now exiting
6552 * cpu_util >= task_util
6553 * cpu_util > util_est (== 0)
6554 * and thus we discount *p's blocked utilization to return:
6555 * cpu_util_without = (cpu_util - task_util) >= 0
6557 * c) if other tasks are RUNNABLE on that CPU and
6558 * util_est > cpu_util
6559 * then we use util_est since it returns a more restrictive
6560 * estimation of the spare capacity on that CPU, by just
6561 * considering the expected utilization of tasks already
6562 * runnable on that CPU.
6564 * Cases a) and b) are covered by the above code, while case c) is
6565 * covered by the following code when estimated utilization is
6568 if (sched_feat(UTIL_EST)) {
6569 unsigned int estimated =
6570 READ_ONCE(cfs_rq->avg.util_est.enqueued);
6573 * Despite the following checks we still have a small window
6574 * for a possible race, when an execl's select_task_rq_fair()
6575 * races with LB's detach_task():
6578 * p->on_rq = TASK_ON_RQ_MIGRATING;
6579 * ---------------------------------- A
6580 * deactivate_task() \
6581 * dequeue_task() + RaceTime
6582 * util_est_dequeue() /
6583 * ---------------------------------- B
6585 * The additional check on "current == p" it's required to
6586 * properly fix the execl regression and it helps in further
6587 * reducing the chances for the above race.
6589 if (unlikely(task_on_rq_queued(p) || current == p))
6590 lsub_positive(&estimated, _task_util_est(p));
6592 util = max(util, estimated);
6596 * Utilization (estimated) can exceed the CPU capacity, thus let's
6597 * clamp to the maximum CPU capacity to ensure consistency with
6598 * the cpu_util call.
6600 return min_t(unsigned long, util, capacity_orig_of(cpu));
6604 * Predicts what cpu_util(@cpu) would return if @p was migrated (and enqueued)
6607 static unsigned long cpu_util_next(int cpu, struct task_struct *p, int dst_cpu)
6609 struct cfs_rq *cfs_rq = &cpu_rq(cpu)->cfs;
6610 unsigned long util_est, util = READ_ONCE(cfs_rq->avg.util_avg);
6613 * If @p migrates from @cpu to another, remove its contribution. Or,
6614 * if @p migrates from another CPU to @cpu, add its contribution. In
6615 * the other cases, @cpu is not impacted by the migration, so the
6616 * util_avg should already be correct.
6618 if (task_cpu(p) == cpu && dst_cpu != cpu)
6619 lsub_positive(&util, task_util(p));
6620 else if (task_cpu(p) != cpu && dst_cpu == cpu)
6621 util += task_util(p);
6623 if (sched_feat(UTIL_EST)) {
6624 util_est = READ_ONCE(cfs_rq->avg.util_est.enqueued);
6627 * During wake-up, the task isn't enqueued yet and doesn't
6628 * appear in the cfs_rq->avg.util_est.enqueued of any rq,
6629 * so just add it (if needed) to "simulate" what will be
6630 * cpu_util() after the task has been enqueued.
6633 util_est += _task_util_est(p);
6635 util = max(util, util_est);
6638 return min(util, capacity_orig_of(cpu));
6642 * compute_energy(): Estimates the energy that @pd would consume if @p was
6643 * migrated to @dst_cpu. compute_energy() predicts what will be the utilization
6644 * landscape of @pd's CPUs after the task migration, and uses the Energy Model
6645 * to compute what would be the energy if we decided to actually migrate that
6649 compute_energy(struct task_struct *p, int dst_cpu, struct perf_domain *pd)
6651 struct cpumask *pd_mask = perf_domain_span(pd);
6652 unsigned long cpu_cap = arch_scale_cpu_capacity(cpumask_first(pd_mask));
6653 unsigned long max_util = 0, sum_util = 0;
6654 unsigned long _cpu_cap = cpu_cap;
6657 _cpu_cap -= arch_scale_thermal_pressure(cpumask_first(pd_mask));
6660 * The capacity state of CPUs of the current rd can be driven by CPUs
6661 * of another rd if they belong to the same pd. So, account for the
6662 * utilization of these CPUs too by masking pd with cpu_online_mask
6663 * instead of the rd span.
6665 * If an entire pd is outside of the current rd, it will not appear in
6666 * its pd list and will not be accounted by compute_energy().
6668 for_each_cpu_and(cpu, pd_mask, cpu_online_mask) {
6669 unsigned long util_freq = cpu_util_next(cpu, p, dst_cpu);
6670 unsigned long cpu_util, util_running = util_freq;
6671 struct task_struct *tsk = NULL;
6674 * When @p is placed on @cpu:
6676 * util_running = max(cpu_util, cpu_util_est) +
6677 * max(task_util, _task_util_est)
6679 * while cpu_util_next is: max(cpu_util + task_util,
6680 * cpu_util_est + _task_util_est)
6682 if (cpu == dst_cpu) {
6685 cpu_util_next(cpu, p, -1) + task_util_est(p);
6689 * Busy time computation: utilization clamping is not
6690 * required since the ratio (sum_util / cpu_capacity)
6691 * is already enough to scale the EM reported power
6692 * consumption at the (eventually clamped) cpu_capacity.
6694 cpu_util = effective_cpu_util(cpu, util_running, cpu_cap,
6697 sum_util += min(cpu_util, _cpu_cap);
6700 * Performance domain frequency: utilization clamping
6701 * must be considered since it affects the selection
6702 * of the performance domain frequency.
6703 * NOTE: in case RT tasks are running, by default the
6704 * FREQUENCY_UTIL's utilization can be max OPP.
6706 cpu_util = effective_cpu_util(cpu, util_freq, cpu_cap,
6707 FREQUENCY_UTIL, tsk);
6708 max_util = max(max_util, min(cpu_util, _cpu_cap));
6711 return em_cpu_energy(pd->em_pd, max_util, sum_util, _cpu_cap);
6715 * find_energy_efficient_cpu(): Find most energy-efficient target CPU for the
6716 * waking task. find_energy_efficient_cpu() looks for the CPU with maximum
6717 * spare capacity in each performance domain and uses it as a potential
6718 * candidate to execute the task. Then, it uses the Energy Model to figure
6719 * out which of the CPU candidates is the most energy-efficient.
6721 * The rationale for this heuristic is as follows. In a performance domain,
6722 * all the most energy efficient CPU candidates (according to the Energy
6723 * Model) are those for which we'll request a low frequency. When there are
6724 * several CPUs for which the frequency request will be the same, we don't
6725 * have enough data to break the tie between them, because the Energy Model
6726 * only includes active power costs. With this model, if we assume that
6727 * frequency requests follow utilization (e.g. using schedutil), the CPU with
6728 * the maximum spare capacity in a performance domain is guaranteed to be among
6729 * the best candidates of the performance domain.
6731 * In practice, it could be preferable from an energy standpoint to pack
6732 * small tasks on a CPU in order to let other CPUs go in deeper idle states,
6733 * but that could also hurt our chances to go cluster idle, and we have no
6734 * ways to tell with the current Energy Model if this is actually a good
6735 * idea or not. So, find_energy_efficient_cpu() basically favors
6736 * cluster-packing, and spreading inside a cluster. That should at least be
6737 * a good thing for latency, and this is consistent with the idea that most
6738 * of the energy savings of EAS come from the asymmetry of the system, and
6739 * not so much from breaking the tie between identical CPUs. That's also the
6740 * reason why EAS is enabled in the topology code only for systems where
6741 * SD_ASYM_CPUCAPACITY is set.
6743 * NOTE: Forkees are not accepted in the energy-aware wake-up path because
6744 * they don't have any useful utilization data yet and it's not possible to
6745 * forecast their impact on energy consumption. Consequently, they will be
6746 * placed by find_idlest_cpu() on the least loaded CPU, which might turn out
6747 * to be energy-inefficient in some use-cases. The alternative would be to
6748 * bias new tasks towards specific types of CPUs first, or to try to infer
6749 * their util_avg from the parent task, but those heuristics could hurt
6750 * other use-cases too. So, until someone finds a better way to solve this,
6751 * let's keep things simple by re-using the existing slow path.
6753 static int find_energy_efficient_cpu(struct task_struct *p, int prev_cpu)
6755 unsigned long prev_delta = ULONG_MAX, best_delta = ULONG_MAX;
6756 struct root_domain *rd = cpu_rq(smp_processor_id())->rd;
6757 int cpu, best_energy_cpu = prev_cpu, target = -1;
6758 unsigned long cpu_cap, util, base_energy = 0;
6759 struct sched_domain *sd;
6760 struct perf_domain *pd;
6763 pd = rcu_dereference(rd->pd);
6764 if (!pd || READ_ONCE(rd->overutilized))
6768 * Energy-aware wake-up happens on the lowest sched_domain starting
6769 * from sd_asym_cpucapacity spanning over this_cpu and prev_cpu.
6771 sd = rcu_dereference(*this_cpu_ptr(&sd_asym_cpucapacity));
6772 while (sd && !cpumask_test_cpu(prev_cpu, sched_domain_span(sd)))
6779 sync_entity_load_avg(&p->se);
6780 if (!task_util_est(p))
6783 for (; pd; pd = pd->next) {
6784 unsigned long cur_delta, spare_cap, max_spare_cap = 0;
6785 bool compute_prev_delta = false;
6786 unsigned long base_energy_pd;
6787 int max_spare_cap_cpu = -1;
6789 for_each_cpu_and(cpu, perf_domain_span(pd), sched_domain_span(sd)) {
6790 if (!cpumask_test_cpu(cpu, p->cpus_ptr))
6793 util = cpu_util_next(cpu, p, cpu);
6794 cpu_cap = capacity_of(cpu);
6795 spare_cap = cpu_cap;
6796 lsub_positive(&spare_cap, util);
6799 * Skip CPUs that cannot satisfy the capacity request.
6800 * IOW, placing the task there would make the CPU
6801 * overutilized. Take uclamp into account to see how
6802 * much capacity we can get out of the CPU; this is
6803 * aligned with sched_cpu_util().
6805 util = uclamp_rq_util_with(cpu_rq(cpu), util, p);
6806 if (!fits_capacity(util, cpu_cap))
6809 if (cpu == prev_cpu) {
6810 /* Always use prev_cpu as a candidate. */
6811 compute_prev_delta = true;
6812 } else if (spare_cap > max_spare_cap) {
6814 * Find the CPU with the maximum spare capacity
6815 * in the performance domain.
6817 max_spare_cap = spare_cap;
6818 max_spare_cap_cpu = cpu;
6822 if (max_spare_cap_cpu < 0 && !compute_prev_delta)
6825 /* Compute the 'base' energy of the pd, without @p */
6826 base_energy_pd = compute_energy(p, -1, pd);
6827 base_energy += base_energy_pd;
6829 /* Evaluate the energy impact of using prev_cpu. */
6830 if (compute_prev_delta) {
6831 prev_delta = compute_energy(p, prev_cpu, pd);
6832 if (prev_delta < base_energy_pd)
6834 prev_delta -= base_energy_pd;
6835 best_delta = min(best_delta, prev_delta);
6838 /* Evaluate the energy impact of using max_spare_cap_cpu. */
6839 if (max_spare_cap_cpu >= 0) {
6840 cur_delta = compute_energy(p, max_spare_cap_cpu, pd);
6841 if (cur_delta < base_energy_pd)
6843 cur_delta -= base_energy_pd;
6844 if (cur_delta < best_delta) {
6845 best_delta = cur_delta;
6846 best_energy_cpu = max_spare_cap_cpu;
6853 * Pick the best CPU if prev_cpu cannot be used, or if it saves at
6854 * least 6% of the energy used by prev_cpu.
6856 if ((prev_delta == ULONG_MAX) ||
6857 (prev_delta - best_delta) > ((prev_delta + base_energy) >> 4))
6858 target = best_energy_cpu;
6869 * select_task_rq_fair: Select target runqueue for the waking task in domains
6870 * that have the relevant SD flag set. In practice, this is SD_BALANCE_WAKE,
6871 * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
6873 * Balances load by selecting the idlest CPU in the idlest group, or under
6874 * certain conditions an idle sibling CPU if the domain has SD_WAKE_AFFINE set.
6876 * Returns the target CPU number.
6879 select_task_rq_fair(struct task_struct *p, int prev_cpu, int wake_flags)
6881 int sync = (wake_flags & WF_SYNC) && !(current->flags & PF_EXITING);
6882 struct sched_domain *tmp, *sd = NULL;
6883 int cpu = smp_processor_id();
6884 int new_cpu = prev_cpu;
6885 int want_affine = 0;
6886 /* SD_flags and WF_flags share the first nibble */
6887 int sd_flag = wake_flags & 0xF;
6890 * required for stable ->cpus_allowed
6892 lockdep_assert_held(&p->pi_lock);
6893 if (wake_flags & WF_TTWU) {
6896 if (sched_energy_enabled()) {
6897 new_cpu = find_energy_efficient_cpu(p, prev_cpu);
6903 want_affine = !wake_wide(p) && cpumask_test_cpu(cpu, p->cpus_ptr);
6907 for_each_domain(cpu, tmp) {
6909 * If both 'cpu' and 'prev_cpu' are part of this domain,
6910 * cpu is a valid SD_WAKE_AFFINE target.
6912 if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
6913 cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
6914 if (cpu != prev_cpu)
6915 new_cpu = wake_affine(tmp, p, cpu, prev_cpu, sync);
6917 sd = NULL; /* Prefer wake_affine over balance flags */
6921 if (tmp->flags & sd_flag)
6923 else if (!want_affine)
6929 new_cpu = find_idlest_cpu(sd, p, cpu, prev_cpu, sd_flag);
6930 } else if (wake_flags & WF_TTWU) { /* XXX always ? */
6932 new_cpu = select_idle_sibling(p, prev_cpu, new_cpu);
6939 static void detach_entity_cfs_rq(struct sched_entity *se);
6942 * Called immediately before a task is migrated to a new CPU; task_cpu(p) and
6943 * cfs_rq_of(p) references at time of call are still valid and identify the
6944 * previous CPU. The caller guarantees p->pi_lock or task_rq(p)->lock is held.
6946 static void migrate_task_rq_fair(struct task_struct *p, int new_cpu)
6949 * As blocked tasks retain absolute vruntime the migration needs to
6950 * deal with this by subtracting the old and adding the new
6951 * min_vruntime -- the latter is done by enqueue_entity() when placing
6952 * the task on the new runqueue.
6954 if (READ_ONCE(p->__state) == TASK_WAKING) {
6955 struct sched_entity *se = &p->se;
6956 struct cfs_rq *cfs_rq = cfs_rq_of(se);
6959 #ifndef CONFIG_64BIT
6960 u64 min_vruntime_copy;
6963 min_vruntime_copy = cfs_rq->min_vruntime_copy;
6965 min_vruntime = cfs_rq->min_vruntime;
6966 } while (min_vruntime != min_vruntime_copy);
6968 min_vruntime = cfs_rq->min_vruntime;
6971 se->vruntime -= min_vruntime;
6974 if (p->on_rq == TASK_ON_RQ_MIGRATING) {
6976 * In case of TASK_ON_RQ_MIGRATING we in fact hold the 'old'
6977 * rq->lock and can modify state directly.
6979 lockdep_assert_rq_held(task_rq(p));
6980 detach_entity_cfs_rq(&p->se);
6984 * We are supposed to update the task to "current" time, then
6985 * its up to date and ready to go to new CPU/cfs_rq. But we
6986 * have difficulty in getting what current time is, so simply
6987 * throw away the out-of-date time. This will result in the
6988 * wakee task is less decayed, but giving the wakee more load
6991 remove_entity_load_avg(&p->se);
6994 /* Tell new CPU we are migrated */
6995 p->se.avg.last_update_time = 0;
6997 /* We have migrated, no longer consider this task hot */
6998 p->se.exec_start = 0;
7000 update_scan_period(p, new_cpu);
7003 static void task_dead_fair(struct task_struct *p)
7005 remove_entity_load_avg(&p->se);
7009 balance_fair(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
7014 return newidle_balance(rq, rf) != 0;
7016 #endif /* CONFIG_SMP */
7018 static unsigned long wakeup_gran(struct sched_entity *se)
7020 unsigned long gran = sysctl_sched_wakeup_granularity;
7023 * Since its curr running now, convert the gran from real-time
7024 * to virtual-time in his units.
7026 * By using 'se' instead of 'curr' we penalize light tasks, so
7027 * they get preempted easier. That is, if 'se' < 'curr' then
7028 * the resulting gran will be larger, therefore penalizing the
7029 * lighter, if otoh 'se' > 'curr' then the resulting gran will
7030 * be smaller, again penalizing the lighter task.
7032 * This is especially important for buddies when the leftmost
7033 * task is higher priority than the buddy.
7035 return calc_delta_fair(gran, se);
7039 * Should 'se' preempt 'curr'.
7053 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
7055 s64 gran, vdiff = curr->vruntime - se->vruntime;
7060 gran = wakeup_gran(se);
7067 static void set_last_buddy(struct sched_entity *se)
7069 for_each_sched_entity(se) {
7070 if (SCHED_WARN_ON(!se->on_rq))
7074 cfs_rq_of(se)->last = se;
7078 static void set_next_buddy(struct sched_entity *se)
7080 for_each_sched_entity(se) {
7081 if (SCHED_WARN_ON(!se->on_rq))
7085 cfs_rq_of(se)->next = se;
7089 static void set_skip_buddy(struct sched_entity *se)
7091 for_each_sched_entity(se)
7092 cfs_rq_of(se)->skip = se;
7096 * Preempt the current task with a newly woken task if needed:
7098 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
7100 struct task_struct *curr = rq->curr;
7101 struct sched_entity *se = &curr->se, *pse = &p->se;
7102 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
7103 int scale = cfs_rq->nr_running >= sched_nr_latency;
7104 int next_buddy_marked = 0;
7105 int cse_is_idle, pse_is_idle;
7107 if (unlikely(se == pse))
7111 * This is possible from callers such as attach_tasks(), in which we
7112 * unconditionally check_preempt_curr() after an enqueue (which may have
7113 * lead to a throttle). This both saves work and prevents false
7114 * next-buddy nomination below.
7116 if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
7119 if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
7120 set_next_buddy(pse);
7121 next_buddy_marked = 1;
7125 * We can come here with TIF_NEED_RESCHED already set from new task
7128 * Note: this also catches the edge-case of curr being in a throttled
7129 * group (e.g. via set_curr_task), since update_curr() (in the
7130 * enqueue of curr) will have resulted in resched being set. This
7131 * prevents us from potentially nominating it as a false LAST_BUDDY
7134 if (test_tsk_need_resched(curr))
7137 /* Idle tasks are by definition preempted by non-idle tasks. */
7138 if (unlikely(task_has_idle_policy(curr)) &&
7139 likely(!task_has_idle_policy(p)))
7143 * Batch and idle tasks do not preempt non-idle tasks (their preemption
7144 * is driven by the tick):
7146 if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
7149 find_matching_se(&se, &pse);
7152 cse_is_idle = se_is_idle(se);
7153 pse_is_idle = se_is_idle(pse);
7156 * Preempt an idle group in favor of a non-idle group (and don't preempt
7157 * in the inverse case).
7159 if (cse_is_idle && !pse_is_idle)
7161 if (cse_is_idle != pse_is_idle)
7164 update_curr(cfs_rq_of(se));
7165 if (wakeup_preempt_entity(se, pse) == 1) {
7167 * Bias pick_next to pick the sched entity that is
7168 * triggering this preemption.
7170 if (!next_buddy_marked)
7171 set_next_buddy(pse);
7180 * Only set the backward buddy when the current task is still
7181 * on the rq. This can happen when a wakeup gets interleaved
7182 * with schedule on the ->pre_schedule() or idle_balance()
7183 * point, either of which can * drop the rq lock.
7185 * Also, during early boot the idle thread is in the fair class,
7186 * for obvious reasons its a bad idea to schedule back to it.
7188 if (unlikely(!se->on_rq || curr == rq->idle))
7191 if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
7196 static struct task_struct *pick_task_fair(struct rq *rq)
7198 struct sched_entity *se;
7199 struct cfs_rq *cfs_rq;
7203 if (!cfs_rq->nr_running)
7207 struct sched_entity *curr = cfs_rq->curr;
7209 /* When we pick for a remote RQ, we'll not have done put_prev_entity() */
7212 update_curr(cfs_rq);
7216 if (unlikely(check_cfs_rq_runtime(cfs_rq)))
7220 se = pick_next_entity(cfs_rq, curr);
7221 cfs_rq = group_cfs_rq(se);
7228 struct task_struct *
7229 pick_next_task_fair(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
7231 struct cfs_rq *cfs_rq = &rq->cfs;
7232 struct sched_entity *se;
7233 struct task_struct *p;
7237 if (!sched_fair_runnable(rq))
7240 #ifdef CONFIG_FAIR_GROUP_SCHED
7241 if (!prev || prev->sched_class != &fair_sched_class)
7245 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
7246 * likely that a next task is from the same cgroup as the current.
7248 * Therefore attempt to avoid putting and setting the entire cgroup
7249 * hierarchy, only change the part that actually changes.
7253 struct sched_entity *curr = cfs_rq->curr;
7256 * Since we got here without doing put_prev_entity() we also
7257 * have to consider cfs_rq->curr. If it is still a runnable
7258 * entity, update_curr() will update its vruntime, otherwise
7259 * forget we've ever seen it.
7263 update_curr(cfs_rq);
7268 * This call to check_cfs_rq_runtime() will do the
7269 * throttle and dequeue its entity in the parent(s).
7270 * Therefore the nr_running test will indeed
7273 if (unlikely(check_cfs_rq_runtime(cfs_rq))) {
7276 if (!cfs_rq->nr_running)
7283 se = pick_next_entity(cfs_rq, curr);
7284 cfs_rq = group_cfs_rq(se);
7290 * Since we haven't yet done put_prev_entity and if the selected task
7291 * is a different task than we started out with, try and touch the
7292 * least amount of cfs_rqs.
7295 struct sched_entity *pse = &prev->se;
7297 while (!(cfs_rq = is_same_group(se, pse))) {
7298 int se_depth = se->depth;
7299 int pse_depth = pse->depth;
7301 if (se_depth <= pse_depth) {
7302 put_prev_entity(cfs_rq_of(pse), pse);
7303 pse = parent_entity(pse);
7305 if (se_depth >= pse_depth) {
7306 set_next_entity(cfs_rq_of(se), se);
7307 se = parent_entity(se);
7311 put_prev_entity(cfs_rq, pse);
7312 set_next_entity(cfs_rq, se);
7319 put_prev_task(rq, prev);
7322 se = pick_next_entity(cfs_rq, NULL);
7323 set_next_entity(cfs_rq, se);
7324 cfs_rq = group_cfs_rq(se);
7329 done: __maybe_unused;
7332 * Move the next running task to the front of
7333 * the list, so our cfs_tasks list becomes MRU
7336 list_move(&p->se.group_node, &rq->cfs_tasks);
7339 if (hrtick_enabled_fair(rq))
7340 hrtick_start_fair(rq, p);
7342 update_misfit_status(p, rq);
7350 new_tasks = newidle_balance(rq, rf);
7353 * Because newidle_balance() releases (and re-acquires) rq->lock, it is
7354 * possible for any higher priority task to appear. In that case we
7355 * must re-start the pick_next_entity() loop.
7364 * rq is about to be idle, check if we need to update the
7365 * lost_idle_time of clock_pelt
7367 update_idle_rq_clock_pelt(rq);
7372 static struct task_struct *__pick_next_task_fair(struct rq *rq)
7374 return pick_next_task_fair(rq, NULL, NULL);
7378 * Account for a descheduled task:
7380 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
7382 struct sched_entity *se = &prev->se;
7383 struct cfs_rq *cfs_rq;
7385 for_each_sched_entity(se) {
7386 cfs_rq = cfs_rq_of(se);
7387 put_prev_entity(cfs_rq, se);
7392 * sched_yield() is very simple
7394 * The magic of dealing with the ->skip buddy is in pick_next_entity.
7396 static void yield_task_fair(struct rq *rq)
7398 struct task_struct *curr = rq->curr;
7399 struct cfs_rq *cfs_rq = task_cfs_rq(curr);
7400 struct sched_entity *se = &curr->se;
7403 * Are we the only task in the tree?
7405 if (unlikely(rq->nr_running == 1))
7408 clear_buddies(cfs_rq, se);
7410 if (curr->policy != SCHED_BATCH) {
7411 update_rq_clock(rq);
7413 * Update run-time statistics of the 'current'.
7415 update_curr(cfs_rq);
7417 * Tell update_rq_clock() that we've just updated,
7418 * so we don't do microscopic update in schedule()
7419 * and double the fastpath cost.
7421 rq_clock_skip_update(rq);
7427 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p)
7429 struct sched_entity *se = &p->se;
7431 /* throttled hierarchies are not runnable */
7432 if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
7435 /* Tell the scheduler that we'd really like pse to run next. */
7438 yield_task_fair(rq);
7444 /**************************************************
7445 * Fair scheduling class load-balancing methods.
7449 * The purpose of load-balancing is to achieve the same basic fairness the
7450 * per-CPU scheduler provides, namely provide a proportional amount of compute
7451 * time to each task. This is expressed in the following equation:
7453 * W_i,n/P_i == W_j,n/P_j for all i,j (1)
7455 * Where W_i,n is the n-th weight average for CPU i. The instantaneous weight
7456 * W_i,0 is defined as:
7458 * W_i,0 = \Sum_j w_i,j (2)
7460 * Where w_i,j is the weight of the j-th runnable task on CPU i. This weight
7461 * is derived from the nice value as per sched_prio_to_weight[].
7463 * The weight average is an exponential decay average of the instantaneous
7466 * W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0 (3)
7468 * C_i is the compute capacity of CPU i, typically it is the
7469 * fraction of 'recent' time available for SCHED_OTHER task execution. But it
7470 * can also include other factors [XXX].
7472 * To achieve this balance we define a measure of imbalance which follows
7473 * directly from (1):
7475 * imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j } (4)
7477 * We them move tasks around to minimize the imbalance. In the continuous
7478 * function space it is obvious this converges, in the discrete case we get
7479 * a few fun cases generally called infeasible weight scenarios.
7482 * - infeasible weights;
7483 * - local vs global optima in the discrete case. ]
7488 * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
7489 * for all i,j solution, we create a tree of CPUs that follows the hardware
7490 * topology where each level pairs two lower groups (or better). This results
7491 * in O(log n) layers. Furthermore we reduce the number of CPUs going up the
7492 * tree to only the first of the previous level and we decrease the frequency
7493 * of load-balance at each level inv. proportional to the number of CPUs in
7499 * \Sum { --- * --- * 2^i } = O(n) (5)
7501 * `- size of each group
7502 * | | `- number of CPUs doing load-balance
7504 * `- sum over all levels
7506 * Coupled with a limit on how many tasks we can migrate every balance pass,
7507 * this makes (5) the runtime complexity of the balancer.
7509 * An important property here is that each CPU is still (indirectly) connected
7510 * to every other CPU in at most O(log n) steps:
7512 * The adjacency matrix of the resulting graph is given by:
7515 * A_i,j = \Union (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1) (6)
7518 * And you'll find that:
7520 * A^(log_2 n)_i,j != 0 for all i,j (7)
7522 * Showing there's indeed a path between every CPU in at most O(log n) steps.
7523 * The task movement gives a factor of O(m), giving a convergence complexity
7526 * O(nm log n), n := nr_cpus, m := nr_tasks (8)
7531 * In order to avoid CPUs going idle while there's still work to do, new idle
7532 * balancing is more aggressive and has the newly idle CPU iterate up the domain
7533 * tree itself instead of relying on other CPUs to bring it work.
7535 * This adds some complexity to both (5) and (8) but it reduces the total idle
7543 * Cgroups make a horror show out of (2), instead of a simple sum we get:
7546 * W_i,0 = \Sum_j \Prod_k w_k * ----- (9)
7551 * s_k,i = \Sum_j w_i,j,k and S_k = \Sum_i s_k,i (10)
7553 * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on CPU i.
7555 * The big problem is S_k, its a global sum needed to compute a local (W_i)
7558 * [XXX write more on how we solve this.. _after_ merging pjt's patches that
7559 * rewrite all of this once again.]
7562 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
7564 enum fbq_type { regular, remote, all };
7567 * 'group_type' describes the group of CPUs at the moment of load balancing.
7569 * The enum is ordered by pulling priority, with the group with lowest priority
7570 * first so the group_type can simply be compared when selecting the busiest
7571 * group. See update_sd_pick_busiest().
7574 /* The group has spare capacity that can be used to run more tasks. */
7575 group_has_spare = 0,
7577 * The group is fully used and the tasks don't compete for more CPU
7578 * cycles. Nevertheless, some tasks might wait before running.
7582 * SD_ASYM_CPUCAPACITY only: One task doesn't fit with CPU's capacity
7583 * and must be migrated to a more powerful CPU.
7587 * SD_ASYM_PACKING only: One local CPU with higher capacity is available,
7588 * and the task should be migrated to it instead of running on the
7593 * The tasks' affinity constraints previously prevented the scheduler
7594 * from balancing the load across the system.
7598 * The CPU is overloaded and can't provide expected CPU cycles to all
7604 enum migration_type {
7611 #define LBF_ALL_PINNED 0x01
7612 #define LBF_NEED_BREAK 0x02
7613 #define LBF_DST_PINNED 0x04
7614 #define LBF_SOME_PINNED 0x08
7615 #define LBF_ACTIVE_LB 0x10
7618 struct sched_domain *sd;
7626 struct cpumask *dst_grpmask;
7628 enum cpu_idle_type idle;
7630 /* The set of CPUs under consideration for load-balancing */
7631 struct cpumask *cpus;
7636 unsigned int loop_break;
7637 unsigned int loop_max;
7639 enum fbq_type fbq_type;
7640 enum migration_type migration_type;
7641 struct list_head tasks;
7645 * Is this task likely cache-hot:
7647 static int task_hot(struct task_struct *p, struct lb_env *env)
7651 lockdep_assert_rq_held(env->src_rq);
7653 if (p->sched_class != &fair_sched_class)
7656 if (unlikely(task_has_idle_policy(p)))
7659 /* SMT siblings share cache */
7660 if (env->sd->flags & SD_SHARE_CPUCAPACITY)
7664 * Buddy candidates are cache hot:
7666 if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running &&
7667 (&p->se == cfs_rq_of(&p->se)->next ||
7668 &p->se == cfs_rq_of(&p->se)->last))
7671 if (sysctl_sched_migration_cost == -1)
7675 * Don't migrate task if the task's cookie does not match
7676 * with the destination CPU's core cookie.
7678 if (!sched_core_cookie_match(cpu_rq(env->dst_cpu), p))
7681 if (sysctl_sched_migration_cost == 0)
7684 delta = rq_clock_task(env->src_rq) - p->se.exec_start;
7686 return delta < (s64)sysctl_sched_migration_cost;
7689 #ifdef CONFIG_NUMA_BALANCING
7691 * Returns 1, if task migration degrades locality
7692 * Returns 0, if task migration improves locality i.e migration preferred.
7693 * Returns -1, if task migration is not affected by locality.
7695 static int migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
7697 struct numa_group *numa_group = rcu_dereference(p->numa_group);
7698 unsigned long src_weight, dst_weight;
7699 int src_nid, dst_nid, dist;
7701 if (!static_branch_likely(&sched_numa_balancing))
7704 if (!p->numa_faults || !(env->sd->flags & SD_NUMA))
7707 src_nid = cpu_to_node(env->src_cpu);
7708 dst_nid = cpu_to_node(env->dst_cpu);
7710 if (src_nid == dst_nid)
7713 /* Migrating away from the preferred node is always bad. */
7714 if (src_nid == p->numa_preferred_nid) {
7715 if (env->src_rq->nr_running > env->src_rq->nr_preferred_running)
7721 /* Encourage migration to the preferred node. */
7722 if (dst_nid == p->numa_preferred_nid)
7725 /* Leaving a core idle is often worse than degrading locality. */
7726 if (env->idle == CPU_IDLE)
7729 dist = node_distance(src_nid, dst_nid);
7731 src_weight = group_weight(p, src_nid, dist);
7732 dst_weight = group_weight(p, dst_nid, dist);
7734 src_weight = task_weight(p, src_nid, dist);
7735 dst_weight = task_weight(p, dst_nid, dist);
7738 return dst_weight < src_weight;
7742 static inline int migrate_degrades_locality(struct task_struct *p,
7750 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
7753 int can_migrate_task(struct task_struct *p, struct lb_env *env)
7757 lockdep_assert_rq_held(env->src_rq);
7760 * We do not migrate tasks that are:
7761 * 1) throttled_lb_pair, or
7762 * 2) cannot be migrated to this CPU due to cpus_ptr, or
7763 * 3) running (obviously), or
7764 * 4) are cache-hot on their current CPU.
7766 if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
7769 /* Disregard pcpu kthreads; they are where they need to be. */
7770 if (kthread_is_per_cpu(p))
7773 if (!cpumask_test_cpu(env->dst_cpu, p->cpus_ptr)) {
7776 schedstat_inc(p->se.statistics.nr_failed_migrations_affine);
7778 env->flags |= LBF_SOME_PINNED;
7781 * Remember if this task can be migrated to any other CPU in
7782 * our sched_group. We may want to revisit it if we couldn't
7783 * meet load balance goals by pulling other tasks on src_cpu.
7785 * Avoid computing new_dst_cpu
7787 * - if we have already computed one in current iteration
7788 * - if it's an active balance
7790 if (env->idle == CPU_NEWLY_IDLE ||
7791 env->flags & (LBF_DST_PINNED | LBF_ACTIVE_LB))
7794 /* Prevent to re-select dst_cpu via env's CPUs: */
7795 for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
7796 if (cpumask_test_cpu(cpu, p->cpus_ptr)) {
7797 env->flags |= LBF_DST_PINNED;
7798 env->new_dst_cpu = cpu;
7806 /* Record that we found at least one task that could run on dst_cpu */
7807 env->flags &= ~LBF_ALL_PINNED;
7809 if (task_running(env->src_rq, p)) {
7810 schedstat_inc(p->se.statistics.nr_failed_migrations_running);
7815 * Aggressive migration if:
7817 * 2) destination numa is preferred
7818 * 3) task is cache cold, or
7819 * 4) too many balance attempts have failed.
7821 if (env->flags & LBF_ACTIVE_LB)
7824 tsk_cache_hot = migrate_degrades_locality(p, env);
7825 if (tsk_cache_hot == -1)
7826 tsk_cache_hot = task_hot(p, env);
7828 if (tsk_cache_hot <= 0 ||
7829 env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
7830 if (tsk_cache_hot == 1) {
7831 schedstat_inc(env->sd->lb_hot_gained[env->idle]);
7832 schedstat_inc(p->se.statistics.nr_forced_migrations);
7837 schedstat_inc(p->se.statistics.nr_failed_migrations_hot);
7842 * detach_task() -- detach the task for the migration specified in env
7844 static void detach_task(struct task_struct *p, struct lb_env *env)
7846 lockdep_assert_rq_held(env->src_rq);
7848 deactivate_task(env->src_rq, p, DEQUEUE_NOCLOCK);
7849 set_task_cpu(p, env->dst_cpu);
7853 * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
7854 * part of active balancing operations within "domain".
7856 * Returns a task if successful and NULL otherwise.
7858 static struct task_struct *detach_one_task(struct lb_env *env)
7860 struct task_struct *p;
7862 lockdep_assert_rq_held(env->src_rq);
7864 list_for_each_entry_reverse(p,
7865 &env->src_rq->cfs_tasks, se.group_node) {
7866 if (!can_migrate_task(p, env))
7869 detach_task(p, env);
7872 * Right now, this is only the second place where
7873 * lb_gained[env->idle] is updated (other is detach_tasks)
7874 * so we can safely collect stats here rather than
7875 * inside detach_tasks().
7877 schedstat_inc(env->sd->lb_gained[env->idle]);
7883 static const unsigned int sched_nr_migrate_break = 32;
7886 * detach_tasks() -- tries to detach up to imbalance load/util/tasks from
7887 * busiest_rq, as part of a balancing operation within domain "sd".
7889 * Returns number of detached tasks if successful and 0 otherwise.
7891 static int detach_tasks(struct lb_env *env)
7893 struct list_head *tasks = &env->src_rq->cfs_tasks;
7894 unsigned long util, load;
7895 struct task_struct *p;
7898 lockdep_assert_rq_held(env->src_rq);
7901 * Source run queue has been emptied by another CPU, clear
7902 * LBF_ALL_PINNED flag as we will not test any task.
7904 if (env->src_rq->nr_running <= 1) {
7905 env->flags &= ~LBF_ALL_PINNED;
7909 if (env->imbalance <= 0)
7912 while (!list_empty(tasks)) {
7914 * We don't want to steal all, otherwise we may be treated likewise,
7915 * which could at worst lead to a livelock crash.
7917 if (env->idle != CPU_NOT_IDLE && env->src_rq->nr_running <= 1)
7920 p = list_last_entry(tasks, struct task_struct, se.group_node);
7923 /* We've more or less seen every task there is, call it quits */
7924 if (env->loop > env->loop_max)
7927 /* take a breather every nr_migrate tasks */
7928 if (env->loop > env->loop_break) {
7929 env->loop_break += sched_nr_migrate_break;
7930 env->flags |= LBF_NEED_BREAK;
7934 if (!can_migrate_task(p, env))
7937 switch (env->migration_type) {
7940 * Depending of the number of CPUs and tasks and the
7941 * cgroup hierarchy, task_h_load() can return a null
7942 * value. Make sure that env->imbalance decreases
7943 * otherwise detach_tasks() will stop only after
7944 * detaching up to loop_max tasks.
7946 load = max_t(unsigned long, task_h_load(p), 1);
7948 if (sched_feat(LB_MIN) &&
7949 load < 16 && !env->sd->nr_balance_failed)
7953 * Make sure that we don't migrate too much load.
7954 * Nevertheless, let relax the constraint if
7955 * scheduler fails to find a good waiting task to
7958 if (shr_bound(load, env->sd->nr_balance_failed) > env->imbalance)
7961 env->imbalance -= load;
7965 util = task_util_est(p);
7967 if (util > env->imbalance)
7970 env->imbalance -= util;
7977 case migrate_misfit:
7978 /* This is not a misfit task */
7979 if (task_fits_capacity(p, capacity_of(env->src_cpu)))
7986 detach_task(p, env);
7987 list_add(&p->se.group_node, &env->tasks);
7991 #ifdef CONFIG_PREEMPTION
7993 * NEWIDLE balancing is a source of latency, so preemptible
7994 * kernels will stop after the first task is detached to minimize
7995 * the critical section.
7997 if (env->idle == CPU_NEWLY_IDLE)
8002 * We only want to steal up to the prescribed amount of
8005 if (env->imbalance <= 0)
8010 list_move(&p->se.group_node, tasks);
8014 * Right now, this is one of only two places we collect this stat
8015 * so we can safely collect detach_one_task() stats here rather
8016 * than inside detach_one_task().
8018 schedstat_add(env->sd->lb_gained[env->idle], detached);
8024 * attach_task() -- attach the task detached by detach_task() to its new rq.
8026 static void attach_task(struct rq *rq, struct task_struct *p)
8028 lockdep_assert_rq_held(rq);
8030 BUG_ON(task_rq(p) != rq);
8031 activate_task(rq, p, ENQUEUE_NOCLOCK);
8032 check_preempt_curr(rq, p, 0);
8036 * attach_one_task() -- attaches the task returned from detach_one_task() to
8039 static void attach_one_task(struct rq *rq, struct task_struct *p)
8044 update_rq_clock(rq);
8050 * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
8053 static void attach_tasks(struct lb_env *env)
8055 struct list_head *tasks = &env->tasks;
8056 struct task_struct *p;
8059 rq_lock(env->dst_rq, &rf);
8060 update_rq_clock(env->dst_rq);
8062 while (!list_empty(tasks)) {
8063 p = list_first_entry(tasks, struct task_struct, se.group_node);
8064 list_del_init(&p->se.group_node);
8066 attach_task(env->dst_rq, p);
8069 rq_unlock(env->dst_rq, &rf);
8072 #ifdef CONFIG_NO_HZ_COMMON
8073 static inline bool cfs_rq_has_blocked(struct cfs_rq *cfs_rq)
8075 if (cfs_rq->avg.load_avg)
8078 if (cfs_rq->avg.util_avg)
8084 static inline bool others_have_blocked(struct rq *rq)
8086 if (READ_ONCE(rq->avg_rt.util_avg))
8089 if (READ_ONCE(rq->avg_dl.util_avg))
8092 if (thermal_load_avg(rq))
8095 #ifdef CONFIG_HAVE_SCHED_AVG_IRQ
8096 if (READ_ONCE(rq->avg_irq.util_avg))
8103 static inline void update_blocked_load_tick(struct rq *rq)
8105 WRITE_ONCE(rq->last_blocked_load_update_tick, jiffies);
8108 static inline void update_blocked_load_status(struct rq *rq, bool has_blocked)
8111 rq->has_blocked_load = 0;
8114 static inline bool cfs_rq_has_blocked(struct cfs_rq *cfs_rq) { return false; }
8115 static inline bool others_have_blocked(struct rq *rq) { return false; }
8116 static inline void update_blocked_load_tick(struct rq *rq) {}
8117 static inline void update_blocked_load_status(struct rq *rq, bool has_blocked) {}
8120 static bool __update_blocked_others(struct rq *rq, bool *done)
8122 const struct sched_class *curr_class;
8123 u64 now = rq_clock_pelt(rq);
8124 unsigned long thermal_pressure;
8128 * update_load_avg() can call cpufreq_update_util(). Make sure that RT,
8129 * DL and IRQ signals have been updated before updating CFS.
8131 curr_class = rq->curr->sched_class;
8133 thermal_pressure = arch_scale_thermal_pressure(cpu_of(rq));
8135 decayed = update_rt_rq_load_avg(now, rq, curr_class == &rt_sched_class) |
8136 update_dl_rq_load_avg(now, rq, curr_class == &dl_sched_class) |
8137 update_thermal_load_avg(rq_clock_thermal(rq), rq, thermal_pressure) |
8138 update_irq_load_avg(rq, 0);
8140 if (others_have_blocked(rq))
8146 #ifdef CONFIG_FAIR_GROUP_SCHED
8148 static bool __update_blocked_fair(struct rq *rq, bool *done)
8150 struct cfs_rq *cfs_rq, *pos;
8151 bool decayed = false;
8152 int cpu = cpu_of(rq);
8155 * Iterates the task_group tree in a bottom up fashion, see
8156 * list_add_leaf_cfs_rq() for details.
8158 for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) {
8159 struct sched_entity *se;
8161 if (update_cfs_rq_load_avg(cfs_rq_clock_pelt(cfs_rq), cfs_rq)) {
8162 update_tg_load_avg(cfs_rq);
8164 if (cfs_rq == &rq->cfs)
8168 /* Propagate pending load changes to the parent, if any: */
8169 se = cfs_rq->tg->se[cpu];
8170 if (se && !skip_blocked_update(se))
8171 update_load_avg(cfs_rq_of(se), se, UPDATE_TG);
8174 * There can be a lot of idle CPU cgroups. Don't let fully
8175 * decayed cfs_rqs linger on the list.
8177 if (cfs_rq_is_decayed(cfs_rq))
8178 list_del_leaf_cfs_rq(cfs_rq);
8180 /* Don't need periodic decay once load/util_avg are null */
8181 if (cfs_rq_has_blocked(cfs_rq))
8189 * Compute the hierarchical load factor for cfs_rq and all its ascendants.
8190 * This needs to be done in a top-down fashion because the load of a child
8191 * group is a fraction of its parents load.
8193 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
8195 struct rq *rq = rq_of(cfs_rq);
8196 struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
8197 unsigned long now = jiffies;
8200 if (cfs_rq->last_h_load_update == now)
8203 WRITE_ONCE(cfs_rq->h_load_next, NULL);
8204 for_each_sched_entity(se) {
8205 cfs_rq = cfs_rq_of(se);
8206 WRITE_ONCE(cfs_rq->h_load_next, se);
8207 if (cfs_rq->last_h_load_update == now)
8212 cfs_rq->h_load = cfs_rq_load_avg(cfs_rq);
8213 cfs_rq->last_h_load_update = now;
8216 while ((se = READ_ONCE(cfs_rq->h_load_next)) != NULL) {
8217 load = cfs_rq->h_load;
8218 load = div64_ul(load * se->avg.load_avg,
8219 cfs_rq_load_avg(cfs_rq) + 1);
8220 cfs_rq = group_cfs_rq(se);
8221 cfs_rq->h_load = load;
8222 cfs_rq->last_h_load_update = now;
8226 static unsigned long task_h_load(struct task_struct *p)
8228 struct cfs_rq *cfs_rq = task_cfs_rq(p);
8230 update_cfs_rq_h_load(cfs_rq);
8231 return div64_ul(p->se.avg.load_avg * cfs_rq->h_load,
8232 cfs_rq_load_avg(cfs_rq) + 1);
8235 static bool __update_blocked_fair(struct rq *rq, bool *done)
8237 struct cfs_rq *cfs_rq = &rq->cfs;
8240 decayed = update_cfs_rq_load_avg(cfs_rq_clock_pelt(cfs_rq), cfs_rq);
8241 if (cfs_rq_has_blocked(cfs_rq))
8247 static unsigned long task_h_load(struct task_struct *p)
8249 return p->se.avg.load_avg;
8253 static void update_blocked_averages(int cpu)
8255 bool decayed = false, done = true;
8256 struct rq *rq = cpu_rq(cpu);
8259 rq_lock_irqsave(rq, &rf);
8260 update_blocked_load_tick(rq);
8261 update_rq_clock(rq);
8263 decayed |= __update_blocked_others(rq, &done);
8264 decayed |= __update_blocked_fair(rq, &done);
8266 update_blocked_load_status(rq, !done);
8268 cpufreq_update_util(rq, 0);
8269 rq_unlock_irqrestore(rq, &rf);
8272 /********** Helpers for find_busiest_group ************************/
8275 * sg_lb_stats - stats of a sched_group required for load_balancing
8277 struct sg_lb_stats {
8278 unsigned long avg_load; /*Avg load across the CPUs of the group */
8279 unsigned long group_load; /* Total load over the CPUs of the group */
8280 unsigned long group_capacity;
8281 unsigned long group_util; /* Total utilization over the CPUs of the group */
8282 unsigned long group_runnable; /* Total runnable time over the CPUs of the group */
8283 unsigned int sum_nr_running; /* Nr of tasks running in the group */
8284 unsigned int sum_h_nr_running; /* Nr of CFS tasks running in the group */
8285 unsigned int idle_cpus;
8286 unsigned int group_weight;
8287 enum group_type group_type;
8288 unsigned int group_asym_packing; /* Tasks should be moved to preferred CPU */
8289 unsigned long group_misfit_task_load; /* A CPU has a task too big for its capacity */
8290 #ifdef CONFIG_NUMA_BALANCING
8291 unsigned int nr_numa_running;
8292 unsigned int nr_preferred_running;
8297 * sd_lb_stats - Structure to store the statistics of a sched_domain
8298 * during load balancing.
8300 struct sd_lb_stats {
8301 struct sched_group *busiest; /* Busiest group in this sd */
8302 struct sched_group *local; /* Local group in this sd */
8303 unsigned long total_load; /* Total load of all groups in sd */
8304 unsigned long total_capacity; /* Total capacity of all groups in sd */
8305 unsigned long avg_load; /* Average load across all groups in sd */
8306 unsigned int prefer_sibling; /* tasks should go to sibling first */
8308 struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
8309 struct sg_lb_stats local_stat; /* Statistics of the local group */
8312 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
8315 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
8316 * local_stat because update_sg_lb_stats() does a full clear/assignment.
8317 * We must however set busiest_stat::group_type and
8318 * busiest_stat::idle_cpus to the worst busiest group because
8319 * update_sd_pick_busiest() reads these before assignment.
8321 *sds = (struct sd_lb_stats){
8325 .total_capacity = 0UL,
8327 .idle_cpus = UINT_MAX,
8328 .group_type = group_has_spare,
8333 static unsigned long scale_rt_capacity(int cpu)
8335 struct rq *rq = cpu_rq(cpu);
8336 unsigned long max = arch_scale_cpu_capacity(cpu);
8337 unsigned long used, free;
8340 irq = cpu_util_irq(rq);
8342 if (unlikely(irq >= max))
8346 * avg_rt.util_avg and avg_dl.util_avg track binary signals
8347 * (running and not running) with weights 0 and 1024 respectively.
8348 * avg_thermal.load_avg tracks thermal pressure and the weighted
8349 * average uses the actual delta max capacity(load).
8351 used = READ_ONCE(rq->avg_rt.util_avg);
8352 used += READ_ONCE(rq->avg_dl.util_avg);
8353 used += thermal_load_avg(rq);
8355 if (unlikely(used >= max))
8360 return scale_irq_capacity(free, irq, max);
8363 static void update_cpu_capacity(struct sched_domain *sd, int cpu)
8365 unsigned long capacity = scale_rt_capacity(cpu);
8366 struct sched_group *sdg = sd->groups;
8368 cpu_rq(cpu)->cpu_capacity_orig = arch_scale_cpu_capacity(cpu);
8373 cpu_rq(cpu)->cpu_capacity = capacity;
8374 trace_sched_cpu_capacity_tp(cpu_rq(cpu));
8376 sdg->sgc->capacity = capacity;
8377 sdg->sgc->min_capacity = capacity;
8378 sdg->sgc->max_capacity = capacity;
8381 void update_group_capacity(struct sched_domain *sd, int cpu)
8383 struct sched_domain *child = sd->child;
8384 struct sched_group *group, *sdg = sd->groups;
8385 unsigned long capacity, min_capacity, max_capacity;
8386 unsigned long interval;
8388 interval = msecs_to_jiffies(sd->balance_interval);
8389 interval = clamp(interval, 1UL, max_load_balance_interval);
8390 sdg->sgc->next_update = jiffies + interval;
8393 update_cpu_capacity(sd, cpu);
8398 min_capacity = ULONG_MAX;
8401 if (child->flags & SD_OVERLAP) {
8403 * SD_OVERLAP domains cannot assume that child groups
8404 * span the current group.
8407 for_each_cpu(cpu, sched_group_span(sdg)) {
8408 unsigned long cpu_cap = capacity_of(cpu);
8410 capacity += cpu_cap;
8411 min_capacity = min(cpu_cap, min_capacity);
8412 max_capacity = max(cpu_cap, max_capacity);
8416 * !SD_OVERLAP domains can assume that child groups
8417 * span the current group.
8420 group = child->groups;
8422 struct sched_group_capacity *sgc = group->sgc;
8424 capacity += sgc->capacity;
8425 min_capacity = min(sgc->min_capacity, min_capacity);
8426 max_capacity = max(sgc->max_capacity, max_capacity);
8427 group = group->next;
8428 } while (group != child->groups);
8431 sdg->sgc->capacity = capacity;
8432 sdg->sgc->min_capacity = min_capacity;
8433 sdg->sgc->max_capacity = max_capacity;
8437 * Check whether the capacity of the rq has been noticeably reduced by side
8438 * activity. The imbalance_pct is used for the threshold.
8439 * Return true is the capacity is reduced
8442 check_cpu_capacity(struct rq *rq, struct sched_domain *sd)
8444 return ((rq->cpu_capacity * sd->imbalance_pct) <
8445 (rq->cpu_capacity_orig * 100));
8449 * Check whether a rq has a misfit task and if it looks like we can actually
8450 * help that task: we can migrate the task to a CPU of higher capacity, or
8451 * the task's current CPU is heavily pressured.
8453 static inline int check_misfit_status(struct rq *rq, struct sched_domain *sd)
8455 return rq->misfit_task_load &&
8456 (rq->cpu_capacity_orig < rq->rd->max_cpu_capacity ||
8457 check_cpu_capacity(rq, sd));
8461 * Group imbalance indicates (and tries to solve) the problem where balancing
8462 * groups is inadequate due to ->cpus_ptr constraints.
8464 * Imagine a situation of two groups of 4 CPUs each and 4 tasks each with a
8465 * cpumask covering 1 CPU of the first group and 3 CPUs of the second group.
8468 * { 0 1 2 3 } { 4 5 6 7 }
8471 * If we were to balance group-wise we'd place two tasks in the first group and
8472 * two tasks in the second group. Clearly this is undesired as it will overload
8473 * cpu 3 and leave one of the CPUs in the second group unused.
8475 * The current solution to this issue is detecting the skew in the first group
8476 * by noticing the lower domain failed to reach balance and had difficulty
8477 * moving tasks due to affinity constraints.
8479 * When this is so detected; this group becomes a candidate for busiest; see
8480 * update_sd_pick_busiest(). And calculate_imbalance() and
8481 * find_busiest_group() avoid some of the usual balance conditions to allow it
8482 * to create an effective group imbalance.
8484 * This is a somewhat tricky proposition since the next run might not find the
8485 * group imbalance and decide the groups need to be balanced again. A most
8486 * subtle and fragile situation.
8489 static inline int sg_imbalanced(struct sched_group *group)
8491 return group->sgc->imbalance;
8495 * group_has_capacity returns true if the group has spare capacity that could
8496 * be used by some tasks.
8497 * We consider that a group has spare capacity if the * number of task is
8498 * smaller than the number of CPUs or if the utilization is lower than the
8499 * available capacity for CFS tasks.
8500 * For the latter, we use a threshold to stabilize the state, to take into
8501 * account the variance of the tasks' load and to return true if the available
8502 * capacity in meaningful for the load balancer.
8503 * As an example, an available capacity of 1% can appear but it doesn't make
8504 * any benefit for the load balance.
8507 group_has_capacity(unsigned int imbalance_pct, struct sg_lb_stats *sgs)
8509 if (sgs->sum_nr_running < sgs->group_weight)
8512 if ((sgs->group_capacity * imbalance_pct) <
8513 (sgs->group_runnable * 100))
8516 if ((sgs->group_capacity * 100) >
8517 (sgs->group_util * imbalance_pct))
8524 * group_is_overloaded returns true if the group has more tasks than it can
8526 * group_is_overloaded is not equals to !group_has_capacity because a group
8527 * with the exact right number of tasks, has no more spare capacity but is not
8528 * overloaded so both group_has_capacity and group_is_overloaded return
8532 group_is_overloaded(unsigned int imbalance_pct, struct sg_lb_stats *sgs)
8534 if (sgs->sum_nr_running <= sgs->group_weight)
8537 if ((sgs->group_capacity * 100) <
8538 (sgs->group_util * imbalance_pct))
8541 if ((sgs->group_capacity * imbalance_pct) <
8542 (sgs->group_runnable * 100))
8549 group_type group_classify(unsigned int imbalance_pct,
8550 struct sched_group *group,
8551 struct sg_lb_stats *sgs)
8553 if (group_is_overloaded(imbalance_pct, sgs))
8554 return group_overloaded;
8556 if (sg_imbalanced(group))
8557 return group_imbalanced;
8559 if (sgs->group_asym_packing)
8560 return group_asym_packing;
8562 if (sgs->group_misfit_task_load)
8563 return group_misfit_task;
8565 if (!group_has_capacity(imbalance_pct, sgs))
8566 return group_fully_busy;
8568 return group_has_spare;
8572 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
8573 * @env: The load balancing environment.
8574 * @group: sched_group whose statistics are to be updated.
8575 * @sgs: variable to hold the statistics for this group.
8576 * @sg_status: Holds flag indicating the status of the sched_group
8578 static inline void update_sg_lb_stats(struct lb_env *env,
8579 struct sched_group *group,
8580 struct sg_lb_stats *sgs,
8583 int i, nr_running, local_group;
8585 memset(sgs, 0, sizeof(*sgs));
8587 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_span(group));
8589 for_each_cpu_and(i, sched_group_span(group), env->cpus) {
8590 struct rq *rq = cpu_rq(i);
8592 sgs->group_load += cpu_load(rq);
8593 sgs->group_util += cpu_util(i);
8594 sgs->group_runnable += cpu_runnable(rq);
8595 sgs->sum_h_nr_running += rq->cfs.h_nr_running;
8597 nr_running = rq->nr_running;
8598 sgs->sum_nr_running += nr_running;
8601 *sg_status |= SG_OVERLOAD;
8603 if (cpu_overutilized(i))
8604 *sg_status |= SG_OVERUTILIZED;
8606 #ifdef CONFIG_NUMA_BALANCING
8607 sgs->nr_numa_running += rq->nr_numa_running;
8608 sgs->nr_preferred_running += rq->nr_preferred_running;
8611 * No need to call idle_cpu() if nr_running is not 0
8613 if (!nr_running && idle_cpu(i)) {
8615 /* Idle cpu can't have misfit task */
8622 /* Check for a misfit task on the cpu */
8623 if (env->sd->flags & SD_ASYM_CPUCAPACITY &&
8624 sgs->group_misfit_task_load < rq->misfit_task_load) {
8625 sgs->group_misfit_task_load = rq->misfit_task_load;
8626 *sg_status |= SG_OVERLOAD;
8630 /* Check if dst CPU is idle and preferred to this group */
8631 if (env->sd->flags & SD_ASYM_PACKING &&
8632 env->idle != CPU_NOT_IDLE &&
8633 sgs->sum_h_nr_running &&
8634 sched_asym_prefer(env->dst_cpu, group->asym_prefer_cpu)) {
8635 sgs->group_asym_packing = 1;
8638 sgs->group_capacity = group->sgc->capacity;
8640 sgs->group_weight = group->group_weight;
8642 sgs->group_type = group_classify(env->sd->imbalance_pct, group, sgs);
8644 /* Computing avg_load makes sense only when group is overloaded */
8645 if (sgs->group_type == group_overloaded)
8646 sgs->avg_load = (sgs->group_load * SCHED_CAPACITY_SCALE) /
8647 sgs->group_capacity;
8651 * update_sd_pick_busiest - return 1 on busiest group
8652 * @env: The load balancing environment.
8653 * @sds: sched_domain statistics
8654 * @sg: sched_group candidate to be checked for being the busiest
8655 * @sgs: sched_group statistics
8657 * Determine if @sg is a busier group than the previously selected
8660 * Return: %true if @sg is a busier group than the previously selected
8661 * busiest group. %false otherwise.
8663 static bool update_sd_pick_busiest(struct lb_env *env,
8664 struct sd_lb_stats *sds,
8665 struct sched_group *sg,
8666 struct sg_lb_stats *sgs)
8668 struct sg_lb_stats *busiest = &sds->busiest_stat;
8670 /* Make sure that there is at least one task to pull */
8671 if (!sgs->sum_h_nr_running)
8675 * Don't try to pull misfit tasks we can't help.
8676 * We can use max_capacity here as reduction in capacity on some
8677 * CPUs in the group should either be possible to resolve
8678 * internally or be covered by avg_load imbalance (eventually).
8680 if (sgs->group_type == group_misfit_task &&
8681 (!capacity_greater(capacity_of(env->dst_cpu), sg->sgc->max_capacity) ||
8682 sds->local_stat.group_type != group_has_spare))
8685 if (sgs->group_type > busiest->group_type)
8688 if (sgs->group_type < busiest->group_type)
8692 * The candidate and the current busiest group are the same type of
8693 * group. Let check which one is the busiest according to the type.
8696 switch (sgs->group_type) {
8697 case group_overloaded:
8698 /* Select the overloaded group with highest avg_load. */
8699 if (sgs->avg_load <= busiest->avg_load)
8703 case group_imbalanced:
8705 * Select the 1st imbalanced group as we don't have any way to
8706 * choose one more than another.
8710 case group_asym_packing:
8711 /* Prefer to move from lowest priority CPU's work */
8712 if (sched_asym_prefer(sg->asym_prefer_cpu, sds->busiest->asym_prefer_cpu))
8716 case group_misfit_task:
8718 * If we have more than one misfit sg go with the biggest
8721 if (sgs->group_misfit_task_load < busiest->group_misfit_task_load)
8725 case group_fully_busy:
8727 * Select the fully busy group with highest avg_load. In
8728 * theory, there is no need to pull task from such kind of
8729 * group because tasks have all compute capacity that they need
8730 * but we can still improve the overall throughput by reducing
8731 * contention when accessing shared HW resources.
8733 * XXX for now avg_load is not computed and always 0 so we
8734 * select the 1st one.
8736 if (sgs->avg_load <= busiest->avg_load)
8740 case group_has_spare:
8742 * Select not overloaded group with lowest number of idle cpus
8743 * and highest number of running tasks. We could also compare
8744 * the spare capacity which is more stable but it can end up
8745 * that the group has less spare capacity but finally more idle
8746 * CPUs which means less opportunity to pull tasks.
8748 if (sgs->idle_cpus > busiest->idle_cpus)
8750 else if ((sgs->idle_cpus == busiest->idle_cpus) &&
8751 (sgs->sum_nr_running <= busiest->sum_nr_running))
8758 * Candidate sg has no more than one task per CPU and has higher
8759 * per-CPU capacity. Migrating tasks to less capable CPUs may harm
8760 * throughput. Maximize throughput, power/energy consequences are not
8763 if ((env->sd->flags & SD_ASYM_CPUCAPACITY) &&
8764 (sgs->group_type <= group_fully_busy) &&
8765 (capacity_greater(sg->sgc->min_capacity, capacity_of(env->dst_cpu))))
8771 #ifdef CONFIG_NUMA_BALANCING
8772 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
8774 if (sgs->sum_h_nr_running > sgs->nr_numa_running)
8776 if (sgs->sum_h_nr_running > sgs->nr_preferred_running)
8781 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
8783 if (rq->nr_running > rq->nr_numa_running)
8785 if (rq->nr_running > rq->nr_preferred_running)
8790 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
8795 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
8799 #endif /* CONFIG_NUMA_BALANCING */
8805 * task_running_on_cpu - return 1 if @p is running on @cpu.
8808 static unsigned int task_running_on_cpu(int cpu, struct task_struct *p)
8810 /* Task has no contribution or is new */
8811 if (cpu != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time))
8814 if (task_on_rq_queued(p))
8821 * idle_cpu_without - would a given CPU be idle without p ?
8822 * @cpu: the processor on which idleness is tested.
8823 * @p: task which should be ignored.
8825 * Return: 1 if the CPU would be idle. 0 otherwise.
8827 static int idle_cpu_without(int cpu, struct task_struct *p)
8829 struct rq *rq = cpu_rq(cpu);
8831 if (rq->curr != rq->idle && rq->curr != p)
8835 * rq->nr_running can't be used but an updated version without the
8836 * impact of p on cpu must be used instead. The updated nr_running
8837 * be computed and tested before calling idle_cpu_without().
8841 if (rq->ttwu_pending)
8849 * update_sg_wakeup_stats - Update sched_group's statistics for wakeup.
8850 * @sd: The sched_domain level to look for idlest group.
8851 * @group: sched_group whose statistics are to be updated.
8852 * @sgs: variable to hold the statistics for this group.
8853 * @p: The task for which we look for the idlest group/CPU.
8855 static inline void update_sg_wakeup_stats(struct sched_domain *sd,
8856 struct sched_group *group,
8857 struct sg_lb_stats *sgs,
8858 struct task_struct *p)
8862 memset(sgs, 0, sizeof(*sgs));
8864 for_each_cpu(i, sched_group_span(group)) {
8865 struct rq *rq = cpu_rq(i);
8868 sgs->group_load += cpu_load_without(rq, p);
8869 sgs->group_util += cpu_util_without(i, p);
8870 sgs->group_runnable += cpu_runnable_without(rq, p);
8871 local = task_running_on_cpu(i, p);
8872 sgs->sum_h_nr_running += rq->cfs.h_nr_running - local;
8874 nr_running = rq->nr_running - local;
8875 sgs->sum_nr_running += nr_running;
8878 * No need to call idle_cpu_without() if nr_running is not 0
8880 if (!nr_running && idle_cpu_without(i, p))
8885 /* Check if task fits in the group */
8886 if (sd->flags & SD_ASYM_CPUCAPACITY &&
8887 !task_fits_capacity(p, group->sgc->max_capacity)) {
8888 sgs->group_misfit_task_load = 1;
8891 sgs->group_capacity = group->sgc->capacity;
8893 sgs->group_weight = group->group_weight;
8895 sgs->group_type = group_classify(sd->imbalance_pct, group, sgs);
8898 * Computing avg_load makes sense only when group is fully busy or
8901 if (sgs->group_type == group_fully_busy ||
8902 sgs->group_type == group_overloaded)
8903 sgs->avg_load = (sgs->group_load * SCHED_CAPACITY_SCALE) /
8904 sgs->group_capacity;
8907 static bool update_pick_idlest(struct sched_group *idlest,
8908 struct sg_lb_stats *idlest_sgs,
8909 struct sched_group *group,
8910 struct sg_lb_stats *sgs)
8912 if (sgs->group_type < idlest_sgs->group_type)
8915 if (sgs->group_type > idlest_sgs->group_type)
8919 * The candidate and the current idlest group are the same type of
8920 * group. Let check which one is the idlest according to the type.
8923 switch (sgs->group_type) {
8924 case group_overloaded:
8925 case group_fully_busy:
8926 /* Select the group with lowest avg_load. */
8927 if (idlest_sgs->avg_load <= sgs->avg_load)
8931 case group_imbalanced:
8932 case group_asym_packing:
8933 /* Those types are not used in the slow wakeup path */
8936 case group_misfit_task:
8937 /* Select group with the highest max capacity */
8938 if (idlest->sgc->max_capacity >= group->sgc->max_capacity)
8942 case group_has_spare:
8943 /* Select group with most idle CPUs */
8944 if (idlest_sgs->idle_cpus > sgs->idle_cpus)
8947 /* Select group with lowest group_util */
8948 if (idlest_sgs->idle_cpus == sgs->idle_cpus &&
8949 idlest_sgs->group_util <= sgs->group_util)
8959 * Allow a NUMA imbalance if busy CPUs is less than 25% of the domain.
8960 * This is an approximation as the number of running tasks may not be
8961 * related to the number of busy CPUs due to sched_setaffinity.
8963 static inline bool allow_numa_imbalance(int dst_running, int dst_weight)
8965 return (dst_running < (dst_weight >> 2));
8969 * find_idlest_group() finds and returns the least busy CPU group within the
8972 * Assumes p is allowed on at least one CPU in sd.
8974 static struct sched_group *
8975 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
8977 struct sched_group *idlest = NULL, *local = NULL, *group = sd->groups;
8978 struct sg_lb_stats local_sgs, tmp_sgs;
8979 struct sg_lb_stats *sgs;
8980 unsigned long imbalance;
8981 struct sg_lb_stats idlest_sgs = {
8982 .avg_load = UINT_MAX,
8983 .group_type = group_overloaded,
8989 /* Skip over this group if it has no CPUs allowed */
8990 if (!cpumask_intersects(sched_group_span(group),
8994 /* Skip over this group if no cookie matched */
8995 if (!sched_group_cookie_match(cpu_rq(this_cpu), p, group))
8998 local_group = cpumask_test_cpu(this_cpu,
8999 sched_group_span(group));
9008 update_sg_wakeup_stats(sd, group, sgs, p);
9010 if (!local_group && update_pick_idlest(idlest, &idlest_sgs, group, sgs)) {
9015 } while (group = group->next, group != sd->groups);
9018 /* There is no idlest group to push tasks to */
9022 /* The local group has been skipped because of CPU affinity */
9027 * If the local group is idler than the selected idlest group
9028 * don't try and push the task.
9030 if (local_sgs.group_type < idlest_sgs.group_type)
9034 * If the local group is busier than the selected idlest group
9035 * try and push the task.
9037 if (local_sgs.group_type > idlest_sgs.group_type)
9040 switch (local_sgs.group_type) {
9041 case group_overloaded:
9042 case group_fully_busy:
9044 /* Calculate allowed imbalance based on load */
9045 imbalance = scale_load_down(NICE_0_LOAD) *
9046 (sd->imbalance_pct-100) / 100;
9049 * When comparing groups across NUMA domains, it's possible for
9050 * the local domain to be very lightly loaded relative to the
9051 * remote domains but "imbalance" skews the comparison making
9052 * remote CPUs look much more favourable. When considering
9053 * cross-domain, add imbalance to the load on the remote node
9054 * and consider staying local.
9057 if ((sd->flags & SD_NUMA) &&
9058 ((idlest_sgs.avg_load + imbalance) >= local_sgs.avg_load))
9062 * If the local group is less loaded than the selected
9063 * idlest group don't try and push any tasks.
9065 if (idlest_sgs.avg_load >= (local_sgs.avg_load + imbalance))
9068 if (100 * local_sgs.avg_load <= sd->imbalance_pct * idlest_sgs.avg_load)
9072 case group_imbalanced:
9073 case group_asym_packing:
9074 /* Those type are not used in the slow wakeup path */
9077 case group_misfit_task:
9078 /* Select group with the highest max capacity */
9079 if (local->sgc->max_capacity >= idlest->sgc->max_capacity)
9083 case group_has_spare:
9084 if (sd->flags & SD_NUMA) {
9085 #ifdef CONFIG_NUMA_BALANCING
9088 * If there is spare capacity at NUMA, try to select
9089 * the preferred node
9091 if (cpu_to_node(this_cpu) == p->numa_preferred_nid)
9094 idlest_cpu = cpumask_first(sched_group_span(idlest));
9095 if (cpu_to_node(idlest_cpu) == p->numa_preferred_nid)
9099 * Otherwise, keep the task on this node to stay close
9100 * its wakeup source and improve locality. If there is
9101 * a real need of migration, periodic load balance will
9104 if (allow_numa_imbalance(local_sgs.sum_nr_running, sd->span_weight))
9109 * Select group with highest number of idle CPUs. We could also
9110 * compare the utilization which is more stable but it can end
9111 * up that the group has less spare capacity but finally more
9112 * idle CPUs which means more opportunity to run task.
9114 if (local_sgs.idle_cpus >= idlest_sgs.idle_cpus)
9123 * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
9124 * @env: The load balancing environment.
9125 * @sds: variable to hold the statistics for this sched_domain.
9128 static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
9130 struct sched_domain *child = env->sd->child;
9131 struct sched_group *sg = env->sd->groups;
9132 struct sg_lb_stats *local = &sds->local_stat;
9133 struct sg_lb_stats tmp_sgs;
9137 struct sg_lb_stats *sgs = &tmp_sgs;
9140 local_group = cpumask_test_cpu(env->dst_cpu, sched_group_span(sg));
9145 if (env->idle != CPU_NEWLY_IDLE ||
9146 time_after_eq(jiffies, sg->sgc->next_update))
9147 update_group_capacity(env->sd, env->dst_cpu);
9150 update_sg_lb_stats(env, sg, sgs, &sg_status);
9156 if (update_sd_pick_busiest(env, sds, sg, sgs)) {
9158 sds->busiest_stat = *sgs;
9162 /* Now, start updating sd_lb_stats */
9163 sds->total_load += sgs->group_load;
9164 sds->total_capacity += sgs->group_capacity;
9167 } while (sg != env->sd->groups);
9169 /* Tag domain that child domain prefers tasks go to siblings first */
9170 sds->prefer_sibling = child && child->flags & SD_PREFER_SIBLING;
9173 if (env->sd->flags & SD_NUMA)
9174 env->fbq_type = fbq_classify_group(&sds->busiest_stat);
9176 if (!env->sd->parent) {
9177 struct root_domain *rd = env->dst_rq->rd;
9179 /* update overload indicator if we are at root domain */
9180 WRITE_ONCE(rd->overload, sg_status & SG_OVERLOAD);
9182 /* Update over-utilization (tipping point, U >= 0) indicator */
9183 WRITE_ONCE(rd->overutilized, sg_status & SG_OVERUTILIZED);
9184 trace_sched_overutilized_tp(rd, sg_status & SG_OVERUTILIZED);
9185 } else if (sg_status & SG_OVERUTILIZED) {
9186 struct root_domain *rd = env->dst_rq->rd;
9188 WRITE_ONCE(rd->overutilized, SG_OVERUTILIZED);
9189 trace_sched_overutilized_tp(rd, SG_OVERUTILIZED);
9193 #define NUMA_IMBALANCE_MIN 2
9195 static inline long adjust_numa_imbalance(int imbalance,
9196 int dst_running, int dst_weight)
9198 if (!allow_numa_imbalance(dst_running, dst_weight))
9202 * Allow a small imbalance based on a simple pair of communicating
9203 * tasks that remain local when the destination is lightly loaded.
9205 if (imbalance <= NUMA_IMBALANCE_MIN)
9212 * calculate_imbalance - Calculate the amount of imbalance present within the
9213 * groups of a given sched_domain during load balance.
9214 * @env: load balance environment
9215 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
9217 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
9219 struct sg_lb_stats *local, *busiest;
9221 local = &sds->local_stat;
9222 busiest = &sds->busiest_stat;
9224 if (busiest->group_type == group_misfit_task) {
9225 /* Set imbalance to allow misfit tasks to be balanced. */
9226 env->migration_type = migrate_misfit;
9231 if (busiest->group_type == group_asym_packing) {
9233 * In case of asym capacity, we will try to migrate all load to
9234 * the preferred CPU.
9236 env->migration_type = migrate_task;
9237 env->imbalance = busiest->sum_h_nr_running;
9241 if (busiest->group_type == group_imbalanced) {
9243 * In the group_imb case we cannot rely on group-wide averages
9244 * to ensure CPU-load equilibrium, try to move any task to fix
9245 * the imbalance. The next load balance will take care of
9246 * balancing back the system.
9248 env->migration_type = migrate_task;
9254 * Try to use spare capacity of local group without overloading it or
9257 if (local->group_type == group_has_spare) {
9258 if ((busiest->group_type > group_fully_busy) &&
9259 !(env->sd->flags & SD_SHARE_PKG_RESOURCES)) {
9261 * If busiest is overloaded, try to fill spare
9262 * capacity. This might end up creating spare capacity
9263 * in busiest or busiest still being overloaded but
9264 * there is no simple way to directly compute the
9265 * amount of load to migrate in order to balance the
9268 env->migration_type = migrate_util;
9269 env->imbalance = max(local->group_capacity, local->group_util) -
9273 * In some cases, the group's utilization is max or even
9274 * higher than capacity because of migrations but the
9275 * local CPU is (newly) idle. There is at least one
9276 * waiting task in this overloaded busiest group. Let's
9279 if (env->idle != CPU_NOT_IDLE && env->imbalance == 0) {
9280 env->migration_type = migrate_task;
9287 if (busiest->group_weight == 1 || sds->prefer_sibling) {
9288 unsigned int nr_diff = busiest->sum_nr_running;
9290 * When prefer sibling, evenly spread running tasks on
9293 env->migration_type = migrate_task;
9294 lsub_positive(&nr_diff, local->sum_nr_running);
9295 env->imbalance = nr_diff >> 1;
9299 * If there is no overload, we just want to even the number of
9302 env->migration_type = migrate_task;
9303 env->imbalance = max_t(long, 0, (local->idle_cpus -
9304 busiest->idle_cpus) >> 1);
9307 /* Consider allowing a small imbalance between NUMA groups */
9308 if (env->sd->flags & SD_NUMA) {
9309 env->imbalance = adjust_numa_imbalance(env->imbalance,
9310 busiest->sum_nr_running, busiest->group_weight);
9317 * Local is fully busy but has to take more load to relieve the
9320 if (local->group_type < group_overloaded) {
9322 * Local will become overloaded so the avg_load metrics are
9326 local->avg_load = (local->group_load * SCHED_CAPACITY_SCALE) /
9327 local->group_capacity;
9329 sds->avg_load = (sds->total_load * SCHED_CAPACITY_SCALE) /
9330 sds->total_capacity;
9332 * If the local group is more loaded than the selected
9333 * busiest group don't try to pull any tasks.
9335 if (local->avg_load >= busiest->avg_load) {
9342 * Both group are or will become overloaded and we're trying to get all
9343 * the CPUs to the average_load, so we don't want to push ourselves
9344 * above the average load, nor do we wish to reduce the max loaded CPU
9345 * below the average load. At the same time, we also don't want to
9346 * reduce the group load below the group capacity. Thus we look for
9347 * the minimum possible imbalance.
9349 env->migration_type = migrate_load;
9350 env->imbalance = min(
9351 (busiest->avg_load - sds->avg_load) * busiest->group_capacity,
9352 (sds->avg_load - local->avg_load) * local->group_capacity
9353 ) / SCHED_CAPACITY_SCALE;
9356 /******* find_busiest_group() helpers end here *********************/
9359 * Decision matrix according to the local and busiest group type:
9361 * busiest \ local has_spare fully_busy misfit asym imbalanced overloaded
9362 * has_spare nr_idle balanced N/A N/A balanced balanced
9363 * fully_busy nr_idle nr_idle N/A N/A balanced balanced
9364 * misfit_task force N/A N/A N/A force force
9365 * asym_packing force force N/A N/A force force
9366 * imbalanced force force N/A N/A force force
9367 * overloaded force force N/A N/A force avg_load
9369 * N/A : Not Applicable because already filtered while updating
9371 * balanced : The system is balanced for these 2 groups.
9372 * force : Calculate the imbalance as load migration is probably needed.
9373 * avg_load : Only if imbalance is significant enough.
9374 * nr_idle : dst_cpu is not busy and the number of idle CPUs is quite
9375 * different in groups.
9379 * find_busiest_group - Returns the busiest group within the sched_domain
9380 * if there is an imbalance.
9382 * Also calculates the amount of runnable load which should be moved
9383 * to restore balance.
9385 * @env: The load balancing environment.
9387 * Return: - The busiest group if imbalance exists.
9389 static struct sched_group *find_busiest_group(struct lb_env *env)
9391 struct sg_lb_stats *local, *busiest;
9392 struct sd_lb_stats sds;
9394 init_sd_lb_stats(&sds);
9397 * Compute the various statistics relevant for load balancing at
9400 update_sd_lb_stats(env, &sds);
9402 if (sched_energy_enabled()) {
9403 struct root_domain *rd = env->dst_rq->rd;
9405 if (rcu_dereference(rd->pd) && !READ_ONCE(rd->overutilized))
9409 local = &sds.local_stat;
9410 busiest = &sds.busiest_stat;
9412 /* There is no busy sibling group to pull tasks from */
9416 /* Misfit tasks should be dealt with regardless of the avg load */
9417 if (busiest->group_type == group_misfit_task)
9420 /* ASYM feature bypasses nice load balance check */
9421 if (busiest->group_type == group_asym_packing)
9425 * If the busiest group is imbalanced the below checks don't
9426 * work because they assume all things are equal, which typically
9427 * isn't true due to cpus_ptr constraints and the like.
9429 if (busiest->group_type == group_imbalanced)
9433 * If the local group is busier than the selected busiest group
9434 * don't try and pull any tasks.
9436 if (local->group_type > busiest->group_type)
9440 * When groups are overloaded, use the avg_load to ensure fairness
9443 if (local->group_type == group_overloaded) {
9445 * If the local group is more loaded than the selected
9446 * busiest group don't try to pull any tasks.
9448 if (local->avg_load >= busiest->avg_load)
9451 /* XXX broken for overlapping NUMA groups */
9452 sds.avg_load = (sds.total_load * SCHED_CAPACITY_SCALE) /
9456 * Don't pull any tasks if this group is already above the
9457 * domain average load.
9459 if (local->avg_load >= sds.avg_load)
9463 * If the busiest group is more loaded, use imbalance_pct to be
9466 if (100 * busiest->avg_load <=
9467 env->sd->imbalance_pct * local->avg_load)
9471 /* Try to move all excess tasks to child's sibling domain */
9472 if (sds.prefer_sibling && local->group_type == group_has_spare &&
9473 busiest->sum_nr_running > local->sum_nr_running + 1)
9476 if (busiest->group_type != group_overloaded) {
9477 if (env->idle == CPU_NOT_IDLE)
9479 * If the busiest group is not overloaded (and as a
9480 * result the local one too) but this CPU is already
9481 * busy, let another idle CPU try to pull task.
9485 if (busiest->group_weight > 1 &&
9486 local->idle_cpus <= (busiest->idle_cpus + 1))
9488 * If the busiest group is not overloaded
9489 * and there is no imbalance between this and busiest
9490 * group wrt idle CPUs, it is balanced. The imbalance
9491 * becomes significant if the diff is greater than 1
9492 * otherwise we might end up to just move the imbalance
9493 * on another group. Of course this applies only if
9494 * there is more than 1 CPU per group.
9498 if (busiest->sum_h_nr_running == 1)
9500 * busiest doesn't have any tasks waiting to run
9506 /* Looks like there is an imbalance. Compute it */
9507 calculate_imbalance(env, &sds);
9508 return env->imbalance ? sds.busiest : NULL;
9516 * find_busiest_queue - find the busiest runqueue among the CPUs in the group.
9518 static struct rq *find_busiest_queue(struct lb_env *env,
9519 struct sched_group *group)
9521 struct rq *busiest = NULL, *rq;
9522 unsigned long busiest_util = 0, busiest_load = 0, busiest_capacity = 1;
9523 unsigned int busiest_nr = 0;
9526 for_each_cpu_and(i, sched_group_span(group), env->cpus) {
9527 unsigned long capacity, load, util;
9528 unsigned int nr_running;
9532 rt = fbq_classify_rq(rq);
9535 * We classify groups/runqueues into three groups:
9536 * - regular: there are !numa tasks
9537 * - remote: there are numa tasks that run on the 'wrong' node
9538 * - all: there is no distinction
9540 * In order to avoid migrating ideally placed numa tasks,
9541 * ignore those when there's better options.
9543 * If we ignore the actual busiest queue to migrate another
9544 * task, the next balance pass can still reduce the busiest
9545 * queue by moving tasks around inside the node.
9547 * If we cannot move enough load due to this classification
9548 * the next pass will adjust the group classification and
9549 * allow migration of more tasks.
9551 * Both cases only affect the total convergence complexity.
9553 if (rt > env->fbq_type)
9556 nr_running = rq->cfs.h_nr_running;
9560 capacity = capacity_of(i);
9563 * For ASYM_CPUCAPACITY domains, don't pick a CPU that could
9564 * eventually lead to active_balancing high->low capacity.
9565 * Higher per-CPU capacity is considered better than balancing
9568 if (env->sd->flags & SD_ASYM_CPUCAPACITY &&
9569 !capacity_greater(capacity_of(env->dst_cpu), capacity) &&
9573 switch (env->migration_type) {
9576 * When comparing with load imbalance, use cpu_load()
9577 * which is not scaled with the CPU capacity.
9579 load = cpu_load(rq);
9581 if (nr_running == 1 && load > env->imbalance &&
9582 !check_cpu_capacity(rq, env->sd))
9586 * For the load comparisons with the other CPUs,
9587 * consider the cpu_load() scaled with the CPU
9588 * capacity, so that the load can be moved away
9589 * from the CPU that is potentially running at a
9592 * Thus we're looking for max(load_i / capacity_i),
9593 * crosswise multiplication to rid ourselves of the
9594 * division works out to:
9595 * load_i * capacity_j > load_j * capacity_i;
9596 * where j is our previous maximum.
9598 if (load * busiest_capacity > busiest_load * capacity) {
9599 busiest_load = load;
9600 busiest_capacity = capacity;
9606 util = cpu_util(cpu_of(rq));
9609 * Don't try to pull utilization from a CPU with one
9610 * running task. Whatever its utilization, we will fail
9613 if (nr_running <= 1)
9616 if (busiest_util < util) {
9617 busiest_util = util;
9623 if (busiest_nr < nr_running) {
9624 busiest_nr = nr_running;
9629 case migrate_misfit:
9631 * For ASYM_CPUCAPACITY domains with misfit tasks we
9632 * simply seek the "biggest" misfit task.
9634 if (rq->misfit_task_load > busiest_load) {
9635 busiest_load = rq->misfit_task_load;
9648 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
9649 * so long as it is large enough.
9651 #define MAX_PINNED_INTERVAL 512
9654 asym_active_balance(struct lb_env *env)
9657 * ASYM_PACKING needs to force migrate tasks from busy but
9658 * lower priority CPUs in order to pack all tasks in the
9659 * highest priority CPUs.
9661 return env->idle != CPU_NOT_IDLE && (env->sd->flags & SD_ASYM_PACKING) &&
9662 sched_asym_prefer(env->dst_cpu, env->src_cpu);
9666 imbalanced_active_balance(struct lb_env *env)
9668 struct sched_domain *sd = env->sd;
9671 * The imbalanced case includes the case of pinned tasks preventing a fair
9672 * distribution of the load on the system but also the even distribution of the
9673 * threads on a system with spare capacity
9675 if ((env->migration_type == migrate_task) &&
9676 (sd->nr_balance_failed > sd->cache_nice_tries+2))
9682 static int need_active_balance(struct lb_env *env)
9684 struct sched_domain *sd = env->sd;
9686 if (asym_active_balance(env))
9689 if (imbalanced_active_balance(env))
9693 * The dst_cpu is idle and the src_cpu CPU has only 1 CFS task.
9694 * It's worth migrating the task if the src_cpu's capacity is reduced
9695 * because of other sched_class or IRQs if more capacity stays
9696 * available on dst_cpu.
9698 if ((env->idle != CPU_NOT_IDLE) &&
9699 (env->src_rq->cfs.h_nr_running == 1)) {
9700 if ((check_cpu_capacity(env->src_rq, sd)) &&
9701 (capacity_of(env->src_cpu)*sd->imbalance_pct < capacity_of(env->dst_cpu)*100))
9705 if (env->migration_type == migrate_misfit)
9711 static int active_load_balance_cpu_stop(void *data);
9713 static int should_we_balance(struct lb_env *env)
9715 struct sched_group *sg = env->sd->groups;
9719 * Ensure the balancing environment is consistent; can happen
9720 * when the softirq triggers 'during' hotplug.
9722 if (!cpumask_test_cpu(env->dst_cpu, env->cpus))
9726 * In the newly idle case, we will allow all the CPUs
9727 * to do the newly idle load balance.
9729 if (env->idle == CPU_NEWLY_IDLE)
9732 /* Try to find first idle CPU */
9733 for_each_cpu_and(cpu, group_balance_mask(sg), env->cpus) {
9737 /* Are we the first idle CPU? */
9738 return cpu == env->dst_cpu;
9741 /* Are we the first CPU of this group ? */
9742 return group_balance_cpu(sg) == env->dst_cpu;
9746 * Check this_cpu to ensure it is balanced within domain. Attempt to move
9747 * tasks if there is an imbalance.
9749 static int load_balance(int this_cpu, struct rq *this_rq,
9750 struct sched_domain *sd, enum cpu_idle_type idle,
9751 int *continue_balancing)
9753 int ld_moved, cur_ld_moved, active_balance = 0;
9754 struct sched_domain *sd_parent = sd->parent;
9755 struct sched_group *group;
9758 struct cpumask *cpus = this_cpu_cpumask_var_ptr(load_balance_mask);
9760 struct lb_env env = {
9762 .dst_cpu = this_cpu,
9764 .dst_grpmask = sched_group_span(sd->groups),
9766 .loop_break = sched_nr_migrate_break,
9769 .tasks = LIST_HEAD_INIT(env.tasks),
9772 cpumask_and(cpus, sched_domain_span(sd), cpu_active_mask);
9774 schedstat_inc(sd->lb_count[idle]);
9777 if (!should_we_balance(&env)) {
9778 *continue_balancing = 0;
9782 group = find_busiest_group(&env);
9784 schedstat_inc(sd->lb_nobusyg[idle]);
9788 busiest = find_busiest_queue(&env, group);
9790 schedstat_inc(sd->lb_nobusyq[idle]);
9794 BUG_ON(busiest == env.dst_rq);
9796 schedstat_add(sd->lb_imbalance[idle], env.imbalance);
9798 env.src_cpu = busiest->cpu;
9799 env.src_rq = busiest;
9802 /* Clear this flag as soon as we find a pullable task */
9803 env.flags |= LBF_ALL_PINNED;
9804 if (busiest->nr_running > 1) {
9806 * Attempt to move tasks. If find_busiest_group has found
9807 * an imbalance but busiest->nr_running <= 1, the group is
9808 * still unbalanced. ld_moved simply stays zero, so it is
9809 * correctly treated as an imbalance.
9811 env.loop_max = min(sysctl_sched_nr_migrate, busiest->nr_running);
9814 rq_lock_irqsave(busiest, &rf);
9815 update_rq_clock(busiest);
9818 * cur_ld_moved - load moved in current iteration
9819 * ld_moved - cumulative load moved across iterations
9821 cur_ld_moved = detach_tasks(&env);
9824 * We've detached some tasks from busiest_rq. Every
9825 * task is masked "TASK_ON_RQ_MIGRATING", so we can safely
9826 * unlock busiest->lock, and we are able to be sure
9827 * that nobody can manipulate the tasks in parallel.
9828 * See task_rq_lock() family for the details.
9831 rq_unlock(busiest, &rf);
9835 ld_moved += cur_ld_moved;
9838 local_irq_restore(rf.flags);
9840 if (env.flags & LBF_NEED_BREAK) {
9841 env.flags &= ~LBF_NEED_BREAK;
9846 * Revisit (affine) tasks on src_cpu that couldn't be moved to
9847 * us and move them to an alternate dst_cpu in our sched_group
9848 * where they can run. The upper limit on how many times we
9849 * iterate on same src_cpu is dependent on number of CPUs in our
9852 * This changes load balance semantics a bit on who can move
9853 * load to a given_cpu. In addition to the given_cpu itself
9854 * (or a ilb_cpu acting on its behalf where given_cpu is
9855 * nohz-idle), we now have balance_cpu in a position to move
9856 * load to given_cpu. In rare situations, this may cause
9857 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
9858 * _independently_ and at _same_ time to move some load to
9859 * given_cpu) causing excess load to be moved to given_cpu.
9860 * This however should not happen so much in practice and
9861 * moreover subsequent load balance cycles should correct the
9862 * excess load moved.
9864 if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
9866 /* Prevent to re-select dst_cpu via env's CPUs */
9867 __cpumask_clear_cpu(env.dst_cpu, env.cpus);
9869 env.dst_rq = cpu_rq(env.new_dst_cpu);
9870 env.dst_cpu = env.new_dst_cpu;
9871 env.flags &= ~LBF_DST_PINNED;
9873 env.loop_break = sched_nr_migrate_break;
9876 * Go back to "more_balance" rather than "redo" since we
9877 * need to continue with same src_cpu.
9883 * We failed to reach balance because of affinity.
9886 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
9888 if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0)
9889 *group_imbalance = 1;
9892 /* All tasks on this runqueue were pinned by CPU affinity */
9893 if (unlikely(env.flags & LBF_ALL_PINNED)) {
9894 __cpumask_clear_cpu(cpu_of(busiest), cpus);
9896 * Attempting to continue load balancing at the current
9897 * sched_domain level only makes sense if there are
9898 * active CPUs remaining as possible busiest CPUs to
9899 * pull load from which are not contained within the
9900 * destination group that is receiving any migrated
9903 if (!cpumask_subset(cpus, env.dst_grpmask)) {
9905 env.loop_break = sched_nr_migrate_break;
9908 goto out_all_pinned;
9913 schedstat_inc(sd->lb_failed[idle]);
9915 * Increment the failure counter only on periodic balance.
9916 * We do not want newidle balance, which can be very
9917 * frequent, pollute the failure counter causing
9918 * excessive cache_hot migrations and active balances.
9920 if (idle != CPU_NEWLY_IDLE)
9921 sd->nr_balance_failed++;
9923 if (need_active_balance(&env)) {
9924 unsigned long flags;
9926 raw_spin_rq_lock_irqsave(busiest, flags);
9929 * Don't kick the active_load_balance_cpu_stop,
9930 * if the curr task on busiest CPU can't be
9931 * moved to this_cpu:
9933 if (!cpumask_test_cpu(this_cpu, busiest->curr->cpus_ptr)) {
9934 raw_spin_rq_unlock_irqrestore(busiest, flags);
9935 goto out_one_pinned;
9938 /* Record that we found at least one task that could run on this_cpu */
9939 env.flags &= ~LBF_ALL_PINNED;
9942 * ->active_balance synchronizes accesses to
9943 * ->active_balance_work. Once set, it's cleared
9944 * only after active load balance is finished.
9946 if (!busiest->active_balance) {
9947 busiest->active_balance = 1;
9948 busiest->push_cpu = this_cpu;
9951 raw_spin_rq_unlock_irqrestore(busiest, flags);
9953 if (active_balance) {
9954 stop_one_cpu_nowait(cpu_of(busiest),
9955 active_load_balance_cpu_stop, busiest,
9956 &busiest->active_balance_work);
9960 sd->nr_balance_failed = 0;
9963 if (likely(!active_balance) || need_active_balance(&env)) {
9964 /* We were unbalanced, so reset the balancing interval */
9965 sd->balance_interval = sd->min_interval;
9972 * We reach balance although we may have faced some affinity
9973 * constraints. Clear the imbalance flag only if other tasks got
9974 * a chance to move and fix the imbalance.
9976 if (sd_parent && !(env.flags & LBF_ALL_PINNED)) {
9977 int *group_imbalance = &sd_parent->groups->sgc->imbalance;
9979 if (*group_imbalance)
9980 *group_imbalance = 0;
9985 * We reach balance because all tasks are pinned at this level so
9986 * we can't migrate them. Let the imbalance flag set so parent level
9987 * can try to migrate them.
9989 schedstat_inc(sd->lb_balanced[idle]);
9991 sd->nr_balance_failed = 0;
9997 * newidle_balance() disregards balance intervals, so we could
9998 * repeatedly reach this code, which would lead to balance_interval
9999 * skyrocketing in a short amount of time. Skip the balance_interval
10000 * increase logic to avoid that.
10002 if (env.idle == CPU_NEWLY_IDLE)
10005 /* tune up the balancing interval */
10006 if ((env.flags & LBF_ALL_PINNED &&
10007 sd->balance_interval < MAX_PINNED_INTERVAL) ||
10008 sd->balance_interval < sd->max_interval)
10009 sd->balance_interval *= 2;
10014 static inline unsigned long
10015 get_sd_balance_interval(struct sched_domain *sd, int cpu_busy)
10017 unsigned long interval = sd->balance_interval;
10020 interval *= sd->busy_factor;
10022 /* scale ms to jiffies */
10023 interval = msecs_to_jiffies(interval);
10026 * Reduce likelihood of busy balancing at higher domains racing with
10027 * balancing at lower domains by preventing their balancing periods
10028 * from being multiples of each other.
10033 interval = clamp(interval, 1UL, max_load_balance_interval);
10039 update_next_balance(struct sched_domain *sd, unsigned long *next_balance)
10041 unsigned long interval, next;
10043 /* used by idle balance, so cpu_busy = 0 */
10044 interval = get_sd_balance_interval(sd, 0);
10045 next = sd->last_balance + interval;
10047 if (time_after(*next_balance, next))
10048 *next_balance = next;
10052 * active_load_balance_cpu_stop is run by the CPU stopper. It pushes
10053 * running tasks off the busiest CPU onto idle CPUs. It requires at
10054 * least 1 task to be running on each physical CPU where possible, and
10055 * avoids physical / logical imbalances.
10057 static int active_load_balance_cpu_stop(void *data)
10059 struct rq *busiest_rq = data;
10060 int busiest_cpu = cpu_of(busiest_rq);
10061 int target_cpu = busiest_rq->push_cpu;
10062 struct rq *target_rq = cpu_rq(target_cpu);
10063 struct sched_domain *sd;
10064 struct task_struct *p = NULL;
10065 struct rq_flags rf;
10067 rq_lock_irq(busiest_rq, &rf);
10069 * Between queueing the stop-work and running it is a hole in which
10070 * CPUs can become inactive. We should not move tasks from or to
10073 if (!cpu_active(busiest_cpu) || !cpu_active(target_cpu))
10076 /* Make sure the requested CPU hasn't gone down in the meantime: */
10077 if (unlikely(busiest_cpu != smp_processor_id() ||
10078 !busiest_rq->active_balance))
10081 /* Is there any task to move? */
10082 if (busiest_rq->nr_running <= 1)
10086 * This condition is "impossible", if it occurs
10087 * we need to fix it. Originally reported by
10088 * Bjorn Helgaas on a 128-CPU setup.
10090 BUG_ON(busiest_rq == target_rq);
10092 /* Search for an sd spanning us and the target CPU. */
10094 for_each_domain(target_cpu, sd) {
10095 if (cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
10100 struct lb_env env = {
10102 .dst_cpu = target_cpu,
10103 .dst_rq = target_rq,
10104 .src_cpu = busiest_rq->cpu,
10105 .src_rq = busiest_rq,
10107 .flags = LBF_ACTIVE_LB,
10110 schedstat_inc(sd->alb_count);
10111 update_rq_clock(busiest_rq);
10113 p = detach_one_task(&env);
10115 schedstat_inc(sd->alb_pushed);
10116 /* Active balancing done, reset the failure counter. */
10117 sd->nr_balance_failed = 0;
10119 schedstat_inc(sd->alb_failed);
10124 busiest_rq->active_balance = 0;
10125 rq_unlock(busiest_rq, &rf);
10128 attach_one_task(target_rq, p);
10130 local_irq_enable();
10135 static DEFINE_SPINLOCK(balancing);
10138 * Scale the max load_balance interval with the number of CPUs in the system.
10139 * This trades load-balance latency on larger machines for less cross talk.
10141 void update_max_interval(void)
10143 max_load_balance_interval = HZ*num_online_cpus()/10;
10147 * It checks each scheduling domain to see if it is due to be balanced,
10148 * and initiates a balancing operation if so.
10150 * Balancing parameters are set up in init_sched_domains.
10152 static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle)
10154 int continue_balancing = 1;
10156 int busy = idle != CPU_IDLE && !sched_idle_cpu(cpu);
10157 unsigned long interval;
10158 struct sched_domain *sd;
10159 /* Earliest time when we have to do rebalance again */
10160 unsigned long next_balance = jiffies + 60*HZ;
10161 int update_next_balance = 0;
10162 int need_serialize, need_decay = 0;
10166 for_each_domain(cpu, sd) {
10168 * Decay the newidle max times here because this is a regular
10169 * visit to all the domains. Decay ~1% per second.
10171 if (time_after(jiffies, sd->next_decay_max_lb_cost)) {
10172 sd->max_newidle_lb_cost =
10173 (sd->max_newidle_lb_cost * 253) / 256;
10174 sd->next_decay_max_lb_cost = jiffies + HZ;
10177 max_cost += sd->max_newidle_lb_cost;
10180 * Stop the load balance at this level. There is another
10181 * CPU in our sched group which is doing load balancing more
10184 if (!continue_balancing) {
10190 interval = get_sd_balance_interval(sd, busy);
10192 need_serialize = sd->flags & SD_SERIALIZE;
10193 if (need_serialize) {
10194 if (!spin_trylock(&balancing))
10198 if (time_after_eq(jiffies, sd->last_balance + interval)) {
10199 if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
10201 * The LBF_DST_PINNED logic could have changed
10202 * env->dst_cpu, so we can't know our idle
10203 * state even if we migrated tasks. Update it.
10205 idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
10206 busy = idle != CPU_IDLE && !sched_idle_cpu(cpu);
10208 sd->last_balance = jiffies;
10209 interval = get_sd_balance_interval(sd, busy);
10211 if (need_serialize)
10212 spin_unlock(&balancing);
10214 if (time_after(next_balance, sd->last_balance + interval)) {
10215 next_balance = sd->last_balance + interval;
10216 update_next_balance = 1;
10221 * Ensure the rq-wide value also decays but keep it at a
10222 * reasonable floor to avoid funnies with rq->avg_idle.
10224 rq->max_idle_balance_cost =
10225 max((u64)sysctl_sched_migration_cost, max_cost);
10230 * next_balance will be updated only when there is a need.
10231 * When the cpu is attached to null domain for ex, it will not be
10234 if (likely(update_next_balance))
10235 rq->next_balance = next_balance;
10239 static inline int on_null_domain(struct rq *rq)
10241 return unlikely(!rcu_dereference_sched(rq->sd));
10244 #ifdef CONFIG_NO_HZ_COMMON
10246 * idle load balancing details
10247 * - When one of the busy CPUs notice that there may be an idle rebalancing
10248 * needed, they will kick the idle load balancer, which then does idle
10249 * load balancing for all the idle CPUs.
10250 * - HK_FLAG_MISC CPUs are used for this task, because HK_FLAG_SCHED not set
10254 static inline int find_new_ilb(void)
10257 const struct cpumask *hk_mask;
10259 hk_mask = housekeeping_cpumask(HK_FLAG_MISC);
10261 for_each_cpu_and(ilb, nohz.idle_cpus_mask, hk_mask) {
10263 if (ilb == smp_processor_id())
10274 * Kick a CPU to do the nohz balancing, if it is time for it. We pick any
10275 * idle CPU in the HK_FLAG_MISC housekeeping set (if there is one).
10277 static void kick_ilb(unsigned int flags)
10282 * Increase nohz.next_balance only when if full ilb is triggered but
10283 * not if we only update stats.
10285 if (flags & NOHZ_BALANCE_KICK)
10286 nohz.next_balance = jiffies+1;
10288 ilb_cpu = find_new_ilb();
10290 if (ilb_cpu >= nr_cpu_ids)
10294 * Access to rq::nohz_csd is serialized by NOHZ_KICK_MASK; he who sets
10295 * the first flag owns it; cleared by nohz_csd_func().
10297 flags = atomic_fetch_or(flags, nohz_flags(ilb_cpu));
10298 if (flags & NOHZ_KICK_MASK)
10302 * This way we generate an IPI on the target CPU which
10303 * is idle. And the softirq performing nohz idle load balance
10304 * will be run before returning from the IPI.
10306 smp_call_function_single_async(ilb_cpu, &cpu_rq(ilb_cpu)->nohz_csd);
10310 * Current decision point for kicking the idle load balancer in the presence
10311 * of idle CPUs in the system.
10313 static void nohz_balancer_kick(struct rq *rq)
10315 unsigned long now = jiffies;
10316 struct sched_domain_shared *sds;
10317 struct sched_domain *sd;
10318 int nr_busy, i, cpu = rq->cpu;
10319 unsigned int flags = 0;
10321 if (unlikely(rq->idle_balance))
10325 * We may be recently in ticked or tickless idle mode. At the first
10326 * busy tick after returning from idle, we will update the busy stats.
10328 nohz_balance_exit_idle(rq);
10331 * None are in tickless mode and hence no need for NOHZ idle load
10334 if (likely(!atomic_read(&nohz.nr_cpus)))
10337 if (READ_ONCE(nohz.has_blocked) &&
10338 time_after(now, READ_ONCE(nohz.next_blocked)))
10339 flags = NOHZ_STATS_KICK;
10341 if (time_before(now, nohz.next_balance))
10344 if (rq->nr_running >= 2) {
10345 flags = NOHZ_KICK_MASK;
10351 sd = rcu_dereference(rq->sd);
10354 * If there's a CFS task and the current CPU has reduced
10355 * capacity; kick the ILB to see if there's a better CPU to run
10358 if (rq->cfs.h_nr_running >= 1 && check_cpu_capacity(rq, sd)) {
10359 flags = NOHZ_KICK_MASK;
10364 sd = rcu_dereference(per_cpu(sd_asym_packing, cpu));
10367 * When ASYM_PACKING; see if there's a more preferred CPU
10368 * currently idle; in which case, kick the ILB to move tasks
10371 for_each_cpu_and(i, sched_domain_span(sd), nohz.idle_cpus_mask) {
10372 if (sched_asym_prefer(i, cpu)) {
10373 flags = NOHZ_KICK_MASK;
10379 sd = rcu_dereference(per_cpu(sd_asym_cpucapacity, cpu));
10382 * When ASYM_CPUCAPACITY; see if there's a higher capacity CPU
10383 * to run the misfit task on.
10385 if (check_misfit_status(rq, sd)) {
10386 flags = NOHZ_KICK_MASK;
10391 * For asymmetric systems, we do not want to nicely balance
10392 * cache use, instead we want to embrace asymmetry and only
10393 * ensure tasks have enough CPU capacity.
10395 * Skip the LLC logic because it's not relevant in that case.
10400 sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
10403 * If there is an imbalance between LLC domains (IOW we could
10404 * increase the overall cache use), we need some less-loaded LLC
10405 * domain to pull some load. Likewise, we may need to spread
10406 * load within the current LLC domain (e.g. packed SMT cores but
10407 * other CPUs are idle). We can't really know from here how busy
10408 * the others are - so just get a nohz balance going if it looks
10409 * like this LLC domain has tasks we could move.
10411 nr_busy = atomic_read(&sds->nr_busy_cpus);
10413 flags = NOHZ_KICK_MASK;
10424 static void set_cpu_sd_state_busy(int cpu)
10426 struct sched_domain *sd;
10429 sd = rcu_dereference(per_cpu(sd_llc, cpu));
10431 if (!sd || !sd->nohz_idle)
10435 atomic_inc(&sd->shared->nr_busy_cpus);
10440 void nohz_balance_exit_idle(struct rq *rq)
10442 SCHED_WARN_ON(rq != this_rq());
10444 if (likely(!rq->nohz_tick_stopped))
10447 rq->nohz_tick_stopped = 0;
10448 cpumask_clear_cpu(rq->cpu, nohz.idle_cpus_mask);
10449 atomic_dec(&nohz.nr_cpus);
10451 set_cpu_sd_state_busy(rq->cpu);
10454 static void set_cpu_sd_state_idle(int cpu)
10456 struct sched_domain *sd;
10459 sd = rcu_dereference(per_cpu(sd_llc, cpu));
10461 if (!sd || sd->nohz_idle)
10465 atomic_dec(&sd->shared->nr_busy_cpus);
10471 * This routine will record that the CPU is going idle with tick stopped.
10472 * This info will be used in performing idle load balancing in the future.
10474 void nohz_balance_enter_idle(int cpu)
10476 struct rq *rq = cpu_rq(cpu);
10478 SCHED_WARN_ON(cpu != smp_processor_id());
10480 /* If this CPU is going down, then nothing needs to be done: */
10481 if (!cpu_active(cpu))
10484 /* Spare idle load balancing on CPUs that don't want to be disturbed: */
10485 if (!housekeeping_cpu(cpu, HK_FLAG_SCHED))
10489 * Can be set safely without rq->lock held
10490 * If a clear happens, it will have evaluated last additions because
10491 * rq->lock is held during the check and the clear
10493 rq->has_blocked_load = 1;
10496 * The tick is still stopped but load could have been added in the
10497 * meantime. We set the nohz.has_blocked flag to trig a check of the
10498 * *_avg. The CPU is already part of nohz.idle_cpus_mask so the clear
10499 * of nohz.has_blocked can only happen after checking the new load
10501 if (rq->nohz_tick_stopped)
10504 /* If we're a completely isolated CPU, we don't play: */
10505 if (on_null_domain(rq))
10508 rq->nohz_tick_stopped = 1;
10510 cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
10511 atomic_inc(&nohz.nr_cpus);
10514 * Ensures that if nohz_idle_balance() fails to observe our
10515 * @idle_cpus_mask store, it must observe the @has_blocked
10518 smp_mb__after_atomic();
10520 set_cpu_sd_state_idle(cpu);
10524 * Each time a cpu enter idle, we assume that it has blocked load and
10525 * enable the periodic update of the load of idle cpus
10527 WRITE_ONCE(nohz.has_blocked, 1);
10530 static bool update_nohz_stats(struct rq *rq)
10532 unsigned int cpu = rq->cpu;
10534 if (!rq->has_blocked_load)
10537 if (!cpumask_test_cpu(cpu, nohz.idle_cpus_mask))
10540 if (!time_after(jiffies, READ_ONCE(rq->last_blocked_load_update_tick)))
10543 update_blocked_averages(cpu);
10545 return rq->has_blocked_load;
10549 * Internal function that runs load balance for all idle cpus. The load balance
10550 * can be a simple update of blocked load or a complete load balance with
10551 * tasks movement depending of flags.
10553 static void _nohz_idle_balance(struct rq *this_rq, unsigned int flags,
10554 enum cpu_idle_type idle)
10556 /* Earliest time when we have to do rebalance again */
10557 unsigned long now = jiffies;
10558 unsigned long next_balance = now + 60*HZ;
10559 bool has_blocked_load = false;
10560 int update_next_balance = 0;
10561 int this_cpu = this_rq->cpu;
10565 SCHED_WARN_ON((flags & NOHZ_KICK_MASK) == NOHZ_BALANCE_KICK);
10568 * We assume there will be no idle load after this update and clear
10569 * the has_blocked flag. If a cpu enters idle in the mean time, it will
10570 * set the has_blocked flag and trig another update of idle load.
10571 * Because a cpu that becomes idle, is added to idle_cpus_mask before
10572 * setting the flag, we are sure to not clear the state and not
10573 * check the load of an idle cpu.
10575 WRITE_ONCE(nohz.has_blocked, 0);
10578 * Ensures that if we miss the CPU, we must see the has_blocked
10579 * store from nohz_balance_enter_idle().
10584 * Start with the next CPU after this_cpu so we will end with this_cpu and let a
10585 * chance for other idle cpu to pull load.
10587 for_each_cpu_wrap(balance_cpu, nohz.idle_cpus_mask, this_cpu+1) {
10588 if (!idle_cpu(balance_cpu))
10592 * If this CPU gets work to do, stop the load balancing
10593 * work being done for other CPUs. Next load
10594 * balancing owner will pick it up.
10596 if (need_resched()) {
10597 has_blocked_load = true;
10601 rq = cpu_rq(balance_cpu);
10603 has_blocked_load |= update_nohz_stats(rq);
10606 * If time for next balance is due,
10609 if (time_after_eq(jiffies, rq->next_balance)) {
10610 struct rq_flags rf;
10612 rq_lock_irqsave(rq, &rf);
10613 update_rq_clock(rq);
10614 rq_unlock_irqrestore(rq, &rf);
10616 if (flags & NOHZ_BALANCE_KICK)
10617 rebalance_domains(rq, CPU_IDLE);
10620 if (time_after(next_balance, rq->next_balance)) {
10621 next_balance = rq->next_balance;
10622 update_next_balance = 1;
10627 * next_balance will be updated only when there is a need.
10628 * When the CPU is attached to null domain for ex, it will not be
10631 if (likely(update_next_balance))
10632 nohz.next_balance = next_balance;
10634 WRITE_ONCE(nohz.next_blocked,
10635 now + msecs_to_jiffies(LOAD_AVG_PERIOD));
10638 /* There is still blocked load, enable periodic update */
10639 if (has_blocked_load)
10640 WRITE_ONCE(nohz.has_blocked, 1);
10644 * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
10645 * rebalancing for all the cpus for whom scheduler ticks are stopped.
10647 static bool nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
10649 unsigned int flags = this_rq->nohz_idle_balance;
10654 this_rq->nohz_idle_balance = 0;
10656 if (idle != CPU_IDLE)
10659 _nohz_idle_balance(this_rq, flags, idle);
10665 * Check if we need to run the ILB for updating blocked load before entering
10668 void nohz_run_idle_balance(int cpu)
10670 unsigned int flags;
10672 flags = atomic_fetch_andnot(NOHZ_NEWILB_KICK, nohz_flags(cpu));
10675 * Update the blocked load only if no SCHED_SOFTIRQ is about to happen
10676 * (ie NOHZ_STATS_KICK set) and will do the same.
10678 if ((flags == NOHZ_NEWILB_KICK) && !need_resched())
10679 _nohz_idle_balance(cpu_rq(cpu), NOHZ_STATS_KICK, CPU_IDLE);
10682 static void nohz_newidle_balance(struct rq *this_rq)
10684 int this_cpu = this_rq->cpu;
10687 * This CPU doesn't want to be disturbed by scheduler
10690 if (!housekeeping_cpu(this_cpu, HK_FLAG_SCHED))
10693 /* Will wake up very soon. No time for doing anything else*/
10694 if (this_rq->avg_idle < sysctl_sched_migration_cost)
10697 /* Don't need to update blocked load of idle CPUs*/
10698 if (!READ_ONCE(nohz.has_blocked) ||
10699 time_before(jiffies, READ_ONCE(nohz.next_blocked)))
10703 * Set the need to trigger ILB in order to update blocked load
10704 * before entering idle state.
10706 atomic_or(NOHZ_NEWILB_KICK, nohz_flags(this_cpu));
10709 #else /* !CONFIG_NO_HZ_COMMON */
10710 static inline void nohz_balancer_kick(struct rq *rq) { }
10712 static inline bool nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
10717 static inline void nohz_newidle_balance(struct rq *this_rq) { }
10718 #endif /* CONFIG_NO_HZ_COMMON */
10721 * newidle_balance is called by schedule() if this_cpu is about to become
10722 * idle. Attempts to pull tasks from other CPUs.
10725 * < 0 - we released the lock and there are !fair tasks present
10726 * 0 - failed, no new tasks
10727 * > 0 - success, new (fair) tasks present
10729 static int newidle_balance(struct rq *this_rq, struct rq_flags *rf)
10731 unsigned long next_balance = jiffies + HZ;
10732 int this_cpu = this_rq->cpu;
10733 struct sched_domain *sd;
10734 int pulled_task = 0;
10737 update_misfit_status(NULL, this_rq);
10740 * There is a task waiting to run. No need to search for one.
10741 * Return 0; the task will be enqueued when switching to idle.
10743 if (this_rq->ttwu_pending)
10747 * We must set idle_stamp _before_ calling idle_balance(), such that we
10748 * measure the duration of idle_balance() as idle time.
10750 this_rq->idle_stamp = rq_clock(this_rq);
10753 * Do not pull tasks towards !active CPUs...
10755 if (!cpu_active(this_cpu))
10759 * This is OK, because current is on_cpu, which avoids it being picked
10760 * for load-balance and preemption/IRQs are still disabled avoiding
10761 * further scheduler activity on it and we're being very careful to
10762 * re-start the picking loop.
10764 rq_unpin_lock(this_rq, rf);
10766 if (this_rq->avg_idle < sysctl_sched_migration_cost ||
10767 !READ_ONCE(this_rq->rd->overload)) {
10770 sd = rcu_dereference_check_sched_domain(this_rq->sd);
10772 update_next_balance(sd, &next_balance);
10778 raw_spin_rq_unlock(this_rq);
10780 update_blocked_averages(this_cpu);
10782 for_each_domain(this_cpu, sd) {
10783 int continue_balancing = 1;
10784 u64 t0, domain_cost;
10786 if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost) {
10787 update_next_balance(sd, &next_balance);
10791 if (sd->flags & SD_BALANCE_NEWIDLE) {
10792 t0 = sched_clock_cpu(this_cpu);
10794 pulled_task = load_balance(this_cpu, this_rq,
10795 sd, CPU_NEWLY_IDLE,
10796 &continue_balancing);
10798 domain_cost = sched_clock_cpu(this_cpu) - t0;
10799 if (domain_cost > sd->max_newidle_lb_cost)
10800 sd->max_newidle_lb_cost = domain_cost;
10802 curr_cost += domain_cost;
10805 update_next_balance(sd, &next_balance);
10808 * Stop searching for tasks to pull if there are
10809 * now runnable tasks on this rq.
10811 if (pulled_task || this_rq->nr_running > 0 ||
10812 this_rq->ttwu_pending)
10817 raw_spin_rq_lock(this_rq);
10819 if (curr_cost > this_rq->max_idle_balance_cost)
10820 this_rq->max_idle_balance_cost = curr_cost;
10823 * While browsing the domains, we released the rq lock, a task could
10824 * have been enqueued in the meantime. Since we're not going idle,
10825 * pretend we pulled a task.
10827 if (this_rq->cfs.h_nr_running && !pulled_task)
10830 /* Is there a task of a high priority class? */
10831 if (this_rq->nr_running != this_rq->cfs.h_nr_running)
10835 /* Move the next balance forward */
10836 if (time_after(this_rq->next_balance, next_balance))
10837 this_rq->next_balance = next_balance;
10840 this_rq->idle_stamp = 0;
10842 nohz_newidle_balance(this_rq);
10844 rq_repin_lock(this_rq, rf);
10846 return pulled_task;
10850 * run_rebalance_domains is triggered when needed from the scheduler tick.
10851 * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
10853 static __latent_entropy void run_rebalance_domains(struct softirq_action *h)
10855 struct rq *this_rq = this_rq();
10856 enum cpu_idle_type idle = this_rq->idle_balance ?
10857 CPU_IDLE : CPU_NOT_IDLE;
10860 * If this CPU has a pending nohz_balance_kick, then do the
10861 * balancing on behalf of the other idle CPUs whose ticks are
10862 * stopped. Do nohz_idle_balance *before* rebalance_domains to
10863 * give the idle CPUs a chance to load balance. Else we may
10864 * load balance only within the local sched_domain hierarchy
10865 * and abort nohz_idle_balance altogether if we pull some load.
10867 if (nohz_idle_balance(this_rq, idle))
10870 /* normal load balance */
10871 update_blocked_averages(this_rq->cpu);
10872 rebalance_domains(this_rq, idle);
10876 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
10878 void trigger_load_balance(struct rq *rq)
10881 * Don't need to rebalance while attached to NULL domain or
10882 * runqueue CPU is not active
10884 if (unlikely(on_null_domain(rq) || !cpu_active(cpu_of(rq))))
10887 if (time_after_eq(jiffies, rq->next_balance))
10888 raise_softirq(SCHED_SOFTIRQ);
10890 nohz_balancer_kick(rq);
10893 static void rq_online_fair(struct rq *rq)
10897 update_runtime_enabled(rq);
10900 static void rq_offline_fair(struct rq *rq)
10904 /* Ensure any throttled groups are reachable by pick_next_task */
10905 unthrottle_offline_cfs_rqs(rq);
10908 #endif /* CONFIG_SMP */
10910 #ifdef CONFIG_SCHED_CORE
10912 __entity_slice_used(struct sched_entity *se, int min_nr_tasks)
10914 u64 slice = sched_slice(cfs_rq_of(se), se);
10915 u64 rtime = se->sum_exec_runtime - se->prev_sum_exec_runtime;
10917 return (rtime * min_nr_tasks > slice);
10920 #define MIN_NR_TASKS_DURING_FORCEIDLE 2
10921 static inline void task_tick_core(struct rq *rq, struct task_struct *curr)
10923 if (!sched_core_enabled(rq))
10927 * If runqueue has only one task which used up its slice and
10928 * if the sibling is forced idle, then trigger schedule to
10929 * give forced idle task a chance.
10931 * sched_slice() considers only this active rq and it gets the
10932 * whole slice. But during force idle, we have siblings acting
10933 * like a single runqueue and hence we need to consider runnable
10934 * tasks on this CPU and the forced idle CPU. Ideally, we should
10935 * go through the forced idle rq, but that would be a perf hit.
10936 * We can assume that the forced idle CPU has at least
10937 * MIN_NR_TASKS_DURING_FORCEIDLE - 1 tasks and use that to check
10938 * if we need to give up the CPU.
10940 if (rq->core->core_forceidle && rq->cfs.nr_running == 1 &&
10941 __entity_slice_used(&curr->se, MIN_NR_TASKS_DURING_FORCEIDLE))
10946 * se_fi_update - Update the cfs_rq->min_vruntime_fi in a CFS hierarchy if needed.
10948 static void se_fi_update(struct sched_entity *se, unsigned int fi_seq, bool forceidle)
10950 for_each_sched_entity(se) {
10951 struct cfs_rq *cfs_rq = cfs_rq_of(se);
10954 if (cfs_rq->forceidle_seq == fi_seq)
10956 cfs_rq->forceidle_seq = fi_seq;
10959 cfs_rq->min_vruntime_fi = cfs_rq->min_vruntime;
10963 void task_vruntime_update(struct rq *rq, struct task_struct *p, bool in_fi)
10965 struct sched_entity *se = &p->se;
10967 if (p->sched_class != &fair_sched_class)
10970 se_fi_update(se, rq->core->core_forceidle_seq, in_fi);
10973 bool cfs_prio_less(struct task_struct *a, struct task_struct *b, bool in_fi)
10975 struct rq *rq = task_rq(a);
10976 struct sched_entity *sea = &a->se;
10977 struct sched_entity *seb = &b->se;
10978 struct cfs_rq *cfs_rqa;
10979 struct cfs_rq *cfs_rqb;
10982 SCHED_WARN_ON(task_rq(b)->core != rq->core);
10984 #ifdef CONFIG_FAIR_GROUP_SCHED
10986 * Find an se in the hierarchy for tasks a and b, such that the se's
10987 * are immediate siblings.
10989 while (sea->cfs_rq->tg != seb->cfs_rq->tg) {
10990 int sea_depth = sea->depth;
10991 int seb_depth = seb->depth;
10993 if (sea_depth >= seb_depth)
10994 sea = parent_entity(sea);
10995 if (sea_depth <= seb_depth)
10996 seb = parent_entity(seb);
10999 se_fi_update(sea, rq->core->core_forceidle_seq, in_fi);
11000 se_fi_update(seb, rq->core->core_forceidle_seq, in_fi);
11002 cfs_rqa = sea->cfs_rq;
11003 cfs_rqb = seb->cfs_rq;
11005 cfs_rqa = &task_rq(a)->cfs;
11006 cfs_rqb = &task_rq(b)->cfs;
11010 * Find delta after normalizing se's vruntime with its cfs_rq's
11011 * min_vruntime_fi, which would have been updated in prior calls
11012 * to se_fi_update().
11014 delta = (s64)(sea->vruntime - seb->vruntime) +
11015 (s64)(cfs_rqb->min_vruntime_fi - cfs_rqa->min_vruntime_fi);
11020 static inline void task_tick_core(struct rq *rq, struct task_struct *curr) {}
11024 * scheduler tick hitting a task of our scheduling class.
11026 * NOTE: This function can be called remotely by the tick offload that
11027 * goes along full dynticks. Therefore no local assumption can be made
11028 * and everything must be accessed through the @rq and @curr passed in
11031 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
11033 struct cfs_rq *cfs_rq;
11034 struct sched_entity *se = &curr->se;
11036 for_each_sched_entity(se) {
11037 cfs_rq = cfs_rq_of(se);
11038 entity_tick(cfs_rq, se, queued);
11041 if (static_branch_unlikely(&sched_numa_balancing))
11042 task_tick_numa(rq, curr);
11044 update_misfit_status(curr, rq);
11045 update_overutilized_status(task_rq(curr));
11047 task_tick_core(rq, curr);
11051 * called on fork with the child task as argument from the parent's context
11052 * - child not yet on the tasklist
11053 * - preemption disabled
11055 static void task_fork_fair(struct task_struct *p)
11057 struct cfs_rq *cfs_rq;
11058 struct sched_entity *se = &p->se, *curr;
11059 struct rq *rq = this_rq();
11060 struct rq_flags rf;
11063 update_rq_clock(rq);
11065 cfs_rq = task_cfs_rq(current);
11066 curr = cfs_rq->curr;
11068 update_curr(cfs_rq);
11069 se->vruntime = curr->vruntime;
11071 place_entity(cfs_rq, se, 1);
11073 if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
11075 * Upon rescheduling, sched_class::put_prev_task() will place
11076 * 'current' within the tree based on its new key value.
11078 swap(curr->vruntime, se->vruntime);
11082 se->vruntime -= cfs_rq->min_vruntime;
11083 rq_unlock(rq, &rf);
11087 * Priority of the task has changed. Check to see if we preempt
11088 * the current task.
11091 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
11093 if (!task_on_rq_queued(p))
11096 if (rq->cfs.nr_running == 1)
11100 * Reschedule if we are currently running on this runqueue and
11101 * our priority decreased, or if we are not currently running on
11102 * this runqueue and our priority is higher than the current's
11104 if (task_current(rq, p)) {
11105 if (p->prio > oldprio)
11108 check_preempt_curr(rq, p, 0);
11111 static inline bool vruntime_normalized(struct task_struct *p)
11113 struct sched_entity *se = &p->se;
11116 * In both the TASK_ON_RQ_QUEUED and TASK_ON_RQ_MIGRATING cases,
11117 * the dequeue_entity(.flags=0) will already have normalized the
11124 * When !on_rq, vruntime of the task has usually NOT been normalized.
11125 * But there are some cases where it has already been normalized:
11127 * - A forked child which is waiting for being woken up by
11128 * wake_up_new_task().
11129 * - A task which has been woken up by try_to_wake_up() and
11130 * waiting for actually being woken up by sched_ttwu_pending().
11132 if (!se->sum_exec_runtime ||
11133 (READ_ONCE(p->__state) == TASK_WAKING && p->sched_remote_wakeup))
11139 #ifdef CONFIG_FAIR_GROUP_SCHED
11141 * Propagate the changes of the sched_entity across the tg tree to make it
11142 * visible to the root
11144 static void propagate_entity_cfs_rq(struct sched_entity *se)
11146 struct cfs_rq *cfs_rq;
11148 list_add_leaf_cfs_rq(cfs_rq_of(se));
11150 /* Start to propagate at parent */
11153 for_each_sched_entity(se) {
11154 cfs_rq = cfs_rq_of(se);
11156 if (!cfs_rq_throttled(cfs_rq)){
11157 update_load_avg(cfs_rq, se, UPDATE_TG);
11158 list_add_leaf_cfs_rq(cfs_rq);
11162 if (list_add_leaf_cfs_rq(cfs_rq))
11167 static void propagate_entity_cfs_rq(struct sched_entity *se) { }
11170 static void detach_entity_cfs_rq(struct sched_entity *se)
11172 struct cfs_rq *cfs_rq = cfs_rq_of(se);
11174 /* Catch up with the cfs_rq and remove our load when we leave */
11175 update_load_avg(cfs_rq, se, 0);
11176 detach_entity_load_avg(cfs_rq, se);
11177 update_tg_load_avg(cfs_rq);
11178 propagate_entity_cfs_rq(se);
11181 static void attach_entity_cfs_rq(struct sched_entity *se)
11183 struct cfs_rq *cfs_rq = cfs_rq_of(se);
11185 #ifdef CONFIG_FAIR_GROUP_SCHED
11187 * Since the real-depth could have been changed (only FAIR
11188 * class maintain depth value), reset depth properly.
11190 se->depth = se->parent ? se->parent->depth + 1 : 0;
11193 /* Synchronize entity with its cfs_rq */
11194 update_load_avg(cfs_rq, se, sched_feat(ATTACH_AGE_LOAD) ? 0 : SKIP_AGE_LOAD);
11195 attach_entity_load_avg(cfs_rq, se);
11196 update_tg_load_avg(cfs_rq);
11197 propagate_entity_cfs_rq(se);
11200 static void detach_task_cfs_rq(struct task_struct *p)
11202 struct sched_entity *se = &p->se;
11203 struct cfs_rq *cfs_rq = cfs_rq_of(se);
11205 if (!vruntime_normalized(p)) {
11207 * Fix up our vruntime so that the current sleep doesn't
11208 * cause 'unlimited' sleep bonus.
11210 place_entity(cfs_rq, se, 0);
11211 se->vruntime -= cfs_rq->min_vruntime;
11214 detach_entity_cfs_rq(se);
11217 static void attach_task_cfs_rq(struct task_struct *p)
11219 struct sched_entity *se = &p->se;
11220 struct cfs_rq *cfs_rq = cfs_rq_of(se);
11222 attach_entity_cfs_rq(se);
11224 if (!vruntime_normalized(p))
11225 se->vruntime += cfs_rq->min_vruntime;
11228 static void switched_from_fair(struct rq *rq, struct task_struct *p)
11230 detach_task_cfs_rq(p);
11233 static void switched_to_fair(struct rq *rq, struct task_struct *p)
11235 attach_task_cfs_rq(p);
11237 if (task_on_rq_queued(p)) {
11239 * We were most likely switched from sched_rt, so
11240 * kick off the schedule if running, otherwise just see
11241 * if we can still preempt the current task.
11243 if (task_current(rq, p))
11246 check_preempt_curr(rq, p, 0);
11250 /* Account for a task changing its policy or group.
11252 * This routine is mostly called to set cfs_rq->curr field when a task
11253 * migrates between groups/classes.
11255 static void set_next_task_fair(struct rq *rq, struct task_struct *p, bool first)
11257 struct sched_entity *se = &p->se;
11260 if (task_on_rq_queued(p)) {
11262 * Move the next running task to the front of the list, so our
11263 * cfs_tasks list becomes MRU one.
11265 list_move(&se->group_node, &rq->cfs_tasks);
11269 for_each_sched_entity(se) {
11270 struct cfs_rq *cfs_rq = cfs_rq_of(se);
11272 set_next_entity(cfs_rq, se);
11273 /* ensure bandwidth has been allocated on our new cfs_rq */
11274 account_cfs_rq_runtime(cfs_rq, 0);
11278 void init_cfs_rq(struct cfs_rq *cfs_rq)
11280 cfs_rq->tasks_timeline = RB_ROOT_CACHED;
11281 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
11282 #ifndef CONFIG_64BIT
11283 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
11286 raw_spin_lock_init(&cfs_rq->removed.lock);
11290 #ifdef CONFIG_FAIR_GROUP_SCHED
11291 static void task_set_group_fair(struct task_struct *p)
11293 struct sched_entity *se = &p->se;
11295 set_task_rq(p, task_cpu(p));
11296 se->depth = se->parent ? se->parent->depth + 1 : 0;
11299 static void task_move_group_fair(struct task_struct *p)
11301 detach_task_cfs_rq(p);
11302 set_task_rq(p, task_cpu(p));
11305 /* Tell se's cfs_rq has been changed -- migrated */
11306 p->se.avg.last_update_time = 0;
11308 attach_task_cfs_rq(p);
11311 static void task_change_group_fair(struct task_struct *p, int type)
11314 case TASK_SET_GROUP:
11315 task_set_group_fair(p);
11318 case TASK_MOVE_GROUP:
11319 task_move_group_fair(p);
11324 void free_fair_sched_group(struct task_group *tg)
11328 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
11330 for_each_possible_cpu(i) {
11332 kfree(tg->cfs_rq[i]);
11341 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
11343 struct sched_entity *se;
11344 struct cfs_rq *cfs_rq;
11347 tg->cfs_rq = kcalloc(nr_cpu_ids, sizeof(cfs_rq), GFP_KERNEL);
11350 tg->se = kcalloc(nr_cpu_ids, sizeof(se), GFP_KERNEL);
11354 tg->shares = NICE_0_LOAD;
11356 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
11358 for_each_possible_cpu(i) {
11359 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
11360 GFP_KERNEL, cpu_to_node(i));
11364 se = kzalloc_node(sizeof(struct sched_entity),
11365 GFP_KERNEL, cpu_to_node(i));
11369 init_cfs_rq(cfs_rq);
11370 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
11371 init_entity_runnable_average(se);
11382 void online_fair_sched_group(struct task_group *tg)
11384 struct sched_entity *se;
11385 struct rq_flags rf;
11389 for_each_possible_cpu(i) {
11392 rq_lock_irq(rq, &rf);
11393 update_rq_clock(rq);
11394 attach_entity_cfs_rq(se);
11395 sync_throttle(tg, i);
11396 rq_unlock_irq(rq, &rf);
11400 void unregister_fair_sched_group(struct task_group *tg)
11402 unsigned long flags;
11406 for_each_possible_cpu(cpu) {
11408 remove_entity_load_avg(tg->se[cpu]);
11411 * Only empty task groups can be destroyed; so we can speculatively
11412 * check on_list without danger of it being re-added.
11414 if (!tg->cfs_rq[cpu]->on_list)
11419 raw_spin_rq_lock_irqsave(rq, flags);
11420 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
11421 raw_spin_rq_unlock_irqrestore(rq, flags);
11425 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
11426 struct sched_entity *se, int cpu,
11427 struct sched_entity *parent)
11429 struct rq *rq = cpu_rq(cpu);
11433 init_cfs_rq_runtime(cfs_rq);
11435 tg->cfs_rq[cpu] = cfs_rq;
11438 /* se could be NULL for root_task_group */
11443 se->cfs_rq = &rq->cfs;
11446 se->cfs_rq = parent->my_q;
11447 se->depth = parent->depth + 1;
11451 /* guarantee group entities always have weight */
11452 update_load_set(&se->load, NICE_0_LOAD);
11453 se->parent = parent;
11456 static DEFINE_MUTEX(shares_mutex);
11458 static int __sched_group_set_shares(struct task_group *tg, unsigned long shares)
11462 lockdep_assert_held(&shares_mutex);
11465 * We can't change the weight of the root cgroup.
11470 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
11472 if (tg->shares == shares)
11475 tg->shares = shares;
11476 for_each_possible_cpu(i) {
11477 struct rq *rq = cpu_rq(i);
11478 struct sched_entity *se = tg->se[i];
11479 struct rq_flags rf;
11481 /* Propagate contribution to hierarchy */
11482 rq_lock_irqsave(rq, &rf);
11483 update_rq_clock(rq);
11484 for_each_sched_entity(se) {
11485 update_load_avg(cfs_rq_of(se), se, UPDATE_TG);
11486 update_cfs_group(se);
11488 rq_unlock_irqrestore(rq, &rf);
11494 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
11498 mutex_lock(&shares_mutex);
11499 if (tg_is_idle(tg))
11502 ret = __sched_group_set_shares(tg, shares);
11503 mutex_unlock(&shares_mutex);
11508 int sched_group_set_idle(struct task_group *tg, long idle)
11512 if (tg == &root_task_group)
11515 if (idle < 0 || idle > 1)
11518 mutex_lock(&shares_mutex);
11520 if (tg->idle == idle) {
11521 mutex_unlock(&shares_mutex);
11527 for_each_possible_cpu(i) {
11528 struct rq *rq = cpu_rq(i);
11529 struct sched_entity *se = tg->se[i];
11530 struct cfs_rq *grp_cfs_rq = tg->cfs_rq[i];
11531 bool was_idle = cfs_rq_is_idle(grp_cfs_rq);
11532 long idle_task_delta;
11533 struct rq_flags rf;
11535 rq_lock_irqsave(rq, &rf);
11537 grp_cfs_rq->idle = idle;
11538 if (WARN_ON_ONCE(was_idle == cfs_rq_is_idle(grp_cfs_rq)))
11541 idle_task_delta = grp_cfs_rq->h_nr_running -
11542 grp_cfs_rq->idle_h_nr_running;
11543 if (!cfs_rq_is_idle(grp_cfs_rq))
11544 idle_task_delta *= -1;
11546 for_each_sched_entity(se) {
11547 struct cfs_rq *cfs_rq = cfs_rq_of(se);
11552 cfs_rq->idle_h_nr_running += idle_task_delta;
11554 /* Already accounted at parent level and above. */
11555 if (cfs_rq_is_idle(cfs_rq))
11560 rq_unlock_irqrestore(rq, &rf);
11563 /* Idle groups have minimum weight. */
11564 if (tg_is_idle(tg))
11565 __sched_group_set_shares(tg, scale_load(WEIGHT_IDLEPRIO));
11567 __sched_group_set_shares(tg, NICE_0_LOAD);
11569 mutex_unlock(&shares_mutex);
11573 #else /* CONFIG_FAIR_GROUP_SCHED */
11575 void free_fair_sched_group(struct task_group *tg) { }
11577 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
11582 void online_fair_sched_group(struct task_group *tg) { }
11584 void unregister_fair_sched_group(struct task_group *tg) { }
11586 #endif /* CONFIG_FAIR_GROUP_SCHED */
11589 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
11591 struct sched_entity *se = &task->se;
11592 unsigned int rr_interval = 0;
11595 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
11598 if (rq->cfs.load.weight)
11599 rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
11601 return rr_interval;
11605 * All the scheduling class methods:
11607 DEFINE_SCHED_CLASS(fair) = {
11609 .enqueue_task = enqueue_task_fair,
11610 .dequeue_task = dequeue_task_fair,
11611 .yield_task = yield_task_fair,
11612 .yield_to_task = yield_to_task_fair,
11614 .check_preempt_curr = check_preempt_wakeup,
11616 .pick_next_task = __pick_next_task_fair,
11617 .put_prev_task = put_prev_task_fair,
11618 .set_next_task = set_next_task_fair,
11621 .balance = balance_fair,
11622 .pick_task = pick_task_fair,
11623 .select_task_rq = select_task_rq_fair,
11624 .migrate_task_rq = migrate_task_rq_fair,
11626 .rq_online = rq_online_fair,
11627 .rq_offline = rq_offline_fair,
11629 .task_dead = task_dead_fair,
11630 .set_cpus_allowed = set_cpus_allowed_common,
11633 .task_tick = task_tick_fair,
11634 .task_fork = task_fork_fair,
11636 .prio_changed = prio_changed_fair,
11637 .switched_from = switched_from_fair,
11638 .switched_to = switched_to_fair,
11640 .get_rr_interval = get_rr_interval_fair,
11642 .update_curr = update_curr_fair,
11644 #ifdef CONFIG_FAIR_GROUP_SCHED
11645 .task_change_group = task_change_group_fair,
11648 #ifdef CONFIG_UCLAMP_TASK
11649 .uclamp_enabled = 1,
11653 #ifdef CONFIG_SCHED_DEBUG
11654 void print_cfs_stats(struct seq_file *m, int cpu)
11656 struct cfs_rq *cfs_rq, *pos;
11659 for_each_leaf_cfs_rq_safe(cpu_rq(cpu), cfs_rq, pos)
11660 print_cfs_rq(m, cpu, cfs_rq);
11664 #ifdef CONFIG_NUMA_BALANCING
11665 void show_numa_stats(struct task_struct *p, struct seq_file *m)
11668 unsigned long tsf = 0, tpf = 0, gsf = 0, gpf = 0;
11669 struct numa_group *ng;
11672 ng = rcu_dereference(p->numa_group);
11673 for_each_online_node(node) {
11674 if (p->numa_faults) {
11675 tsf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 0)];
11676 tpf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 1)];
11679 gsf = ng->faults[task_faults_idx(NUMA_MEM, node, 0)],
11680 gpf = ng->faults[task_faults_idx(NUMA_MEM, node, 1)];
11682 print_numa_stats(m, node, tsf, tpf, gsf, gpf);
11686 #endif /* CONFIG_NUMA_BALANCING */
11687 #endif /* CONFIG_SCHED_DEBUG */
11689 __init void init_sched_fair_class(void)
11692 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
11694 #ifdef CONFIG_NO_HZ_COMMON
11695 nohz.next_balance = jiffies;
11696 nohz.next_blocked = jiffies;
11697 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
11704 * Helper functions to facilitate extracting info from tracepoints.
11707 const struct sched_avg *sched_trace_cfs_rq_avg(struct cfs_rq *cfs_rq)
11710 return cfs_rq ? &cfs_rq->avg : NULL;
11715 EXPORT_SYMBOL_GPL(sched_trace_cfs_rq_avg);
11717 char *sched_trace_cfs_rq_path(struct cfs_rq *cfs_rq, char *str, int len)
11721 strlcpy(str, "(null)", len);
11726 cfs_rq_tg_path(cfs_rq, str, len);
11729 EXPORT_SYMBOL_GPL(sched_trace_cfs_rq_path);
11731 int sched_trace_cfs_rq_cpu(struct cfs_rq *cfs_rq)
11733 return cfs_rq ? cpu_of(rq_of(cfs_rq)) : -1;
11735 EXPORT_SYMBOL_GPL(sched_trace_cfs_rq_cpu);
11737 const struct sched_avg *sched_trace_rq_avg_rt(struct rq *rq)
11740 return rq ? &rq->avg_rt : NULL;
11745 EXPORT_SYMBOL_GPL(sched_trace_rq_avg_rt);
11747 const struct sched_avg *sched_trace_rq_avg_dl(struct rq *rq)
11750 return rq ? &rq->avg_dl : NULL;
11755 EXPORT_SYMBOL_GPL(sched_trace_rq_avg_dl);
11757 const struct sched_avg *sched_trace_rq_avg_irq(struct rq *rq)
11759 #if defined(CONFIG_SMP) && defined(CONFIG_HAVE_SCHED_AVG_IRQ)
11760 return rq ? &rq->avg_irq : NULL;
11765 EXPORT_SYMBOL_GPL(sched_trace_rq_avg_irq);
11767 int sched_trace_rq_cpu(struct rq *rq)
11769 return rq ? cpu_of(rq) : -1;
11771 EXPORT_SYMBOL_GPL(sched_trace_rq_cpu);
11773 int sched_trace_rq_cpu_capacity(struct rq *rq)
11779 SCHED_CAPACITY_SCALE
11783 EXPORT_SYMBOL_GPL(sched_trace_rq_cpu_capacity);
11785 const struct cpumask *sched_trace_rd_span(struct root_domain *rd)
11788 return rd ? rd->span : NULL;
11793 EXPORT_SYMBOL_GPL(sched_trace_rd_span);
11795 int sched_trace_rq_nr_running(struct rq *rq)
11797 return rq ? rq->nr_running : -1;
11799 EXPORT_SYMBOL_GPL(sched_trace_rq_nr_running);