1 // SPDX-License-Identifier: GPL-2.0
44 #include <linux/memory-tiers.h>
62  * The initial- and re-scaling of tunables is configurable
66  *   SCHED_TUNABLESCALING_NONE - unscaled, always *1
67  *   SCHED_TUNABLESCALING_LOG - scaled logarithmically, *1+ilog(ncpus)
68  *   SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus
75  * Minimal preemption granularity for CPU-bound tasks:
97 	return -cpu;  in arch_asym_cpu_priority()
118  * Amount of runtime to allocate from global (tg) to local (per-cfs_rq) pool
169 	lw->weight += inc;  in update_load_add()
170 	lw->inv_weight = 0;  in update_load_add()
175 	lw->weight -= dec;  in update_load_sub()
176 	lw->inv_weight = 0;  in update_load_sub()
181 	lw->weight = w;  in update_load_set()
182 	lw->inv_weight = 0;  in update_load_set()
189  * so pick a second-best guess by going with the log2 of the
237 	if (likely(lw->inv_weight))  in __update_inv_weight()
240 	w = scale_load_down(lw->weight);  in __update_inv_weight()
243 		lw->inv_weight = 1;  in __update_inv_weight()
245 		lw->inv_weight = WMULT_CONST;  in __update_inv_weight()
247 		lw->inv_weight = WMULT_CONST / w;  in __update_inv_weight()
253  * (delta_exec * (weight * lw->inv_weight)) >> WMULT_SHIFT
273 		shift -= fs;  in __calc_delta()
277 	fact = mul_u32_u32(fact, lw->inv_weight);  in __calc_delta()
282 		shift -= fs;  in __calc_delta()
294 	if (unlikely(se->load.weight != NICE_0_LOAD))  in calc_delta_fair()
295 		delta = __calc_delta(delta, NICE_0_LOAD, &se->load);  in calc_delta_fair()
310 		for (; se; se = se->parent)
317 	if (cfs_rq->on_list)  in list_add_leaf_cfs_rq()
318 		return rq->tmp_alone_branch == &rq->leaf_cfs_rq_list;  in list_add_leaf_cfs_rq()
320 	cfs_rq->on_list = 1;  in list_add_leaf_cfs_rq()
325 	 * enqueued. The fact that we always enqueue bottom-up  in list_add_leaf_cfs_rq()
331 	if (cfs_rq->tg->parent &&  in list_add_leaf_cfs_rq()
332 	    cfs_rq->tg->parent->cfs_rq[cpu]->on_list) {  in list_add_leaf_cfs_rq()
339 		list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,  in list_add_leaf_cfs_rq()
340 			&(cfs_rq->tg->parent->cfs_rq[cpu]->leaf_cfs_rq_list));  in list_add_leaf_cfs_rq()
346 		rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;  in list_add_leaf_cfs_rq()
350 	if (!cfs_rq->tg->parent) {  in list_add_leaf_cfs_rq()
355 		list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,  in list_add_leaf_cfs_rq()
356 			&rq->leaf_cfs_rq_list);  in list_add_leaf_cfs_rq()
361 		rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;  in list_add_leaf_cfs_rq()
371 	list_add_rcu(&cfs_rq->leaf_cfs_rq_list, rq->tmp_alone_branch);  in list_add_leaf_cfs_rq()
376 	rq->tmp_alone_branch = &cfs_rq->leaf_cfs_rq_list;  in list_add_leaf_cfs_rq()
382 	if (cfs_rq->on_list) {  in list_del_leaf_cfs_rq()
389 		 * to the prev element but it will point to rq->leaf_cfs_rq_list  in list_del_leaf_cfs_rq()
392 		if (rq->tmp_alone_branch == &cfs_rq->leaf_cfs_rq_list)  in list_del_leaf_cfs_rq()
393 			rq->tmp_alone_branch = cfs_rq->leaf_cfs_rq_list.prev;  in list_del_leaf_cfs_rq()
395 		list_del_rcu(&cfs_rq->leaf_cfs_rq_list);  in list_del_leaf_cfs_rq()
396 		cfs_rq->on_list = 0;  in list_del_leaf_cfs_rq()
402 	SCHED_WARN_ON(rq->tmp_alone_branch != &rq->leaf_cfs_rq_list);  in assert_list_leaf_cfs_rq()
407 	list_for_each_entry_safe(cfs_rq, pos, &rq->leaf_cfs_rq_list,	\
414 	if (se->cfs_rq == pse->cfs_rq)  in is_same_group()
415 		return se->cfs_rq;  in is_same_group()
422 	return se->parent;  in parent_entity()
438 	se_depth = (*se)->depth;  in find_matching_se()
439 	pse_depth = (*pse)->depth;  in find_matching_se()
442 		se_depth--;  in find_matching_se()
447 		pse_depth--;  in find_matching_se()
459 	return tg->idle > 0;  in tg_is_idle()
464 	return cfs_rq->idle > 0;  in cfs_rq_is_idle()
493 		for (cfs_rq = &rq->cfs, pos = NULL; cfs_rq; cfs_rq = pos)
531 	s64 delta = (s64)(vruntime - max_vruntime);  in max_vruntime()
540 	s64 delta = (s64)(vruntime - min_vruntime);  in min_vruntime()
554 	return (s64)(a->deadline - b->deadline) < 0;  in entity_before()
559 	return (s64)(se->vruntime - cfs_rq->min_vruntime);  in entity_key()
566  * Compute virtual time from the per-task service numbers:
574  *   lag_i = S - s_i = w_i * (V - v_i)
580  *   \Sum w_i * (V - v_i) = 0
581  *   \Sum w_i * V - w_i * v_i = 0
584  * se->vruntime):
587  *   V = -------------- = --------------
594  *          virtual time has non-contiguous motion equivalent to:
596  *	      V +-= lag_i / W
603  * Substitute: v_i == (v_i - v0) + v0
605  *     \Sum ((v_i - v0) + v0) * w_i   \Sum (v_i - v0) * w_i
606  * V = ---------------------------- = --------------------- + v0
611  *                    v0 := cfs_rq->min_vruntime
612  * \Sum (v_i - v0) * w_i := cfs_rq->avg_vruntime
613  *              \Sum w_i := cfs_rq->avg_load
616  * the per-task service, these deltas: (v_i - v), will be in the order of the
626 	unsigned long weight = scale_load_down(se->load.weight);  in avg_vruntime_add()
629 	cfs_rq->avg_vruntime += key * weight;  in avg_vruntime_add()
630 	cfs_rq->avg_load += weight;  in avg_vruntime_add()
636 	unsigned long weight = scale_load_down(se->load.weight);  in avg_vruntime_sub()
639 	cfs_rq->avg_vruntime -= key * weight;  in avg_vruntime_sub()
640 	cfs_rq->avg_load -= weight;  in avg_vruntime_sub()
647 	 * v' = v + d ==> avg_vruntime' = avg_runtime - d*avg_load  in avg_vruntime_update()
649 	cfs_rq->avg_vruntime -= cfs_rq->avg_load * delta;  in avg_vruntime_update()
658 	struct sched_entity *curr = cfs_rq->curr;  in avg_vruntime()
659 	s64 avg = cfs_rq->avg_vruntime;  in avg_vruntime()
660 	long load = cfs_rq->avg_load;  in avg_vruntime()
662 	if (curr && curr->on_rq) {  in avg_vruntime()
663 		unsigned long weight = scale_load_down(curr->load.weight);  in avg_vruntime()
672 			avg -= (load - 1);  in avg_vruntime()
676 	return cfs_rq->min_vruntime + avg;  in avg_vruntime()
680  * lag_i = S - s_i = w_i * (V - v_i)
683  * is possible -- by addition/removal/reweight to the tree -- to move V around
691  *   -r_max < lag < max(r_max, q)
699 	SCHED_WARN_ON(!se->on_rq);  in update_entity_lag()
701 	vlag = avg_vruntime(cfs_rq) - se->vruntime;  in update_entity_lag()
702 	limit = calc_delta_fair(max_t(u64, 2*se->slice, TICK_NSEC), se);  in update_entity_lag()
704 	se->vlag = clamp(vlag, -limit, limit);  in update_entity_lag()
711  * lag_i = S - s_i = w_i*(V - v_i)
713  * lag_i >= 0 -> V >= v_i
715  *     \Sum (v_i - v)*w_i
716  * V = ------------------ + v
719  * lag_i >= 0 -> \Sum (v_i - v)*w_i >= (v_i - v)*(\Sum w_i)
721  * Note: using 'avg_vruntime() > se->vruntime' is inaccurate due
726 	struct sched_entity *curr = cfs_rq->curr;  in vruntime_eligible()
727 	s64 avg = cfs_rq->avg_vruntime;  in vruntime_eligible()
728 	long load = cfs_rq->avg_load;  in vruntime_eligible()
730 	if (curr && curr->on_rq) {  in vruntime_eligible()
731 		unsigned long weight = scale_load_down(curr->load.weight);  in vruntime_eligible()
737 	return avg >= (s64)(vruntime - cfs_rq->min_vruntime) * load;  in vruntime_eligible()
742 	return vruntime_eligible(cfs_rq, se->vruntime);  in entity_eligible()
747 	u64 min_vruntime = cfs_rq->min_vruntime;  in __update_min_vruntime()
751 	s64 delta = (s64)(vruntime - min_vruntime);  in __update_min_vruntime()
762 	struct sched_entity *curr = cfs_rq->curr;  in update_min_vruntime()
763 	u64 vruntime = cfs_rq->min_vruntime;  in update_min_vruntime()
766 		if (curr->on_rq)  in update_min_vruntime()
767 			vruntime = curr->vruntime;  in update_min_vruntime()
774 			vruntime = se->min_vruntime;  in update_min_vruntime()
776 			vruntime = min_vruntime(vruntime, se->min_vruntime);  in update_min_vruntime()
780 	cfs_rq->min_vruntime = __update_min_vruntime(cfs_rq, vruntime);  in update_min_vruntime()
786 	struct sched_entity *curr = cfs_rq->curr;  in cfs_rq_min_slice()
789 	if (curr && curr->on_rq)  in cfs_rq_min_slice()
790 		min_slice = curr->slice;  in cfs_rq_min_slice()
793 		min_slice = min(min_slice, root->min_slice);  in cfs_rq_min_slice()
803 #define vruntime_gt(field, lse, rse) ({ (s64)((lse)->field - (rse)->field) > 0; })
810 			se->min_vruntime = rse->min_vruntime;  in __min_vruntime_update()
818 		if (rse->min_slice < se->min_slice)  in __min_slice_update()
819 			se->min_slice = rse->min_slice;  in __min_slice_update()
824  * se->min_vruntime = min(se->vruntime, {left,right}->min_vruntime)
828 	u64 old_min_vruntime = se->min_vruntime;  in min_vruntime_update()
829 	u64 old_min_slice = se->min_slice;  in min_vruntime_update()
830 	struct rb_node *node = &se->run_node;  in min_vruntime_update()
832 	se->min_vruntime = se->vruntime;  in min_vruntime_update()
833 	__min_vruntime_update(se, node->rb_right);  in min_vruntime_update()
834 	__min_vruntime_update(se, node->rb_left);  in min_vruntime_update()
836 	se->min_slice = se->slice;  in min_vruntime_update()
837 	__min_slice_update(se, node->rb_right);  in min_vruntime_update()
838 	__min_slice_update(se, node->rb_left);  in min_vruntime_update()
840 	return se->min_vruntime == old_min_vruntime &&  in min_vruntime_update()
841 	       se->min_slice == old_min_slice;  in min_vruntime_update()
848  * Enqueue an entity into the rb-tree:
853 	se->min_vruntime = se->vruntime;  in __enqueue_entity()
854 	se->min_slice = se->slice;  in __enqueue_entity()
855 	rb_add_augmented_cached(&se->run_node, &cfs_rq->tasks_timeline,  in __enqueue_entity()
861 	rb_erase_augmented_cached(&se->run_node, &cfs_rq->tasks_timeline,  in __dequeue_entity()
868 	struct rb_node *root = cfs_rq->tasks_timeline.rb_root.rb_node;  in __pick_root_entity()
878 	struct rb_node *left = rb_first_cached(&cfs_rq->tasks_timeline);  in __pick_first_entity()
892 	se->vlag = se->deadline;  in set_protect_slice()
897 	return se->vlag == se->deadline;  in protect_slice()
903 		se->vlag = se->deadline + 1;  in cancel_protect_slice()
917  * We can do this in O(log n) time due to an augmented RB-tree. The
921  *  se->min_vruntime = min(se->vruntime, se->{left,right}->min_vruntime)
927 	struct rb_node *node = cfs_rq->tasks_timeline.rb_root.rb_node;  in pick_eevdf()
929 	struct sched_entity *curr = cfs_rq->curr;  in pick_eevdf()
936 	if (cfs_rq->nr_queued == 1)  in pick_eevdf()
937 		return curr && curr->on_rq ? curr : se;  in pick_eevdf()
939 	if (curr && (!curr->on_rq || !entity_eligible(cfs_rq, curr)))  in pick_eevdf()
953 		struct rb_node *left = node->rb_left;  in pick_eevdf()
960 					__node_2_se(left)->min_vruntime)) {  in pick_eevdf()
977 		node = node->rb_right;  in pick_eevdf()
989 	struct rb_node *last = rb_last(&cfs_rq->tasks_timeline.rb_root);  in __pick_last_entity()
1023 	if ((s64)(se->vruntime - se->deadline) < 0)  in update_deadline()
1031 	if (!se->custom_slice)  in update_deadline()
1032 		se->slice = sysctl_sched_base_slice;  in update_deadline()
1037 	se->deadline = se->vruntime + calc_delta_fair(se->slice, se);  in update_deadline()
1055 	struct sched_avg *sa = &se->avg;  in init_entity_runnable_average()
1066 		sa->load_avg = scale_load_down(se->load.weight);  in init_entity_runnable_average()
1075  *   util_avg = cfs_rq->avg.util_avg / (cfs_rq->avg.load_avg + 1)
1085  *   util_avg_cap = (cpu_scale - cfs_rq->avg.util_avg) / 2^n
1100 	struct sched_entity *se = &p->se;  in post_init_entity_util_avg()
1102 	struct sched_avg *sa = &se->avg;  in post_init_entity_util_avg()
1104 	long cap = (long)(cpu_scale - cfs_rq->avg.util_avg) / 2;  in post_init_entity_util_avg()
1106 	if (p->sched_class != &fair_sched_class) {  in post_init_entity_util_avg()
1117 		se->avg.last_update_time = cfs_rq_clock_pelt(cfs_rq);  in post_init_entity_util_avg()
1122 		if (cfs_rq->avg.util_avg != 0) {  in post_init_entity_util_avg()
1123 			sa->util_avg  = cfs_rq->avg.util_avg * se_weight(se);  in post_init_entity_util_avg()
1124 			sa->util_avg /= (cfs_rq->avg.load_avg + 1);  in post_init_entity_util_avg()
1126 			if (sa->util_avg > cap)  in post_init_entity_util_avg()
1127 				sa->util_avg = cap;  in post_init_entity_util_avg()
1129 			sa->util_avg = cap;  in post_init_entity_util_avg()
1133 	sa->runnable_avg = sa->util_avg;  in post_init_entity_util_avg()
1153 	delta_exec = now - curr->exec_start;  in update_curr_se()
1157 	curr->exec_start = now;  in update_curr_se()
1158 	curr->sum_exec_runtime += delta_exec;  in update_curr_se()
1164 		__schedstat_set(stats->exec_max,  in update_curr_se()
1165 				max(delta_exec, stats->exec_max));  in update_curr_se()
1183 	if (curr->vlag == curr->deadline)  in did_preempt_short()
1195 	if (pse->slice >= se->slice)  in do_preempt_short()
1215 	struct task_struct *donor = rq->donor;  in update_curr_common()
1218 	delta_exec = update_curr_se(rq, &donor->se);  in update_curr_common()
1230 	struct sched_entity *curr = cfs_rq->curr;  in update_curr()
1242 	curr->vruntime += calc_delta_fair(delta_exec, curr);  in update_curr()
1255 		 *  - If the task is running on behalf of fair_server, we need  in update_curr()
1257 		 *  - Fair task that runs outside of fair_server should account  in update_curr()
1261 		if (dl_server_active(&rq->fair_server))  in update_curr()
1262 			dl_server_update(&rq->fair_server, delta_exec);  in update_curr()
1267 	if (cfs_rq->nr_queued == 1)  in update_curr()
1278 	update_curr(cfs_rq_of(&rq->donor->se));  in update_curr_fair()
1311 	 * maybe already in the runqueue, the se->statistics.wait_start  in update_stats_wait_end_fair()
1315 	if (unlikely(!schedstat_val(stats->wait_start)))  in update_stats_wait_end_fair()
1342  * Task is being enqueued - update stats:
1354 	if (se != cfs_rq->curr)  in update_stats_enqueue_fair()
1372 	if (se != cfs_rq->curr)  in update_stats_dequeue_fair()
1380 		state = READ_ONCE(tsk->__state);  in update_stats_dequeue_fair()
1382 			__schedstat_set(tsk->stats.sleep_start,  in update_stats_dequeue_fair()
1385 			__schedstat_set(tsk->stats.block_start,  in update_stats_dequeue_fair()
1391  * We are picking a new current task - update its stats:
1399 	se->exec_start = rq_clock_task(rq_of(cfs_rq));  in update_stats_curr_start()
1440 	 * Allow a small imbalance based on a simple pair of communicating  in adjust_numa_imbalance()
1491  * ->numa_group (see struct task_struct for locking rules).
1495 	return rcu_dereference_check(p->numa_group, p == current ||  in deref_task_numa_group()
1496 		(lockdep_is_held(__rq_lockp(task_rq(p))) && !READ_ONCE(p->on_cpu)));  in deref_task_numa_group()
1501 	return rcu_dereference_protected(p->numa_group, p == current);  in deref_curr_numa_group()
1513 	 * Calculations based on RSS as non-present and empty pages are skipped  in task_nr_scan_windows()
1517 	nr_scan_pages = sysctl_numa_balancing_scan_size << (20 - PAGE_SHIFT);  in task_nr_scan_windows()
1518 	rss = get_mm_rss(p->mm);  in task_nr_scan_windows()
1551 	ng = rcu_dereference(p->numa_group);  in task_scan_start()
1556 		period *= refcount_read(&ng->refcount);  in task_scan_start()
1581 		period *= refcount_read(&ng->refcount);  in task_scan_max()
1593 	rq->nr_numa_running += (p->numa_preferred_nid != NUMA_NO_NODE);  in account_numa_enqueue()
1594 	rq->nr_preferred_running += (p->numa_preferred_nid == task_node(p));  in account_numa_enqueue()
1599 	rq->nr_numa_running -= (p->numa_preferred_nid != NUMA_NO_NODE);  in account_numa_dequeue()
1600 	rq->nr_preferred_running -= (p->numa_preferred_nid == task_node(p));  in account_numa_dequeue()
1618 	ng = rcu_dereference(p->numa_group);  in task_numa_group_id()
1620 		gid = ng->gid;  in task_numa_group_id()
1639 	if (!p->numa_faults)  in task_faults()
1642 	return p->numa_faults[task_faults_idx(NUMA_MEM, nid, 0)] +  in task_faults()
1643 		p->numa_faults[task_faults_idx(NUMA_MEM, nid, 1)];  in task_faults()
1653 	return ng->faults[task_faults_idx(NUMA_MEM, nid, 0)] +  in group_faults()
1654 		ng->faults[task_faults_idx(NUMA_MEM, nid, 1)];  in group_faults()
1659 	return group->faults[task_faults_idx(NUMA_CPU, nid, 0)] +  in group_faults_cpu()
1660 		group->faults[task_faults_idx(NUMA_CPU, nid, 1)];  in group_faults_cpu()
1669 		faults += ng->faults[task_faults_idx(NUMA_MEM, node, 1)];  in group_faults_priv()
1681 		faults += ng->faults[task_faults_idx(NUMA_MEM, node, 0)];  in group_faults_shared()
1689  * considered part of a numa group's pseudo-interleaving set. Migrations
1696 	return group_faults_cpu(ng, nid) * ACTIVE_NODE_FRACTION > ng->max_faults_cpu;  in numa_is_active_node()
1755 			faults *= (max_dist - dist);  in score_nearby_nodes()
1756 			faults /= (max_dist - LOCAL_DISTANCE);  in score_nearby_nodes()
1776 	if (!p->numa_faults)  in task_weight()
1779 	total_faults = p->total_numa_faults;  in task_weight()
1799 	total_faults = ng->total_faults;  in group_weight()
1835 			   pgdat->node_present_pages >> 4);  in pgdat_free_space_enough()
1836 	for (z = pgdat->nr_zones - 1; z >= 0; z--) {  in pgdat_free_space_enough()
1837 		struct zone *zone = pgdat->node_zones + z;  in pgdat_free_space_enough()
1857  *	hint page fault latency = hint page fault time - scan time
1869 	return (time - last_time) & PAGE_ACCESS_TIME_MASK;  in numa_hint_fault_latency()
1886 	start = pgdat->nbp_rl_start;  in numa_promotion_rate_limit()
1887 	if (now - start > MSEC_PER_SEC &&  in numa_promotion_rate_limit()
1888 	    cmpxchg(&pgdat->nbp_rl_start, start, now) == start)  in numa_promotion_rate_limit()
1889 		pgdat->nbp_rl_nr_cand = nr_cand;  in numa_promotion_rate_limit()
1890 	if (nr_cand - pgdat->nbp_rl_nr_cand >= rate_limit)  in numa_promotion_rate_limit()
1906 	start = pgdat->nbp_th_start;  in numa_promotion_adjust_threshold()
1907 	if (now - start > th_period &&  in numa_promotion_adjust_threshold()
1908 	    cmpxchg(&pgdat->nbp_th_start, start, now) == start) {  in numa_promotion_adjust_threshold()
1912 		diff_cand = nr_cand - pgdat->nbp_th_nr_cand;  in numa_promotion_adjust_threshold()
1914 		th = pgdat->nbp_threshold ? : ref_th;  in numa_promotion_adjust_threshold()
1916 			th = max(th - unit_th, unit_th);  in numa_promotion_adjust_threshold()
1919 		pgdat->nbp_th_nr_cand = nr_cand;  in numa_promotion_adjust_threshold()
1920 		pgdat->nbp_threshold = th;  in numa_promotion_adjust_threshold()
1949 			pgdat->nbp_threshold = 0;  in should_numa_migrate_memory()
1955 			(20 - PAGE_SHIFT);  in should_numa_migrate_memory()
1958 		th = pgdat->nbp_threshold ? : def_th;  in should_numa_migrate_memory()
1967 	this_cpupid = cpu_pid_to_cpupid(dst_cpu, current->pid);  in should_numa_migrate_memory()
1977 	 * two full passes of the "multi-stage node selection" test that is  in should_numa_migrate_memory()
1980 	if ((p->numa_preferred_nid == NUMA_NO_NODE || p->numa_scan_seq <= 4) &&  in should_numa_migrate_memory()
1985 	 * Multi-stage node selection is used in conjunction with a periodic  in should_numa_migrate_memory()
1986 	 * migration fault to build a temporal task<->page relation. By using  in should_numa_migrate_memory()
1987 	 * a two-stage filter we remove short/unlikely relations.  in should_numa_migrate_memory()
1991 	 * page (n_t) (in a given time-span) to a probability.  in should_numa_migrate_memory()
1999 	 * act on an unlikely task<->page relation.  in should_numa_migrate_memory()
2009 	/* A shared fault, but p->numa_group has not been set up yet. */  in should_numa_migrate_memory()
2026 	 * --------------- * - > ---------------  in should_numa_migrate_memory()
2088 	if ((ns->nr_running > ns->weight) &&  in numa_classify()
2089 	    (((ns->compute_capacity * 100) < (ns->util * imbalance_pct)) ||  in numa_classify()
2090 	     ((ns->compute_capacity * imbalance_pct) < (ns->runnable * 100))))  in numa_classify()
2093 	if ((ns->nr_running < ns->weight) ||  in numa_classify()
2094 	    (((ns->compute_capacity * 100) > (ns->util * imbalance_pct)) &&  in numa_classify()
2095 	     ((ns->compute_capacity * imbalance_pct) > (ns->runnable * 100))))  in numa_classify()
2136 	int cpu, idle_core = -1;  in update_numa_stats()
2139 	ns->idle_cpu = -1;  in update_numa_stats()
2145 		ns->load += cpu_load(rq);  in update_numa_stats()
2146 		ns->runnable += cpu_runnable(rq);  in update_numa_stats()
2147 		ns->util += cpu_util_cfs(cpu);  in update_numa_stats()
2148 		ns->nr_running += rq->cfs.h_nr_runnable;  in update_numa_stats()
2149 		ns->compute_capacity += capacity_of(cpu);  in update_numa_stats()
2151 		if (find_idle && idle_core < 0 && !rq->nr_running && idle_cpu(cpu)) {  in update_numa_stats()
2152 			if (READ_ONCE(rq->numa_migrate_on) ||  in update_numa_stats()
2153 			    !cpumask_test_cpu(cpu, env->p->cpus_ptr))  in update_numa_stats()
2156 			if (ns->idle_cpu == -1)  in update_numa_stats()
2157 				ns->idle_cpu = cpu;  in update_numa_stats()
2164 	ns->weight = cpumask_weight(cpumask_of_node(nid));  in update_numa_stats()
2166 	ns->node_type = numa_classify(env->imbalance_pct, ns);  in update_numa_stats()
2169 		ns->idle_cpu = idle_core;  in update_numa_stats()
2175 	struct rq *rq = cpu_rq(env->dst_cpu);  in task_numa_assign()
2177 	/* Check if run-queue part of active NUMA balance. */  in task_numa_assign()
2178 	if (env->best_cpu != env->dst_cpu && xchg(&rq->numa_migrate_on, 1)) {  in task_numa_assign()
2180 		int start = env->dst_cpu;  in task_numa_assign()
2183 		for_each_cpu_wrap(cpu, cpumask_of_node(env->dst_nid), start + 1) {  in task_numa_assign()
2184 			if (cpu == env->best_cpu || !idle_cpu(cpu) ||  in task_numa_assign()
2185 			    !cpumask_test_cpu(cpu, env->p->cpus_ptr)) {  in task_numa_assign()
2189 			env->dst_cpu = cpu;  in task_numa_assign()
2190 			rq = cpu_rq(env->dst_cpu);  in task_numa_assign()
2191 			if (!xchg(&rq->numa_migrate_on, 1))  in task_numa_assign()
2201 	 * Clear previous best_cpu/rq numa-migrate flag, since task now  in task_numa_assign()
2204 	if (env->best_cpu != -1 && env->best_cpu != env->dst_cpu) {  in task_numa_assign()
2205 		rq = cpu_rq(env->best_cpu);  in task_numa_assign()
2206 		WRITE_ONCE(rq->numa_migrate_on, 0);  in task_numa_assign()
2209 	if (env->best_task)  in task_numa_assign()
2210 		put_task_struct(env->best_task);  in task_numa_assign()
2214 	env->best_task = p;  in task_numa_assign()
2215 	env->best_imp = imp;  in task_numa_assign()
2216 	env->best_cpu = env->dst_cpu;  in task_numa_assign()
2230 	 * ------------ vs ---------  in load_too_imbalanced()
2233 	src_capacity = env->src_stats.compute_capacity;  in load_too_imbalanced()
2234 	dst_capacity = env->dst_stats.compute_capacity;  in load_too_imbalanced()
2236 	imb = abs(dst_load * src_capacity - src_load * dst_capacity);  in load_too_imbalanced()
2238 	orig_src_load = env->src_stats.load;  in load_too_imbalanced()
2239 	orig_dst_load = env->dst_stats.load;  in load_too_imbalanced()
2241 	old_imb = abs(orig_dst_load * src_capacity - orig_src_load * dst_capacity);  in load_too_imbalanced()
2263 	struct numa_group *cur_ng, *p_ng = deref_curr_numa_group(env->p);  in task_numa_compare()
2264 	struct rq *dst_rq = cpu_rq(env->dst_cpu);  in task_numa_compare()
2268 	int dist = env->dist;  in task_numa_compare()
2273 	if (READ_ONCE(dst_rq->numa_migrate_on))  in task_numa_compare()
2277 	cur = rcu_dereference(dst_rq->curr);  in task_numa_compare()
2278 	if (cur && ((cur->flags & PF_EXITING) || is_idle_task(cur)))  in task_numa_compare()
2283 	 * end try selecting ourselves (current == env->p) as a swap candidate.  in task_numa_compare()
2285 	if (cur == env->p) {  in task_numa_compare()
2291 		if (maymove && moveimp >= env->best_imp)  in task_numa_compare()
2298 	if (!cpumask_test_cpu(env->src_cpu, cur->cpus_ptr))  in task_numa_compare()
2305 	if (env->best_task &&  in task_numa_compare()
2306 	    env->best_task->numa_preferred_nid == env->src_nid &&  in task_numa_compare()
2307 	    cur->numa_preferred_nid != env->src_nid) {  in task_numa_compare()
2312 	 * "imp" is the fault differential for the source task between the  in task_numa_compare()
2313 	 * source and destination node. Calculate the total differential for  in task_numa_compare()
2321 	cur_ng = rcu_dereference(cur->numa_group);  in task_numa_compare()
2329 		if (env->dst_stats.node_type == node_has_spare)  in task_numa_compare()
2332 		imp = taskimp + task_weight(cur, env->src_nid, dist) -  in task_numa_compare()
2333 		      task_weight(cur, env->dst_nid, dist);  in task_numa_compare()
2339 			imp -= imp / 16;  in task_numa_compare()
2346 			imp += group_weight(cur, env->src_nid, dist) -  in task_numa_compare()
2347 			       group_weight(cur, env->dst_nid, dist);  in task_numa_compare()
2349 			imp += task_weight(cur, env->src_nid, dist) -  in task_numa_compare()
2350 			       task_weight(cur, env->dst_nid, dist);  in task_numa_compare()
2354 	if (cur->numa_preferred_nid == env->dst_nid)  in task_numa_compare()
2355 		imp -= imp / 16;  in task_numa_compare()
2363 	if (cur->numa_preferred_nid == env->src_nid)  in task_numa_compare()
2366 	if (maymove && moveimp > imp && moveimp > env->best_imp) {  in task_numa_compare()
2376 	if (env->best_task && cur->numa_preferred_nid == env->src_nid &&  in task_numa_compare()
2377 	    env->best_task->numa_preferred_nid != env->src_nid) {  in task_numa_compare()
2387 	if (imp < SMALLIMP || imp <= env->best_imp + SMALLIMP / 2)  in task_numa_compare()
2393 	load = task_h_load(env->p) - task_h_load(cur);  in task_numa_compare()
2397 	dst_load = env->dst_stats.load + load;  in task_numa_compare()
2398 	src_load = env->src_stats.load - load;  in task_numa_compare()
2406 		int cpu = env->dst_stats.idle_cpu;  in task_numa_compare()
2410 			cpu = env->dst_cpu;  in task_numa_compare()
2416 		if (!idle_cpu(cpu) && env->best_cpu >= 0 &&  in task_numa_compare()
2417 		    idle_cpu(env->best_cpu)) {  in task_numa_compare()
2418 			cpu = env->best_cpu;  in task_numa_compare()
2421 		env->dst_cpu = cpu;  in task_numa_compare()
2431 	if (maymove && !cur && env->best_cpu >= 0 && idle_cpu(env->best_cpu))  in task_numa_compare()
2438 	if (!maymove && env->best_task &&  in task_numa_compare()
2439 	    env->best_task->numa_preferred_nid == env->src_nid) {  in task_numa_compare()
2458 	if (env->dst_stats.node_type == node_has_spare) {  in task_numa_find_cpu()
2468 		src_running = env->src_stats.nr_running - 1;  in task_numa_find_cpu()
2469 		dst_running = env->dst_stats.nr_running + 1;  in task_numa_find_cpu()
2470 		imbalance = max(0, dst_running - src_running);  in task_numa_find_cpu()
2472 						  env->imb_numa_nr);  in task_numa_find_cpu()
2477 			if (env->dst_stats.idle_cpu >= 0) {  in task_numa_find_cpu()
2478 				env->dst_cpu = env->dst_stats.idle_cpu;  in task_numa_find_cpu()
2486 		 * If the improvement from just moving env->p direction is better  in task_numa_find_cpu()
2489 		load = task_h_load(env->p);  in task_numa_find_cpu()
2490 		dst_load = env->dst_stats.load + load;  in task_numa_find_cpu()
2491 		src_load = env->src_stats.load - load;  in task_numa_find_cpu()
2495 	for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) {  in task_numa_find_cpu()
2497 		if (!cpumask_test_cpu(cpu, env->p->cpus_ptr))  in task_numa_find_cpu()
2500 		env->dst_cpu = cpu;  in task_numa_find_cpu()
2518 		.best_cpu = -1,  in task_numa_migrate()
2532 	 * random movement of tasks -- counter the numa conditions we're trying  in task_numa_migrate()
2538 		env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2;  in task_numa_migrate()
2539 		env.imb_numa_nr = sd->imb_numa_nr;  in task_numa_migrate()
2551 		return -EINVAL;  in task_numa_migrate()
2554 	env.dst_nid = p->numa_preferred_nid;  in task_numa_migrate()
2559 	taskimp = task_weight(p, env.dst_nid, dist) - taskweight;  in task_numa_migrate()
2560 	groupimp = group_weight(p, env.dst_nid, dist) - groupweight;  in task_numa_migrate()
2568 	 * - there is no space available on the preferred_nid  in task_numa_migrate()
2569 	 * - the task is part of a numa_group that is interleaved across  in task_numa_migrate()
2574 	if (env.best_cpu == -1 || (ng && ng->active_nodes > 1)) {  in task_numa_migrate()
2576 			if (nid == env.src_nid || nid == p->numa_preferred_nid)  in task_numa_migrate()
2587 			taskimp = task_weight(p, nid, dist) - taskweight;  in task_numa_migrate()
2588 			groupimp = group_weight(p, nid, dist) - groupweight;  in task_numa_migrate()
2608 		if (env.best_cpu == -1)  in task_numa_migrate()
2613 		if (nid != p->numa_preferred_nid)  in task_numa_migrate()
2618 	if (env.best_cpu == -1) {  in task_numa_migrate()
2619 		trace_sched_stick_numa(p, env.src_cpu, NULL, -1);  in task_numa_migrate()
2620 		return -EAGAIN;  in task_numa_migrate()
2626 		WRITE_ONCE(best_rq->numa_migrate_on, 0);  in task_numa_migrate()
2633 	WRITE_ONCE(best_rq->numa_migrate_on, 0);  in task_numa_migrate()
2647 	if (unlikely(p->numa_preferred_nid == NUMA_NO_NODE || !p->numa_faults))  in numa_migrate_preferred()
2651 	interval = min(interval, msecs_to_jiffies(p->numa_scan_period) / 16);  in numa_migrate_preferred()
2652 	p->numa_migrate_retry = jiffies + interval;  in numa_migrate_preferred()
2655 	if (task_node(p) == p->numa_preferred_nid)  in numa_migrate_preferred()
2685 	numa_group->max_faults_cpu = max_faults;  in numa_group_count_active_nodes()
2686 	numa_group->active_nodes = active_nodes;  in numa_group_count_active_nodes()
2712 	unsigned long remote = p->numa_faults_locality[0];  in update_task_scan_period()
2713 	unsigned long local = p->numa_faults_locality[1];  in update_task_scan_period()
2722 	if (local + shared == 0 || p->numa_faults_locality[2]) {  in update_task_scan_period()
2723 		p->numa_scan_period = min(p->numa_scan_period_max,  in update_task_scan_period()
2724 			p->numa_scan_period << 1);  in update_task_scan_period()
2726 		p->mm->numa_next_scan = jiffies +  in update_task_scan_period()
2727 			msecs_to_jiffies(p->numa_scan_period);  in update_task_scan_period()
2738 	period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS);  in update_task_scan_period()
2747 		int slot = ps_ratio - NUMA_PERIOD_THRESHOLD;  in update_task_scan_period()
2757 		int slot = lr_ratio - NUMA_PERIOD_THRESHOLD;  in update_task_scan_period()
2763 		 * Private memory faults exceed (SLOTS-THRESHOLD)/SLOTS,  in update_task_scan_period()
2768 		diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot;  in update_task_scan_period()
2771 	p->numa_scan_period = clamp(p->numa_scan_period + diff,  in update_task_scan_period()
2773 	memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));  in update_task_scan_period()
2780  * from the dozens-of-seconds NUMA balancing period. Use the scheduler
2787 	now = p->se.exec_start;  in numa_get_avg_runtime()
2788 	runtime = p->se.sum_exec_runtime;  in numa_get_avg_runtime()
2790 	if (p->last_task_numa_placement) {  in numa_get_avg_runtime()
2791 		delta = runtime - p->last_sum_exec_runtime;  in numa_get_avg_runtime()
2792 		*period = now - p->last_task_numa_placement;  in numa_get_avg_runtime()
2798 		delta = p->se.avg.load_sum;  in numa_get_avg_runtime()
2802 	p->last_sum_exec_runtime = runtime;  in numa_get_avg_runtime()
2803 	p->last_task_numa_placement = now;  in numa_get_avg_runtime()
2853 	for (dist = sched_max_numa_distance; dist > LOCAL_DISTANCE; dist--) {  in preferred_group_nid()
2907 	 * The p->mm->numa_scan_seq field gets updated without  in task_numa_placement()
2911 	seq = READ_ONCE(p->mm->numa_scan_seq);  in task_numa_placement()
2912 	if (p->numa_scan_seq == seq)  in task_numa_placement()
2914 	p->numa_scan_seq = seq;  in task_numa_placement()
2915 	p->numa_scan_period_max = task_scan_max(p);  in task_numa_placement()
2917 	total_faults = p->numa_faults_locality[0] +  in task_numa_placement()
2918 		       p->numa_faults_locality[1];  in task_numa_placement()
2924 		group_lock = &ng->lock;  in task_numa_placement()
2944 			diff = p->numa_faults[membuf_idx] - p->numa_faults[mem_idx] / 2;  in task_numa_placement()
2945 			fault_types[priv] += p->numa_faults[membuf_idx];  in task_numa_placement()
2946 			p->numa_faults[membuf_idx] = 0;  in task_numa_placement()
2952 			 * little over-all impact on throughput, and thus their  in task_numa_placement()
2956 			f_weight = (f_weight * p->numa_faults[cpubuf_idx]) /  in task_numa_placement()
2958 			f_diff = f_weight - p->numa_faults[cpu_idx] / 2;  in task_numa_placement()
2959 			p->numa_faults[cpubuf_idx] = 0;  in task_numa_placement()
2961 			p->numa_faults[mem_idx] += diff;  in task_numa_placement()
2962 			p->numa_faults[cpu_idx] += f_diff;  in task_numa_placement()
2963 			faults += p->numa_faults[mem_idx];  in task_numa_placement()
2964 			p->total_numa_faults += diff;  in task_numa_placement()
2973 				ng->faults[mem_idx] += diff;  in task_numa_placement()
2974 				ng->faults[cpu_idx] += f_diff;  in task_numa_placement()
2975 				ng->total_faults += diff;  in task_numa_placement()
2976 				group_faults += ng->faults[mem_idx];  in task_numa_placement()
2991 	/* Cannot migrate task to CPU-less node */  in task_numa_placement()
3002 		if (max_nid != p->numa_preferred_nid)  in task_numa_placement()
3011 	return refcount_inc_not_zero(&grp->refcount);  in get_numa_group()
3016 	if (refcount_dec_and_test(&grp->refcount))  in put_numa_group()
3038 		refcount_set(&grp->refcount, 1);  in task_numa_group()
3039 		grp->active_nodes = 1;  in task_numa_group()
3040 		grp->max_faults_cpu = 0;  in task_numa_group()
3041 		spin_lock_init(&grp->lock);  in task_numa_group()
3042 		grp->gid = p->pid;  in task_numa_group()
3045 			grp->faults[i] = p->numa_faults[i];  in task_numa_group()
3047 		grp->total_faults = p->total_numa_faults;  in task_numa_group()
3049 		grp->nr_tasks++;  in task_numa_group()
3050 		rcu_assign_pointer(p->numa_group, grp);  in task_numa_group()
3054 	tsk = READ_ONCE(cpu_rq(cpu)->curr);  in task_numa_group()
3059 	grp = rcu_dereference(tsk->numa_group);  in task_numa_group()
3071 	if (my_grp->nr_tasks > grp->nr_tasks)  in task_numa_group()
3075 	 * Tie-break on the grp address.  in task_numa_group()
3077 	if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp)  in task_numa_group()
3081 	if (tsk->mm == current->mm)  in task_numa_group()
3100 	double_lock_irq(&my_grp->lock, &grp->lock);  in task_numa_group()
3103 		my_grp->faults[i] -= p->numa_faults[i];  in task_numa_group()
3104 		grp->faults[i] += p->numa_faults[i];  in task_numa_group()
3106 	my_grp->total_faults -= p->total_numa_faults;  in task_numa_group()
3107 	grp->total_faults += p->total_numa_faults;  in task_numa_group()
3109 	my_grp->nr_tasks--;  in task_numa_group()
3110 	grp->nr_tasks++;  in task_numa_group()
3112 	spin_unlock(&my_grp->lock);  in task_numa_group()
3113 	spin_unlock_irq(&grp->lock);  in task_numa_group()
3115 	rcu_assign_pointer(p->numa_group, grp);  in task_numa_group()
3130  * reset the data back to default state without freeing ->numa_faults.
3135 	struct numa_group *grp = rcu_dereference_raw(p->numa_group);  in task_numa_free()
3136 	unsigned long *numa_faults = p->numa_faults;  in task_numa_free()
3144 		spin_lock_irqsave(&grp->lock, flags);  in task_numa_free()
3146 			grp->faults[i] -= p->numa_faults[i];  in task_numa_free()
3147 		grp->total_faults -= p->total_numa_faults;  in task_numa_free()
3149 		grp->nr_tasks--;  in task_numa_free()
3150 		spin_unlock_irqrestore(&grp->lock, flags);  in task_numa_free()
3151 		RCU_INIT_POINTER(p->numa_group, NULL);  in task_numa_free()
3156 		p->numa_faults = NULL;  in task_numa_free()
3159 		p->total_numa_faults = 0;  in task_numa_free()
3181 	if (!p->mm)  in task_numa_fault()
3193 	/* Allocate buffer to track faults on a per-node basis */  in task_numa_fault()
3194 	if (unlikely(!p->numa_faults)) {  in task_numa_fault()
3195 		int size = sizeof(*p->numa_faults) *  in task_numa_fault()
3198 		p->numa_faults = kzalloc(size, GFP_KERNEL|__GFP_NOWARN);  in task_numa_fault()
3199 		if (!p->numa_faults)  in task_numa_fault()
3202 		p->total_numa_faults = 0;  in task_numa_fault()
3203 		memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));  in task_numa_fault()
3210 	if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) {  in task_numa_fault()
3225 	if (!priv && !local && ng && ng->active_nodes > 1 &&  in task_numa_fault()
3234 	if (time_after(jiffies, p->numa_migrate_retry)) {  in task_numa_fault()
3240 		p->numa_pages_migrated += pages;  in task_numa_fault()
3242 		p->numa_faults_locality[2] += pages;  in task_numa_fault()
3244 	p->numa_faults[task_faults_idx(NUMA_MEMBUF, mem_node, priv)] += pages;  in task_numa_fault()
3245 	p->numa_faults[task_faults_idx(NUMA_CPUBUF, cpu_node, priv)] += pages;  in task_numa_fault()
3246 	p->numa_faults_locality[local] += pages;  in task_numa_fault()
3253 	 * p->mm->numa_scan_seq is written to without exclusive access  in reset_ptenuma_scan()
3259 	WRITE_ONCE(p->mm->numa_scan_seq, READ_ONCE(p->mm->numa_scan_seq) + 1);  in reset_ptenuma_scan()
3260 	p->mm->numa_scan_offset = 0;  in reset_ptenuma_scan()
3272 	if ((READ_ONCE(current->mm->numa_scan_seq) - vma->numab_state->start_scan_seq) < 2)  in vma_is_accessed()
3275 	pids = vma->numab_state->pids_active[0] | vma->numab_state->pids_active[1];  in vma_is_accessed()
3276 	if (test_bit(hash_32(current->pid, ilog2(BITS_PER_LONG)), &pids))  in vma_is_accessed()
3281 	 * some VMAs may never be scanned in multi-threaded applications:  in vma_is_accessed()
3283 	if (mm->numa_scan_offset > vma->vm_start) {  in vma_is_accessed()
3293 	if (READ_ONCE(mm->numa_scan_seq) >  in vma_is_accessed()
3294 	   (vma->numab_state->prev_scan_seq + get_nr_threads(current)))  in vma_is_accessed()
3310 	struct mm_struct *mm = p->mm;  in task_numa_work()
3311 	u64 runtime = p->se.sum_exec_runtime;  in task_numa_work()
3322 	work->next = work;  in task_numa_work()
3326 	 * NOTE: make sure not to dereference p->mm before this check,  in task_numa_work()
3328 	 * without p->mm even though we still had it when we enqueued this  in task_numa_work()
3331 	if (p->flags & PF_EXITING)  in task_numa_work()
3334 	if (!mm->numa_next_scan) {  in task_numa_work()
3335 		mm->numa_next_scan = now +  in task_numa_work()
3342 	migrate = mm->numa_next_scan;  in task_numa_work()
3346 	if (p->numa_scan_period == 0) {  in task_numa_work()
3347 		p->numa_scan_period_max = task_scan_max(p);  in task_numa_work()
3348 		p->numa_scan_period = task_scan_start(p);  in task_numa_work()
3351 	next_scan = now + msecs_to_jiffies(p->numa_scan_period);  in task_numa_work()
3352 	if (!try_cmpxchg(&mm->numa_next_scan, &migrate, next_scan))  in task_numa_work()
3359 	p->node_stamp += 2 * TICK_NSEC;  in task_numa_work()
3362 	pages <<= 20 - PAGE_SHIFT; /* MB in pages */  in task_numa_work()
3379 	start = mm->numa_scan_offset;  in task_numa_work()
3391 			is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_MIXEDMAP)) {  in task_numa_work()
3399 		 * hinting faults in read-only file-backed mappings or the vDSO  in task_numa_work()
3402 		if (!vma->vm_mm ||  in task_numa_work()
3403 		    (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ))) {  in task_numa_work()
3417 		/* Initialise new per-VMA NUMAB state. */  in task_numa_work()
3418 		if (!vma->numab_state) {  in task_numa_work()
3425 			if (cmpxchg(&vma->numab_state, NULL, ptr)) {  in task_numa_work()
3430 			vma->numab_state->start_scan_seq = mm->numa_scan_seq;  in task_numa_work()
3432 			vma->numab_state->next_scan = now +  in task_numa_work()
3436 			vma->numab_state->pids_active_reset =  vma->numab_state->next_scan +  in task_numa_work()
3444 			 vma->numab_state->prev_scan_seq = mm->numa_scan_seq - 1;  in task_numa_work()
3451 		if (mm->numa_scan_seq && time_before(jiffies,  in task_numa_work()
3452 						vma->numab_state->next_scan)) {  in task_numa_work()
3458 		if (mm->numa_scan_seq &&  in task_numa_work()
3459 				time_after(jiffies, vma->numab_state->pids_active_reset)) {  in task_numa_work()
3460 			vma->numab_state->pids_active_reset = vma->numab_state->pids_active_reset +  in task_numa_work()
3462 			vma->numab_state->pids_active[0] = READ_ONCE(vma->numab_state->pids_active[1]);  in task_numa_work()
3463 			vma->numab_state->pids_active[1] = 0;  in task_numa_work()
3467 		if (vma->numab_state->prev_scan_seq == mm->numa_scan_seq) {  in task_numa_work()
3468 			mm->numa_scan_offset = vma->vm_end;  in task_numa_work()
3484 			start = max(start, vma->vm_start);  in task_numa_work()
3486 			end = min(end, vma->vm_end);  in task_numa_work()
3492 			 * is not already PTE-numa. If the VMA contains  in task_numa_work()
3498 				pages -= (end - start) >> PAGE_SHIFT;  in task_numa_work()
3499 			virtpages -= (end - start) >> PAGE_SHIFT;  in task_numa_work()
3506 		} while (end != vma->vm_end);  in task_numa_work()
3509 		vma->numab_state->prev_scan_seq = mm->numa_scan_seq;  in task_numa_work()
3537 		mm->numa_scan_offset = start;  in task_numa_work()
3548 	if (unlikely(p->se.sum_exec_runtime != runtime)) {  in task_numa_work()
3549 		u64 diff = p->se.sum_exec_runtime - runtime;  in task_numa_work()
3550 		p->node_stamp += 32 * diff;  in task_numa_work()
3557 	struct mm_struct *mm = p->mm;  in init_numa_balancing()
3560 		mm_users = atomic_read(&mm->mm_users);  in init_numa_balancing()
3562 			mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay);  in init_numa_balancing()
3563 			mm->numa_scan_seq = 0;  in init_numa_balancing()
3566 	p->node_stamp			= 0;  in init_numa_balancing()
3567 	p->numa_scan_seq		= mm ? mm->numa_scan_seq : 0;  in init_numa_balancing()
3568 	p->numa_scan_period		= sysctl_numa_balancing_scan_delay;  in init_numa_balancing()
3569 	p->numa_migrate_retry		= 0;  in init_numa_balancing()
3571 	p->numa_work.next		= &p->numa_work;  in init_numa_balancing()
3572 	p->numa_faults			= NULL;  in init_numa_balancing()
3573 	p->numa_pages_migrated		= 0;  in init_numa_balancing()
3574 	p->total_numa_faults		= 0;  in init_numa_balancing()
3575 	RCU_INIT_POINTER(p->numa_group, NULL);  in init_numa_balancing()
3576 	p->last_task_numa_placement	= 0;  in init_numa_balancing()
3577 	p->last_sum_exec_runtime	= 0;  in init_numa_balancing()
3579 	init_task_work(&p->numa_work, task_numa_work);  in init_numa_balancing()
3583 		p->numa_preferred_nid = NUMA_NO_NODE;  in init_numa_balancing()
3595 			current->numa_scan_period * mm_users * NSEC_PER_MSEC);  in init_numa_balancing()
3597 		p->node_stamp = delay;  in init_numa_balancing()
3606 	struct callback_head *work = &curr->numa_work;  in task_tick_numa()
3612 	if (!curr->mm || (curr->flags & (PF_EXITING | PF_KTHREAD)) || work->next != work)  in task_tick_numa()
3621 	now = curr->se.sum_exec_runtime;  in task_tick_numa()
3622 	period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;  in task_tick_numa()
3624 	if (now > curr->node_stamp + period) {  in task_tick_numa()
3625 		if (!curr->node_stamp)  in task_tick_numa()
3626 			curr->numa_scan_period = task_scan_start(curr);  in task_tick_numa()
3627 		curr->node_stamp += period;  in task_tick_numa()
3629 		if (!time_before(jiffies, curr->mm->numa_next_scan))  in task_tick_numa()
3642 	if (!p->mm || !p->numa_faults || (p->flags & PF_EXITING))  in update_scan_period()
3651 	 * is pulled cross-node due to wakeups or load balancing.  in update_scan_period()
3653 	if (p->numa_scan_seq) {  in update_scan_period()
3659 		if (dst_nid == p->numa_preferred_nid ||  in update_scan_period()
3660 		    (p->numa_preferred_nid != NUMA_NO_NODE &&  in update_scan_period()
3661 			src_nid != p->numa_preferred_nid))  in update_scan_period()
3665 	p->numa_scan_period = task_scan_start(p);  in update_scan_period()
3690 	update_load_add(&cfs_rq->load, se->load.weight);  in account_entity_enqueue()
3696 		list_add(&se->group_node, &rq->cfs_tasks);  in account_entity_enqueue()
3699 	cfs_rq->nr_queued++;  in account_entity_enqueue()
3705 	update_load_sub(&cfs_rq->load, se->load.weight);  in account_entity_dequeue()
3709 		list_del_init(&se->group_node);  in account_entity_dequeue()
3712 	cfs_rq->nr_queued--;  in account_entity_dequeue()
3718  * Explicitly do a load-store to ensure the intermediate value never hits
3738  * Explicitly do a load-store to ensure the intermediate value never hits
3746 	res = var - val;					\
3755  * A variant of sub_positive(), which does not use explicit load-store
3760 	*ptr -= min_t(typeof(*ptr), *ptr, _val);		\
3767 	cfs_rq->avg.load_avg += se->avg.load_avg;  in enqueue_load_avg()
3768 	cfs_rq->avg.load_sum += se_weight(se) * se->avg.load_sum;  in enqueue_load_avg()
3774 	sub_positive(&cfs_rq->avg.load_avg, se->avg.load_avg);  in dequeue_load_avg()
3775 	sub_positive(&cfs_rq->avg.load_sum, se_weight(se) * se->avg.load_sum);  in dequeue_load_avg()
3777 	cfs_rq->avg.load_sum = max_t(u32, cfs_rq->avg.load_sum,  in dequeue_load_avg()
3778 					  cfs_rq->avg.load_avg * PELT_MIN_DIVIDER);  in dequeue_load_avg()
3792 	bool curr = cfs_rq->curr == se;  in reweight_entity()
3794 	if (se->on_rq) {  in reweight_entity()
3798 		se->deadline -= se->vruntime;  in reweight_entity()
3799 		se->rel_deadline = 1;  in reweight_entity()
3802 		update_load_sub(&cfs_rq->load, se->load.weight);  in reweight_entity()
3807 	 * Because we keep se->vlag = V - v_i, while: lag_i = w_i*(V - v_i),  in reweight_entity()
3808 	 * we need to scale se->vlag when w_i changes.  in reweight_entity()
3810 	se->vlag = div_s64(se->vlag * se->load.weight, weight);  in reweight_entity()
3811 	if (se->rel_deadline)  in reweight_entity()
3812 		se->deadline = div_s64(se->deadline * se->load.weight, weight);  in reweight_entity()
3814 	update_load_set(&se->load, weight);  in reweight_entity()
3818 		u32 divider = get_pelt_divider(&se->avg);  in reweight_entity()
3820 		se->avg.load_avg = div_u64(se_weight(se) * se->avg.load_sum, divider);  in reweight_entity()
3825 	if (se->on_rq) {  in reweight_entity()
3826 		update_load_add(&cfs_rq->load, se->load.weight);  in reweight_entity()
3833 		 * whether the rq-wide min_vruntime needs updated too. Since  in reweight_entity()
3835 		 * than up-to-date one, we do the update at the end of the  in reweight_entity()
3845 	struct sched_entity *se = &p->se;  in reweight_task_fair()
3847 	struct load_weight *load = &se->load;  in reweight_task_fair()
3849 	reweight_entity(cfs_rq, se, lw->weight);  in reweight_task_fair()
3850 	load->inv_weight = lw->inv_weight;  in reweight_task_fair()
3864  *                     tg->weight * grq->load.weight
3865  *   ge->load.weight = -----------------------------               (1)
3866  *                       \Sum grq->load.weight
3874  *   grq->load.weight -> grq->avg.load_avg                         (2)
3878  *                     tg->weight * grq->avg.load_avg
3879  *   ge->load.weight = ------------------------------              (3)
3880  *                             tg->load_avg
3882  * Where: tg->load_avg ~= \Sum grq->avg.load_avg
3886  * The problem with it is that because the average is slow -- it was designed
3887  * to be exactly that of course -- this leads to transients in boundary
3889  * one task. It takes time for our CPU's grq->avg.load_avg to build up,
3894  *                     tg->weight * grq->load.weight
3895  *   ge->load.weight = ----------------------------- = tg->weight   (4)
3896  *                         grp->load.weight
3903  *   ge->load.weight =
3905  *              tg->weight * grq->load.weight
3906  *     ---------------------------------------------------         (5)
3907  *     tg->load_avg - grq->avg.load_avg + grq->load.weight
3909  * But because grq->load.weight can drop to 0, resulting in a divide by zero,
3910  * we need to use grq->avg.load_avg as its lower bound, which then gives:
3913  *                     tg->weight * grq->load.weight
3914  *   ge->load.weight = -----------------------------		   (6)
3919  *   tg_load_avg' = tg->load_avg - grq->avg.load_avg +
3920  *                  max(grq->load.weight, grq->avg.load_avg)
3924  * overestimates the ge->load.weight and therefore:
3926  *   \Sum ge->load.weight >= tg->weight
3933 	struct task_group *tg = cfs_rq->tg;  in calc_group_shares()
3935 	tg_shares = READ_ONCE(tg->shares);  in calc_group_shares()
3937 	load = max(scale_load_down(cfs_rq->load.weight), cfs_rq->avg.load_avg);  in calc_group_shares()
3939 	tg_weight = atomic_long_read(&tg->load_avg);  in calc_group_shares()
3942 	tg_weight -= cfs_rq->tg_load_avg_contrib;  in calc_group_shares()
3950 	 * MIN_SHARES has to be unscaled here to support per-CPU partitioning  in calc_group_shares()
3951 	 * of a group with small tg->shares value. It is a floor value which is  in calc_group_shares()
3955 	 * E.g. on 64-bit for a group with tg->shares of scale_load(15)=15*1024  in calc_group_shares()
3956 	 * on an 8-core system with 8 tasks each runnable on one CPU shares has  in calc_group_shares()
3978 	if (!gcfs_rq || !gcfs_rq->load.weight)  in update_cfs_group()
3985 	shares = READ_ONCE(gcfs_rq->tg->shares);  in update_cfs_group()
3989 	if (unlikely(se->load.weight != shares))  in update_cfs_group()
4003 	if (&rq->cfs == cfs_rq) {  in cfs_rq_util_change()
4013 		 * As is, the util number is not freq-invariant (we'd have to  in cfs_rq_util_change()
4025 	if (sa->load_sum)  in load_avg_is_decayed()
4028 	if (sa->util_sum)  in load_avg_is_decayed()
4031 	if (sa->runnable_sum)  in load_avg_is_decayed()
4039 	SCHED_WARN_ON(sa->load_avg ||  in load_avg_is_decayed()
4040 		      sa->util_avg ||  in load_avg_is_decayed()
4041 		      sa->runnable_avg);  in load_avg_is_decayed()
4048 	return u64_u32_load_copy(cfs_rq->avg.last_update_time,  in cfs_rq_last_update_time()
4049 				 cfs_rq->last_update_time_copy);  in cfs_rq_last_update_time()
4055  * bottom-up, we only have to test whether the cfs_rq before us on the list
4066 	if (cfs_rq->on_list) {  in child_cfs_rq_on_list()
4067 		prev = cfs_rq->leaf_cfs_rq_list.prev;  in child_cfs_rq_on_list()
4069 		prev = rq->tmp_alone_branch;  in child_cfs_rq_on_list()
4072 	if (prev == &rq->leaf_cfs_rq_list)  in child_cfs_rq_on_list()
4077 	return (prev_cfs_rq->tg->parent == cfs_rq->tg);  in child_cfs_rq_on_list()
4082 	if (cfs_rq->load.weight)  in cfs_rq_is_decayed()
4085 	if (!load_avg_is_decayed(&cfs_rq->avg))  in cfs_rq_is_decayed()
4095  * update_tg_load_avg - update the tg's load avg
4098  * This function 'ensures': tg->load_avg := \Sum tg->cfs_rq[]->avg.load.
4099  * However, because tg->load_avg is a global value there are performance
4103  * differential update where we store the last value we propagated. This in
4104  * turn allows skipping updates if the differential is 'small'.
4116 	if (cfs_rq->tg == &root_task_group)  in update_tg_load_avg()
4124 	 * For migration heavy workloads, access to tg->load_avg can be  in update_tg_load_avg()
4128 	if (now - cfs_rq->last_update_tg_load_avg < NSEC_PER_MSEC)  in update_tg_load_avg()
4131 	delta = cfs_rq->avg.load_avg - cfs_rq->tg_load_avg_contrib;  in update_tg_load_avg()
4132 	if (abs(delta) > cfs_rq->tg_load_avg_contrib / 64) {  in update_tg_load_avg()
4133 		atomic_long_add(delta, &cfs_rq->tg->load_avg);  in update_tg_load_avg()
4134 		cfs_rq->tg_load_avg_contrib = cfs_rq->avg.load_avg;  in update_tg_load_avg()
4135 		cfs_rq->last_update_tg_load_avg = now;  in update_tg_load_avg()
4147 	if (cfs_rq->tg == &root_task_group)  in clear_tg_load_avg()
4151 	delta = 0 - cfs_rq->tg_load_avg_contrib;  in clear_tg_load_avg()
4152 	atomic_long_add(delta, &cfs_rq->tg->load_avg);  in clear_tg_load_avg()
4153 	cfs_rq->tg_load_avg_contrib = 0;  in clear_tg_load_avg()
4154 	cfs_rq->last_update_tg_load_avg = now;  in clear_tg_load_avg()
4173 		struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];  in clear_tg_offline_cfs_rqs()
4184  * caller only guarantees p->pi_lock is held; no other assumptions,
4185  * including the state of rq->lock, should be made.
4199 	 * getting what current time is, so simply throw away the out-of-date  in set_task_rq_fair()
4203 	if (!(se->avg.last_update_time && prev))  in set_task_rq_fair()
4210 	se->avg.last_update_time = n_last_update_time;  in set_task_rq_fair()
4218  *   ge->avg == grq->avg						(1)
4229  *   ge->avg.load_avg = ge->load.weight * ge->avg.runnable_avg		(2)
4234  *   grq->avg.load_avg = grq->load.weight * grq->avg.runnable_avg	(3)
4238  *   ge->avg.runnable_avg == grq->avg.runnable_avg
4242  *                      ge->load.weight * grq->avg.load_avg
4243  *   ge->avg.load_avg = -----------------------------------		(4)
4244  *                               grq->load.weight
4257  * Another reason this doesn't work is that runnable isn't a 0-sum entity.
4268  *   ge->avg.running_sum <= ge->avg.runnable_sum <= LOAD_AVG_MAX
4275  *   grq->avg.runnable_sum = grq->avg.load_sum / grq->load.weight
4283 	long delta_sum, delta_avg = gcfs_rq->avg.util_avg - se->avg.util_avg;  in update_tg_cfs_util()
4291 	 * cfs_rq->avg.period_contrib can be used for both cfs_rq and se.  in update_tg_cfs_util()
4294 	divider = get_pelt_divider(&cfs_rq->avg);  in update_tg_cfs_util()
4298 	se->avg.util_avg = gcfs_rq->avg.util_avg;  in update_tg_cfs_util()
4299 	new_sum = se->avg.util_avg * divider;  in update_tg_cfs_util()
4300 	delta_sum = (long)new_sum - (long)se->avg.util_sum;  in update_tg_cfs_util()
4301 	se->avg.util_sum = new_sum;  in update_tg_cfs_util()
4304 	add_positive(&cfs_rq->avg.util_avg, delta_avg);  in update_tg_cfs_util()
4305 	add_positive(&cfs_rq->avg.util_sum, delta_sum);  in update_tg_cfs_util()
4308 	cfs_rq->avg.util_sum = max_t(u32, cfs_rq->avg.util_sum,  in update_tg_cfs_util()
4309 					  cfs_rq->avg.util_avg * PELT_MIN_DIVIDER);  in update_tg_cfs_util()
4315 	long delta_sum, delta_avg = gcfs_rq->avg.runnable_avg - se->avg.runnable_avg;  in update_tg_cfs_runnable()
4323 	 * cfs_rq->avg.period_contrib can be used for both cfs_rq and se.  in update_tg_cfs_runnable()
4326 	divider = get_pelt_divider(&cfs_rq->avg);  in update_tg_cfs_runnable()
4329 	se->avg.runnable_avg = gcfs_rq->avg.runnable_avg;  in update_tg_cfs_runnable()
4330 	new_sum = se->avg.runnable_avg * divider;  in update_tg_cfs_runnable()
4331 	delta_sum = (long)new_sum - (long)se->avg.runnable_sum;  in update_tg_cfs_runnable()
4332 	se->avg.runnable_sum = new_sum;  in update_tg_cfs_runnable()
4335 	add_positive(&cfs_rq->avg.runnable_avg, delta_avg);  in update_tg_cfs_runnable()
4336 	add_positive(&cfs_rq->avg.runnable_sum, delta_sum);  in update_tg_cfs_runnable()
4338 	cfs_rq->avg.runnable_sum = max_t(u32, cfs_rq->avg.runnable_sum,  in update_tg_cfs_runnable()
4339 					      cfs_rq->avg.runnable_avg * PELT_MIN_DIVIDER);  in update_tg_cfs_runnable()
4345 	long delta_avg, running_sum, runnable_sum = gcfs_rq->prop_runnable_sum;  in update_tg_cfs_load()
4354 	gcfs_rq->prop_runnable_sum = 0;  in update_tg_cfs_load()
4357 	 * cfs_rq->avg.period_contrib can be used for both cfs_rq and se.  in update_tg_cfs_load()
4360 	divider = get_pelt_divider(&cfs_rq->avg);  in update_tg_cfs_load()
4367 		runnable_sum += se->avg.load_sum;  in update_tg_cfs_load()
4374 		if (scale_load_down(gcfs_rq->load.weight)) {  in update_tg_cfs_load()
4375 			load_sum = div_u64(gcfs_rq->avg.load_sum,  in update_tg_cfs_load()
4376 				scale_load_down(gcfs_rq->load.weight));  in update_tg_cfs_load()
4380 		runnable_sum = min(se->avg.load_sum, load_sum);  in update_tg_cfs_load()
4389 	running_sum = se->avg.util_sum >> SCHED_CAPACITY_SHIFT;  in update_tg_cfs_load()
4395 	delta_avg = load_avg - se->avg.load_avg;  in update_tg_cfs_load()
4399 	delta_sum = load_sum - (s64)se_weight(se) * se->avg.load_sum;  in update_tg_cfs_load()
4401 	se->avg.load_sum = runnable_sum;  in update_tg_cfs_load()
4402 	se->avg.load_avg = load_avg;  in update_tg_cfs_load()
4403 	add_positive(&cfs_rq->avg.load_avg, delta_avg);  in update_tg_cfs_load()
4404 	add_positive(&cfs_rq->avg.load_sum, delta_sum);  in update_tg_cfs_load()
4406 	cfs_rq->avg.load_sum = max_t(u32, cfs_rq->avg.load_sum,  in update_tg_cfs_load()
4407 					  cfs_rq->avg.load_avg * PELT_MIN_DIVIDER);  in update_tg_cfs_load()
4412 	cfs_rq->propagate = 1;  in add_tg_cfs_propagate()
4413 	cfs_rq->prop_runnable_sum += runnable_sum;  in add_tg_cfs_propagate()
4425 	if (!gcfs_rq->propagate)  in propagate_entity_load_avg()
4428 	gcfs_rq->propagate = 0;  in propagate_entity_load_avg()
4432 	add_tg_cfs_propagate(cfs_rq, gcfs_rq->prop_runnable_sum);  in propagate_entity_load_avg()
4456 	if (se->avg.load_avg || se->avg.util_avg)  in skip_blocked_update()
4463 	if (gcfs_rq->propagate)  in skip_blocked_update()
4497 	if (load_avg_is_decayed(&se->avg))  in migrate_se_pelt_lag()
4504 	is_idle = is_idle_task(rcu_dereference(rq->curr));  in migrate_se_pelt_lag()
4521 	 *        - cfs->throttled_clock_pelt_time@cfs_rq_idle  in migrate_se_pelt_lag()
4524 	 *      = rq_clock_pelt()@rq_idle - rq_clock_pelt()@cfs_rq_idle  in migrate_se_pelt_lag()
4527 	 *      = sched_clock_cpu() - rq_clock()@rq_idle  in migrate_se_pelt_lag()
4531 	 *    now = rq_clock_pelt()@rq_idle - cfs->throttled_clock_pelt_time +  in migrate_se_pelt_lag()
4532 	 *          sched_clock_cpu() - rq_clock()@rq_idle  in migrate_se_pelt_lag()
4534 	 *      rq_clock_pelt()@rq_idle is rq->clock_pelt_idle  in migrate_se_pelt_lag()
4535 	 *      rq_clock()@rq_idle      is rq->clock_idle  in migrate_se_pelt_lag()
4536 	 *      cfs->throttled_clock_pelt_time@cfs_rq_idle  in migrate_se_pelt_lag()
4537 	 *                              is cfs_rq->throttled_pelt_idle  in migrate_se_pelt_lag()
4541 	throttled = u64_u32_load(cfs_rq->throttled_pelt_idle);  in migrate_se_pelt_lag()
4546 	now = u64_u32_load(rq->clock_pelt_idle);  in migrate_se_pelt_lag()
4556 	now -= throttled;  in migrate_se_pelt_lag()
4559 		 * cfs_rq->avg.last_update_time is more recent than our  in migrate_se_pelt_lag()
4564 		now += sched_clock_cpu(cpu_of(rq)) - u64_u32_load(rq->clock_idle);  in migrate_se_pelt_lag()
4573  * update_cfs_rq_load_avg - update the cfs_rq's load/util averages
4580  * cfs_rq->avg is used for task_h_load() and update_cfs_share() for example.
4584  * Since both these conditions indicate a changed cfs_rq->avg.load we should
4591 	struct sched_avg *sa = &cfs_rq->avg;  in update_cfs_rq_load_avg()
4594 	if (cfs_rq->removed.nr) {  in update_cfs_rq_load_avg()
4596 		u32 divider = get_pelt_divider(&cfs_rq->avg);  in update_cfs_rq_load_avg()
4598 		raw_spin_lock(&cfs_rq->removed.lock);  in update_cfs_rq_load_avg()
4599 		swap(cfs_rq->removed.util_avg, removed_util);  in update_cfs_rq_load_avg()
4600 		swap(cfs_rq->removed.load_avg, removed_load);  in update_cfs_rq_load_avg()
4601 		swap(cfs_rq->removed.runnable_avg, removed_runnable);  in update_cfs_rq_load_avg()
4602 		cfs_rq->removed.nr = 0;  in update_cfs_rq_load_avg()
4603 		raw_spin_unlock(&cfs_rq->removed.lock);  in update_cfs_rq_load_avg()
4606 		sub_positive(&sa->load_avg, r);  in update_cfs_rq_load_avg()
4607 		sub_positive(&sa->load_sum, r * divider);  in update_cfs_rq_load_avg()
4608 		/* See sa->util_sum below */  in update_cfs_rq_load_avg()
4609 		sa->load_sum = max_t(u32, sa->load_sum, sa->load_avg * PELT_MIN_DIVIDER);  in update_cfs_rq_load_avg()
4612 		sub_positive(&sa->util_avg, r);  in update_cfs_rq_load_avg()
4613 		sub_positive(&sa->util_sum, r * divider);  in update_cfs_rq_load_avg()
4615 		 * Because of rounding, se->util_sum might ends up being +1 more than  in update_cfs_rq_load_avg()
4616 		 * cfs->util_sum. Although this is not a problem by itself, detaching  in update_cfs_rq_load_avg()
4618 		 * util_avg (~1ms) can make cfs->util_sum becoming null whereas  in update_cfs_rq_load_avg()
4625 		sa->util_sum = max_t(u32, sa->util_sum, sa->util_avg * PELT_MIN_DIVIDER);  in update_cfs_rq_load_avg()
4628 		sub_positive(&sa->runnable_avg, r);  in update_cfs_rq_load_avg()
4629 		sub_positive(&sa->runnable_sum, r * divider);  in update_cfs_rq_load_avg()
4630 		/* See sa->util_sum above */  in update_cfs_rq_load_avg()
4631 		sa->runnable_sum = max_t(u32, sa->runnable_sum,  in update_cfs_rq_load_avg()
4632 					      sa->runnable_avg * PELT_MIN_DIVIDER);  in update_cfs_rq_load_avg()
4639 			-(long)(removed_runnable * divider) >> SCHED_CAPACITY_SHIFT);  in update_cfs_rq_load_avg()
4645 	u64_u32_store_copy(sa->last_update_time,  in update_cfs_rq_load_avg()
4646 			   cfs_rq->last_update_time_copy,  in update_cfs_rq_load_avg()
4647 			   sa->last_update_time);  in update_cfs_rq_load_avg()
4652  * attach_entity_load_avg - attach this entity to its cfs_rq load avg
4657  * cfs_rq->avg.last_update_time being current.
4662 	 * cfs_rq->avg.period_contrib can be used for both cfs_rq and se.  in attach_entity_load_avg()
4665 	u32 divider = get_pelt_divider(&cfs_rq->avg);  in attach_entity_load_avg()
4674 	se->avg.last_update_time = cfs_rq->avg.last_update_time;  in attach_entity_load_avg()
4675 	se->avg.period_contrib = cfs_rq->avg.period_contrib;  in attach_entity_load_avg()
4683 	se->avg.util_sum = se->avg.util_avg * divider;  in attach_entity_load_avg()
4685 	se->avg.runnable_sum = se->avg.runnable_avg * divider;  in attach_entity_load_avg()
4687 	se->avg.load_sum = se->avg.load_avg * divider;  in attach_entity_load_avg()
4688 	if (se_weight(se) < se->avg.load_sum)  in attach_entity_load_avg()
4689 		se->avg.load_sum = div_u64(se->avg.load_sum, se_weight(se));  in attach_entity_load_avg()
4691 		se->avg.load_sum = 1;  in attach_entity_load_avg()
4694 	cfs_rq->avg.util_avg += se->avg.util_avg;  in attach_entity_load_avg()
4695 	cfs_rq->avg.util_sum += se->avg.util_sum;  in attach_entity_load_avg()
4696 	cfs_rq->avg.runnable_avg += se->avg.runnable_avg;  in attach_entity_load_avg()
4697 	cfs_rq->avg.runnable_sum += se->avg.runnable_sum;  in attach_entity_load_avg()
4699 	add_tg_cfs_propagate(cfs_rq, se->avg.load_sum);  in attach_entity_load_avg()
4707  * detach_entity_load_avg - detach this entity from its cfs_rq load avg
4712  * cfs_rq->avg.last_update_time being current.
4717 	sub_positive(&cfs_rq->avg.util_avg, se->avg.util_avg);  in detach_entity_load_avg()
4718 	sub_positive(&cfs_rq->avg.util_sum, se->avg.util_sum);  in detach_entity_load_avg()
4720 	cfs_rq->avg.util_sum = max_t(u32, cfs_rq->avg.util_sum,  in detach_entity_load_avg()
4721 					  cfs_rq->avg.util_avg * PELT_MIN_DIVIDER);  in detach_entity_load_avg()
4723 	sub_positive(&cfs_rq->avg.runnable_avg, se->avg.runnable_avg);  in detach_entity_load_avg()
4724 	sub_positive(&cfs_rq->avg.runnable_sum, se->avg.runnable_sum);  in detach_entity_load_avg()
4726 	cfs_rq->avg.runnable_sum = max_t(u32, cfs_rq->avg.runnable_sum,  in detach_entity_load_avg()
4727 					      cfs_rq->avg.runnable_avg * PELT_MIN_DIVIDER);  in detach_entity_load_avg()
4729 	add_tg_cfs_propagate(cfs_rq, -se->avg.load_sum);  in detach_entity_load_avg()
4754 	if (se->avg.last_update_time && !(flags & SKIP_AGE_LOAD))  in update_load_avg()
4760 	if (!se->avg.last_update_time && (flags & DO_ATTACH)) {  in update_load_avg()
4810 	 * tasks cannot exit without having gone through wake_up_new_task() ->  in remove_entity_load_avg()
4817 	raw_spin_lock_irqsave(&cfs_rq->removed.lock, flags);  in remove_entity_load_avg()
4818 	++cfs_rq->removed.nr;  in remove_entity_load_avg()
4819 	cfs_rq->removed.util_avg	+= se->avg.util_avg;  in remove_entity_load_avg()
4820 	cfs_rq->removed.load_avg	+= se->avg.load_avg;  in remove_entity_load_avg()
4821 	cfs_rq->removed.runnable_avg	+= se->avg.runnable_avg;  in remove_entity_load_avg()
4822 	raw_spin_unlock_irqrestore(&cfs_rq->removed.lock, flags);  in remove_entity_load_avg()
4827 	return cfs_rq->avg.runnable_avg;  in cfs_rq_runnable_avg()
4832 	return cfs_rq->avg.load_avg;  in cfs_rq_load_avg()
4839 	return READ_ONCE(p->se.avg.util_avg);  in task_util()
4844 	return READ_ONCE(p->se.avg.runnable_avg);  in task_runnable()
4849 	return READ_ONCE(p->se.avg.util_est) & ~UTIL_AVG_UNCHANGED;  in _task_util_est()
4866 	enqueued  = cfs_rq->avg.util_est;  in util_est_enqueue()
4868 	WRITE_ONCE(cfs_rq->avg.util_est, enqueued);  in util_est_enqueue()
4882 	enqueued  = cfs_rq->avg.util_est;  in util_est_dequeue()
4883 	enqueued -= min_t(unsigned int, enqueued, _task_util_est(p));  in util_est_dequeue()
4884 	WRITE_ONCE(cfs_rq->avg.util_est, enqueued);  in util_est_dequeue()
4908 	ewma = READ_ONCE(p->se.avg.util_est);  in util_est_update()
4933 	last_ewma_diff = ewma - dequeued;  in util_est_update()
4959 	 *  ewma(t) = w *  task_util(p) + (1-w) * ewma(t-1)  in util_est_update()
4960 	 *          = w *  task_util(p) +         ewma(t-1)  - w * ewma(t-1)  in util_est_update()
4961 	 *          = w * (task_util(p) -         ewma(t-1)) +     ewma(t-1)  in util_est_update()
4962 	 *          = w * (      -last_ewma_diff           ) +     ewma(t-1)  in util_est_update()
4963 	 *          = w * (-last_ewma_diff +  ewma(t-1) / w)  in util_est_update()
4969 	ewma  -= last_ewma_diff;  in util_est_update()
4973 	WRITE_ONCE(p->se.avg.util_est, ewma);  in util_est_update()
4975 	trace_sched_util_est_se_tp(&p->se);  in util_est_update()
4982 	capacity -= max(hw_load_avg(cpu_rq(cpu)), cpufreq_get_pressure(cpu));  in get_actual_cpu_capacity()
5040 	 *   +----------------------------------------  in util_fits_cpu()
5078 	 *   +----------------------------------------  in util_fits_cpu()
5101 		return -1;  in util_fits_cpu()
5129 	if (!p || (p->nr_cpus_allowed == 1) ||  in update_misfit_status()
5130 	    (arch_scale_cpu_capacity(cpu) == p->max_allowed_capacity) ||  in update_misfit_status()
5133 		rq->misfit_task_load = 0;  in update_misfit_status()
5141 	rq->misfit_task_load = max_t(unsigned long, task_h_load(p), 1);  in update_misfit_status()
5148 	return !cfs_rq->nr_queued;  in cfs_rq_is_decayed()
5188 	struct sched_entity *se = &p->se;  in __setparam_fair()
5190 	p->static_prio = NICE_TO_PRIO(attr->sched_nice);  in __setparam_fair()
5191 	if (attr->sched_runtime) {  in __setparam_fair()
5192 		se->custom_slice = 1;  in __setparam_fair()
5193 		se->slice = clamp_t(u64, attr->sched_runtime,  in __setparam_fair()
5197 		se->custom_slice = 0;  in __setparam_fair()
5198 		se->slice = sysctl_sched_base_slice;  in __setparam_fair()
5208 	if (!se->custom_slice)  in place_entity()
5209 		se->slice = sysctl_sched_base_slice;  in place_entity()
5210 	vslice = calc_delta_fair(se->slice, se);  in place_entity()
5220 	if (sched_feat(PLACE_LAG) && cfs_rq->nr_queued && se->vlag) {  in place_entity()
5221 		struct sched_entity *curr = cfs_rq->curr;  in place_entity()
5224 		lag = se->vlag;  in place_entity()
5234 		 *   lag_i = S - s_i = w_i * (V - v_i)  in place_entity()
5239 		 *   vl_i = V - v_i <=> v_i = V - vl_i  in place_entity()
5251 		 *      = (W*V + w_i*(V - vl_i)) / (W + w_i)  in place_entity()
5252 		 *      = (W*V + w_i*V - w_i*vl_i) / (W + w_i)  in place_entity()
5253 		 *      = (V*(W + w_i) - w_i*l) / (W + w_i)  in place_entity()
5254 		 *      = V - w_i*vl_i / (W + w_i)  in place_entity()
5258 		 *   vl'_i = V' - v_i  in place_entity()
5259 		 *         = V - w_i*vl_i / (W + w_i) - (V - vl_i)  in place_entity()
5260 		 *         = vl_i - w_i*vl_i / (W + w_i)  in place_entity()
5270 		 *   vl'_i = vl_i - w_i*vl_i / (W + w_i)  in place_entity()
5271 		 *         = ((W + w_i)*vl_i - w_i*vl_i) / (W + w_i)  in place_entity()
5273 		 *   (W + w_i)*vl'_i = (W + w_i)*vl_i - w_i*vl_i  in place_entity()
5278 		load = cfs_rq->avg_load;  in place_entity()
5279 		if (curr && curr->on_rq)  in place_entity()
5280 			load += scale_load_down(curr->load.weight);  in place_entity()
5282 		lag *= load + scale_load_down(se->load.weight);  in place_entity()
5288 	se->vruntime = vruntime - lag;  in place_entity()
5290 	if (se->rel_deadline) {  in place_entity()
5291 		se->deadline += se->vruntime;  in place_entity()
5292 		se->rel_deadline = 0;  in place_entity()
5307 	se->deadline = se->vruntime + vslice;  in place_entity()
5319 	bool curr = cfs_rq->curr == se;  in enqueue_entity()
5332 	 *   - Update loads to have both entity and cfs_rq synced with now.  in enqueue_entity()
5333 	 *   - For group_entity, update its runnable_weight to reflect the new  in enqueue_entity()
5335 	 *   - For group_entity, update its weight to reflect the new share of  in enqueue_entity()
5337 	 *   - Add its new weight to cfs_rq->load.weight  in enqueue_entity()
5343 	 * but update_cfs_group() here will re-adjust the weight and have to  in enqueue_entity()
5349 	 * XXX now that the entity has been re-weighted, and it's lag adjusted,  in enqueue_entity()
5359 		se->exec_start = 0;  in enqueue_entity()
5365 	se->on_rq = 1;  in enqueue_entity()
5367 	if (cfs_rq->nr_queued == 1) {  in enqueue_entity()
5375 			if (cfs_rq_throttled(cfs_rq) && !cfs_rq->throttled_clock)  in enqueue_entity()
5376 				cfs_rq->throttled_clock = rq_clock(rq);  in enqueue_entity()
5377 			if (!cfs_rq->throttled_clock_self)  in enqueue_entity()
5378 				cfs_rq->throttled_clock_self = rq_clock(rq);  in enqueue_entity()
5388 		if (cfs_rq->next != se)  in __clear_buddies_next()
5391 		cfs_rq->next = NULL;  in __clear_buddies_next()
5397 	if (cfs_rq->next == se)  in clear_buddies()
5405 	se->sched_delayed = 1;  in set_delayed()
5418 		cfs_rq->h_nr_runnable--;  in set_delayed()
5426 	se->sched_delayed = 0;  in clear_delayed()
5440 		cfs_rq->h_nr_runnable++;  in clear_delayed()
5449 	if (sched_feat(DELAY_ZERO) && se->vlag > 0)  in finish_delayed_dequeue_entity()
5450 		se->vlag = 0;  in finish_delayed_dequeue_entity()
5463 		SCHED_WARN_ON(!se->sched_delayed);  in dequeue_entity()
5473 		SCHED_WARN_ON(delay && se->sched_delayed);  in dequeue_entity()
5488 	 *   - Update loads to have both entity and cfs_rq synced with now.  in dequeue_entity()
5489 	 *   - For group_entity, update its runnable_weight to reflect the new  in dequeue_entity()
5491 	 *   - Subtract its previous weight from cfs_rq->load.weight.  in dequeue_entity()
5492 	 *   - For group entity, update its weight to reflect the new share  in dequeue_entity()
5502 		se->deadline -= se->vruntime;  in dequeue_entity()
5503 		se->rel_deadline = 1;  in dequeue_entity()
5506 	if (se != cfs_rq->curr)  in dequeue_entity()
5508 	se->on_rq = 0;  in dequeue_entity()
5520 	 * further than we started -- i.e. we'll be penalized.  in dequeue_entity()
5528 	if (cfs_rq->nr_queued == 0)  in dequeue_entity()
5540 	if (se->on_rq) {  in set_next_entity()
5554 	SCHED_WARN_ON(cfs_rq->curr);  in set_next_entity()
5555 	cfs_rq->curr = se;  in set_next_entity()
5560 	 * when there are only lesser-weight tasks around):  in set_next_entity()
5563 	    rq_of(cfs_rq)->cfs.load.weight >= 2*se->load.weight) {  in set_next_entity()
5567 		__schedstat_set(stats->slice_max,  in set_next_entity()
5568 				max((u64)stats->slice_max,  in set_next_entity()
5569 				    se->sum_exec_runtime - se->prev_sum_exec_runtime));  in set_next_entity()
5572 	se->prev_sum_exec_runtime = se->sum_exec_runtime;  in set_next_entity()
5590 	 * Picking the ->next buddy will affect latency but not fairness.  in pick_next_entity()
5593 	    cfs_rq->next && entity_eligible(cfs_rq, cfs_rq->next)) {  in pick_next_entity()
5594 		/* ->next will never be delayed */  in pick_next_entity()
5595 		SCHED_WARN_ON(cfs_rq->next->sched_delayed);  in pick_next_entity()
5596 		return cfs_rq->next;  in pick_next_entity()
5600 	if (se->sched_delayed) {  in pick_next_entity()
5618 	if (prev->on_rq)  in put_prev_entity()
5624 	if (prev->on_rq) {  in put_prev_entity()
5631 	SCHED_WARN_ON(cfs_rq->curr != prev);  in put_prev_entity()
5632 	cfs_rq->curr = NULL;  in put_prev_entity()
5639 	 * Update run-time statistics of the 'current'.  in entity_tick()
5711  * directly instead of rq->clock to avoid adding additional synchronization
5712  * around rq->lock.
5714  * requires cfs_b->lock
5720 	if (unlikely(cfs_b->quota == RUNTIME_INF))  in __refill_cfs_bandwidth_runtime()
5723 	cfs_b->runtime += cfs_b->quota;  in __refill_cfs_bandwidth_runtime()
5724 	runtime = cfs_b->runtime_snap - cfs_b->runtime;  in __refill_cfs_bandwidth_runtime()
5726 		cfs_b->burst_time += runtime;  in __refill_cfs_bandwidth_runtime()
5727 		cfs_b->nr_burst++;  in __refill_cfs_bandwidth_runtime()
5730 	cfs_b->runtime = min(cfs_b->runtime, cfs_b->quota + cfs_b->burst);  in __refill_cfs_bandwidth_runtime()
5731 	cfs_b->runtime_snap = cfs_b->runtime;  in __refill_cfs_bandwidth_runtime()
5736 	return &tg->cfs_bandwidth;  in tg_cfs_bandwidth()
5745 	lockdep_assert_held(&cfs_b->lock);  in __assign_cfs_rq_runtime()
5748 	min_amount = target_runtime - cfs_rq->runtime_remaining;  in __assign_cfs_rq_runtime()
5750 	if (cfs_b->quota == RUNTIME_INF)  in __assign_cfs_rq_runtime()
5755 		if (cfs_b->runtime > 0) {  in __assign_cfs_rq_runtime()
5756 			amount = min(cfs_b->runtime, min_amount);  in __assign_cfs_rq_runtime()
5757 			cfs_b->runtime -= amount;  in __assign_cfs_rq_runtime()
5758 			cfs_b->idle = 0;  in __assign_cfs_rq_runtime()
5762 	cfs_rq->runtime_remaining += amount;  in __assign_cfs_rq_runtime()
5764 	return cfs_rq->runtime_remaining > 0;  in __assign_cfs_rq_runtime()
5770 	struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);  in assign_cfs_rq_runtime()
5773 	raw_spin_lock(&cfs_b->lock);  in assign_cfs_rq_runtime()
5775 	raw_spin_unlock(&cfs_b->lock);  in assign_cfs_rq_runtime()
5783 	cfs_rq->runtime_remaining -= delta_exec;  in __account_cfs_rq_runtime()
5785 	if (likely(cfs_rq->runtime_remaining > 0))  in __account_cfs_rq_runtime()
5788 	if (cfs_rq->throttled)  in __account_cfs_rq_runtime()
5794 	if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))  in __account_cfs_rq_runtime()
5801 	if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)  in account_cfs_rq_runtime()
5809 	return cfs_bandwidth_used() && cfs_rq->throttled;  in cfs_rq_throttled()
5815 	return cfs_bandwidth_used() && cfs_rq->throttle_count;  in throttled_hierarchy()
5821  * load-balance operations.
5828 	src_cfs_rq = tg->cfs_rq[src_cpu];  in throttled_lb_pair()
5829 	dest_cfs_rq = tg->cfs_rq[dest_cpu];  in throttled_lb_pair()
5838 	struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];  in tg_unthrottle_up()
5840 	cfs_rq->throttle_count--;  in tg_unthrottle_up()
5841 	if (!cfs_rq->throttle_count) {  in tg_unthrottle_up()
5842 		cfs_rq->throttled_clock_pelt_time += rq_clock_pelt(rq) -  in tg_unthrottle_up()
5843 					     cfs_rq->throttled_clock_pelt;  in tg_unthrottle_up()
5849 		if (cfs_rq->throttled_clock_self) {  in tg_unthrottle_up()
5850 			u64 delta = rq_clock(rq) - cfs_rq->throttled_clock_self;  in tg_unthrottle_up()
5852 			cfs_rq->throttled_clock_self = 0;  in tg_unthrottle_up()
5857 			cfs_rq->throttled_clock_self_time += delta;  in tg_unthrottle_up()
5867 	struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];  in tg_throttle_down()
5870 	if (!cfs_rq->throttle_count) {  in tg_throttle_down()
5871 		cfs_rq->throttled_clock_pelt = rq_clock_pelt(rq);  in tg_throttle_down()
5874 		SCHED_WARN_ON(cfs_rq->throttled_clock_self);  in tg_throttle_down()
5875 		if (cfs_rq->nr_queued)  in tg_throttle_down()
5876 			cfs_rq->throttled_clock_self = rq_clock(rq);  in tg_throttle_down()
5878 	cfs_rq->throttle_count++;  in tg_throttle_down()
5886 	struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);  in throttle_cfs_rq()
5889 	long rq_h_nr_queued = rq->cfs.h_nr_queued;  in throttle_cfs_rq()
5891 	raw_spin_lock(&cfs_b->lock);  in throttle_cfs_rq()
5904 		list_add_tail_rcu(&cfs_rq->throttled_list,  in throttle_cfs_rq()
5905 				  &cfs_b->throttled_cfs_rq);  in throttle_cfs_rq()
5907 	raw_spin_unlock(&cfs_b->lock);  in throttle_cfs_rq()
5912 	se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];  in throttle_cfs_rq()
5916 	walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);  in throttle_cfs_rq()
5919 	queued_delta = cfs_rq->h_nr_queued;  in throttle_cfs_rq()
5920 	runnable_delta = cfs_rq->h_nr_runnable;  in throttle_cfs_rq()
5921 	idle_delta = cfs_rq->h_nr_idle;  in throttle_cfs_rq()
5926 		/* throttled entity or throttle-on-deactivate */  in throttle_cfs_rq()
5927 		if (!se->on_rq)  in throttle_cfs_rq()
5936 		if (se->sched_delayed)  in throttle_cfs_rq()
5941 			idle_delta = cfs_rq->h_nr_queued;  in throttle_cfs_rq()
5943 		qcfs_rq->h_nr_queued -= queued_delta;  in throttle_cfs_rq()
5944 		qcfs_rq->h_nr_runnable -= runnable_delta;  in throttle_cfs_rq()
5945 		qcfs_rq->h_nr_idle -= idle_delta;  in throttle_cfs_rq()
5947 		if (qcfs_rq->load.weight) {  in throttle_cfs_rq()
5948 			/* Avoid re-evaluating load for this entity: */  in throttle_cfs_rq()
5956 		/* throttled entity or throttle-on-deactivate */  in throttle_cfs_rq()
5957 		if (!se->on_rq)  in throttle_cfs_rq()
5964 			idle_delta = cfs_rq->h_nr_queued;  in throttle_cfs_rq()
5966 		qcfs_rq->h_nr_queued -= queued_delta;  in throttle_cfs_rq()
5967 		qcfs_rq->h_nr_runnable -= runnable_delta;  in throttle_cfs_rq()
5968 		qcfs_rq->h_nr_idle -= idle_delta;  in throttle_cfs_rq()
5975 	if (rq_h_nr_queued && !rq->cfs.h_nr_queued)  in throttle_cfs_rq()
5976 		dl_server_stop(&rq->fair_server);  in throttle_cfs_rq()
5980 	 * throttled-list.  rq->lock protects completion.  in throttle_cfs_rq()
5982 	cfs_rq->throttled = 1;  in throttle_cfs_rq()
5983 	SCHED_WARN_ON(cfs_rq->throttled_clock);  in throttle_cfs_rq()
5984 	if (cfs_rq->nr_queued)  in throttle_cfs_rq()
5985 		cfs_rq->throttled_clock = rq_clock(rq);  in throttle_cfs_rq()
5992 	struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);  in unthrottle_cfs_rq()
5995 	long rq_h_nr_queued = rq->cfs.h_nr_queued;  in unthrottle_cfs_rq()
5997 	se = cfs_rq->tg->se[cpu_of(rq)];  in unthrottle_cfs_rq()
5999 	cfs_rq->throttled = 0;  in unthrottle_cfs_rq()
6003 	raw_spin_lock(&cfs_b->lock);  in unthrottle_cfs_rq()
6004 	if (cfs_rq->throttled_clock) {  in unthrottle_cfs_rq()
6005 		cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;  in unthrottle_cfs_rq()
6006 		cfs_rq->throttled_clock = 0;  in unthrottle_cfs_rq()
6008 	list_del_rcu(&cfs_rq->throttled_list);  in unthrottle_cfs_rq()
6009 	raw_spin_unlock(&cfs_b->lock);  in unthrottle_cfs_rq()
6012 	walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);  in unthrottle_cfs_rq()
6014 	if (!cfs_rq->load.weight) {  in unthrottle_cfs_rq()
6015 		if (!cfs_rq->on_list)  in unthrottle_cfs_rq()
6028 	queued_delta = cfs_rq->h_nr_queued;  in unthrottle_cfs_rq()
6029 	runnable_delta = cfs_rq->h_nr_runnable;  in unthrottle_cfs_rq()
6030 	idle_delta = cfs_rq->h_nr_idle;  in unthrottle_cfs_rq()
6035 		if (se->sched_delayed) {  in unthrottle_cfs_rq()
6039 		} else if (se->on_rq)  in unthrottle_cfs_rq()
6044 			idle_delta = cfs_rq->h_nr_queued;  in unthrottle_cfs_rq()
6046 		qcfs_rq->h_nr_queued += queued_delta;  in unthrottle_cfs_rq()
6047 		qcfs_rq->h_nr_runnable += runnable_delta;  in unthrottle_cfs_rq()
6048 		qcfs_rq->h_nr_idle += idle_delta;  in unthrottle_cfs_rq()
6062 			idle_delta = cfs_rq->h_nr_queued;  in unthrottle_cfs_rq()
6064 		qcfs_rq->h_nr_queued += queued_delta;  in unthrottle_cfs_rq()
6065 		qcfs_rq->h_nr_runnable += runnable_delta;  in unthrottle_cfs_rq()
6066 		qcfs_rq->h_nr_idle += idle_delta;  in unthrottle_cfs_rq()
6073 	/* Start the fair server if un-throttling resulted in new runnable tasks */  in unthrottle_cfs_rq()
6074 	if (!rq_h_nr_queued && rq->cfs.h_nr_queued)  in unthrottle_cfs_rq()
6075 		dl_server_start(&rq->fair_server);  in unthrottle_cfs_rq()
6084 	if (rq->curr == rq->idle && rq->cfs.nr_queued)  in unthrottle_cfs_rq()
6108 	 * fact that we pair with sched_free_group_rcu(), so that we cannot  in __cfsb_csd_unthrottle()
6114 	list_for_each_entry_safe(cursor, tmp, &rq->cfsb_csd_list,  in __cfsb_csd_unthrottle()
6116 		list_del_init(&cursor->throttled_csd_list);  in __cfsb_csd_unthrottle()
6139 	if (SCHED_WARN_ON(!list_empty(&cfs_rq->throttled_csd_list)))  in __unthrottle_cfs_rq_async()
6142 	first = list_empty(&rq->cfsb_csd_list);  in __unthrottle_cfs_rq_async()
6143 	list_add_tail(&cfs_rq->throttled_csd_list, &rq->cfsb_csd_list);  in __unthrottle_cfs_rq_async()
6145 		smp_call_function_single_async(cpu_of(rq), &rq->cfsb_csd);  in __unthrottle_cfs_rq_async()
6159 	    cfs_rq->runtime_remaining <= 0))  in unthrottle_cfs_rq_async()
6176 	list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,  in distribute_cfs_runtime()
6190 		if (!list_empty(&cfs_rq->throttled_csd_list))  in distribute_cfs_runtime()
6194 		SCHED_WARN_ON(cfs_rq->runtime_remaining > 0);  in distribute_cfs_runtime()
6196 		raw_spin_lock(&cfs_b->lock);  in distribute_cfs_runtime()
6197 		runtime = -cfs_rq->runtime_remaining + 1;  in distribute_cfs_runtime()
6198 		if (runtime > cfs_b->runtime)  in distribute_cfs_runtime()
6199 			runtime = cfs_b->runtime;  in distribute_cfs_runtime()
6200 		cfs_b->runtime -= runtime;  in distribute_cfs_runtime()
6201 		remaining = cfs_b->runtime;  in distribute_cfs_runtime()
6202 		raw_spin_unlock(&cfs_b->lock);  in distribute_cfs_runtime()
6204 		cfs_rq->runtime_remaining += runtime;  in distribute_cfs_runtime()
6207 		if (cfs_rq->runtime_remaining > 0) {  in distribute_cfs_runtime()
6216 				list_add_tail(&cfs_rq->throttled_csd_list,  in distribute_cfs_runtime()
6233 		list_del_init(&cfs_rq->throttled_csd_list);  in distribute_cfs_runtime()
6250  * period the timer is deactivated until scheduling resumes; cfs_b->idle is
6258 	if (cfs_b->quota == RUNTIME_INF)  in do_sched_cfs_period_timer()
6261 	throttled = !list_empty(&cfs_b->throttled_cfs_rq);  in do_sched_cfs_period_timer()
6262 	cfs_b->nr_periods += overrun;  in do_sched_cfs_period_timer()
6264 	/* Refill extra burst quota even if cfs_b->idle */  in do_sched_cfs_period_timer()
6271 	if (cfs_b->idle && !throttled)  in do_sched_cfs_period_timer()
6276 		cfs_b->idle = 1;  in do_sched_cfs_period_timer()
6281 	cfs_b->nr_throttled += overrun;  in do_sched_cfs_period_timer()
6284 	 * This check is repeated as we release cfs_b->lock while we unthrottle.  in do_sched_cfs_period_timer()
6286 	while (throttled && cfs_b->runtime > 0) {  in do_sched_cfs_period_timer()
6287 		raw_spin_unlock_irqrestore(&cfs_b->lock, flags);  in do_sched_cfs_period_timer()
6288 		/* we can't nest cfs_b->lock while distributing bandwidth */  in do_sched_cfs_period_timer()
6290 		raw_spin_lock_irqsave(&cfs_b->lock, flags);  in do_sched_cfs_period_timer()
6299 	cfs_b->idle = 0;  in do_sched_cfs_period_timer()
6317  * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
6323 	struct hrtimer *refresh_timer = &cfs_b->period_timer;  in runtime_refresh_within()
6326 	/* if the call-back is running a quota refresh is already occurring */  in runtime_refresh_within()
6347 	if (cfs_b->slack_started)  in start_cfs_slack_bandwidth()
6349 	cfs_b->slack_started = true;  in start_cfs_slack_bandwidth()
6351 	hrtimer_start(&cfs_b->slack_timer,  in start_cfs_slack_bandwidth()
6359 	struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);  in __return_cfs_rq_runtime()
6360 	s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;  in __return_cfs_rq_runtime()
6365 	raw_spin_lock(&cfs_b->lock);  in __return_cfs_rq_runtime()
6366 	if (cfs_b->quota != RUNTIME_INF) {  in __return_cfs_rq_runtime()
6367 		cfs_b->runtime += slack_runtime;  in __return_cfs_rq_runtime()
6369 		/* we are under rq->lock, defer unthrottling using a timer */  in __return_cfs_rq_runtime()
6370 		if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&  in __return_cfs_rq_runtime()
6371 		    !list_empty(&cfs_b->throttled_cfs_rq))  in __return_cfs_rq_runtime()
6374 	raw_spin_unlock(&cfs_b->lock);  in __return_cfs_rq_runtime()
6377 	cfs_rq->runtime_remaining -= slack_runtime;  in __return_cfs_rq_runtime()
6385 	if (!cfs_rq->runtime_enabled || cfs_rq->nr_queued)  in return_cfs_rq_runtime()
6393  * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
6401 	raw_spin_lock_irqsave(&cfs_b->lock, flags);  in do_sched_cfs_slack_timer()
6402 	cfs_b->slack_started = false;  in do_sched_cfs_slack_timer()
6405 		raw_spin_unlock_irqrestore(&cfs_b->lock, flags);  in do_sched_cfs_slack_timer()
6409 	if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice)  in do_sched_cfs_slack_timer()
6410 		runtime = cfs_b->runtime;  in do_sched_cfs_slack_timer()
6412 	raw_spin_unlock_irqrestore(&cfs_b->lock, flags);  in do_sched_cfs_slack_timer()
6423  * runtime as update_curr() throttling can not trigger until it's on-rq.
6430 	/* an active group must be handled by the update_curr()->put() path */  in check_enqueue_throttle()
6431 	if (!cfs_rq->runtime_enabled || cfs_rq->curr)  in check_enqueue_throttle()
6440 	if (cfs_rq->runtime_remaining <= 0)  in check_enqueue_throttle()
6451 	if (!tg->parent)  in sync_throttle()
6454 	cfs_rq = tg->cfs_rq[cpu];  in sync_throttle()
6455 	pcfs_rq = tg->parent->cfs_rq[cpu];  in sync_throttle()
6457 	cfs_rq->throttle_count = pcfs_rq->throttle_count;  in sync_throttle()
6458 	cfs_rq->throttled_clock_pelt = rq_clock_pelt(cpu_rq(cpu));  in sync_throttle()
6467 	if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))  in check_cfs_rq_runtime()
6501 	raw_spin_lock_irqsave(&cfs_b->lock, flags);  in sched_cfs_period_timer()
6503 		overrun = hrtimer_forward_now(timer, cfs_b->period);  in sched_cfs_period_timer()
6510 			u64 new, old = ktime_to_ns(cfs_b->period);  in sched_cfs_period_timer()
6519 				cfs_b->period = ns_to_ktime(new);  in sched_cfs_period_timer()
6520 				cfs_b->quota *= 2;  in sched_cfs_period_timer()
6521 				cfs_b->burst *= 2;  in sched_cfs_period_timer()
6527 					div_u64(cfs_b->quota, NSEC_PER_USEC));  in sched_cfs_period_timer()
6533 					div_u64(cfs_b->quota, NSEC_PER_USEC));  in sched_cfs_period_timer()
6541 		cfs_b->period_active = 0;  in sched_cfs_period_timer()
6542 	raw_spin_unlock_irqrestore(&cfs_b->lock, flags);  in sched_cfs_period_timer()
6549 	raw_spin_lock_init(&cfs_b->lock);  in init_cfs_bandwidth()
6550 	cfs_b->runtime = 0;  in init_cfs_bandwidth()
6551 	cfs_b->quota = RUNTIME_INF;  in init_cfs_bandwidth()
6552 	cfs_b->period = ns_to_ktime(default_cfs_period());  in init_cfs_bandwidth()
6553 	cfs_b->burst = 0;  in init_cfs_bandwidth()
6554 	cfs_b->hierarchical_quota = parent ? parent->hierarchical_quota : RUNTIME_INF;  in init_cfs_bandwidth()
6556 	INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);  in init_cfs_bandwidth()
6557 	hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED);  in init_cfs_bandwidth()
6558 	cfs_b->period_timer.function = sched_cfs_period_timer;  in init_cfs_bandwidth()
6561 	hrtimer_set_expires(&cfs_b->period_timer,  in init_cfs_bandwidth()
6562 			    get_random_u32_below(cfs_b->period));  in init_cfs_bandwidth()
6563 	hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);  in init_cfs_bandwidth()
6564 	cfs_b->slack_timer.function = sched_cfs_slack_timer;  in init_cfs_bandwidth()
6565 	cfs_b->slack_started = false;  in init_cfs_bandwidth()
6570 	cfs_rq->runtime_enabled = 0;  in init_cfs_rq_runtime()
6571 	INIT_LIST_HEAD(&cfs_rq->throttled_list);  in init_cfs_rq_runtime()
6572 	INIT_LIST_HEAD(&cfs_rq->throttled_csd_list);  in init_cfs_rq_runtime()
6577 	lockdep_assert_held(&cfs_b->lock);  in start_cfs_bandwidth()
6579 	if (cfs_b->period_active)  in start_cfs_bandwidth()
6582 	cfs_b->period_active = 1;  in start_cfs_bandwidth()
6583 	hrtimer_forward_now(&cfs_b->period_timer, cfs_b->period);  in start_cfs_bandwidth()
6584 	hrtimer_start_expires(&cfs_b->period_timer, HRTIMER_MODE_ABS_PINNED);  in start_cfs_bandwidth()
6592 	if (!cfs_b->throttled_cfs_rq.next)  in destroy_cfs_bandwidth()
6595 	hrtimer_cancel(&cfs_b->period_timer);  in destroy_cfs_bandwidth()
6596 	hrtimer_cancel(&cfs_b->slack_timer);  in destroy_cfs_bandwidth()
6613 		if (list_empty(&rq->cfsb_csd_list))  in destroy_cfs_bandwidth()
6639 		struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;  in update_runtime_enabled()
6640 		struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];  in update_runtime_enabled()
6642 		raw_spin_lock(&cfs_b->lock);  in update_runtime_enabled()
6643 		cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF;  in update_runtime_enabled()
6644 		raw_spin_unlock(&cfs_b->lock);  in update_runtime_enabled()
6669 		struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];  in unthrottle_offline_cfs_rqs()
6671 		if (!cfs_rq->runtime_enabled)  in unthrottle_offline_cfs_rqs()
6678 		cfs_rq->runtime_enabled = 0;  in unthrottle_offline_cfs_rqs()
6687 		cfs_rq->runtime_remaining = 1;  in unthrottle_offline_cfs_rqs()
6702 	if (cfs_rq->runtime_enabled ||  in cfs_task_bw_constrained()
6703 	    tg_cfs_bandwidth(cfs_rq->tg)->hierarchical_quota != RUNTIME_INF)  in cfs_task_bw_constrained()
6721 	if (rq->nr_running != 1)  in sched_fair_update_stop_tick()
6790 	struct sched_entity *se = &p->se;  in hrtick_start_fair()
6794 	if (rq->cfs.h_nr_queued > 1) {  in hrtick_start_fair()
6795 		u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;  in hrtick_start_fair()
6796 		u64 slice = se->slice;  in hrtick_start_fair()
6797 		s64 delta = slice - ran;  in hrtick_start_fair()
6815 	struct task_struct *donor = rq->donor;  in hrtick_update()
6817 	if (!hrtick_enabled_fair(rq) || donor->sched_class != &fair_sched_class)  in hrtick_update()
6853 	return !sched_energy_enabled() || READ_ONCE(rd->overutilized);  in is_rd_overutilized()
6861 	WRITE_ONCE(rd->overutilized, flag);  in set_rd_overutilized()
6872 	if (!is_rd_overutilized(rq->rd) && cpu_overutilized(rq->cpu))  in check_update_overutilized_status()
6873 		set_rd_overutilized(rq->rd, 1);  in check_update_overutilized_status()
6882 	return unlikely(rq->nr_running == rq->cfs.h_nr_idle &&  in sched_idle_rq()
6883 			rq->nr_running);  in sched_idle_rq()
6899 	 * se->sched_delayed should imply: se->on_rq == 1.  in requeue_delayed_entity()
6903 	SCHED_WARN_ON(!se->sched_delayed);  in requeue_delayed_entity()
6904 	SCHED_WARN_ON(!se->on_rq);  in requeue_delayed_entity()
6908 		if (se->vlag > 0) {  in requeue_delayed_entity()
6909 			cfs_rq->nr_queued--;  in requeue_delayed_entity()
6910 			if (se != cfs_rq->curr)  in requeue_delayed_entity()
6912 			se->vlag = 0;  in requeue_delayed_entity()
6914 			if (se != cfs_rq->curr)  in requeue_delayed_entity()
6916 			cfs_rq->nr_queued++;  in requeue_delayed_entity()
6933 	struct sched_entity *se = &p->se;  in enqueue_task_fair()
6937 	int rq_h_nr_queued = rq->cfs.h_nr_queued;  in enqueue_task_fair()
6946 	if (!(p->se.sched_delayed && (task_on_rq_migrating(p) || (flags & ENQUEUE_RESTORE))))  in enqueue_task_fair()
6947 		util_est_enqueue(&rq->cfs, p);  in enqueue_task_fair()
6959 	if (p->in_iowait)  in enqueue_task_fair()
6962 	if (task_new && se->sched_delayed)  in enqueue_task_fair()
6966 		if (se->on_rq) {  in enqueue_task_fair()
6967 			if (se->sched_delayed)  in enqueue_task_fair()
6976 		 * its entities in the desired time-frame.  in enqueue_task_fair()
6979 			se->slice = slice;  in enqueue_task_fair()
6980 			se->custom_slice = 1;  in enqueue_task_fair()
6985 		cfs_rq->h_nr_runnable += h_nr_runnable;  in enqueue_task_fair()
6986 		cfs_rq->h_nr_queued++;  in enqueue_task_fair()
6987 		cfs_rq->h_nr_idle += h_nr_idle;  in enqueue_task_fair()
7006 		se->slice = slice;  in enqueue_task_fair()
7007 		if (se != cfs_rq->curr)  in enqueue_task_fair()
7008 			min_vruntime_cb_propagate(&se->run_node, NULL);  in enqueue_task_fair()
7011 		cfs_rq->h_nr_runnable += h_nr_runnable;  in enqueue_task_fair()
7012 		cfs_rq->h_nr_queued++;  in enqueue_task_fair()
7013 		cfs_rq->h_nr_idle += h_nr_idle;  in enqueue_task_fair()
7023 	if (!rq_h_nr_queued && rq->cfs.h_nr_queued) {  in enqueue_task_fair()
7025 		if (!rq->nr_running)  in enqueue_task_fair()
7026 			dl_server_update_idle_time(rq, rq->curr);  in enqueue_task_fair()
7027 		dl_server_start(&rq->fair_server);  in enqueue_task_fair()
7060  * failing half-way through and resume the dequeue later.
7063  * -1 - dequeue delayed
7064  *  0 - dequeue throttled
7065  *  1 - dequeue complete
7070 	int rq_h_nr_queued = rq->cfs.h_nr_queued;  in dequeue_entities()
7084 		if (task_sleep || task_delayed || !se->sched_delayed)  in dequeue_entities()
7095 			if (p && &p->se == se)  in dequeue_entities()
7096 				return -1;  in dequeue_entities()
7101 		cfs_rq->h_nr_runnable -= h_nr_runnable;  in dequeue_entities()
7102 		cfs_rq->h_nr_queued -= h_nr_queued;  in dequeue_entities()
7103 		cfs_rq->h_nr_idle -= h_nr_idle;  in dequeue_entities()
7113 		if (cfs_rq->load.weight) {  in dequeue_entities()
7116 			/* Avoid re-evaluating load for this entity: */  in dequeue_entities()
7137 		se->slice = slice;  in dequeue_entities()
7138 		if (se != cfs_rq->curr)  in dequeue_entities()
7139 			min_vruntime_cb_propagate(&se->run_node, NULL);  in dequeue_entities()
7142 		cfs_rq->h_nr_runnable -= h_nr_runnable;  in dequeue_entities()
7143 		cfs_rq->h_nr_queued -= h_nr_queued;  in dequeue_entities()
7144 		cfs_rq->h_nr_idle -= h_nr_idle;  in dequeue_entities()
7156 	if (rq_h_nr_queued && !rq->cfs.h_nr_queued)  in dequeue_entities()
7157 		dl_server_stop(&rq->fair_server);  in dequeue_entities()
7161 		rq->next_balance = jiffies;  in dequeue_entities()
7165 		SCHED_WARN_ON(p->on_rq != 1);  in dequeue_entities()
7167 		/* Fix-up what dequeue_task_fair() skipped */  in dequeue_entities()
7171 		 * Fix-up what block_task() skipped.  in dequeue_entities()
7188 	if (!(p->se.sched_delayed && (task_on_rq_migrating(p) || (flags & DEQUEUE_SAVE))))  in dequeue_task_fair()
7189 		util_est_dequeue(&rq->cfs, p);  in dequeue_task_fair()
7191 	util_est_update(&rq->cfs, p, flags & DEQUEUE_SLEEP);  in dequeue_task_fair()
7192 	if (dequeue_entities(rq, &p->se, flags) < 0)  in dequeue_task_fair()
7225 	return cfs_rq_load_avg(&rq->cfs);  in cpu_load()
7229  * cpu_load_without - compute CPU load without any contributions from *p
7247 	if (cpu_of(rq) != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time))  in cpu_load_without()
7250 	cfs_rq = &rq->cfs;  in cpu_load_without()
7251 	load = READ_ONCE(cfs_rq->avg.load_avg);  in cpu_load_without()
7261 	return cfs_rq_runnable_avg(&rq->cfs);  in cpu_runnable()
7270 	if (cpu_of(rq) != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time))  in cpu_runnable_without()
7273 	cfs_rq = &rq->cfs;  in cpu_runnable_without()
7274 	runnable = READ_ONCE(cfs_rq->avg.runnable_avg);  in cpu_runnable_without()
7277 	lsub_positive(&runnable, p->se.avg.runnable_avg);  in cpu_runnable_without()
7284 	return cpu_rq(cpu)->cpu_capacity;  in capacity_of()
7293 	if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) {  in record_wakee()
7294 		current->wakee_flips >>= 1;  in record_wakee()
7295 		current->wakee_flip_decay_ts = jiffies;  in record_wakee()
7298 	if (current->last_wakee != p) {  in record_wakee()
7299 		current->last_wakee = p;  in record_wakee()
7300 		current->wakee_flips++;  in record_wakee()
7305  * Detect M:N waker/wakee relationships via a switching-frequency heuristic.
7315  * non-monogamous, with partner count exceeding socket size.
7323 	unsigned int master = current->wakee_flips;  in wake_wide()
7324 	unsigned int slave = p->wakee_flips;  in wake_wide()
7339  * wake_affine_idle() - only considers 'now', it check if the waking CPU is
7340  *			cache-affine and is (or	will be) idle.
7342  * wake_affine_weight() - considers the weight to reflect the average
7364 	if (sync && cpu_rq(this_cpu)->nr_running == 1)  in wake_affine_idle()
7388 		this_eff_load -= current_load;  in wake_affine_weight()
7399 	prev_eff_load -= task_load;  in wake_affine_weight()
7401 		prev_eff_load *= 100 + (sd->imbalance_pct - 100) / 2;  in wake_affine_weight()
7427 	schedstat_inc(p->stats.nr_wakeups_affine_attempts);  in wake_affine()
7431 	schedstat_inc(sd->ttwu_move_affine);  in wake_affine()
7432 	schedstat_inc(p->stats.nr_wakeups_affine);  in wake_affine()
7440  * sched_balance_find_dst_group_cpu - find the idlest CPU among the CPUs in the group.
7449 	int shallowest_idle_cpu = -1;  in sched_balance_find_dst_group_cpu()
7453 	if (group->group_weight == 1)  in sched_balance_find_dst_group_cpu()
7457 	for_each_cpu_and(i, sched_group_span(group), p->cpus_ptr) {  in sched_balance_find_dst_group_cpu()
7468 			if (idle && idle->exit_latency < min_exit_latency) {  in sched_balance_find_dst_group_cpu()
7474 				min_exit_latency = idle->exit_latency;  in sched_balance_find_dst_group_cpu()
7475 				latest_idle_timestamp = rq->idle_stamp;  in sched_balance_find_dst_group_cpu()
7477 			} else if ((!idle || idle->exit_latency == min_exit_latency) &&  in sched_balance_find_dst_group_cpu()
7478 				   rq->idle_stamp > latest_idle_timestamp) {  in sched_balance_find_dst_group_cpu()
7484 				latest_idle_timestamp = rq->idle_stamp;  in sched_balance_find_dst_group_cpu()
7487 		} else if (shallowest_idle_cpu == -1) {  in sched_balance_find_dst_group_cpu()
7496 	return shallowest_idle_cpu != -1 ? shallowest_idle_cpu : least_loaded_cpu;  in sched_balance_find_dst_group_cpu()
7504 	if (!cpumask_intersects(sched_domain_span(sd), p->cpus_ptr))  in sched_balance_find_dst_cpu()
7512 		sync_entity_load_avg(&p->se);  in sched_balance_find_dst_cpu()
7519 		if (!(sd->flags & sd_flag)) {  in sched_balance_find_dst_cpu()
7520 			sd = sd->child;  in sched_balance_find_dst_cpu()
7526 			sd = sd->child;  in sched_balance_find_dst_cpu()
7533 			sd = sd->child;  in sched_balance_find_dst_cpu()
7539 		weight = sd->span_weight;  in sched_balance_find_dst_cpu()
7542 			if (weight <= tmp->span_weight)  in sched_balance_find_dst_cpu()
7544 			if (tmp->flags & sd_flag)  in sched_balance_find_dst_cpu()
7558 	return -1;  in __select_idle_cpu()
7571 		WRITE_ONCE(sds->has_idle_cores, val);  in set_idle_cores()
7580 		return READ_ONCE(sds->has_idle_cores);  in test_idle_cores()
7587  * information in sd_llc_shared->has_idle_cores.
7617  * sd_llc->shared->has_idle_cores and enabled through update_idle_core() above.
7627 			if (*idle_cpu == -1) {  in select_idle_core()
7636 		if (*idle_cpu == -1 && cpumask_test_cpu(cpu, cpus))  in select_idle_core()
7644 	return -1;  in select_idle_core()
7654 	for_each_cpu_and(cpu, cpu_smt_mask(target), p->cpus_ptr) {  in select_idle_smt()
7667 	return -1;  in select_idle_smt()
7688 	return -1;  in select_idle_smt()
7695  * comparing the average scan cost (tracked in sd->avg_scan_cost) against the
7696  * average idle time for this rq (as found in rq->avg_idle).
7701 	int i, cpu, idle_cpu = -1, nr = INT_MAX;  in select_idle_cpu()
7704 	cpumask_and(cpus, sched_domain_span(sd), p->cpus_ptr);  in select_idle_cpu()
7709 			/* because !--nr is the condition to stop scan */  in select_idle_cpu()
7710 			nr = READ_ONCE(sd_share->nr_idle_scan) + 1;  in select_idle_cpu()
7713 				return -1;  in select_idle_cpu()
7718 		struct sched_group *sg = sd->groups;  in select_idle_cpu()
7720 		if (sg->flags & SD_CLUSTER) {  in select_idle_cpu()
7730 					if (--nr <= 0)  in select_idle_cpu()
7731 						return -1;  in select_idle_cpu()
7748 			if (--nr <= 0)  in select_idle_cpu()
7749 				return -1;  in select_idle_cpu()
7772 	int cpu, best_cpu = -1;  in select_idle_capacity()
7776 	cpumask_and(cpus, sched_domain_span(sd), p->cpus_ptr);  in select_idle_capacity()
7801 		 * First, select CPU which fits better (-1 being better than 0).  in select_idle_capacity()
7838 	int i, recent_used_cpu, prev_aff = -1;  in select_idle_sibling()
7845 		sync_entity_load_avg(&p->se);  in select_idle_sibling()
7852 	 * per-cpu select_rq_mask usage  in select_idle_sibling()
7875 	 * Allow a per-cpu kthread to stack with the wakee if the  in select_idle_sibling()
7878 	 * per-cpu kthread that is now complete and the wakeup is  in select_idle_sibling()
7885 	    this_rq()->nr_running <= 1 &&  in select_idle_sibling()
7891 	recent_used_cpu = p->recent_used_cpu;  in select_idle_sibling()
7892 	p->recent_used_cpu = prev;  in select_idle_sibling()
7897 	    cpumask_test_cpu(recent_used_cpu, p->cpus_ptr) &&  in select_idle_sibling()
7905 		recent_used_cpu = -1;  in select_idle_sibling()
7961  * cpu_util() - Estimates the amount of CPU capacity used by CFS tasks.
7964  * @dst_cpu: CPU @p migrates to, -1 if @p moves from @cpu or @p == NULL
7971  * recent utilization of currently non-runnable tasks on that CPU.
7979  * previously-executed tasks, which helps better deduce how busy a CPU will
7980  * be when a long-sleeping task wakes up. The contribution to CPU utilization
7994  * could be seen as over-utilized even though CPU1 has 20% of spare CPU
7997  * after task migrations (scheduler-driven DVFS).
8004 	struct cfs_rq *cfs_rq = &cpu_rq(cpu)->cfs;  in cpu_util()
8005 	unsigned long util = READ_ONCE(cfs_rq->avg.util_avg);  in cpu_util()
8009 		runnable = READ_ONCE(cfs_rq->avg.runnable_avg);  in cpu_util()
8014 	 * If @dst_cpu is -1 or @p migrates from @cpu to @dst_cpu remove its  in cpu_util()
8027 		util_est = READ_ONCE(cfs_rq->avg.util_est);  in cpu_util()
8030 		 * During wake-up @p isn't enqueued yet and doesn't contribute  in cpu_util()
8031 		 * to any cpu_rq(cpu)->cfs.avg.util_est.  in cpu_util()
8035 		 * During exec (@dst_cpu = -1) @p is enqueued and does  in cpu_util()
8036 		 * contribute to cpu_rq(cpu)->cfs.util_est.  in cpu_util()
8045 		 *       p->on_rq = TASK_ON_RQ_MIGRATING;  in cpu_util()
8046 		 *       -------------------------------- A  in cpu_util()
8050 		 *       -------------------------------- B  in cpu_util()
8068 	return cpu_util(cpu, NULL, -1, 0);  in cpu_util_cfs()
8073 	return cpu_util(cpu, NULL, -1, 1);  in cpu_util_cfs_boost()
8092 	if (cpu != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time))  in cpu_util_without()
8095 	return cpu_util(cpu, p, -1, 0);  in cpu_util_without()
8110  * The cfs,rt,dl utilization are the running times measured with rq->clock_task
8111  * which excludes things like IRQ and steal-time. These latter are then accrued
8129 	 * because of inaccuracies in how we track these -- see  in effective_cpu_util()
8144 		 * - the computed DL bandwidth needed with the IRQ pressure which  in effective_cpu_util()
8146 		 * - The minimum performance requirement for CFS and/or RT.  in effective_cpu_util()
8154 		if (!uclamp_is_used() && rt_rq_is_runnable(&rq->rt))  in effective_cpu_util()
8182 	 *              max - irq  in effective_cpu_util()
8183 	 *   U' = irq + --------- * U  in effective_cpu_util()
8198  * energy_env - Utilization landscape for energy estimation.
8204  * @pd_cap:         Entire perf domain capacity. (pd->nr_cpus * cpu_cap).
8230 	eenv->task_busy_time = busy_time;  in eenv_task_busy_time()
8244  *   - A stable PD utilization, no matter which CPU of that PD we want to place
8247  *   - A fair comparison between CPUs as the task contribution (task_util())
8252  * exceed @eenv->pd_cap.
8262 		unsigned long util = cpu_util(cpu, p, -1, 0);  in eenv_pd_busy_time()
8267 	eenv->pd_busy_time = min(eenv->pd_cap, busy_time);  in eenv_pd_busy_time()
8274  * Returns the maximum utilization among @eenv->cpus. This utilization can't
8275  * exceed @eenv->cpu_cap.
8316 	return min(max_util, eenv->cpu_cap);  in eenv_pd_max_util()
8329 	unsigned long busy_time = eenv->pd_busy_time;  in compute_energy()
8333 		busy_time = min(eenv->pd_cap, busy_time + eenv->task_busy_time);  in compute_energy()
8335 	energy = em_cpu_energy(pd->em_pd, max_util, busy_time, eenv->cpu_cap);  in compute_energy()
8343  * find_energy_efficient_cpu(): Find most energy-efficient target CPU for the
8347  * out which of the CPU candidates is the most energy-efficient.
8364  * cluster-packing, and spreading inside a cluster. That should at least be
8371  * NOTE: Forkees are not accepted in the energy-aware wake-up path because
8375  * to be energy-inefficient in some use-cases. The alternative would be to
8378  * other use-cases too. So, until someone finds a better way to solve this,
8379  * let's keep things simple by re-using the existing slow path.
8387 	struct root_domain *rd = this_rq()->rd;  in find_energy_efficient_cpu()
8388 	int cpu, best_energy_cpu, target = -1;  in find_energy_efficient_cpu()
8389 	int prev_fits = -1, best_fits = -1;  in find_energy_efficient_cpu()
8397 	pd = rcu_dereference(rd->pd);  in find_energy_efficient_cpu()
8402 	 * Energy-aware wake-up happens on the lowest sched_domain starting  in find_energy_efficient_cpu()
8407 		sd = sd->parent;  in find_energy_efficient_cpu()
8413 	sync_entity_load_avg(&p->se);  in find_energy_efficient_cpu()
8419 	for (; pd; pd = pd->next) {  in find_energy_efficient_cpu()
8422 		long prev_spare_cap = -1, max_spare_cap = -1;  in find_energy_efficient_cpu()
8425 		int max_spare_cap_cpu = -1;  in find_energy_efficient_cpu()
8426 		int fits, max_fits = -1;  in find_energy_efficient_cpu()
8448 			if (!cpumask_test_cpu(cpu, p->cpus_ptr))  in find_energy_efficient_cpu()
8467 				 * max-aggregated uclamp_{min, max}.  in find_energy_efficient_cpu()
8504 		base_energy = compute_energy(&eenv, pd, cpus, p, -1);  in find_energy_efficient_cpu()
8507 		if (prev_spare_cap > -1) {  in find_energy_efficient_cpu()
8513 			prev_delta -= base_energy;  in find_energy_efficient_cpu()
8537 			cur_delta -= base_energy;  in find_energy_efficient_cpu()
8581 	int sync = (wake_flags & WF_SYNC) && !(current->flags & PF_EXITING);  in select_task_rq_fair()
8590 	 * required for stable ->cpus_allowed  in select_task_rq_fair()
8592 	lockdep_assert_held(&p->pi_lock);  in select_task_rq_fair()
8597 		    cpumask_test_cpu(cpu, p->cpus_ptr))  in select_task_rq_fair()
8600 		if (!is_rd_overutilized(this_rq()->rd)) {  in select_task_rq_fair()
8607 		want_affine = !wake_wide(p) && cpumask_test_cpu(cpu, p->cpus_ptr);  in select_task_rq_fair()
8616 		if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&  in select_task_rq_fair()
8630 		if (tmp->flags & sd_flag)  in select_task_rq_fair()
8651  * previous CPU. The caller guarantees p->pi_lock or task_rq(p)->lock is held.
8655 	struct sched_entity *se = &p->se;  in migrate_task_rq_fair()
8664 		 * leading to an inflation after wake-up on the new rq.  in migrate_task_rq_fair()
8674 	se->avg.last_update_time = 0;  in migrate_task_rq_fair()
8681 	struct sched_entity *se = &p->se;  in task_dead_fair()
8683 	if (se->sched_delayed) {  in task_dead_fair()
8688 		if (se->sched_delayed) {  in task_dead_fair()
8713 		if (!cpumask_intersects(p->cpus_ptr, cpumask))  in set_task_max_allowed_capacity()
8716 		p->max_allowed_capacity = entry->capacity;  in set_task_max_allowed_capacity()
8743 		if (SCHED_WARN_ON(!se->on_rq))  in set_next_buddy()
8747 		cfs_rq_of(se)->next = se;  in set_next_buddy()
8756 	struct task_struct *donor = rq->donor;  in check_preempt_wakeup_fair()
8757 	struct sched_entity *se = &donor->se, *pse = &p->se;  in check_preempt_wakeup_fair()
8768 	 * next-buddy nomination below.  in check_preempt_wakeup_fair()
8773 	if (sched_feat(NEXT_BUDDY) && !(wake_flags & WF_FORK) && !pse->sched_delayed) {  in check_preempt_wakeup_fair()
8781 	 * Note: this also catches the edge-case of curr being in a throttled  in check_preempt_wakeup_fair()
8787 	if (test_tsk_need_resched(rq->curr))  in check_preempt_wakeup_fair()
8800 	 * Preempt an idle entity in favor of a non-idle entity (and don't preempt  in check_preempt_wakeup_fair()
8805 		 * When non-idle entity preempt an idle entity,  in check_preempt_wakeup_fair()
8818 	if (unlikely(!normal_policy(p->policy)))  in check_preempt_wakeup_fair()
8851 	cfs_rq = &rq->cfs;  in pick_task_fair()
8852 	if (!cfs_rq->nr_queued)  in pick_task_fair()
8857 		if (cfs_rq->curr && cfs_rq->curr->on_rq)  in pick_task_fair()
8886 	se = &p->se;  in pick_next_task_fair()
8889 	if (prev->sched_class != &fair_sched_class)  in pick_next_task_fair()
8906 		struct sched_entity *pse = &prev->se;  in pick_next_task_fair()
8910 			int se_depth = se->depth;  in pick_next_task_fair()
8911 			int pse_depth = pse->depth;  in pick_next_task_fair()
8943 	 * Because sched_balance_newidle() releases (and re-acquires) rq->lock, it is  in pick_next_task_fair()
8945 	 * must re-start the pick_next_entity() loop.  in pick_next_task_fair()
8969 	return !!dl_se->rq->cfs.nr_queued;  in fair_server_has_tasks()
8974 	return pick_task_fair(dl_se->rq);  in fair_server_pick_task()
8979 	struct sched_dl_entity *dl_se = &rq->fair_server;  in fair_server_init()
8991 	struct sched_entity *se = &prev->se;  in put_prev_task_fair()
9005 	struct task_struct *curr = rq->curr;  in yield_task_fair()
9007 	struct sched_entity *se = &curr->se;  in yield_task_fair()
9012 	if (unlikely(rq->nr_running == 1))  in yield_task_fair()
9019 	 * Update run-time statistics of the 'current'.  in yield_task_fair()
9029 	se->deadline += calc_delta_fair(se->slice, se);  in yield_task_fair()
9034 	struct sched_entity *se = &p->se;  in yield_to_task_fair()
9037 	if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))  in yield_to_task_fair()
9050  * Fair scheduling class load-balancing methods.
9054  * The purpose of load-balancing is to achieve the same basic fairness the
9055  * per-CPU scheduler provides, namely provide a proportional amount of compute
9060  * Where W_i,n is the n-th weight average for CPU i. The instantaneous weight
9065  * Where w_i,j is the weight of the j-th runnable task on CPU i. This weight
9071  *   W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0               (3)
9080  *   imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j }    (4)
9087  *     - infeasible weights;
9088  *     - local vs global optima in the discrete case. ]
9098  * of load-balance at each level inversely proportional to the number of CPUs in
9104  *   \Sum       { --- * --- * 2^i } = O(n)                            (5)
9106  *                               `- size of each group
9107  *         |         |     `- number of CPUs doing load-balance
9108  *         |         `- freq
9109  *         `- sum over all levels
9151  *   W_i,0 = \Sum_j \Prod_k w_k * -----                               (9)
9158  * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on CPU i.
9240 	/* The set of CPUs under consideration for load-balancing */
9255  * Is this task likely cache-hot:
9261 	lockdep_assert_rq_held(env->src_rq);  in task_hot()
9263 	if (p->sched_class != &fair_sched_class)  in task_hot()
9270 	if (env->sd->flags & SD_SHARE_CPUCAPACITY)  in task_hot()
9276 	if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running &&  in task_hot()
9277 	    (&p->se == cfs_rq_of(&p->se)->next))  in task_hot()
9280 	if (sysctl_sched_migration_cost == -1)  in task_hot()
9287 	if (!sched_core_cookie_match(cpu_rq(env->dst_cpu), p))  in task_hot()
9293 	delta = rq_clock_task(env->src_rq) - p->se.exec_start;  in task_hot()
9306 	struct numa_group *numa_group = rcu_dereference(p->numa_group);  in migrate_degrades_locality()
9313 	if (!p->numa_faults || !(env->sd->flags & SD_NUMA))  in migrate_degrades_locality()
9316 	src_nid = cpu_to_node(env->src_cpu);  in migrate_degrades_locality()
9317 	dst_nid = cpu_to_node(env->dst_cpu);  in migrate_degrades_locality()
9323 	if (src_nid == p->numa_preferred_nid) {  in migrate_degrades_locality()
9324 		if (env->src_rq->nr_running > env->src_rq->nr_preferred_running)  in migrate_degrades_locality()
9331 	if (dst_nid == p->numa_preferred_nid)  in migrate_degrades_locality()
9332 		return -1;  in migrate_degrades_locality()
9335 	if (env->idle == CPU_IDLE)  in migrate_degrades_locality()
9347 	return src_weight - dst_weight;  in migrate_degrades_locality()
9362  * dst_cfs_rq->nr_queued is greater than 1, if the task
9371 	dst_cfs_rq = task_group(p)->cfs_rq[dest_cpu];  in task_is_ineligible_on_dst_cpu()
9373 	dst_cfs_rq = &cpu_rq(dest_cpu)->cfs;  in task_is_ineligible_on_dst_cpu()
9375 	if (sched_feat(PLACE_LAG) && dst_cfs_rq->nr_queued &&  in task_is_ineligible_on_dst_cpu()
9376 	    !entity_eligible(task_cfs_rq(p), &p->se))  in task_is_ineligible_on_dst_cpu()
9383  * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
9390 	lockdep_assert_rq_held(env->src_rq);  in can_migrate_task()
9391 	if (p->sched_task_hot)  in can_migrate_task()
9392 		p->sched_task_hot = 0;  in can_migrate_task()
9400 	 * 5) are cache-hot on their current CPU.  in can_migrate_task()
9402 	if ((p->se.sched_delayed) && (env->migration_type != migrate_load))  in can_migrate_task()
9405 	if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))  in can_migrate_task()
9410 	 * For ineligible tasks we soft-limit them and only allow  in can_migrate_task()
9411 	 * them to migrate when nr_balance_failed is non-zero to  in can_migrate_task()
9412 	 * avoid load-balancing trying very hard to balance the load.  in can_migrate_task()
9414 	if (!env->sd->nr_balance_failed &&  in can_migrate_task()
9415 	    task_is_ineligible_on_dst_cpu(p, env->dst_cpu))  in can_migrate_task()
9422 	if (!cpumask_test_cpu(env->dst_cpu, p->cpus_ptr)) {  in can_migrate_task()
9425 		schedstat_inc(p->stats.nr_failed_migrations_affine);  in can_migrate_task()
9427 		env->flags |= LBF_SOME_PINNED;  in can_migrate_task()
9435 		 * - for NEWLY_IDLE  in can_migrate_task()
9436 		 * - if we have already computed one in current iteration  in can_migrate_task()
9437 		 * - if it's an active balance  in can_migrate_task()
9439 		if (env->idle == CPU_NEWLY_IDLE ||  in can_migrate_task()
9440 		    env->flags & (LBF_DST_PINNED | LBF_ACTIVE_LB))  in can_migrate_task()
9443 		/* Prevent to re-select dst_cpu via env's CPUs: */  in can_migrate_task()
9444 		for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {  in can_migrate_task()
9445 			if (cpumask_test_cpu(cpu, p->cpus_ptr)) {  in can_migrate_task()
9446 				env->flags |= LBF_DST_PINNED;  in can_migrate_task()
9447 				env->new_dst_cpu = cpu;  in can_migrate_task()
9456 	env->flags &= ~LBF_ALL_PINNED;  in can_migrate_task()
9458 	if (task_on_cpu(env->src_rq, p)) {  in can_migrate_task()
9459 		schedstat_inc(p->stats.nr_failed_migrations_running);  in can_migrate_task()
9470 	if (env->flags & LBF_ACTIVE_LB)  in can_migrate_task()
9479 	if (!hot || env->sd->nr_balance_failed > env->sd->cache_nice_tries) {  in can_migrate_task()
9481 			p->sched_task_hot = 1;  in can_migrate_task()
9485 	schedstat_inc(p->stats.nr_failed_migrations_hot);  in can_migrate_task()
9490  * detach_task() -- detach the task for the migration specified in env
9494 	lockdep_assert_rq_held(env->src_rq);  in detach_task()
9496 	if (p->sched_task_hot) {  in detach_task()
9497 		p->sched_task_hot = 0;  in detach_task()
9498 		schedstat_inc(env->sd->lb_hot_gained[env->idle]);  in detach_task()
9499 		schedstat_inc(p->stats.nr_forced_migrations);  in detach_task()
9502 	deactivate_task(env->src_rq, p, DEQUEUE_NOCLOCK);  in detach_task()
9503 	set_task_cpu(p, env->dst_cpu);  in detach_task()
9507  * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
9516 	lockdep_assert_rq_held(env->src_rq);  in detach_one_task()
9519 			&env->src_rq->cfs_tasks, se.group_node) {  in detach_one_task()
9527 		 * lb_gained[env->idle] is updated (other is detach_tasks)  in detach_one_task()
9531 		schedstat_inc(env->sd->lb_gained[env->idle]);  in detach_one_task()
9538  * detach_tasks() -- tries to detach up to imbalance load/util/tasks from
9545 	struct list_head *tasks = &env->src_rq->cfs_tasks;  in detach_tasks()
9550 	lockdep_assert_rq_held(env->src_rq);  in detach_tasks()
9556 	if (env->src_rq->nr_running <= 1) {  in detach_tasks()
9557 		env->flags &= ~LBF_ALL_PINNED;  in detach_tasks()
9561 	if (env->imbalance <= 0)  in detach_tasks()
9569 		if (env->idle && env->src_rq->nr_running <= 1)  in detach_tasks()
9572 		env->loop++;  in detach_tasks()
9574 		if (env->loop > env->loop_max)  in detach_tasks()
9578 		if (env->loop > env->loop_break) {  in detach_tasks()
9579 			env->loop_break += SCHED_NR_MIGRATE_BREAK;  in detach_tasks()
9580 			env->flags |= LBF_NEED_BREAK;  in detach_tasks()
9589 		switch (env->migration_type) {  in detach_tasks()
9594 			 * value. Make sure that env->imbalance decreases  in detach_tasks()
9601 			    load < 16 && !env->sd->nr_balance_failed)  in detach_tasks()
9610 			if (shr_bound(load, env->sd->nr_balance_failed) > env->imbalance)  in detach_tasks()
9613 			env->imbalance -= load;  in detach_tasks()
9619 			if (shr_bound(util, env->sd->nr_balance_failed) > env->imbalance)  in detach_tasks()
9622 			env->imbalance -= util;  in detach_tasks()
9626 			env->imbalance--;  in detach_tasks()
9631 			if (task_fits_cpu(p, env->src_cpu))  in detach_tasks()
9634 			env->imbalance = 0;  in detach_tasks()
9639 		list_add(&p->se.group_node, &env->tasks);  in detach_tasks()
9649 		if (env->idle == CPU_NEWLY_IDLE)  in detach_tasks()
9657 		if (env->imbalance <= 0)  in detach_tasks()
9662 		if (p->sched_task_hot)  in detach_tasks()
9663 			schedstat_inc(p->stats.nr_failed_migrations_hot);  in detach_tasks()
9665 		list_move(&p->se.group_node, tasks);  in detach_tasks()
9673 	schedstat_add(env->sd->lb_gained[env->idle], detached);  in detach_tasks()
9679  * attach_task() -- attach the task detached by detach_task() to its new rq.
9691  * attach_one_task() -- attaches the task returned from detach_one_task() to
9705  * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
9710 	struct list_head *tasks = &env->tasks;  in attach_tasks()
9714 	rq_lock(env->dst_rq, &rf);  in attach_tasks()
9715 	update_rq_clock(env->dst_rq);  in attach_tasks()
9719 		list_del_init(&p->se.group_node);  in attach_tasks()
9721 		attach_task(env->dst_rq, p);  in attach_tasks()
9724 	rq_unlock(env->dst_rq, &rf);  in attach_tasks()
9730 	if (cfs_rq->avg.load_avg)  in cfs_rq_has_blocked()
9733 	if (cfs_rq->avg.util_avg)  in cfs_rq_has_blocked()
9758 	WRITE_ONCE(rq->last_blocked_load_update_tick, jiffies);  in update_blocked_load_tick()
9764 		rq->has_blocked_load = 0;  in update_blocked_load_status()
9807 			if (cfs_rq->nr_queued == 0)  in __update_blocked_fair()
9810 			if (cfs_rq == &rq->cfs)  in __update_blocked_fair()
9815 		se = cfs_rq->tg->se[cpu];  in __update_blocked_fair()
9836  * This needs to be done in a top-down fashion because the load of a child
9842 	struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];  in update_cfs_rq_h_load()
9846 	if (cfs_rq->last_h_load_update == now)  in update_cfs_rq_h_load()
9849 	WRITE_ONCE(cfs_rq->h_load_next, NULL);  in update_cfs_rq_h_load()
9852 		WRITE_ONCE(cfs_rq->h_load_next, se);  in update_cfs_rq_h_load()
9853 		if (cfs_rq->last_h_load_update == now)  in update_cfs_rq_h_load()
9858 		cfs_rq->h_load = cfs_rq_load_avg(cfs_rq);  in update_cfs_rq_h_load()
9859 		cfs_rq->last_h_load_update = now;  in update_cfs_rq_h_load()
9862 	while ((se = READ_ONCE(cfs_rq->h_load_next)) != NULL) {  in update_cfs_rq_h_load()
9863 		load = cfs_rq->h_load;  in update_cfs_rq_h_load()
9864 		load = div64_ul(load * se->avg.load_avg,  in update_cfs_rq_h_load()
9867 		cfs_rq->h_load = load;  in update_cfs_rq_h_load()
9868 		cfs_rq->last_h_load_update = now;  in update_cfs_rq_h_load()
9877 	return div64_ul(p->se.avg.load_avg * cfs_rq->h_load,  in task_h_load()
9883 	struct cfs_rq *cfs_rq = &rq->cfs;  in __update_blocked_fair()
9895 	return p->se.avg.load_avg;  in task_h_load()
9921  * sg_lb_stats - stats of a sched_group required for load-balancing:
9944  * sd_lb_stats - stats of a sched_domain required for load-balancing:
10001 	free = max - used;  in scale_rt_capacity()
10009 	struct sched_group *sdg = sd->groups;  in update_cpu_capacity()
10014 	cpu_rq(cpu)->cpu_capacity = capacity;  in update_cpu_capacity()
10017 	sdg->sgc->capacity = capacity;  in update_cpu_capacity()
10018 	sdg->sgc->min_capacity = capacity;  in update_cpu_capacity()
10019 	sdg->sgc->max_capacity = capacity;  in update_cpu_capacity()
10024 	struct sched_domain *child = sd->child;  in update_group_capacity()
10025 	struct sched_group *group, *sdg = sd->groups;  in update_group_capacity()
10029 	interval = msecs_to_jiffies(sd->balance_interval);  in update_group_capacity()
10031 	sdg->sgc->next_update = jiffies + interval;  in update_group_capacity()
10042 	if (child->flags & SD_OVERLAP) {  in update_group_capacity()
10061 		group = child->groups;  in update_group_capacity()
10063 			struct sched_group_capacity *sgc = group->sgc;  in update_group_capacity()
10065 			capacity += sgc->capacity;  in update_group_capacity()
10066 			min_capacity = min(sgc->min_capacity, min_capacity);  in update_group_capacity()
10067 			max_capacity = max(sgc->max_capacity, max_capacity);  in update_group_capacity()
10068 			group = group->next;  in update_group_capacity()
10069 		} while (group != child->groups);  in update_group_capacity()
10072 	sdg->sgc->capacity = capacity;  in update_group_capacity()
10073 	sdg->sgc->min_capacity = min_capacity;  in update_group_capacity()
10074 	sdg->sgc->max_capacity = max_capacity;  in update_group_capacity()
10085 	return ((rq->cpu_capacity * sd->imbalance_pct) <  in check_cpu_capacity()
10092 	return rq->misfit_task_load;  in check_misfit_status()
10097  * groups is inadequate due to ->cpus_ptr constraints.
10106  * If we were to balance group-wise we'd place two tasks in the first group and
10126 	return group->sgc->imbalance;  in sg_imbalanced()
10144 	if (sgs->sum_nr_running < sgs->group_weight)  in group_has_capacity()
10147 	if ((sgs->group_capacity * imbalance_pct) <  in group_has_capacity()
10148 			(sgs->group_runnable * 100))  in group_has_capacity()
10151 	if ((sgs->group_capacity * 100) >  in group_has_capacity()
10152 			(sgs->group_util * imbalance_pct))  in group_has_capacity()
10169 	if (sgs->sum_nr_running <= sgs->group_weight)  in group_is_overloaded()
10172 	if ((sgs->group_capacity * 100) <  in group_is_overloaded()
10173 			(sgs->group_util * imbalance_pct))  in group_is_overloaded()
10176 	if ((sgs->group_capacity * imbalance_pct) <  in group_is_overloaded()
10177 			(sgs->group_runnable * 100))  in group_is_overloaded()
10194 	if (sgs->group_asym_packing)  in group_classify()
10197 	if (sgs->group_smt_balance)  in group_classify()
10200 	if (sgs->group_misfit_task_load)  in group_classify()
10210  * sched_use_asym_prio - Check whether asym_packing priority must be used
10222 	if (!(sd->flags & SD_ASYM_PACKING))  in sched_use_asym_prio()
10228 	return sd->flags & SD_SHARE_CPUCAPACITY || is_core_idle(cpu);  in sched_use_asym_prio()
10242  * sched_group_asym - Check if the destination CPU can do asym_packing balance
10244  * @sgs:	Load-balancing statistics of the candidate busiest group
10260 	if ((group->flags & SD_SHARE_CPUCAPACITY) &&  in sched_group_asym()
10261 	    (sgs->group_weight - sgs->idle_cpus != 1))  in sched_group_asym()
10264 	return sched_asym(env->sd, env->dst_cpu, group->asym_prefer_cpu);  in sched_group_asym()
10274 	return (sg1->flags & SD_SHARE_CPUCAPACITY) !=  in smt_vs_nonsmt_groups()
10275 		(sg2->flags & SD_SHARE_CPUCAPACITY);  in smt_vs_nonsmt_groups()
10281 	if (!env->idle)  in smt_balance()
10290 	if (group->flags & SD_SHARE_CPUCAPACITY &&  in smt_balance()
10291 	    sgs->sum_h_nr_running > 1)  in smt_balance()
10305 	if (!env->idle || !busiest->sum_nr_running)  in sibling_imbalance()
10308 	ncores_busiest = sds->busiest->cores;  in sibling_imbalance()
10309 	ncores_local = sds->local->cores;  in sibling_imbalance()
10312 		imbalance = busiest->sum_nr_running;  in sibling_imbalance()
10313 		lsub_positive(&imbalance, local->sum_nr_running);  in sibling_imbalance()
10318 	imbalance = ncores_local * busiest->sum_nr_running;  in sibling_imbalance()
10319 	lsub_positive(&imbalance, ncores_busiest * local->sum_nr_running);  in sibling_imbalance()
10325 	if (imbalance <= 1 && local->sum_nr_running == 0 &&  in sibling_imbalance()
10326 	    busiest->sum_nr_running > 1)  in sibling_imbalance()
10339 	if (rq->cfs.h_nr_runnable != 1)  in sched_reduced_capacity()
10346  * update_sg_lb_stats - Update sched_group's statistics for load balancing.
10348  * @sds: Load-balancing data with statistics of the local group.
10361 	int i, nr_running, local_group, sd_flags = env->sd->flags;  in update_sg_lb_stats()
10362 	bool balancing_at_rd = !env->sd->parent;  in update_sg_lb_stats()
10366 	local_group = group == sds->local;  in update_sg_lb_stats()
10368 	for_each_cpu_and(i, sched_group_span(group), env->cpus) {  in update_sg_lb_stats()
10372 		sgs->group_load += load;  in update_sg_lb_stats()
10373 		sgs->group_util += cpu_util_cfs(i);  in update_sg_lb_stats()
10374 		sgs->group_runnable += cpu_runnable(rq);  in update_sg_lb_stats()
10375 		sgs->sum_h_nr_running += rq->cfs.h_nr_runnable;  in update_sg_lb_stats()
10377 		nr_running = rq->nr_running;  in update_sg_lb_stats()
10378 		sgs->sum_nr_running += nr_running;  in update_sg_lb_stats()
10387 			sgs->idle_cpus++;  in update_sg_lb_stats()
10399 			sgs->nr_numa_running += rq->nr_numa_running;  in update_sg_lb_stats()
10400 			sgs->nr_preferred_running += rq->nr_preferred_running;  in update_sg_lb_stats()
10408 			if (sgs->group_misfit_task_load < rq->misfit_task_load) {  in update_sg_lb_stats()
10409 				sgs->group_misfit_task_load = rq->misfit_task_load;  in update_sg_lb_stats()
10412 		} else if (env->idle && sched_reduced_capacity(rq, env->sd)) {  in update_sg_lb_stats()
10414 			if (sgs->group_misfit_task_load < load)  in update_sg_lb_stats()
10415 				sgs->group_misfit_task_load = load;  in update_sg_lb_stats()
10419 	sgs->group_capacity = group->sgc->capacity;  in update_sg_lb_stats()
10421 	sgs->group_weight = group->group_weight;  in update_sg_lb_stats()
10424 	if (!local_group && env->idle && sgs->sum_h_nr_running &&  in update_sg_lb_stats()
10426 		sgs->group_asym_packing = 1;  in update_sg_lb_stats()
10430 		sgs->group_smt_balance = 1;  in update_sg_lb_stats()
10432 	sgs->group_type = group_classify(env->sd->imbalance_pct, group, sgs);  in update_sg_lb_stats()
10435 	if (sgs->group_type == group_overloaded)  in update_sg_lb_stats()
10436 		sgs->avg_load = (sgs->group_load * SCHED_CAPACITY_SCALE) /  in update_sg_lb_stats()
10437 				sgs->group_capacity;  in update_sg_lb_stats()
10441  * update_sd_pick_busiest - return 1 on busiest group
10458 	struct sg_lb_stats *busiest = &sds->busiest_stat;  in update_sd_pick_busiest()
10461 	if (!sgs->sum_h_nr_running)  in update_sd_pick_busiest()
10470 	if ((env->sd->flags & SD_ASYM_CPUCAPACITY) &&  in update_sd_pick_busiest()
10471 	    (sgs->group_type == group_misfit_task) &&  in update_sd_pick_busiest()
10472 	    (!capacity_greater(capacity_of(env->dst_cpu), sg->sgc->max_capacity) ||  in update_sd_pick_busiest()
10473 	     sds->local_stat.group_type != group_has_spare))  in update_sd_pick_busiest()
10476 	if (sgs->group_type > busiest->group_type)  in update_sd_pick_busiest()
10479 	if (sgs->group_type < busiest->group_type)  in update_sd_pick_busiest()
10487 	switch (sgs->group_type) {  in update_sd_pick_busiest()
10490 		return sgs->avg_load > busiest->avg_load;  in update_sd_pick_busiest()
10501 		return sched_asym_prefer(sds->busiest->asym_prefer_cpu, sg->asym_prefer_cpu);  in update_sd_pick_busiest()
10508 		return sgs->group_misfit_task_load > busiest->group_misfit_task_load;  in update_sd_pick_busiest()
10515 		if (sgs->idle_cpus != 0 || busiest->idle_cpus != 0)  in update_sd_pick_busiest()
10533 		if (sgs->avg_load < busiest->avg_load)  in update_sd_pick_busiest()
10536 		if (sgs->avg_load == busiest->avg_load) {  in update_sd_pick_busiest()
10538 			 * SMT sched groups need more help than non-SMT groups.  in update_sd_pick_busiest()
10541 			if (sds->busiest->flags & SD_SHARE_CPUCAPACITY)  in update_sd_pick_busiest()
10553 		if (smt_vs_nonsmt_groups(sds->busiest, sg)) {  in update_sd_pick_busiest()
10554 			if (sg->flags & SD_SHARE_CPUCAPACITY && sgs->sum_h_nr_running <= 1)  in update_sd_pick_busiest()
10568 		if (sgs->idle_cpus > busiest->idle_cpus)  in update_sd_pick_busiest()
10570 		else if ((sgs->idle_cpus == busiest->idle_cpus) &&  in update_sd_pick_busiest()
10571 			 (sgs->sum_nr_running <= busiest->sum_nr_running))  in update_sd_pick_busiest()
10579 	 * per-CPU capacity. Migrating tasks to less capable CPUs may harm  in update_sd_pick_busiest()
10583 	if ((env->sd->flags & SD_ASYM_CPUCAPACITY) &&  in update_sd_pick_busiest()
10584 	    (sgs->group_type <= group_fully_busy) &&  in update_sd_pick_busiest()
10585 	    (capacity_greater(sg->sgc->min_capacity, capacity_of(env->dst_cpu))))  in update_sd_pick_busiest()
10594 	if (sgs->sum_h_nr_running > sgs->nr_numa_running)  in fbq_classify_group()
10596 	if (sgs->sum_h_nr_running > sgs->nr_preferred_running)  in fbq_classify_group()
10603 	if (rq->nr_running > rq->nr_numa_running)  in fbq_classify_rq()
10605 	if (rq->nr_running > rq->nr_preferred_running)  in fbq_classify_rq()
10625  * task_running_on_cpu - return 1 if @p is running on @cpu.
10631 	if (cpu != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time))  in task_running_on_cpu()
10641  * idle_cpu_without - would a given CPU be idle without p ?
10651 	if (rq->curr != rq->idle && rq->curr != p)  in idle_cpu_without()
10655 	 * rq->nr_running can't be used but an updated version without the  in idle_cpu_without()
10660 	if (rq->ttwu_pending)  in idle_cpu_without()
10667  * update_sg_wakeup_stats - Update sched_group's statistics for wakeup.
10683 	if (sd->flags & SD_ASYM_CPUCAPACITY)  in update_sg_wakeup_stats()
10684 		sgs->group_misfit_task_load = 1;  in update_sg_wakeup_stats()
10690 		sgs->group_load += cpu_load_without(rq, p);  in update_sg_wakeup_stats()
10691 		sgs->group_util += cpu_util_without(i, p);  in update_sg_wakeup_stats()
10692 		sgs->group_runnable += cpu_runnable_without(rq, p);  in update_sg_wakeup_stats()
10694 		sgs->sum_h_nr_running += rq->cfs.h_nr_runnable - local;  in update_sg_wakeup_stats()
10696 		nr_running = rq->nr_running - local;  in update_sg_wakeup_stats()
10697 		sgs->sum_nr_running += nr_running;  in update_sg_wakeup_stats()
10703 			sgs->idle_cpus++;  in update_sg_wakeup_stats()
10706 		if (sd->flags & SD_ASYM_CPUCAPACITY &&  in update_sg_wakeup_stats()
10707 		    sgs->group_misfit_task_load &&  in update_sg_wakeup_stats()
10709 			sgs->group_misfit_task_load = 0;  in update_sg_wakeup_stats()
10713 	sgs->group_capacity = group->sgc->capacity;  in update_sg_wakeup_stats()
10715 	sgs->group_weight = group->group_weight;  in update_sg_wakeup_stats()
10717 	sgs->group_type = group_classify(sd->imbalance_pct, group, sgs);  in update_sg_wakeup_stats()
10723 	if (sgs->group_type == group_fully_busy ||  in update_sg_wakeup_stats()
10724 		sgs->group_type == group_overloaded)  in update_sg_wakeup_stats()
10725 		sgs->avg_load = (sgs->group_load * SCHED_CAPACITY_SCALE) /  in update_sg_wakeup_stats()
10726 				sgs->group_capacity;  in update_sg_wakeup_stats()
10734 	if (sgs->group_type < idlest_sgs->group_type)  in update_pick_idlest()
10737 	if (sgs->group_type > idlest_sgs->group_type)  in update_pick_idlest()
10745 	switch (sgs->group_type) {  in update_pick_idlest()
10749 		if (idlest_sgs->avg_load <= sgs->avg_load)  in update_pick_idlest()
10761 		if (idlest->sgc->max_capacity >= group->sgc->max_capacity)  in update_pick_idlest()
10767 		if (idlest_sgs->idle_cpus > sgs->idle_cpus)  in update_pick_idlest()
10771 		if (idlest_sgs->idle_cpus == sgs->idle_cpus &&  in update_pick_idlest()
10772 			idlest_sgs->group_util <= sgs->group_util)  in update_pick_idlest()
10790 	struct sched_group *idlest = NULL, *local = NULL, *group = sd->groups;  in sched_balance_find_dst_group()
10804 					p->cpus_ptr))  in sched_balance_find_dst_group()
10828 	} while (group = group->next, group != sd->groups);  in sched_balance_find_dst_group()
10859 				(sd->imbalance_pct-100) / 100;  in sched_balance_find_dst_group()
10866 		 * cross-domain, add imbalance to the load on the remote node  in sched_balance_find_dst_group()
10870 		if ((sd->flags & SD_NUMA) &&  in sched_balance_find_dst_group()
10881 		if (100 * local_sgs.avg_load <= sd->imbalance_pct * idlest_sgs.avg_load)  in sched_balance_find_dst_group()
10893 		if (local->sgc->max_capacity >= idlest->sgc->max_capacity)  in sched_balance_find_dst_group()
10899 		if (sd->flags & SD_NUMA) {  in sched_balance_find_dst_group()
10900 			int imb_numa_nr = sd->imb_numa_nr;  in sched_balance_find_dst_group()
10907 			if (cpu_to_node(this_cpu) == p->numa_preferred_nid)  in sched_balance_find_dst_group()
10911 			if (cpu_to_node(idlest_cpu) == p->numa_preferred_nid)  in sched_balance_find_dst_group()
10923 			if (p->nr_cpus_allowed != NR_CPUS) {  in sched_balance_find_dst_group()
10926 				cpumask_and(cpus, sched_group_span(local), p->cpus_ptr);  in sched_balance_find_dst_group()
10927 				imb_numa_nr = min(cpumask_weight(cpus), sd->imb_numa_nr);  in sched_balance_find_dst_group()
10930 			imbalance = abs(local_sgs.idle_cpus - idlest_sgs.idle_cpus);  in sched_balance_find_dst_group()
10967 	if (!sched_feat(SIS_UTIL) || env->idle == CPU_NEWLY_IDLE)  in update_idle_cpu_scan()
10970 	llc_weight = per_cpu(sd_llc_size, env->dst_cpu);  in update_idle_cpu_scan()
10971 	if (env->sd->span_weight != llc_weight)  in update_idle_cpu_scan()
10974 	sd_share = rcu_dereference(per_cpu(sd_llc_shared, env->dst_cpu));  in update_idle_cpu_scan()
10984 	 * let y = SCHED_CAPACITY_SCALE - p * x^2                       [1]  in update_idle_cpu_scan()
11002 	 * y = SCHED_CAPACITY_SCALE -  in update_idle_cpu_scan()
11011 	pct = env->sd->imbalance_pct;  in update_idle_cpu_scan()
11015 	y = SCHED_CAPACITY_SCALE - tmp;  in update_idle_cpu_scan()
11020 	if ((int)y != sd_share->nr_idle_scan)  in update_idle_cpu_scan()
11021 		WRITE_ONCE(sd_share->nr_idle_scan, (int)y);  in update_idle_cpu_scan()
11025  * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
11032 	struct sched_group *sg = env->sd->groups;  in update_sd_lb_stats()
11033 	struct sg_lb_stats *local = &sds->local_stat;  in update_sd_lb_stats()
11042 		local_group = cpumask_test_cpu(env->dst_cpu, sched_group_span(sg));  in update_sd_lb_stats()
11044 			sds->local = sg;  in update_sd_lb_stats()
11047 			if (env->idle != CPU_NEWLY_IDLE ||  in update_sd_lb_stats()
11048 			    time_after_eq(jiffies, sg->sgc->next_update))  in update_sd_lb_stats()
11049 				update_group_capacity(env->sd, env->dst_cpu);  in update_sd_lb_stats()
11055 			sds->busiest = sg;  in update_sd_lb_stats()
11056 			sds->busiest_stat = *sgs;  in update_sd_lb_stats()
11060 		sds->total_load += sgs->group_load;  in update_sd_lb_stats()
11061 		sds->total_capacity += sgs->group_capacity;  in update_sd_lb_stats()
11063 		sum_util += sgs->group_util;  in update_sd_lb_stats()
11064 		sg = sg->next;  in update_sd_lb_stats()
11065 	} while (sg != env->sd->groups);  in update_sd_lb_stats()
11072 	if (sds->busiest)  in update_sd_lb_stats()
11073 		sds->prefer_sibling = !!(sds->busiest->flags & SD_PREFER_SIBLING);  in update_sd_lb_stats()
11076 	if (env->sd->flags & SD_NUMA)  in update_sd_lb_stats()
11077 		env->fbq_type = fbq_classify_group(&sds->busiest_stat);  in update_sd_lb_stats()
11079 	if (!env->sd->parent) {  in update_sd_lb_stats()
11081 		set_rd_overloaded(env->dst_rq->rd, sg_overloaded);  in update_sd_lb_stats()
11083 		/* Update over-utilization (tipping point, U >= 0) indicator */  in update_sd_lb_stats()
11084 		set_rd_overutilized(env->dst_rq->rd, sg_overutilized);  in update_sd_lb_stats()
11086 		set_rd_overutilized(env->dst_rq->rd, sg_overutilized);  in update_sd_lb_stats()
11093  * calculate_imbalance - Calculate the amount of imbalance present within the
11102 	local = &sds->local_stat;  in calculate_imbalance()
11103 	busiest = &sds->busiest_stat;  in calculate_imbalance()
11105 	if (busiest->group_type == group_misfit_task) {  in calculate_imbalance()
11106 		if (env->sd->flags & SD_ASYM_CPUCAPACITY) {  in calculate_imbalance()
11108 			env->migration_type = migrate_misfit;  in calculate_imbalance()
11109 			env->imbalance = 1;  in calculate_imbalance()
11115 			env->migration_type = migrate_load;  in calculate_imbalance()
11116 			env->imbalance = busiest->group_misfit_task_load;  in calculate_imbalance()
11121 	if (busiest->group_type == group_asym_packing) {  in calculate_imbalance()
11126 		env->migration_type = migrate_task;  in calculate_imbalance()
11127 		env->imbalance = busiest->sum_h_nr_running;  in calculate_imbalance()
11131 	if (busiest->group_type == group_smt_balance) {  in calculate_imbalance()
11133 		env->migration_type = migrate_task;  in calculate_imbalance()
11134 		env->imbalance = 1;  in calculate_imbalance()
11138 	if (busiest->group_type == group_imbalanced) {  in calculate_imbalance()
11140 		 * In the group_imb case we cannot rely on group-wide averages  in calculate_imbalance()
11141 		 * to ensure CPU-load equilibrium, try to move any task to fix  in calculate_imbalance()
11145 		env->migration_type = migrate_task;  in calculate_imbalance()
11146 		env->imbalance = 1;  in calculate_imbalance()
11154 	if (local->group_type == group_has_spare) {  in calculate_imbalance()
11155 		if ((busiest->group_type > group_fully_busy) &&  in calculate_imbalance()
11156 		    !(env->sd->flags & SD_SHARE_LLC)) {  in calculate_imbalance()
11165 			env->migration_type = migrate_util;  in calculate_imbalance()
11166 			env->imbalance = max(local->group_capacity, local->group_util) -  in calculate_imbalance()
11167 					 local->group_util;  in calculate_imbalance()
11176 			if (env->idle && env->imbalance == 0) {  in calculate_imbalance()
11177 				env->migration_type = migrate_task;  in calculate_imbalance()
11178 				env->imbalance = 1;  in calculate_imbalance()
11184 		if (busiest->group_weight == 1 || sds->prefer_sibling) {  in calculate_imbalance()
11189 			env->migration_type = migrate_task;  in calculate_imbalance()
11190 			env->imbalance = sibling_imbalance(env, sds, busiest, local);  in calculate_imbalance()
11197 			env->migration_type = migrate_task;  in calculate_imbalance()
11198 			env->imbalance = max_t(long, 0,  in calculate_imbalance()
11199 					       (local->idle_cpus - busiest->idle_cpus));  in calculate_imbalance()
11204 		if (env->sd->flags & SD_NUMA) {  in calculate_imbalance()
11205 			env->imbalance = adjust_numa_imbalance(env->imbalance,  in calculate_imbalance()
11206 							       local->sum_nr_running + 1,  in calculate_imbalance()
11207 							       env->sd->imb_numa_nr);  in calculate_imbalance()
11212 		env->imbalance >>= 1;  in calculate_imbalance()
11221 	if (local->group_type < group_overloaded) {  in calculate_imbalance()
11227 		local->avg_load = (local->group_load * SCHED_CAPACITY_SCALE) /  in calculate_imbalance()
11228 				  local->group_capacity;  in calculate_imbalance()
11234 		if (local->avg_load >= busiest->avg_load) {  in calculate_imbalance()
11235 			env->imbalance = 0;  in calculate_imbalance()
11239 		sds->avg_load = (sds->total_load * SCHED_CAPACITY_SCALE) /  in calculate_imbalance()
11240 				sds->total_capacity;  in calculate_imbalance()
11246 		if (local->avg_load >= sds->avg_load) {  in calculate_imbalance()
11247 			env->imbalance = 0;  in calculate_imbalance()
11261 	env->migration_type = migrate_load;  in calculate_imbalance()
11262 	env->imbalance = min(  in calculate_imbalance()
11263 		(busiest->avg_load - sds->avg_load) * busiest->group_capacity,  in calculate_imbalance()
11264 		(sds->avg_load - local->avg_load) * local->group_capacity  in calculate_imbalance()
11291  * sched_balance_find_src_group - Returns the busiest group within the sched_domain
11298  * Return:	- The busiest group if imbalance exists.
11320 	if (busiest->group_type == group_misfit_task)  in sched_balance_find_src_group()
11323 	if (!is_rd_overutilized(env->dst_rq->rd) &&  in sched_balance_find_src_group()
11324 	    rcu_dereference(env->dst_rq->rd->pd))  in sched_balance_find_src_group()
11328 	if (busiest->group_type == group_asym_packing)  in sched_balance_find_src_group()
11336 	if (busiest->group_type == group_imbalanced)  in sched_balance_find_src_group()
11344 	if (local->group_type > busiest->group_type)  in sched_balance_find_src_group()
11351 	if (local->group_type == group_overloaded) {  in sched_balance_find_src_group()
11356 		if (local->avg_load >= busiest->avg_load)  in sched_balance_find_src_group()
11367 		if (local->avg_load >= sds.avg_load)  in sched_balance_find_src_group()
11374 		if (100 * busiest->avg_load <=  in sched_balance_find_src_group()
11375 				env->sd->imbalance_pct * local->avg_load)  in sched_balance_find_src_group()
11383 	if (sds.prefer_sibling && local->group_type == group_has_spare &&  in sched_balance_find_src_group()
11387 	if (busiest->group_type != group_overloaded) {  in sched_balance_find_src_group()
11388 		if (!env->idle) {  in sched_balance_find_src_group()
11397 		if (busiest->group_type == group_smt_balance &&  in sched_balance_find_src_group()
11403 		if (busiest->group_weight > 1 &&  in sched_balance_find_src_group()
11404 		    local->idle_cpus <= (busiest->idle_cpus + 1)) {  in sched_balance_find_src_group()
11417 		if (busiest->sum_h_nr_running == 1) {  in sched_balance_find_src_group()
11428 	return env->imbalance ? sds.busiest : NULL;  in sched_balance_find_src_group()
11431 	env->imbalance = 0;  in sched_balance_find_src_group()
11436  * sched_balance_find_src_rq - find the busiest runqueue among the CPUs in the group.
11446 	for_each_cpu_and(i, sched_group_span(group), env->cpus) {  in sched_balance_find_src_rq()
11456 		 *  - regular: there are !numa tasks  in sched_balance_find_src_rq()
11457 		 *  - remote:  there are numa tasks that run on the 'wrong' node  in sched_balance_find_src_rq()
11458 		 *  - all:     there is no distinction  in sched_balance_find_src_rq()
11473 		if (rt > env->fbq_type)  in sched_balance_find_src_rq()
11476 		nr_running = rq->cfs.h_nr_runnable;  in sched_balance_find_src_rq()
11484 		 * eventually lead to active_balancing high->low capacity.  in sched_balance_find_src_rq()
11485 		 * Higher per-CPU capacity is considered better than balancing  in sched_balance_find_src_rq()
11488 		if (env->sd->flags & SD_ASYM_CPUCAPACITY &&  in sched_balance_find_src_rq()
11489 		    !capacity_greater(capacity_of(env->dst_cpu), capacity) &&  in sched_balance_find_src_rq()
11500 		if (sched_asym(env->sd, i, env->dst_cpu) && nr_running == 1)  in sched_balance_find_src_rq()
11503 		switch (env->migration_type) {  in sched_balance_find_src_rq()
11511 			if (nr_running == 1 && load > env->imbalance &&  in sched_balance_find_src_rq()
11512 			    !check_cpu_capacity(rq, env->sd))  in sched_balance_find_src_rq()
11564 			if (rq->misfit_task_load > busiest_load) {  in sched_balance_find_src_rq()
11565 				busiest_load = rq->misfit_task_load;  in sched_balance_find_src_rq()
11596 	return env->idle && sched_use_asym_prio(env->sd, env->dst_cpu) &&  in asym_active_balance()
11597 	       (sched_asym_prefer(env->dst_cpu, env->src_cpu) ||  in asym_active_balance()
11598 		!sched_use_asym_prio(env->sd, env->src_cpu));  in asym_active_balance()
11604 	struct sched_domain *sd = env->sd;  in imbalanced_active_balance()
11611 	if ((env->migration_type == migrate_task) &&  in imbalanced_active_balance()
11612 	    (sd->nr_balance_failed > sd->cache_nice_tries+2))  in imbalanced_active_balance()
11620 	struct sched_domain *sd = env->sd;  in need_active_balance()
11634 	if (env->idle &&  in need_active_balance()
11635 	    (env->src_rq->cfs.h_nr_runnable == 1)) {  in need_active_balance()
11636 		if ((check_cpu_capacity(env->src_rq, sd)) &&  in need_active_balance()
11637 		    (capacity_of(env->src_cpu)*sd->imbalance_pct < capacity_of(env->dst_cpu)*100))  in need_active_balance()
11641 	if (env->migration_type == migrate_misfit)  in need_active_balance()
11652 	struct sched_group *sg = env->sd->groups;  in should_we_balance()
11653 	int cpu, idle_smt = -1;  in should_we_balance()
11659 	if (!cpumask_test_cpu(env->dst_cpu, env->cpus))  in should_we_balance()
11669 	if (env->idle == CPU_NEWLY_IDLE) {  in should_we_balance()
11670 		if (env->dst_rq->nr_running > 0 || env->dst_rq->ttwu_pending)  in should_we_balance()
11677 	for_each_cpu_and(cpu, swb_cpus, env->cpus) {  in should_we_balance()
11686 		if (!(env->sd->flags & SD_SHARE_CPUCAPACITY) && !is_core_idle(cpu)) {  in should_we_balance()
11687 			if (idle_smt == -1)  in should_we_balance()
11701 		 * Are we the first idle core in a non-SMT domain or higher,  in should_we_balance()
11704 		return cpu == env->dst_cpu;  in should_we_balance()
11708 	if (idle_smt != -1)  in should_we_balance()
11709 		return idle_smt == env->dst_cpu;  in should_we_balance()
11712 	return group_balance_cpu(sg) == env->dst_cpu;  in should_we_balance()
11721 	switch (env->migration_type) {  in update_lb_imbalance_stat()
11723 		__schedstat_add(sd->lb_imbalance_load[idle], env->imbalance);  in update_lb_imbalance_stat()
11726 		__schedstat_add(sd->lb_imbalance_util[idle], env->imbalance);  in update_lb_imbalance_stat()
11729 		__schedstat_add(sd->lb_imbalance_task[idle], env->imbalance);  in update_lb_imbalance_stat()
11732 		__schedstat_add(sd->lb_imbalance_misfit[idle], env->imbalance);  in update_lb_imbalance_stat()
11746 	struct sched_domain *sd_parent = sd->parent;  in sched_balance_rq()
11755 		.dst_grpmask    = group_balance_mask(sd->groups),  in sched_balance_rq()
11765 	schedstat_inc(sd->lb_count[idle]);  in sched_balance_rq()
11775 		schedstat_inc(sd->lb_nobusyg[idle]);  in sched_balance_rq()
11781 		schedstat_inc(sd->lb_nobusyq[idle]);  in sched_balance_rq()
11789 	env.src_cpu = busiest->cpu;  in sched_balance_rq()
11795 	if (busiest->nr_running > 1) {  in sched_balance_rq()
11798 		 * an imbalance but busiest->nr_running <= 1, the group is  in sched_balance_rq()
11802 		env.loop_max  = min(sysctl_sched_nr_migrate, busiest->nr_running);  in sched_balance_rq()
11809 		 * cur_ld_moved - load moved in current iteration  in sched_balance_rq()
11810 		 * ld_moved     - cumulative load moved across iterations  in sched_balance_rq()
11817 		 * unlock busiest->lock, and we are able to be sure  in sched_balance_rq()
11846 		 * nohz-idle), we now have balance_cpu in a position to move  in sched_balance_rq()
11857 			/* Prevent to re-select dst_cpu via env's CPUs */  in sched_balance_rq()
11877 			int *group_imbalance = &sd_parent->groups->sgc->imbalance;  in sched_balance_rq()
11904 		schedstat_inc(sd->lb_failed[idle]);  in sched_balance_rq()
11916 			sd->nr_balance_failed++;  in sched_balance_rq()
11928 			if (!cpumask_test_cpu(this_cpu, busiest->curr->cpus_ptr)) {  in sched_balance_rq()
11937 			 * ->active_balance synchronizes accesses to  in sched_balance_rq()
11938 			 * ->active_balance_work.  Once set, it's cleared  in sched_balance_rq()
11941 			if (!busiest->active_balance) {  in sched_balance_rq()
11942 				busiest->active_balance = 1;  in sched_balance_rq()
11943 				busiest->push_cpu = this_cpu;  in sched_balance_rq()
11952 					&busiest->active_balance_work);  in sched_balance_rq()
11957 		sd->nr_balance_failed = 0;  in sched_balance_rq()
11962 		sd->balance_interval = sd->min_interval;  in sched_balance_rq()
11974 		int *group_imbalance = &sd_parent->groups->sgc->imbalance;  in sched_balance_rq()
11986 	schedstat_inc(sd->lb_balanced[idle]);  in sched_balance_rq()
11988 	sd->nr_balance_failed = 0;  in sched_balance_rq()
12009 	     sd->balance_interval < MAX_PINNED_INTERVAL) ||  in sched_balance_rq()
12010 	    sd->balance_interval < sd->max_interval)  in sched_balance_rq()
12011 		sd->balance_interval *= 2;  in sched_balance_rq()
12019 	unsigned long interval = sd->balance_interval;  in get_sd_balance_interval()
12022 		interval *= sd->busy_factor;  in get_sd_balance_interval()
12033 		interval -= 1;  in get_sd_balance_interval()
12047 	next = sd->last_balance + interval;  in update_next_balance()
12063 	int target_cpu = busiest_rq->push_cpu;  in active_load_balance_cpu_stop()
12071 	 * Between queueing the stop-work and running it is a hole in which  in active_load_balance_cpu_stop()
12080 		     !busiest_rq->active_balance))  in active_load_balance_cpu_stop()
12084 	if (busiest_rq->nr_running <= 1)  in active_load_balance_cpu_stop()
12090 	 * Bjorn Helgaas on a 128-CPU setup.  in active_load_balance_cpu_stop()
12106 			.src_cpu	= busiest_rq->cpu,  in active_load_balance_cpu_stop()
12112 		schedstat_inc(sd->alb_count);  in active_load_balance_cpu_stop()
12117 			schedstat_inc(sd->alb_pushed);  in active_load_balance_cpu_stop()
12119 			sd->nr_balance_failed = 0;  in active_load_balance_cpu_stop()
12121 			schedstat_inc(sd->alb_failed);  in active_load_balance_cpu_stop()
12126 	busiest_rq->active_balance = 0;  in active_load_balance_cpu_stop()
12138  * This flag serializes load-balancing passes over large domains
12139  * (above the NODE topology level) - only one load-balancing instance
12143  * - Note that load-balancing passes triggered while another one
12144  *   is executing are skipped and not re-tried.
12146  * - Also note that this does not serialize rebalance_domains()
12147  *   execution, as non-SD_SERIALIZE domains will still be
12148  *   load-balanced in parallel.
12154  * This trades load-balance latency on larger machines for less cross talk.
12163 	if (cost > sd->max_newidle_lb_cost) {  in update_newidle_cost()
12168 		sd->max_newidle_lb_cost = cost;  in update_newidle_cost()
12169 		sd->last_decay_max_lb_cost = jiffies;  in update_newidle_cost()
12170 	} else if (time_after(jiffies, sd->last_decay_max_lb_cost + HZ)) {  in update_newidle_cost()
12176 		sd->max_newidle_lb_cost = (sd->max_newidle_lb_cost * 253) / 256;  in update_newidle_cost()
12177 		sd->last_decay_max_lb_cost = jiffies;  in update_newidle_cost()
12194 	int cpu = rq->cpu;  in sched_balance_domains()
12211 		max_cost += sd->max_newidle_lb_cost;  in sched_balance_domains()
12226 		need_serialize = sd->flags & SD_SERIALIZE;  in sched_balance_domains()
12232 		if (time_after_eq(jiffies, sd->last_balance + interval)) {  in sched_balance_domains()
12236 				 * env->dst_cpu, so we can't know our idle  in sched_balance_domains()
12242 			sd->last_balance = jiffies;  in sched_balance_domains()
12248 		if (time_after(next_balance, sd->last_balance + interval)) {  in sched_balance_domains()
12249 			next_balance = sd->last_balance + interval;  in sched_balance_domains()
12255 		 * Ensure the rq-wide value also decays but keep it at a  in sched_balance_domains()
12256 		 * reasonable floor to avoid funnies with rq->avg_idle.  in sched_balance_domains()
12258 		rq->max_idle_balance_cost =  in sched_balance_domains()
12269 		rq->next_balance = next_balance;  in sched_balance_domains()
12275 	return unlikely(!rcu_dereference_sched(rq->sd));  in on_null_domain()
12282  * - When one of the busy CPUs notices that there may be an idle rebalancing
12302 	return -1;  in find_new_ilb()
12306  * Kick a CPU to do the NOHZ balancing, if it is time for it, via a cross-CPU
12347 	smp_call_function_single_async(ilb_cpu, &cpu_rq(ilb_cpu)->nohz_csd);  in kick_ilb()
12359 	int nr_busy, i, cpu = rq->cpu;  in nohz_balancer_kick()
12362 	if (unlikely(rq->idle_balance))  in nohz_balancer_kick()
12385 	if (rq->nr_running >= 2) {  in nohz_balancer_kick()
12392 	sd = rcu_dereference(rq->sd);  in nohz_balancer_kick()
12398 		if (rq->cfs.h_nr_runnable >= 1 && check_cpu_capacity(rq, sd)) {  in nohz_balancer_kick()
12447 		 * increase the overall cache utilization), we need a less-loaded LLC  in nohz_balancer_kick()
12451 		 * the others are - so just get a NOHZ balance going if it looks  in nohz_balancer_kick()
12454 		nr_busy = atomic_read(&sds->nr_busy_cpus);  in nohz_balancer_kick()
12477 	if (!sd || !sd->nohz_idle)  in set_cpu_sd_state_busy()
12479 	sd->nohz_idle = 0;  in set_cpu_sd_state_busy()
12481 	atomic_inc(&sd->shared->nr_busy_cpus);  in set_cpu_sd_state_busy()
12490 	if (likely(!rq->nohz_tick_stopped))  in nohz_balance_exit_idle()
12493 	rq->nohz_tick_stopped = 0;  in nohz_balance_exit_idle()
12494 	cpumask_clear_cpu(rq->cpu, nohz.idle_cpus_mask);  in nohz_balance_exit_idle()
12497 	set_cpu_sd_state_busy(rq->cpu);  in nohz_balance_exit_idle()
12507 	if (!sd || sd->nohz_idle)  in set_cpu_sd_state_idle()
12509 	sd->nohz_idle = 1;  in set_cpu_sd_state_idle()
12511 	atomic_dec(&sd->shared->nr_busy_cpus);  in set_cpu_sd_state_idle()
12531 	 * Can be set safely without rq->lock held  in nohz_balance_enter_idle()
12533 	 * rq->lock is held during the check and the clear  in nohz_balance_enter_idle()
12535 	rq->has_blocked_load = 1;  in nohz_balance_enter_idle()
12543 	if (rq->nohz_tick_stopped)  in nohz_balance_enter_idle()
12550 	rq->nohz_tick_stopped = 1;  in nohz_balance_enter_idle()
12575 	unsigned int cpu = rq->cpu;  in update_nohz_stats()
12577 	if (!rq->has_blocked_load)  in update_nohz_stats()
12583 	if (!time_after(jiffies, READ_ONCE(rq->last_blocked_load_update_tick)))  in update_nohz_stats()
12588 	return rq->has_blocked_load;  in update_nohz_stats()
12603 	int this_cpu = this_rq->cpu;  in _nohz_idle_balance()
12660 		if (time_after_eq(jiffies, rq->next_balance)) {  in _nohz_idle_balance()
12671 		if (time_after(next_balance, rq->next_balance)) {  in _nohz_idle_balance()
12672 			next_balance = rq->next_balance;  in _nohz_idle_balance()
12701 	unsigned int flags = this_rq->nohz_idle_balance;  in nohz_idle_balance()
12706 	this_rq->nohz_idle_balance = 0;  in nohz_idle_balance()
12747 	int this_cpu = this_rq->cpu;  in nohz_newidle_balance()
12750 	if (this_rq->avg_idle < sysctl_sched_migration_cost)  in nohz_newidle_balance()
12781  *   < 0 - we released the lock and there are !fair tasks present
12782  *     0 - failed, no new tasks
12783  *   > 0 - success, new (fair) tasks present
12788 	int this_cpu = this_rq->cpu;  in sched_balance_newidle()
12800 	if (this_rq->ttwu_pending)  in sched_balance_newidle()
12808 	this_rq->idle_stamp = rq_clock(this_rq);  in sched_balance_newidle()
12818 	 * for load-balance and preemption/IRQs are still disabled avoiding  in sched_balance_newidle()
12820 	 * re-start the picking loop.  in sched_balance_newidle()
12825 	sd = rcu_dereference_check_sched_domain(this_rq->sd);  in sched_balance_newidle()
12827 	if (!get_rd_overloaded(this_rq->rd) ||  in sched_balance_newidle()
12828 	    (sd && this_rq->avg_idle < sd->max_newidle_lb_cost)) {  in sched_balance_newidle()
12849 		if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost)  in sched_balance_newidle()
12852 		if (sd->flags & SD_BALANCE_NEWIDLE) {  in sched_balance_newidle()
12859 			domain_cost = t1 - t0;  in sched_balance_newidle()
12877 	if (curr_cost > this_rq->max_idle_balance_cost)  in sched_balance_newidle()
12878 		this_rq->max_idle_balance_cost = curr_cost;  in sched_balance_newidle()
12885 	if (this_rq->cfs.h_nr_queued && !pulled_task)  in sched_balance_newidle()
12889 	if (this_rq->nr_running != this_rq->cfs.h_nr_queued)  in sched_balance_newidle()
12890 		pulled_task = -1;  in sched_balance_newidle()
12894 	if (time_after(this_rq->next_balance, next_balance))  in sched_balance_newidle()
12895 		this_rq->next_balance = next_balance;  in sched_balance_newidle()
12898 		this_rq->idle_stamp = 0;  in sched_balance_newidle()
12910  * - directly from the local sched_tick() for periodic load balancing
12912  * - indirectly from a remote sched_tick() for NOHZ idle balancing
12913  *   through the SMP cross-call nohz_csd_func()
12918 	enum cpu_idle_type idle = this_rq->idle_balance;  in sched_balance_softirq()
12931 	sched_balance_update_blocked_averages(this_rq->cpu);  in sched_balance_softirq()
12947 	if (time_after_eq(jiffies, rq->next_balance))  in sched_balance_trigger()
12977 	u64 rtime = se->sum_exec_runtime - se->prev_sum_exec_runtime;  in __entity_slice_used()
12978 	u64 slice = se->slice;  in __entity_slice_used()
13000 	 * MIN_NR_TASKS_DURING_FORCEIDLE - 1 tasks and use that to check  in task_tick_core()
13003 	if (rq->core->core_forceidle_count && rq->cfs.nr_queued == 1 &&  in task_tick_core()
13004 	    __entity_slice_used(&curr->se, MIN_NR_TASKS_DURING_FORCEIDLE))  in task_tick_core()
13009  * se_fi_update - Update the cfs_rq->min_vruntime_fi in a CFS hierarchy if needed.
13018 			if (cfs_rq->forceidle_seq == fi_seq)  in se_fi_update()
13020 			cfs_rq->forceidle_seq = fi_seq;  in se_fi_update()
13023 		cfs_rq->min_vruntime_fi = cfs_rq->min_vruntime;  in se_fi_update()
13029 	struct sched_entity *se = &p->se;  in task_vruntime_update()
13031 	if (p->sched_class != &fair_sched_class)  in task_vruntime_update()
13034 	se_fi_update(se, rq->core->core_forceidle_seq, in_fi);  in task_vruntime_update()
13041 	const struct sched_entity *sea = &a->se;  in cfs_prio_less()
13042 	const struct sched_entity *seb = &b->se;  in cfs_prio_less()
13047 	SCHED_WARN_ON(task_rq(b)->core != rq->core);  in cfs_prio_less()
13054 	while (sea->cfs_rq->tg != seb->cfs_rq->tg) {  in cfs_prio_less()
13055 		int sea_depth = sea->depth;  in cfs_prio_less()
13056 		int seb_depth = seb->depth;  in cfs_prio_less()
13064 	se_fi_update(sea, rq->core->core_forceidle_seq, in_fi);  in cfs_prio_less()
13065 	se_fi_update(seb, rq->core->core_forceidle_seq, in_fi);  in cfs_prio_less()
13067 	cfs_rqa = sea->cfs_rq;  in cfs_prio_less()
13068 	cfs_rqb = seb->cfs_rq;  in cfs_prio_less()
13070 	cfs_rqa = &task_rq(a)->cfs;  in cfs_prio_less()
13071 	cfs_rqb = &task_rq(b)->cfs;  in cfs_prio_less()
13079 	delta = (s64)(sea->vruntime - seb->vruntime) +  in cfs_prio_less()
13080 		(s64)(cfs_rqb->min_vruntime_fi - cfs_rqa->min_vruntime_fi);  in cfs_prio_less()
13090 	cfs_rq = task_group(p)->cfs_rq[cpu];  in task_is_throttled_fair()
13092 	cfs_rq = &cpu_rq(cpu)->cfs;  in task_is_throttled_fair()
13111 	struct sched_entity *se = &curr->se;  in task_tick_fair()
13129  *  - child not yet on the tasklist
13130  *  - preemption disabled
13147 	if (rq->cfs.nr_queued == 1)  in prio_changed_fair()
13156 		if (p->prio > oldprio)  in prio_changed_fair()
13178 	se = se->parent;  in propagate_entity_cfs_rq()
13203 	 * - A forked task which hasn't been woken up by wake_up_new_task().  in detach_entity_cfs_rq()
13204 	 * - A task which has been woken up by try_to_wake_up() but is  in detach_entity_cfs_rq()
13207 	if (!se->avg.last_update_time)  in detach_entity_cfs_rq()
13231 	struct sched_entity *se = &p->se;  in detach_task_cfs_rq()
13238 	struct sched_entity *se = &p->se;  in attach_task_cfs_rq()
13250 	SCHED_WARN_ON(p->se.sched_delayed);  in switched_to_fair()
13271 	struct sched_entity *se = &p->se;  in __set_next_task_fair()
13279 		list_move(&se->group_node, &rq->cfs_tasks);  in __set_next_task_fair()
13285 	SCHED_WARN_ON(se->sched_delayed);  in __set_next_task_fair()
13297  * This routine is mostly called to set cfs_rq->curr field when a task
13302 	struct sched_entity *se = &p->se;  in set_next_task_fair()
13317 	cfs_rq->tasks_timeline = RB_ROOT_CACHED;  in init_cfs_rq()
13318 	cfs_rq->min_vruntime = (u64)(-(1LL << 20));  in init_cfs_rq()
13320 	raw_spin_lock_init(&cfs_rq->removed.lock);  in init_cfs_rq()
13331 	if (READ_ONCE(p->__state) == TASK_NEW)  in task_change_group_fair()
13337 	/* Tell se's cfs_rq has been changed -- migrated */  in task_change_group_fair()
13338 	p->se.avg.last_update_time = 0;  in task_change_group_fair()
13349 		if (tg->cfs_rq)  in free_fair_sched_group()
13350 			kfree(tg->cfs_rq[i]);  in free_fair_sched_group()
13351 		if (tg->se)  in free_fair_sched_group()
13352 			kfree(tg->se[i]);  in free_fair_sched_group()
13355 	kfree(tg->cfs_rq);  in free_fair_sched_group()
13356 	kfree(tg->se);  in free_fair_sched_group()
13365 	tg->cfs_rq = kcalloc(nr_cpu_ids, sizeof(cfs_rq), GFP_KERNEL);  in alloc_fair_sched_group()
13366 	if (!tg->cfs_rq)  in alloc_fair_sched_group()
13368 	tg->se = kcalloc(nr_cpu_ids, sizeof(se), GFP_KERNEL);  in alloc_fair_sched_group()
13369 	if (!tg->se)  in alloc_fair_sched_group()
13372 	tg->shares = NICE_0_LOAD;  in alloc_fair_sched_group()
13388 		init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);  in alloc_fair_sched_group()
13409 		se = tg->se[i];  in online_fair_sched_group()
13425 		struct cfs_rq *cfs_rq = tg->cfs_rq[cpu];  in unregister_fair_sched_group()
13426 		struct sched_entity *se = tg->se[cpu];  in unregister_fair_sched_group()
13430 			if (se->sched_delayed) {  in unregister_fair_sched_group()
13432 				if (se->sched_delayed) {  in unregister_fair_sched_group()
13443 		 * check on_list without danger of it being re-added.  in unregister_fair_sched_group()
13445 		if (cfs_rq->on_list) {  in unregister_fair_sched_group()
13458 	cfs_rq->tg = tg;  in init_tg_cfs_entry()
13459 	cfs_rq->rq = rq;  in init_tg_cfs_entry()
13462 	tg->cfs_rq[cpu] = cfs_rq;  in init_tg_cfs_entry()
13463 	tg->se[cpu] = se;  in init_tg_cfs_entry()
13470 		se->cfs_rq = &rq->cfs;  in init_tg_cfs_entry()
13471 		se->depth = 0;  in init_tg_cfs_entry()
13473 		se->cfs_rq = parent->my_q;  in init_tg_cfs_entry()
13474 		se->depth = parent->depth + 1;  in init_tg_cfs_entry()
13477 	se->my_q = cfs_rq;  in init_tg_cfs_entry()
13479 	update_load_set(&se->load, NICE_0_LOAD);  in init_tg_cfs_entry()
13480 	se->parent = parent;  in init_tg_cfs_entry()
13494 	if (!tg->se[0])  in __sched_group_set_shares()
13495 		return -EINVAL;  in __sched_group_set_shares()
13499 	if (tg->shares == shares)  in __sched_group_set_shares()
13502 	tg->shares = shares;  in __sched_group_set_shares()
13505 		struct sched_entity *se = tg->se[i];  in __sched_group_set_shares()
13527 		ret = -EINVAL;  in sched_group_set_shares()
13540 		return -EINVAL;  in sched_group_set_idle()
13543 		return -EINVAL;  in sched_group_set_idle()
13547 	if (tg->idle == idle) {  in sched_group_set_idle()
13552 	tg->idle = idle;  in sched_group_set_idle()
13556 		struct sched_entity *se = tg->se[i];  in sched_group_set_idle()
13557 		struct cfs_rq *grp_cfs_rq = tg->cfs_rq[i];  in sched_group_set_idle()
13564 		grp_cfs_rq->idle = idle;  in sched_group_set_idle()
13568 		idle_task_delta = grp_cfs_rq->h_nr_queued -  in sched_group_set_idle()
13569 				  grp_cfs_rq->h_nr_idle;  in sched_group_set_idle()
13571 			idle_task_delta *= -1;  in sched_group_set_idle()
13576 			if (!se->on_rq)  in sched_group_set_idle()
13579 			cfs_rq->h_nr_idle += idle_task_delta;  in sched_group_set_idle()
13605 	struct sched_entity *se = &task->se;  in get_rr_interval_fair()
13612 	if (rq->cfs.load.weight)  in get_rr_interval_fair()
13613 		rr_interval = NS_TO_JIFFIES(se->slice);  in get_rr_interval_fair()
13691 	ng = rcu_dereference(p->numa_group);  in show_numa_stats()
13693 		if (p->numa_faults) {  in show_numa_stats()
13694 			tsf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 0)];  in show_numa_stats()
13695 			tpf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 1)];  in show_numa_stats()
13698 			gsf = ng->faults[task_faults_idx(NUMA_MEM, node, 0)],  in show_numa_stats()
13699 			gpf = ng->faults[task_faults_idx(NUMA_MEM, node, 1)];  in show_numa_stats()
13720 		INIT_CSD(&cpu_rq(i)->cfsb_csd, __cfsb_csd_unthrottle, cpu_rq(i));  in init_sched_fair_class()
13721 		INIT_LIST_HEAD(&cpu_rq(i)->cfsb_csd_list);  in init_sched_fair_class()