xref: /openbmc/linux/kernel/sched/fair.c (revision e92606fa)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * Completely Fair Scheduling (CFS) Class (SCHED_NORMAL/SCHED_BATCH)
4  *
5  *  Copyright (C) 2007 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
6  *
7  *  Interactivity improvements by Mike Galbraith
8  *  (C) 2007 Mike Galbraith <efault@gmx.de>
9  *
10  *  Various enhancements by Dmitry Adamushko.
11  *  (C) 2007 Dmitry Adamushko <dmitry.adamushko@gmail.com>
12  *
13  *  Group scheduling enhancements by Srivatsa Vaddagiri
14  *  Copyright IBM Corporation, 2007
15  *  Author: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
16  *
17  *  Scaled math optimizations by Thomas Gleixner
18  *  Copyright (C) 2007, Thomas Gleixner <tglx@linutronix.de>
19  *
20  *  Adaptive scheduling granularity, math enhancements by Peter Zijlstra
21  *  Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra
22  */
23 #include <linux/energy_model.h>
24 #include <linux/mmap_lock.h>
25 #include <linux/hugetlb_inline.h>
26 #include <linux/jiffies.h>
27 #include <linux/mm_api.h>
28 #include <linux/highmem.h>
29 #include <linux/spinlock_api.h>
30 #include <linux/cpumask_api.h>
31 #include <linux/lockdep_api.h>
32 #include <linux/softirq.h>
33 #include <linux/refcount_api.h>
34 #include <linux/topology.h>
35 #include <linux/sched/clock.h>
36 #include <linux/sched/cond_resched.h>
37 #include <linux/sched/cputime.h>
38 #include <linux/sched/isolation.h>
39 #include <linux/sched/nohz.h>
40 
41 #include <linux/cpuidle.h>
42 #include <linux/interrupt.h>
43 #include <linux/memory-tiers.h>
44 #include <linux/mempolicy.h>
45 #include <linux/mutex_api.h>
46 #include <linux/profile.h>
47 #include <linux/psi.h>
48 #include <linux/ratelimit.h>
49 #include <linux/task_work.h>
50 
51 #include <asm/switch_to.h>
52 
53 #include <linux/sched/cond_resched.h>
54 
55 #include "sched.h"
56 #include "stats.h"
57 #include "autogroup.h"
58 
59 /*
60  * Targeted preemption latency for CPU-bound tasks:
61  *
62  * NOTE: this latency value is not the same as the concept of
63  * 'timeslice length' - timeslices in CFS are of variable length
64  * and have no persistent notion like in traditional, time-slice
65  * based scheduling concepts.
66  *
67  * (to see the precise effective timeslice length of your workload,
68  *  run vmstat and monitor the context-switches (cs) field)
69  *
70  * (default: 6ms * (1 + ilog(ncpus)), units: nanoseconds)
71  */
72 unsigned int sysctl_sched_latency			= 6000000ULL;
73 static unsigned int normalized_sysctl_sched_latency	= 6000000ULL;
74 
75 /*
76  * The initial- and re-scaling of tunables is configurable
77  *
78  * Options are:
79  *
80  *   SCHED_TUNABLESCALING_NONE - unscaled, always *1
81  *   SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus)
82  *   SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus
83  *
84  * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus))
85  */
86 unsigned int sysctl_sched_tunable_scaling = SCHED_TUNABLESCALING_LOG;
87 
88 /*
89  * Minimal preemption granularity for CPU-bound tasks:
90  *
91  * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds)
92  */
93 unsigned int sysctl_sched_min_granularity			= 750000ULL;
94 static unsigned int normalized_sysctl_sched_min_granularity	= 750000ULL;
95 
96 /*
97  * Minimal preemption granularity for CPU-bound SCHED_IDLE tasks.
98  * Applies only when SCHED_IDLE tasks compete with normal tasks.
99  *
100  * (default: 0.75 msec)
101  */
102 unsigned int sysctl_sched_idle_min_granularity			= 750000ULL;
103 
104 /*
105  * This value is kept at sysctl_sched_latency/sysctl_sched_min_granularity
106  */
107 static unsigned int sched_nr_latency = 8;
108 
109 /*
110  * After fork, child runs first. If set to 0 (default) then
111  * parent will (try to) run first.
112  */
113 unsigned int sysctl_sched_child_runs_first __read_mostly;
114 
115 /*
116  * SCHED_OTHER wake-up granularity.
117  *
118  * This option delays the preemption effects of decoupled workloads
119  * and reduces their over-scheduling. Synchronous workloads will still
120  * have immediate wakeup/sleep latencies.
121  *
122  * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds)
123  */
124 unsigned int sysctl_sched_wakeup_granularity			= 1000000UL;
125 static unsigned int normalized_sysctl_sched_wakeup_granularity	= 1000000UL;
126 
127 const_debug unsigned int sysctl_sched_migration_cost	= 500000UL;
128 
129 int sched_thermal_decay_shift;
130 static int __init setup_sched_thermal_decay_shift(char *str)
131 {
132 	int _shift = 0;
133 
134 	if (kstrtoint(str, 0, &_shift))
135 		pr_warn("Unable to set scheduler thermal pressure decay shift parameter\n");
136 
137 	sched_thermal_decay_shift = clamp(_shift, 0, 10);
138 	return 1;
139 }
140 __setup("sched_thermal_decay_shift=", setup_sched_thermal_decay_shift);
141 
142 #ifdef CONFIG_SMP
143 /*
144  * For asym packing, by default the lower numbered CPU has higher priority.
145  */
146 int __weak arch_asym_cpu_priority(int cpu)
147 {
148 	return -cpu;
149 }
150 
151 /*
152  * The margin used when comparing utilization with CPU capacity.
153  *
154  * (default: ~20%)
155  */
156 #define fits_capacity(cap, max)	((cap) * 1280 < (max) * 1024)
157 
158 /*
159  * The margin used when comparing CPU capacities.
160  * is 'cap1' noticeably greater than 'cap2'
161  *
162  * (default: ~5%)
163  */
164 #define capacity_greater(cap1, cap2) ((cap1) * 1024 > (cap2) * 1078)
165 #endif
166 
167 #ifdef CONFIG_CFS_BANDWIDTH
168 /*
169  * Amount of runtime to allocate from global (tg) to local (per-cfs_rq) pool
170  * each time a cfs_rq requests quota.
171  *
172  * Note: in the case that the slice exceeds the runtime remaining (either due
173  * to consumption or the quota being specified to be smaller than the slice)
174  * we will always only issue the remaining available time.
175  *
176  * (default: 5 msec, units: microseconds)
177  */
178 static unsigned int sysctl_sched_cfs_bandwidth_slice		= 5000UL;
179 #endif
180 
181 #ifdef CONFIG_NUMA_BALANCING
182 /* Restrict the NUMA promotion throughput (MB/s) for each target node. */
183 static unsigned int sysctl_numa_balancing_promote_rate_limit = 65536;
184 #endif
185 
186 #ifdef CONFIG_SYSCTL
187 static struct ctl_table sched_fair_sysctls[] = {
188 	{
189 		.procname       = "sched_child_runs_first",
190 		.data           = &sysctl_sched_child_runs_first,
191 		.maxlen         = sizeof(unsigned int),
192 		.mode           = 0644,
193 		.proc_handler   = proc_dointvec,
194 	},
195 #ifdef CONFIG_CFS_BANDWIDTH
196 	{
197 		.procname       = "sched_cfs_bandwidth_slice_us",
198 		.data           = &sysctl_sched_cfs_bandwidth_slice,
199 		.maxlen         = sizeof(unsigned int),
200 		.mode           = 0644,
201 		.proc_handler   = proc_dointvec_minmax,
202 		.extra1         = SYSCTL_ONE,
203 	},
204 #endif
205 #ifdef CONFIG_NUMA_BALANCING
206 	{
207 		.procname	= "numa_balancing_promote_rate_limit_MBps",
208 		.data		= &sysctl_numa_balancing_promote_rate_limit,
209 		.maxlen		= sizeof(unsigned int),
210 		.mode		= 0644,
211 		.proc_handler	= proc_dointvec_minmax,
212 		.extra1		= SYSCTL_ZERO,
213 	},
214 #endif /* CONFIG_NUMA_BALANCING */
215 	{}
216 };
217 
218 static int __init sched_fair_sysctl_init(void)
219 {
220 	register_sysctl_init("kernel", sched_fair_sysctls);
221 	return 0;
222 }
223 late_initcall(sched_fair_sysctl_init);
224 #endif
225 
226 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
227 {
228 	lw->weight += inc;
229 	lw->inv_weight = 0;
230 }
231 
232 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
233 {
234 	lw->weight -= dec;
235 	lw->inv_weight = 0;
236 }
237 
238 static inline void update_load_set(struct load_weight *lw, unsigned long w)
239 {
240 	lw->weight = w;
241 	lw->inv_weight = 0;
242 }
243 
244 /*
245  * Increase the granularity value when there are more CPUs,
246  * because with more CPUs the 'effective latency' as visible
247  * to users decreases. But the relationship is not linear,
248  * so pick a second-best guess by going with the log2 of the
249  * number of CPUs.
250  *
251  * This idea comes from the SD scheduler of Con Kolivas:
252  */
253 static unsigned int get_update_sysctl_factor(void)
254 {
255 	unsigned int cpus = min_t(unsigned int, num_online_cpus(), 8);
256 	unsigned int factor;
257 
258 	switch (sysctl_sched_tunable_scaling) {
259 	case SCHED_TUNABLESCALING_NONE:
260 		factor = 1;
261 		break;
262 	case SCHED_TUNABLESCALING_LINEAR:
263 		factor = cpus;
264 		break;
265 	case SCHED_TUNABLESCALING_LOG:
266 	default:
267 		factor = 1 + ilog2(cpus);
268 		break;
269 	}
270 
271 	return factor;
272 }
273 
274 static void update_sysctl(void)
275 {
276 	unsigned int factor = get_update_sysctl_factor();
277 
278 #define SET_SYSCTL(name) \
279 	(sysctl_##name = (factor) * normalized_sysctl_##name)
280 	SET_SYSCTL(sched_min_granularity);
281 	SET_SYSCTL(sched_latency);
282 	SET_SYSCTL(sched_wakeup_granularity);
283 #undef SET_SYSCTL
284 }
285 
286 void __init sched_init_granularity(void)
287 {
288 	update_sysctl();
289 }
290 
291 #define WMULT_CONST	(~0U)
292 #define WMULT_SHIFT	32
293 
294 static void __update_inv_weight(struct load_weight *lw)
295 {
296 	unsigned long w;
297 
298 	if (likely(lw->inv_weight))
299 		return;
300 
301 	w = scale_load_down(lw->weight);
302 
303 	if (BITS_PER_LONG > 32 && unlikely(w >= WMULT_CONST))
304 		lw->inv_weight = 1;
305 	else if (unlikely(!w))
306 		lw->inv_weight = WMULT_CONST;
307 	else
308 		lw->inv_weight = WMULT_CONST / w;
309 }
310 
311 /*
312  * delta_exec * weight / lw.weight
313  *   OR
314  * (delta_exec * (weight * lw->inv_weight)) >> WMULT_SHIFT
315  *
316  * Either weight := NICE_0_LOAD and lw \e sched_prio_to_wmult[], in which case
317  * we're guaranteed shift stays positive because inv_weight is guaranteed to
318  * fit 32 bits, and NICE_0_LOAD gives another 10 bits; therefore shift >= 22.
319  *
320  * Or, weight =< lw.weight (because lw.weight is the runqueue weight), thus
321  * weight/lw.weight <= 1, and therefore our shift will also be positive.
322  */
323 static u64 __calc_delta(u64 delta_exec, unsigned long weight, struct load_weight *lw)
324 {
325 	u64 fact = scale_load_down(weight);
326 	u32 fact_hi = (u32)(fact >> 32);
327 	int shift = WMULT_SHIFT;
328 	int fs;
329 
330 	__update_inv_weight(lw);
331 
332 	if (unlikely(fact_hi)) {
333 		fs = fls(fact_hi);
334 		shift -= fs;
335 		fact >>= fs;
336 	}
337 
338 	fact = mul_u32_u32(fact, lw->inv_weight);
339 
340 	fact_hi = (u32)(fact >> 32);
341 	if (fact_hi) {
342 		fs = fls(fact_hi);
343 		shift -= fs;
344 		fact >>= fs;
345 	}
346 
347 	return mul_u64_u32_shr(delta_exec, fact, shift);
348 }
349 
350 
351 const struct sched_class fair_sched_class;
352 
353 /**************************************************************
354  * CFS operations on generic schedulable entities:
355  */
356 
357 #ifdef CONFIG_FAIR_GROUP_SCHED
358 
359 /* Walk up scheduling entities hierarchy */
360 #define for_each_sched_entity(se) \
361 		for (; se; se = se->parent)
362 
363 static inline bool list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
364 {
365 	struct rq *rq = rq_of(cfs_rq);
366 	int cpu = cpu_of(rq);
367 
368 	if (cfs_rq->on_list)
369 		return rq->tmp_alone_branch == &rq->leaf_cfs_rq_list;
370 
371 	cfs_rq->on_list = 1;
372 
373 	/*
374 	 * Ensure we either appear before our parent (if already
375 	 * enqueued) or force our parent to appear after us when it is
376 	 * enqueued. The fact that we always enqueue bottom-up
377 	 * reduces this to two cases and a special case for the root
378 	 * cfs_rq. Furthermore, it also means that we will always reset
379 	 * tmp_alone_branch either when the branch is connected
380 	 * to a tree or when we reach the top of the tree
381 	 */
382 	if (cfs_rq->tg->parent &&
383 	    cfs_rq->tg->parent->cfs_rq[cpu]->on_list) {
384 		/*
385 		 * If parent is already on the list, we add the child
386 		 * just before. Thanks to circular linked property of
387 		 * the list, this means to put the child at the tail
388 		 * of the list that starts by parent.
389 		 */
390 		list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
391 			&(cfs_rq->tg->parent->cfs_rq[cpu]->leaf_cfs_rq_list));
392 		/*
393 		 * The branch is now connected to its tree so we can
394 		 * reset tmp_alone_branch to the beginning of the
395 		 * list.
396 		 */
397 		rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
398 		return true;
399 	}
400 
401 	if (!cfs_rq->tg->parent) {
402 		/*
403 		 * cfs rq without parent should be put
404 		 * at the tail of the list.
405 		 */
406 		list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
407 			&rq->leaf_cfs_rq_list);
408 		/*
409 		 * We have reach the top of a tree so we can reset
410 		 * tmp_alone_branch to the beginning of the list.
411 		 */
412 		rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
413 		return true;
414 	}
415 
416 	/*
417 	 * The parent has not already been added so we want to
418 	 * make sure that it will be put after us.
419 	 * tmp_alone_branch points to the begin of the branch
420 	 * where we will add parent.
421 	 */
422 	list_add_rcu(&cfs_rq->leaf_cfs_rq_list, rq->tmp_alone_branch);
423 	/*
424 	 * update tmp_alone_branch to points to the new begin
425 	 * of the branch
426 	 */
427 	rq->tmp_alone_branch = &cfs_rq->leaf_cfs_rq_list;
428 	return false;
429 }
430 
431 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
432 {
433 	if (cfs_rq->on_list) {
434 		struct rq *rq = rq_of(cfs_rq);
435 
436 		/*
437 		 * With cfs_rq being unthrottled/throttled during an enqueue,
438 		 * it can happen the tmp_alone_branch points the a leaf that
439 		 * we finally want to del. In this case, tmp_alone_branch moves
440 		 * to the prev element but it will point to rq->leaf_cfs_rq_list
441 		 * at the end of the enqueue.
442 		 */
443 		if (rq->tmp_alone_branch == &cfs_rq->leaf_cfs_rq_list)
444 			rq->tmp_alone_branch = cfs_rq->leaf_cfs_rq_list.prev;
445 
446 		list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
447 		cfs_rq->on_list = 0;
448 	}
449 }
450 
451 static inline void assert_list_leaf_cfs_rq(struct rq *rq)
452 {
453 	SCHED_WARN_ON(rq->tmp_alone_branch != &rq->leaf_cfs_rq_list);
454 }
455 
456 /* Iterate thr' all leaf cfs_rq's on a runqueue */
457 #define for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos)			\
458 	list_for_each_entry_safe(cfs_rq, pos, &rq->leaf_cfs_rq_list,	\
459 				 leaf_cfs_rq_list)
460 
461 /* Do the two (enqueued) entities belong to the same group ? */
462 static inline struct cfs_rq *
463 is_same_group(struct sched_entity *se, struct sched_entity *pse)
464 {
465 	if (se->cfs_rq == pse->cfs_rq)
466 		return se->cfs_rq;
467 
468 	return NULL;
469 }
470 
471 static inline struct sched_entity *parent_entity(const struct sched_entity *se)
472 {
473 	return se->parent;
474 }
475 
476 static void
477 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
478 {
479 	int se_depth, pse_depth;
480 
481 	/*
482 	 * preemption test can be made between sibling entities who are in the
483 	 * same cfs_rq i.e who have a common parent. Walk up the hierarchy of
484 	 * both tasks until we find their ancestors who are siblings of common
485 	 * parent.
486 	 */
487 
488 	/* First walk up until both entities are at same depth */
489 	se_depth = (*se)->depth;
490 	pse_depth = (*pse)->depth;
491 
492 	while (se_depth > pse_depth) {
493 		se_depth--;
494 		*se = parent_entity(*se);
495 	}
496 
497 	while (pse_depth > se_depth) {
498 		pse_depth--;
499 		*pse = parent_entity(*pse);
500 	}
501 
502 	while (!is_same_group(*se, *pse)) {
503 		*se = parent_entity(*se);
504 		*pse = parent_entity(*pse);
505 	}
506 }
507 
508 static int tg_is_idle(struct task_group *tg)
509 {
510 	return tg->idle > 0;
511 }
512 
513 static int cfs_rq_is_idle(struct cfs_rq *cfs_rq)
514 {
515 	return cfs_rq->idle > 0;
516 }
517 
518 static int se_is_idle(struct sched_entity *se)
519 {
520 	if (entity_is_task(se))
521 		return task_has_idle_policy(task_of(se));
522 	return cfs_rq_is_idle(group_cfs_rq(se));
523 }
524 
525 #else	/* !CONFIG_FAIR_GROUP_SCHED */
526 
527 #define for_each_sched_entity(se) \
528 		for (; se; se = NULL)
529 
530 static inline bool list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
531 {
532 	return true;
533 }
534 
535 static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
536 {
537 }
538 
539 static inline void assert_list_leaf_cfs_rq(struct rq *rq)
540 {
541 }
542 
543 #define for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos)	\
544 		for (cfs_rq = &rq->cfs, pos = NULL; cfs_rq; cfs_rq = pos)
545 
546 static inline struct sched_entity *parent_entity(struct sched_entity *se)
547 {
548 	return NULL;
549 }
550 
551 static inline void
552 find_matching_se(struct sched_entity **se, struct sched_entity **pse)
553 {
554 }
555 
556 static inline int tg_is_idle(struct task_group *tg)
557 {
558 	return 0;
559 }
560 
561 static int cfs_rq_is_idle(struct cfs_rq *cfs_rq)
562 {
563 	return 0;
564 }
565 
566 static int se_is_idle(struct sched_entity *se)
567 {
568 	return 0;
569 }
570 
571 #endif	/* CONFIG_FAIR_GROUP_SCHED */
572 
573 static __always_inline
574 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec);
575 
576 /**************************************************************
577  * Scheduling class tree data structure manipulation methods:
578  */
579 
580 static inline u64 max_vruntime(u64 max_vruntime, u64 vruntime)
581 {
582 	s64 delta = (s64)(vruntime - max_vruntime);
583 	if (delta > 0)
584 		max_vruntime = vruntime;
585 
586 	return max_vruntime;
587 }
588 
589 static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime)
590 {
591 	s64 delta = (s64)(vruntime - min_vruntime);
592 	if (delta < 0)
593 		min_vruntime = vruntime;
594 
595 	return min_vruntime;
596 }
597 
598 static inline bool entity_before(const struct sched_entity *a,
599 				 const struct sched_entity *b)
600 {
601 	return (s64)(a->vruntime - b->vruntime) < 0;
602 }
603 
604 #define __node_2_se(node) \
605 	rb_entry((node), struct sched_entity, run_node)
606 
607 static void update_min_vruntime(struct cfs_rq *cfs_rq)
608 {
609 	struct sched_entity *curr = cfs_rq->curr;
610 	struct rb_node *leftmost = rb_first_cached(&cfs_rq->tasks_timeline);
611 
612 	u64 vruntime = cfs_rq->min_vruntime;
613 
614 	if (curr) {
615 		if (curr->on_rq)
616 			vruntime = curr->vruntime;
617 		else
618 			curr = NULL;
619 	}
620 
621 	if (leftmost) { /* non-empty tree */
622 		struct sched_entity *se = __node_2_se(leftmost);
623 
624 		if (!curr)
625 			vruntime = se->vruntime;
626 		else
627 			vruntime = min_vruntime(vruntime, se->vruntime);
628 	}
629 
630 	/* ensure we never gain time by being placed backwards. */
631 	u64_u32_store(cfs_rq->min_vruntime,
632 		      max_vruntime(cfs_rq->min_vruntime, vruntime));
633 }
634 
635 static inline bool __entity_less(struct rb_node *a, const struct rb_node *b)
636 {
637 	return entity_before(__node_2_se(a), __node_2_se(b));
638 }
639 
640 /*
641  * Enqueue an entity into the rb-tree:
642  */
643 static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
644 {
645 	rb_add_cached(&se->run_node, &cfs_rq->tasks_timeline, __entity_less);
646 }
647 
648 static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
649 {
650 	rb_erase_cached(&se->run_node, &cfs_rq->tasks_timeline);
651 }
652 
653 struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq)
654 {
655 	struct rb_node *left = rb_first_cached(&cfs_rq->tasks_timeline);
656 
657 	if (!left)
658 		return NULL;
659 
660 	return __node_2_se(left);
661 }
662 
663 static struct sched_entity *__pick_next_entity(struct sched_entity *se)
664 {
665 	struct rb_node *next = rb_next(&se->run_node);
666 
667 	if (!next)
668 		return NULL;
669 
670 	return __node_2_se(next);
671 }
672 
673 #ifdef CONFIG_SCHED_DEBUG
674 struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
675 {
676 	struct rb_node *last = rb_last(&cfs_rq->tasks_timeline.rb_root);
677 
678 	if (!last)
679 		return NULL;
680 
681 	return __node_2_se(last);
682 }
683 
684 /**************************************************************
685  * Scheduling class statistics methods:
686  */
687 
688 int sched_update_scaling(void)
689 {
690 	unsigned int factor = get_update_sysctl_factor();
691 
692 	sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency,
693 					sysctl_sched_min_granularity);
694 
695 #define WRT_SYSCTL(name) \
696 	(normalized_sysctl_##name = sysctl_##name / (factor))
697 	WRT_SYSCTL(sched_min_granularity);
698 	WRT_SYSCTL(sched_latency);
699 	WRT_SYSCTL(sched_wakeup_granularity);
700 #undef WRT_SYSCTL
701 
702 	return 0;
703 }
704 #endif
705 
706 /*
707  * delta /= w
708  */
709 static inline u64 calc_delta_fair(u64 delta, struct sched_entity *se)
710 {
711 	if (unlikely(se->load.weight != NICE_0_LOAD))
712 		delta = __calc_delta(delta, NICE_0_LOAD, &se->load);
713 
714 	return delta;
715 }
716 
717 /*
718  * The idea is to set a period in which each task runs once.
719  *
720  * When there are too many tasks (sched_nr_latency) we have to stretch
721  * this period because otherwise the slices get too small.
722  *
723  * p = (nr <= nl) ? l : l*nr/nl
724  */
725 static u64 __sched_period(unsigned long nr_running)
726 {
727 	if (unlikely(nr_running > sched_nr_latency))
728 		return nr_running * sysctl_sched_min_granularity;
729 	else
730 		return sysctl_sched_latency;
731 }
732 
733 static bool sched_idle_cfs_rq(struct cfs_rq *cfs_rq);
734 
735 /*
736  * We calculate the wall-time slice from the period by taking a part
737  * proportional to the weight.
738  *
739  * s = p*P[w/rw]
740  */
741 static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
742 {
743 	unsigned int nr_running = cfs_rq->nr_running;
744 	struct sched_entity *init_se = se;
745 	unsigned int min_gran;
746 	u64 slice;
747 
748 	if (sched_feat(ALT_PERIOD))
749 		nr_running = rq_of(cfs_rq)->cfs.h_nr_running;
750 
751 	slice = __sched_period(nr_running + !se->on_rq);
752 
753 	for_each_sched_entity(se) {
754 		struct load_weight *load;
755 		struct load_weight lw;
756 		struct cfs_rq *qcfs_rq;
757 
758 		qcfs_rq = cfs_rq_of(se);
759 		load = &qcfs_rq->load;
760 
761 		if (unlikely(!se->on_rq)) {
762 			lw = qcfs_rq->load;
763 
764 			update_load_add(&lw, se->load.weight);
765 			load = &lw;
766 		}
767 		slice = __calc_delta(slice, se->load.weight, load);
768 	}
769 
770 	if (sched_feat(BASE_SLICE)) {
771 		if (se_is_idle(init_se) && !sched_idle_cfs_rq(cfs_rq))
772 			min_gran = sysctl_sched_idle_min_granularity;
773 		else
774 			min_gran = sysctl_sched_min_granularity;
775 
776 		slice = max_t(u64, slice, min_gran);
777 	}
778 
779 	return slice;
780 }
781 
782 /*
783  * We calculate the vruntime slice of a to-be-inserted task.
784  *
785  * vs = s/w
786  */
787 static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se)
788 {
789 	return calc_delta_fair(sched_slice(cfs_rq, se), se);
790 }
791 
792 #include "pelt.h"
793 #ifdef CONFIG_SMP
794 
795 static int select_idle_sibling(struct task_struct *p, int prev_cpu, int cpu);
796 static unsigned long task_h_load(struct task_struct *p);
797 static unsigned long capacity_of(int cpu);
798 
799 /* Give new sched_entity start runnable values to heavy its load in infant time */
800 void init_entity_runnable_average(struct sched_entity *se)
801 {
802 	struct sched_avg *sa = &se->avg;
803 
804 	memset(sa, 0, sizeof(*sa));
805 
806 	/*
807 	 * Tasks are initialized with full load to be seen as heavy tasks until
808 	 * they get a chance to stabilize to their real load level.
809 	 * Group entities are initialized with zero load to reflect the fact that
810 	 * nothing has been attached to the task group yet.
811 	 */
812 	if (entity_is_task(se))
813 		sa->load_avg = scale_load_down(se->load.weight);
814 
815 	/* when this task enqueue'ed, it will contribute to its cfs_rq's load_avg */
816 }
817 
818 /*
819  * With new tasks being created, their initial util_avgs are extrapolated
820  * based on the cfs_rq's current util_avg:
821  *
822  *   util_avg = cfs_rq->util_avg / (cfs_rq->load_avg + 1) * se.load.weight
823  *
824  * However, in many cases, the above util_avg does not give a desired
825  * value. Moreover, the sum of the util_avgs may be divergent, such
826  * as when the series is a harmonic series.
827  *
828  * To solve this problem, we also cap the util_avg of successive tasks to
829  * only 1/2 of the left utilization budget:
830  *
831  *   util_avg_cap = (cpu_scale - cfs_rq->avg.util_avg) / 2^n
832  *
833  * where n denotes the nth task and cpu_scale the CPU capacity.
834  *
835  * For example, for a CPU with 1024 of capacity, a simplest series from
836  * the beginning would be like:
837  *
838  *  task  util_avg: 512, 256, 128,  64,  32,   16,    8, ...
839  * cfs_rq util_avg: 512, 768, 896, 960, 992, 1008, 1016, ...
840  *
841  * Finally, that extrapolated util_avg is clamped to the cap (util_avg_cap)
842  * if util_avg > util_avg_cap.
843  */
844 void post_init_entity_util_avg(struct task_struct *p)
845 {
846 	struct sched_entity *se = &p->se;
847 	struct cfs_rq *cfs_rq = cfs_rq_of(se);
848 	struct sched_avg *sa = &se->avg;
849 	long cpu_scale = arch_scale_cpu_capacity(cpu_of(rq_of(cfs_rq)));
850 	long cap = (long)(cpu_scale - cfs_rq->avg.util_avg) / 2;
851 
852 	if (p->sched_class != &fair_sched_class) {
853 		/*
854 		 * For !fair tasks do:
855 		 *
856 		update_cfs_rq_load_avg(now, cfs_rq);
857 		attach_entity_load_avg(cfs_rq, se);
858 		switched_from_fair(rq, p);
859 		 *
860 		 * such that the next switched_to_fair() has the
861 		 * expected state.
862 		 */
863 		se->avg.last_update_time = cfs_rq_clock_pelt(cfs_rq);
864 		return;
865 	}
866 
867 	if (cap > 0) {
868 		if (cfs_rq->avg.util_avg != 0) {
869 			sa->util_avg  = cfs_rq->avg.util_avg * se->load.weight;
870 			sa->util_avg /= (cfs_rq->avg.load_avg + 1);
871 
872 			if (sa->util_avg > cap)
873 				sa->util_avg = cap;
874 		} else {
875 			sa->util_avg = cap;
876 		}
877 	}
878 
879 	sa->runnable_avg = sa->util_avg;
880 }
881 
882 #else /* !CONFIG_SMP */
883 void init_entity_runnable_average(struct sched_entity *se)
884 {
885 }
886 void post_init_entity_util_avg(struct task_struct *p)
887 {
888 }
889 static void update_tg_load_avg(struct cfs_rq *cfs_rq)
890 {
891 }
892 #endif /* CONFIG_SMP */
893 
894 /*
895  * Update the current task's runtime statistics.
896  */
897 static void update_curr(struct cfs_rq *cfs_rq)
898 {
899 	struct sched_entity *curr = cfs_rq->curr;
900 	u64 now = rq_clock_task(rq_of(cfs_rq));
901 	u64 delta_exec;
902 
903 	if (unlikely(!curr))
904 		return;
905 
906 	delta_exec = now - curr->exec_start;
907 	if (unlikely((s64)delta_exec <= 0))
908 		return;
909 
910 	curr->exec_start = now;
911 
912 	if (schedstat_enabled()) {
913 		struct sched_statistics *stats;
914 
915 		stats = __schedstats_from_se(curr);
916 		__schedstat_set(stats->exec_max,
917 				max(delta_exec, stats->exec_max));
918 	}
919 
920 	curr->sum_exec_runtime += delta_exec;
921 	schedstat_add(cfs_rq->exec_clock, delta_exec);
922 
923 	curr->vruntime += calc_delta_fair(delta_exec, curr);
924 	update_min_vruntime(cfs_rq);
925 
926 	if (entity_is_task(curr)) {
927 		struct task_struct *curtask = task_of(curr);
928 
929 		trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime);
930 		cgroup_account_cputime(curtask, delta_exec);
931 		account_group_exec_runtime(curtask, delta_exec);
932 	}
933 
934 	account_cfs_rq_runtime(cfs_rq, delta_exec);
935 }
936 
937 static void update_curr_fair(struct rq *rq)
938 {
939 	update_curr(cfs_rq_of(&rq->curr->se));
940 }
941 
942 static inline void
943 update_stats_wait_start_fair(struct cfs_rq *cfs_rq, struct sched_entity *se)
944 {
945 	struct sched_statistics *stats;
946 	struct task_struct *p = NULL;
947 
948 	if (!schedstat_enabled())
949 		return;
950 
951 	stats = __schedstats_from_se(se);
952 
953 	if (entity_is_task(se))
954 		p = task_of(se);
955 
956 	__update_stats_wait_start(rq_of(cfs_rq), p, stats);
957 }
958 
959 static inline void
960 update_stats_wait_end_fair(struct cfs_rq *cfs_rq, struct sched_entity *se)
961 {
962 	struct sched_statistics *stats;
963 	struct task_struct *p = NULL;
964 
965 	if (!schedstat_enabled())
966 		return;
967 
968 	stats = __schedstats_from_se(se);
969 
970 	/*
971 	 * When the sched_schedstat changes from 0 to 1, some sched se
972 	 * maybe already in the runqueue, the se->statistics.wait_start
973 	 * will be 0.So it will let the delta wrong. We need to avoid this
974 	 * scenario.
975 	 */
976 	if (unlikely(!schedstat_val(stats->wait_start)))
977 		return;
978 
979 	if (entity_is_task(se))
980 		p = task_of(se);
981 
982 	__update_stats_wait_end(rq_of(cfs_rq), p, stats);
983 }
984 
985 static inline void
986 update_stats_enqueue_sleeper_fair(struct cfs_rq *cfs_rq, struct sched_entity *se)
987 {
988 	struct sched_statistics *stats;
989 	struct task_struct *tsk = NULL;
990 
991 	if (!schedstat_enabled())
992 		return;
993 
994 	stats = __schedstats_from_se(se);
995 
996 	if (entity_is_task(se))
997 		tsk = task_of(se);
998 
999 	__update_stats_enqueue_sleeper(rq_of(cfs_rq), tsk, stats);
1000 }
1001 
1002 /*
1003  * Task is being enqueued - update stats:
1004  */
1005 static inline void
1006 update_stats_enqueue_fair(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
1007 {
1008 	if (!schedstat_enabled())
1009 		return;
1010 
1011 	/*
1012 	 * Are we enqueueing a waiting task? (for current tasks
1013 	 * a dequeue/enqueue event is a NOP)
1014 	 */
1015 	if (se != cfs_rq->curr)
1016 		update_stats_wait_start_fair(cfs_rq, se);
1017 
1018 	if (flags & ENQUEUE_WAKEUP)
1019 		update_stats_enqueue_sleeper_fair(cfs_rq, se);
1020 }
1021 
1022 static inline void
1023 update_stats_dequeue_fair(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
1024 {
1025 
1026 	if (!schedstat_enabled())
1027 		return;
1028 
1029 	/*
1030 	 * Mark the end of the wait period if dequeueing a
1031 	 * waiting task:
1032 	 */
1033 	if (se != cfs_rq->curr)
1034 		update_stats_wait_end_fair(cfs_rq, se);
1035 
1036 	if ((flags & DEQUEUE_SLEEP) && entity_is_task(se)) {
1037 		struct task_struct *tsk = task_of(se);
1038 		unsigned int state;
1039 
1040 		/* XXX racy against TTWU */
1041 		state = READ_ONCE(tsk->__state);
1042 		if (state & TASK_INTERRUPTIBLE)
1043 			__schedstat_set(tsk->stats.sleep_start,
1044 				      rq_clock(rq_of(cfs_rq)));
1045 		if (state & TASK_UNINTERRUPTIBLE)
1046 			__schedstat_set(tsk->stats.block_start,
1047 				      rq_clock(rq_of(cfs_rq)));
1048 	}
1049 }
1050 
1051 /*
1052  * We are picking a new current task - update its stats:
1053  */
1054 static inline void
1055 update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
1056 {
1057 	/*
1058 	 * We are starting a new run period:
1059 	 */
1060 	se->exec_start = rq_clock_task(rq_of(cfs_rq));
1061 }
1062 
1063 /**************************************************
1064  * Scheduling class queueing methods:
1065  */
1066 
1067 #ifdef CONFIG_NUMA
1068 #define NUMA_IMBALANCE_MIN 2
1069 
1070 static inline long
1071 adjust_numa_imbalance(int imbalance, int dst_running, int imb_numa_nr)
1072 {
1073 	/*
1074 	 * Allow a NUMA imbalance if busy CPUs is less than the maximum
1075 	 * threshold. Above this threshold, individual tasks may be contending
1076 	 * for both memory bandwidth and any shared HT resources.  This is an
1077 	 * approximation as the number of running tasks may not be related to
1078 	 * the number of busy CPUs due to sched_setaffinity.
1079 	 */
1080 	if (dst_running > imb_numa_nr)
1081 		return imbalance;
1082 
1083 	/*
1084 	 * Allow a small imbalance based on a simple pair of communicating
1085 	 * tasks that remain local when the destination is lightly loaded.
1086 	 */
1087 	if (imbalance <= NUMA_IMBALANCE_MIN)
1088 		return 0;
1089 
1090 	return imbalance;
1091 }
1092 #endif /* CONFIG_NUMA */
1093 
1094 #ifdef CONFIG_NUMA_BALANCING
1095 /*
1096  * Approximate time to scan a full NUMA task in ms. The task scan period is
1097  * calculated based on the tasks virtual memory size and
1098  * numa_balancing_scan_size.
1099  */
1100 unsigned int sysctl_numa_balancing_scan_period_min = 1000;
1101 unsigned int sysctl_numa_balancing_scan_period_max = 60000;
1102 
1103 /* Portion of address space to scan in MB */
1104 unsigned int sysctl_numa_balancing_scan_size = 256;
1105 
1106 /* Scan @scan_size MB every @scan_period after an initial @scan_delay in ms */
1107 unsigned int sysctl_numa_balancing_scan_delay = 1000;
1108 
1109 /* The page with hint page fault latency < threshold in ms is considered hot */
1110 unsigned int sysctl_numa_balancing_hot_threshold = MSEC_PER_SEC;
1111 
1112 struct numa_group {
1113 	refcount_t refcount;
1114 
1115 	spinlock_t lock; /* nr_tasks, tasks */
1116 	int nr_tasks;
1117 	pid_t gid;
1118 	int active_nodes;
1119 
1120 	struct rcu_head rcu;
1121 	unsigned long total_faults;
1122 	unsigned long max_faults_cpu;
1123 	/*
1124 	 * faults[] array is split into two regions: faults_mem and faults_cpu.
1125 	 *
1126 	 * Faults_cpu is used to decide whether memory should move
1127 	 * towards the CPU. As a consequence, these stats are weighted
1128 	 * more by CPU use than by memory faults.
1129 	 */
1130 	unsigned long faults[];
1131 };
1132 
1133 /*
1134  * For functions that can be called in multiple contexts that permit reading
1135  * ->numa_group (see struct task_struct for locking rules).
1136  */
1137 static struct numa_group *deref_task_numa_group(struct task_struct *p)
1138 {
1139 	return rcu_dereference_check(p->numa_group, p == current ||
1140 		(lockdep_is_held(__rq_lockp(task_rq(p))) && !READ_ONCE(p->on_cpu)));
1141 }
1142 
1143 static struct numa_group *deref_curr_numa_group(struct task_struct *p)
1144 {
1145 	return rcu_dereference_protected(p->numa_group, p == current);
1146 }
1147 
1148 static inline unsigned long group_faults_priv(struct numa_group *ng);
1149 static inline unsigned long group_faults_shared(struct numa_group *ng);
1150 
1151 static unsigned int task_nr_scan_windows(struct task_struct *p)
1152 {
1153 	unsigned long rss = 0;
1154 	unsigned long nr_scan_pages;
1155 
1156 	/*
1157 	 * Calculations based on RSS as non-present and empty pages are skipped
1158 	 * by the PTE scanner and NUMA hinting faults should be trapped based
1159 	 * on resident pages
1160 	 */
1161 	nr_scan_pages = sysctl_numa_balancing_scan_size << (20 - PAGE_SHIFT);
1162 	rss = get_mm_rss(p->mm);
1163 	if (!rss)
1164 		rss = nr_scan_pages;
1165 
1166 	rss = round_up(rss, nr_scan_pages);
1167 	return rss / nr_scan_pages;
1168 }
1169 
1170 /* For sanity's sake, never scan more PTEs than MAX_SCAN_WINDOW MB/sec. */
1171 #define MAX_SCAN_WINDOW 2560
1172 
1173 static unsigned int task_scan_min(struct task_struct *p)
1174 {
1175 	unsigned int scan_size = READ_ONCE(sysctl_numa_balancing_scan_size);
1176 	unsigned int scan, floor;
1177 	unsigned int windows = 1;
1178 
1179 	if (scan_size < MAX_SCAN_WINDOW)
1180 		windows = MAX_SCAN_WINDOW / scan_size;
1181 	floor = 1000 / windows;
1182 
1183 	scan = sysctl_numa_balancing_scan_period_min / task_nr_scan_windows(p);
1184 	return max_t(unsigned int, floor, scan);
1185 }
1186 
1187 static unsigned int task_scan_start(struct task_struct *p)
1188 {
1189 	unsigned long smin = task_scan_min(p);
1190 	unsigned long period = smin;
1191 	struct numa_group *ng;
1192 
1193 	/* Scale the maximum scan period with the amount of shared memory. */
1194 	rcu_read_lock();
1195 	ng = rcu_dereference(p->numa_group);
1196 	if (ng) {
1197 		unsigned long shared = group_faults_shared(ng);
1198 		unsigned long private = group_faults_priv(ng);
1199 
1200 		period *= refcount_read(&ng->refcount);
1201 		period *= shared + 1;
1202 		period /= private + shared + 1;
1203 	}
1204 	rcu_read_unlock();
1205 
1206 	return max(smin, period);
1207 }
1208 
1209 static unsigned int task_scan_max(struct task_struct *p)
1210 {
1211 	unsigned long smin = task_scan_min(p);
1212 	unsigned long smax;
1213 	struct numa_group *ng;
1214 
1215 	/* Watch for min being lower than max due to floor calculations */
1216 	smax = sysctl_numa_balancing_scan_period_max / task_nr_scan_windows(p);
1217 
1218 	/* Scale the maximum scan period with the amount of shared memory. */
1219 	ng = deref_curr_numa_group(p);
1220 	if (ng) {
1221 		unsigned long shared = group_faults_shared(ng);
1222 		unsigned long private = group_faults_priv(ng);
1223 		unsigned long period = smax;
1224 
1225 		period *= refcount_read(&ng->refcount);
1226 		period *= shared + 1;
1227 		period /= private + shared + 1;
1228 
1229 		smax = max(smax, period);
1230 	}
1231 
1232 	return max(smin, smax);
1233 }
1234 
1235 static void account_numa_enqueue(struct rq *rq, struct task_struct *p)
1236 {
1237 	rq->nr_numa_running += (p->numa_preferred_nid != NUMA_NO_NODE);
1238 	rq->nr_preferred_running += (p->numa_preferred_nid == task_node(p));
1239 }
1240 
1241 static void account_numa_dequeue(struct rq *rq, struct task_struct *p)
1242 {
1243 	rq->nr_numa_running -= (p->numa_preferred_nid != NUMA_NO_NODE);
1244 	rq->nr_preferred_running -= (p->numa_preferred_nid == task_node(p));
1245 }
1246 
1247 /* Shared or private faults. */
1248 #define NR_NUMA_HINT_FAULT_TYPES 2
1249 
1250 /* Memory and CPU locality */
1251 #define NR_NUMA_HINT_FAULT_STATS (NR_NUMA_HINT_FAULT_TYPES * 2)
1252 
1253 /* Averaged statistics, and temporary buffers. */
1254 #define NR_NUMA_HINT_FAULT_BUCKETS (NR_NUMA_HINT_FAULT_STATS * 2)
1255 
1256 pid_t task_numa_group_id(struct task_struct *p)
1257 {
1258 	struct numa_group *ng;
1259 	pid_t gid = 0;
1260 
1261 	rcu_read_lock();
1262 	ng = rcu_dereference(p->numa_group);
1263 	if (ng)
1264 		gid = ng->gid;
1265 	rcu_read_unlock();
1266 
1267 	return gid;
1268 }
1269 
1270 /*
1271  * The averaged statistics, shared & private, memory & CPU,
1272  * occupy the first half of the array. The second half of the
1273  * array is for current counters, which are averaged into the
1274  * first set by task_numa_placement.
1275  */
1276 static inline int task_faults_idx(enum numa_faults_stats s, int nid, int priv)
1277 {
1278 	return NR_NUMA_HINT_FAULT_TYPES * (s * nr_node_ids + nid) + priv;
1279 }
1280 
1281 static inline unsigned long task_faults(struct task_struct *p, int nid)
1282 {
1283 	if (!p->numa_faults)
1284 		return 0;
1285 
1286 	return p->numa_faults[task_faults_idx(NUMA_MEM, nid, 0)] +
1287 		p->numa_faults[task_faults_idx(NUMA_MEM, nid, 1)];
1288 }
1289 
1290 static inline unsigned long group_faults(struct task_struct *p, int nid)
1291 {
1292 	struct numa_group *ng = deref_task_numa_group(p);
1293 
1294 	if (!ng)
1295 		return 0;
1296 
1297 	return ng->faults[task_faults_idx(NUMA_MEM, nid, 0)] +
1298 		ng->faults[task_faults_idx(NUMA_MEM, nid, 1)];
1299 }
1300 
1301 static inline unsigned long group_faults_cpu(struct numa_group *group, int nid)
1302 {
1303 	return group->faults[task_faults_idx(NUMA_CPU, nid, 0)] +
1304 		group->faults[task_faults_idx(NUMA_CPU, nid, 1)];
1305 }
1306 
1307 static inline unsigned long group_faults_priv(struct numa_group *ng)
1308 {
1309 	unsigned long faults = 0;
1310 	int node;
1311 
1312 	for_each_online_node(node) {
1313 		faults += ng->faults[task_faults_idx(NUMA_MEM, node, 1)];
1314 	}
1315 
1316 	return faults;
1317 }
1318 
1319 static inline unsigned long group_faults_shared(struct numa_group *ng)
1320 {
1321 	unsigned long faults = 0;
1322 	int node;
1323 
1324 	for_each_online_node(node) {
1325 		faults += ng->faults[task_faults_idx(NUMA_MEM, node, 0)];
1326 	}
1327 
1328 	return faults;
1329 }
1330 
1331 /*
1332  * A node triggering more than 1/3 as many NUMA faults as the maximum is
1333  * considered part of a numa group's pseudo-interleaving set. Migrations
1334  * between these nodes are slowed down, to allow things to settle down.
1335  */
1336 #define ACTIVE_NODE_FRACTION 3
1337 
1338 static bool numa_is_active_node(int nid, struct numa_group *ng)
1339 {
1340 	return group_faults_cpu(ng, nid) * ACTIVE_NODE_FRACTION > ng->max_faults_cpu;
1341 }
1342 
1343 /* Handle placement on systems where not all nodes are directly connected. */
1344 static unsigned long score_nearby_nodes(struct task_struct *p, int nid,
1345 					int lim_dist, bool task)
1346 {
1347 	unsigned long score = 0;
1348 	int node, max_dist;
1349 
1350 	/*
1351 	 * All nodes are directly connected, and the same distance
1352 	 * from each other. No need for fancy placement algorithms.
1353 	 */
1354 	if (sched_numa_topology_type == NUMA_DIRECT)
1355 		return 0;
1356 
1357 	/* sched_max_numa_distance may be changed in parallel. */
1358 	max_dist = READ_ONCE(sched_max_numa_distance);
1359 	/*
1360 	 * This code is called for each node, introducing N^2 complexity,
1361 	 * which should be ok given the number of nodes rarely exceeds 8.
1362 	 */
1363 	for_each_online_node(node) {
1364 		unsigned long faults;
1365 		int dist = node_distance(nid, node);
1366 
1367 		/*
1368 		 * The furthest away nodes in the system are not interesting
1369 		 * for placement; nid was already counted.
1370 		 */
1371 		if (dist >= max_dist || node == nid)
1372 			continue;
1373 
1374 		/*
1375 		 * On systems with a backplane NUMA topology, compare groups
1376 		 * of nodes, and move tasks towards the group with the most
1377 		 * memory accesses. When comparing two nodes at distance
1378 		 * "hoplimit", only nodes closer by than "hoplimit" are part
1379 		 * of each group. Skip other nodes.
1380 		 */
1381 		if (sched_numa_topology_type == NUMA_BACKPLANE && dist >= lim_dist)
1382 			continue;
1383 
1384 		/* Add up the faults from nearby nodes. */
1385 		if (task)
1386 			faults = task_faults(p, node);
1387 		else
1388 			faults = group_faults(p, node);
1389 
1390 		/*
1391 		 * On systems with a glueless mesh NUMA topology, there are
1392 		 * no fixed "groups of nodes". Instead, nodes that are not
1393 		 * directly connected bounce traffic through intermediate
1394 		 * nodes; a numa_group can occupy any set of nodes.
1395 		 * The further away a node is, the less the faults count.
1396 		 * This seems to result in good task placement.
1397 		 */
1398 		if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
1399 			faults *= (max_dist - dist);
1400 			faults /= (max_dist - LOCAL_DISTANCE);
1401 		}
1402 
1403 		score += faults;
1404 	}
1405 
1406 	return score;
1407 }
1408 
1409 /*
1410  * These return the fraction of accesses done by a particular task, or
1411  * task group, on a particular numa node.  The group weight is given a
1412  * larger multiplier, in order to group tasks together that are almost
1413  * evenly spread out between numa nodes.
1414  */
1415 static inline unsigned long task_weight(struct task_struct *p, int nid,
1416 					int dist)
1417 {
1418 	unsigned long faults, total_faults;
1419 
1420 	if (!p->numa_faults)
1421 		return 0;
1422 
1423 	total_faults = p->total_numa_faults;
1424 
1425 	if (!total_faults)
1426 		return 0;
1427 
1428 	faults = task_faults(p, nid);
1429 	faults += score_nearby_nodes(p, nid, dist, true);
1430 
1431 	return 1000 * faults / total_faults;
1432 }
1433 
1434 static inline unsigned long group_weight(struct task_struct *p, int nid,
1435 					 int dist)
1436 {
1437 	struct numa_group *ng = deref_task_numa_group(p);
1438 	unsigned long faults, total_faults;
1439 
1440 	if (!ng)
1441 		return 0;
1442 
1443 	total_faults = ng->total_faults;
1444 
1445 	if (!total_faults)
1446 		return 0;
1447 
1448 	faults = group_faults(p, nid);
1449 	faults += score_nearby_nodes(p, nid, dist, false);
1450 
1451 	return 1000 * faults / total_faults;
1452 }
1453 
1454 /*
1455  * If memory tiering mode is enabled, cpupid of slow memory page is
1456  * used to record scan time instead of CPU and PID.  When tiering mode
1457  * is disabled at run time, the scan time (in cpupid) will be
1458  * interpreted as CPU and PID.  So CPU needs to be checked to avoid to
1459  * access out of array bound.
1460  */
1461 static inline bool cpupid_valid(int cpupid)
1462 {
1463 	return cpupid_to_cpu(cpupid) < nr_cpu_ids;
1464 }
1465 
1466 /*
1467  * For memory tiering mode, if there are enough free pages (more than
1468  * enough watermark defined here) in fast memory node, to take full
1469  * advantage of fast memory capacity, all recently accessed slow
1470  * memory pages will be migrated to fast memory node without
1471  * considering hot threshold.
1472  */
1473 static bool pgdat_free_space_enough(struct pglist_data *pgdat)
1474 {
1475 	int z;
1476 	unsigned long enough_wmark;
1477 
1478 	enough_wmark = max(1UL * 1024 * 1024 * 1024 >> PAGE_SHIFT,
1479 			   pgdat->node_present_pages >> 4);
1480 	for (z = pgdat->nr_zones - 1; z >= 0; z--) {
1481 		struct zone *zone = pgdat->node_zones + z;
1482 
1483 		if (!populated_zone(zone))
1484 			continue;
1485 
1486 		if (zone_watermark_ok(zone, 0,
1487 				      wmark_pages(zone, WMARK_PROMO) + enough_wmark,
1488 				      ZONE_MOVABLE, 0))
1489 			return true;
1490 	}
1491 	return false;
1492 }
1493 
1494 /*
1495  * For memory tiering mode, when page tables are scanned, the scan
1496  * time will be recorded in struct page in addition to make page
1497  * PROT_NONE for slow memory page.  So when the page is accessed, in
1498  * hint page fault handler, the hint page fault latency is calculated
1499  * via,
1500  *
1501  *	hint page fault latency = hint page fault time - scan time
1502  *
1503  * The smaller the hint page fault latency, the higher the possibility
1504  * for the page to be hot.
1505  */
1506 static int numa_hint_fault_latency(struct page *page)
1507 {
1508 	int last_time, time;
1509 
1510 	time = jiffies_to_msecs(jiffies);
1511 	last_time = xchg_page_access_time(page, time);
1512 
1513 	return (time - last_time) & PAGE_ACCESS_TIME_MASK;
1514 }
1515 
1516 /*
1517  * For memory tiering mode, too high promotion/demotion throughput may
1518  * hurt application latency.  So we provide a mechanism to rate limit
1519  * the number of pages that are tried to be promoted.
1520  */
1521 static bool numa_promotion_rate_limit(struct pglist_data *pgdat,
1522 				      unsigned long rate_limit, int nr)
1523 {
1524 	unsigned long nr_cand;
1525 	unsigned int now, start;
1526 
1527 	now = jiffies_to_msecs(jiffies);
1528 	mod_node_page_state(pgdat, PGPROMOTE_CANDIDATE, nr);
1529 	nr_cand = node_page_state(pgdat, PGPROMOTE_CANDIDATE);
1530 	start = pgdat->nbp_rl_start;
1531 	if (now - start > MSEC_PER_SEC &&
1532 	    cmpxchg(&pgdat->nbp_rl_start, start, now) == start)
1533 		pgdat->nbp_rl_nr_cand = nr_cand;
1534 	if (nr_cand - pgdat->nbp_rl_nr_cand >= rate_limit)
1535 		return true;
1536 	return false;
1537 }
1538 
1539 #define NUMA_MIGRATION_ADJUST_STEPS	16
1540 
1541 static void numa_promotion_adjust_threshold(struct pglist_data *pgdat,
1542 					    unsigned long rate_limit,
1543 					    unsigned int ref_th)
1544 {
1545 	unsigned int now, start, th_period, unit_th, th;
1546 	unsigned long nr_cand, ref_cand, diff_cand;
1547 
1548 	now = jiffies_to_msecs(jiffies);
1549 	th_period = sysctl_numa_balancing_scan_period_max;
1550 	start = pgdat->nbp_th_start;
1551 	if (now - start > th_period &&
1552 	    cmpxchg(&pgdat->nbp_th_start, start, now) == start) {
1553 		ref_cand = rate_limit *
1554 			sysctl_numa_balancing_scan_period_max / MSEC_PER_SEC;
1555 		nr_cand = node_page_state(pgdat, PGPROMOTE_CANDIDATE);
1556 		diff_cand = nr_cand - pgdat->nbp_th_nr_cand;
1557 		unit_th = ref_th * 2 / NUMA_MIGRATION_ADJUST_STEPS;
1558 		th = pgdat->nbp_threshold ? : ref_th;
1559 		if (diff_cand > ref_cand * 11 / 10)
1560 			th = max(th - unit_th, unit_th);
1561 		else if (diff_cand < ref_cand * 9 / 10)
1562 			th = min(th + unit_th, ref_th * 2);
1563 		pgdat->nbp_th_nr_cand = nr_cand;
1564 		pgdat->nbp_threshold = th;
1565 	}
1566 }
1567 
1568 bool should_numa_migrate_memory(struct task_struct *p, struct page * page,
1569 				int src_nid, int dst_cpu)
1570 {
1571 	struct numa_group *ng = deref_curr_numa_group(p);
1572 	int dst_nid = cpu_to_node(dst_cpu);
1573 	int last_cpupid, this_cpupid;
1574 
1575 	/*
1576 	 * The pages in slow memory node should be migrated according
1577 	 * to hot/cold instead of private/shared.
1578 	 */
1579 	if (sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING &&
1580 	    !node_is_toptier(src_nid)) {
1581 		struct pglist_data *pgdat;
1582 		unsigned long rate_limit;
1583 		unsigned int latency, th, def_th;
1584 
1585 		pgdat = NODE_DATA(dst_nid);
1586 		if (pgdat_free_space_enough(pgdat)) {
1587 			/* workload changed, reset hot threshold */
1588 			pgdat->nbp_threshold = 0;
1589 			return true;
1590 		}
1591 
1592 		def_th = sysctl_numa_balancing_hot_threshold;
1593 		rate_limit = sysctl_numa_balancing_promote_rate_limit << \
1594 			(20 - PAGE_SHIFT);
1595 		numa_promotion_adjust_threshold(pgdat, rate_limit, def_th);
1596 
1597 		th = pgdat->nbp_threshold ? : def_th;
1598 		latency = numa_hint_fault_latency(page);
1599 		if (latency >= th)
1600 			return false;
1601 
1602 		return !numa_promotion_rate_limit(pgdat, rate_limit,
1603 						  thp_nr_pages(page));
1604 	}
1605 
1606 	this_cpupid = cpu_pid_to_cpupid(dst_cpu, current->pid);
1607 	last_cpupid = page_cpupid_xchg_last(page, this_cpupid);
1608 
1609 	if (!(sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING) &&
1610 	    !node_is_toptier(src_nid) && !cpupid_valid(last_cpupid))
1611 		return false;
1612 
1613 	/*
1614 	 * Allow first faults or private faults to migrate immediately early in
1615 	 * the lifetime of a task. The magic number 4 is based on waiting for
1616 	 * two full passes of the "multi-stage node selection" test that is
1617 	 * executed below.
1618 	 */
1619 	if ((p->numa_preferred_nid == NUMA_NO_NODE || p->numa_scan_seq <= 4) &&
1620 	    (cpupid_pid_unset(last_cpupid) || cpupid_match_pid(p, last_cpupid)))
1621 		return true;
1622 
1623 	/*
1624 	 * Multi-stage node selection is used in conjunction with a periodic
1625 	 * migration fault to build a temporal task<->page relation. By using
1626 	 * a two-stage filter we remove short/unlikely relations.
1627 	 *
1628 	 * Using P(p) ~ n_p / n_t as per frequentist probability, we can equate
1629 	 * a task's usage of a particular page (n_p) per total usage of this
1630 	 * page (n_t) (in a given time-span) to a probability.
1631 	 *
1632 	 * Our periodic faults will sample this probability and getting the
1633 	 * same result twice in a row, given these samples are fully
1634 	 * independent, is then given by P(n)^2, provided our sample period
1635 	 * is sufficiently short compared to the usage pattern.
1636 	 *
1637 	 * This quadric squishes small probabilities, making it less likely we
1638 	 * act on an unlikely task<->page relation.
1639 	 */
1640 	if (!cpupid_pid_unset(last_cpupid) &&
1641 				cpupid_to_nid(last_cpupid) != dst_nid)
1642 		return false;
1643 
1644 	/* Always allow migrate on private faults */
1645 	if (cpupid_match_pid(p, last_cpupid))
1646 		return true;
1647 
1648 	/* A shared fault, but p->numa_group has not been set up yet. */
1649 	if (!ng)
1650 		return true;
1651 
1652 	/*
1653 	 * Destination node is much more heavily used than the source
1654 	 * node? Allow migration.
1655 	 */
1656 	if (group_faults_cpu(ng, dst_nid) > group_faults_cpu(ng, src_nid) *
1657 					ACTIVE_NODE_FRACTION)
1658 		return true;
1659 
1660 	/*
1661 	 * Distribute memory according to CPU & memory use on each node,
1662 	 * with 3/4 hysteresis to avoid unnecessary memory migrations:
1663 	 *
1664 	 * faults_cpu(dst)   3   faults_cpu(src)
1665 	 * --------------- * - > ---------------
1666 	 * faults_mem(dst)   4   faults_mem(src)
1667 	 */
1668 	return group_faults_cpu(ng, dst_nid) * group_faults(p, src_nid) * 3 >
1669 	       group_faults_cpu(ng, src_nid) * group_faults(p, dst_nid) * 4;
1670 }
1671 
1672 /*
1673  * 'numa_type' describes the node at the moment of load balancing.
1674  */
1675 enum numa_type {
1676 	/* The node has spare capacity that can be used to run more tasks.  */
1677 	node_has_spare = 0,
1678 	/*
1679 	 * The node is fully used and the tasks don't compete for more CPU
1680 	 * cycles. Nevertheless, some tasks might wait before running.
1681 	 */
1682 	node_fully_busy,
1683 	/*
1684 	 * The node is overloaded and can't provide expected CPU cycles to all
1685 	 * tasks.
1686 	 */
1687 	node_overloaded
1688 };
1689 
1690 /* Cached statistics for all CPUs within a node */
1691 struct numa_stats {
1692 	unsigned long load;
1693 	unsigned long runnable;
1694 	unsigned long util;
1695 	/* Total compute capacity of CPUs on a node */
1696 	unsigned long compute_capacity;
1697 	unsigned int nr_running;
1698 	unsigned int weight;
1699 	enum numa_type node_type;
1700 	int idle_cpu;
1701 };
1702 
1703 static inline bool is_core_idle(int cpu)
1704 {
1705 #ifdef CONFIG_SCHED_SMT
1706 	int sibling;
1707 
1708 	for_each_cpu(sibling, cpu_smt_mask(cpu)) {
1709 		if (cpu == sibling)
1710 			continue;
1711 
1712 		if (!idle_cpu(sibling))
1713 			return false;
1714 	}
1715 #endif
1716 
1717 	return true;
1718 }
1719 
1720 struct task_numa_env {
1721 	struct task_struct *p;
1722 
1723 	int src_cpu, src_nid;
1724 	int dst_cpu, dst_nid;
1725 	int imb_numa_nr;
1726 
1727 	struct numa_stats src_stats, dst_stats;
1728 
1729 	int imbalance_pct;
1730 	int dist;
1731 
1732 	struct task_struct *best_task;
1733 	long best_imp;
1734 	int best_cpu;
1735 };
1736 
1737 static unsigned long cpu_load(struct rq *rq);
1738 static unsigned long cpu_runnable(struct rq *rq);
1739 
1740 static inline enum
1741 numa_type numa_classify(unsigned int imbalance_pct,
1742 			 struct numa_stats *ns)
1743 {
1744 	if ((ns->nr_running > ns->weight) &&
1745 	    (((ns->compute_capacity * 100) < (ns->util * imbalance_pct)) ||
1746 	     ((ns->compute_capacity * imbalance_pct) < (ns->runnable * 100))))
1747 		return node_overloaded;
1748 
1749 	if ((ns->nr_running < ns->weight) ||
1750 	    (((ns->compute_capacity * 100) > (ns->util * imbalance_pct)) &&
1751 	     ((ns->compute_capacity * imbalance_pct) > (ns->runnable * 100))))
1752 		return node_has_spare;
1753 
1754 	return node_fully_busy;
1755 }
1756 
1757 #ifdef CONFIG_SCHED_SMT
1758 /* Forward declarations of select_idle_sibling helpers */
1759 static inline bool test_idle_cores(int cpu);
1760 static inline int numa_idle_core(int idle_core, int cpu)
1761 {
1762 	if (!static_branch_likely(&sched_smt_present) ||
1763 	    idle_core >= 0 || !test_idle_cores(cpu))
1764 		return idle_core;
1765 
1766 	/*
1767 	 * Prefer cores instead of packing HT siblings
1768 	 * and triggering future load balancing.
1769 	 */
1770 	if (is_core_idle(cpu))
1771 		idle_core = cpu;
1772 
1773 	return idle_core;
1774 }
1775 #else
1776 static inline int numa_idle_core(int idle_core, int cpu)
1777 {
1778 	return idle_core;
1779 }
1780 #endif
1781 
1782 /*
1783  * Gather all necessary information to make NUMA balancing placement
1784  * decisions that are compatible with standard load balancer. This
1785  * borrows code and logic from update_sg_lb_stats but sharing a
1786  * common implementation is impractical.
1787  */
1788 static void update_numa_stats(struct task_numa_env *env,
1789 			      struct numa_stats *ns, int nid,
1790 			      bool find_idle)
1791 {
1792 	int cpu, idle_core = -1;
1793 
1794 	memset(ns, 0, sizeof(*ns));
1795 	ns->idle_cpu = -1;
1796 
1797 	rcu_read_lock();
1798 	for_each_cpu(cpu, cpumask_of_node(nid)) {
1799 		struct rq *rq = cpu_rq(cpu);
1800 
1801 		ns->load += cpu_load(rq);
1802 		ns->runnable += cpu_runnable(rq);
1803 		ns->util += cpu_util_cfs(cpu);
1804 		ns->nr_running += rq->cfs.h_nr_running;
1805 		ns->compute_capacity += capacity_of(cpu);
1806 
1807 		if (find_idle && idle_core < 0 && !rq->nr_running && idle_cpu(cpu)) {
1808 			if (READ_ONCE(rq->numa_migrate_on) ||
1809 			    !cpumask_test_cpu(cpu, env->p->cpus_ptr))
1810 				continue;
1811 
1812 			if (ns->idle_cpu == -1)
1813 				ns->idle_cpu = cpu;
1814 
1815 			idle_core = numa_idle_core(idle_core, cpu);
1816 		}
1817 	}
1818 	rcu_read_unlock();
1819 
1820 	ns->weight = cpumask_weight(cpumask_of_node(nid));
1821 
1822 	ns->node_type = numa_classify(env->imbalance_pct, ns);
1823 
1824 	if (idle_core >= 0)
1825 		ns->idle_cpu = idle_core;
1826 }
1827 
1828 static void task_numa_assign(struct task_numa_env *env,
1829 			     struct task_struct *p, long imp)
1830 {
1831 	struct rq *rq = cpu_rq(env->dst_cpu);
1832 
1833 	/* Check if run-queue part of active NUMA balance. */
1834 	if (env->best_cpu != env->dst_cpu && xchg(&rq->numa_migrate_on, 1)) {
1835 		int cpu;
1836 		int start = env->dst_cpu;
1837 
1838 		/* Find alternative idle CPU. */
1839 		for_each_cpu_wrap(cpu, cpumask_of_node(env->dst_nid), start + 1) {
1840 			if (cpu == env->best_cpu || !idle_cpu(cpu) ||
1841 			    !cpumask_test_cpu(cpu, env->p->cpus_ptr)) {
1842 				continue;
1843 			}
1844 
1845 			env->dst_cpu = cpu;
1846 			rq = cpu_rq(env->dst_cpu);
1847 			if (!xchg(&rq->numa_migrate_on, 1))
1848 				goto assign;
1849 		}
1850 
1851 		/* Failed to find an alternative idle CPU */
1852 		return;
1853 	}
1854 
1855 assign:
1856 	/*
1857 	 * Clear previous best_cpu/rq numa-migrate flag, since task now
1858 	 * found a better CPU to move/swap.
1859 	 */
1860 	if (env->best_cpu != -1 && env->best_cpu != env->dst_cpu) {
1861 		rq = cpu_rq(env->best_cpu);
1862 		WRITE_ONCE(rq->numa_migrate_on, 0);
1863 	}
1864 
1865 	if (env->best_task)
1866 		put_task_struct(env->best_task);
1867 	if (p)
1868 		get_task_struct(p);
1869 
1870 	env->best_task = p;
1871 	env->best_imp = imp;
1872 	env->best_cpu = env->dst_cpu;
1873 }
1874 
1875 static bool load_too_imbalanced(long src_load, long dst_load,
1876 				struct task_numa_env *env)
1877 {
1878 	long imb, old_imb;
1879 	long orig_src_load, orig_dst_load;
1880 	long src_capacity, dst_capacity;
1881 
1882 	/*
1883 	 * The load is corrected for the CPU capacity available on each node.
1884 	 *
1885 	 * src_load        dst_load
1886 	 * ------------ vs ---------
1887 	 * src_capacity    dst_capacity
1888 	 */
1889 	src_capacity = env->src_stats.compute_capacity;
1890 	dst_capacity = env->dst_stats.compute_capacity;
1891 
1892 	imb = abs(dst_load * src_capacity - src_load * dst_capacity);
1893 
1894 	orig_src_load = env->src_stats.load;
1895 	orig_dst_load = env->dst_stats.load;
1896 
1897 	old_imb = abs(orig_dst_load * src_capacity - orig_src_load * dst_capacity);
1898 
1899 	/* Would this change make things worse? */
1900 	return (imb > old_imb);
1901 }
1902 
1903 /*
1904  * Maximum NUMA importance can be 1998 (2*999);
1905  * SMALLIMP @ 30 would be close to 1998/64.
1906  * Used to deter task migration.
1907  */
1908 #define SMALLIMP	30
1909 
1910 /*
1911  * This checks if the overall compute and NUMA accesses of the system would
1912  * be improved if the source tasks was migrated to the target dst_cpu taking
1913  * into account that it might be best if task running on the dst_cpu should
1914  * be exchanged with the source task
1915  */
1916 static bool task_numa_compare(struct task_numa_env *env,
1917 			      long taskimp, long groupimp, bool maymove)
1918 {
1919 	struct numa_group *cur_ng, *p_ng = deref_curr_numa_group(env->p);
1920 	struct rq *dst_rq = cpu_rq(env->dst_cpu);
1921 	long imp = p_ng ? groupimp : taskimp;
1922 	struct task_struct *cur;
1923 	long src_load, dst_load;
1924 	int dist = env->dist;
1925 	long moveimp = imp;
1926 	long load;
1927 	bool stopsearch = false;
1928 
1929 	if (READ_ONCE(dst_rq->numa_migrate_on))
1930 		return false;
1931 
1932 	rcu_read_lock();
1933 	cur = rcu_dereference(dst_rq->curr);
1934 	if (cur && ((cur->flags & PF_EXITING) || is_idle_task(cur)))
1935 		cur = NULL;
1936 
1937 	/*
1938 	 * Because we have preemption enabled we can get migrated around and
1939 	 * end try selecting ourselves (current == env->p) as a swap candidate.
1940 	 */
1941 	if (cur == env->p) {
1942 		stopsearch = true;
1943 		goto unlock;
1944 	}
1945 
1946 	if (!cur) {
1947 		if (maymove && moveimp >= env->best_imp)
1948 			goto assign;
1949 		else
1950 			goto unlock;
1951 	}
1952 
1953 	/* Skip this swap candidate if cannot move to the source cpu. */
1954 	if (!cpumask_test_cpu(env->src_cpu, cur->cpus_ptr))
1955 		goto unlock;
1956 
1957 	/*
1958 	 * Skip this swap candidate if it is not moving to its preferred
1959 	 * node and the best task is.
1960 	 */
1961 	if (env->best_task &&
1962 	    env->best_task->numa_preferred_nid == env->src_nid &&
1963 	    cur->numa_preferred_nid != env->src_nid) {
1964 		goto unlock;
1965 	}
1966 
1967 	/*
1968 	 * "imp" is the fault differential for the source task between the
1969 	 * source and destination node. Calculate the total differential for
1970 	 * the source task and potential destination task. The more negative
1971 	 * the value is, the more remote accesses that would be expected to
1972 	 * be incurred if the tasks were swapped.
1973 	 *
1974 	 * If dst and source tasks are in the same NUMA group, or not
1975 	 * in any group then look only at task weights.
1976 	 */
1977 	cur_ng = rcu_dereference(cur->numa_group);
1978 	if (cur_ng == p_ng) {
1979 		/*
1980 		 * Do not swap within a group or between tasks that have
1981 		 * no group if there is spare capacity. Swapping does
1982 		 * not address the load imbalance and helps one task at
1983 		 * the cost of punishing another.
1984 		 */
1985 		if (env->dst_stats.node_type == node_has_spare)
1986 			goto unlock;
1987 
1988 		imp = taskimp + task_weight(cur, env->src_nid, dist) -
1989 		      task_weight(cur, env->dst_nid, dist);
1990 		/*
1991 		 * Add some hysteresis to prevent swapping the
1992 		 * tasks within a group over tiny differences.
1993 		 */
1994 		if (cur_ng)
1995 			imp -= imp / 16;
1996 	} else {
1997 		/*
1998 		 * Compare the group weights. If a task is all by itself
1999 		 * (not part of a group), use the task weight instead.
2000 		 */
2001 		if (cur_ng && p_ng)
2002 			imp += group_weight(cur, env->src_nid, dist) -
2003 			       group_weight(cur, env->dst_nid, dist);
2004 		else
2005 			imp += task_weight(cur, env->src_nid, dist) -
2006 			       task_weight(cur, env->dst_nid, dist);
2007 	}
2008 
2009 	/* Discourage picking a task already on its preferred node */
2010 	if (cur->numa_preferred_nid == env->dst_nid)
2011 		imp -= imp / 16;
2012 
2013 	/*
2014 	 * Encourage picking a task that moves to its preferred node.
2015 	 * This potentially makes imp larger than it's maximum of
2016 	 * 1998 (see SMALLIMP and task_weight for why) but in this
2017 	 * case, it does not matter.
2018 	 */
2019 	if (cur->numa_preferred_nid == env->src_nid)
2020 		imp += imp / 8;
2021 
2022 	if (maymove && moveimp > imp && moveimp > env->best_imp) {
2023 		imp = moveimp;
2024 		cur = NULL;
2025 		goto assign;
2026 	}
2027 
2028 	/*
2029 	 * Prefer swapping with a task moving to its preferred node over a
2030 	 * task that is not.
2031 	 */
2032 	if (env->best_task && cur->numa_preferred_nid == env->src_nid &&
2033 	    env->best_task->numa_preferred_nid != env->src_nid) {
2034 		goto assign;
2035 	}
2036 
2037 	/*
2038 	 * If the NUMA importance is less than SMALLIMP,
2039 	 * task migration might only result in ping pong
2040 	 * of tasks and also hurt performance due to cache
2041 	 * misses.
2042 	 */
2043 	if (imp < SMALLIMP || imp <= env->best_imp + SMALLIMP / 2)
2044 		goto unlock;
2045 
2046 	/*
2047 	 * In the overloaded case, try and keep the load balanced.
2048 	 */
2049 	load = task_h_load(env->p) - task_h_load(cur);
2050 	if (!load)
2051 		goto assign;
2052 
2053 	dst_load = env->dst_stats.load + load;
2054 	src_load = env->src_stats.load - load;
2055 
2056 	if (load_too_imbalanced(src_load, dst_load, env))
2057 		goto unlock;
2058 
2059 assign:
2060 	/* Evaluate an idle CPU for a task numa move. */
2061 	if (!cur) {
2062 		int cpu = env->dst_stats.idle_cpu;
2063 
2064 		/* Nothing cached so current CPU went idle since the search. */
2065 		if (cpu < 0)
2066 			cpu = env->dst_cpu;
2067 
2068 		/*
2069 		 * If the CPU is no longer truly idle and the previous best CPU
2070 		 * is, keep using it.
2071 		 */
2072 		if (!idle_cpu(cpu) && env->best_cpu >= 0 &&
2073 		    idle_cpu(env->best_cpu)) {
2074 			cpu = env->best_cpu;
2075 		}
2076 
2077 		env->dst_cpu = cpu;
2078 	}
2079 
2080 	task_numa_assign(env, cur, imp);
2081 
2082 	/*
2083 	 * If a move to idle is allowed because there is capacity or load
2084 	 * balance improves then stop the search. While a better swap
2085 	 * candidate may exist, a search is not free.
2086 	 */
2087 	if (maymove && !cur && env->best_cpu >= 0 && idle_cpu(env->best_cpu))
2088 		stopsearch = true;
2089 
2090 	/*
2091 	 * If a swap candidate must be identified and the current best task
2092 	 * moves its preferred node then stop the search.
2093 	 */
2094 	if (!maymove && env->best_task &&
2095 	    env->best_task->numa_preferred_nid == env->src_nid) {
2096 		stopsearch = true;
2097 	}
2098 unlock:
2099 	rcu_read_unlock();
2100 
2101 	return stopsearch;
2102 }
2103 
2104 static void task_numa_find_cpu(struct task_numa_env *env,
2105 				long taskimp, long groupimp)
2106 {
2107 	bool maymove = false;
2108 	int cpu;
2109 
2110 	/*
2111 	 * If dst node has spare capacity, then check if there is an
2112 	 * imbalance that would be overruled by the load balancer.
2113 	 */
2114 	if (env->dst_stats.node_type == node_has_spare) {
2115 		unsigned int imbalance;
2116 		int src_running, dst_running;
2117 
2118 		/*
2119 		 * Would movement cause an imbalance? Note that if src has
2120 		 * more running tasks that the imbalance is ignored as the
2121 		 * move improves the imbalance from the perspective of the
2122 		 * CPU load balancer.
2123 		 * */
2124 		src_running = env->src_stats.nr_running - 1;
2125 		dst_running = env->dst_stats.nr_running + 1;
2126 		imbalance = max(0, dst_running - src_running);
2127 		imbalance = adjust_numa_imbalance(imbalance, dst_running,
2128 						  env->imb_numa_nr);
2129 
2130 		/* Use idle CPU if there is no imbalance */
2131 		if (!imbalance) {
2132 			maymove = true;
2133 			if (env->dst_stats.idle_cpu >= 0) {
2134 				env->dst_cpu = env->dst_stats.idle_cpu;
2135 				task_numa_assign(env, NULL, 0);
2136 				return;
2137 			}
2138 		}
2139 	} else {
2140 		long src_load, dst_load, load;
2141 		/*
2142 		 * If the improvement from just moving env->p direction is better
2143 		 * than swapping tasks around, check if a move is possible.
2144 		 */
2145 		load = task_h_load(env->p);
2146 		dst_load = env->dst_stats.load + load;
2147 		src_load = env->src_stats.load - load;
2148 		maymove = !load_too_imbalanced(src_load, dst_load, env);
2149 	}
2150 
2151 	for_each_cpu(cpu, cpumask_of_node(env->dst_nid)) {
2152 		/* Skip this CPU if the source task cannot migrate */
2153 		if (!cpumask_test_cpu(cpu, env->p->cpus_ptr))
2154 			continue;
2155 
2156 		env->dst_cpu = cpu;
2157 		if (task_numa_compare(env, taskimp, groupimp, maymove))
2158 			break;
2159 	}
2160 }
2161 
2162 static int task_numa_migrate(struct task_struct *p)
2163 {
2164 	struct task_numa_env env = {
2165 		.p = p,
2166 
2167 		.src_cpu = task_cpu(p),
2168 		.src_nid = task_node(p),
2169 
2170 		.imbalance_pct = 112,
2171 
2172 		.best_task = NULL,
2173 		.best_imp = 0,
2174 		.best_cpu = -1,
2175 	};
2176 	unsigned long taskweight, groupweight;
2177 	struct sched_domain *sd;
2178 	long taskimp, groupimp;
2179 	struct numa_group *ng;
2180 	struct rq *best_rq;
2181 	int nid, ret, dist;
2182 
2183 	/*
2184 	 * Pick the lowest SD_NUMA domain, as that would have the smallest
2185 	 * imbalance and would be the first to start moving tasks about.
2186 	 *
2187 	 * And we want to avoid any moving of tasks about, as that would create
2188 	 * random movement of tasks -- counter the numa conditions we're trying
2189 	 * to satisfy here.
2190 	 */
2191 	rcu_read_lock();
2192 	sd = rcu_dereference(per_cpu(sd_numa, env.src_cpu));
2193 	if (sd) {
2194 		env.imbalance_pct = 100 + (sd->imbalance_pct - 100) / 2;
2195 		env.imb_numa_nr = sd->imb_numa_nr;
2196 	}
2197 	rcu_read_unlock();
2198 
2199 	/*
2200 	 * Cpusets can break the scheduler domain tree into smaller
2201 	 * balance domains, some of which do not cross NUMA boundaries.
2202 	 * Tasks that are "trapped" in such domains cannot be migrated
2203 	 * elsewhere, so there is no point in (re)trying.
2204 	 */
2205 	if (unlikely(!sd)) {
2206 		sched_setnuma(p, task_node(p));
2207 		return -EINVAL;
2208 	}
2209 
2210 	env.dst_nid = p->numa_preferred_nid;
2211 	dist = env.dist = node_distance(env.src_nid, env.dst_nid);
2212 	taskweight = task_weight(p, env.src_nid, dist);
2213 	groupweight = group_weight(p, env.src_nid, dist);
2214 	update_numa_stats(&env, &env.src_stats, env.src_nid, false);
2215 	taskimp = task_weight(p, env.dst_nid, dist) - taskweight;
2216 	groupimp = group_weight(p, env.dst_nid, dist) - groupweight;
2217 	update_numa_stats(&env, &env.dst_stats, env.dst_nid, true);
2218 
2219 	/* Try to find a spot on the preferred nid. */
2220 	task_numa_find_cpu(&env, taskimp, groupimp);
2221 
2222 	/*
2223 	 * Look at other nodes in these cases:
2224 	 * - there is no space available on the preferred_nid
2225 	 * - the task is part of a numa_group that is interleaved across
2226 	 *   multiple NUMA nodes; in order to better consolidate the group,
2227 	 *   we need to check other locations.
2228 	 */
2229 	ng = deref_curr_numa_group(p);
2230 	if (env.best_cpu == -1 || (ng && ng->active_nodes > 1)) {
2231 		for_each_node_state(nid, N_CPU) {
2232 			if (nid == env.src_nid || nid == p->numa_preferred_nid)
2233 				continue;
2234 
2235 			dist = node_distance(env.src_nid, env.dst_nid);
2236 			if (sched_numa_topology_type == NUMA_BACKPLANE &&
2237 						dist != env.dist) {
2238 				taskweight = task_weight(p, env.src_nid, dist);
2239 				groupweight = group_weight(p, env.src_nid, dist);
2240 			}
2241 
2242 			/* Only consider nodes where both task and groups benefit */
2243 			taskimp = task_weight(p, nid, dist) - taskweight;
2244 			groupimp = group_weight(p, nid, dist) - groupweight;
2245 			if (taskimp < 0 && groupimp < 0)
2246 				continue;
2247 
2248 			env.dist = dist;
2249 			env.dst_nid = nid;
2250 			update_numa_stats(&env, &env.dst_stats, env.dst_nid, true);
2251 			task_numa_find_cpu(&env, taskimp, groupimp);
2252 		}
2253 	}
2254 
2255 	/*
2256 	 * If the task is part of a workload that spans multiple NUMA nodes,
2257 	 * and is migrating into one of the workload's active nodes, remember
2258 	 * this node as the task's preferred numa node, so the workload can
2259 	 * settle down.
2260 	 * A task that migrated to a second choice node will be better off
2261 	 * trying for a better one later. Do not set the preferred node here.
2262 	 */
2263 	if (ng) {
2264 		if (env.best_cpu == -1)
2265 			nid = env.src_nid;
2266 		else
2267 			nid = cpu_to_node(env.best_cpu);
2268 
2269 		if (nid != p->numa_preferred_nid)
2270 			sched_setnuma(p, nid);
2271 	}
2272 
2273 	/* No better CPU than the current one was found. */
2274 	if (env.best_cpu == -1) {
2275 		trace_sched_stick_numa(p, env.src_cpu, NULL, -1);
2276 		return -EAGAIN;
2277 	}
2278 
2279 	best_rq = cpu_rq(env.best_cpu);
2280 	if (env.best_task == NULL) {
2281 		ret = migrate_task_to(p, env.best_cpu);
2282 		WRITE_ONCE(best_rq->numa_migrate_on, 0);
2283 		if (ret != 0)
2284 			trace_sched_stick_numa(p, env.src_cpu, NULL, env.best_cpu);
2285 		return ret;
2286 	}
2287 
2288 	ret = migrate_swap(p, env.best_task, env.best_cpu, env.src_cpu);
2289 	WRITE_ONCE(best_rq->numa_migrate_on, 0);
2290 
2291 	if (ret != 0)
2292 		trace_sched_stick_numa(p, env.src_cpu, env.best_task, env.best_cpu);
2293 	put_task_struct(env.best_task);
2294 	return ret;
2295 }
2296 
2297 /* Attempt to migrate a task to a CPU on the preferred node. */
2298 static void numa_migrate_preferred(struct task_struct *p)
2299 {
2300 	unsigned long interval = HZ;
2301 
2302 	/* This task has no NUMA fault statistics yet */
2303 	if (unlikely(p->numa_preferred_nid == NUMA_NO_NODE || !p->numa_faults))
2304 		return;
2305 
2306 	/* Periodically retry migrating the task to the preferred node */
2307 	interval = min(interval, msecs_to_jiffies(p->numa_scan_period) / 16);
2308 	p->numa_migrate_retry = jiffies + interval;
2309 
2310 	/* Success if task is already running on preferred CPU */
2311 	if (task_node(p) == p->numa_preferred_nid)
2312 		return;
2313 
2314 	/* Otherwise, try migrate to a CPU on the preferred node */
2315 	task_numa_migrate(p);
2316 }
2317 
2318 /*
2319  * Find out how many nodes the workload is actively running on. Do this by
2320  * tracking the nodes from which NUMA hinting faults are triggered. This can
2321  * be different from the set of nodes where the workload's memory is currently
2322  * located.
2323  */
2324 static void numa_group_count_active_nodes(struct numa_group *numa_group)
2325 {
2326 	unsigned long faults, max_faults = 0;
2327 	int nid, active_nodes = 0;
2328 
2329 	for_each_node_state(nid, N_CPU) {
2330 		faults = group_faults_cpu(numa_group, nid);
2331 		if (faults > max_faults)
2332 			max_faults = faults;
2333 	}
2334 
2335 	for_each_node_state(nid, N_CPU) {
2336 		faults = group_faults_cpu(numa_group, nid);
2337 		if (faults * ACTIVE_NODE_FRACTION > max_faults)
2338 			active_nodes++;
2339 	}
2340 
2341 	numa_group->max_faults_cpu = max_faults;
2342 	numa_group->active_nodes = active_nodes;
2343 }
2344 
2345 /*
2346  * When adapting the scan rate, the period is divided into NUMA_PERIOD_SLOTS
2347  * increments. The more local the fault statistics are, the higher the scan
2348  * period will be for the next scan window. If local/(local+remote) ratio is
2349  * below NUMA_PERIOD_THRESHOLD (where range of ratio is 1..NUMA_PERIOD_SLOTS)
2350  * the scan period will decrease. Aim for 70% local accesses.
2351  */
2352 #define NUMA_PERIOD_SLOTS 10
2353 #define NUMA_PERIOD_THRESHOLD 7
2354 
2355 /*
2356  * Increase the scan period (slow down scanning) if the majority of
2357  * our memory is already on our local node, or if the majority of
2358  * the page accesses are shared with other processes.
2359  * Otherwise, decrease the scan period.
2360  */
2361 static void update_task_scan_period(struct task_struct *p,
2362 			unsigned long shared, unsigned long private)
2363 {
2364 	unsigned int period_slot;
2365 	int lr_ratio, ps_ratio;
2366 	int diff;
2367 
2368 	unsigned long remote = p->numa_faults_locality[0];
2369 	unsigned long local = p->numa_faults_locality[1];
2370 
2371 	/*
2372 	 * If there were no record hinting faults then either the task is
2373 	 * completely idle or all activity is in areas that are not of interest
2374 	 * to automatic numa balancing. Related to that, if there were failed
2375 	 * migration then it implies we are migrating too quickly or the local
2376 	 * node is overloaded. In either case, scan slower
2377 	 */
2378 	if (local + shared == 0 || p->numa_faults_locality[2]) {
2379 		p->numa_scan_period = min(p->numa_scan_period_max,
2380 			p->numa_scan_period << 1);
2381 
2382 		p->mm->numa_next_scan = jiffies +
2383 			msecs_to_jiffies(p->numa_scan_period);
2384 
2385 		return;
2386 	}
2387 
2388 	/*
2389 	 * Prepare to scale scan period relative to the current period.
2390 	 *	 == NUMA_PERIOD_THRESHOLD scan period stays the same
2391 	 *       <  NUMA_PERIOD_THRESHOLD scan period decreases (scan faster)
2392 	 *	 >= NUMA_PERIOD_THRESHOLD scan period increases (scan slower)
2393 	 */
2394 	period_slot = DIV_ROUND_UP(p->numa_scan_period, NUMA_PERIOD_SLOTS);
2395 	lr_ratio = (local * NUMA_PERIOD_SLOTS) / (local + remote);
2396 	ps_ratio = (private * NUMA_PERIOD_SLOTS) / (private + shared);
2397 
2398 	if (ps_ratio >= NUMA_PERIOD_THRESHOLD) {
2399 		/*
2400 		 * Most memory accesses are local. There is no need to
2401 		 * do fast NUMA scanning, since memory is already local.
2402 		 */
2403 		int slot = ps_ratio - NUMA_PERIOD_THRESHOLD;
2404 		if (!slot)
2405 			slot = 1;
2406 		diff = slot * period_slot;
2407 	} else if (lr_ratio >= NUMA_PERIOD_THRESHOLD) {
2408 		/*
2409 		 * Most memory accesses are shared with other tasks.
2410 		 * There is no point in continuing fast NUMA scanning,
2411 		 * since other tasks may just move the memory elsewhere.
2412 		 */
2413 		int slot = lr_ratio - NUMA_PERIOD_THRESHOLD;
2414 		if (!slot)
2415 			slot = 1;
2416 		diff = slot * period_slot;
2417 	} else {
2418 		/*
2419 		 * Private memory faults exceed (SLOTS-THRESHOLD)/SLOTS,
2420 		 * yet they are not on the local NUMA node. Speed up
2421 		 * NUMA scanning to get the memory moved over.
2422 		 */
2423 		int ratio = max(lr_ratio, ps_ratio);
2424 		diff = -(NUMA_PERIOD_THRESHOLD - ratio) * period_slot;
2425 	}
2426 
2427 	p->numa_scan_period = clamp(p->numa_scan_period + diff,
2428 			task_scan_min(p), task_scan_max(p));
2429 	memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
2430 }
2431 
2432 /*
2433  * Get the fraction of time the task has been running since the last
2434  * NUMA placement cycle. The scheduler keeps similar statistics, but
2435  * decays those on a 32ms period, which is orders of magnitude off
2436  * from the dozens-of-seconds NUMA balancing period. Use the scheduler
2437  * stats only if the task is so new there are no NUMA statistics yet.
2438  */
2439 static u64 numa_get_avg_runtime(struct task_struct *p, u64 *period)
2440 {
2441 	u64 runtime, delta, now;
2442 	/* Use the start of this time slice to avoid calculations. */
2443 	now = p->se.exec_start;
2444 	runtime = p->se.sum_exec_runtime;
2445 
2446 	if (p->last_task_numa_placement) {
2447 		delta = runtime - p->last_sum_exec_runtime;
2448 		*period = now - p->last_task_numa_placement;
2449 
2450 		/* Avoid time going backwards, prevent potential divide error: */
2451 		if (unlikely((s64)*period < 0))
2452 			*period = 0;
2453 	} else {
2454 		delta = p->se.avg.load_sum;
2455 		*period = LOAD_AVG_MAX;
2456 	}
2457 
2458 	p->last_sum_exec_runtime = runtime;
2459 	p->last_task_numa_placement = now;
2460 
2461 	return delta;
2462 }
2463 
2464 /*
2465  * Determine the preferred nid for a task in a numa_group. This needs to
2466  * be done in a way that produces consistent results with group_weight,
2467  * otherwise workloads might not converge.
2468  */
2469 static int preferred_group_nid(struct task_struct *p, int nid)
2470 {
2471 	nodemask_t nodes;
2472 	int dist;
2473 
2474 	/* Direct connections between all NUMA nodes. */
2475 	if (sched_numa_topology_type == NUMA_DIRECT)
2476 		return nid;
2477 
2478 	/*
2479 	 * On a system with glueless mesh NUMA topology, group_weight
2480 	 * scores nodes according to the number of NUMA hinting faults on
2481 	 * both the node itself, and on nearby nodes.
2482 	 */
2483 	if (sched_numa_topology_type == NUMA_GLUELESS_MESH) {
2484 		unsigned long score, max_score = 0;
2485 		int node, max_node = nid;
2486 
2487 		dist = sched_max_numa_distance;
2488 
2489 		for_each_node_state(node, N_CPU) {
2490 			score = group_weight(p, node, dist);
2491 			if (score > max_score) {
2492 				max_score = score;
2493 				max_node = node;
2494 			}
2495 		}
2496 		return max_node;
2497 	}
2498 
2499 	/*
2500 	 * Finding the preferred nid in a system with NUMA backplane
2501 	 * interconnect topology is more involved. The goal is to locate
2502 	 * tasks from numa_groups near each other in the system, and
2503 	 * untangle workloads from different sides of the system. This requires
2504 	 * searching down the hierarchy of node groups, recursively searching
2505 	 * inside the highest scoring group of nodes. The nodemask tricks
2506 	 * keep the complexity of the search down.
2507 	 */
2508 	nodes = node_states[N_CPU];
2509 	for (dist = sched_max_numa_distance; dist > LOCAL_DISTANCE; dist--) {
2510 		unsigned long max_faults = 0;
2511 		nodemask_t max_group = NODE_MASK_NONE;
2512 		int a, b;
2513 
2514 		/* Are there nodes at this distance from each other? */
2515 		if (!find_numa_distance(dist))
2516 			continue;
2517 
2518 		for_each_node_mask(a, nodes) {
2519 			unsigned long faults = 0;
2520 			nodemask_t this_group;
2521 			nodes_clear(this_group);
2522 
2523 			/* Sum group's NUMA faults; includes a==b case. */
2524 			for_each_node_mask(b, nodes) {
2525 				if (node_distance(a, b) < dist) {
2526 					faults += group_faults(p, b);
2527 					node_set(b, this_group);
2528 					node_clear(b, nodes);
2529 				}
2530 			}
2531 
2532 			/* Remember the top group. */
2533 			if (faults > max_faults) {
2534 				max_faults = faults;
2535 				max_group = this_group;
2536 				/*
2537 				 * subtle: at the smallest distance there is
2538 				 * just one node left in each "group", the
2539 				 * winner is the preferred nid.
2540 				 */
2541 				nid = a;
2542 			}
2543 		}
2544 		/* Next round, evaluate the nodes within max_group. */
2545 		if (!max_faults)
2546 			break;
2547 		nodes = max_group;
2548 	}
2549 	return nid;
2550 }
2551 
2552 static void task_numa_placement(struct task_struct *p)
2553 {
2554 	int seq, nid, max_nid = NUMA_NO_NODE;
2555 	unsigned long max_faults = 0;
2556 	unsigned long fault_types[2] = { 0, 0 };
2557 	unsigned long total_faults;
2558 	u64 runtime, period;
2559 	spinlock_t *group_lock = NULL;
2560 	struct numa_group *ng;
2561 
2562 	/*
2563 	 * The p->mm->numa_scan_seq field gets updated without
2564 	 * exclusive access. Use READ_ONCE() here to ensure
2565 	 * that the field is read in a single access:
2566 	 */
2567 	seq = READ_ONCE(p->mm->numa_scan_seq);
2568 	if (p->numa_scan_seq == seq)
2569 		return;
2570 	p->numa_scan_seq = seq;
2571 	p->numa_scan_period_max = task_scan_max(p);
2572 
2573 	total_faults = p->numa_faults_locality[0] +
2574 		       p->numa_faults_locality[1];
2575 	runtime = numa_get_avg_runtime(p, &period);
2576 
2577 	/* If the task is part of a group prevent parallel updates to group stats */
2578 	ng = deref_curr_numa_group(p);
2579 	if (ng) {
2580 		group_lock = &ng->lock;
2581 		spin_lock_irq(group_lock);
2582 	}
2583 
2584 	/* Find the node with the highest number of faults */
2585 	for_each_online_node(nid) {
2586 		/* Keep track of the offsets in numa_faults array */
2587 		int mem_idx, membuf_idx, cpu_idx, cpubuf_idx;
2588 		unsigned long faults = 0, group_faults = 0;
2589 		int priv;
2590 
2591 		for (priv = 0; priv < NR_NUMA_HINT_FAULT_TYPES; priv++) {
2592 			long diff, f_diff, f_weight;
2593 
2594 			mem_idx = task_faults_idx(NUMA_MEM, nid, priv);
2595 			membuf_idx = task_faults_idx(NUMA_MEMBUF, nid, priv);
2596 			cpu_idx = task_faults_idx(NUMA_CPU, nid, priv);
2597 			cpubuf_idx = task_faults_idx(NUMA_CPUBUF, nid, priv);
2598 
2599 			/* Decay existing window, copy faults since last scan */
2600 			diff = p->numa_faults[membuf_idx] - p->numa_faults[mem_idx] / 2;
2601 			fault_types[priv] += p->numa_faults[membuf_idx];
2602 			p->numa_faults[membuf_idx] = 0;
2603 
2604 			/*
2605 			 * Normalize the faults_from, so all tasks in a group
2606 			 * count according to CPU use, instead of by the raw
2607 			 * number of faults. Tasks with little runtime have
2608 			 * little over-all impact on throughput, and thus their
2609 			 * faults are less important.
2610 			 */
2611 			f_weight = div64_u64(runtime << 16, period + 1);
2612 			f_weight = (f_weight * p->numa_faults[cpubuf_idx]) /
2613 				   (total_faults + 1);
2614 			f_diff = f_weight - p->numa_faults[cpu_idx] / 2;
2615 			p->numa_faults[cpubuf_idx] = 0;
2616 
2617 			p->numa_faults[mem_idx] += diff;
2618 			p->numa_faults[cpu_idx] += f_diff;
2619 			faults += p->numa_faults[mem_idx];
2620 			p->total_numa_faults += diff;
2621 			if (ng) {
2622 				/*
2623 				 * safe because we can only change our own group
2624 				 *
2625 				 * mem_idx represents the offset for a given
2626 				 * nid and priv in a specific region because it
2627 				 * is at the beginning of the numa_faults array.
2628 				 */
2629 				ng->faults[mem_idx] += diff;
2630 				ng->faults[cpu_idx] += f_diff;
2631 				ng->total_faults += diff;
2632 				group_faults += ng->faults[mem_idx];
2633 			}
2634 		}
2635 
2636 		if (!ng) {
2637 			if (faults > max_faults) {
2638 				max_faults = faults;
2639 				max_nid = nid;
2640 			}
2641 		} else if (group_faults > max_faults) {
2642 			max_faults = group_faults;
2643 			max_nid = nid;
2644 		}
2645 	}
2646 
2647 	/* Cannot migrate task to CPU-less node */
2648 	if (max_nid != NUMA_NO_NODE && !node_state(max_nid, N_CPU)) {
2649 		int near_nid = max_nid;
2650 		int distance, near_distance = INT_MAX;
2651 
2652 		for_each_node_state(nid, N_CPU) {
2653 			distance = node_distance(max_nid, nid);
2654 			if (distance < near_distance) {
2655 				near_nid = nid;
2656 				near_distance = distance;
2657 			}
2658 		}
2659 		max_nid = near_nid;
2660 	}
2661 
2662 	if (ng) {
2663 		numa_group_count_active_nodes(ng);
2664 		spin_unlock_irq(group_lock);
2665 		max_nid = preferred_group_nid(p, max_nid);
2666 	}
2667 
2668 	if (max_faults) {
2669 		/* Set the new preferred node */
2670 		if (max_nid != p->numa_preferred_nid)
2671 			sched_setnuma(p, max_nid);
2672 	}
2673 
2674 	update_task_scan_period(p, fault_types[0], fault_types[1]);
2675 }
2676 
2677 static inline int get_numa_group(struct numa_group *grp)
2678 {
2679 	return refcount_inc_not_zero(&grp->refcount);
2680 }
2681 
2682 static inline void put_numa_group(struct numa_group *grp)
2683 {
2684 	if (refcount_dec_and_test(&grp->refcount))
2685 		kfree_rcu(grp, rcu);
2686 }
2687 
2688 static void task_numa_group(struct task_struct *p, int cpupid, int flags,
2689 			int *priv)
2690 {
2691 	struct numa_group *grp, *my_grp;
2692 	struct task_struct *tsk;
2693 	bool join = false;
2694 	int cpu = cpupid_to_cpu(cpupid);
2695 	int i;
2696 
2697 	if (unlikely(!deref_curr_numa_group(p))) {
2698 		unsigned int size = sizeof(struct numa_group) +
2699 				    NR_NUMA_HINT_FAULT_STATS *
2700 				    nr_node_ids * sizeof(unsigned long);
2701 
2702 		grp = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
2703 		if (!grp)
2704 			return;
2705 
2706 		refcount_set(&grp->refcount, 1);
2707 		grp->active_nodes = 1;
2708 		grp->max_faults_cpu = 0;
2709 		spin_lock_init(&grp->lock);
2710 		grp->gid = p->pid;
2711 
2712 		for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2713 			grp->faults[i] = p->numa_faults[i];
2714 
2715 		grp->total_faults = p->total_numa_faults;
2716 
2717 		grp->nr_tasks++;
2718 		rcu_assign_pointer(p->numa_group, grp);
2719 	}
2720 
2721 	rcu_read_lock();
2722 	tsk = READ_ONCE(cpu_rq(cpu)->curr);
2723 
2724 	if (!cpupid_match_pid(tsk, cpupid))
2725 		goto no_join;
2726 
2727 	grp = rcu_dereference(tsk->numa_group);
2728 	if (!grp)
2729 		goto no_join;
2730 
2731 	my_grp = deref_curr_numa_group(p);
2732 	if (grp == my_grp)
2733 		goto no_join;
2734 
2735 	/*
2736 	 * Only join the other group if its bigger; if we're the bigger group,
2737 	 * the other task will join us.
2738 	 */
2739 	if (my_grp->nr_tasks > grp->nr_tasks)
2740 		goto no_join;
2741 
2742 	/*
2743 	 * Tie-break on the grp address.
2744 	 */
2745 	if (my_grp->nr_tasks == grp->nr_tasks && my_grp > grp)
2746 		goto no_join;
2747 
2748 	/* Always join threads in the same process. */
2749 	if (tsk->mm == current->mm)
2750 		join = true;
2751 
2752 	/* Simple filter to avoid false positives due to PID collisions */
2753 	if (flags & TNF_SHARED)
2754 		join = true;
2755 
2756 	/* Update priv based on whether false sharing was detected */
2757 	*priv = !join;
2758 
2759 	if (join && !get_numa_group(grp))
2760 		goto no_join;
2761 
2762 	rcu_read_unlock();
2763 
2764 	if (!join)
2765 		return;
2766 
2767 	WARN_ON_ONCE(irqs_disabled());
2768 	double_lock_irq(&my_grp->lock, &grp->lock);
2769 
2770 	for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++) {
2771 		my_grp->faults[i] -= p->numa_faults[i];
2772 		grp->faults[i] += p->numa_faults[i];
2773 	}
2774 	my_grp->total_faults -= p->total_numa_faults;
2775 	grp->total_faults += p->total_numa_faults;
2776 
2777 	my_grp->nr_tasks--;
2778 	grp->nr_tasks++;
2779 
2780 	spin_unlock(&my_grp->lock);
2781 	spin_unlock_irq(&grp->lock);
2782 
2783 	rcu_assign_pointer(p->numa_group, grp);
2784 
2785 	put_numa_group(my_grp);
2786 	return;
2787 
2788 no_join:
2789 	rcu_read_unlock();
2790 	return;
2791 }
2792 
2793 /*
2794  * Get rid of NUMA statistics associated with a task (either current or dead).
2795  * If @final is set, the task is dead and has reached refcount zero, so we can
2796  * safely free all relevant data structures. Otherwise, there might be
2797  * concurrent reads from places like load balancing and procfs, and we should
2798  * reset the data back to default state without freeing ->numa_faults.
2799  */
2800 void task_numa_free(struct task_struct *p, bool final)
2801 {
2802 	/* safe: p either is current or is being freed by current */
2803 	struct numa_group *grp = rcu_dereference_raw(p->numa_group);
2804 	unsigned long *numa_faults = p->numa_faults;
2805 	unsigned long flags;
2806 	int i;
2807 
2808 	if (!numa_faults)
2809 		return;
2810 
2811 	if (grp) {
2812 		spin_lock_irqsave(&grp->lock, flags);
2813 		for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2814 			grp->faults[i] -= p->numa_faults[i];
2815 		grp->total_faults -= p->total_numa_faults;
2816 
2817 		grp->nr_tasks--;
2818 		spin_unlock_irqrestore(&grp->lock, flags);
2819 		RCU_INIT_POINTER(p->numa_group, NULL);
2820 		put_numa_group(grp);
2821 	}
2822 
2823 	if (final) {
2824 		p->numa_faults = NULL;
2825 		kfree(numa_faults);
2826 	} else {
2827 		p->total_numa_faults = 0;
2828 		for (i = 0; i < NR_NUMA_HINT_FAULT_STATS * nr_node_ids; i++)
2829 			numa_faults[i] = 0;
2830 	}
2831 }
2832 
2833 /*
2834  * Got a PROT_NONE fault for a page on @node.
2835  */
2836 void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags)
2837 {
2838 	struct task_struct *p = current;
2839 	bool migrated = flags & TNF_MIGRATED;
2840 	int cpu_node = task_node(current);
2841 	int local = !!(flags & TNF_FAULT_LOCAL);
2842 	struct numa_group *ng;
2843 	int priv;
2844 
2845 	if (!static_branch_likely(&sched_numa_balancing))
2846 		return;
2847 
2848 	/* for example, ksmd faulting in a user's mm */
2849 	if (!p->mm)
2850 		return;
2851 
2852 	/*
2853 	 * NUMA faults statistics are unnecessary for the slow memory
2854 	 * node for memory tiering mode.
2855 	 */
2856 	if (!node_is_toptier(mem_node) &&
2857 	    (sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING ||
2858 	     !cpupid_valid(last_cpupid)))
2859 		return;
2860 
2861 	/* Allocate buffer to track faults on a per-node basis */
2862 	if (unlikely(!p->numa_faults)) {
2863 		int size = sizeof(*p->numa_faults) *
2864 			   NR_NUMA_HINT_FAULT_BUCKETS * nr_node_ids;
2865 
2866 		p->numa_faults = kzalloc(size, GFP_KERNEL|__GFP_NOWARN);
2867 		if (!p->numa_faults)
2868 			return;
2869 
2870 		p->total_numa_faults = 0;
2871 		memset(p->numa_faults_locality, 0, sizeof(p->numa_faults_locality));
2872 	}
2873 
2874 	/*
2875 	 * First accesses are treated as private, otherwise consider accesses
2876 	 * to be private if the accessing pid has not changed
2877 	 */
2878 	if (unlikely(last_cpupid == (-1 & LAST_CPUPID_MASK))) {
2879 		priv = 1;
2880 	} else {
2881 		priv = cpupid_match_pid(p, last_cpupid);
2882 		if (!priv && !(flags & TNF_NO_GROUP))
2883 			task_numa_group(p, last_cpupid, flags, &priv);
2884 	}
2885 
2886 	/*
2887 	 * If a workload spans multiple NUMA nodes, a shared fault that
2888 	 * occurs wholly within the set of nodes that the workload is
2889 	 * actively using should be counted as local. This allows the
2890 	 * scan rate to slow down when a workload has settled down.
2891 	 */
2892 	ng = deref_curr_numa_group(p);
2893 	if (!priv && !local && ng && ng->active_nodes > 1 &&
2894 				numa_is_active_node(cpu_node, ng) &&
2895 				numa_is_active_node(mem_node, ng))
2896 		local = 1;
2897 
2898 	/*
2899 	 * Retry to migrate task to preferred node periodically, in case it
2900 	 * previously failed, or the scheduler moved us.
2901 	 */
2902 	if (time_after(jiffies, p->numa_migrate_retry)) {
2903 		task_numa_placement(p);
2904 		numa_migrate_preferred(p);
2905 	}
2906 
2907 	if (migrated)
2908 		p->numa_pages_migrated += pages;
2909 	if (flags & TNF_MIGRATE_FAIL)
2910 		p->numa_faults_locality[2] += pages;
2911 
2912 	p->numa_faults[task_faults_idx(NUMA_MEMBUF, mem_node, priv)] += pages;
2913 	p->numa_faults[task_faults_idx(NUMA_CPUBUF, cpu_node, priv)] += pages;
2914 	p->numa_faults_locality[local] += pages;
2915 }
2916 
2917 static void reset_ptenuma_scan(struct task_struct *p)
2918 {
2919 	/*
2920 	 * We only did a read acquisition of the mmap sem, so
2921 	 * p->mm->numa_scan_seq is written to without exclusive access
2922 	 * and the update is not guaranteed to be atomic. That's not
2923 	 * much of an issue though, since this is just used for
2924 	 * statistical sampling. Use READ_ONCE/WRITE_ONCE, which are not
2925 	 * expensive, to avoid any form of compiler optimizations:
2926 	 */
2927 	WRITE_ONCE(p->mm->numa_scan_seq, READ_ONCE(p->mm->numa_scan_seq) + 1);
2928 	p->mm->numa_scan_offset = 0;
2929 }
2930 
2931 /*
2932  * The expensive part of numa migration is done from task_work context.
2933  * Triggered from task_tick_numa().
2934  */
2935 static void task_numa_work(struct callback_head *work)
2936 {
2937 	unsigned long migrate, next_scan, now = jiffies;
2938 	struct task_struct *p = current;
2939 	struct mm_struct *mm = p->mm;
2940 	u64 runtime = p->se.sum_exec_runtime;
2941 	struct vm_area_struct *vma;
2942 	unsigned long start, end;
2943 	unsigned long nr_pte_updates = 0;
2944 	long pages, virtpages;
2945 	struct vma_iterator vmi;
2946 
2947 	SCHED_WARN_ON(p != container_of(work, struct task_struct, numa_work));
2948 
2949 	work->next = work;
2950 	/*
2951 	 * Who cares about NUMA placement when they're dying.
2952 	 *
2953 	 * NOTE: make sure not to dereference p->mm before this check,
2954 	 * exit_task_work() happens _after_ exit_mm() so we could be called
2955 	 * without p->mm even though we still had it when we enqueued this
2956 	 * work.
2957 	 */
2958 	if (p->flags & PF_EXITING)
2959 		return;
2960 
2961 	if (!mm->numa_next_scan) {
2962 		mm->numa_next_scan = now +
2963 			msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2964 	}
2965 
2966 	/*
2967 	 * Enforce maximal scan/migration frequency..
2968 	 */
2969 	migrate = mm->numa_next_scan;
2970 	if (time_before(now, migrate))
2971 		return;
2972 
2973 	if (p->numa_scan_period == 0) {
2974 		p->numa_scan_period_max = task_scan_max(p);
2975 		p->numa_scan_period = task_scan_start(p);
2976 	}
2977 
2978 	next_scan = now + msecs_to_jiffies(p->numa_scan_period);
2979 	if (!try_cmpxchg(&mm->numa_next_scan, &migrate, next_scan))
2980 		return;
2981 
2982 	/*
2983 	 * Delay this task enough that another task of this mm will likely win
2984 	 * the next time around.
2985 	 */
2986 	p->node_stamp += 2 * TICK_NSEC;
2987 
2988 	start = mm->numa_scan_offset;
2989 	pages = sysctl_numa_balancing_scan_size;
2990 	pages <<= 20 - PAGE_SHIFT; /* MB in pages */
2991 	virtpages = pages * 8;	   /* Scan up to this much virtual space */
2992 	if (!pages)
2993 		return;
2994 
2995 
2996 	if (!mmap_read_trylock(mm))
2997 		return;
2998 	vma_iter_init(&vmi, mm, start);
2999 	vma = vma_next(&vmi);
3000 	if (!vma) {
3001 		reset_ptenuma_scan(p);
3002 		start = 0;
3003 		vma_iter_set(&vmi, start);
3004 		vma = vma_next(&vmi);
3005 	}
3006 
3007 	do {
3008 		if (!vma_migratable(vma) || !vma_policy_mof(vma) ||
3009 			is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_MIXEDMAP)) {
3010 			continue;
3011 		}
3012 
3013 		/*
3014 		 * Shared library pages mapped by multiple processes are not
3015 		 * migrated as it is expected they are cache replicated. Avoid
3016 		 * hinting faults in read-only file-backed mappings or the vdso
3017 		 * as migrating the pages will be of marginal benefit.
3018 		 */
3019 		if (!vma->vm_mm ||
3020 		    (vma->vm_file && (vma->vm_flags & (VM_READ|VM_WRITE)) == (VM_READ)))
3021 			continue;
3022 
3023 		/*
3024 		 * Skip inaccessible VMAs to avoid any confusion between
3025 		 * PROT_NONE and NUMA hinting ptes
3026 		 */
3027 		if (!vma_is_accessible(vma))
3028 			continue;
3029 
3030 		do {
3031 			start = max(start, vma->vm_start);
3032 			end = ALIGN(start + (pages << PAGE_SHIFT), HPAGE_SIZE);
3033 			end = min(end, vma->vm_end);
3034 			nr_pte_updates = change_prot_numa(vma, start, end);
3035 
3036 			/*
3037 			 * Try to scan sysctl_numa_balancing_size worth of
3038 			 * hpages that have at least one present PTE that
3039 			 * is not already pte-numa. If the VMA contains
3040 			 * areas that are unused or already full of prot_numa
3041 			 * PTEs, scan up to virtpages, to skip through those
3042 			 * areas faster.
3043 			 */
3044 			if (nr_pte_updates)
3045 				pages -= (end - start) >> PAGE_SHIFT;
3046 			virtpages -= (end - start) >> PAGE_SHIFT;
3047 
3048 			start = end;
3049 			if (pages <= 0 || virtpages <= 0)
3050 				goto out;
3051 
3052 			cond_resched();
3053 		} while (end != vma->vm_end);
3054 	} for_each_vma(vmi, vma);
3055 
3056 out:
3057 	/*
3058 	 * It is possible to reach the end of the VMA list but the last few
3059 	 * VMAs are not guaranteed to the vma_migratable. If they are not, we
3060 	 * would find the !migratable VMA on the next scan but not reset the
3061 	 * scanner to the start so check it now.
3062 	 */
3063 	if (vma)
3064 		mm->numa_scan_offset = start;
3065 	else
3066 		reset_ptenuma_scan(p);
3067 	mmap_read_unlock(mm);
3068 
3069 	/*
3070 	 * Make sure tasks use at least 32x as much time to run other code
3071 	 * than they used here, to limit NUMA PTE scanning overhead to 3% max.
3072 	 * Usually update_task_scan_period slows down scanning enough; on an
3073 	 * overloaded system we need to limit overhead on a per task basis.
3074 	 */
3075 	if (unlikely(p->se.sum_exec_runtime != runtime)) {
3076 		u64 diff = p->se.sum_exec_runtime - runtime;
3077 		p->node_stamp += 32 * diff;
3078 	}
3079 }
3080 
3081 void init_numa_balancing(unsigned long clone_flags, struct task_struct *p)
3082 {
3083 	int mm_users = 0;
3084 	struct mm_struct *mm = p->mm;
3085 
3086 	if (mm) {
3087 		mm_users = atomic_read(&mm->mm_users);
3088 		if (mm_users == 1) {
3089 			mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
3090 			mm->numa_scan_seq = 0;
3091 		}
3092 	}
3093 	p->node_stamp			= 0;
3094 	p->numa_scan_seq		= mm ? mm->numa_scan_seq : 0;
3095 	p->numa_scan_period		= sysctl_numa_balancing_scan_delay;
3096 	p->numa_migrate_retry		= 0;
3097 	/* Protect against double add, see task_tick_numa and task_numa_work */
3098 	p->numa_work.next		= &p->numa_work;
3099 	p->numa_faults			= NULL;
3100 	p->numa_pages_migrated		= 0;
3101 	p->total_numa_faults		= 0;
3102 	RCU_INIT_POINTER(p->numa_group, NULL);
3103 	p->last_task_numa_placement	= 0;
3104 	p->last_sum_exec_runtime	= 0;
3105 
3106 	init_task_work(&p->numa_work, task_numa_work);
3107 
3108 	/* New address space, reset the preferred nid */
3109 	if (!(clone_flags & CLONE_VM)) {
3110 		p->numa_preferred_nid = NUMA_NO_NODE;
3111 		return;
3112 	}
3113 
3114 	/*
3115 	 * New thread, keep existing numa_preferred_nid which should be copied
3116 	 * already by arch_dup_task_struct but stagger when scans start.
3117 	 */
3118 	if (mm) {
3119 		unsigned int delay;
3120 
3121 		delay = min_t(unsigned int, task_scan_max(current),
3122 			current->numa_scan_period * mm_users * NSEC_PER_MSEC);
3123 		delay += 2 * TICK_NSEC;
3124 		p->node_stamp = delay;
3125 	}
3126 }
3127 
3128 /*
3129  * Drive the periodic memory faults..
3130  */
3131 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
3132 {
3133 	struct callback_head *work = &curr->numa_work;
3134 	u64 period, now;
3135 
3136 	/*
3137 	 * We don't care about NUMA placement if we don't have memory.
3138 	 */
3139 	if (!curr->mm || (curr->flags & (PF_EXITING | PF_KTHREAD)) || work->next != work)
3140 		return;
3141 
3142 	/*
3143 	 * Using runtime rather than walltime has the dual advantage that
3144 	 * we (mostly) drive the selection from busy threads and that the
3145 	 * task needs to have done some actual work before we bother with
3146 	 * NUMA placement.
3147 	 */
3148 	now = curr->se.sum_exec_runtime;
3149 	period = (u64)curr->numa_scan_period * NSEC_PER_MSEC;
3150 
3151 	if (now > curr->node_stamp + period) {
3152 		if (!curr->node_stamp)
3153 			curr->numa_scan_period = task_scan_start(curr);
3154 		curr->node_stamp += period;
3155 
3156 		if (!time_before(jiffies, curr->mm->numa_next_scan))
3157 			task_work_add(curr, work, TWA_RESUME);
3158 	}
3159 }
3160 
3161 static void update_scan_period(struct task_struct *p, int new_cpu)
3162 {
3163 	int src_nid = cpu_to_node(task_cpu(p));
3164 	int dst_nid = cpu_to_node(new_cpu);
3165 
3166 	if (!static_branch_likely(&sched_numa_balancing))
3167 		return;
3168 
3169 	if (!p->mm || !p->numa_faults || (p->flags & PF_EXITING))
3170 		return;
3171 
3172 	if (src_nid == dst_nid)
3173 		return;
3174 
3175 	/*
3176 	 * Allow resets if faults have been trapped before one scan
3177 	 * has completed. This is most likely due to a new task that
3178 	 * is pulled cross-node due to wakeups or load balancing.
3179 	 */
3180 	if (p->numa_scan_seq) {
3181 		/*
3182 		 * Avoid scan adjustments if moving to the preferred
3183 		 * node or if the task was not previously running on
3184 		 * the preferred node.
3185 		 */
3186 		if (dst_nid == p->numa_preferred_nid ||
3187 		    (p->numa_preferred_nid != NUMA_NO_NODE &&
3188 			src_nid != p->numa_preferred_nid))
3189 			return;
3190 	}
3191 
3192 	p->numa_scan_period = task_scan_start(p);
3193 }
3194 
3195 #else
3196 static void task_tick_numa(struct rq *rq, struct task_struct *curr)
3197 {
3198 }
3199 
3200 static inline void account_numa_enqueue(struct rq *rq, struct task_struct *p)
3201 {
3202 }
3203 
3204 static inline void account_numa_dequeue(struct rq *rq, struct task_struct *p)
3205 {
3206 }
3207 
3208 static inline void update_scan_period(struct task_struct *p, int new_cpu)
3209 {
3210 }
3211 
3212 #endif /* CONFIG_NUMA_BALANCING */
3213 
3214 static void
3215 account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
3216 {
3217 	update_load_add(&cfs_rq->load, se->load.weight);
3218 #ifdef CONFIG_SMP
3219 	if (entity_is_task(se)) {
3220 		struct rq *rq = rq_of(cfs_rq);
3221 
3222 		account_numa_enqueue(rq, task_of(se));
3223 		list_add(&se->group_node, &rq->cfs_tasks);
3224 	}
3225 #endif
3226 	cfs_rq->nr_running++;
3227 	if (se_is_idle(se))
3228 		cfs_rq->idle_nr_running++;
3229 }
3230 
3231 static void
3232 account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
3233 {
3234 	update_load_sub(&cfs_rq->load, se->load.weight);
3235 #ifdef CONFIG_SMP
3236 	if (entity_is_task(se)) {
3237 		account_numa_dequeue(rq_of(cfs_rq), task_of(se));
3238 		list_del_init(&se->group_node);
3239 	}
3240 #endif
3241 	cfs_rq->nr_running--;
3242 	if (se_is_idle(se))
3243 		cfs_rq->idle_nr_running--;
3244 }
3245 
3246 /*
3247  * Signed add and clamp on underflow.
3248  *
3249  * Explicitly do a load-store to ensure the intermediate value never hits
3250  * memory. This allows lockless observations without ever seeing the negative
3251  * values.
3252  */
3253 #define add_positive(_ptr, _val) do {                           \
3254 	typeof(_ptr) ptr = (_ptr);                              \
3255 	typeof(_val) val = (_val);                              \
3256 	typeof(*ptr) res, var = READ_ONCE(*ptr);                \
3257 								\
3258 	res = var + val;                                        \
3259 								\
3260 	if (val < 0 && res > var)                               \
3261 		res = 0;                                        \
3262 								\
3263 	WRITE_ONCE(*ptr, res);                                  \
3264 } while (0)
3265 
3266 /*
3267  * Unsigned subtract and clamp on underflow.
3268  *
3269  * Explicitly do a load-store to ensure the intermediate value never hits
3270  * memory. This allows lockless observations without ever seeing the negative
3271  * values.
3272  */
3273 #define sub_positive(_ptr, _val) do {				\
3274 	typeof(_ptr) ptr = (_ptr);				\
3275 	typeof(*ptr) val = (_val);				\
3276 	typeof(*ptr) res, var = READ_ONCE(*ptr);		\
3277 	res = var - val;					\
3278 	if (res > var)						\
3279 		res = 0;					\
3280 	WRITE_ONCE(*ptr, res);					\
3281 } while (0)
3282 
3283 /*
3284  * Remove and clamp on negative, from a local variable.
3285  *
3286  * A variant of sub_positive(), which does not use explicit load-store
3287  * and is thus optimized for local variable updates.
3288  */
3289 #define lsub_positive(_ptr, _val) do {				\
3290 	typeof(_ptr) ptr = (_ptr);				\
3291 	*ptr -= min_t(typeof(*ptr), *ptr, _val);		\
3292 } while (0)
3293 
3294 #ifdef CONFIG_SMP
3295 static inline void
3296 enqueue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3297 {
3298 	cfs_rq->avg.load_avg += se->avg.load_avg;
3299 	cfs_rq->avg.load_sum += se_weight(se) * se->avg.load_sum;
3300 }
3301 
3302 static inline void
3303 dequeue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
3304 {
3305 	sub_positive(&cfs_rq->avg.load_avg, se->avg.load_avg);
3306 	sub_positive(&cfs_rq->avg.load_sum, se_weight(se) * se->avg.load_sum);
3307 	/* See update_cfs_rq_load_avg() */
3308 	cfs_rq->avg.load_sum = max_t(u32, cfs_rq->avg.load_sum,
3309 					  cfs_rq->avg.load_avg * PELT_MIN_DIVIDER);
3310 }
3311 #else
3312 static inline void
3313 enqueue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
3314 static inline void
3315 dequeue_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
3316 #endif
3317 
3318 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
3319 			    unsigned long weight)
3320 {
3321 	if (se->on_rq) {
3322 		/* commit outstanding execution time */
3323 		if (cfs_rq->curr == se)
3324 			update_curr(cfs_rq);
3325 		update_load_sub(&cfs_rq->load, se->load.weight);
3326 	}
3327 	dequeue_load_avg(cfs_rq, se);
3328 
3329 	update_load_set(&se->load, weight);
3330 
3331 #ifdef CONFIG_SMP
3332 	do {
3333 		u32 divider = get_pelt_divider(&se->avg);
3334 
3335 		se->avg.load_avg = div_u64(se_weight(se) * se->avg.load_sum, divider);
3336 	} while (0);
3337 #endif
3338 
3339 	enqueue_load_avg(cfs_rq, se);
3340 	if (se->on_rq)
3341 		update_load_add(&cfs_rq->load, se->load.weight);
3342 
3343 }
3344 
3345 void reweight_task(struct task_struct *p, int prio)
3346 {
3347 	struct sched_entity *se = &p->se;
3348 	struct cfs_rq *cfs_rq = cfs_rq_of(se);
3349 	struct load_weight *load = &se->load;
3350 	unsigned long weight = scale_load(sched_prio_to_weight[prio]);
3351 
3352 	reweight_entity(cfs_rq, se, weight);
3353 	load->inv_weight = sched_prio_to_wmult[prio];
3354 }
3355 
3356 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq);
3357 
3358 #ifdef CONFIG_FAIR_GROUP_SCHED
3359 #ifdef CONFIG_SMP
3360 /*
3361  * All this does is approximate the hierarchical proportion which includes that
3362  * global sum we all love to hate.
3363  *
3364  * That is, the weight of a group entity, is the proportional share of the
3365  * group weight based on the group runqueue weights. That is:
3366  *
3367  *                     tg->weight * grq->load.weight
3368  *   ge->load.weight = -----------------------------               (1)
3369  *                       \Sum grq->load.weight
3370  *
3371  * Now, because computing that sum is prohibitively expensive to compute (been
3372  * there, done that) we approximate it with this average stuff. The average
3373  * moves slower and therefore the approximation is cheaper and more stable.
3374  *
3375  * So instead of the above, we substitute:
3376  *
3377  *   grq->load.weight -> grq->avg.load_avg                         (2)
3378  *
3379  * which yields the following:
3380  *
3381  *                     tg->weight * grq->avg.load_avg
3382  *   ge->load.weight = ------------------------------              (3)
3383  *                             tg->load_avg
3384  *
3385  * Where: tg->load_avg ~= \Sum grq->avg.load_avg
3386  *
3387  * That is shares_avg, and it is right (given the approximation (2)).
3388  *
3389  * The problem with it is that because the average is slow -- it was designed
3390  * to be exactly that of course -- this leads to transients in boundary
3391  * conditions. In specific, the case where the group was idle and we start the
3392  * one task. It takes time for our CPU's grq->avg.load_avg to build up,
3393  * yielding bad latency etc..
3394  *
3395  * Now, in that special case (1) reduces to:
3396  *
3397  *                     tg->weight * grq->load.weight
3398  *   ge->load.weight = ----------------------------- = tg->weight   (4)
3399  *                         grp->load.weight
3400  *
3401  * That is, the sum collapses because all other CPUs are idle; the UP scenario.
3402  *
3403  * So what we do is modify our approximation (3) to approach (4) in the (near)
3404  * UP case, like:
3405  *
3406  *   ge->load.weight =
3407  *
3408  *              tg->weight * grq->load.weight
3409  *     ---------------------------------------------------         (5)
3410  *     tg->load_avg - grq->avg.load_avg + grq->load.weight
3411  *
3412  * But because grq->load.weight can drop to 0, resulting in a divide by zero,
3413  * we need to use grq->avg.load_avg as its lower bound, which then gives:
3414  *
3415  *
3416  *                     tg->weight * grq->load.weight
3417  *   ge->load.weight = -----------------------------		   (6)
3418  *                             tg_load_avg'
3419  *
3420  * Where:
3421  *
3422  *   tg_load_avg' = tg->load_avg - grq->avg.load_avg +
3423  *                  max(grq->load.weight, grq->avg.load_avg)
3424  *
3425  * And that is shares_weight and is icky. In the (near) UP case it approaches
3426  * (4) while in the normal case it approaches (3). It consistently
3427  * overestimates the ge->load.weight and therefore:
3428  *
3429  *   \Sum ge->load.weight >= tg->weight
3430  *
3431  * hence icky!
3432  */
3433 static long calc_group_shares(struct cfs_rq *cfs_rq)
3434 {
3435 	long tg_weight, tg_shares, load, shares;
3436 	struct task_group *tg = cfs_rq->tg;
3437 
3438 	tg_shares = READ_ONCE(tg->shares);
3439 
3440 	load = max(scale_load_down(cfs_rq->load.weight), cfs_rq->avg.load_avg);
3441 
3442 	tg_weight = atomic_long_read(&tg->load_avg);
3443 
3444 	/* Ensure tg_weight >= load */
3445 	tg_weight -= cfs_rq->tg_load_avg_contrib;
3446 	tg_weight += load;
3447 
3448 	shares = (tg_shares * load);
3449 	if (tg_weight)
3450 		shares /= tg_weight;
3451 
3452 	/*
3453 	 * MIN_SHARES has to be unscaled here to support per-CPU partitioning
3454 	 * of a group with small tg->shares value. It is a floor value which is
3455 	 * assigned as a minimum load.weight to the sched_entity representing
3456 	 * the group on a CPU.
3457 	 *
3458 	 * E.g. on 64-bit for a group with tg->shares of scale_load(15)=15*1024
3459 	 * on an 8-core system with 8 tasks each runnable on one CPU shares has
3460 	 * to be 15*1024*1/8=1920 instead of scale_load(MIN_SHARES)=2*1024. In
3461 	 * case no task is runnable on a CPU MIN_SHARES=2 should be returned
3462 	 * instead of 0.
3463 	 */
3464 	return clamp_t(long, shares, MIN_SHARES, tg_shares);
3465 }
3466 #endif /* CONFIG_SMP */
3467 
3468 /*
3469  * Recomputes the group entity based on the current state of its group
3470  * runqueue.
3471  */
3472 static void update_cfs_group(struct sched_entity *se)
3473 {
3474 	struct cfs_rq *gcfs_rq = group_cfs_rq(se);
3475 	long shares;
3476 
3477 	if (!gcfs_rq)
3478 		return;
3479 
3480 	if (throttled_hierarchy(gcfs_rq))
3481 		return;
3482 
3483 #ifndef CONFIG_SMP
3484 	shares = READ_ONCE(gcfs_rq->tg->shares);
3485 
3486 	if (likely(se->load.weight == shares))
3487 		return;
3488 #else
3489 	shares   = calc_group_shares(gcfs_rq);
3490 #endif
3491 
3492 	reweight_entity(cfs_rq_of(se), se, shares);
3493 }
3494 
3495 #else /* CONFIG_FAIR_GROUP_SCHED */
3496 static inline void update_cfs_group(struct sched_entity *se)
3497 {
3498 }
3499 #endif /* CONFIG_FAIR_GROUP_SCHED */
3500 
3501 static inline void cfs_rq_util_change(struct cfs_rq *cfs_rq, int flags)
3502 {
3503 	struct rq *rq = rq_of(cfs_rq);
3504 
3505 	if (&rq->cfs == cfs_rq) {
3506 		/*
3507 		 * There are a few boundary cases this might miss but it should
3508 		 * get called often enough that that should (hopefully) not be
3509 		 * a real problem.
3510 		 *
3511 		 * It will not get called when we go idle, because the idle
3512 		 * thread is a different class (!fair), nor will the utilization
3513 		 * number include things like RT tasks.
3514 		 *
3515 		 * As is, the util number is not freq-invariant (we'd have to
3516 		 * implement arch_scale_freq_capacity() for that).
3517 		 *
3518 		 * See cpu_util_cfs().
3519 		 */
3520 		cpufreq_update_util(rq, flags);
3521 	}
3522 }
3523 
3524 #ifdef CONFIG_SMP
3525 static inline bool load_avg_is_decayed(struct sched_avg *sa)
3526 {
3527 	if (sa->load_sum)
3528 		return false;
3529 
3530 	if (sa->util_sum)
3531 		return false;
3532 
3533 	if (sa->runnable_sum)
3534 		return false;
3535 
3536 	/*
3537 	 * _avg must be null when _sum are null because _avg = _sum / divider
3538 	 * Make sure that rounding and/or propagation of PELT values never
3539 	 * break this.
3540 	 */
3541 	SCHED_WARN_ON(sa->load_avg ||
3542 		      sa->util_avg ||
3543 		      sa->runnable_avg);
3544 
3545 	return true;
3546 }
3547 
3548 static inline u64 cfs_rq_last_update_time(struct cfs_rq *cfs_rq)
3549 {
3550 	return u64_u32_load_copy(cfs_rq->avg.last_update_time,
3551 				 cfs_rq->last_update_time_copy);
3552 }
3553 #ifdef CONFIG_FAIR_GROUP_SCHED
3554 /*
3555  * Because list_add_leaf_cfs_rq always places a child cfs_rq on the list
3556  * immediately before a parent cfs_rq, and cfs_rqs are removed from the list
3557  * bottom-up, we only have to test whether the cfs_rq before us on the list
3558  * is our child.
3559  * If cfs_rq is not on the list, test whether a child needs its to be added to
3560  * connect a branch to the tree  * (see list_add_leaf_cfs_rq() for details).
3561  */
3562 static inline bool child_cfs_rq_on_list(struct cfs_rq *cfs_rq)
3563 {
3564 	struct cfs_rq *prev_cfs_rq;
3565 	struct list_head *prev;
3566 
3567 	if (cfs_rq->on_list) {
3568 		prev = cfs_rq->leaf_cfs_rq_list.prev;
3569 	} else {
3570 		struct rq *rq = rq_of(cfs_rq);
3571 
3572 		prev = rq->tmp_alone_branch;
3573 	}
3574 
3575 	prev_cfs_rq = container_of(prev, struct cfs_rq, leaf_cfs_rq_list);
3576 
3577 	return (prev_cfs_rq->tg->parent == cfs_rq->tg);
3578 }
3579 
3580 static inline bool cfs_rq_is_decayed(struct cfs_rq *cfs_rq)
3581 {
3582 	if (cfs_rq->load.weight)
3583 		return false;
3584 
3585 	if (!load_avg_is_decayed(&cfs_rq->avg))
3586 		return false;
3587 
3588 	if (child_cfs_rq_on_list(cfs_rq))
3589 		return false;
3590 
3591 	return true;
3592 }
3593 
3594 /**
3595  * update_tg_load_avg - update the tg's load avg
3596  * @cfs_rq: the cfs_rq whose avg changed
3597  *
3598  * This function 'ensures': tg->load_avg := \Sum tg->cfs_rq[]->avg.load.
3599  * However, because tg->load_avg is a global value there are performance
3600  * considerations.
3601  *
3602  * In order to avoid having to look at the other cfs_rq's, we use a
3603  * differential update where we store the last value we propagated. This in
3604  * turn allows skipping updates if the differential is 'small'.
3605  *
3606  * Updating tg's load_avg is necessary before update_cfs_share().
3607  */
3608 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq)
3609 {
3610 	long delta = cfs_rq->avg.load_avg - cfs_rq->tg_load_avg_contrib;
3611 
3612 	/*
3613 	 * No need to update load_avg for root_task_group as it is not used.
3614 	 */
3615 	if (cfs_rq->tg == &root_task_group)
3616 		return;
3617 
3618 	if (abs(delta) > cfs_rq->tg_load_avg_contrib / 64) {
3619 		atomic_long_add(delta, &cfs_rq->tg->load_avg);
3620 		cfs_rq->tg_load_avg_contrib = cfs_rq->avg.load_avg;
3621 	}
3622 }
3623 
3624 /*
3625  * Called within set_task_rq() right before setting a task's CPU. The
3626  * caller only guarantees p->pi_lock is held; no other assumptions,
3627  * including the state of rq->lock, should be made.
3628  */
3629 void set_task_rq_fair(struct sched_entity *se,
3630 		      struct cfs_rq *prev, struct cfs_rq *next)
3631 {
3632 	u64 p_last_update_time;
3633 	u64 n_last_update_time;
3634 
3635 	if (!sched_feat(ATTACH_AGE_LOAD))
3636 		return;
3637 
3638 	/*
3639 	 * We are supposed to update the task to "current" time, then its up to
3640 	 * date and ready to go to new CPU/cfs_rq. But we have difficulty in
3641 	 * getting what current time is, so simply throw away the out-of-date
3642 	 * time. This will result in the wakee task is less decayed, but giving
3643 	 * the wakee more load sounds not bad.
3644 	 */
3645 	if (!(se->avg.last_update_time && prev))
3646 		return;
3647 
3648 	p_last_update_time = cfs_rq_last_update_time(prev);
3649 	n_last_update_time = cfs_rq_last_update_time(next);
3650 
3651 	__update_load_avg_blocked_se(p_last_update_time, se);
3652 	se->avg.last_update_time = n_last_update_time;
3653 }
3654 
3655 /*
3656  * When on migration a sched_entity joins/leaves the PELT hierarchy, we need to
3657  * propagate its contribution. The key to this propagation is the invariant
3658  * that for each group:
3659  *
3660  *   ge->avg == grq->avg						(1)
3661  *
3662  * _IFF_ we look at the pure running and runnable sums. Because they
3663  * represent the very same entity, just at different points in the hierarchy.
3664  *
3665  * Per the above update_tg_cfs_util() and update_tg_cfs_runnable() are trivial
3666  * and simply copies the running/runnable sum over (but still wrong, because
3667  * the group entity and group rq do not have their PELT windows aligned).
3668  *
3669  * However, update_tg_cfs_load() is more complex. So we have:
3670  *
3671  *   ge->avg.load_avg = ge->load.weight * ge->avg.runnable_avg		(2)
3672  *
3673  * And since, like util, the runnable part should be directly transferable,
3674  * the following would _appear_ to be the straight forward approach:
3675  *
3676  *   grq->avg.load_avg = grq->load.weight * grq->avg.runnable_avg	(3)
3677  *
3678  * And per (1) we have:
3679  *
3680  *   ge->avg.runnable_avg == grq->avg.runnable_avg
3681  *
3682  * Which gives:
3683  *
3684  *                      ge->load.weight * grq->avg.load_avg
3685  *   ge->avg.load_avg = -----------------------------------		(4)
3686  *                               grq->load.weight
3687  *
3688  * Except that is wrong!
3689  *
3690  * Because while for entities historical weight is not important and we
3691  * really only care about our future and therefore can consider a pure
3692  * runnable sum, runqueues can NOT do this.
3693  *
3694  * We specifically want runqueues to have a load_avg that includes
3695  * historical weights. Those represent the blocked load, the load we expect
3696  * to (shortly) return to us. This only works by keeping the weights as
3697  * integral part of the sum. We therefore cannot decompose as per (3).
3698  *
3699  * Another reason this doesn't work is that runnable isn't a 0-sum entity.
3700  * Imagine a rq with 2 tasks that each are runnable 2/3 of the time. Then the
3701  * rq itself is runnable anywhere between 2/3 and 1 depending on how the
3702  * runnable section of these tasks overlap (or not). If they were to perfectly
3703  * align the rq as a whole would be runnable 2/3 of the time. If however we
3704  * always have at least 1 runnable task, the rq as a whole is always runnable.
3705  *
3706  * So we'll have to approximate.. :/
3707  *
3708  * Given the constraint:
3709  *
3710  *   ge->avg.running_sum <= ge->avg.runnable_sum <= LOAD_AVG_MAX
3711  *
3712  * We can construct a rule that adds runnable to a rq by assuming minimal
3713  * overlap.
3714  *
3715  * On removal, we'll assume each task is equally runnable; which yields:
3716  *
3717  *   grq->avg.runnable_sum = grq->avg.load_sum / grq->load.weight
3718  *
3719  * XXX: only do this for the part of runnable > running ?
3720  *
3721  */
3722 static inline void
3723 update_tg_cfs_util(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq)
3724 {
3725 	long delta_sum, delta_avg = gcfs_rq->avg.util_avg - se->avg.util_avg;
3726 	u32 new_sum, divider;
3727 
3728 	/* Nothing to update */
3729 	if (!delta_avg)
3730 		return;
3731 
3732 	/*
3733 	 * cfs_rq->avg.period_contrib can be used for both cfs_rq and se.
3734 	 * See ___update_load_avg() for details.
3735 	 */
3736 	divider = get_pelt_divider(&cfs_rq->avg);
3737 
3738 
3739 	/* Set new sched_entity's utilization */
3740 	se->avg.util_avg = gcfs_rq->avg.util_avg;
3741 	new_sum = se->avg.util_avg * divider;
3742 	delta_sum = (long)new_sum - (long)se->avg.util_sum;
3743 	se->avg.util_sum = new_sum;
3744 
3745 	/* Update parent cfs_rq utilization */
3746 	add_positive(&cfs_rq->avg.util_avg, delta_avg);
3747 	add_positive(&cfs_rq->avg.util_sum, delta_sum);
3748 
3749 	/* See update_cfs_rq_load_avg() */
3750 	cfs_rq->avg.util_sum = max_t(u32, cfs_rq->avg.util_sum,
3751 					  cfs_rq->avg.util_avg * PELT_MIN_DIVIDER);
3752 }
3753 
3754 static inline void
3755 update_tg_cfs_runnable(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq)
3756 {
3757 	long delta_sum, delta_avg = gcfs_rq->avg.runnable_avg - se->avg.runnable_avg;
3758 	u32 new_sum, divider;
3759 
3760 	/* Nothing to update */
3761 	if (!delta_avg)
3762 		return;
3763 
3764 	/*
3765 	 * cfs_rq->avg.period_contrib can be used for both cfs_rq and se.
3766 	 * See ___update_load_avg() for details.
3767 	 */
3768 	divider = get_pelt_divider(&cfs_rq->avg);
3769 
3770 	/* Set new sched_entity's runnable */
3771 	se->avg.runnable_avg = gcfs_rq->avg.runnable_avg;
3772 	new_sum = se->avg.runnable_avg * divider;
3773 	delta_sum = (long)new_sum - (long)se->avg.runnable_sum;
3774 	se->avg.runnable_sum = new_sum;
3775 
3776 	/* Update parent cfs_rq runnable */
3777 	add_positive(&cfs_rq->avg.runnable_avg, delta_avg);
3778 	add_positive(&cfs_rq->avg.runnable_sum, delta_sum);
3779 	/* See update_cfs_rq_load_avg() */
3780 	cfs_rq->avg.runnable_sum = max_t(u32, cfs_rq->avg.runnable_sum,
3781 					      cfs_rq->avg.runnable_avg * PELT_MIN_DIVIDER);
3782 }
3783 
3784 static inline void
3785 update_tg_cfs_load(struct cfs_rq *cfs_rq, struct sched_entity *se, struct cfs_rq *gcfs_rq)
3786 {
3787 	long delta_avg, running_sum, runnable_sum = gcfs_rq->prop_runnable_sum;
3788 	unsigned long load_avg;
3789 	u64 load_sum = 0;
3790 	s64 delta_sum;
3791 	u32 divider;
3792 
3793 	if (!runnable_sum)
3794 		return;
3795 
3796 	gcfs_rq->prop_runnable_sum = 0;
3797 
3798 	/*
3799 	 * cfs_rq->avg.period_contrib can be used for both cfs_rq and se.
3800 	 * See ___update_load_avg() for details.
3801 	 */
3802 	divider = get_pelt_divider(&cfs_rq->avg);
3803 
3804 	if (runnable_sum >= 0) {
3805 		/*
3806 		 * Add runnable; clip at LOAD_AVG_MAX. Reflects that until
3807 		 * the CPU is saturated running == runnable.
3808 		 */
3809 		runnable_sum += se->avg.load_sum;
3810 		runnable_sum = min_t(long, runnable_sum, divider);
3811 	} else {
3812 		/*
3813 		 * Estimate the new unweighted runnable_sum of the gcfs_rq by
3814 		 * assuming all tasks are equally runnable.
3815 		 */
3816 		if (scale_load_down(gcfs_rq->load.weight)) {
3817 			load_sum = div_u64(gcfs_rq->avg.load_sum,
3818 				scale_load_down(gcfs_rq->load.weight));
3819 		}
3820 
3821 		/* But make sure to not inflate se's runnable */
3822 		runnable_sum = min(se->avg.load_sum, load_sum);
3823 	}
3824 
3825 	/*
3826 	 * runnable_sum can't be lower than running_sum
3827 	 * Rescale running sum to be in the same range as runnable sum
3828 	 * running_sum is in [0 : LOAD_AVG_MAX <<  SCHED_CAPACITY_SHIFT]
3829 	 * runnable_sum is in [0 : LOAD_AVG_MAX]
3830 	 */
3831 	running_sum = se->avg.util_sum >> SCHED_CAPACITY_SHIFT;
3832 	runnable_sum = max(runnable_sum, running_sum);
3833 
3834 	load_sum = se_weight(se) * runnable_sum;
3835 	load_avg = div_u64(load_sum, divider);
3836 
3837 	delta_avg = load_avg - se->avg.load_avg;
3838 	if (!delta_avg)
3839 		return;
3840 
3841 	delta_sum = load_sum - (s64)se_weight(se) * se->avg.load_sum;
3842 
3843 	se->avg.load_sum = runnable_sum;
3844 	se->avg.load_avg = load_avg;
3845 	add_positive(&cfs_rq->avg.load_avg, delta_avg);
3846 	add_positive(&cfs_rq->avg.load_sum, delta_sum);
3847 	/* See update_cfs_rq_load_avg() */
3848 	cfs_rq->avg.load_sum = max_t(u32, cfs_rq->avg.load_sum,
3849 					  cfs_rq->avg.load_avg * PELT_MIN_DIVIDER);
3850 }
3851 
3852 static inline void add_tg_cfs_propagate(struct cfs_rq *cfs_rq, long runnable_sum)
3853 {
3854 	cfs_rq->propagate = 1;
3855 	cfs_rq->prop_runnable_sum += runnable_sum;
3856 }
3857 
3858 /* Update task and its cfs_rq load average */
3859 static inline int propagate_entity_load_avg(struct sched_entity *se)
3860 {
3861 	struct cfs_rq *cfs_rq, *gcfs_rq;
3862 
3863 	if (entity_is_task(se))
3864 		return 0;
3865 
3866 	gcfs_rq = group_cfs_rq(se);
3867 	if (!gcfs_rq->propagate)
3868 		return 0;
3869 
3870 	gcfs_rq->propagate = 0;
3871 
3872 	cfs_rq = cfs_rq_of(se);
3873 
3874 	add_tg_cfs_propagate(cfs_rq, gcfs_rq->prop_runnable_sum);
3875 
3876 	update_tg_cfs_util(cfs_rq, se, gcfs_rq);
3877 	update_tg_cfs_runnable(cfs_rq, se, gcfs_rq);
3878 	update_tg_cfs_load(cfs_rq, se, gcfs_rq);
3879 
3880 	trace_pelt_cfs_tp(cfs_rq);
3881 	trace_pelt_se_tp(se);
3882 
3883 	return 1;
3884 }
3885 
3886 /*
3887  * Check if we need to update the load and the utilization of a blocked
3888  * group_entity:
3889  */
3890 static inline bool skip_blocked_update(struct sched_entity *se)
3891 {
3892 	struct cfs_rq *gcfs_rq = group_cfs_rq(se);
3893 
3894 	/*
3895 	 * If sched_entity still have not zero load or utilization, we have to
3896 	 * decay it:
3897 	 */
3898 	if (se->avg.load_avg || se->avg.util_avg)
3899 		return false;
3900 
3901 	/*
3902 	 * If there is a pending propagation, we have to update the load and
3903 	 * the utilization of the sched_entity:
3904 	 */
3905 	if (gcfs_rq->propagate)
3906 		return false;
3907 
3908 	/*
3909 	 * Otherwise, the load and the utilization of the sched_entity is
3910 	 * already zero and there is no pending propagation, so it will be a
3911 	 * waste of time to try to decay it:
3912 	 */
3913 	return true;
3914 }
3915 
3916 #else /* CONFIG_FAIR_GROUP_SCHED */
3917 
3918 static inline void update_tg_load_avg(struct cfs_rq *cfs_rq) {}
3919 
3920 static inline int propagate_entity_load_avg(struct sched_entity *se)
3921 {
3922 	return 0;
3923 }
3924 
3925 static inline void add_tg_cfs_propagate(struct cfs_rq *cfs_rq, long runnable_sum) {}
3926 
3927 #endif /* CONFIG_FAIR_GROUP_SCHED */
3928 
3929 #ifdef CONFIG_NO_HZ_COMMON
3930 static inline void migrate_se_pelt_lag(struct sched_entity *se)
3931 {
3932 	u64 throttled = 0, now, lut;
3933 	struct cfs_rq *cfs_rq;
3934 	struct rq *rq;
3935 	bool is_idle;
3936 
3937 	if (load_avg_is_decayed(&se->avg))
3938 		return;
3939 
3940 	cfs_rq = cfs_rq_of(se);
3941 	rq = rq_of(cfs_rq);
3942 
3943 	rcu_read_lock();
3944 	is_idle = is_idle_task(rcu_dereference(rq->curr));
3945 	rcu_read_unlock();
3946 
3947 	/*
3948 	 * The lag estimation comes with a cost we don't want to pay all the
3949 	 * time. Hence, limiting to the case where the source CPU is idle and
3950 	 * we know we are at the greatest risk to have an outdated clock.
3951 	 */
3952 	if (!is_idle)
3953 		return;
3954 
3955 	/*
3956 	 * Estimated "now" is: last_update_time + cfs_idle_lag + rq_idle_lag, where:
3957 	 *
3958 	 *   last_update_time (the cfs_rq's last_update_time)
3959 	 *	= cfs_rq_clock_pelt()@cfs_rq_idle
3960 	 *      = rq_clock_pelt()@cfs_rq_idle
3961 	 *        - cfs->throttled_clock_pelt_time@cfs_rq_idle
3962 	 *
3963 	 *   cfs_idle_lag (delta between rq's update and cfs_rq's update)
3964 	 *      = rq_clock_pelt()@rq_idle - rq_clock_pelt()@cfs_rq_idle
3965 	 *
3966 	 *   rq_idle_lag (delta between now and rq's update)
3967 	 *      = sched_clock_cpu() - rq_clock()@rq_idle
3968 	 *
3969 	 * We can then write:
3970 	 *
3971 	 *    now = rq_clock_pelt()@rq_idle - cfs->throttled_clock_pelt_time +
3972 	 *          sched_clock_cpu() - rq_clock()@rq_idle
3973 	 * Where:
3974 	 *      rq_clock_pelt()@rq_idle is rq->clock_pelt_idle
3975 	 *      rq_clock()@rq_idle      is rq->clock_idle
3976 	 *      cfs->throttled_clock_pelt_time@cfs_rq_idle
3977 	 *                              is cfs_rq->throttled_pelt_idle
3978 	 */
3979 
3980 #ifdef CONFIG_CFS_BANDWIDTH
3981 	throttled = u64_u32_load(cfs_rq->throttled_pelt_idle);
3982 	/* The clock has been stopped for throttling */
3983 	if (throttled == U64_MAX)
3984 		return;
3985 #endif
3986 	now = u64_u32_load(rq->clock_pelt_idle);
3987 	/*
3988 	 * Paired with _update_idle_rq_clock_pelt(). It ensures at the worst case
3989 	 * is observed the old clock_pelt_idle value and the new clock_idle,
3990 	 * which lead to an underestimation. The opposite would lead to an
3991 	 * overestimation.
3992 	 */
3993 	smp_rmb();
3994 	lut = cfs_rq_last_update_time(cfs_rq);
3995 
3996 	now -= throttled;
3997 	if (now < lut)
3998 		/*
3999 		 * cfs_rq->avg.last_update_time is more recent than our
4000 		 * estimation, let's use it.
4001 		 */
4002 		now = lut;
4003 	else
4004 		now += sched_clock_cpu(cpu_of(rq)) - u64_u32_load(rq->clock_idle);
4005 
4006 	__update_load_avg_blocked_se(now, se);
4007 }
4008 #else
4009 static void migrate_se_pelt_lag(struct sched_entity *se) {}
4010 #endif
4011 
4012 /**
4013  * update_cfs_rq_load_avg - update the cfs_rq's load/util averages
4014  * @now: current time, as per cfs_rq_clock_pelt()
4015  * @cfs_rq: cfs_rq to update
4016  *
4017  * The cfs_rq avg is the direct sum of all its entities (blocked and runnable)
4018  * avg. The immediate corollary is that all (fair) tasks must be attached.
4019  *
4020  * cfs_rq->avg is used for task_h_load() and update_cfs_share() for example.
4021  *
4022  * Return: true if the load decayed or we removed load.
4023  *
4024  * Since both these conditions indicate a changed cfs_rq->avg.load we should
4025  * call update_tg_load_avg() when this function returns true.
4026  */
4027 static inline int
4028 update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq)
4029 {
4030 	unsigned long removed_load = 0, removed_util = 0, removed_runnable = 0;
4031 	struct sched_avg *sa = &cfs_rq->avg;
4032 	int decayed = 0;
4033 
4034 	if (cfs_rq->removed.nr) {
4035 		unsigned long r;
4036 		u32 divider = get_pelt_divider(&cfs_rq->avg);
4037 
4038 		raw_spin_lock(&cfs_rq->removed.lock);
4039 		swap(cfs_rq->removed.util_avg, removed_util);
4040 		swap(cfs_rq->removed.load_avg, removed_load);
4041 		swap(cfs_rq->removed.runnable_avg, removed_runnable);
4042 		cfs_rq->removed.nr = 0;
4043 		raw_spin_unlock(&cfs_rq->removed.lock);
4044 
4045 		r = removed_load;
4046 		sub_positive(&sa->load_avg, r);
4047 		sub_positive(&sa->load_sum, r * divider);
4048 		/* See sa->util_sum below */
4049 		sa->load_sum = max_t(u32, sa->load_sum, sa->load_avg * PELT_MIN_DIVIDER);
4050 
4051 		r = removed_util;
4052 		sub_positive(&sa->util_avg, r);
4053 		sub_positive(&sa->util_sum, r * divider);
4054 		/*
4055 		 * Because of rounding, se->util_sum might ends up being +1 more than
4056 		 * cfs->util_sum. Although this is not a problem by itself, detaching
4057 		 * a lot of tasks with the rounding problem between 2 updates of
4058 		 * util_avg (~1ms) can make cfs->util_sum becoming null whereas
4059 		 * cfs_util_avg is not.
4060 		 * Check that util_sum is still above its lower bound for the new
4061 		 * util_avg. Given that period_contrib might have moved since the last
4062 		 * sync, we are only sure that util_sum must be above or equal to
4063 		 *    util_avg * minimum possible divider
4064 		 */
4065 		sa->util_sum = max_t(u32, sa->util_sum, sa->util_avg * PELT_MIN_DIVIDER);
4066 
4067 		r = removed_runnable;
4068 		sub_positive(&sa->runnable_avg, r);
4069 		sub_positive(&sa->runnable_sum, r * divider);
4070 		/* See sa->util_sum above */
4071 		sa->runnable_sum = max_t(u32, sa->runnable_sum,
4072 					      sa->runnable_avg * PELT_MIN_DIVIDER);
4073 
4074 		/*
4075 		 * removed_runnable is the unweighted version of removed_load so we
4076 		 * can use it to estimate removed_load_sum.
4077 		 */
4078 		add_tg_cfs_propagate(cfs_rq,
4079 			-(long)(removed_runnable * divider) >> SCHED_CAPACITY_SHIFT);
4080 
4081 		decayed = 1;
4082 	}
4083 
4084 	decayed |= __update_load_avg_cfs_rq(now, cfs_rq);
4085 	u64_u32_store_copy(sa->last_update_time,
4086 			   cfs_rq->last_update_time_copy,
4087 			   sa->last_update_time);
4088 	return decayed;
4089 }
4090 
4091 /**
4092  * attach_entity_load_avg - attach this entity to its cfs_rq load avg
4093  * @cfs_rq: cfs_rq to attach to
4094  * @se: sched_entity to attach
4095  *
4096  * Must call update_cfs_rq_load_avg() before this, since we rely on
4097  * cfs_rq->avg.last_update_time being current.
4098  */
4099 static void attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
4100 {
4101 	/*
4102 	 * cfs_rq->avg.period_contrib can be used for both cfs_rq and se.
4103 	 * See ___update_load_avg() for details.
4104 	 */
4105 	u32 divider = get_pelt_divider(&cfs_rq->avg);
4106 
4107 	/*
4108 	 * When we attach the @se to the @cfs_rq, we must align the decay
4109 	 * window because without that, really weird and wonderful things can
4110 	 * happen.
4111 	 *
4112 	 * XXX illustrate
4113 	 */
4114 	se->avg.last_update_time = cfs_rq->avg.last_update_time;
4115 	se->avg.period_contrib = cfs_rq->avg.period_contrib;
4116 
4117 	/*
4118 	 * Hell(o) Nasty stuff.. we need to recompute _sum based on the new
4119 	 * period_contrib. This isn't strictly correct, but since we're
4120 	 * entirely outside of the PELT hierarchy, nobody cares if we truncate
4121 	 * _sum a little.
4122 	 */
4123 	se->avg.util_sum = se->avg.util_avg * divider;
4124 
4125 	se->avg.runnable_sum = se->avg.runnable_avg * divider;
4126 
4127 	se->avg.load_sum = se->avg.load_avg * divider;
4128 	if (se_weight(se) < se->avg.load_sum)
4129 		se->avg.load_sum = div_u64(se->avg.load_sum, se_weight(se));
4130 	else
4131 		se->avg.load_sum = 1;
4132 
4133 	enqueue_load_avg(cfs_rq, se);
4134 	cfs_rq->avg.util_avg += se->avg.util_avg;
4135 	cfs_rq->avg.util_sum += se->avg.util_sum;
4136 	cfs_rq->avg.runnable_avg += se->avg.runnable_avg;
4137 	cfs_rq->avg.runnable_sum += se->avg.runnable_sum;
4138 
4139 	add_tg_cfs_propagate(cfs_rq, se->avg.load_sum);
4140 
4141 	cfs_rq_util_change(cfs_rq, 0);
4142 
4143 	trace_pelt_cfs_tp(cfs_rq);
4144 }
4145 
4146 /**
4147  * detach_entity_load_avg - detach this entity from its cfs_rq load avg
4148  * @cfs_rq: cfs_rq to detach from
4149  * @se: sched_entity to detach
4150  *
4151  * Must call update_cfs_rq_load_avg() before this, since we rely on
4152  * cfs_rq->avg.last_update_time being current.
4153  */
4154 static void detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
4155 {
4156 	dequeue_load_avg(cfs_rq, se);
4157 	sub_positive(&cfs_rq->avg.util_avg, se->avg.util_avg);
4158 	sub_positive(&cfs_rq->avg.util_sum, se->avg.util_sum);
4159 	/* See update_cfs_rq_load_avg() */
4160 	cfs_rq->avg.util_sum = max_t(u32, cfs_rq->avg.util_sum,
4161 					  cfs_rq->avg.util_avg * PELT_MIN_DIVIDER);
4162 
4163 	sub_positive(&cfs_rq->avg.runnable_avg, se->avg.runnable_avg);
4164 	sub_positive(&cfs_rq->avg.runnable_sum, se->avg.runnable_sum);
4165 	/* See update_cfs_rq_load_avg() */
4166 	cfs_rq->avg.runnable_sum = max_t(u32, cfs_rq->avg.runnable_sum,
4167 					      cfs_rq->avg.runnable_avg * PELT_MIN_DIVIDER);
4168 
4169 	add_tg_cfs_propagate(cfs_rq, -se->avg.load_sum);
4170 
4171 	cfs_rq_util_change(cfs_rq, 0);
4172 
4173 	trace_pelt_cfs_tp(cfs_rq);
4174 }
4175 
4176 /*
4177  * Optional action to be done while updating the load average
4178  */
4179 #define UPDATE_TG	0x1
4180 #define SKIP_AGE_LOAD	0x2
4181 #define DO_ATTACH	0x4
4182 #define DO_DETACH	0x8
4183 
4184 /* Update task and its cfs_rq load average */
4185 static inline void update_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
4186 {
4187 	u64 now = cfs_rq_clock_pelt(cfs_rq);
4188 	int decayed;
4189 
4190 	/*
4191 	 * Track task load average for carrying it to new CPU after migrated, and
4192 	 * track group sched_entity load average for task_h_load calc in migration
4193 	 */
4194 	if (se->avg.last_update_time && !(flags & SKIP_AGE_LOAD))
4195 		__update_load_avg_se(now, cfs_rq, se);
4196 
4197 	decayed  = update_cfs_rq_load_avg(now, cfs_rq);
4198 	decayed |= propagate_entity_load_avg(se);
4199 
4200 	if (!se->avg.last_update_time && (flags & DO_ATTACH)) {
4201 
4202 		/*
4203 		 * DO_ATTACH means we're here from enqueue_entity().
4204 		 * !last_update_time means we've passed through
4205 		 * migrate_task_rq_fair() indicating we migrated.
4206 		 *
4207 		 * IOW we're enqueueing a task on a new CPU.
4208 		 */
4209 		attach_entity_load_avg(cfs_rq, se);
4210 		update_tg_load_avg(cfs_rq);
4211 
4212 	} else if (flags & DO_DETACH) {
4213 		/*
4214 		 * DO_DETACH means we're here from dequeue_entity()
4215 		 * and we are migrating task out of the CPU.
4216 		 */
4217 		detach_entity_load_avg(cfs_rq, se);
4218 		update_tg_load_avg(cfs_rq);
4219 	} else if (decayed) {
4220 		cfs_rq_util_change(cfs_rq, 0);
4221 
4222 		if (flags & UPDATE_TG)
4223 			update_tg_load_avg(cfs_rq);
4224 	}
4225 }
4226 
4227 /*
4228  * Synchronize entity load avg of dequeued entity without locking
4229  * the previous rq.
4230  */
4231 static void sync_entity_load_avg(struct sched_entity *se)
4232 {
4233 	struct cfs_rq *cfs_rq = cfs_rq_of(se);
4234 	u64 last_update_time;
4235 
4236 	last_update_time = cfs_rq_last_update_time(cfs_rq);
4237 	__update_load_avg_blocked_se(last_update_time, se);
4238 }
4239 
4240 /*
4241  * Task first catches up with cfs_rq, and then subtract
4242  * itself from the cfs_rq (task must be off the queue now).
4243  */
4244 static void remove_entity_load_avg(struct sched_entity *se)
4245 {
4246 	struct cfs_rq *cfs_rq = cfs_rq_of(se);
4247 	unsigned long flags;
4248 
4249 	/*
4250 	 * tasks cannot exit without having gone through wake_up_new_task() ->
4251 	 * enqueue_task_fair() which will have added things to the cfs_rq,
4252 	 * so we can remove unconditionally.
4253 	 */
4254 
4255 	sync_entity_load_avg(se);
4256 
4257 	raw_spin_lock_irqsave(&cfs_rq->removed.lock, flags);
4258 	++cfs_rq->removed.nr;
4259 	cfs_rq->removed.util_avg	+= se->avg.util_avg;
4260 	cfs_rq->removed.load_avg	+= se->avg.load_avg;
4261 	cfs_rq->removed.runnable_avg	+= se->avg.runnable_avg;
4262 	raw_spin_unlock_irqrestore(&cfs_rq->removed.lock, flags);
4263 }
4264 
4265 static inline unsigned long cfs_rq_runnable_avg(struct cfs_rq *cfs_rq)
4266 {
4267 	return cfs_rq->avg.runnable_avg;
4268 }
4269 
4270 static inline unsigned long cfs_rq_load_avg(struct cfs_rq *cfs_rq)
4271 {
4272 	return cfs_rq->avg.load_avg;
4273 }
4274 
4275 static int newidle_balance(struct rq *this_rq, struct rq_flags *rf);
4276 
4277 static inline unsigned long task_util(struct task_struct *p)
4278 {
4279 	return READ_ONCE(p->se.avg.util_avg);
4280 }
4281 
4282 static inline unsigned long _task_util_est(struct task_struct *p)
4283 {
4284 	struct util_est ue = READ_ONCE(p->se.avg.util_est);
4285 
4286 	return max(ue.ewma, (ue.enqueued & ~UTIL_AVG_UNCHANGED));
4287 }
4288 
4289 static inline unsigned long task_util_est(struct task_struct *p)
4290 {
4291 	return max(task_util(p), _task_util_est(p));
4292 }
4293 
4294 #ifdef CONFIG_UCLAMP_TASK
4295 static inline unsigned long uclamp_task_util(struct task_struct *p,
4296 					     unsigned long uclamp_min,
4297 					     unsigned long uclamp_max)
4298 {
4299 	return clamp(task_util_est(p), uclamp_min, uclamp_max);
4300 }
4301 #else
4302 static inline unsigned long uclamp_task_util(struct task_struct *p,
4303 					     unsigned long uclamp_min,
4304 					     unsigned long uclamp_max)
4305 {
4306 	return task_util_est(p);
4307 }
4308 #endif
4309 
4310 static inline void util_est_enqueue(struct cfs_rq *cfs_rq,
4311 				    struct task_struct *p)
4312 {
4313 	unsigned int enqueued;
4314 
4315 	if (!sched_feat(UTIL_EST))
4316 		return;
4317 
4318 	/* Update root cfs_rq's estimated utilization */
4319 	enqueued  = cfs_rq->avg.util_est.enqueued;
4320 	enqueued += _task_util_est(p);
4321 	WRITE_ONCE(cfs_rq->avg.util_est.enqueued, enqueued);
4322 
4323 	trace_sched_util_est_cfs_tp(cfs_rq);
4324 }
4325 
4326 static inline void util_est_dequeue(struct cfs_rq *cfs_rq,
4327 				    struct task_struct *p)
4328 {
4329 	unsigned int enqueued;
4330 
4331 	if (!sched_feat(UTIL_EST))
4332 		return;
4333 
4334 	/* Update root cfs_rq's estimated utilization */
4335 	enqueued  = cfs_rq->avg.util_est.enqueued;
4336 	enqueued -= min_t(unsigned int, enqueued, _task_util_est(p));
4337 	WRITE_ONCE(cfs_rq->avg.util_est.enqueued, enqueued);
4338 
4339 	trace_sched_util_est_cfs_tp(cfs_rq);
4340 }
4341 
4342 #define UTIL_EST_MARGIN (SCHED_CAPACITY_SCALE / 100)
4343 
4344 /*
4345  * Check if a (signed) value is within a specified (unsigned) margin,
4346  * based on the observation that:
4347  *
4348  *     abs(x) < y := (unsigned)(x + y - 1) < (2 * y - 1)
4349  *
4350  * NOTE: this only works when value + margin < INT_MAX.
4351  */
4352 static inline bool within_margin(int value, int margin)
4353 {
4354 	return ((unsigned int)(value + margin - 1) < (2 * margin - 1));
4355 }
4356 
4357 static inline void util_est_update(struct cfs_rq *cfs_rq,
4358 				   struct task_struct *p,
4359 				   bool task_sleep)
4360 {
4361 	long last_ewma_diff, last_enqueued_diff;
4362 	struct util_est ue;
4363 
4364 	if (!sched_feat(UTIL_EST))
4365 		return;
4366 
4367 	/*
4368 	 * Skip update of task's estimated utilization when the task has not
4369 	 * yet completed an activation, e.g. being migrated.
4370 	 */
4371 	if (!task_sleep)
4372 		return;
4373 
4374 	/*
4375 	 * If the PELT values haven't changed since enqueue time,
4376 	 * skip the util_est update.
4377 	 */
4378 	ue = p->se.avg.util_est;
4379 	if (ue.enqueued & UTIL_AVG_UNCHANGED)
4380 		return;
4381 
4382 	last_enqueued_diff = ue.enqueued;
4383 
4384 	/*
4385 	 * Reset EWMA on utilization increases, the moving average is used only
4386 	 * to smooth utilization decreases.
4387 	 */
4388 	ue.enqueued = task_util(p);
4389 	if (sched_feat(UTIL_EST_FASTUP)) {
4390 		if (ue.ewma < ue.enqueued) {
4391 			ue.ewma = ue.enqueued;
4392 			goto done;
4393 		}
4394 	}
4395 
4396 	/*
4397 	 * Skip update of task's estimated utilization when its members are
4398 	 * already ~1% close to its last activation value.
4399 	 */
4400 	last_ewma_diff = ue.enqueued - ue.ewma;
4401 	last_enqueued_diff -= ue.enqueued;
4402 	if (within_margin(last_ewma_diff, UTIL_EST_MARGIN)) {
4403 		if (!within_margin(last_enqueued_diff, UTIL_EST_MARGIN))
4404 			goto done;
4405 
4406 		return;
4407 	}
4408 
4409 	/*
4410 	 * To avoid overestimation of actual task utilization, skip updates if
4411 	 * we cannot grant there is idle time in this CPU.
4412 	 */
4413 	if (task_util(p) > capacity_orig_of(cpu_of(rq_of(cfs_rq))))
4414 		return;
4415 
4416 	/*
4417 	 * Update Task's estimated utilization
4418 	 *
4419 	 * When *p completes an activation we can consolidate another sample
4420 	 * of the task size. This is done by storing the current PELT value
4421 	 * as ue.enqueued and by using this value to update the Exponential
4422 	 * Weighted Moving Average (EWMA):
4423 	 *
4424 	 *  ewma(t) = w *  task_util(p) + (1-w) * ewma(t-1)
4425 	 *          = w *  task_util(p) +         ewma(t-1)  - w * ewma(t-1)
4426 	 *          = w * (task_util(p) -         ewma(t-1)) +     ewma(t-1)
4427 	 *          = w * (      last_ewma_diff            ) +     ewma(t-1)
4428 	 *          = w * (last_ewma_diff  +  ewma(t-1) / w)
4429 	 *
4430 	 * Where 'w' is the weight of new samples, which is configured to be
4431 	 * 0.25, thus making w=1/4 ( >>= UTIL_EST_WEIGHT_SHIFT)
4432 	 */
4433 	ue.ewma <<= UTIL_EST_WEIGHT_SHIFT;
4434 	ue.ewma  += last_ewma_diff;
4435 	ue.ewma >>= UTIL_EST_WEIGHT_SHIFT;
4436 done:
4437 	ue.enqueued |= UTIL_AVG_UNCHANGED;
4438 	WRITE_ONCE(p->se.avg.util_est, ue);
4439 
4440 	trace_sched_util_est_se_tp(&p->se);
4441 }
4442 
4443 static inline int util_fits_cpu(unsigned long util,
4444 				unsigned long uclamp_min,
4445 				unsigned long uclamp_max,
4446 				int cpu)
4447 {
4448 	unsigned long capacity_orig, capacity_orig_thermal;
4449 	unsigned long capacity = capacity_of(cpu);
4450 	bool fits, uclamp_max_fits;
4451 
4452 	/*
4453 	 * Check if the real util fits without any uclamp boost/cap applied.
4454 	 */
4455 	fits = fits_capacity(util, capacity);
4456 
4457 	if (!uclamp_is_used())
4458 		return fits;
4459 
4460 	/*
4461 	 * We must use capacity_orig_of() for comparing against uclamp_min and
4462 	 * uclamp_max. We only care about capacity pressure (by using
4463 	 * capacity_of()) for comparing against the real util.
4464 	 *
4465 	 * If a task is boosted to 1024 for example, we don't want a tiny
4466 	 * pressure to skew the check whether it fits a CPU or not.
4467 	 *
4468 	 * Similarly if a task is capped to capacity_orig_of(little_cpu), it
4469 	 * should fit a little cpu even if there's some pressure.
4470 	 *
4471 	 * Only exception is for thermal pressure since it has a direct impact
4472 	 * on available OPP of the system.
4473 	 *
4474 	 * We honour it for uclamp_min only as a drop in performance level
4475 	 * could result in not getting the requested minimum performance level.
4476 	 *
4477 	 * For uclamp_max, we can tolerate a drop in performance level as the
4478 	 * goal is to cap the task. So it's okay if it's getting less.
4479 	 */
4480 	capacity_orig = capacity_orig_of(cpu);
4481 	capacity_orig_thermal = capacity_orig - arch_scale_thermal_pressure(cpu);
4482 
4483 	/*
4484 	 * We want to force a task to fit a cpu as implied by uclamp_max.
4485 	 * But we do have some corner cases to cater for..
4486 	 *
4487 	 *
4488 	 *                                 C=z
4489 	 *   |                             ___
4490 	 *   |                  C=y       |   |
4491 	 *   |_ _ _ _ _ _ _ _ _ ___ _ _ _ | _ | _ _ _ _ _  uclamp_max
4492 	 *   |      C=x        |   |      |   |
4493 	 *   |      ___        |   |      |   |
4494 	 *   |     |   |       |   |      |   |    (util somewhere in this region)
4495 	 *   |     |   |       |   |      |   |
4496 	 *   |     |   |       |   |      |   |
4497 	 *   +----------------------------------------
4498 	 *         cpu0        cpu1       cpu2
4499 	 *
4500 	 *   In the above example if a task is capped to a specific performance
4501 	 *   point, y, then when:
4502 	 *
4503 	 *   * util = 80% of x then it does not fit on cpu0 and should migrate
4504 	 *     to cpu1
4505 	 *   * util = 80% of y then it is forced to fit on cpu1 to honour
4506 	 *     uclamp_max request.
4507 	 *
4508 	 *   which is what we're enforcing here. A task always fits if
4509 	 *   uclamp_max <= capacity_orig. But when uclamp_max > capacity_orig,
4510 	 *   the normal upmigration rules should withhold still.
4511 	 *
4512 	 *   Only exception is when we are on max capacity, then we need to be
4513 	 *   careful not to block overutilized state. This is so because:
4514 	 *
4515 	 *     1. There's no concept of capping at max_capacity! We can't go
4516 	 *        beyond this performance level anyway.
4517 	 *     2. The system is being saturated when we're operating near
4518 	 *        max capacity, it doesn't make sense to block overutilized.
4519 	 */
4520 	uclamp_max_fits = (capacity_orig == SCHED_CAPACITY_SCALE) && (uclamp_max == SCHED_CAPACITY_SCALE);
4521 	uclamp_max_fits = !uclamp_max_fits && (uclamp_max <= capacity_orig);
4522 	fits = fits || uclamp_max_fits;
4523 
4524 	/*
4525 	 *
4526 	 *                                 C=z
4527 	 *   |                             ___       (region a, capped, util >= uclamp_max)
4528 	 *   |                  C=y       |   |
4529 	 *   |_ _ _ _ _ _ _ _ _ ___ _ _ _ | _ | _ _ _ _ _ uclamp_max
4530 	 *   |      C=x        |   |      |   |
4531 	 *   |      ___        |   |      |   |      (region b, uclamp_min <= util <= uclamp_max)
4532 	 *   |_ _ _|_ _|_ _ _ _| _ | _ _ _| _ | _ _ _ _ _ uclamp_min
4533 	 *   |     |   |       |   |      |   |
4534 	 *   |     |   |       |   |      |   |      (region c, boosted, util < uclamp_min)
4535 	 *   +----------------------------------------
4536 	 *         cpu0        cpu1       cpu2
4537 	 *
4538 	 * a) If util > uclamp_max, then we're capped, we don't care about
4539 	 *    actual fitness value here. We only care if uclamp_max fits
4540 	 *    capacity without taking margin/pressure into account.
4541 	 *    See comment above.
4542 	 *
4543 	 * b) If uclamp_min <= util <= uclamp_max, then the normal
4544 	 *    fits_capacity() rules apply. Except we need to ensure that we
4545 	 *    enforce we remain within uclamp_max, see comment above.
4546 	 *
4547 	 * c) If util < uclamp_min, then we are boosted. Same as (b) but we
4548 	 *    need to take into account the boosted value fits the CPU without
4549 	 *    taking margin/pressure into account.
4550 	 *
4551 	 * Cases (a) and (b) are handled in the 'fits' variable already. We
4552 	 * just need to consider an extra check for case (c) after ensuring we
4553 	 * handle the case uclamp_min > uclamp_max.
4554 	 */
4555 	uclamp_min = min(uclamp_min, uclamp_max);
4556 	if (fits && (util < uclamp_min) && (uclamp_min > capacity_orig_thermal))
4557 		return -1;
4558 
4559 	return fits;
4560 }
4561 
4562 static inline int task_fits_cpu(struct task_struct *p, int cpu)
4563 {
4564 	unsigned long uclamp_min = uclamp_eff_value(p, UCLAMP_MIN);
4565 	unsigned long uclamp_max = uclamp_eff_value(p, UCLAMP_MAX);
4566 	unsigned long util = task_util_est(p);
4567 	/*
4568 	 * Return true only if the cpu fully fits the task requirements, which
4569 	 * include the utilization but also the performance hints.
4570 	 */
4571 	return (util_fits_cpu(util, uclamp_min, uclamp_max, cpu) > 0);
4572 }
4573 
4574 static inline void update_misfit_status(struct task_struct *p, struct rq *rq)
4575 {
4576 	if (!sched_asym_cpucap_active())
4577 		return;
4578 
4579 	if (!p || p->nr_cpus_allowed == 1) {
4580 		rq->misfit_task_load = 0;
4581 		return;
4582 	}
4583 
4584 	if (task_fits_cpu(p, cpu_of(rq))) {
4585 		rq->misfit_task_load = 0;
4586 		return;
4587 	}
4588 
4589 	/*
4590 	 * Make sure that misfit_task_load will not be null even if
4591 	 * task_h_load() returns 0.
4592 	 */
4593 	rq->misfit_task_load = max_t(unsigned long, task_h_load(p), 1);
4594 }
4595 
4596 #else /* CONFIG_SMP */
4597 
4598 static inline bool cfs_rq_is_decayed(struct cfs_rq *cfs_rq)
4599 {
4600 	return true;
4601 }
4602 
4603 #define UPDATE_TG	0x0
4604 #define SKIP_AGE_LOAD	0x0
4605 #define DO_ATTACH	0x0
4606 #define DO_DETACH	0x0
4607 
4608 static inline void update_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int not_used1)
4609 {
4610 	cfs_rq_util_change(cfs_rq, 0);
4611 }
4612 
4613 static inline void remove_entity_load_avg(struct sched_entity *se) {}
4614 
4615 static inline void
4616 attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
4617 static inline void
4618 detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se) {}
4619 
4620 static inline int newidle_balance(struct rq *rq, struct rq_flags *rf)
4621 {
4622 	return 0;
4623 }
4624 
4625 static inline void
4626 util_est_enqueue(struct cfs_rq *cfs_rq, struct task_struct *p) {}
4627 
4628 static inline void
4629 util_est_dequeue(struct cfs_rq *cfs_rq, struct task_struct *p) {}
4630 
4631 static inline void
4632 util_est_update(struct cfs_rq *cfs_rq, struct task_struct *p,
4633 		bool task_sleep) {}
4634 static inline void update_misfit_status(struct task_struct *p, struct rq *rq) {}
4635 
4636 #endif /* CONFIG_SMP */
4637 
4638 static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
4639 {
4640 #ifdef CONFIG_SCHED_DEBUG
4641 	s64 d = se->vruntime - cfs_rq->min_vruntime;
4642 
4643 	if (d < 0)
4644 		d = -d;
4645 
4646 	if (d > 3*sysctl_sched_latency)
4647 		schedstat_inc(cfs_rq->nr_spread_over);
4648 #endif
4649 }
4650 
4651 static void
4652 place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
4653 {
4654 	u64 vruntime = cfs_rq->min_vruntime;
4655 	u64 sleep_time;
4656 
4657 	/*
4658 	 * The 'current' period is already promised to the current tasks,
4659 	 * however the extra weight of the new task will slow them down a
4660 	 * little, place the new task so that it fits in the slot that
4661 	 * stays open at the end.
4662 	 */
4663 	if (initial && sched_feat(START_DEBIT))
4664 		vruntime += sched_vslice(cfs_rq, se);
4665 
4666 	/* sleeps up to a single latency don't count. */
4667 	if (!initial) {
4668 		unsigned long thresh;
4669 
4670 		if (se_is_idle(se))
4671 			thresh = sysctl_sched_min_granularity;
4672 		else
4673 			thresh = sysctl_sched_latency;
4674 
4675 		/*
4676 		 * Halve their sleep time's effect, to allow
4677 		 * for a gentler effect of sleepers:
4678 		 */
4679 		if (sched_feat(GENTLE_FAIR_SLEEPERS))
4680 			thresh >>= 1;
4681 
4682 		vruntime -= thresh;
4683 	}
4684 
4685 	/*
4686 	 * Pull vruntime of the entity being placed to the base level of
4687 	 * cfs_rq, to prevent boosting it if placed backwards.  If the entity
4688 	 * slept for a long time, don't even try to compare its vruntime with
4689 	 * the base as it may be too far off and the comparison may get
4690 	 * inversed due to s64 overflow.
4691 	 */
4692 	sleep_time = rq_clock_task(rq_of(cfs_rq)) - se->exec_start;
4693 	if ((s64)sleep_time > 60LL * NSEC_PER_SEC)
4694 		se->vruntime = vruntime;
4695 	else
4696 		se->vruntime = max_vruntime(se->vruntime, vruntime);
4697 }
4698 
4699 static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
4700 
4701 static inline bool cfs_bandwidth_used(void);
4702 
4703 /*
4704  * MIGRATION
4705  *
4706  *	dequeue
4707  *	  update_curr()
4708  *	    update_min_vruntime()
4709  *	  vruntime -= min_vruntime
4710  *
4711  *	enqueue
4712  *	  update_curr()
4713  *	    update_min_vruntime()
4714  *	  vruntime += min_vruntime
4715  *
4716  * this way the vruntime transition between RQs is done when both
4717  * min_vruntime are up-to-date.
4718  *
4719  * WAKEUP (remote)
4720  *
4721  *	->migrate_task_rq_fair() (p->state == TASK_WAKING)
4722  *	  vruntime -= min_vruntime
4723  *
4724  *	enqueue
4725  *	  update_curr()
4726  *	    update_min_vruntime()
4727  *	  vruntime += min_vruntime
4728  *
4729  * this way we don't have the most up-to-date min_vruntime on the originating
4730  * CPU and an up-to-date min_vruntime on the destination CPU.
4731  */
4732 
4733 static void
4734 enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
4735 {
4736 	bool renorm = !(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_MIGRATED);
4737 	bool curr = cfs_rq->curr == se;
4738 
4739 	/*
4740 	 * If we're the current task, we must renormalise before calling
4741 	 * update_curr().
4742 	 */
4743 	if (renorm && curr)
4744 		se->vruntime += cfs_rq->min_vruntime;
4745 
4746 	update_curr(cfs_rq);
4747 
4748 	/*
4749 	 * Otherwise, renormalise after, such that we're placed at the current
4750 	 * moment in time, instead of some random moment in the past. Being
4751 	 * placed in the past could significantly boost this task to the
4752 	 * fairness detriment of existing tasks.
4753 	 */
4754 	if (renorm && !curr)
4755 		se->vruntime += cfs_rq->min_vruntime;
4756 
4757 	/*
4758 	 * When enqueuing a sched_entity, we must:
4759 	 *   - Update loads to have both entity and cfs_rq synced with now.
4760 	 *   - For group_entity, update its runnable_weight to reflect the new
4761 	 *     h_nr_running of its group cfs_rq.
4762 	 *   - For group_entity, update its weight to reflect the new share of
4763 	 *     its group cfs_rq
4764 	 *   - Add its new weight to cfs_rq->load.weight
4765 	 */
4766 	update_load_avg(cfs_rq, se, UPDATE_TG | DO_ATTACH);
4767 	se_update_runnable(se);
4768 	update_cfs_group(se);
4769 	account_entity_enqueue(cfs_rq, se);
4770 
4771 	if (flags & ENQUEUE_WAKEUP)
4772 		place_entity(cfs_rq, se, 0);
4773 
4774 	check_schedstat_required();
4775 	update_stats_enqueue_fair(cfs_rq, se, flags);
4776 	check_spread(cfs_rq, se);
4777 	if (!curr)
4778 		__enqueue_entity(cfs_rq, se);
4779 	se->on_rq = 1;
4780 
4781 	if (cfs_rq->nr_running == 1) {
4782 		check_enqueue_throttle(cfs_rq);
4783 		if (!throttled_hierarchy(cfs_rq))
4784 			list_add_leaf_cfs_rq(cfs_rq);
4785 	}
4786 }
4787 
4788 static void __clear_buddies_last(struct sched_entity *se)
4789 {
4790 	for_each_sched_entity(se) {
4791 		struct cfs_rq *cfs_rq = cfs_rq_of(se);
4792 		if (cfs_rq->last != se)
4793 			break;
4794 
4795 		cfs_rq->last = NULL;
4796 	}
4797 }
4798 
4799 static void __clear_buddies_next(struct sched_entity *se)
4800 {
4801 	for_each_sched_entity(se) {
4802 		struct cfs_rq *cfs_rq = cfs_rq_of(se);
4803 		if (cfs_rq->next != se)
4804 			break;
4805 
4806 		cfs_rq->next = NULL;
4807 	}
4808 }
4809 
4810 static void __clear_buddies_skip(struct sched_entity *se)
4811 {
4812 	for_each_sched_entity(se) {
4813 		struct cfs_rq *cfs_rq = cfs_rq_of(se);
4814 		if (cfs_rq->skip != se)
4815 			break;
4816 
4817 		cfs_rq->skip = NULL;
4818 	}
4819 }
4820 
4821 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
4822 {
4823 	if (cfs_rq->last == se)
4824 		__clear_buddies_last(se);
4825 
4826 	if (cfs_rq->next == se)
4827 		__clear_buddies_next(se);
4828 
4829 	if (cfs_rq->skip == se)
4830 		__clear_buddies_skip(se);
4831 }
4832 
4833 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq);
4834 
4835 static void
4836 dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
4837 {
4838 	int action = UPDATE_TG;
4839 
4840 	if (entity_is_task(se) && task_on_rq_migrating(task_of(se)))
4841 		action |= DO_DETACH;
4842 
4843 	/*
4844 	 * Update run-time statistics of the 'current'.
4845 	 */
4846 	update_curr(cfs_rq);
4847 
4848 	/*
4849 	 * When dequeuing a sched_entity, we must:
4850 	 *   - Update loads to have both entity and cfs_rq synced with now.
4851 	 *   - For group_entity, update its runnable_weight to reflect the new
4852 	 *     h_nr_running of its group cfs_rq.
4853 	 *   - Subtract its previous weight from cfs_rq->load.weight.
4854 	 *   - For group entity, update its weight to reflect the new share
4855 	 *     of its group cfs_rq.
4856 	 */
4857 	update_load_avg(cfs_rq, se, action);
4858 	se_update_runnable(se);
4859 
4860 	update_stats_dequeue_fair(cfs_rq, se, flags);
4861 
4862 	clear_buddies(cfs_rq, se);
4863 
4864 	if (se != cfs_rq->curr)
4865 		__dequeue_entity(cfs_rq, se);
4866 	se->on_rq = 0;
4867 	account_entity_dequeue(cfs_rq, se);
4868 
4869 	/*
4870 	 * Normalize after update_curr(); which will also have moved
4871 	 * min_vruntime if @se is the one holding it back. But before doing
4872 	 * update_min_vruntime() again, which will discount @se's position and
4873 	 * can move min_vruntime forward still more.
4874 	 */
4875 	if (!(flags & DEQUEUE_SLEEP))
4876 		se->vruntime -= cfs_rq->min_vruntime;
4877 
4878 	/* return excess runtime on last dequeue */
4879 	return_cfs_rq_runtime(cfs_rq);
4880 
4881 	update_cfs_group(se);
4882 
4883 	/*
4884 	 * Now advance min_vruntime if @se was the entity holding it back,
4885 	 * except when: DEQUEUE_SAVE && !DEQUEUE_MOVE, in this case we'll be
4886 	 * put back on, and if we advance min_vruntime, we'll be placed back
4887 	 * further than we started -- ie. we'll be penalized.
4888 	 */
4889 	if ((flags & (DEQUEUE_SAVE | DEQUEUE_MOVE)) != DEQUEUE_SAVE)
4890 		update_min_vruntime(cfs_rq);
4891 
4892 	if (cfs_rq->nr_running == 0)
4893 		update_idle_cfs_rq_clock_pelt(cfs_rq);
4894 }
4895 
4896 /*
4897  * Preempt the current task with a newly woken task if needed:
4898  */
4899 static void
4900 check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
4901 {
4902 	unsigned long ideal_runtime, delta_exec;
4903 	struct sched_entity *se;
4904 	s64 delta;
4905 
4906 	/*
4907 	 * When many tasks blow up the sched_period; it is possible that
4908 	 * sched_slice() reports unusually large results (when many tasks are
4909 	 * very light for example). Therefore impose a maximum.
4910 	 */
4911 	ideal_runtime = min_t(u64, sched_slice(cfs_rq, curr), sysctl_sched_latency);
4912 
4913 	delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
4914 	if (delta_exec > ideal_runtime) {
4915 		resched_curr(rq_of(cfs_rq));
4916 		/*
4917 		 * The current task ran long enough, ensure it doesn't get
4918 		 * re-elected due to buddy favours.
4919 		 */
4920 		clear_buddies(cfs_rq, curr);
4921 		return;
4922 	}
4923 
4924 	/*
4925 	 * Ensure that a task that missed wakeup preemption by a
4926 	 * narrow margin doesn't have to wait for a full slice.
4927 	 * This also mitigates buddy induced latencies under load.
4928 	 */
4929 	if (delta_exec < sysctl_sched_min_granularity)
4930 		return;
4931 
4932 	se = __pick_first_entity(cfs_rq);
4933 	delta = curr->vruntime - se->vruntime;
4934 
4935 	if (delta < 0)
4936 		return;
4937 
4938 	if (delta > ideal_runtime)
4939 		resched_curr(rq_of(cfs_rq));
4940 }
4941 
4942 static void
4943 set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
4944 {
4945 	clear_buddies(cfs_rq, se);
4946 
4947 	/* 'current' is not kept within the tree. */
4948 	if (se->on_rq) {
4949 		/*
4950 		 * Any task has to be enqueued before it get to execute on
4951 		 * a CPU. So account for the time it spent waiting on the
4952 		 * runqueue.
4953 		 */
4954 		update_stats_wait_end_fair(cfs_rq, se);
4955 		__dequeue_entity(cfs_rq, se);
4956 		update_load_avg(cfs_rq, se, UPDATE_TG);
4957 	}
4958 
4959 	update_stats_curr_start(cfs_rq, se);
4960 	cfs_rq->curr = se;
4961 
4962 	/*
4963 	 * Track our maximum slice length, if the CPU's load is at
4964 	 * least twice that of our own weight (i.e. dont track it
4965 	 * when there are only lesser-weight tasks around):
4966 	 */
4967 	if (schedstat_enabled() &&
4968 	    rq_of(cfs_rq)->cfs.load.weight >= 2*se->load.weight) {
4969 		struct sched_statistics *stats;
4970 
4971 		stats = __schedstats_from_se(se);
4972 		__schedstat_set(stats->slice_max,
4973 				max((u64)stats->slice_max,
4974 				    se->sum_exec_runtime - se->prev_sum_exec_runtime));
4975 	}
4976 
4977 	se->prev_sum_exec_runtime = se->sum_exec_runtime;
4978 }
4979 
4980 static int
4981 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
4982 
4983 /*
4984  * Pick the next process, keeping these things in mind, in this order:
4985  * 1) keep things fair between processes/task groups
4986  * 2) pick the "next" process, since someone really wants that to run
4987  * 3) pick the "last" process, for cache locality
4988  * 4) do not run the "skip" process, if something else is available
4989  */
4990 static struct sched_entity *
4991 pick_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *curr)
4992 {
4993 	struct sched_entity *left = __pick_first_entity(cfs_rq);
4994 	struct sched_entity *se;
4995 
4996 	/*
4997 	 * If curr is set we have to see if its left of the leftmost entity
4998 	 * still in the tree, provided there was anything in the tree at all.
4999 	 */
5000 	if (!left || (curr && entity_before(curr, left)))
5001 		left = curr;
5002 
5003 	se = left; /* ideally we run the leftmost entity */
5004 
5005 	/*
5006 	 * Avoid running the skip buddy, if running something else can
5007 	 * be done without getting too unfair.
5008 	 */
5009 	if (cfs_rq->skip && cfs_rq->skip == se) {
5010 		struct sched_entity *second;
5011 
5012 		if (se == curr) {
5013 			second = __pick_first_entity(cfs_rq);
5014 		} else {
5015 			second = __pick_next_entity(se);
5016 			if (!second || (curr && entity_before(curr, second)))
5017 				second = curr;
5018 		}
5019 
5020 		if (second && wakeup_preempt_entity(second, left) < 1)
5021 			se = second;
5022 	}
5023 
5024 	if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1) {
5025 		/*
5026 		 * Someone really wants this to run. If it's not unfair, run it.
5027 		 */
5028 		se = cfs_rq->next;
5029 	} else if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1) {
5030 		/*
5031 		 * Prefer last buddy, try to return the CPU to a preempted task.
5032 		 */
5033 		se = cfs_rq->last;
5034 	}
5035 
5036 	return se;
5037 }
5038 
5039 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq);
5040 
5041 static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
5042 {
5043 	/*
5044 	 * If still on the runqueue then deactivate_task()
5045 	 * was not called and update_curr() has to be done:
5046 	 */
5047 	if (prev->on_rq)
5048 		update_curr(cfs_rq);
5049 
5050 	/* throttle cfs_rqs exceeding runtime */
5051 	check_cfs_rq_runtime(cfs_rq);
5052 
5053 	check_spread(cfs_rq, prev);
5054 
5055 	if (prev->on_rq) {
5056 		update_stats_wait_start_fair(cfs_rq, prev);
5057 		/* Put 'current' back into the tree. */
5058 		__enqueue_entity(cfs_rq, prev);
5059 		/* in !on_rq case, update occurred at dequeue */
5060 		update_load_avg(cfs_rq, prev, 0);
5061 	}
5062 	cfs_rq->curr = NULL;
5063 }
5064 
5065 static void
5066 entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
5067 {
5068 	/*
5069 	 * Update run-time statistics of the 'current'.
5070 	 */
5071 	update_curr(cfs_rq);
5072 
5073 	/*
5074 	 * Ensure that runnable average is periodically updated.
5075 	 */
5076 	update_load_avg(cfs_rq, curr, UPDATE_TG);
5077 	update_cfs_group(curr);
5078 
5079 #ifdef CONFIG_SCHED_HRTICK
5080 	/*
5081 	 * queued ticks are scheduled to match the slice, so don't bother
5082 	 * validating it and just reschedule.
5083 	 */
5084 	if (queued) {
5085 		resched_curr(rq_of(cfs_rq));
5086 		return;
5087 	}
5088 	/*
5089 	 * don't let the period tick interfere with the hrtick preemption
5090 	 */
5091 	if (!sched_feat(DOUBLE_TICK) &&
5092 			hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
5093 		return;
5094 #endif
5095 
5096 	if (cfs_rq->nr_running > 1)
5097 		check_preempt_tick(cfs_rq, curr);
5098 }
5099 
5100 
5101 /**************************************************
5102  * CFS bandwidth control machinery
5103  */
5104 
5105 #ifdef CONFIG_CFS_BANDWIDTH
5106 
5107 #ifdef CONFIG_JUMP_LABEL
5108 static struct static_key __cfs_bandwidth_used;
5109 
5110 static inline bool cfs_bandwidth_used(void)
5111 {
5112 	return static_key_false(&__cfs_bandwidth_used);
5113 }
5114 
5115 void cfs_bandwidth_usage_inc(void)
5116 {
5117 	static_key_slow_inc_cpuslocked(&__cfs_bandwidth_used);
5118 }
5119 
5120 void cfs_bandwidth_usage_dec(void)
5121 {
5122 	static_key_slow_dec_cpuslocked(&__cfs_bandwidth_used);
5123 }
5124 #else /* CONFIG_JUMP_LABEL */
5125 static bool cfs_bandwidth_used(void)
5126 {
5127 	return true;
5128 }
5129 
5130 void cfs_bandwidth_usage_inc(void) {}
5131 void cfs_bandwidth_usage_dec(void) {}
5132 #endif /* CONFIG_JUMP_LABEL */
5133 
5134 /*
5135  * default period for cfs group bandwidth.
5136  * default: 0.1s, units: nanoseconds
5137  */
5138 static inline u64 default_cfs_period(void)
5139 {
5140 	return 100000000ULL;
5141 }
5142 
5143 static inline u64 sched_cfs_bandwidth_slice(void)
5144 {
5145 	return (u64)sysctl_sched_cfs_bandwidth_slice * NSEC_PER_USEC;
5146 }
5147 
5148 /*
5149  * Replenish runtime according to assigned quota. We use sched_clock_cpu
5150  * directly instead of rq->clock to avoid adding additional synchronization
5151  * around rq->lock.
5152  *
5153  * requires cfs_b->lock
5154  */
5155 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
5156 {
5157 	s64 runtime;
5158 
5159 	if (unlikely(cfs_b->quota == RUNTIME_INF))
5160 		return;
5161 
5162 	cfs_b->runtime += cfs_b->quota;
5163 	runtime = cfs_b->runtime_snap - cfs_b->runtime;
5164 	if (runtime > 0) {
5165 		cfs_b->burst_time += runtime;
5166 		cfs_b->nr_burst++;
5167 	}
5168 
5169 	cfs_b->runtime = min(cfs_b->runtime, cfs_b->quota + cfs_b->burst);
5170 	cfs_b->runtime_snap = cfs_b->runtime;
5171 }
5172 
5173 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
5174 {
5175 	return &tg->cfs_bandwidth;
5176 }
5177 
5178 /* returns 0 on failure to allocate runtime */
5179 static int __assign_cfs_rq_runtime(struct cfs_bandwidth *cfs_b,
5180 				   struct cfs_rq *cfs_rq, u64 target_runtime)
5181 {
5182 	u64 min_amount, amount = 0;
5183 
5184 	lockdep_assert_held(&cfs_b->lock);
5185 
5186 	/* note: this is a positive sum as runtime_remaining <= 0 */
5187 	min_amount = target_runtime - cfs_rq->runtime_remaining;
5188 
5189 	if (cfs_b->quota == RUNTIME_INF)
5190 		amount = min_amount;
5191 	else {
5192 		start_cfs_bandwidth(cfs_b);
5193 
5194 		if (cfs_b->runtime > 0) {
5195 			amount = min(cfs_b->runtime, min_amount);
5196 			cfs_b->runtime -= amount;
5197 			cfs_b->idle = 0;
5198 		}
5199 	}
5200 
5201 	cfs_rq->runtime_remaining += amount;
5202 
5203 	return cfs_rq->runtime_remaining > 0;
5204 }
5205 
5206 /* returns 0 on failure to allocate runtime */
5207 static int assign_cfs_rq_runtime(struct cfs_rq *cfs_rq)
5208 {
5209 	struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
5210 	int ret;
5211 
5212 	raw_spin_lock(&cfs_b->lock);
5213 	ret = __assign_cfs_rq_runtime(cfs_b, cfs_rq, sched_cfs_bandwidth_slice());
5214 	raw_spin_unlock(&cfs_b->lock);
5215 
5216 	return ret;
5217 }
5218 
5219 static void __account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
5220 {
5221 	/* dock delta_exec before expiring quota (as it could span periods) */
5222 	cfs_rq->runtime_remaining -= delta_exec;
5223 
5224 	if (likely(cfs_rq->runtime_remaining > 0))
5225 		return;
5226 
5227 	if (cfs_rq->throttled)
5228 		return;
5229 	/*
5230 	 * if we're unable to extend our runtime we resched so that the active
5231 	 * hierarchy can be throttled
5232 	 */
5233 	if (!assign_cfs_rq_runtime(cfs_rq) && likely(cfs_rq->curr))
5234 		resched_curr(rq_of(cfs_rq));
5235 }
5236 
5237 static __always_inline
5238 void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec)
5239 {
5240 	if (!cfs_bandwidth_used() || !cfs_rq->runtime_enabled)
5241 		return;
5242 
5243 	__account_cfs_rq_runtime(cfs_rq, delta_exec);
5244 }
5245 
5246 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
5247 {
5248 	return cfs_bandwidth_used() && cfs_rq->throttled;
5249 }
5250 
5251 /* check whether cfs_rq, or any parent, is throttled */
5252 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
5253 {
5254 	return cfs_bandwidth_used() && cfs_rq->throttle_count;
5255 }
5256 
5257 /*
5258  * Ensure that neither of the group entities corresponding to src_cpu or
5259  * dest_cpu are members of a throttled hierarchy when performing group
5260  * load-balance operations.
5261  */
5262 static inline int throttled_lb_pair(struct task_group *tg,
5263 				    int src_cpu, int dest_cpu)
5264 {
5265 	struct cfs_rq *src_cfs_rq, *dest_cfs_rq;
5266 
5267 	src_cfs_rq = tg->cfs_rq[src_cpu];
5268 	dest_cfs_rq = tg->cfs_rq[dest_cpu];
5269 
5270 	return throttled_hierarchy(src_cfs_rq) ||
5271 	       throttled_hierarchy(dest_cfs_rq);
5272 }
5273 
5274 static int tg_unthrottle_up(struct task_group *tg, void *data)
5275 {
5276 	struct rq *rq = data;
5277 	struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
5278 
5279 	cfs_rq->throttle_count--;
5280 	if (!cfs_rq->throttle_count) {
5281 		cfs_rq->throttled_clock_pelt_time += rq_clock_pelt(rq) -
5282 					     cfs_rq->throttled_clock_pelt;
5283 
5284 		/* Add cfs_rq with load or one or more already running entities to the list */
5285 		if (!cfs_rq_is_decayed(cfs_rq))
5286 			list_add_leaf_cfs_rq(cfs_rq);
5287 	}
5288 
5289 	return 0;
5290 }
5291 
5292 static int tg_throttle_down(struct task_group *tg, void *data)
5293 {
5294 	struct rq *rq = data;
5295 	struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
5296 
5297 	/* group is entering throttled state, stop time */
5298 	if (!cfs_rq->throttle_count) {
5299 		cfs_rq->throttled_clock_pelt = rq_clock_pelt(rq);
5300 		list_del_leaf_cfs_rq(cfs_rq);
5301 	}
5302 	cfs_rq->throttle_count++;
5303 
5304 	return 0;
5305 }
5306 
5307 static bool throttle_cfs_rq(struct cfs_rq *cfs_rq)
5308 {
5309 	struct rq *rq = rq_of(cfs_rq);
5310 	struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
5311 	struct sched_entity *se;
5312 	long task_delta, idle_task_delta, dequeue = 1;
5313 
5314 	raw_spin_lock(&cfs_b->lock);
5315 	/* This will start the period timer if necessary */
5316 	if (__assign_cfs_rq_runtime(cfs_b, cfs_rq, 1)) {
5317 		/*
5318 		 * We have raced with bandwidth becoming available, and if we
5319 		 * actually throttled the timer might not unthrottle us for an
5320 		 * entire period. We additionally needed to make sure that any
5321 		 * subsequent check_cfs_rq_runtime calls agree not to throttle
5322 		 * us, as we may commit to do cfs put_prev+pick_next, so we ask
5323 		 * for 1ns of runtime rather than just check cfs_b.
5324 		 */
5325 		dequeue = 0;
5326 	} else {
5327 		list_add_tail_rcu(&cfs_rq->throttled_list,
5328 				  &cfs_b->throttled_cfs_rq);
5329 	}
5330 	raw_spin_unlock(&cfs_b->lock);
5331 
5332 	if (!dequeue)
5333 		return false;  /* Throttle no longer required. */
5334 
5335 	se = cfs_rq->tg->se[cpu_of(rq_of(cfs_rq))];
5336 
5337 	/* freeze hierarchy runnable averages while throttled */
5338 	rcu_read_lock();
5339 	walk_tg_tree_from(cfs_rq->tg, tg_throttle_down, tg_nop, (void *)rq);
5340 	rcu_read_unlock();
5341 
5342 	task_delta = cfs_rq->h_nr_running;
5343 	idle_task_delta = cfs_rq->idle_h_nr_running;
5344 	for_each_sched_entity(se) {
5345 		struct cfs_rq *qcfs_rq = cfs_rq_of(se);
5346 		/* throttled entity or throttle-on-deactivate */
5347 		if (!se->on_rq)
5348 			goto done;
5349 
5350 		dequeue_entity(qcfs_rq, se, DEQUEUE_SLEEP);
5351 
5352 		if (cfs_rq_is_idle(group_cfs_rq(se)))
5353 			idle_task_delta = cfs_rq->h_nr_running;
5354 
5355 		qcfs_rq->h_nr_running -= task_delta;
5356 		qcfs_rq->idle_h_nr_running -= idle_task_delta;
5357 
5358 		if (qcfs_rq->load.weight) {
5359 			/* Avoid re-evaluating load for this entity: */
5360 			se = parent_entity(se);
5361 			break;
5362 		}
5363 	}
5364 
5365 	for_each_sched_entity(se) {
5366 		struct cfs_rq *qcfs_rq = cfs_rq_of(se);
5367 		/* throttled entity or throttle-on-deactivate */
5368 		if (!se->on_rq)
5369 			goto done;
5370 
5371 		update_load_avg(qcfs_rq, se, 0);
5372 		se_update_runnable(se);
5373 
5374 		if (cfs_rq_is_idle(group_cfs_rq(se)))
5375 			idle_task_delta = cfs_rq->h_nr_running;
5376 
5377 		qcfs_rq->h_nr_running -= task_delta;
5378 		qcfs_rq->idle_h_nr_running -= idle_task_delta;
5379 	}
5380 
5381 	/* At this point se is NULL and we are at root level*/
5382 	sub_nr_running(rq, task_delta);
5383 
5384 done:
5385 	/*
5386 	 * Note: distribution will already see us throttled via the
5387 	 * throttled-list.  rq->lock protects completion.
5388 	 */
5389 	cfs_rq->throttled = 1;
5390 	cfs_rq->throttled_clock = rq_clock(rq);
5391 	return true;
5392 }
5393 
5394 void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
5395 {
5396 	struct rq *rq = rq_of(cfs_rq);
5397 	struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
5398 	struct sched_entity *se;
5399 	long task_delta, idle_task_delta;
5400 
5401 	se = cfs_rq->tg->se[cpu_of(rq)];
5402 
5403 	cfs_rq->throttled = 0;
5404 
5405 	update_rq_clock(rq);
5406 
5407 	raw_spin_lock(&cfs_b->lock);
5408 	cfs_b->throttled_time += rq_clock(rq) - cfs_rq->throttled_clock;
5409 	list_del_rcu(&cfs_rq->throttled_list);
5410 	raw_spin_unlock(&cfs_b->lock);
5411 
5412 	/* update hierarchical throttle state */
5413 	walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
5414 
5415 	if (!cfs_rq->load.weight) {
5416 		if (!cfs_rq->on_list)
5417 			return;
5418 		/*
5419 		 * Nothing to run but something to decay (on_list)?
5420 		 * Complete the branch.
5421 		 */
5422 		for_each_sched_entity(se) {
5423 			if (list_add_leaf_cfs_rq(cfs_rq_of(se)))
5424 				break;
5425 		}
5426 		goto unthrottle_throttle;
5427 	}
5428 
5429 	task_delta = cfs_rq->h_nr_running;
5430 	idle_task_delta = cfs_rq->idle_h_nr_running;
5431 	for_each_sched_entity(se) {
5432 		struct cfs_rq *qcfs_rq = cfs_rq_of(se);
5433 
5434 		if (se->on_rq)
5435 			break;
5436 		enqueue_entity(qcfs_rq, se, ENQUEUE_WAKEUP);
5437 
5438 		if (cfs_rq_is_idle(group_cfs_rq(se)))
5439 			idle_task_delta = cfs_rq->h_nr_running;
5440 
5441 		qcfs_rq->h_nr_running += task_delta;
5442 		qcfs_rq->idle_h_nr_running += idle_task_delta;
5443 
5444 		/* end evaluation on encountering a throttled cfs_rq */
5445 		if (cfs_rq_throttled(qcfs_rq))
5446 			goto unthrottle_throttle;
5447 	}
5448 
5449 	for_each_sched_entity(se) {
5450 		struct cfs_rq *qcfs_rq = cfs_rq_of(se);
5451 
5452 		update_load_avg(qcfs_rq, se, UPDATE_TG);
5453 		se_update_runnable(se);
5454 
5455 		if (cfs_rq_is_idle(group_cfs_rq(se)))
5456 			idle_task_delta = cfs_rq->h_nr_running;
5457 
5458 		qcfs_rq->h_nr_running += task_delta;
5459 		qcfs_rq->idle_h_nr_running += idle_task_delta;
5460 
5461 		/* end evaluation on encountering a throttled cfs_rq */
5462 		if (cfs_rq_throttled(qcfs_rq))
5463 			goto unthrottle_throttle;
5464 	}
5465 
5466 	/* At this point se is NULL and we are at root level*/
5467 	add_nr_running(rq, task_delta);
5468 
5469 unthrottle_throttle:
5470 	assert_list_leaf_cfs_rq(rq);
5471 
5472 	/* Determine whether we need to wake up potentially idle CPU: */
5473 	if (rq->curr == rq->idle && rq->cfs.nr_running)
5474 		resched_curr(rq);
5475 }
5476 
5477 #ifdef CONFIG_SMP
5478 static void __cfsb_csd_unthrottle(void *arg)
5479 {
5480 	struct cfs_rq *cursor, *tmp;
5481 	struct rq *rq = arg;
5482 	struct rq_flags rf;
5483 
5484 	rq_lock(rq, &rf);
5485 
5486 	/*
5487 	 * Since we hold rq lock we're safe from concurrent manipulation of
5488 	 * the CSD list. However, this RCU critical section annotates the
5489 	 * fact that we pair with sched_free_group_rcu(), so that we cannot
5490 	 * race with group being freed in the window between removing it
5491 	 * from the list and advancing to the next entry in the list.
5492 	 */
5493 	rcu_read_lock();
5494 
5495 	list_for_each_entry_safe(cursor, tmp, &rq->cfsb_csd_list,
5496 				 throttled_csd_list) {
5497 		list_del_init(&cursor->throttled_csd_list);
5498 
5499 		if (cfs_rq_throttled(cursor))
5500 			unthrottle_cfs_rq(cursor);
5501 	}
5502 
5503 	rcu_read_unlock();
5504 
5505 	rq_unlock(rq, &rf);
5506 }
5507 
5508 static inline void __unthrottle_cfs_rq_async(struct cfs_rq *cfs_rq)
5509 {
5510 	struct rq *rq = rq_of(cfs_rq);
5511 	bool first;
5512 
5513 	if (rq == this_rq()) {
5514 		unthrottle_cfs_rq(cfs_rq);
5515 		return;
5516 	}
5517 
5518 	/* Already enqueued */
5519 	if (SCHED_WARN_ON(!list_empty(&cfs_rq->throttled_csd_list)))
5520 		return;
5521 
5522 	first = list_empty(&rq->cfsb_csd_list);
5523 	list_add_tail(&cfs_rq->throttled_csd_list, &rq->cfsb_csd_list);
5524 	if (first)
5525 		smp_call_function_single_async(cpu_of(rq), &rq->cfsb_csd);
5526 }
5527 #else
5528 static inline void __unthrottle_cfs_rq_async(struct cfs_rq *cfs_rq)
5529 {
5530 	unthrottle_cfs_rq(cfs_rq);
5531 }
5532 #endif
5533 
5534 static void unthrottle_cfs_rq_async(struct cfs_rq *cfs_rq)
5535 {
5536 	lockdep_assert_rq_held(rq_of(cfs_rq));
5537 
5538 	if (SCHED_WARN_ON(!cfs_rq_throttled(cfs_rq) ||
5539 	    cfs_rq->runtime_remaining <= 0))
5540 		return;
5541 
5542 	__unthrottle_cfs_rq_async(cfs_rq);
5543 }
5544 
5545 static bool distribute_cfs_runtime(struct cfs_bandwidth *cfs_b)
5546 {
5547 	struct cfs_rq *local_unthrottle = NULL;
5548 	int this_cpu = smp_processor_id();
5549 	u64 runtime, remaining = 1;
5550 	bool throttled = false;
5551 	struct cfs_rq *cfs_rq;
5552 	struct rq_flags rf;
5553 	struct rq *rq;
5554 
5555 	rcu_read_lock();
5556 	list_for_each_entry_rcu(cfs_rq, &cfs_b->throttled_cfs_rq,
5557 				throttled_list) {
5558 		rq = rq_of(cfs_rq);
5559 
5560 		if (!remaining) {
5561 			throttled = true;
5562 			break;
5563 		}
5564 
5565 		rq_lock_irqsave(rq, &rf);
5566 		if (!cfs_rq_throttled(cfs_rq))
5567 			goto next;
5568 
5569 #ifdef CONFIG_SMP
5570 		/* Already queued for async unthrottle */
5571 		if (!list_empty(&cfs_rq->throttled_csd_list))
5572 			goto next;
5573 #endif
5574 
5575 		/* By the above checks, this should never be true */
5576 		SCHED_WARN_ON(cfs_rq->runtime_remaining > 0);
5577 
5578 		raw_spin_lock(&cfs_b->lock);
5579 		runtime = -cfs_rq->runtime_remaining + 1;
5580 		if (runtime > cfs_b->runtime)
5581 			runtime = cfs_b->runtime;
5582 		cfs_b->runtime -= runtime;
5583 		remaining = cfs_b->runtime;
5584 		raw_spin_unlock(&cfs_b->lock);
5585 
5586 		cfs_rq->runtime_remaining += runtime;
5587 
5588 		/* we check whether we're throttled above */
5589 		if (cfs_rq->runtime_remaining > 0) {
5590 			if (cpu_of(rq) != this_cpu ||
5591 			    SCHED_WARN_ON(local_unthrottle))
5592 				unthrottle_cfs_rq_async(cfs_rq);
5593 			else
5594 				local_unthrottle = cfs_rq;
5595 		} else {
5596 			throttled = true;
5597 		}
5598 
5599 next:
5600 		rq_unlock_irqrestore(rq, &rf);
5601 	}
5602 	rcu_read_unlock();
5603 
5604 	if (local_unthrottle) {
5605 		rq = cpu_rq(this_cpu);
5606 		rq_lock_irqsave(rq, &rf);
5607 		if (cfs_rq_throttled(local_unthrottle))
5608 			unthrottle_cfs_rq(local_unthrottle);
5609 		rq_unlock_irqrestore(rq, &rf);
5610 	}
5611 
5612 	return throttled;
5613 }
5614 
5615 /*
5616  * Responsible for refilling a task_group's bandwidth and unthrottling its
5617  * cfs_rqs as appropriate. If there has been no activity within the last
5618  * period the timer is deactivated until scheduling resumes; cfs_b->idle is
5619  * used to track this state.
5620  */
5621 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun, unsigned long flags)
5622 {
5623 	int throttled;
5624 
5625 	/* no need to continue the timer with no bandwidth constraint */
5626 	if (cfs_b->quota == RUNTIME_INF)
5627 		goto out_deactivate;
5628 
5629 	throttled = !list_empty(&cfs_b->throttled_cfs_rq);
5630 	cfs_b->nr_periods += overrun;
5631 
5632 	/* Refill extra burst quota even if cfs_b->idle */
5633 	__refill_cfs_bandwidth_runtime(cfs_b);
5634 
5635 	/*
5636 	 * idle depends on !throttled (for the case of a large deficit), and if
5637 	 * we're going inactive then everything else can be deferred
5638 	 */
5639 	if (cfs_b->idle && !throttled)
5640 		goto out_deactivate;
5641 
5642 	if (!throttled) {
5643 		/* mark as potentially idle for the upcoming period */
5644 		cfs_b->idle = 1;
5645 		return 0;
5646 	}
5647 
5648 	/* account preceding periods in which throttling occurred */
5649 	cfs_b->nr_throttled += overrun;
5650 
5651 	/*
5652 	 * This check is repeated as we release cfs_b->lock while we unthrottle.
5653 	 */
5654 	while (throttled && cfs_b->runtime > 0) {
5655 		raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
5656 		/* we can't nest cfs_b->lock while distributing bandwidth */
5657 		throttled = distribute_cfs_runtime(cfs_b);
5658 		raw_spin_lock_irqsave(&cfs_b->lock, flags);
5659 	}
5660 
5661 	/*
5662 	 * While we are ensured activity in the period following an
5663 	 * unthrottle, this also covers the case in which the new bandwidth is
5664 	 * insufficient to cover the existing bandwidth deficit.  (Forcing the
5665 	 * timer to remain active while there are any throttled entities.)
5666 	 */
5667 	cfs_b->idle = 0;
5668 
5669 	return 0;
5670 
5671 out_deactivate:
5672 	return 1;
5673 }
5674 
5675 /* a cfs_rq won't donate quota below this amount */
5676 static const u64 min_cfs_rq_runtime = 1 * NSEC_PER_MSEC;
5677 /* minimum remaining period time to redistribute slack quota */
5678 static const u64 min_bandwidth_expiration = 2 * NSEC_PER_MSEC;
5679 /* how long we wait to gather additional slack before distributing */
5680 static const u64 cfs_bandwidth_slack_period = 5 * NSEC_PER_MSEC;
5681 
5682 /*
5683  * Are we near the end of the current quota period?
5684  *
5685  * Requires cfs_b->lock for hrtimer_expires_remaining to be safe against the
5686  * hrtimer base being cleared by hrtimer_start. In the case of
5687  * migrate_hrtimers, base is never cleared, so we are fine.
5688  */
5689 static int runtime_refresh_within(struct cfs_bandwidth *cfs_b, u64 min_expire)
5690 {
5691 	struct hrtimer *refresh_timer = &cfs_b->period_timer;
5692 	s64 remaining;
5693 
5694 	/* if the call-back is running a quota refresh is already occurring */
5695 	if (hrtimer_callback_running(refresh_timer))
5696 		return 1;
5697 
5698 	/* is a quota refresh about to occur? */
5699 	remaining = ktime_to_ns(hrtimer_expires_remaining(refresh_timer));
5700 	if (remaining < (s64)min_expire)
5701 		return 1;
5702 
5703 	return 0;
5704 }
5705 
5706 static void start_cfs_slack_bandwidth(struct cfs_bandwidth *cfs_b)
5707 {
5708 	u64 min_left = cfs_bandwidth_slack_period + min_bandwidth_expiration;
5709 
5710 	/* if there's a quota refresh soon don't bother with slack */
5711 	if (runtime_refresh_within(cfs_b, min_left))
5712 		return;
5713 
5714 	/* don't push forwards an existing deferred unthrottle */
5715 	if (cfs_b->slack_started)
5716 		return;
5717 	cfs_b->slack_started = true;
5718 
5719 	hrtimer_start(&cfs_b->slack_timer,
5720 			ns_to_ktime(cfs_bandwidth_slack_period),
5721 			HRTIMER_MODE_REL);
5722 }
5723 
5724 /* we know any runtime found here is valid as update_curr() precedes return */
5725 static void __return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
5726 {
5727 	struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
5728 	s64 slack_runtime = cfs_rq->runtime_remaining - min_cfs_rq_runtime;
5729 
5730 	if (slack_runtime <= 0)
5731 		return;
5732 
5733 	raw_spin_lock(&cfs_b->lock);
5734 	if (cfs_b->quota != RUNTIME_INF) {
5735 		cfs_b->runtime += slack_runtime;
5736 
5737 		/* we are under rq->lock, defer unthrottling using a timer */
5738 		if (cfs_b->runtime > sched_cfs_bandwidth_slice() &&
5739 		    !list_empty(&cfs_b->throttled_cfs_rq))
5740 			start_cfs_slack_bandwidth(cfs_b);
5741 	}
5742 	raw_spin_unlock(&cfs_b->lock);
5743 
5744 	/* even if it's not valid for return we don't want to try again */
5745 	cfs_rq->runtime_remaining -= slack_runtime;
5746 }
5747 
5748 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq)
5749 {
5750 	if (!cfs_bandwidth_used())
5751 		return;
5752 
5753 	if (!cfs_rq->runtime_enabled || cfs_rq->nr_running)
5754 		return;
5755 
5756 	__return_cfs_rq_runtime(cfs_rq);
5757 }
5758 
5759 /*
5760  * This is done with a timer (instead of inline with bandwidth return) since
5761  * it's necessary to juggle rq->locks to unthrottle their respective cfs_rqs.
5762  */
5763 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b)
5764 {
5765 	u64 runtime = 0, slice = sched_cfs_bandwidth_slice();
5766 	unsigned long flags;
5767 
5768 	/* confirm we're still not at a refresh boundary */
5769 	raw_spin_lock_irqsave(&cfs_b->lock, flags);
5770 	cfs_b->slack_started = false;
5771 
5772 	if (runtime_refresh_within(cfs_b, min_bandwidth_expiration)) {
5773 		raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
5774 		return;
5775 	}
5776 
5777 	if (cfs_b->quota != RUNTIME_INF && cfs_b->runtime > slice)
5778 		runtime = cfs_b->runtime;
5779 
5780 	raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
5781 
5782 	if (!runtime)
5783 		return;
5784 
5785 	distribute_cfs_runtime(cfs_b);
5786 }
5787 
5788 /*
5789  * When a group wakes up we want to make sure that its quota is not already
5790  * expired/exceeded, otherwise it may be allowed to steal additional ticks of
5791  * runtime as update_curr() throttling can not trigger until it's on-rq.
5792  */
5793 static void check_enqueue_throttle(struct cfs_rq *cfs_rq)
5794 {
5795 	if (!cfs_bandwidth_used())
5796 		return;
5797 
5798 	/* an active group must be handled by the update_curr()->put() path */
5799 	if (!cfs_rq->runtime_enabled || cfs_rq->curr)
5800 		return;
5801 
5802 	/* ensure the group is not already throttled */
5803 	if (cfs_rq_throttled(cfs_rq))
5804 		return;
5805 
5806 	/* update runtime allocation */
5807 	account_cfs_rq_runtime(cfs_rq, 0);
5808 	if (cfs_rq->runtime_remaining <= 0)
5809 		throttle_cfs_rq(cfs_rq);
5810 }
5811 
5812 static void sync_throttle(struct task_group *tg, int cpu)
5813 {
5814 	struct cfs_rq *pcfs_rq, *cfs_rq;
5815 
5816 	if (!cfs_bandwidth_used())
5817 		return;
5818 
5819 	if (!tg->parent)
5820 		return;
5821 
5822 	cfs_rq = tg->cfs_rq[cpu];
5823 	pcfs_rq = tg->parent->cfs_rq[cpu];
5824 
5825 	cfs_rq->throttle_count = pcfs_rq->throttle_count;
5826 	cfs_rq->throttled_clock_pelt = rq_clock_pelt(cpu_rq(cpu));
5827 }
5828 
5829 /* conditionally throttle active cfs_rq's from put_prev_entity() */
5830 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq)
5831 {
5832 	if (!cfs_bandwidth_used())
5833 		return false;
5834 
5835 	if (likely(!cfs_rq->runtime_enabled || cfs_rq->runtime_remaining > 0))
5836 		return false;
5837 
5838 	/*
5839 	 * it's possible for a throttled entity to be forced into a running
5840 	 * state (e.g. set_curr_task), in this case we're finished.
5841 	 */
5842 	if (cfs_rq_throttled(cfs_rq))
5843 		return true;
5844 
5845 	return throttle_cfs_rq(cfs_rq);
5846 }
5847 
5848 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
5849 {
5850 	struct cfs_bandwidth *cfs_b =
5851 		container_of(timer, struct cfs_bandwidth, slack_timer);
5852 
5853 	do_sched_cfs_slack_timer(cfs_b);
5854 
5855 	return HRTIMER_NORESTART;
5856 }
5857 
5858 extern const u64 max_cfs_quota_period;
5859 
5860 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
5861 {
5862 	struct cfs_bandwidth *cfs_b =
5863 		container_of(timer, struct cfs_bandwidth, period_timer);
5864 	unsigned long flags;
5865 	int overrun;
5866 	int idle = 0;
5867 	int count = 0;
5868 
5869 	raw_spin_lock_irqsave(&cfs_b->lock, flags);
5870 	for (;;) {
5871 		overrun = hrtimer_forward_now(timer, cfs_b->period);
5872 		if (!overrun)
5873 			break;
5874 
5875 		idle = do_sched_cfs_period_timer(cfs_b, overrun, flags);
5876 
5877 		if (++count > 3) {
5878 			u64 new, old = ktime_to_ns(cfs_b->period);
5879 
5880 			/*
5881 			 * Grow period by a factor of 2 to avoid losing precision.
5882 			 * Precision loss in the quota/period ratio can cause __cfs_schedulable
5883 			 * to fail.
5884 			 */
5885 			new = old * 2;
5886 			if (new < max_cfs_quota_period) {
5887 				cfs_b->period = ns_to_ktime(new);
5888 				cfs_b->quota *= 2;
5889 				cfs_b->burst *= 2;
5890 
5891 				pr_warn_ratelimited(
5892 	"cfs_period_timer[cpu%d]: period too short, scaling up (new cfs_period_us = %lld, cfs_quota_us = %lld)\n",
5893 					smp_processor_id(),
5894 					div_u64(new, NSEC_PER_USEC),
5895 					div_u64(cfs_b->quota, NSEC_PER_USEC));
5896 			} else {
5897 				pr_warn_ratelimited(
5898 	"cfs_period_timer[cpu%d]: period too short, but cannot scale up without losing precision (cfs_period_us = %lld, cfs_quota_us = %lld)\n",
5899 					smp_processor_id(),
5900 					div_u64(old, NSEC_PER_USEC),
5901 					div_u64(cfs_b->quota, NSEC_PER_USEC));
5902 			}
5903 
5904 			/* reset count so we don't come right back in here */
5905 			count = 0;
5906 		}
5907 	}
5908 	if (idle)
5909 		cfs_b->period_active = 0;
5910 	raw_spin_unlock_irqrestore(&cfs_b->lock, flags);
5911 
5912 	return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
5913 }
5914 
5915 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
5916 {
5917 	raw_spin_lock_init(&cfs_b->lock);
5918 	cfs_b->runtime = 0;
5919 	cfs_b->quota = RUNTIME_INF;
5920 	cfs_b->period = ns_to_ktime(default_cfs_period());
5921 	cfs_b->burst = 0;
5922 
5923 	INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
5924 	hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS_PINNED);
5925 	cfs_b->period_timer.function = sched_cfs_period_timer;
5926 
5927 	/* Add a random offset so that timers interleave */
5928 	hrtimer_set_expires(&cfs_b->period_timer,
5929 			    get_random_u32_below(cfs_b->period));
5930 	hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
5931 	cfs_b->slack_timer.function = sched_cfs_slack_timer;
5932 	cfs_b->slack_started = false;
5933 }
5934 
5935 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
5936 {
5937 	cfs_rq->runtime_enabled = 0;
5938 	INIT_LIST_HEAD(&cfs_rq->throttled_list);
5939 #ifdef CONFIG_SMP
5940 	INIT_LIST_HEAD(&cfs_rq->throttled_csd_list);
5941 #endif
5942 }
5943 
5944 void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
5945 {
5946 	lockdep_assert_held(&cfs_b->lock);
5947 
5948 	if (cfs_b->period_active)
5949 		return;
5950 
5951 	cfs_b->period_active = 1;
5952 	hrtimer_forward_now(&cfs_b->period_timer, cfs_b->period);
5953 	hrtimer_start_expires(&cfs_b->period_timer, HRTIMER_MODE_ABS_PINNED);
5954 }
5955 
5956 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
5957 {
5958 	int __maybe_unused i;
5959 
5960 	/* init_cfs_bandwidth() was not called */
5961 	if (!cfs_b->throttled_cfs_rq.next)
5962 		return;
5963 
5964 	hrtimer_cancel(&cfs_b->period_timer);
5965 	hrtimer_cancel(&cfs_b->slack_timer);
5966 
5967 	/*
5968 	 * It is possible that we still have some cfs_rq's pending on a CSD
5969 	 * list, though this race is very rare. In order for this to occur, we
5970 	 * must have raced with the last task leaving the group while there
5971 	 * exist throttled cfs_rq(s), and the period_timer must have queued the
5972 	 * CSD item but the remote cpu has not yet processed it. To handle this,
5973 	 * we can simply flush all pending CSD work inline here. We're
5974 	 * guaranteed at this point that no additional cfs_rq of this group can
5975 	 * join a CSD list.
5976 	 */
5977 #ifdef CONFIG_SMP
5978 	for_each_possible_cpu(i) {
5979 		struct rq *rq = cpu_rq(i);
5980 		unsigned long flags;
5981 
5982 		if (list_empty(&rq->cfsb_csd_list))
5983 			continue;
5984 
5985 		local_irq_save(flags);
5986 		__cfsb_csd_unthrottle(rq);
5987 		local_irq_restore(flags);
5988 	}
5989 #endif
5990 }
5991 
5992 /*
5993  * Both these CPU hotplug callbacks race against unregister_fair_sched_group()
5994  *
5995  * The race is harmless, since modifying bandwidth settings of unhooked group
5996  * bits doesn't do much.
5997  */
5998 
5999 /* cpu online callback */
6000 static void __maybe_unused update_runtime_enabled(struct rq *rq)
6001 {
6002 	struct task_group *tg;
6003 
6004 	lockdep_assert_rq_held(rq);
6005 
6006 	rcu_read_lock();
6007 	list_for_each_entry_rcu(tg, &task_groups, list) {
6008 		struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
6009 		struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
6010 
6011 		raw_spin_lock(&cfs_b->lock);
6012 		cfs_rq->runtime_enabled = cfs_b->quota != RUNTIME_INF;
6013 		raw_spin_unlock(&cfs_b->lock);
6014 	}
6015 	rcu_read_unlock();
6016 }
6017 
6018 /* cpu offline callback */
6019 static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
6020 {
6021 	struct task_group *tg;
6022 
6023 	lockdep_assert_rq_held(rq);
6024 
6025 	rcu_read_lock();
6026 	list_for_each_entry_rcu(tg, &task_groups, list) {
6027 		struct cfs_rq *cfs_rq = tg->cfs_rq[cpu_of(rq)];
6028 
6029 		if (!cfs_rq->runtime_enabled)
6030 			continue;
6031 
6032 		/*
6033 		 * clock_task is not advancing so we just need to make sure
6034 		 * there's some valid quota amount
6035 		 */
6036 		cfs_rq->runtime_remaining = 1;
6037 		/*
6038 		 * Offline rq is schedulable till CPU is completely disabled
6039 		 * in take_cpu_down(), so we prevent new cfs throttling here.
6040 		 */
6041 		cfs_rq->runtime_enabled = 0;
6042 
6043 		if (cfs_rq_throttled(cfs_rq))
6044 			unthrottle_cfs_rq(cfs_rq);
6045 	}
6046 	rcu_read_unlock();
6047 }
6048 
6049 #else /* CONFIG_CFS_BANDWIDTH */
6050 
6051 static inline bool cfs_bandwidth_used(void)
6052 {
6053 	return false;
6054 }
6055 
6056 static void account_cfs_rq_runtime(struct cfs_rq *cfs_rq, u64 delta_exec) {}
6057 static bool check_cfs_rq_runtime(struct cfs_rq *cfs_rq) { return false; }
6058 static void check_enqueue_throttle(struct cfs_rq *cfs_rq) {}
6059 static inline void sync_throttle(struct task_group *tg, int cpu) {}
6060 static __always_inline void return_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
6061 
6062 static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
6063 {
6064 	return 0;
6065 }
6066 
6067 static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
6068 {
6069 	return 0;
6070 }
6071 
6072 static inline int throttled_lb_pair(struct task_group *tg,
6073 				    int src_cpu, int dest_cpu)
6074 {
6075 	return 0;
6076 }
6077 
6078 void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
6079 
6080 #ifdef CONFIG_FAIR_GROUP_SCHED
6081 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
6082 #endif
6083 
6084 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
6085 {
6086 	return NULL;
6087 }
6088 static inline void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
6089 static inline void update_runtime_enabled(struct rq *rq) {}
6090 static inline void unthrottle_offline_cfs_rqs(struct rq *rq) {}
6091 
6092 #endif /* CONFIG_CFS_BANDWIDTH */
6093 
6094 /**************************************************
6095  * CFS operations on tasks:
6096  */
6097 
6098 #ifdef CONFIG_SCHED_HRTICK
6099 static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
6100 {
6101 	struct sched_entity *se = &p->se;
6102 	struct cfs_rq *cfs_rq = cfs_rq_of(se);
6103 
6104 	SCHED_WARN_ON(task_rq(p) != rq);
6105 
6106 	if (rq->cfs.h_nr_running > 1) {
6107 		u64 slice = sched_slice(cfs_rq, se);
6108 		u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
6109 		s64 delta = slice - ran;
6110 
6111 		if (delta < 0) {
6112 			if (task_current(rq, p))
6113 				resched_curr(rq);
6114 			return;
6115 		}
6116 		hrtick_start(rq, delta);
6117 	}
6118 }
6119 
6120 /*
6121  * called from enqueue/dequeue and updates the hrtick when the
6122  * current task is from our class and nr_running is low enough
6123  * to matter.
6124  */
6125 static void hrtick_update(struct rq *rq)
6126 {
6127 	struct task_struct *curr = rq->curr;
6128 
6129 	if (!hrtick_enabled_fair(rq) || curr->sched_class != &fair_sched_class)
6130 		return;
6131 
6132 	if (cfs_rq_of(&curr->se)->nr_running < sched_nr_latency)
6133 		hrtick_start_fair(rq, curr);
6134 }
6135 #else /* !CONFIG_SCHED_HRTICK */
6136 static inline void
6137 hrtick_start_fair(struct rq *rq, struct task_struct *p)
6138 {
6139 }
6140 
6141 static inline void hrtick_update(struct rq *rq)
6142 {
6143 }
6144 #endif
6145 
6146 #ifdef CONFIG_SMP
6147 static inline bool cpu_overutilized(int cpu)
6148 {
6149 	unsigned long rq_util_min = uclamp_rq_get(cpu_rq(cpu), UCLAMP_MIN);
6150 	unsigned long rq_util_max = uclamp_rq_get(cpu_rq(cpu), UCLAMP_MAX);
6151 
6152 	/* Return true only if the utilization doesn't fit CPU's capacity */
6153 	return !util_fits_cpu(cpu_util_cfs(cpu), rq_util_min, rq_util_max, cpu);
6154 }
6155 
6156 static inline void update_overutilized_status(struct rq *rq)
6157 {
6158 	if (!READ_ONCE(rq->rd->overutilized) && cpu_overutilized(rq->cpu)) {
6159 		WRITE_ONCE(rq->rd->overutilized, SG_OVERUTILIZED);
6160 		trace_sched_overutilized_tp(rq->rd, SG_OVERUTILIZED);
6161 	}
6162 }
6163 #else
6164 static inline void update_overutilized_status(struct rq *rq) { }
6165 #endif
6166 
6167 /* Runqueue only has SCHED_IDLE tasks enqueued */
6168 static int sched_idle_rq(struct rq *rq)
6169 {
6170 	return unlikely(rq->nr_running == rq->cfs.idle_h_nr_running &&
6171 			rq->nr_running);
6172 }
6173 
6174 /*
6175  * Returns true if cfs_rq only has SCHED_IDLE entities enqueued. Note the use
6176  * of idle_nr_running, which does not consider idle descendants of normal
6177  * entities.
6178  */
6179 static bool sched_idle_cfs_rq(struct cfs_rq *cfs_rq)
6180 {
6181 	return cfs_rq->nr_running &&
6182 		cfs_rq->nr_running == cfs_rq->idle_nr_running;
6183 }
6184 
6185 #ifdef CONFIG_SMP
6186 static int sched_idle_cpu(int cpu)
6187 {
6188 	return sched_idle_rq(cpu_rq(cpu));
6189 }
6190 #endif
6191 
6192 /*
6193  * The enqueue_task method is called before nr_running is
6194  * increased. Here we update the fair scheduling stats and
6195  * then put the task into the rbtree:
6196  */
6197 static void
6198 enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
6199 {
6200 	struct cfs_rq *cfs_rq;
6201 	struct sched_entity *se = &p->se;
6202 	int idle_h_nr_running = task_has_idle_policy(p);
6203 	int task_new = !(flags & ENQUEUE_WAKEUP);
6204 
6205 	/*
6206 	 * The code below (indirectly) updates schedutil which looks at
6207 	 * the cfs_rq utilization to select a frequency.
6208 	 * Let's add the task's estimated utilization to the cfs_rq's
6209 	 * estimated utilization, before we update schedutil.
6210 	 */
6211 	util_est_enqueue(&rq->cfs, p);
6212 
6213 	/*
6214 	 * If in_iowait is set, the code below may not trigger any cpufreq
6215 	 * utilization updates, so do it here explicitly with the IOWAIT flag
6216 	 * passed.
6217 	 */
6218 	if (p->in_iowait)
6219 		cpufreq_update_util(rq, SCHED_CPUFREQ_IOWAIT);
6220 
6221 	for_each_sched_entity(se) {
6222 		if (se->on_rq)
6223 			break;
6224 		cfs_rq = cfs_rq_of(se);
6225 		enqueue_entity(cfs_rq, se, flags);
6226 
6227 		cfs_rq->h_nr_running++;
6228 		cfs_rq->idle_h_nr_running += idle_h_nr_running;
6229 
6230 		if (cfs_rq_is_idle(cfs_rq))
6231 			idle_h_nr_running = 1;
6232 
6233 		/* end evaluation on encountering a throttled cfs_rq */
6234 		if (cfs_rq_throttled(cfs_rq))
6235 			goto enqueue_throttle;
6236 
6237 		flags = ENQUEUE_WAKEUP;
6238 	}
6239 
6240 	for_each_sched_entity(se) {
6241 		cfs_rq = cfs_rq_of(se);
6242 
6243 		update_load_avg(cfs_rq, se, UPDATE_TG);
6244 		se_update_runnable(se);
6245 		update_cfs_group(se);
6246 
6247 		cfs_rq->h_nr_running++;
6248 		cfs_rq->idle_h_nr_running += idle_h_nr_running;
6249 
6250 		if (cfs_rq_is_idle(cfs_rq))
6251 			idle_h_nr_running = 1;
6252 
6253 		/* end evaluation on encountering a throttled cfs_rq */
6254 		if (cfs_rq_throttled(cfs_rq))
6255 			goto enqueue_throttle;
6256 	}
6257 
6258 	/* At this point se is NULL and we are at root level*/
6259 	add_nr_running(rq, 1);
6260 
6261 	/*
6262 	 * Since new tasks are assigned an initial util_avg equal to
6263 	 * half of the spare capacity of their CPU, tiny tasks have the
6264 	 * ability to cross the overutilized threshold, which will
6265 	 * result in the load balancer ruining all the task placement
6266 	 * done by EAS. As a way to mitigate that effect, do not account
6267 	 * for the first enqueue operation of new tasks during the
6268 	 * overutilized flag detection.
6269 	 *
6270 	 * A better way of solving this problem would be to wait for
6271 	 * the PELT signals of tasks to converge before taking them
6272 	 * into account, but that is not straightforward to implement,
6273 	 * and the following generally works well enough in practice.
6274 	 */
6275 	if (!task_new)
6276 		update_overutilized_status(rq);
6277 
6278 enqueue_throttle:
6279 	assert_list_leaf_cfs_rq(rq);
6280 
6281 	hrtick_update(rq);
6282 }
6283 
6284 static void set_next_buddy(struct sched_entity *se);
6285 
6286 /*
6287  * The dequeue_task method is called before nr_running is
6288  * decreased. We remove the task from the rbtree and
6289  * update the fair scheduling stats:
6290  */
6291 static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
6292 {
6293 	struct cfs_rq *cfs_rq;
6294 	struct sched_entity *se = &p->se;
6295 	int task_sleep = flags & DEQUEUE_SLEEP;
6296 	int idle_h_nr_running = task_has_idle_policy(p);
6297 	bool was_sched_idle = sched_idle_rq(rq);
6298 
6299 	util_est_dequeue(&rq->cfs, p);
6300 
6301 	for_each_sched_entity(se) {
6302 		cfs_rq = cfs_rq_of(se);
6303 		dequeue_entity(cfs_rq, se, flags);
6304 
6305 		cfs_rq->h_nr_running--;
6306 		cfs_rq->idle_h_nr_running -= idle_h_nr_running;
6307 
6308 		if (cfs_rq_is_idle(cfs_rq))
6309 			idle_h_nr_running = 1;
6310 
6311 		/* end evaluation on encountering a throttled cfs_rq */
6312 		if (cfs_rq_throttled(cfs_rq))
6313 			goto dequeue_throttle;
6314 
6315 		/* Don't dequeue parent if it has other entities besides us */
6316 		if (cfs_rq->load.weight) {
6317 			/* Avoid re-evaluating load for this entity: */
6318 			se = parent_entity(se);
6319 			/*
6320 			 * Bias pick_next to pick a task from this cfs_rq, as
6321 			 * p is sleeping when it is within its sched_slice.
6322 			 */
6323 			if (task_sleep && se && !throttled_hierarchy(cfs_rq))
6324 				set_next_buddy(se);
6325 			break;
6326 		}
6327 		flags |= DEQUEUE_SLEEP;
6328 	}
6329 
6330 	for_each_sched_entity(se) {
6331 		cfs_rq = cfs_rq_of(se);
6332 
6333 		update_load_avg(cfs_rq, se, UPDATE_TG);
6334 		se_update_runnable(se);
6335 		update_cfs_group(se);
6336 
6337 		cfs_rq->h_nr_running--;
6338 		cfs_rq->idle_h_nr_running -= idle_h_nr_running;
6339 
6340 		if (cfs_rq_is_idle(cfs_rq))
6341 			idle_h_nr_running = 1;
6342 
6343 		/* end evaluation on encountering a throttled cfs_rq */
6344 		if (cfs_rq_throttled(cfs_rq))
6345 			goto dequeue_throttle;
6346 
6347 	}
6348 
6349 	/* At this point se is NULL and we are at root level*/
6350 	sub_nr_running(rq, 1);
6351 
6352 	/* balance early to pull high priority tasks */
6353 	if (unlikely(!was_sched_idle && sched_idle_rq(rq)))
6354 		rq->next_balance = jiffies;
6355 
6356 dequeue_throttle:
6357 	util_est_update(&rq->cfs, p, task_sleep);
6358 	hrtick_update(rq);
6359 }
6360 
6361 #ifdef CONFIG_SMP
6362 
6363 /* Working cpumask for: load_balance, load_balance_newidle. */
6364 static DEFINE_PER_CPU(cpumask_var_t, load_balance_mask);
6365 static DEFINE_PER_CPU(cpumask_var_t, select_rq_mask);
6366 
6367 #ifdef CONFIG_NO_HZ_COMMON
6368 
6369 static struct {
6370 	cpumask_var_t idle_cpus_mask;
6371 	atomic_t nr_cpus;
6372 	int has_blocked;		/* Idle CPUS has blocked load */
6373 	int needs_update;		/* Newly idle CPUs need their next_balance collated */
6374 	unsigned long next_balance;     /* in jiffy units */
6375 	unsigned long next_blocked;	/* Next update of blocked load in jiffies */
6376 } nohz ____cacheline_aligned;
6377 
6378 #endif /* CONFIG_NO_HZ_COMMON */
6379 
6380 static unsigned long cpu_load(struct rq *rq)
6381 {
6382 	return cfs_rq_load_avg(&rq->cfs);
6383 }
6384 
6385 /*
6386  * cpu_load_without - compute CPU load without any contributions from *p
6387  * @cpu: the CPU which load is requested
6388  * @p: the task which load should be discounted
6389  *
6390  * The load of a CPU is defined by the load of tasks currently enqueued on that
6391  * CPU as well as tasks which are currently sleeping after an execution on that
6392  * CPU.
6393  *
6394  * This method returns the load of the specified CPU by discounting the load of
6395  * the specified task, whenever the task is currently contributing to the CPU
6396  * load.
6397  */
6398 static unsigned long cpu_load_without(struct rq *rq, struct task_struct *p)
6399 {
6400 	struct cfs_rq *cfs_rq;
6401 	unsigned int load;
6402 
6403 	/* Task has no contribution or is new */
6404 	if (cpu_of(rq) != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time))
6405 		return cpu_load(rq);
6406 
6407 	cfs_rq = &rq->cfs;
6408 	load = READ_ONCE(cfs_rq->avg.load_avg);
6409 
6410 	/* Discount task's util from CPU's util */
6411 	lsub_positive(&load, task_h_load(p));
6412 
6413 	return load;
6414 }
6415 
6416 static unsigned long cpu_runnable(struct rq *rq)
6417 {
6418 	return cfs_rq_runnable_avg(&rq->cfs);
6419 }
6420 
6421 static unsigned long cpu_runnable_without(struct rq *rq, struct task_struct *p)
6422 {
6423 	struct cfs_rq *cfs_rq;
6424 	unsigned int runnable;
6425 
6426 	/* Task has no contribution or is new */
6427 	if (cpu_of(rq) != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time))
6428 		return cpu_runnable(rq);
6429 
6430 	cfs_rq = &rq->cfs;
6431 	runnable = READ_ONCE(cfs_rq->avg.runnable_avg);
6432 
6433 	/* Discount task's runnable from CPU's runnable */
6434 	lsub_positive(&runnable, p->se.avg.runnable_avg);
6435 
6436 	return runnable;
6437 }
6438 
6439 static unsigned long capacity_of(int cpu)
6440 {
6441 	return cpu_rq(cpu)->cpu_capacity;
6442 }
6443 
6444 static void record_wakee(struct task_struct *p)
6445 {
6446 	/*
6447 	 * Only decay a single time; tasks that have less then 1 wakeup per
6448 	 * jiffy will not have built up many flips.
6449 	 */
6450 	if (time_after(jiffies, current->wakee_flip_decay_ts + HZ)) {
6451 		current->wakee_flips >>= 1;
6452 		current->wakee_flip_decay_ts = jiffies;
6453 	}
6454 
6455 	if (current->last_wakee != p) {
6456 		current->last_wakee = p;
6457 		current->wakee_flips++;
6458 	}
6459 }
6460 
6461 /*
6462  * Detect M:N waker/wakee relationships via a switching-frequency heuristic.
6463  *
6464  * A waker of many should wake a different task than the one last awakened
6465  * at a frequency roughly N times higher than one of its wakees.
6466  *
6467  * In order to determine whether we should let the load spread vs consolidating
6468  * to shared cache, we look for a minimum 'flip' frequency of llc_size in one
6469  * partner, and a factor of lls_size higher frequency in the other.
6470  *
6471  * With both conditions met, we can be relatively sure that the relationship is
6472  * non-monogamous, with partner count exceeding socket size.
6473  *
6474  * Waker/wakee being client/server, worker/dispatcher, interrupt source or
6475  * whatever is irrelevant, spread criteria is apparent partner count exceeds
6476  * socket size.
6477  */
6478 static int wake_wide(struct task_struct *p)
6479 {
6480 	unsigned int master = current->wakee_flips;
6481 	unsigned int slave = p->wakee_flips;
6482 	int factor = __this_cpu_read(sd_llc_size);
6483 
6484 	if (master < slave)
6485 		swap(master, slave);
6486 	if (slave < factor || master < slave * factor)
6487 		return 0;
6488 	return 1;
6489 }
6490 
6491 /*
6492  * The purpose of wake_affine() is to quickly determine on which CPU we can run
6493  * soonest. For the purpose of speed we only consider the waking and previous
6494  * CPU.
6495  *
6496  * wake_affine_idle() - only considers 'now', it check if the waking CPU is
6497  *			cache-affine and is (or	will be) idle.
6498  *
6499  * wake_affine_weight() - considers the weight to reflect the average
6500  *			  scheduling latency of the CPUs. This seems to work
6501  *			  for the overloaded case.
6502  */
6503 static int
6504 wake_affine_idle(int this_cpu, int prev_cpu, int sync)
6505 {
6506 	/*
6507 	 * If this_cpu is idle, it implies the wakeup is from interrupt
6508 	 * context. Only allow the move if cache is shared. Otherwise an
6509 	 * interrupt intensive workload could force all tasks onto one
6510 	 * node depending on the IO topology or IRQ affinity settings.
6511 	 *
6512 	 * If the prev_cpu is idle and cache affine then avoid a migration.
6513 	 * There is no guarantee that the cache hot data from an interrupt
6514 	 * is more important than cache hot data on the prev_cpu and from
6515 	 * a cpufreq perspective, it's better to have higher utilisation
6516 	 * on one CPU.
6517 	 */
6518 	if (available_idle_cpu(this_cpu) && cpus_share_cache(this_cpu, prev_cpu))
6519 		return available_idle_cpu(prev_cpu) ? prev_cpu : this_cpu;
6520 
6521 	if (sync && cpu_rq(this_cpu)->nr_running == 1)
6522 		return this_cpu;
6523 
6524 	if (available_idle_cpu(prev_cpu))
6525 		return prev_cpu;
6526 
6527 	return nr_cpumask_bits;
6528 }
6529 
6530 static int
6531 wake_affine_weight(struct sched_domain *sd, struct task_struct *p,
6532 		   int this_cpu, int prev_cpu, int sync)
6533 {
6534 	s64 this_eff_load, prev_eff_load;
6535 	unsigned long task_load;
6536 
6537 	this_eff_load = cpu_load(cpu_rq(this_cpu));
6538 
6539 	if (sync) {
6540 		unsigned long current_load = task_h_load(current);
6541 
6542 		if (current_load > this_eff_load)
6543 			return this_cpu;
6544 
6545 		this_eff_load -= current_load;
6546 	}
6547 
6548 	task_load = task_h_load(p);
6549 
6550 	this_eff_load += task_load;
6551 	if (sched_feat(WA_BIAS))
6552 		this_eff_load *= 100;
6553 	this_eff_load *= capacity_of(prev_cpu);
6554 
6555 	prev_eff_load = cpu_load(cpu_rq(prev_cpu));
6556 	prev_eff_load -= task_load;
6557 	if (sched_feat(WA_BIAS))
6558 		prev_eff_load *= 100 + (sd->imbalance_pct - 100) / 2;
6559 	prev_eff_load *= capacity_of(this_cpu);
6560 
6561 	/*
6562 	 * If sync, adjust the weight of prev_eff_load such that if
6563 	 * prev_eff == this_eff that select_idle_sibling() will consider
6564 	 * stacking the wakee on top of the waker if no other CPU is
6565 	 * idle.
6566 	 */
6567 	if (sync)
6568 		prev_eff_load += 1;
6569 
6570 	return this_eff_load < prev_eff_load ? this_cpu : nr_cpumask_bits;
6571 }
6572 
6573 static int wake_affine(struct sched_domain *sd, struct task_struct *p,
6574 		       int this_cpu, int prev_cpu, int sync)
6575 {
6576 	int target = nr_cpumask_bits;
6577 
6578 	if (sched_feat(WA_IDLE))
6579 		target = wake_affine_idle(this_cpu, prev_cpu, sync);
6580 
6581 	if (sched_feat(WA_WEIGHT) && target == nr_cpumask_bits)
6582 		target = wake_affine_weight(sd, p, this_cpu, prev_cpu, sync);
6583 
6584 	schedstat_inc(p->stats.nr_wakeups_affine_attempts);
6585 	if (target == nr_cpumask_bits)
6586 		return prev_cpu;
6587 
6588 	schedstat_inc(sd->ttwu_move_affine);
6589 	schedstat_inc(p->stats.nr_wakeups_affine);
6590 	return target;
6591 }
6592 
6593 static struct sched_group *
6594 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu);
6595 
6596 /*
6597  * find_idlest_group_cpu - find the idlest CPU among the CPUs in the group.
6598  */
6599 static int
6600 find_idlest_group_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
6601 {
6602 	unsigned long load, min_load = ULONG_MAX;
6603 	unsigned int min_exit_latency = UINT_MAX;
6604 	u64 latest_idle_timestamp = 0;
6605 	int least_loaded_cpu = this_cpu;
6606 	int shallowest_idle_cpu = -1;
6607 	int i;
6608 
6609 	/* Check if we have any choice: */
6610 	if (group->group_weight == 1)
6611 		return cpumask_first(sched_group_span(group));
6612 
6613 	/* Traverse only the allowed CPUs */
6614 	for_each_cpu_and(i, sched_group_span(group), p->cpus_ptr) {
6615 		struct rq *rq = cpu_rq(i);
6616 
6617 		if (!sched_core_cookie_match(rq, p))
6618 			continue;
6619 
6620 		if (sched_idle_cpu(i))
6621 			return i;
6622 
6623 		if (available_idle_cpu(i)) {
6624 			struct cpuidle_state *idle = idle_get_state(rq);
6625 			if (idle && idle->exit_latency < min_exit_latency) {
6626 				/*
6627 				 * We give priority to a CPU whose idle state
6628 				 * has the smallest exit latency irrespective
6629 				 * of any idle timestamp.
6630 				 */
6631 				min_exit_latency = idle->exit_latency;
6632 				latest_idle_timestamp = rq->idle_stamp;
6633 				shallowest_idle_cpu = i;
6634 			} else if ((!idle || idle->exit_latency == min_exit_latency) &&
6635 				   rq->idle_stamp > latest_idle_timestamp) {
6636 				/*
6637 				 * If equal or no active idle state, then
6638 				 * the most recently idled CPU might have
6639 				 * a warmer cache.
6640 				 */
6641 				latest_idle_timestamp = rq->idle_stamp;
6642 				shallowest_idle_cpu = i;
6643 			}
6644 		} else if (shallowest_idle_cpu == -1) {
6645 			load = cpu_load(cpu_rq(i));
6646 			if (load < min_load) {
6647 				min_load = load;
6648 				least_loaded_cpu = i;
6649 			}
6650 		}
6651 	}
6652 
6653 	return shallowest_idle_cpu != -1 ? shallowest_idle_cpu : least_loaded_cpu;
6654 }
6655 
6656 static inline int find_idlest_cpu(struct sched_domain *sd, struct task_struct *p,
6657 				  int cpu, int prev_cpu, int sd_flag)
6658 {
6659 	int new_cpu = cpu;
6660 
6661 	if (!cpumask_intersects(sched_domain_span(sd), p->cpus_ptr))
6662 		return prev_cpu;
6663 
6664 	/*
6665 	 * We need task's util for cpu_util_without, sync it up to
6666 	 * prev_cpu's last_update_time.
6667 	 */
6668 	if (!(sd_flag & SD_BALANCE_FORK))
6669 		sync_entity_load_avg(&p->se);
6670 
6671 	while (sd) {
6672 		struct sched_group *group;
6673 		struct sched_domain *tmp;
6674 		int weight;
6675 
6676 		if (!(sd->flags & sd_flag)) {
6677 			sd = sd->child;
6678 			continue;
6679 		}
6680 
6681 		group = find_idlest_group(sd, p, cpu);
6682 		if (!group) {
6683 			sd = sd->child;
6684 			continue;
6685 		}
6686 
6687 		new_cpu = find_idlest_group_cpu(group, p, cpu);
6688 		if (new_cpu == cpu) {
6689 			/* Now try balancing at a lower domain level of 'cpu': */
6690 			sd = sd->child;
6691 			continue;
6692 		}
6693 
6694 		/* Now try balancing at a lower domain level of 'new_cpu': */
6695 		cpu = new_cpu;
6696 		weight = sd->span_weight;
6697 		sd = NULL;
6698 		for_each_domain(cpu, tmp) {
6699 			if (weight <= tmp->span_weight)
6700 				break;
6701 			if (tmp->flags & sd_flag)
6702 				sd = tmp;
6703 		}
6704 	}
6705 
6706 	return new_cpu;
6707 }
6708 
6709 static inline int __select_idle_cpu(int cpu, struct task_struct *p)
6710 {
6711 	if ((available_idle_cpu(cpu) || sched_idle_cpu(cpu)) &&
6712 	    sched_cpu_cookie_match(cpu_rq(cpu), p))
6713 		return cpu;
6714 
6715 	return -1;
6716 }
6717 
6718 #ifdef CONFIG_SCHED_SMT
6719 DEFINE_STATIC_KEY_FALSE(sched_smt_present);
6720 EXPORT_SYMBOL_GPL(sched_smt_present);
6721 
6722 static inline void set_idle_cores(int cpu, int val)
6723 {
6724 	struct sched_domain_shared *sds;
6725 
6726 	sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
6727 	if (sds)
6728 		WRITE_ONCE(sds->has_idle_cores, val);
6729 }
6730 
6731 static inline bool test_idle_cores(int cpu)
6732 {
6733 	struct sched_domain_shared *sds;
6734 
6735 	sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
6736 	if (sds)
6737 		return READ_ONCE(sds->has_idle_cores);
6738 
6739 	return false;
6740 }
6741 
6742 /*
6743  * Scans the local SMT mask to see if the entire core is idle, and records this
6744  * information in sd_llc_shared->has_idle_cores.
6745  *
6746  * Since SMT siblings share all cache levels, inspecting this limited remote
6747  * state should be fairly cheap.
6748  */
6749 void __update_idle_core(struct rq *rq)
6750 {
6751 	int core = cpu_of(rq);
6752 	int cpu;
6753 
6754 	rcu_read_lock();
6755 	if (test_idle_cores(core))
6756 		goto unlock;
6757 
6758 	for_each_cpu(cpu, cpu_smt_mask(core)) {
6759 		if (cpu == core)
6760 			continue;
6761 
6762 		if (!available_idle_cpu(cpu))
6763 			goto unlock;
6764 	}
6765 
6766 	set_idle_cores(core, 1);
6767 unlock:
6768 	rcu_read_unlock();
6769 }
6770 
6771 /*
6772  * Scan the entire LLC domain for idle cores; this dynamically switches off if
6773  * there are no idle cores left in the system; tracked through
6774  * sd_llc->shared->has_idle_cores and enabled through update_idle_core() above.
6775  */
6776 static int select_idle_core(struct task_struct *p, int core, struct cpumask *cpus, int *idle_cpu)
6777 {
6778 	bool idle = true;
6779 	int cpu;
6780 
6781 	for_each_cpu(cpu, cpu_smt_mask(core)) {
6782 		if (!available_idle_cpu(cpu)) {
6783 			idle = false;
6784 			if (*idle_cpu == -1) {
6785 				if (sched_idle_cpu(cpu) && cpumask_test_cpu(cpu, p->cpus_ptr)) {
6786 					*idle_cpu = cpu;
6787 					break;
6788 				}
6789 				continue;
6790 			}
6791 			break;
6792 		}
6793 		if (*idle_cpu == -1 && cpumask_test_cpu(cpu, p->cpus_ptr))
6794 			*idle_cpu = cpu;
6795 	}
6796 
6797 	if (idle)
6798 		return core;
6799 
6800 	cpumask_andnot(cpus, cpus, cpu_smt_mask(core));
6801 	return -1;
6802 }
6803 
6804 /*
6805  * Scan the local SMT mask for idle CPUs.
6806  */
6807 static int select_idle_smt(struct task_struct *p, int target)
6808 {
6809 	int cpu;
6810 
6811 	for_each_cpu_and(cpu, cpu_smt_mask(target), p->cpus_ptr) {
6812 		if (cpu == target)
6813 			continue;
6814 		if (available_idle_cpu(cpu) || sched_idle_cpu(cpu))
6815 			return cpu;
6816 	}
6817 
6818 	return -1;
6819 }
6820 
6821 #else /* CONFIG_SCHED_SMT */
6822 
6823 static inline void set_idle_cores(int cpu, int val)
6824 {
6825 }
6826 
6827 static inline bool test_idle_cores(int cpu)
6828 {
6829 	return false;
6830 }
6831 
6832 static inline int select_idle_core(struct task_struct *p, int core, struct cpumask *cpus, int *idle_cpu)
6833 {
6834 	return __select_idle_cpu(core, p);
6835 }
6836 
6837 static inline int select_idle_smt(struct task_struct *p, int target)
6838 {
6839 	return -1;
6840 }
6841 
6842 #endif /* CONFIG_SCHED_SMT */
6843 
6844 /*
6845  * Scan the LLC domain for idle CPUs; this is dynamically regulated by
6846  * comparing the average scan cost (tracked in sd->avg_scan_cost) against the
6847  * average idle time for this rq (as found in rq->avg_idle).
6848  */
6849 static int select_idle_cpu(struct task_struct *p, struct sched_domain *sd, bool has_idle_core, int target)
6850 {
6851 	struct cpumask *cpus = this_cpu_cpumask_var_ptr(select_rq_mask);
6852 	int i, cpu, idle_cpu = -1, nr = INT_MAX;
6853 	struct sched_domain_shared *sd_share;
6854 	struct rq *this_rq = this_rq();
6855 	int this = smp_processor_id();
6856 	struct sched_domain *this_sd = NULL;
6857 	u64 time = 0;
6858 
6859 	cpumask_and(cpus, sched_domain_span(sd), p->cpus_ptr);
6860 
6861 	if (sched_feat(SIS_PROP) && !has_idle_core) {
6862 		u64 avg_cost, avg_idle, span_avg;
6863 		unsigned long now = jiffies;
6864 
6865 		this_sd = rcu_dereference(*this_cpu_ptr(&sd_llc));
6866 		if (!this_sd)
6867 			return -1;
6868 
6869 		/*
6870 		 * If we're busy, the assumption that the last idle period
6871 		 * predicts the future is flawed; age away the remaining
6872 		 * predicted idle time.
6873 		 */
6874 		if (unlikely(this_rq->wake_stamp < now)) {
6875 			while (this_rq->wake_stamp < now && this_rq->wake_avg_idle) {
6876 				this_rq->wake_stamp++;
6877 				this_rq->wake_avg_idle >>= 1;
6878 			}
6879 		}
6880 
6881 		avg_idle = this_rq->wake_avg_idle;
6882 		avg_cost = this_sd->avg_scan_cost + 1;
6883 
6884 		span_avg = sd->span_weight * avg_idle;
6885 		if (span_avg > 4*avg_cost)
6886 			nr = div_u64(span_avg, avg_cost);
6887 		else
6888 			nr = 4;
6889 
6890 		time = cpu_clock(this);
6891 	}
6892 
6893 	if (sched_feat(SIS_UTIL)) {
6894 		sd_share = rcu_dereference(per_cpu(sd_llc_shared, target));
6895 		if (sd_share) {
6896 			/* because !--nr is the condition to stop scan */
6897 			nr = READ_ONCE(sd_share->nr_idle_scan) + 1;
6898 			/* overloaded LLC is unlikely to have idle cpu/core */
6899 			if (nr == 1)
6900 				return -1;
6901 		}
6902 	}
6903 
6904 	for_each_cpu_wrap(cpu, cpus, target + 1) {
6905 		if (has_idle_core) {
6906 			i = select_idle_core(p, cpu, cpus, &idle_cpu);
6907 			if ((unsigned int)i < nr_cpumask_bits)
6908 				return i;
6909 
6910 		} else {
6911 			if (!--nr)
6912 				return -1;
6913 			idle_cpu = __select_idle_cpu(cpu, p);
6914 			if ((unsigned int)idle_cpu < nr_cpumask_bits)
6915 				break;
6916 		}
6917 	}
6918 
6919 	if (has_idle_core)
6920 		set_idle_cores(target, false);
6921 
6922 	if (sched_feat(SIS_PROP) && this_sd && !has_idle_core) {
6923 		time = cpu_clock(this) - time;
6924 
6925 		/*
6926 		 * Account for the scan cost of wakeups against the average
6927 		 * idle time.
6928 		 */
6929 		this_rq->wake_avg_idle -= min(this_rq->wake_avg_idle, time);
6930 
6931 		update_avg(&this_sd->avg_scan_cost, time);
6932 	}
6933 
6934 	return idle_cpu;
6935 }
6936 
6937 /*
6938  * Scan the asym_capacity domain for idle CPUs; pick the first idle one on which
6939  * the task fits. If no CPU is big enough, but there are idle ones, try to
6940  * maximize capacity.
6941  */
6942 static int
6943 select_idle_capacity(struct task_struct *p, struct sched_domain *sd, int target)
6944 {
6945 	unsigned long task_util, util_min, util_max, best_cap = 0;
6946 	int fits, best_fits = 0;
6947 	int cpu, best_cpu = -1;
6948 	struct cpumask *cpus;
6949 
6950 	cpus = this_cpu_cpumask_var_ptr(select_rq_mask);
6951 	cpumask_and(cpus, sched_domain_span(sd), p->cpus_ptr);
6952 
6953 	task_util = task_util_est(p);
6954 	util_min = uclamp_eff_value(p, UCLAMP_MIN);
6955 	util_max = uclamp_eff_value(p, UCLAMP_MAX);
6956 
6957 	for_each_cpu_wrap(cpu, cpus, target + 1) {
6958 		unsigned long cpu_cap = capacity_of(cpu);
6959 
6960 		if (!available_idle_cpu(cpu) && !sched_idle_cpu(cpu))
6961 			continue;
6962 
6963 		fits = util_fits_cpu(task_util, util_min, util_max, cpu);
6964 
6965 		/* This CPU fits with all requirements */
6966 		if (fits > 0)
6967 			return cpu;
6968 		/*
6969 		 * Only the min performance hint (i.e. uclamp_min) doesn't fit.
6970 		 * Look for the CPU with best capacity.
6971 		 */
6972 		else if (fits < 0)
6973 			cpu_cap = capacity_orig_of(cpu) - thermal_load_avg(cpu_rq(cpu));
6974 
6975 		/*
6976 		 * First, select CPU which fits better (-1 being better than 0).
6977 		 * Then, select the one with best capacity at same level.
6978 		 */
6979 		if ((fits < best_fits) ||
6980 		    ((fits == best_fits) && (cpu_cap > best_cap))) {
6981 			best_cap = cpu_cap;
6982 			best_cpu = cpu;
6983 			best_fits = fits;
6984 		}
6985 	}
6986 
6987 	return best_cpu;
6988 }
6989 
6990 static inline bool asym_fits_cpu(unsigned long util,
6991 				 unsigned long util_min,
6992 				 unsigned long util_max,
6993 				 int cpu)
6994 {
6995 	if (sched_asym_cpucap_active())
6996 		/*
6997 		 * Return true only if the cpu fully fits the task requirements
6998 		 * which include the utilization and the performance hints.
6999 		 */
7000 		return (util_fits_cpu(util, util_min, util_max, cpu) > 0);
7001 
7002 	return true;
7003 }
7004 
7005 /*
7006  * Try and locate an idle core/thread in the LLC cache domain.
7007  */
7008 static int select_idle_sibling(struct task_struct *p, int prev, int target)
7009 {
7010 	bool has_idle_core = false;
7011 	struct sched_domain *sd;
7012 	unsigned long task_util, util_min, util_max;
7013 	int i, recent_used_cpu;
7014 
7015 	/*
7016 	 * On asymmetric system, update task utilization because we will check
7017 	 * that the task fits with cpu's capacity.
7018 	 */
7019 	if (sched_asym_cpucap_active()) {
7020 		sync_entity_load_avg(&p->se);
7021 		task_util = task_util_est(p);
7022 		util_min = uclamp_eff_value(p, UCLAMP_MIN);
7023 		util_max = uclamp_eff_value(p, UCLAMP_MAX);
7024 	}
7025 
7026 	/*
7027 	 * per-cpu select_rq_mask usage
7028 	 */
7029 	lockdep_assert_irqs_disabled();
7030 
7031 	if ((available_idle_cpu(target) || sched_idle_cpu(target)) &&
7032 	    asym_fits_cpu(task_util, util_min, util_max, target))
7033 		return target;
7034 
7035 	/*
7036 	 * If the previous CPU is cache affine and idle, don't be stupid:
7037 	 */
7038 	if (prev != target && cpus_share_cache(prev, target) &&
7039 	    (available_idle_cpu(prev) || sched_idle_cpu(prev)) &&
7040 	    asym_fits_cpu(task_util, util_min, util_max, prev))
7041 		return prev;
7042 
7043 	/*
7044 	 * Allow a per-cpu kthread to stack with the wakee if the
7045 	 * kworker thread and the tasks previous CPUs are the same.
7046 	 * The assumption is that the wakee queued work for the
7047 	 * per-cpu kthread that is now complete and the wakeup is
7048 	 * essentially a sync wakeup. An obvious example of this
7049 	 * pattern is IO completions.
7050 	 */
7051 	if (is_per_cpu_kthread(current) &&
7052 	    in_task() &&
7053 	    prev == smp_processor_id() &&
7054 	    this_rq()->nr_running <= 1 &&
7055 	    asym_fits_cpu(task_util, util_min, util_max, prev)) {
7056 		return prev;
7057 	}
7058 
7059 	/* Check a recently used CPU as a potential idle candidate: */
7060 	recent_used_cpu = p->recent_used_cpu;
7061 	p->recent_used_cpu = prev;
7062 	if (recent_used_cpu != prev &&
7063 	    recent_used_cpu != target &&
7064 	    cpus_share_cache(recent_used_cpu, target) &&
7065 	    (available_idle_cpu(recent_used_cpu) || sched_idle_cpu(recent_used_cpu)) &&
7066 	    cpumask_test_cpu(p->recent_used_cpu, p->cpus_ptr) &&
7067 	    asym_fits_cpu(task_util, util_min, util_max, recent_used_cpu)) {
7068 		return recent_used_cpu;
7069 	}
7070 
7071 	/*
7072 	 * For asymmetric CPU capacity systems, our domain of interest is
7073 	 * sd_asym_cpucapacity rather than sd_llc.
7074 	 */
7075 	if (sched_asym_cpucap_active()) {
7076 		sd = rcu_dereference(per_cpu(sd_asym_cpucapacity, target));
7077 		/*
7078 		 * On an asymmetric CPU capacity system where an exclusive
7079 		 * cpuset defines a symmetric island (i.e. one unique
7080 		 * capacity_orig value through the cpuset), the key will be set
7081 		 * but the CPUs within that cpuset will not have a domain with
7082 		 * SD_ASYM_CPUCAPACITY. These should follow the usual symmetric
7083 		 * capacity path.
7084 		 */
7085 		if (sd) {
7086 			i = select_idle_capacity(p, sd, target);
7087 			return ((unsigned)i < nr_cpumask_bits) ? i : target;
7088 		}
7089 	}
7090 
7091 	sd = rcu_dereference(per_cpu(sd_llc, target));
7092 	if (!sd)
7093 		return target;
7094 
7095 	if (sched_smt_active()) {
7096 		has_idle_core = test_idle_cores(target);
7097 
7098 		if (!has_idle_core && cpus_share_cache(prev, target)) {
7099 			i = select_idle_smt(p, prev);
7100 			if ((unsigned int)i < nr_cpumask_bits)
7101 				return i;
7102 		}
7103 	}
7104 
7105 	i = select_idle_cpu(p, sd, has_idle_core, target);
7106 	if ((unsigned)i < nr_cpumask_bits)
7107 		return i;
7108 
7109 	return target;
7110 }
7111 
7112 /*
7113  * Predicts what cpu_util(@cpu) would return if @p was removed from @cpu
7114  * (@dst_cpu = -1) or migrated to @dst_cpu.
7115  */
7116 static unsigned long cpu_util_next(int cpu, struct task_struct *p, int dst_cpu)
7117 {
7118 	struct cfs_rq *cfs_rq = &cpu_rq(cpu)->cfs;
7119 	unsigned long util = READ_ONCE(cfs_rq->avg.util_avg);
7120 
7121 	/*
7122 	 * If @dst_cpu is -1 or @p migrates from @cpu to @dst_cpu remove its
7123 	 * contribution. If @p migrates from another CPU to @cpu add its
7124 	 * contribution. In all the other cases @cpu is not impacted by the
7125 	 * migration so its util_avg is already correct.
7126 	 */
7127 	if (task_cpu(p) == cpu && dst_cpu != cpu)
7128 		lsub_positive(&util, task_util(p));
7129 	else if (task_cpu(p) != cpu && dst_cpu == cpu)
7130 		util += task_util(p);
7131 
7132 	if (sched_feat(UTIL_EST)) {
7133 		unsigned long util_est;
7134 
7135 		util_est = READ_ONCE(cfs_rq->avg.util_est.enqueued);
7136 
7137 		/*
7138 		 * During wake-up @p isn't enqueued yet and doesn't contribute
7139 		 * to any cpu_rq(cpu)->cfs.avg.util_est.enqueued.
7140 		 * If @dst_cpu == @cpu add it to "simulate" cpu_util after @p
7141 		 * has been enqueued.
7142 		 *
7143 		 * During exec (@dst_cpu = -1) @p is enqueued and does
7144 		 * contribute to cpu_rq(cpu)->cfs.util_est.enqueued.
7145 		 * Remove it to "simulate" cpu_util without @p's contribution.
7146 		 *
7147 		 * Despite the task_on_rq_queued(@p) check there is still a
7148 		 * small window for a possible race when an exec
7149 		 * select_task_rq_fair() races with LB's detach_task().
7150 		 *
7151 		 *   detach_task()
7152 		 *     deactivate_task()
7153 		 *       p->on_rq = TASK_ON_RQ_MIGRATING;
7154 		 *       -------------------------------- A
7155 		 *       dequeue_task()                    \
7156 		 *         dequeue_task_fair()              + Race Time
7157 		 *           util_est_dequeue()            /
7158 		 *       -------------------------------- B
7159 		 *
7160 		 * The additional check "current == p" is required to further
7161 		 * reduce the race window.
7162 		 */
7163 		if (dst_cpu == cpu)
7164 			util_est += _task_util_est(p);
7165 		else if (unlikely(task_on_rq_queued(p) || current == p))
7166 			lsub_positive(&util_est, _task_util_est(p));
7167 
7168 		util = max(util, util_est);
7169 	}
7170 
7171 	return min(util, capacity_orig_of(cpu));
7172 }
7173 
7174 /*
7175  * cpu_util_without: compute cpu utilization without any contributions from *p
7176  * @cpu: the CPU which utilization is requested
7177  * @p: the task which utilization should be discounted
7178  *
7179  * The utilization of a CPU is defined by the utilization of tasks currently
7180  * enqueued on that CPU as well as tasks which are currently sleeping after an
7181  * execution on that CPU.
7182  *
7183  * This method returns the utilization of the specified CPU by discounting the
7184  * utilization of the specified task, whenever the task is currently
7185  * contributing to the CPU utilization.
7186  */
7187 static unsigned long cpu_util_without(int cpu, struct task_struct *p)
7188 {
7189 	/* Task has no contribution or is new */
7190 	if (cpu != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time))
7191 		return cpu_util_cfs(cpu);
7192 
7193 	return cpu_util_next(cpu, p, -1);
7194 }
7195 
7196 /*
7197  * energy_env - Utilization landscape for energy estimation.
7198  * @task_busy_time: Utilization contribution by the task for which we test the
7199  *                  placement. Given by eenv_task_busy_time().
7200  * @pd_busy_time:   Utilization of the whole perf domain without the task
7201  *                  contribution. Given by eenv_pd_busy_time().
7202  * @cpu_cap:        Maximum CPU capacity for the perf domain.
7203  * @pd_cap:         Entire perf domain capacity. (pd->nr_cpus * cpu_cap).
7204  */
7205 struct energy_env {
7206 	unsigned long task_busy_time;
7207 	unsigned long pd_busy_time;
7208 	unsigned long cpu_cap;
7209 	unsigned long pd_cap;
7210 };
7211 
7212 /*
7213  * Compute the task busy time for compute_energy(). This time cannot be
7214  * injected directly into effective_cpu_util() because of the IRQ scaling.
7215  * The latter only makes sense with the most recent CPUs where the task has
7216  * run.
7217  */
7218 static inline void eenv_task_busy_time(struct energy_env *eenv,
7219 				       struct task_struct *p, int prev_cpu)
7220 {
7221 	unsigned long busy_time, max_cap = arch_scale_cpu_capacity(prev_cpu);
7222 	unsigned long irq = cpu_util_irq(cpu_rq(prev_cpu));
7223 
7224 	if (unlikely(irq >= max_cap))
7225 		busy_time = max_cap;
7226 	else
7227 		busy_time = scale_irq_capacity(task_util_est(p), irq, max_cap);
7228 
7229 	eenv->task_busy_time = busy_time;
7230 }
7231 
7232 /*
7233  * Compute the perf_domain (PD) busy time for compute_energy(). Based on the
7234  * utilization for each @pd_cpus, it however doesn't take into account
7235  * clamping since the ratio (utilization / cpu_capacity) is already enough to
7236  * scale the EM reported power consumption at the (eventually clamped)
7237  * cpu_capacity.
7238  *
7239  * The contribution of the task @p for which we want to estimate the
7240  * energy cost is removed (by cpu_util_next()) and must be calculated
7241  * separately (see eenv_task_busy_time). This ensures:
7242  *
7243  *   - A stable PD utilization, no matter which CPU of that PD we want to place
7244  *     the task on.
7245  *
7246  *   - A fair comparison between CPUs as the task contribution (task_util())
7247  *     will always be the same no matter which CPU utilization we rely on
7248  *     (util_avg or util_est).
7249  *
7250  * Set @eenv busy time for the PD that spans @pd_cpus. This busy time can't
7251  * exceed @eenv->pd_cap.
7252  */
7253 static inline void eenv_pd_busy_time(struct energy_env *eenv,
7254 				     struct cpumask *pd_cpus,
7255 				     struct task_struct *p)
7256 {
7257 	unsigned long busy_time = 0;
7258 	int cpu;
7259 
7260 	for_each_cpu(cpu, pd_cpus) {
7261 		unsigned long util = cpu_util_next(cpu, p, -1);
7262 
7263 		busy_time += effective_cpu_util(cpu, util, ENERGY_UTIL, NULL);
7264 	}
7265 
7266 	eenv->pd_busy_time = min(eenv->pd_cap, busy_time);
7267 }
7268 
7269 /*
7270  * Compute the maximum utilization for compute_energy() when the task @p
7271  * is placed on the cpu @dst_cpu.
7272  *
7273  * Returns the maximum utilization among @eenv->cpus. This utilization can't
7274  * exceed @eenv->cpu_cap.
7275  */
7276 static inline unsigned long
7277 eenv_pd_max_util(struct energy_env *eenv, struct cpumask *pd_cpus,
7278 		 struct task_struct *p, int dst_cpu)
7279 {
7280 	unsigned long max_util = 0;
7281 	int cpu;
7282 
7283 	for_each_cpu(cpu, pd_cpus) {
7284 		struct task_struct *tsk = (cpu == dst_cpu) ? p : NULL;
7285 		unsigned long util = cpu_util_next(cpu, p, dst_cpu);
7286 		unsigned long cpu_util;
7287 
7288 		/*
7289 		 * Performance domain frequency: utilization clamping
7290 		 * must be considered since it affects the selection
7291 		 * of the performance domain frequency.
7292 		 * NOTE: in case RT tasks are running, by default the
7293 		 * FREQUENCY_UTIL's utilization can be max OPP.
7294 		 */
7295 		cpu_util = effective_cpu_util(cpu, util, FREQUENCY_UTIL, tsk);
7296 		max_util = max(max_util, cpu_util);
7297 	}
7298 
7299 	return min(max_util, eenv->cpu_cap);
7300 }
7301 
7302 /*
7303  * compute_energy(): Use the Energy Model to estimate the energy that @pd would
7304  * consume for a given utilization landscape @eenv. When @dst_cpu < 0, the task
7305  * contribution is ignored.
7306  */
7307 static inline unsigned long
7308 compute_energy(struct energy_env *eenv, struct perf_domain *pd,
7309 	       struct cpumask *pd_cpus, struct task_struct *p, int dst_cpu)
7310 {
7311 	unsigned long max_util = eenv_pd_max_util(eenv, pd_cpus, p, dst_cpu);
7312 	unsigned long busy_time = eenv->pd_busy_time;
7313 
7314 	if (dst_cpu >= 0)
7315 		busy_time = min(eenv->pd_cap, busy_time + eenv->task_busy_time);
7316 
7317 	return em_cpu_energy(pd->em_pd, max_util, busy_time, eenv->cpu_cap);
7318 }
7319 
7320 /*
7321  * find_energy_efficient_cpu(): Find most energy-efficient target CPU for the
7322  * waking task. find_energy_efficient_cpu() looks for the CPU with maximum
7323  * spare capacity in each performance domain and uses it as a potential
7324  * candidate to execute the task. Then, it uses the Energy Model to figure
7325  * out which of the CPU candidates is the most energy-efficient.
7326  *
7327  * The rationale for this heuristic is as follows. In a performance domain,
7328  * all the most energy efficient CPU candidates (according to the Energy
7329  * Model) are those for which we'll request a low frequency. When there are
7330  * several CPUs for which the frequency request will be the same, we don't
7331  * have enough data to break the tie between them, because the Energy Model
7332  * only includes active power costs. With this model, if we assume that
7333  * frequency requests follow utilization (e.g. using schedutil), the CPU with
7334  * the maximum spare capacity in a performance domain is guaranteed to be among
7335  * the best candidates of the performance domain.
7336  *
7337  * In practice, it could be preferable from an energy standpoint to pack
7338  * small tasks on a CPU in order to let other CPUs go in deeper idle states,
7339  * but that could also hurt our chances to go cluster idle, and we have no
7340  * ways to tell with the current Energy Model if this is actually a good
7341  * idea or not. So, find_energy_efficient_cpu() basically favors
7342  * cluster-packing, and spreading inside a cluster. That should at least be
7343  * a good thing for latency, and this is consistent with the idea that most
7344  * of the energy savings of EAS come from the asymmetry of the system, and
7345  * not so much from breaking the tie between identical CPUs. That's also the
7346  * reason why EAS is enabled in the topology code only for systems where
7347  * SD_ASYM_CPUCAPACITY is set.
7348  *
7349  * NOTE: Forkees are not accepted in the energy-aware wake-up path because
7350  * they don't have any useful utilization data yet and it's not possible to
7351  * forecast their impact on energy consumption. Consequently, they will be
7352  * placed by find_idlest_cpu() on the least loaded CPU, which might turn out
7353  * to be energy-inefficient in some use-cases. The alternative would be to
7354  * bias new tasks towards specific types of CPUs first, or to try to infer
7355  * their util_avg from the parent task, but those heuristics could hurt
7356  * other use-cases too. So, until someone finds a better way to solve this,
7357  * let's keep things simple by re-using the existing slow path.
7358  */
7359 static int find_energy_efficient_cpu(struct task_struct *p, int prev_cpu)
7360 {
7361 	struct cpumask *cpus = this_cpu_cpumask_var_ptr(select_rq_mask);
7362 	unsigned long prev_delta = ULONG_MAX, best_delta = ULONG_MAX;
7363 	unsigned long p_util_min = uclamp_is_used() ? uclamp_eff_value(p, UCLAMP_MIN) : 0;
7364 	unsigned long p_util_max = uclamp_is_used() ? uclamp_eff_value(p, UCLAMP_MAX) : 1024;
7365 	struct root_domain *rd = this_rq()->rd;
7366 	int cpu, best_energy_cpu, target = -1;
7367 	int prev_fits = -1, best_fits = -1;
7368 	unsigned long best_thermal_cap = 0;
7369 	unsigned long prev_thermal_cap = 0;
7370 	struct sched_domain *sd;
7371 	struct perf_domain *pd;
7372 	struct energy_env eenv;
7373 
7374 	rcu_read_lock();
7375 	pd = rcu_dereference(rd->pd);
7376 	if (!pd || READ_ONCE(rd->overutilized))
7377 		goto unlock;
7378 
7379 	/*
7380 	 * Energy-aware wake-up happens on the lowest sched_domain starting
7381 	 * from sd_asym_cpucapacity spanning over this_cpu and prev_cpu.
7382 	 */
7383 	sd = rcu_dereference(*this_cpu_ptr(&sd_asym_cpucapacity));
7384 	while (sd && !cpumask_test_cpu(prev_cpu, sched_domain_span(sd)))
7385 		sd = sd->parent;
7386 	if (!sd)
7387 		goto unlock;
7388 
7389 	target = prev_cpu;
7390 
7391 	sync_entity_load_avg(&p->se);
7392 	if (!uclamp_task_util(p, p_util_min, p_util_max))
7393 		goto unlock;
7394 
7395 	eenv_task_busy_time(&eenv, p, prev_cpu);
7396 
7397 	for (; pd; pd = pd->next) {
7398 		unsigned long util_min = p_util_min, util_max = p_util_max;
7399 		unsigned long cpu_cap, cpu_thermal_cap, util;
7400 		unsigned long cur_delta, max_spare_cap = 0;
7401 		unsigned long rq_util_min, rq_util_max;
7402 		unsigned long prev_spare_cap = 0;
7403 		int max_spare_cap_cpu = -1;
7404 		unsigned long base_energy;
7405 		int fits, max_fits = -1;
7406 
7407 		cpumask_and(cpus, perf_domain_span(pd), cpu_online_mask);
7408 
7409 		if (cpumask_empty(cpus))
7410 			continue;
7411 
7412 		/* Account thermal pressure for the energy estimation */
7413 		cpu = cpumask_first(cpus);
7414 		cpu_thermal_cap = arch_scale_cpu_capacity(cpu);
7415 		cpu_thermal_cap -= arch_scale_thermal_pressure(cpu);
7416 
7417 		eenv.cpu_cap = cpu_thermal_cap;
7418 		eenv.pd_cap = 0;
7419 
7420 		for_each_cpu(cpu, cpus) {
7421 			struct rq *rq = cpu_rq(cpu);
7422 
7423 			eenv.pd_cap += cpu_thermal_cap;
7424 
7425 			if (!cpumask_test_cpu(cpu, sched_domain_span(sd)))
7426 				continue;
7427 
7428 			if (!cpumask_test_cpu(cpu, p->cpus_ptr))
7429 				continue;
7430 
7431 			util = cpu_util_next(cpu, p, cpu);
7432 			cpu_cap = capacity_of(cpu);
7433 
7434 			/*
7435 			 * Skip CPUs that cannot satisfy the capacity request.
7436 			 * IOW, placing the task there would make the CPU
7437 			 * overutilized. Take uclamp into account to see how
7438 			 * much capacity we can get out of the CPU; this is
7439 			 * aligned with sched_cpu_util().
7440 			 */
7441 			if (uclamp_is_used() && !uclamp_rq_is_idle(rq)) {
7442 				/*
7443 				 * Open code uclamp_rq_util_with() except for
7444 				 * the clamp() part. Ie: apply max aggregation
7445 				 * only. util_fits_cpu() logic requires to
7446 				 * operate on non clamped util but must use the
7447 				 * max-aggregated uclamp_{min, max}.
7448 				 */
7449 				rq_util_min = uclamp_rq_get(rq, UCLAMP_MIN);
7450 				rq_util_max = uclamp_rq_get(rq, UCLAMP_MAX);
7451 
7452 				util_min = max(rq_util_min, p_util_min);
7453 				util_max = max(rq_util_max, p_util_max);
7454 			}
7455 
7456 			fits = util_fits_cpu(util, util_min, util_max, cpu);
7457 			if (!fits)
7458 				continue;
7459 
7460 			lsub_positive(&cpu_cap, util);
7461 
7462 			if (cpu == prev_cpu) {
7463 				/* Always use prev_cpu as a candidate. */
7464 				prev_spare_cap = cpu_cap;
7465 				prev_fits = fits;
7466 			} else if ((fits > max_fits) ||
7467 				   ((fits == max_fits) && (cpu_cap > max_spare_cap))) {
7468 				/*
7469 				 * Find the CPU with the maximum spare capacity
7470 				 * among the remaining CPUs in the performance
7471 				 * domain.
7472 				 */
7473 				max_spare_cap = cpu_cap;
7474 				max_spare_cap_cpu = cpu;
7475 				max_fits = fits;
7476 			}
7477 		}
7478 
7479 		if (max_spare_cap_cpu < 0 && prev_spare_cap == 0)
7480 			continue;
7481 
7482 		eenv_pd_busy_time(&eenv, cpus, p);
7483 		/* Compute the 'base' energy of the pd, without @p */
7484 		base_energy = compute_energy(&eenv, pd, cpus, p, -1);
7485 
7486 		/* Evaluate the energy impact of using prev_cpu. */
7487 		if (prev_spare_cap > 0) {
7488 			prev_delta = compute_energy(&eenv, pd, cpus, p,
7489 						    prev_cpu);
7490 			/* CPU utilization has changed */
7491 			if (prev_delta < base_energy)
7492 				goto unlock;
7493 			prev_delta -= base_energy;
7494 			prev_thermal_cap = cpu_thermal_cap;
7495 			best_delta = min(best_delta, prev_delta);
7496 		}
7497 
7498 		/* Evaluate the energy impact of using max_spare_cap_cpu. */
7499 		if (max_spare_cap_cpu >= 0 && max_spare_cap > prev_spare_cap) {
7500 			/* Current best energy cpu fits better */
7501 			if (max_fits < best_fits)
7502 				continue;
7503 
7504 			/*
7505 			 * Both don't fit performance hint (i.e. uclamp_min)
7506 			 * but best energy cpu has better capacity.
7507 			 */
7508 			if ((max_fits < 0) &&
7509 			    (cpu_thermal_cap <= best_thermal_cap))
7510 				continue;
7511 
7512 			cur_delta = compute_energy(&eenv, pd, cpus, p,
7513 						   max_spare_cap_cpu);
7514 			/* CPU utilization has changed */
7515 			if (cur_delta < base_energy)
7516 				goto unlock;
7517 			cur_delta -= base_energy;
7518 
7519 			/*
7520 			 * Both fit for the task but best energy cpu has lower
7521 			 * energy impact.
7522 			 */
7523 			if ((max_fits > 0) && (best_fits > 0) &&
7524 			    (cur_delta >= best_delta))
7525 				continue;
7526 
7527 			best_delta = cur_delta;
7528 			best_energy_cpu = max_spare_cap_cpu;
7529 			best_fits = max_fits;
7530 			best_thermal_cap = cpu_thermal_cap;
7531 		}
7532 	}
7533 	rcu_read_unlock();
7534 
7535 	if ((best_fits > prev_fits) ||
7536 	    ((best_fits > 0) && (best_delta < prev_delta)) ||
7537 	    ((best_fits < 0) && (best_thermal_cap > prev_thermal_cap)))
7538 		target = best_energy_cpu;
7539 
7540 	return target;
7541 
7542 unlock:
7543 	rcu_read_unlock();
7544 
7545 	return target;
7546 }
7547 
7548 /*
7549  * select_task_rq_fair: Select target runqueue for the waking task in domains
7550  * that have the relevant SD flag set. In practice, this is SD_BALANCE_WAKE,
7551  * SD_BALANCE_FORK, or SD_BALANCE_EXEC.
7552  *
7553  * Balances load by selecting the idlest CPU in the idlest group, or under
7554  * certain conditions an idle sibling CPU if the domain has SD_WAKE_AFFINE set.
7555  *
7556  * Returns the target CPU number.
7557  */
7558 static int
7559 select_task_rq_fair(struct task_struct *p, int prev_cpu, int wake_flags)
7560 {
7561 	int sync = (wake_flags & WF_SYNC) && !(current->flags & PF_EXITING);
7562 	struct sched_domain *tmp, *sd = NULL;
7563 	int cpu = smp_processor_id();
7564 	int new_cpu = prev_cpu;
7565 	int want_affine = 0;
7566 	/* SD_flags and WF_flags share the first nibble */
7567 	int sd_flag = wake_flags & 0xF;
7568 
7569 	/*
7570 	 * required for stable ->cpus_allowed
7571 	 */
7572 	lockdep_assert_held(&p->pi_lock);
7573 	if (wake_flags & WF_TTWU) {
7574 		record_wakee(p);
7575 
7576 		if (sched_energy_enabled()) {
7577 			new_cpu = find_energy_efficient_cpu(p, prev_cpu);
7578 			if (new_cpu >= 0)
7579 				return new_cpu;
7580 			new_cpu = prev_cpu;
7581 		}
7582 
7583 		want_affine = !wake_wide(p) && cpumask_test_cpu(cpu, p->cpus_ptr);
7584 	}
7585 
7586 	rcu_read_lock();
7587 	for_each_domain(cpu, tmp) {
7588 		/*
7589 		 * If both 'cpu' and 'prev_cpu' are part of this domain,
7590 		 * cpu is a valid SD_WAKE_AFFINE target.
7591 		 */
7592 		if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
7593 		    cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
7594 			if (cpu != prev_cpu)
7595 				new_cpu = wake_affine(tmp, p, cpu, prev_cpu, sync);
7596 
7597 			sd = NULL; /* Prefer wake_affine over balance flags */
7598 			break;
7599 		}
7600 
7601 		/*
7602 		 * Usually only true for WF_EXEC and WF_FORK, as sched_domains
7603 		 * usually do not have SD_BALANCE_WAKE set. That means wakeup
7604 		 * will usually go to the fast path.
7605 		 */
7606 		if (tmp->flags & sd_flag)
7607 			sd = tmp;
7608 		else if (!want_affine)
7609 			break;
7610 	}
7611 
7612 	if (unlikely(sd)) {
7613 		/* Slow path */
7614 		new_cpu = find_idlest_cpu(sd, p, cpu, prev_cpu, sd_flag);
7615 	} else if (wake_flags & WF_TTWU) { /* XXX always ? */
7616 		/* Fast path */
7617 		new_cpu = select_idle_sibling(p, prev_cpu, new_cpu);
7618 	}
7619 	rcu_read_unlock();
7620 
7621 	return new_cpu;
7622 }
7623 
7624 /*
7625  * Called immediately before a task is migrated to a new CPU; task_cpu(p) and
7626  * cfs_rq_of(p) references at time of call are still valid and identify the
7627  * previous CPU. The caller guarantees p->pi_lock or task_rq(p)->lock is held.
7628  */
7629 static void migrate_task_rq_fair(struct task_struct *p, int new_cpu)
7630 {
7631 	struct sched_entity *se = &p->se;
7632 
7633 	/*
7634 	 * As blocked tasks retain absolute vruntime the migration needs to
7635 	 * deal with this by subtracting the old and adding the new
7636 	 * min_vruntime -- the latter is done by enqueue_entity() when placing
7637 	 * the task on the new runqueue.
7638 	 */
7639 	if (READ_ONCE(p->__state) == TASK_WAKING) {
7640 		struct cfs_rq *cfs_rq = cfs_rq_of(se);
7641 
7642 		se->vruntime -= u64_u32_load(cfs_rq->min_vruntime);
7643 	}
7644 
7645 	if (!task_on_rq_migrating(p)) {
7646 		remove_entity_load_avg(se);
7647 
7648 		/*
7649 		 * Here, the task's PELT values have been updated according to
7650 		 * the current rq's clock. But if that clock hasn't been
7651 		 * updated in a while, a substantial idle time will be missed,
7652 		 * leading to an inflation after wake-up on the new rq.
7653 		 *
7654 		 * Estimate the missing time from the cfs_rq last_update_time
7655 		 * and update sched_avg to improve the PELT continuity after
7656 		 * migration.
7657 		 */
7658 		migrate_se_pelt_lag(se);
7659 	}
7660 
7661 	/* Tell new CPU we are migrated */
7662 	se->avg.last_update_time = 0;
7663 
7664 	/* We have migrated, no longer consider this task hot */
7665 	se->exec_start = 0;
7666 
7667 	update_scan_period(p, new_cpu);
7668 }
7669 
7670 static void task_dead_fair(struct task_struct *p)
7671 {
7672 	remove_entity_load_avg(&p->se);
7673 }
7674 
7675 static int
7676 balance_fair(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
7677 {
7678 	if (rq->nr_running)
7679 		return 1;
7680 
7681 	return newidle_balance(rq, rf) != 0;
7682 }
7683 #endif /* CONFIG_SMP */
7684 
7685 static unsigned long wakeup_gran(struct sched_entity *se)
7686 {
7687 	unsigned long gran = sysctl_sched_wakeup_granularity;
7688 
7689 	/*
7690 	 * Since its curr running now, convert the gran from real-time
7691 	 * to virtual-time in his units.
7692 	 *
7693 	 * By using 'se' instead of 'curr' we penalize light tasks, so
7694 	 * they get preempted easier. That is, if 'se' < 'curr' then
7695 	 * the resulting gran will be larger, therefore penalizing the
7696 	 * lighter, if otoh 'se' > 'curr' then the resulting gran will
7697 	 * be smaller, again penalizing the lighter task.
7698 	 *
7699 	 * This is especially important for buddies when the leftmost
7700 	 * task is higher priority than the buddy.
7701 	 */
7702 	return calc_delta_fair(gran, se);
7703 }
7704 
7705 /*
7706  * Should 'se' preempt 'curr'.
7707  *
7708  *             |s1
7709  *        |s2
7710  *   |s3
7711  *         g
7712  *      |<--->|c
7713  *
7714  *  w(c, s1) = -1
7715  *  w(c, s2) =  0
7716  *  w(c, s3) =  1
7717  *
7718  */
7719 static int
7720 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
7721 {
7722 	s64 gran, vdiff = curr->vruntime - se->vruntime;
7723 
7724 	if (vdiff <= 0)
7725 		return -1;
7726 
7727 	gran = wakeup_gran(se);
7728 	if (vdiff > gran)
7729 		return 1;
7730 
7731 	return 0;
7732 }
7733 
7734 static void set_last_buddy(struct sched_entity *se)
7735 {
7736 	for_each_sched_entity(se) {
7737 		if (SCHED_WARN_ON(!se->on_rq))
7738 			return;
7739 		if (se_is_idle(se))
7740 			return;
7741 		cfs_rq_of(se)->last = se;
7742 	}
7743 }
7744 
7745 static void set_next_buddy(struct sched_entity *se)
7746 {
7747 	for_each_sched_entity(se) {
7748 		if (SCHED_WARN_ON(!se->on_rq))
7749 			return;
7750 		if (se_is_idle(se))
7751 			return;
7752 		cfs_rq_of(se)->next = se;
7753 	}
7754 }
7755 
7756 static void set_skip_buddy(struct sched_entity *se)
7757 {
7758 	for_each_sched_entity(se)
7759 		cfs_rq_of(se)->skip = se;
7760 }
7761 
7762 /*
7763  * Preempt the current task with a newly woken task if needed:
7764  */
7765 static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
7766 {
7767 	struct task_struct *curr = rq->curr;
7768 	struct sched_entity *se = &curr->se, *pse = &p->se;
7769 	struct cfs_rq *cfs_rq = task_cfs_rq(curr);
7770 	int scale = cfs_rq->nr_running >= sched_nr_latency;
7771 	int next_buddy_marked = 0;
7772 	int cse_is_idle, pse_is_idle;
7773 
7774 	if (unlikely(se == pse))
7775 		return;
7776 
7777 	/*
7778 	 * This is possible from callers such as attach_tasks(), in which we
7779 	 * unconditionally check_preempt_curr() after an enqueue (which may have
7780 	 * lead to a throttle).  This both saves work and prevents false
7781 	 * next-buddy nomination below.
7782 	 */
7783 	if (unlikely(throttled_hierarchy(cfs_rq_of(pse))))
7784 		return;
7785 
7786 	if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
7787 		set_next_buddy(pse);
7788 		next_buddy_marked = 1;
7789 	}
7790 
7791 	/*
7792 	 * We can come here with TIF_NEED_RESCHED already set from new task
7793 	 * wake up path.
7794 	 *
7795 	 * Note: this also catches the edge-case of curr being in a throttled
7796 	 * group (e.g. via set_curr_task), since update_curr() (in the
7797 	 * enqueue of curr) will have resulted in resched being set.  This
7798 	 * prevents us from potentially nominating it as a false LAST_BUDDY
7799 	 * below.
7800 	 */
7801 	if (test_tsk_need_resched(curr))
7802 		return;
7803 
7804 	/* Idle tasks are by definition preempted by non-idle tasks. */
7805 	if (unlikely(task_has_idle_policy(curr)) &&
7806 	    likely(!task_has_idle_policy(p)))
7807 		goto preempt;
7808 
7809 	/*
7810 	 * Batch and idle tasks do not preempt non-idle tasks (their preemption
7811 	 * is driven by the tick):
7812 	 */
7813 	if (unlikely(p->policy != SCHED_NORMAL) || !sched_feat(WAKEUP_PREEMPTION))
7814 		return;
7815 
7816 	find_matching_se(&se, &pse);
7817 	WARN_ON_ONCE(!pse);
7818 
7819 	cse_is_idle = se_is_idle(se);
7820 	pse_is_idle = se_is_idle(pse);
7821 
7822 	/*
7823 	 * Preempt an idle group in favor of a non-idle group (and don't preempt
7824 	 * in the inverse case).
7825 	 */
7826 	if (cse_is_idle && !pse_is_idle)
7827 		goto preempt;
7828 	if (cse_is_idle != pse_is_idle)
7829 		return;
7830 
7831 	update_curr(cfs_rq_of(se));
7832 	if (wakeup_preempt_entity(se, pse) == 1) {
7833 		/*
7834 		 * Bias pick_next to pick the sched entity that is
7835 		 * triggering this preemption.
7836 		 */
7837 		if (!next_buddy_marked)
7838 			set_next_buddy(pse);
7839 		goto preempt;
7840 	}
7841 
7842 	return;
7843 
7844 preempt:
7845 	resched_curr(rq);
7846 	/*
7847 	 * Only set the backward buddy when the current task is still
7848 	 * on the rq. This can happen when a wakeup gets interleaved
7849 	 * with schedule on the ->pre_schedule() or idle_balance()
7850 	 * point, either of which can * drop the rq lock.
7851 	 *
7852 	 * Also, during early boot the idle thread is in the fair class,
7853 	 * for obvious reasons its a bad idea to schedule back to it.
7854 	 */
7855 	if (unlikely(!se->on_rq || curr == rq->idle))
7856 		return;
7857 
7858 	if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
7859 		set_last_buddy(se);
7860 }
7861 
7862 #ifdef CONFIG_SMP
7863 static struct task_struct *pick_task_fair(struct rq *rq)
7864 {
7865 	struct sched_entity *se;
7866 	struct cfs_rq *cfs_rq;
7867 
7868 again:
7869 	cfs_rq = &rq->cfs;
7870 	if (!cfs_rq->nr_running)
7871 		return NULL;
7872 
7873 	do {
7874 		struct sched_entity *curr = cfs_rq->curr;
7875 
7876 		/* When we pick for a remote RQ, we'll not have done put_prev_entity() */
7877 		if (curr) {
7878 			if (curr->on_rq)
7879 				update_curr(cfs_rq);
7880 			else
7881 				curr = NULL;
7882 
7883 			if (unlikely(check_cfs_rq_runtime(cfs_rq)))
7884 				goto again;
7885 		}
7886 
7887 		se = pick_next_entity(cfs_rq, curr);
7888 		cfs_rq = group_cfs_rq(se);
7889 	} while (cfs_rq);
7890 
7891 	return task_of(se);
7892 }
7893 #endif
7894 
7895 struct task_struct *
7896 pick_next_task_fair(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
7897 {
7898 	struct cfs_rq *cfs_rq = &rq->cfs;
7899 	struct sched_entity *se;
7900 	struct task_struct *p;
7901 	int new_tasks;
7902 
7903 again:
7904 	if (!sched_fair_runnable(rq))
7905 		goto idle;
7906 
7907 #ifdef CONFIG_FAIR_GROUP_SCHED
7908 	if (!prev || prev->sched_class != &fair_sched_class)
7909 		goto simple;
7910 
7911 	/*
7912 	 * Because of the set_next_buddy() in dequeue_task_fair() it is rather
7913 	 * likely that a next task is from the same cgroup as the current.
7914 	 *
7915 	 * Therefore attempt to avoid putting and setting the entire cgroup
7916 	 * hierarchy, only change the part that actually changes.
7917 	 */
7918 
7919 	do {
7920 		struct sched_entity *curr = cfs_rq->curr;
7921 
7922 		/*
7923 		 * Since we got here without doing put_prev_entity() we also
7924 		 * have to consider cfs_rq->curr. If it is still a runnable
7925 		 * entity, update_curr() will update its vruntime, otherwise
7926 		 * forget we've ever seen it.
7927 		 */
7928 		if (curr) {
7929 			if (curr->on_rq)
7930 				update_curr(cfs_rq);
7931 			else
7932 				curr = NULL;
7933 
7934 			/*
7935 			 * This call to check_cfs_rq_runtime() will do the
7936 			 * throttle and dequeue its entity in the parent(s).
7937 			 * Therefore the nr_running test will indeed
7938 			 * be correct.
7939 			 */
7940 			if (unlikely(check_cfs_rq_runtime(cfs_rq))) {
7941 				cfs_rq = &rq->cfs;
7942 
7943 				if (!cfs_rq->nr_running)
7944 					goto idle;
7945 
7946 				goto simple;
7947 			}
7948 		}
7949 
7950 		se = pick_next_entity(cfs_rq, curr);
7951 		cfs_rq = group_cfs_rq(se);
7952 	} while (cfs_rq);
7953 
7954 	p = task_of(se);
7955 
7956 	/*
7957 	 * Since we haven't yet done put_prev_entity and if the selected task
7958 	 * is a different task than we started out with, try and touch the
7959 	 * least amount of cfs_rqs.
7960 	 */
7961 	if (prev != p) {
7962 		struct sched_entity *pse = &prev->se;
7963 
7964 		while (!(cfs_rq = is_same_group(se, pse))) {
7965 			int se_depth = se->depth;
7966 			int pse_depth = pse->depth;
7967 
7968 			if (se_depth <= pse_depth) {
7969 				put_prev_entity(cfs_rq_of(pse), pse);
7970 				pse = parent_entity(pse);
7971 			}
7972 			if (se_depth >= pse_depth) {
7973 				set_next_entity(cfs_rq_of(se), se);
7974 				se = parent_entity(se);
7975 			}
7976 		}
7977 
7978 		put_prev_entity(cfs_rq, pse);
7979 		set_next_entity(cfs_rq, se);
7980 	}
7981 
7982 	goto done;
7983 simple:
7984 #endif
7985 	if (prev)
7986 		put_prev_task(rq, prev);
7987 
7988 	do {
7989 		se = pick_next_entity(cfs_rq, NULL);
7990 		set_next_entity(cfs_rq, se);
7991 		cfs_rq = group_cfs_rq(se);
7992 	} while (cfs_rq);
7993 
7994 	p = task_of(se);
7995 
7996 done: __maybe_unused;
7997 #ifdef CONFIG_SMP
7998 	/*
7999 	 * Move the next running task to the front of
8000 	 * the list, so our cfs_tasks list becomes MRU
8001 	 * one.
8002 	 */
8003 	list_move(&p->se.group_node, &rq->cfs_tasks);
8004 #endif
8005 
8006 	if (hrtick_enabled_fair(rq))
8007 		hrtick_start_fair(rq, p);
8008 
8009 	update_misfit_status(p, rq);
8010 
8011 	return p;
8012 
8013 idle:
8014 	if (!rf)
8015 		return NULL;
8016 
8017 	new_tasks = newidle_balance(rq, rf);
8018 
8019 	/*
8020 	 * Because newidle_balance() releases (and re-acquires) rq->lock, it is
8021 	 * possible for any higher priority task to appear. In that case we
8022 	 * must re-start the pick_next_entity() loop.
8023 	 */
8024 	if (new_tasks < 0)
8025 		return RETRY_TASK;
8026 
8027 	if (new_tasks > 0)
8028 		goto again;
8029 
8030 	/*
8031 	 * rq is about to be idle, check if we need to update the
8032 	 * lost_idle_time of clock_pelt
8033 	 */
8034 	update_idle_rq_clock_pelt(rq);
8035 
8036 	return NULL;
8037 }
8038 
8039 static struct task_struct *__pick_next_task_fair(struct rq *rq)
8040 {
8041 	return pick_next_task_fair(rq, NULL, NULL);
8042 }
8043 
8044 /*
8045  * Account for a descheduled task:
8046  */
8047 static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
8048 {
8049 	struct sched_entity *se = &prev->se;
8050 	struct cfs_rq *cfs_rq;
8051 
8052 	for_each_sched_entity(se) {
8053 		cfs_rq = cfs_rq_of(se);
8054 		put_prev_entity(cfs_rq, se);
8055 	}
8056 }
8057 
8058 /*
8059  * sched_yield() is very simple
8060  *
8061  * The magic of dealing with the ->skip buddy is in pick_next_entity.
8062  */
8063 static void yield_task_fair(struct rq *rq)
8064 {
8065 	struct task_struct *curr = rq->curr;
8066 	struct cfs_rq *cfs_rq = task_cfs_rq(curr);
8067 	struct sched_entity *se = &curr->se;
8068 
8069 	/*
8070 	 * Are we the only task in the tree?
8071 	 */
8072 	if (unlikely(rq->nr_running == 1))
8073 		return;
8074 
8075 	clear_buddies(cfs_rq, se);
8076 
8077 	if (curr->policy != SCHED_BATCH) {
8078 		update_rq_clock(rq);
8079 		/*
8080 		 * Update run-time statistics of the 'current'.
8081 		 */
8082 		update_curr(cfs_rq);
8083 		/*
8084 		 * Tell update_rq_clock() that we've just updated,
8085 		 * so we don't do microscopic update in schedule()
8086 		 * and double the fastpath cost.
8087 		 */
8088 		rq_clock_skip_update(rq);
8089 	}
8090 
8091 	set_skip_buddy(se);
8092 }
8093 
8094 static bool yield_to_task_fair(struct rq *rq, struct task_struct *p)
8095 {
8096 	struct sched_entity *se = &p->se;
8097 
8098 	/* throttled hierarchies are not runnable */
8099 	if (!se->on_rq || throttled_hierarchy(cfs_rq_of(se)))
8100 		return false;
8101 
8102 	/* Tell the scheduler that we'd really like pse to run next. */
8103 	set_next_buddy(se);
8104 
8105 	yield_task_fair(rq);
8106 
8107 	return true;
8108 }
8109 
8110 #ifdef CONFIG_SMP
8111 /**************************************************
8112  * Fair scheduling class load-balancing methods.
8113  *
8114  * BASICS
8115  *
8116  * The purpose of load-balancing is to achieve the same basic fairness the
8117  * per-CPU scheduler provides, namely provide a proportional amount of compute
8118  * time to each task. This is expressed in the following equation:
8119  *
8120  *   W_i,n/P_i == W_j,n/P_j for all i,j                               (1)
8121  *
8122  * Where W_i,n is the n-th weight average for CPU i. The instantaneous weight
8123  * W_i,0 is defined as:
8124  *
8125  *   W_i,0 = \Sum_j w_i,j                                             (2)
8126  *
8127  * Where w_i,j is the weight of the j-th runnable task on CPU i. This weight
8128  * is derived from the nice value as per sched_prio_to_weight[].
8129  *
8130  * The weight average is an exponential decay average of the instantaneous
8131  * weight:
8132  *
8133  *   W'_i,n = (2^n - 1) / 2^n * W_i,n + 1 / 2^n * W_i,0               (3)
8134  *
8135  * C_i is the compute capacity of CPU i, typically it is the
8136  * fraction of 'recent' time available for SCHED_OTHER task execution. But it
8137  * can also include other factors [XXX].
8138  *
8139  * To achieve this balance we define a measure of imbalance which follows
8140  * directly from (1):
8141  *
8142  *   imb_i,j = max{ avg(W/C), W_i/C_i } - min{ avg(W/C), W_j/C_j }    (4)
8143  *
8144  * We them move tasks around to minimize the imbalance. In the continuous
8145  * function space it is obvious this converges, in the discrete case we get
8146  * a few fun cases generally called infeasible weight scenarios.
8147  *
8148  * [XXX expand on:
8149  *     - infeasible weights;
8150  *     - local vs global optima in the discrete case. ]
8151  *
8152  *
8153  * SCHED DOMAINS
8154  *
8155  * In order to solve the imbalance equation (4), and avoid the obvious O(n^2)
8156  * for all i,j solution, we create a tree of CPUs that follows the hardware
8157  * topology where each level pairs two lower groups (or better). This results
8158  * in O(log n) layers. Furthermore we reduce the number of CPUs going up the
8159  * tree to only the first of the previous level and we decrease the frequency
8160  * of load-balance at each level inv. proportional to the number of CPUs in
8161  * the groups.
8162  *
8163  * This yields:
8164  *
8165  *     log_2 n     1     n
8166  *   \Sum       { --- * --- * 2^i } = O(n)                            (5)
8167  *     i = 0      2^i   2^i
8168  *                               `- size of each group
8169  *         |         |     `- number of CPUs doing load-balance
8170  *         |         `- freq
8171  *         `- sum over all levels
8172  *
8173  * Coupled with a limit on how many tasks we can migrate every balance pass,
8174  * this makes (5) the runtime complexity of the balancer.
8175  *
8176  * An important property here is that each CPU is still (indirectly) connected
8177  * to every other CPU in at most O(log n) steps:
8178  *
8179  * The adjacency matrix of the resulting graph is given by:
8180  *
8181  *             log_2 n
8182  *   A_i,j = \Union     (i % 2^k == 0) && i / 2^(k+1) == j / 2^(k+1)  (6)
8183  *             k = 0
8184  *
8185  * And you'll find that:
8186  *
8187  *   A^(log_2 n)_i,j != 0  for all i,j                                (7)
8188  *
8189  * Showing there's indeed a path between every CPU in at most O(log n) steps.
8190  * The task movement gives a factor of O(m), giving a convergence complexity
8191  * of:
8192  *
8193  *   O(nm log n),  n := nr_cpus, m := nr_tasks                        (8)
8194  *
8195  *
8196  * WORK CONSERVING
8197  *
8198  * In order to avoid CPUs going idle while there's still work to do, new idle
8199  * balancing is more aggressive and has the newly idle CPU iterate up the domain
8200  * tree itself instead of relying on other CPUs to bring it work.
8201  *
8202  * This adds some complexity to both (5) and (8) but it reduces the total idle
8203  * time.
8204  *
8205  * [XXX more?]
8206  *
8207  *
8208  * CGROUPS
8209  *
8210  * Cgroups make a horror show out of (2), instead of a simple sum we get:
8211  *
8212  *                                s_k,i
8213  *   W_i,0 = \Sum_j \Prod_k w_k * -----                               (9)
8214  *                                 S_k
8215  *
8216  * Where
8217  *
8218  *   s_k,i = \Sum_j w_i,j,k  and  S_k = \Sum_i s_k,i                 (10)
8219  *
8220  * w_i,j,k is the weight of the j-th runnable task in the k-th cgroup on CPU i.
8221  *
8222  * The big problem is S_k, its a global sum needed to compute a local (W_i)
8223  * property.
8224  *
8225  * [XXX write more on how we solve this.. _after_ merging pjt's patches that
8226  *      rewrite all of this once again.]
8227  */
8228 
8229 static unsigned long __read_mostly max_load_balance_interval = HZ/10;
8230 
8231 enum fbq_type { regular, remote, all };
8232 
8233 /*
8234  * 'group_type' describes the group of CPUs at the moment of load balancing.
8235  *
8236  * The enum is ordered by pulling priority, with the group with lowest priority
8237  * first so the group_type can simply be compared when selecting the busiest
8238  * group. See update_sd_pick_busiest().
8239  */
8240 enum group_type {
8241 	/* The group has spare capacity that can be used to run more tasks.  */
8242 	group_has_spare = 0,
8243 	/*
8244 	 * The group is fully used and the tasks don't compete for more CPU
8245 	 * cycles. Nevertheless, some tasks might wait before running.
8246 	 */
8247 	group_fully_busy,
8248 	/*
8249 	 * One task doesn't fit with CPU's capacity and must be migrated to a
8250 	 * more powerful CPU.
8251 	 */
8252 	group_misfit_task,
8253 	/*
8254 	 * SD_ASYM_PACKING only: One local CPU with higher capacity is available,
8255 	 * and the task should be migrated to it instead of running on the
8256 	 * current CPU.
8257 	 */
8258 	group_asym_packing,
8259 	/*
8260 	 * The tasks' affinity constraints previously prevented the scheduler
8261 	 * from balancing the load across the system.
8262 	 */
8263 	group_imbalanced,
8264 	/*
8265 	 * The CPU is overloaded and can't provide expected CPU cycles to all
8266 	 * tasks.
8267 	 */
8268 	group_overloaded
8269 };
8270 
8271 enum migration_type {
8272 	migrate_load = 0,
8273 	migrate_util,
8274 	migrate_task,
8275 	migrate_misfit
8276 };
8277 
8278 #define LBF_ALL_PINNED	0x01
8279 #define LBF_NEED_BREAK	0x02
8280 #define LBF_DST_PINNED  0x04
8281 #define LBF_SOME_PINNED	0x08
8282 #define LBF_ACTIVE_LB	0x10
8283 
8284 struct lb_env {
8285 	struct sched_domain	*sd;
8286 
8287 	struct rq		*src_rq;
8288 	int			src_cpu;
8289 
8290 	int			dst_cpu;
8291 	struct rq		*dst_rq;
8292 
8293 	struct cpumask		*dst_grpmask;
8294 	int			new_dst_cpu;
8295 	enum cpu_idle_type	idle;
8296 	long			imbalance;
8297 	/* The set of CPUs under consideration for load-balancing */
8298 	struct cpumask		*cpus;
8299 
8300 	unsigned int		flags;
8301 
8302 	unsigned int		loop;
8303 	unsigned int		loop_break;
8304 	unsigned int		loop_max;
8305 
8306 	enum fbq_type		fbq_type;
8307 	enum migration_type	migration_type;
8308 	struct list_head	tasks;
8309 };
8310 
8311 /*
8312  * Is this task likely cache-hot:
8313  */
8314 static int task_hot(struct task_struct *p, struct lb_env *env)
8315 {
8316 	s64 delta;
8317 
8318 	lockdep_assert_rq_held(env->src_rq);
8319 
8320 	if (p->sched_class != &fair_sched_class)
8321 		return 0;
8322 
8323 	if (unlikely(task_has_idle_policy(p)))
8324 		return 0;
8325 
8326 	/* SMT siblings share cache */
8327 	if (env->sd->flags & SD_SHARE_CPUCAPACITY)
8328 		return 0;
8329 
8330 	/*
8331 	 * Buddy candidates are cache hot:
8332 	 */
8333 	if (sched_feat(CACHE_HOT_BUDDY) && env->dst_rq->nr_running &&
8334 			(&p->se == cfs_rq_of(&p->se)->next ||
8335 			 &p->se == cfs_rq_of(&p->se)->last))
8336 		return 1;
8337 
8338 	if (sysctl_sched_migration_cost == -1)
8339 		return 1;
8340 
8341 	/*
8342 	 * Don't migrate task if the task's cookie does not match
8343 	 * with the destination CPU's core cookie.
8344 	 */
8345 	if (!sched_core_cookie_match(cpu_rq(env->dst_cpu), p))
8346 		return 1;
8347 
8348 	if (sysctl_sched_migration_cost == 0)
8349 		return 0;
8350 
8351 	delta = rq_clock_task(env->src_rq) - p->se.exec_start;
8352 
8353 	return delta < (s64)sysctl_sched_migration_cost;
8354 }
8355 
8356 #ifdef CONFIG_NUMA_BALANCING
8357 /*
8358  * Returns 1, if task migration degrades locality
8359  * Returns 0, if task migration improves locality i.e migration preferred.
8360  * Returns -1, if task migration is not affected by locality.
8361  */
8362 static int migrate_degrades_locality(struct task_struct *p, struct lb_env *env)
8363 {
8364 	struct numa_group *numa_group = rcu_dereference(p->numa_group);
8365 	unsigned long src_weight, dst_weight;
8366 	int src_nid, dst_nid, dist;
8367 
8368 	if (!static_branch_likely(&sched_numa_balancing))
8369 		return -1;
8370 
8371 	if (!p->numa_faults || !(env->sd->flags & SD_NUMA))
8372 		return -1;
8373 
8374 	src_nid = cpu_to_node(env->src_cpu);
8375 	dst_nid = cpu_to_node(env->dst_cpu);
8376 
8377 	if (src_nid == dst_nid)
8378 		return -1;
8379 
8380 	/* Migrating away from the preferred node is always bad. */
8381 	if (src_nid == p->numa_preferred_nid) {
8382 		if (env->src_rq->nr_running > env->src_rq->nr_preferred_running)
8383 			return 1;
8384 		else
8385 			return -1;
8386 	}
8387 
8388 	/* Encourage migration to the preferred node. */
8389 	if (dst_nid == p->numa_preferred_nid)
8390 		return 0;
8391 
8392 	/* Leaving a core idle is often worse than degrading locality. */
8393 	if (env->idle == CPU_IDLE)
8394 		return -1;
8395 
8396 	dist = node_distance(src_nid, dst_nid);
8397 	if (numa_group) {
8398 		src_weight = group_weight(p, src_nid, dist);
8399 		dst_weight = group_weight(p, dst_nid, dist);
8400 	} else {
8401 		src_weight = task_weight(p, src_nid, dist);
8402 		dst_weight = task_weight(p, dst_nid, dist);
8403 	}
8404 
8405 	return dst_weight < src_weight;
8406 }
8407 
8408 #else
8409 static inline int migrate_degrades_locality(struct task_struct *p,
8410 					     struct lb_env *env)
8411 {
8412 	return -1;
8413 }
8414 #endif
8415 
8416 /*
8417  * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
8418  */
8419 static
8420 int can_migrate_task(struct task_struct *p, struct lb_env *env)
8421 {
8422 	int tsk_cache_hot;
8423 
8424 	lockdep_assert_rq_held(env->src_rq);
8425 
8426 	/*
8427 	 * We do not migrate tasks that are:
8428 	 * 1) throttled_lb_pair, or
8429 	 * 2) cannot be migrated to this CPU due to cpus_ptr, or
8430 	 * 3) running (obviously), or
8431 	 * 4) are cache-hot on their current CPU.
8432 	 */
8433 	if (throttled_lb_pair(task_group(p), env->src_cpu, env->dst_cpu))
8434 		return 0;
8435 
8436 	/* Disregard pcpu kthreads; they are where they need to be. */
8437 	if (kthread_is_per_cpu(p))
8438 		return 0;
8439 
8440 	if (!cpumask_test_cpu(env->dst_cpu, p->cpus_ptr)) {
8441 		int cpu;
8442 
8443 		schedstat_inc(p->stats.nr_failed_migrations_affine);
8444 
8445 		env->flags |= LBF_SOME_PINNED;
8446 
8447 		/*
8448 		 * Remember if this task can be migrated to any other CPU in
8449 		 * our sched_group. We may want to revisit it if we couldn't
8450 		 * meet load balance goals by pulling other tasks on src_cpu.
8451 		 *
8452 		 * Avoid computing new_dst_cpu
8453 		 * - for NEWLY_IDLE
8454 		 * - if we have already computed one in current iteration
8455 		 * - if it's an active balance
8456 		 */
8457 		if (env->idle == CPU_NEWLY_IDLE ||
8458 		    env->flags & (LBF_DST_PINNED | LBF_ACTIVE_LB))
8459 			return 0;
8460 
8461 		/* Prevent to re-select dst_cpu via env's CPUs: */
8462 		for_each_cpu_and(cpu, env->dst_grpmask, env->cpus) {
8463 			if (cpumask_test_cpu(cpu, p->cpus_ptr)) {
8464 				env->flags |= LBF_DST_PINNED;
8465 				env->new_dst_cpu = cpu;
8466 				break;
8467 			}
8468 		}
8469 
8470 		return 0;
8471 	}
8472 
8473 	/* Record that we found at least one task that could run on dst_cpu */
8474 	env->flags &= ~LBF_ALL_PINNED;
8475 
8476 	if (task_on_cpu(env->src_rq, p)) {
8477 		schedstat_inc(p->stats.nr_failed_migrations_running);
8478 		return 0;
8479 	}
8480 
8481 	/*
8482 	 * Aggressive migration if:
8483 	 * 1) active balance
8484 	 * 2) destination numa is preferred
8485 	 * 3) task is cache cold, or
8486 	 * 4) too many balance attempts have failed.
8487 	 */
8488 	if (env->flags & LBF_ACTIVE_LB)
8489 		return 1;
8490 
8491 	tsk_cache_hot = migrate_degrades_locality(p, env);
8492 	if (tsk_cache_hot == -1)
8493 		tsk_cache_hot = task_hot(p, env);
8494 
8495 	if (tsk_cache_hot <= 0 ||
8496 	    env->sd->nr_balance_failed > env->sd->cache_nice_tries) {
8497 		if (tsk_cache_hot == 1) {
8498 			schedstat_inc(env->sd->lb_hot_gained[env->idle]);
8499 			schedstat_inc(p->stats.nr_forced_migrations);
8500 		}
8501 		return 1;
8502 	}
8503 
8504 	schedstat_inc(p->stats.nr_failed_migrations_hot);
8505 	return 0;
8506 }
8507 
8508 /*
8509  * detach_task() -- detach the task for the migration specified in env
8510  */
8511 static void detach_task(struct task_struct *p, struct lb_env *env)
8512 {
8513 	lockdep_assert_rq_held(env->src_rq);
8514 
8515 	deactivate_task(env->src_rq, p, DEQUEUE_NOCLOCK);
8516 	set_task_cpu(p, env->dst_cpu);
8517 }
8518 
8519 /*
8520  * detach_one_task() -- tries to dequeue exactly one task from env->src_rq, as
8521  * part of active balancing operations within "domain".
8522  *
8523  * Returns a task if successful and NULL otherwise.
8524  */
8525 static struct task_struct *detach_one_task(struct lb_env *env)
8526 {
8527 	struct task_struct *p;
8528 
8529 	lockdep_assert_rq_held(env->src_rq);
8530 
8531 	list_for_each_entry_reverse(p,
8532 			&env->src_rq->cfs_tasks, se.group_node) {
8533 		if (!can_migrate_task(p, env))
8534 			continue;
8535 
8536 		detach_task(p, env);
8537 
8538 		/*
8539 		 * Right now, this is only the second place where
8540 		 * lb_gained[env->idle] is updated (other is detach_tasks)
8541 		 * so we can safely collect stats here rather than
8542 		 * inside detach_tasks().
8543 		 */
8544 		schedstat_inc(env->sd->lb_gained[env->idle]);
8545 		return p;
8546 	}
8547 	return NULL;
8548 }
8549 
8550 /*
8551  * detach_tasks() -- tries to detach up to imbalance load/util/tasks from
8552  * busiest_rq, as part of a balancing operation within domain "sd".
8553  *
8554  * Returns number of detached tasks if successful and 0 otherwise.
8555  */
8556 static int detach_tasks(struct lb_env *env)
8557 {
8558 	struct list_head *tasks = &env->src_rq->cfs_tasks;
8559 	unsigned long util, load;
8560 	struct task_struct *p;
8561 	int detached = 0;
8562 
8563 	lockdep_assert_rq_held(env->src_rq);
8564 
8565 	/*
8566 	 * Source run queue has been emptied by another CPU, clear
8567 	 * LBF_ALL_PINNED flag as we will not test any task.
8568 	 */
8569 	if (env->src_rq->nr_running <= 1) {
8570 		env->flags &= ~LBF_ALL_PINNED;
8571 		return 0;
8572 	}
8573 
8574 	if (env->imbalance <= 0)
8575 		return 0;
8576 
8577 	while (!list_empty(tasks)) {
8578 		/*
8579 		 * We don't want to steal all, otherwise we may be treated likewise,
8580 		 * which could at worst lead to a livelock crash.
8581 		 */
8582 		if (env->idle != CPU_NOT_IDLE && env->src_rq->nr_running <= 1)
8583 			break;
8584 
8585 		env->loop++;
8586 		/*
8587 		 * We've more or less seen every task there is, call it quits
8588 		 * unless we haven't found any movable task yet.
8589 		 */
8590 		if (env->loop > env->loop_max &&
8591 		    !(env->flags & LBF_ALL_PINNED))
8592 			break;
8593 
8594 		/* take a breather every nr_migrate tasks */
8595 		if (env->loop > env->loop_break) {
8596 			env->loop_break += SCHED_NR_MIGRATE_BREAK;
8597 			env->flags |= LBF_NEED_BREAK;
8598 			break;
8599 		}
8600 
8601 		p = list_last_entry(tasks, struct task_struct, se.group_node);
8602 
8603 		if (!can_migrate_task(p, env))
8604 			goto next;
8605 
8606 		switch (env->migration_type) {
8607 		case migrate_load:
8608 			/*
8609 			 * Depending of the number of CPUs and tasks and the
8610 			 * cgroup hierarchy, task_h_load() can return a null
8611 			 * value. Make sure that env->imbalance decreases
8612 			 * otherwise detach_tasks() will stop only after
8613 			 * detaching up to loop_max tasks.
8614 			 */
8615 			load = max_t(unsigned long, task_h_load(p), 1);
8616 
8617 			if (sched_feat(LB_MIN) &&
8618 			    load < 16 && !env->sd->nr_balance_failed)
8619 				goto next;
8620 
8621 			/*
8622 			 * Make sure that we don't migrate too much load.
8623 			 * Nevertheless, let relax the constraint if
8624 			 * scheduler fails to find a good waiting task to
8625 			 * migrate.
8626 			 */
8627 			if (shr_bound(load, env->sd->nr_balance_failed) > env->imbalance)
8628 				goto next;
8629 
8630 			env->imbalance -= load;
8631 			break;
8632 
8633 		case migrate_util:
8634 			util = task_util_est(p);
8635 
8636 			if (util > env->imbalance)
8637 				goto next;
8638 
8639 			env->imbalance -= util;
8640 			break;
8641 
8642 		case migrate_task:
8643 			env->imbalance--;
8644 			break;
8645 
8646 		case migrate_misfit:
8647 			/* This is not a misfit task */
8648 			if (task_fits_cpu(p, env->src_cpu))
8649 				goto next;
8650 
8651 			env->imbalance = 0;
8652 			break;
8653 		}
8654 
8655 		detach_task(p, env);
8656 		list_add(&p->se.group_node, &env->tasks);
8657 
8658 		detached++;
8659 
8660 #ifdef CONFIG_PREEMPTION
8661 		/*
8662 		 * NEWIDLE balancing is a source of latency, so preemptible
8663 		 * kernels will stop after the first task is detached to minimize
8664 		 * the critical section.
8665 		 */
8666 		if (env->idle == CPU_NEWLY_IDLE)
8667 			break;
8668 #endif
8669 
8670 		/*
8671 		 * We only want to steal up to the prescribed amount of
8672 		 * load/util/tasks.
8673 		 */
8674 		if (env->imbalance <= 0)
8675 			break;
8676 
8677 		continue;
8678 next:
8679 		list_move(&p->se.group_node, tasks);
8680 	}
8681 
8682 	/*
8683 	 * Right now, this is one of only two places we collect this stat
8684 	 * so we can safely collect detach_one_task() stats here rather
8685 	 * than inside detach_one_task().
8686 	 */
8687 	schedstat_add(env->sd->lb_gained[env->idle], detached);
8688 
8689 	return detached;
8690 }
8691 
8692 /*
8693  * attach_task() -- attach the task detached by detach_task() to its new rq.
8694  */
8695 static void attach_task(struct rq *rq, struct task_struct *p)
8696 {
8697 	lockdep_assert_rq_held(rq);
8698 
8699 	WARN_ON_ONCE(task_rq(p) != rq);
8700 	activate_task(rq, p, ENQUEUE_NOCLOCK);
8701 	check_preempt_curr(rq, p, 0);
8702 }
8703 
8704 /*
8705  * attach_one_task() -- attaches the task returned from detach_one_task() to
8706  * its new rq.
8707  */
8708 static void attach_one_task(struct rq *rq, struct task_struct *p)
8709 {
8710 	struct rq_flags rf;
8711 
8712 	rq_lock(rq, &rf);
8713 	update_rq_clock(rq);
8714 	attach_task(rq, p);
8715 	rq_unlock(rq, &rf);
8716 }
8717 
8718 /*
8719  * attach_tasks() -- attaches all tasks detached by detach_tasks() to their
8720  * new rq.
8721  */
8722 static void attach_tasks(struct lb_env *env)
8723 {
8724 	struct list_head *tasks = &env->tasks;
8725 	struct task_struct *p;
8726 	struct rq_flags rf;
8727 
8728 	rq_lock(env->dst_rq, &rf);
8729 	update_rq_clock(env->dst_rq);
8730 
8731 	while (!list_empty(tasks)) {
8732 		p = list_first_entry(tasks, struct task_struct, se.group_node);
8733 		list_del_init(&p->se.group_node);
8734 
8735 		attach_task(env->dst_rq, p);
8736 	}
8737 
8738 	rq_unlock(env->dst_rq, &rf);
8739 }
8740 
8741 #ifdef CONFIG_NO_HZ_COMMON
8742 static inline bool cfs_rq_has_blocked(struct cfs_rq *cfs_rq)
8743 {
8744 	if (cfs_rq->avg.load_avg)
8745 		return true;
8746 
8747 	if (cfs_rq->avg.util_avg)
8748 		return true;
8749 
8750 	return false;
8751 }
8752 
8753 static inline bool others_have_blocked(struct rq *rq)
8754 {
8755 	if (READ_ONCE(rq->avg_rt.util_avg))
8756 		return true;
8757 
8758 	if (READ_ONCE(rq->avg_dl.util_avg))
8759 		return true;
8760 
8761 	if (thermal_load_avg(rq))
8762 		return true;
8763 
8764 #ifdef CONFIG_HAVE_SCHED_AVG_IRQ
8765 	if (READ_ONCE(rq->avg_irq.util_avg))
8766 		return true;
8767 #endif
8768 
8769 	return false;
8770 }
8771 
8772 static inline void update_blocked_load_tick(struct rq *rq)
8773 {
8774 	WRITE_ONCE(rq->last_blocked_load_update_tick, jiffies);
8775 }
8776 
8777 static inline void update_blocked_load_status(struct rq *rq, bool has_blocked)
8778 {
8779 	if (!has_blocked)
8780 		rq->has_blocked_load = 0;
8781 }
8782 #else
8783 static inline bool cfs_rq_has_blocked(struct cfs_rq *cfs_rq) { return false; }
8784 static inline bool others_have_blocked(struct rq *rq) { return false; }
8785 static inline void update_blocked_load_tick(struct rq *rq) {}
8786 static inline void update_blocked_load_status(struct rq *rq, bool has_blocked) {}
8787 #endif
8788 
8789 static bool __update_blocked_others(struct rq *rq, bool *done)
8790 {
8791 	const struct sched_class *curr_class;
8792 	u64 now = rq_clock_pelt(rq);
8793 	unsigned long thermal_pressure;
8794 	bool decayed;
8795 
8796 	/*
8797 	 * update_load_avg() can call cpufreq_update_util(). Make sure that RT,
8798 	 * DL and IRQ signals have been updated before updating CFS.
8799 	 */
8800 	curr_class = rq->curr->sched_class;
8801 
8802 	thermal_pressure = arch_scale_thermal_pressure(cpu_of(rq));
8803 
8804 	decayed = update_rt_rq_load_avg(now, rq, curr_class == &rt_sched_class) |
8805 		  update_dl_rq_load_avg(now, rq, curr_class == &dl_sched_class) |
8806 		  update_thermal_load_avg(rq_clock_thermal(rq), rq, thermal_pressure) |
8807 		  update_irq_load_avg(rq, 0);
8808 
8809 	if (others_have_blocked(rq))
8810 		*done = false;
8811 
8812 	return decayed;
8813 }
8814 
8815 #ifdef CONFIG_FAIR_GROUP_SCHED
8816 
8817 static bool __update_blocked_fair(struct rq *rq, bool *done)
8818 {
8819 	struct cfs_rq *cfs_rq, *pos;
8820 	bool decayed = false;
8821 	int cpu = cpu_of(rq);
8822 
8823 	/*
8824 	 * Iterates the task_group tree in a bottom up fashion, see
8825 	 * list_add_leaf_cfs_rq() for details.
8826 	 */
8827 	for_each_leaf_cfs_rq_safe(rq, cfs_rq, pos) {
8828 		struct sched_entity *se;
8829 
8830 		if (update_cfs_rq_load_avg(cfs_rq_clock_pelt(cfs_rq), cfs_rq)) {
8831 			update_tg_load_avg(cfs_rq);
8832 
8833 			if (cfs_rq->nr_running == 0)
8834 				update_idle_cfs_rq_clock_pelt(cfs_rq);
8835 
8836 			if (cfs_rq == &rq->cfs)
8837 				decayed = true;
8838 		}
8839 
8840 		/* Propagate pending load changes to the parent, if any: */
8841 		se = cfs_rq->tg->se[cpu];
8842 		if (se && !skip_blocked_update(se))
8843 			update_load_avg(cfs_rq_of(se), se, UPDATE_TG);
8844 
8845 		/*
8846 		 * There can be a lot of idle CPU cgroups.  Don't let fully
8847 		 * decayed cfs_rqs linger on the list.
8848 		 */
8849 		if (cfs_rq_is_decayed(cfs_rq))
8850 			list_del_leaf_cfs_rq(cfs_rq);
8851 
8852 		/* Don't need periodic decay once load/util_avg are null */
8853 		if (cfs_rq_has_blocked(cfs_rq))
8854 			*done = false;
8855 	}
8856 
8857 	return decayed;
8858 }
8859 
8860 /*
8861  * Compute the hierarchical load factor for cfs_rq and all its ascendants.
8862  * This needs to be done in a top-down fashion because the load of a child
8863  * group is a fraction of its parents load.
8864  */
8865 static void update_cfs_rq_h_load(struct cfs_rq *cfs_rq)
8866 {
8867 	struct rq *rq = rq_of(cfs_rq);
8868 	struct sched_entity *se = cfs_rq->tg->se[cpu_of(rq)];
8869 	unsigned long now = jiffies;
8870 	unsigned long load;
8871 
8872 	if (cfs_rq->last_h_load_update == now)
8873 		return;
8874 
8875 	WRITE_ONCE(cfs_rq->h_load_next, NULL);
8876 	for_each_sched_entity(se) {
8877 		cfs_rq = cfs_rq_of(se);
8878 		WRITE_ONCE(cfs_rq->h_load_next, se);
8879 		if (cfs_rq->last_h_load_update == now)
8880 			break;
8881 	}
8882 
8883 	if (!se) {
8884 		cfs_rq->h_load = cfs_rq_load_avg(cfs_rq);
8885 		cfs_rq->last_h_load_update = now;
8886 	}
8887 
8888 	while ((se = READ_ONCE(cfs_rq->h_load_next)) != NULL) {
8889 		load = cfs_rq->h_load;
8890 		load = div64_ul(load * se->avg.load_avg,
8891 			cfs_rq_load_avg(cfs_rq) + 1);
8892 		cfs_rq = group_cfs_rq(se);
8893 		cfs_rq->h_load = load;
8894 		cfs_rq->last_h_load_update = now;
8895 	}
8896 }
8897 
8898 static unsigned long task_h_load(struct task_struct *p)
8899 {
8900 	struct cfs_rq *cfs_rq = task_cfs_rq(p);
8901 
8902 	update_cfs_rq_h_load(cfs_rq);
8903 	return div64_ul(p->se.avg.load_avg * cfs_rq->h_load,
8904 			cfs_rq_load_avg(cfs_rq) + 1);
8905 }
8906 #else
8907 static bool __update_blocked_fair(struct rq *rq, bool *done)
8908 {
8909 	struct cfs_rq *cfs_rq = &rq->cfs;
8910 	bool decayed;
8911 
8912 	decayed = update_cfs_rq_load_avg(cfs_rq_clock_pelt(cfs_rq), cfs_rq);
8913 	if (cfs_rq_has_blocked(cfs_rq))
8914 		*done = false;
8915 
8916 	return decayed;
8917 }
8918 
8919 static unsigned long task_h_load(struct task_struct *p)
8920 {
8921 	return p->se.avg.load_avg;
8922 }
8923 #endif
8924 
8925 static void update_blocked_averages(int cpu)
8926 {
8927 	bool decayed = false, done = true;
8928 	struct rq *rq = cpu_rq(cpu);
8929 	struct rq_flags rf;
8930 
8931 	rq_lock_irqsave(rq, &rf);
8932 	update_blocked_load_tick(rq);
8933 	update_rq_clock(rq);
8934 
8935 	decayed |= __update_blocked_others(rq, &done);
8936 	decayed |= __update_blocked_fair(rq, &done);
8937 
8938 	update_blocked_load_status(rq, !done);
8939 	if (decayed)
8940 		cpufreq_update_util(rq, 0);
8941 	rq_unlock_irqrestore(rq, &rf);
8942 }
8943 
8944 /********** Helpers for find_busiest_group ************************/
8945 
8946 /*
8947  * sg_lb_stats - stats of a sched_group required for load_balancing
8948  */
8949 struct sg_lb_stats {
8950 	unsigned long avg_load; /*Avg load across the CPUs of the group */
8951 	unsigned long group_load; /* Total load over the CPUs of the group */
8952 	unsigned long group_capacity;
8953 	unsigned long group_util; /* Total utilization over the CPUs of the group */
8954 	unsigned long group_runnable; /* Total runnable time over the CPUs of the group */
8955 	unsigned int sum_nr_running; /* Nr of tasks running in the group */
8956 	unsigned int sum_h_nr_running; /* Nr of CFS tasks running in the group */
8957 	unsigned int idle_cpus;
8958 	unsigned int group_weight;
8959 	enum group_type group_type;
8960 	unsigned int group_asym_packing; /* Tasks should be moved to preferred CPU */
8961 	unsigned long group_misfit_task_load; /* A CPU has a task too big for its capacity */
8962 #ifdef CONFIG_NUMA_BALANCING
8963 	unsigned int nr_numa_running;
8964 	unsigned int nr_preferred_running;
8965 #endif
8966 };
8967 
8968 /*
8969  * sd_lb_stats - Structure to store the statistics of a sched_domain
8970  *		 during load balancing.
8971  */
8972 struct sd_lb_stats {
8973 	struct sched_group *busiest;	/* Busiest group in this sd */
8974 	struct sched_group *local;	/* Local group in this sd */
8975 	unsigned long total_load;	/* Total load of all groups in sd */
8976 	unsigned long total_capacity;	/* Total capacity of all groups in sd */
8977 	unsigned long avg_load;	/* Average load across all groups in sd */
8978 	unsigned int prefer_sibling; /* tasks should go to sibling first */
8979 
8980 	struct sg_lb_stats busiest_stat;/* Statistics of the busiest group */
8981 	struct sg_lb_stats local_stat;	/* Statistics of the local group */
8982 };
8983 
8984 static inline void init_sd_lb_stats(struct sd_lb_stats *sds)
8985 {
8986 	/*
8987 	 * Skimp on the clearing to avoid duplicate work. We can avoid clearing
8988 	 * local_stat because update_sg_lb_stats() does a full clear/assignment.
8989 	 * We must however set busiest_stat::group_type and
8990 	 * busiest_stat::idle_cpus to the worst busiest group because
8991 	 * update_sd_pick_busiest() reads these before assignment.
8992 	 */
8993 	*sds = (struct sd_lb_stats){
8994 		.busiest = NULL,
8995 		.local = NULL,
8996 		.total_load = 0UL,
8997 		.total_capacity = 0UL,
8998 		.busiest_stat = {
8999 			.idle_cpus = UINT_MAX,
9000 			.group_type = group_has_spare,
9001 		},
9002 	};
9003 }
9004 
9005 static unsigned long scale_rt_capacity(int cpu)
9006 {
9007 	struct rq *rq = cpu_rq(cpu);
9008 	unsigned long max = arch_scale_cpu_capacity(cpu);
9009 	unsigned long used, free;
9010 	unsigned long irq;
9011 
9012 	irq = cpu_util_irq(rq);
9013 
9014 	if (unlikely(irq >= max))
9015 		return 1;
9016 
9017 	/*
9018 	 * avg_rt.util_avg and avg_dl.util_avg track binary signals
9019 	 * (running and not running) with weights 0 and 1024 respectively.
9020 	 * avg_thermal.load_avg tracks thermal pressure and the weighted
9021 	 * average uses the actual delta max capacity(load).
9022 	 */
9023 	used = READ_ONCE(rq->avg_rt.util_avg);
9024 	used += READ_ONCE(rq->avg_dl.util_avg);
9025 	used += thermal_load_avg(rq);
9026 
9027 	if (unlikely(used >= max))
9028 		return 1;
9029 
9030 	free = max - used;
9031 
9032 	return scale_irq_capacity(free, irq, max);
9033 }
9034 
9035 static void update_cpu_capacity(struct sched_domain *sd, int cpu)
9036 {
9037 	unsigned long capacity = scale_rt_capacity(cpu);
9038 	struct sched_group *sdg = sd->groups;
9039 
9040 	cpu_rq(cpu)->cpu_capacity_orig = arch_scale_cpu_capacity(cpu);
9041 
9042 	if (!capacity)
9043 		capacity = 1;
9044 
9045 	cpu_rq(cpu)->cpu_capacity = capacity;
9046 	trace_sched_cpu_capacity_tp(cpu_rq(cpu));
9047 
9048 	sdg->sgc->capacity = capacity;
9049 	sdg->sgc->min_capacity = capacity;
9050 	sdg->sgc->max_capacity = capacity;
9051 }
9052 
9053 void update_group_capacity(struct sched_domain *sd, int cpu)
9054 {
9055 	struct sched_domain *child = sd->child;
9056 	struct sched_group *group, *sdg = sd->groups;
9057 	unsigned long capacity, min_capacity, max_capacity;
9058 	unsigned long interval;
9059 
9060 	interval = msecs_to_jiffies(sd->balance_interval);
9061 	interval = clamp(interval, 1UL, max_load_balance_interval);
9062 	sdg->sgc->next_update = jiffies + interval;
9063 
9064 	if (!child) {
9065 		update_cpu_capacity(sd, cpu);
9066 		return;
9067 	}
9068 
9069 	capacity = 0;
9070 	min_capacity = ULONG_MAX;
9071 	max_capacity = 0;
9072 
9073 	if (child->flags & SD_OVERLAP) {
9074 		/*
9075 		 * SD_OVERLAP domains cannot assume that child groups
9076 		 * span the current group.
9077 		 */
9078 
9079 		for_each_cpu(cpu, sched_group_span(sdg)) {
9080 			unsigned long cpu_cap = capacity_of(cpu);
9081 
9082 			capacity += cpu_cap;
9083 			min_capacity = min(cpu_cap, min_capacity);
9084 			max_capacity = max(cpu_cap, max_capacity);
9085 		}
9086 	} else  {
9087 		/*
9088 		 * !SD_OVERLAP domains can assume that child groups
9089 		 * span the current group.
9090 		 */
9091 
9092 		group = child->groups;
9093 		do {
9094 			struct sched_group_capacity *sgc = group->sgc;
9095 
9096 			capacity += sgc->capacity;
9097 			min_capacity = min(sgc->min_capacity, min_capacity);
9098 			max_capacity = max(sgc->max_capacity, max_capacity);
9099 			group = group->next;
9100 		} while (group != child->groups);
9101 	}
9102 
9103 	sdg->sgc->capacity = capacity;
9104 	sdg->sgc->min_capacity = min_capacity;
9105 	sdg->sgc->max_capacity = max_capacity;
9106 }
9107 
9108 /*
9109  * Check whether the capacity of the rq has been noticeably reduced by side
9110  * activity. The imbalance_pct is used for the threshold.
9111  * Return true is the capacity is reduced
9112  */
9113 static inline int
9114 check_cpu_capacity(struct rq *rq, struct sched_domain *sd)
9115 {
9116 	return ((rq->cpu_capacity * sd->imbalance_pct) <
9117 				(rq->cpu_capacity_orig * 100));
9118 }
9119 
9120 /*
9121  * Check whether a rq has a misfit task and if it looks like we can actually
9122  * help that task: we can migrate the task to a CPU of higher capacity, or
9123  * the task's current CPU is heavily pressured.
9124  */
9125 static inline int check_misfit_status(struct rq *rq, struct sched_domain *sd)
9126 {
9127 	return rq->misfit_task_load &&
9128 		(rq->cpu_capacity_orig < rq->rd->max_cpu_capacity ||
9129 		 check_cpu_capacity(rq, sd));
9130 }
9131 
9132 /*
9133  * Group imbalance indicates (and tries to solve) the problem where balancing
9134  * groups is inadequate due to ->cpus_ptr constraints.
9135  *
9136  * Imagine a situation of two groups of 4 CPUs each and 4 tasks each with a
9137  * cpumask covering 1 CPU of the first group and 3 CPUs of the second group.
9138  * Something like:
9139  *
9140  *	{ 0 1 2 3 } { 4 5 6 7 }
9141  *	        *     * * *
9142  *
9143  * If we were to balance group-wise we'd place two tasks in the first group and
9144  * two tasks in the second group. Clearly this is undesired as it will overload
9145  * cpu 3 and leave one of the CPUs in the second group unused.
9146  *
9147  * The current solution to this issue is detecting the skew in the first group
9148  * by noticing the lower domain failed to reach balance and had difficulty
9149  * moving tasks due to affinity constraints.
9150  *
9151  * When this is so detected; this group becomes a candidate for busiest; see
9152  * update_sd_pick_busiest(). And calculate_imbalance() and
9153  * find_busiest_group() avoid some of the usual balance conditions to allow it
9154  * to create an effective group imbalance.
9155  *
9156  * This is a somewhat tricky proposition since the next run might not find the
9157  * group imbalance and decide the groups need to be balanced again. A most
9158  * subtle and fragile situation.
9159  */
9160 
9161 static inline int sg_imbalanced(struct sched_group *group)
9162 {
9163 	return group->sgc->imbalance;
9164 }
9165 
9166 /*
9167  * group_has_capacity returns true if the group has spare capacity that could
9168  * be used by some tasks.
9169  * We consider that a group has spare capacity if the number of task is
9170  * smaller than the number of CPUs or if the utilization is lower than the
9171  * available capacity for CFS tasks.
9172  * For the latter, we use a threshold to stabilize the state, to take into
9173  * account the variance of the tasks' load and to return true if the available
9174  * capacity in meaningful for the load balancer.
9175  * As an example, an available capacity of 1% can appear but it doesn't make
9176  * any benefit for the load balance.
9177  */
9178 static inline bool
9179 group_has_capacity(unsigned int imbalance_pct, struct sg_lb_stats *sgs)
9180 {
9181 	if (sgs->sum_nr_running < sgs->group_weight)
9182 		return true;
9183 
9184 	if ((sgs->group_capacity * imbalance_pct) <
9185 			(sgs->group_runnable * 100))
9186 		return false;
9187 
9188 	if ((sgs->group_capacity * 100) >
9189 			(sgs->group_util * imbalance_pct))
9190 		return true;
9191 
9192 	return false;
9193 }
9194 
9195 /*
9196  *  group_is_overloaded returns true if the group has more tasks than it can
9197  *  handle.
9198  *  group_is_overloaded is not equals to !group_has_capacity because a group
9199  *  with the exact right number of tasks, has no more spare capacity but is not
9200  *  overloaded so both group_has_capacity and group_is_overloaded return
9201  *  false.
9202  */
9203 static inline bool
9204 group_is_overloaded(unsigned int imbalance_pct, struct sg_lb_stats *sgs)
9205 {
9206 	if (sgs->sum_nr_running <= sgs->group_weight)
9207 		return false;
9208 
9209 	if ((sgs->group_capacity * 100) <
9210 			(sgs->group_util * imbalance_pct))
9211 		return true;
9212 
9213 	if ((sgs->group_capacity * imbalance_pct) <
9214 			(sgs->group_runnable * 100))
9215 		return true;
9216 
9217 	return false;
9218 }
9219 
9220 static inline enum
9221 group_type group_classify(unsigned int imbalance_pct,
9222 			  struct sched_group *group,
9223 			  struct sg_lb_stats *sgs)
9224 {
9225 	if (group_is_overloaded(imbalance_pct, sgs))
9226 		return group_overloaded;
9227 
9228 	if (sg_imbalanced(group))
9229 		return group_imbalanced;
9230 
9231 	if (sgs->group_asym_packing)
9232 		return group_asym_packing;
9233 
9234 	if (sgs->group_misfit_task_load)
9235 		return group_misfit_task;
9236 
9237 	if (!group_has_capacity(imbalance_pct, sgs))
9238 		return group_fully_busy;
9239 
9240 	return group_has_spare;
9241 }
9242 
9243 /**
9244  * asym_smt_can_pull_tasks - Check whether the load balancing CPU can pull tasks
9245  * @dst_cpu:	Destination CPU of the load balancing
9246  * @sds:	Load-balancing data with statistics of the local group
9247  * @sgs:	Load-balancing statistics of the candidate busiest group
9248  * @sg:		The candidate busiest group
9249  *
9250  * Check the state of the SMT siblings of both @sds::local and @sg and decide
9251  * if @dst_cpu can pull tasks.
9252  *
9253  * If @dst_cpu does not have SMT siblings, it can pull tasks if two or more of
9254  * the SMT siblings of @sg are busy. If only one CPU in @sg is busy, pull tasks
9255  * only if @dst_cpu has higher priority.
9256  *
9257  * If both @dst_cpu and @sg have SMT siblings, and @sg has exactly one more
9258  * busy CPU than @sds::local, let @dst_cpu pull tasks if it has higher priority.
9259  * Bigger imbalances in the number of busy CPUs will be dealt with in
9260  * update_sd_pick_busiest().
9261  *
9262  * If @sg does not have SMT siblings, only pull tasks if all of the SMT siblings
9263  * of @dst_cpu are idle and @sg has lower priority.
9264  *
9265  * Return: true if @dst_cpu can pull tasks, false otherwise.
9266  */
9267 static bool asym_smt_can_pull_tasks(int dst_cpu, struct sd_lb_stats *sds,
9268 				    struct sg_lb_stats *sgs,
9269 				    struct sched_group *sg)
9270 {
9271 #ifdef CONFIG_SCHED_SMT
9272 	bool local_is_smt, sg_is_smt;
9273 	int sg_busy_cpus;
9274 
9275 	local_is_smt = sds->local->flags & SD_SHARE_CPUCAPACITY;
9276 	sg_is_smt = sg->flags & SD_SHARE_CPUCAPACITY;
9277 
9278 	sg_busy_cpus = sgs->group_weight - sgs->idle_cpus;
9279 
9280 	if (!local_is_smt) {
9281 		/*
9282 		 * If we are here, @dst_cpu is idle and does not have SMT
9283 		 * siblings. Pull tasks if candidate group has two or more
9284 		 * busy CPUs.
9285 		 */
9286 		if (sg_busy_cpus >= 2) /* implies sg_is_smt */
9287 			return true;
9288 
9289 		/*
9290 		 * @dst_cpu does not have SMT siblings. @sg may have SMT
9291 		 * siblings and only one is busy. In such case, @dst_cpu
9292 		 * can help if it has higher priority and is idle (i.e.,
9293 		 * it has no running tasks).
9294 		 */
9295 		return sched_asym_prefer(dst_cpu, sg->asym_prefer_cpu);
9296 	}
9297 
9298 	/* @dst_cpu has SMT siblings. */
9299 
9300 	if (sg_is_smt) {
9301 		int local_busy_cpus = sds->local->group_weight -
9302 				      sds->local_stat.idle_cpus;
9303 		int busy_cpus_delta = sg_busy_cpus - local_busy_cpus;
9304 
9305 		if (busy_cpus_delta == 1)
9306 			return sched_asym_prefer(dst_cpu, sg->asym_prefer_cpu);
9307 
9308 		return false;
9309 	}
9310 
9311 	/*
9312 	 * @sg does not have SMT siblings. Ensure that @sds::local does not end
9313 	 * up with more than one busy SMT sibling and only pull tasks if there
9314 	 * are not busy CPUs (i.e., no CPU has running tasks).
9315 	 */
9316 	if (!sds->local_stat.sum_nr_running)
9317 		return sched_asym_prefer(dst_cpu, sg->asym_prefer_cpu);
9318 
9319 	return false;
9320 #else
9321 	/* Always return false so that callers deal with non-SMT cases. */
9322 	return false;
9323 #endif
9324 }
9325 
9326 static inline bool
9327 sched_asym(struct lb_env *env, struct sd_lb_stats *sds,  struct sg_lb_stats *sgs,
9328 	   struct sched_group *group)
9329 {
9330 	/* Only do SMT checks if either local or candidate have SMT siblings */
9331 	if ((sds->local->flags & SD_SHARE_CPUCAPACITY) ||
9332 	    (group->flags & SD_SHARE_CPUCAPACITY))
9333 		return asym_smt_can_pull_tasks(env->dst_cpu, sds, sgs, group);
9334 
9335 	return sched_asym_prefer(env->dst_cpu, group->asym_prefer_cpu);
9336 }
9337 
9338 static inline bool
9339 sched_reduced_capacity(struct rq *rq, struct sched_domain *sd)
9340 {
9341 	/*
9342 	 * When there is more than 1 task, the group_overloaded case already
9343 	 * takes care of cpu with reduced capacity
9344 	 */
9345 	if (rq->cfs.h_nr_running != 1)
9346 		return false;
9347 
9348 	return check_cpu_capacity(rq, sd);
9349 }
9350 
9351 /**
9352  * update_sg_lb_stats - Update sched_group's statistics for load balancing.
9353  * @env: The load balancing environment.
9354  * @sds: Load-balancing data with statistics of the local group.
9355  * @group: sched_group whose statistics are to be updated.
9356  * @sgs: variable to hold the statistics for this group.
9357  * @sg_status: Holds flag indicating the status of the sched_group
9358  */
9359 static inline void update_sg_lb_stats(struct lb_env *env,
9360 				      struct sd_lb_stats *sds,
9361 				      struct sched_group *group,
9362 				      struct sg_lb_stats *sgs,
9363 				      int *sg_status)
9364 {
9365 	int i, nr_running, local_group;
9366 
9367 	memset(sgs, 0, sizeof(*sgs));
9368 
9369 	local_group = group == sds->local;
9370 
9371 	for_each_cpu_and(i, sched_group_span(group), env->cpus) {
9372 		struct rq *rq = cpu_rq(i);
9373 		unsigned long load = cpu_load(rq);
9374 
9375 		sgs->group_load += load;
9376 		sgs->group_util += cpu_util_cfs(i);
9377 		sgs->group_runnable += cpu_runnable(rq);
9378 		sgs->sum_h_nr_running += rq->cfs.h_nr_running;
9379 
9380 		nr_running = rq->nr_running;
9381 		sgs->sum_nr_running += nr_running;
9382 
9383 		if (nr_running > 1)
9384 			*sg_status |= SG_OVERLOAD;
9385 
9386 		if (cpu_overutilized(i))
9387 			*sg_status |= SG_OVERUTILIZED;
9388 
9389 #ifdef CONFIG_NUMA_BALANCING
9390 		sgs->nr_numa_running += rq->nr_numa_running;
9391 		sgs->nr_preferred_running += rq->nr_preferred_running;
9392 #endif
9393 		/*
9394 		 * No need to call idle_cpu() if nr_running is not 0
9395 		 */
9396 		if (!nr_running && idle_cpu(i)) {
9397 			sgs->idle_cpus++;
9398 			/* Idle cpu can't have misfit task */
9399 			continue;
9400 		}
9401 
9402 		if (local_group)
9403 			continue;
9404 
9405 		if (env->sd->flags & SD_ASYM_CPUCAPACITY) {
9406 			/* Check for a misfit task on the cpu */
9407 			if (sgs->group_misfit_task_load < rq->misfit_task_load) {
9408 				sgs->group_misfit_task_load = rq->misfit_task_load;
9409 				*sg_status |= SG_OVERLOAD;
9410 			}
9411 		} else if ((env->idle != CPU_NOT_IDLE) &&
9412 			   sched_reduced_capacity(rq, env->sd)) {
9413 			/* Check for a task running on a CPU with reduced capacity */
9414 			if (sgs->group_misfit_task_load < load)
9415 				sgs->group_misfit_task_load = load;
9416 		}
9417 	}
9418 
9419 	sgs->group_capacity = group->sgc->capacity;
9420 
9421 	sgs->group_weight = group->group_weight;
9422 
9423 	/* Check if dst CPU is idle and preferred to this group */
9424 	if (!local_group && env->sd->flags & SD_ASYM_PACKING &&
9425 	    env->idle != CPU_NOT_IDLE && sgs->sum_h_nr_running &&
9426 	    sched_asym(env, sds, sgs, group)) {
9427 		sgs->group_asym_packing = 1;
9428 	}
9429 
9430 	sgs->group_type = group_classify(env->sd->imbalance_pct, group, sgs);
9431 
9432 	/* Computing avg_load makes sense only when group is overloaded */
9433 	if (sgs->group_type == group_overloaded)
9434 		sgs->avg_load = (sgs->group_load * SCHED_CAPACITY_SCALE) /
9435 				sgs->group_capacity;
9436 }
9437 
9438 /**
9439  * update_sd_pick_busiest - return 1 on busiest group
9440  * @env: The load balancing environment.
9441  * @sds: sched_domain statistics
9442  * @sg: sched_group candidate to be checked for being the busiest
9443  * @sgs: sched_group statistics
9444  *
9445  * Determine if @sg is a busier group than the previously selected
9446  * busiest group.
9447  *
9448  * Return: %true if @sg is a busier group than the previously selected
9449  * busiest group. %false otherwise.
9450  */
9451 static bool update_sd_pick_busiest(struct lb_env *env,
9452 				   struct sd_lb_stats *sds,
9453 				   struct sched_group *sg,
9454 				   struct sg_lb_stats *sgs)
9455 {
9456 	struct sg_lb_stats *busiest = &sds->busiest_stat;
9457 
9458 	/* Make sure that there is at least one task to pull */
9459 	if (!sgs->sum_h_nr_running)
9460 		return false;
9461 
9462 	/*
9463 	 * Don't try to pull misfit tasks we can't help.
9464 	 * We can use max_capacity here as reduction in capacity on some
9465 	 * CPUs in the group should either be possible to resolve
9466 	 * internally or be covered by avg_load imbalance (eventually).
9467 	 */
9468 	if ((env->sd->flags & SD_ASYM_CPUCAPACITY) &&
9469 	    (sgs->group_type == group_misfit_task) &&
9470 	    (!capacity_greater(capacity_of(env->dst_cpu), sg->sgc->max_capacity) ||
9471 	     sds->local_stat.group_type != group_has_spare))
9472 		return false;
9473 
9474 	if (sgs->group_type > busiest->group_type)
9475 		return true;
9476 
9477 	if (sgs->group_type < busiest->group_type)
9478 		return false;
9479 
9480 	/*
9481 	 * The candidate and the current busiest group are the same type of
9482 	 * group. Let check which one is the busiest according to the type.
9483 	 */
9484 
9485 	switch (sgs->group_type) {
9486 	case group_overloaded:
9487 		/* Select the overloaded group with highest avg_load. */
9488 		if (sgs->avg_load <= busiest->avg_load)
9489 			return false;
9490 		break;
9491 
9492 	case group_imbalanced:
9493 		/*
9494 		 * Select the 1st imbalanced group as we don't have any way to
9495 		 * choose one more than another.
9496 		 */
9497 		return false;
9498 
9499 	case group_asym_packing:
9500 		/* Prefer to move from lowest priority CPU's work */
9501 		if (sched_asym_prefer(sg->asym_prefer_cpu, sds->busiest->asym_prefer_cpu))
9502 			return false;
9503 		break;
9504 
9505 	case group_misfit_task:
9506 		/*
9507 		 * If we have more than one misfit sg go with the biggest
9508 		 * misfit.
9509 		 */
9510 		if (sgs->group_misfit_task_load < busiest->group_misfit_task_load)
9511 			return false;
9512 		break;
9513 
9514 	case group_fully_busy:
9515 		/*
9516 		 * Select the fully busy group with highest avg_load. In
9517 		 * theory, there is no need to pull task from such kind of
9518 		 * group because tasks have all compute capacity that they need
9519 		 * but we can still improve the overall throughput by reducing
9520 		 * contention when accessing shared HW resources.
9521 		 *
9522 		 * XXX for now avg_load is not computed and always 0 so we
9523 		 * select the 1st one.
9524 		 */
9525 		if (sgs->avg_load <= busiest->avg_load)
9526 			return false;
9527 		break;
9528 
9529 	case group_has_spare:
9530 		/*
9531 		 * Select not overloaded group with lowest number of idle cpus
9532 		 * and highest number of running tasks. We could also compare
9533 		 * the spare capacity which is more stable but it can end up
9534 		 * that the group has less spare capacity but finally more idle
9535 		 * CPUs which means less opportunity to pull tasks.
9536 		 */
9537 		if (sgs->idle_cpus > busiest->idle_cpus)
9538 			return false;
9539 		else if ((sgs->idle_cpus == busiest->idle_cpus) &&
9540 			 (sgs->sum_nr_running <= busiest->sum_nr_running))
9541 			return false;
9542 
9543 		break;
9544 	}
9545 
9546 	/*
9547 	 * Candidate sg has no more than one task per CPU and has higher
9548 	 * per-CPU capacity. Migrating tasks to less capable CPUs may harm
9549 	 * throughput. Maximize throughput, power/energy consequences are not
9550 	 * considered.
9551 	 */
9552 	if ((env->sd->flags & SD_ASYM_CPUCAPACITY) &&
9553 	    (sgs->group_type <= group_fully_busy) &&
9554 	    (capacity_greater(sg->sgc->min_capacity, capacity_of(env->dst_cpu))))
9555 		return false;
9556 
9557 	return true;
9558 }
9559 
9560 #ifdef CONFIG_NUMA_BALANCING
9561 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
9562 {
9563 	if (sgs->sum_h_nr_running > sgs->nr_numa_running)
9564 		return regular;
9565 	if (sgs->sum_h_nr_running > sgs->nr_preferred_running)
9566 		return remote;
9567 	return all;
9568 }
9569 
9570 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
9571 {
9572 	if (rq->nr_running > rq->nr_numa_running)
9573 		return regular;
9574 	if (rq->nr_running > rq->nr_preferred_running)
9575 		return remote;
9576 	return all;
9577 }
9578 #else
9579 static inline enum fbq_type fbq_classify_group(struct sg_lb_stats *sgs)
9580 {
9581 	return all;
9582 }
9583 
9584 static inline enum fbq_type fbq_classify_rq(struct rq *rq)
9585 {
9586 	return regular;
9587 }
9588 #endif /* CONFIG_NUMA_BALANCING */
9589 
9590 
9591 struct sg_lb_stats;
9592 
9593 /*
9594  * task_running_on_cpu - return 1 if @p is running on @cpu.
9595  */
9596 
9597 static unsigned int task_running_on_cpu(int cpu, struct task_struct *p)
9598 {
9599 	/* Task has no contribution or is new */
9600 	if (cpu != task_cpu(p) || !READ_ONCE(p->se.avg.last_update_time))
9601 		return 0;
9602 
9603 	if (task_on_rq_queued(p))
9604 		return 1;
9605 
9606 	return 0;
9607 }
9608 
9609 /**
9610  * idle_cpu_without - would a given CPU be idle without p ?
9611  * @cpu: the processor on which idleness is tested.
9612  * @p: task which should be ignored.
9613  *
9614  * Return: 1 if the CPU would be idle. 0 otherwise.
9615  */
9616 static int idle_cpu_without(int cpu, struct task_struct *p)
9617 {
9618 	struct rq *rq = cpu_rq(cpu);
9619 
9620 	if (rq->curr != rq->idle && rq->curr != p)
9621 		return 0;
9622 
9623 	/*
9624 	 * rq->nr_running can't be used but an updated version without the
9625 	 * impact of p on cpu must be used instead. The updated nr_running
9626 	 * be computed and tested before calling idle_cpu_without().
9627 	 */
9628 
9629 #ifdef CONFIG_SMP
9630 	if (rq->ttwu_pending)
9631 		return 0;
9632 #endif
9633 
9634 	return 1;
9635 }
9636 
9637 /*
9638  * update_sg_wakeup_stats - Update sched_group's statistics for wakeup.
9639  * @sd: The sched_domain level to look for idlest group.
9640  * @group: sched_group whose statistics are to be updated.
9641  * @sgs: variable to hold the statistics for this group.
9642  * @p: The task for which we look for the idlest group/CPU.
9643  */
9644 static inline void update_sg_wakeup_stats(struct sched_domain *sd,
9645 					  struct sched_group *group,
9646 					  struct sg_lb_stats *sgs,
9647 					  struct task_struct *p)
9648 {
9649 	int i, nr_running;
9650 
9651 	memset(sgs, 0, sizeof(*sgs));
9652 
9653 	/* Assume that task can't fit any CPU of the group */
9654 	if (sd->flags & SD_ASYM_CPUCAPACITY)
9655 		sgs->group_misfit_task_load = 1;
9656 
9657 	for_each_cpu(i, sched_group_span(group)) {
9658 		struct rq *rq = cpu_rq(i);
9659 		unsigned int local;
9660 
9661 		sgs->group_load += cpu_load_without(rq, p);
9662 		sgs->group_util += cpu_util_without(i, p);
9663 		sgs->group_runnable += cpu_runnable_without(rq, p);
9664 		local = task_running_on_cpu(i, p);
9665 		sgs->sum_h_nr_running += rq->cfs.h_nr_running - local;
9666 
9667 		nr_running = rq->nr_running - local;
9668 		sgs->sum_nr_running += nr_running;
9669 
9670 		/*
9671 		 * No need to call idle_cpu_without() if nr_running is not 0
9672 		 */
9673 		if (!nr_running && idle_cpu_without(i, p))
9674 			sgs->idle_cpus++;
9675 
9676 		/* Check if task fits in the CPU */
9677 		if (sd->flags & SD_ASYM_CPUCAPACITY &&
9678 		    sgs->group_misfit_task_load &&
9679 		    task_fits_cpu(p, i))
9680 			sgs->group_misfit_task_load = 0;
9681 
9682 	}
9683 
9684 	sgs->group_capacity = group->sgc->capacity;
9685 
9686 	sgs->group_weight = group->group_weight;
9687 
9688 	sgs->group_type = group_classify(sd->imbalance_pct, group, sgs);
9689 
9690 	/*
9691 	 * Computing avg_load makes sense only when group is fully busy or
9692 	 * overloaded
9693 	 */
9694 	if (sgs->group_type == group_fully_busy ||
9695 		sgs->group_type == group_overloaded)
9696 		sgs->avg_load = (sgs->group_load * SCHED_CAPACITY_SCALE) /
9697 				sgs->group_capacity;
9698 }
9699 
9700 static bool update_pick_idlest(struct sched_group *idlest,
9701 			       struct sg_lb_stats *idlest_sgs,
9702 			       struct sched_group *group,
9703 			       struct sg_lb_stats *sgs)
9704 {
9705 	if (sgs->group_type < idlest_sgs->group_type)
9706 		return true;
9707 
9708 	if (sgs->group_type > idlest_sgs->group_type)
9709 		return false;
9710 
9711 	/*
9712 	 * The candidate and the current idlest group are the same type of
9713 	 * group. Let check which one is the idlest according to the type.
9714 	 */
9715 
9716 	switch (sgs->group_type) {
9717 	case group_overloaded:
9718 	case group_fully_busy:
9719 		/* Select the group with lowest avg_load. */
9720 		if (idlest_sgs->avg_load <= sgs->avg_load)
9721 			return false;
9722 		break;
9723 
9724 	case group_imbalanced:
9725 	case group_asym_packing:
9726 		/* Those types are not used in the slow wakeup path */
9727 		return false;
9728 
9729 	case group_misfit_task:
9730 		/* Select group with the highest max capacity */
9731 		if (idlest->sgc->max_capacity >= group->sgc->max_capacity)
9732 			return false;
9733 		break;
9734 
9735 	case group_has_spare:
9736 		/* Select group with most idle CPUs */
9737 		if (idlest_sgs->idle_cpus > sgs->idle_cpus)
9738 			return false;
9739 
9740 		/* Select group with lowest group_util */
9741 		if (idlest_sgs->idle_cpus == sgs->idle_cpus &&
9742 			idlest_sgs->group_util <= sgs->group_util)
9743 			return false;
9744 
9745 		break;
9746 	}
9747 
9748 	return true;
9749 }
9750 
9751 /*
9752  * find_idlest_group() finds and returns the least busy CPU group within the
9753  * domain.
9754  *
9755  * Assumes p is allowed on at least one CPU in sd.
9756  */
9757 static struct sched_group *
9758 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
9759 {
9760 	struct sched_group *idlest = NULL, *local = NULL, *group = sd->groups;
9761 	struct sg_lb_stats local_sgs, tmp_sgs;
9762 	struct sg_lb_stats *sgs;
9763 	unsigned long imbalance;
9764 	struct sg_lb_stats idlest_sgs = {
9765 			.avg_load = UINT_MAX,
9766 			.group_type = group_overloaded,
9767 	};
9768 
9769 	do {
9770 		int local_group;
9771 
9772 		/* Skip over this group if it has no CPUs allowed */
9773 		if (!cpumask_intersects(sched_group_span(group),
9774 					p->cpus_ptr))
9775 			continue;
9776 
9777 		/* Skip over this group if no cookie matched */
9778 		if (!sched_group_cookie_match(cpu_rq(this_cpu), p, group))
9779 			continue;
9780 
9781 		local_group = cpumask_test_cpu(this_cpu,
9782 					       sched_group_span(group));
9783 
9784 		if (local_group) {
9785 			sgs = &local_sgs;
9786 			local = group;
9787 		} else {
9788 			sgs = &tmp_sgs;
9789 		}
9790 
9791 		update_sg_wakeup_stats(sd, group, sgs, p);
9792 
9793 		if (!local_group && update_pick_idlest(idlest, &idlest_sgs, group, sgs)) {
9794 			idlest = group;
9795 			idlest_sgs = *sgs;
9796 		}
9797 
9798 	} while (group = group->next, group != sd->groups);
9799 
9800 
9801 	/* There is no idlest group to push tasks to */
9802 	if (!idlest)
9803 		return NULL;
9804 
9805 	/* The local group has been skipped because of CPU affinity */
9806 	if (!local)
9807 		return idlest;
9808 
9809 	/*
9810 	 * If the local group is idler than the selected idlest group
9811 	 * don't try and push the task.
9812 	 */
9813 	if (local_sgs.group_type < idlest_sgs.group_type)
9814 		return NULL;
9815 
9816 	/*
9817 	 * If the local group is busier than the selected idlest group
9818 	 * try and push the task.
9819 	 */
9820 	if (local_sgs.group_type > idlest_sgs.group_type)
9821 		return idlest;
9822 
9823 	switch (local_sgs.group_type) {
9824 	case group_overloaded:
9825 	case group_fully_busy:
9826 
9827 		/* Calculate allowed imbalance based on load */
9828 		imbalance = scale_load_down(NICE_0_LOAD) *
9829 				(sd->imbalance_pct-100) / 100;
9830 
9831 		/*
9832 		 * When comparing groups across NUMA domains, it's possible for
9833 		 * the local domain to be very lightly loaded relative to the
9834 		 * remote domains but "imbalance" skews the comparison making
9835 		 * remote CPUs look much more favourable. When considering
9836 		 * cross-domain, add imbalance to the load on the remote node
9837 		 * and consider staying local.
9838 		 */
9839 
9840 		if ((sd->flags & SD_NUMA) &&
9841 		    ((idlest_sgs.avg_load + imbalance) >= local_sgs.avg_load))
9842 			return NULL;
9843 
9844 		/*
9845 		 * If the local group is less loaded than the selected
9846 		 * idlest group don't try and push any tasks.
9847 		 */
9848 		if (idlest_sgs.avg_load >= (local_sgs.avg_load + imbalance))
9849 			return NULL;
9850 
9851 		if (100 * local_sgs.avg_load <= sd->imbalance_pct * idlest_sgs.avg_load)
9852 			return NULL;
9853 		break;
9854 
9855 	case group_imbalanced:
9856 	case group_asym_packing:
9857 		/* Those type are not used in the slow wakeup path */
9858 		return NULL;
9859 
9860 	case group_misfit_task:
9861 		/* Select group with the highest max capacity */
9862 		if (local->sgc->max_capacity >= idlest->sgc->max_capacity)
9863 			return NULL;
9864 		break;
9865 
9866 	case group_has_spare:
9867 #ifdef CONFIG_NUMA
9868 		if (sd->flags & SD_NUMA) {
9869 			int imb_numa_nr = sd->imb_numa_nr;
9870 #ifdef CONFIG_NUMA_BALANCING
9871 			int idlest_cpu;
9872 			/*
9873 			 * If there is spare capacity at NUMA, try to select
9874 			 * the preferred node
9875 			 */
9876 			if (cpu_to_node(this_cpu) == p->numa_preferred_nid)
9877 				return NULL;
9878 
9879 			idlest_cpu = cpumask_first(sched_group_span(idlest));
9880 			if (cpu_to_node(idlest_cpu) == p->numa_preferred_nid)
9881 				return idlest;
9882 #endif /* CONFIG_NUMA_BALANCING */
9883 			/*
9884 			 * Otherwise, keep the task close to the wakeup source
9885 			 * and improve locality if the number of running tasks
9886 			 * would remain below threshold where an imbalance is
9887 			 * allowed while accounting for the possibility the
9888 			 * task is pinned to a subset of CPUs. If there is a
9889 			 * real need of migration, periodic load balance will
9890 			 * take care of it.
9891 			 */
9892 			if (p->nr_cpus_allowed != NR_CPUS) {
9893 				struct cpumask *cpus = this_cpu_cpumask_var_ptr(select_rq_mask);
9894 
9895 				cpumask_and(cpus, sched_group_span(local), p->cpus_ptr);
9896 				imb_numa_nr = min(cpumask_weight(cpus), sd->imb_numa_nr);
9897 			}
9898 
9899 			imbalance = abs(local_sgs.idle_cpus - idlest_sgs.idle_cpus);
9900 			if (!adjust_numa_imbalance(imbalance,
9901 						   local_sgs.sum_nr_running + 1,
9902 						   imb_numa_nr)) {
9903 				return NULL;
9904 			}
9905 		}
9906 #endif /* CONFIG_NUMA */
9907 
9908 		/*
9909 		 * Select group with highest number of idle CPUs. We could also
9910 		 * compare the utilization which is more stable but it can end
9911 		 * up that the group has less spare capacity but finally more
9912 		 * idle CPUs which means more opportunity to run task.
9913 		 */
9914 		if (local_sgs.idle_cpus >= idlest_sgs.idle_cpus)
9915 			return NULL;
9916 		break;
9917 	}
9918 
9919 	return idlest;
9920 }
9921 
9922 static void update_idle_cpu_scan(struct lb_env *env,
9923 				 unsigned long sum_util)
9924 {
9925 	struct sched_domain_shared *sd_share;
9926 	int llc_weight, pct;
9927 	u64 x, y, tmp;
9928 	/*
9929 	 * Update the number of CPUs to scan in LLC domain, which could
9930 	 * be used as a hint in select_idle_cpu(). The update of sd_share
9931 	 * could be expensive because it is within a shared cache line.
9932 	 * So the write of this hint only occurs during periodic load
9933 	 * balancing, rather than CPU_NEWLY_IDLE, because the latter
9934 	 * can fire way more frequently than the former.
9935 	 */
9936 	if (!sched_feat(SIS_UTIL) || env->idle == CPU_NEWLY_IDLE)
9937 		return;
9938 
9939 	llc_weight = per_cpu(sd_llc_size, env->dst_cpu);
9940 	if (env->sd->span_weight != llc_weight)
9941 		return;
9942 
9943 	sd_share = rcu_dereference(per_cpu(sd_llc_shared, env->dst_cpu));
9944 	if (!sd_share)
9945 		return;
9946 
9947 	/*
9948 	 * The number of CPUs to search drops as sum_util increases, when
9949 	 * sum_util hits 85% or above, the scan stops.
9950 	 * The reason to choose 85% as the threshold is because this is the
9951 	 * imbalance_pct(117) when a LLC sched group is overloaded.
9952 	 *
9953 	 * let y = SCHED_CAPACITY_SCALE - p * x^2                       [1]
9954 	 * and y'= y / SCHED_CAPACITY_SCALE
9955 	 *
9956 	 * x is the ratio of sum_util compared to the CPU capacity:
9957 	 * x = sum_util / (llc_weight * SCHED_CAPACITY_SCALE)
9958 	 * y' is the ratio of CPUs to be scanned in the LLC domain,
9959 	 * and the number of CPUs to scan is calculated by:
9960 	 *
9961 	 * nr_scan = llc_weight * y'                                    [2]
9962 	 *
9963 	 * When x hits the threshold of overloaded, AKA, when
9964 	 * x = 100 / pct, y drops to 0. According to [1],
9965 	 * p should be SCHED_CAPACITY_SCALE * pct^2 / 10000
9966 	 *
9967 	 * Scale x by SCHED_CAPACITY_SCALE:
9968 	 * x' = sum_util / llc_weight;                                  [3]
9969 	 *
9970 	 * and finally [1] becomes:
9971 	 * y = SCHED_CAPACITY_SCALE -
9972 	 *     x'^2 * pct^2 / (10000 * SCHED_CAPACITY_SCALE)            [4]
9973 	 *
9974 	 */
9975 	/* equation [3] */
9976 	x = sum_util;
9977 	do_div(x, llc_weight);
9978 
9979 	/* equation [4] */
9980 	pct = env->sd->imbalance_pct;
9981 	tmp = x * x * pct * pct;
9982 	do_div(tmp, 10000 * SCHED_CAPACITY_SCALE);
9983 	tmp = min_t(long, tmp, SCHED_CAPACITY_SCALE);
9984 	y = SCHED_CAPACITY_SCALE - tmp;
9985 
9986 	/* equation [2] */
9987 	y *= llc_weight;
9988 	do_div(y, SCHED_CAPACITY_SCALE);
9989 	if ((int)y != sd_share->nr_idle_scan)
9990 		WRITE_ONCE(sd_share->nr_idle_scan, (int)y);
9991 }
9992 
9993 /**
9994  * update_sd_lb_stats - Update sched_domain's statistics for load balancing.
9995  * @env: The load balancing environment.
9996  * @sds: variable to hold the statistics for this sched_domain.
9997  */
9998 
9999 static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sds)
10000 {
10001 	struct sched_domain *child = env->sd->child;
10002 	struct sched_group *sg = env->sd->groups;
10003 	struct sg_lb_stats *local = &sds->local_stat;
10004 	struct sg_lb_stats tmp_sgs;
10005 	unsigned long sum_util = 0;
10006 	int sg_status = 0;
10007 
10008 	do {
10009 		struct sg_lb_stats *sgs = &tmp_sgs;
10010 		int local_group;
10011 
10012 		local_group = cpumask_test_cpu(env->dst_cpu, sched_group_span(sg));
10013 		if (local_group) {
10014 			sds->local = sg;
10015 			sgs = local;
10016 
10017 			if (env->idle != CPU_NEWLY_IDLE ||
10018 			    time_after_eq(jiffies, sg->sgc->next_update))
10019 				update_group_capacity(env->sd, env->dst_cpu);
10020 		}
10021 
10022 		update_sg_lb_stats(env, sds, sg, sgs, &sg_status);
10023 
10024 		if (local_group)
10025 			goto next_group;
10026 
10027 
10028 		if (update_sd_pick_busiest(env, sds, sg, sgs)) {
10029 			sds->busiest = sg;
10030 			sds->busiest_stat = *sgs;
10031 		}
10032 
10033 next_group:
10034 		/* Now, start updating sd_lb_stats */
10035 		sds->total_load += sgs->group_load;
10036 		sds->total_capacity += sgs->group_capacity;
10037 
10038 		sum_util += sgs->group_util;
10039 		sg = sg->next;
10040 	} while (sg != env->sd->groups);
10041 
10042 	/* Tag domain that child domain prefers tasks go to siblings first */
10043 	sds->prefer_sibling = child && child->flags & SD_PREFER_SIBLING;
10044 
10045 
10046 	if (env->sd->flags & SD_NUMA)
10047 		env->fbq_type = fbq_classify_group(&sds->busiest_stat);
10048 
10049 	if (!env->sd->parent) {
10050 		struct root_domain *rd = env->dst_rq->rd;
10051 
10052 		/* update overload indicator if we are at root domain */
10053 		WRITE_ONCE(rd->overload, sg_status & SG_OVERLOAD);
10054 
10055 		/* Update over-utilization (tipping point, U >= 0) indicator */
10056 		WRITE_ONCE(rd->overutilized, sg_status & SG_OVERUTILIZED);
10057 		trace_sched_overutilized_tp(rd, sg_status & SG_OVERUTILIZED);
10058 	} else if (sg_status & SG_OVERUTILIZED) {
10059 		struct root_domain *rd = env->dst_rq->rd;
10060 
10061 		WRITE_ONCE(rd->overutilized, SG_OVERUTILIZED);
10062 		trace_sched_overutilized_tp(rd, SG_OVERUTILIZED);
10063 	}
10064 
10065 	update_idle_cpu_scan(env, sum_util);
10066 }
10067 
10068 /**
10069  * calculate_imbalance - Calculate the amount of imbalance present within the
10070  *			 groups of a given sched_domain during load balance.
10071  * @env: load balance environment
10072  * @sds: statistics of the sched_domain whose imbalance is to be calculated.
10073  */
10074 static inline void calculate_imbalance(struct lb_env *env, struct sd_lb_stats *sds)
10075 {
10076 	struct sg_lb_stats *local, *busiest;
10077 
10078 	local = &sds->local_stat;
10079 	busiest = &sds->busiest_stat;
10080 
10081 	if (busiest->group_type == group_misfit_task) {
10082 		if (env->sd->flags & SD_ASYM_CPUCAPACITY) {
10083 			/* Set imbalance to allow misfit tasks to be balanced. */
10084 			env->migration_type = migrate_misfit;
10085 			env->imbalance = 1;
10086 		} else {
10087 			/*
10088 			 * Set load imbalance to allow moving task from cpu
10089 			 * with reduced capacity.
10090 			 */
10091 			env->migration_type = migrate_load;
10092 			env->imbalance = busiest->group_misfit_task_load;
10093 		}
10094 		return;
10095 	}
10096 
10097 	if (busiest->group_type == group_asym_packing) {
10098 		/*
10099 		 * In case of asym capacity, we will try to migrate all load to
10100 		 * the preferred CPU.
10101 		 */
10102 		env->migration_type = migrate_task;
10103 		env->imbalance = busiest->sum_h_nr_running;
10104 		return;
10105 	}
10106 
10107 	if (busiest->group_type == group_imbalanced) {
10108 		/*
10109 		 * In the group_imb case we cannot rely on group-wide averages
10110 		 * to ensure CPU-load equilibrium, try to move any task to fix
10111 		 * the imbalance. The next load balance will take care of
10112 		 * balancing back the system.
10113 		 */
10114 		env->migration_type = migrate_task;
10115 		env->imbalance = 1;
10116 		return;
10117 	}
10118 
10119 	/*
10120 	 * Try to use spare capacity of local group without overloading it or
10121 	 * emptying busiest.
10122 	 */
10123 	if (local->group_type == group_has_spare) {
10124 		if ((busiest->group_type > group_fully_busy) &&
10125 		    !(env->sd->flags & SD_SHARE_PKG_RESOURCES)) {
10126 			/*
10127 			 * If busiest is overloaded, try to fill spare
10128 			 * capacity. This might end up creating spare capacity
10129 			 * in busiest or busiest still being overloaded but
10130 			 * there is no simple way to directly compute the
10131 			 * amount of load to migrate in order to balance the
10132 			 * system.
10133 			 */
10134 			env->migration_type = migrate_util;
10135 			env->imbalance = max(local->group_capacity, local->group_util) -
10136 					 local->group_util;
10137 
10138 			/*
10139 			 * In some cases, the group's utilization is max or even
10140 			 * higher than capacity because of migrations but the
10141 			 * local CPU is (newly) idle. There is at least one
10142 			 * waiting task in this overloaded busiest group. Let's
10143 			 * try to pull it.
10144 			 */
10145 			if (env->idle != CPU_NOT_IDLE && env->imbalance == 0) {
10146 				env->migration_type = migrate_task;
10147 				env->imbalance = 1;
10148 			}
10149 
10150 			return;
10151 		}
10152 
10153 		if (busiest->group_weight == 1 || sds->prefer_sibling) {
10154 			unsigned int nr_diff = busiest->sum_nr_running;
10155 			/*
10156 			 * When prefer sibling, evenly spread running tasks on
10157 			 * groups.
10158 			 */
10159 			env->migration_type = migrate_task;
10160 			lsub_positive(&nr_diff, local->sum_nr_running);
10161 			env->imbalance = nr_diff;
10162 		} else {
10163 
10164 			/*
10165 			 * If there is no overload, we just want to even the number of
10166 			 * idle cpus.
10167 			 */
10168 			env->migration_type = migrate_task;
10169 			env->imbalance = max_t(long, 0,
10170 					       (local->idle_cpus - busiest->idle_cpus));
10171 		}
10172 
10173 #ifdef CONFIG_NUMA
10174 		/* Consider allowing a small imbalance between NUMA groups */
10175 		if (env->sd->flags & SD_NUMA) {
10176 			env->imbalance = adjust_numa_imbalance(env->imbalance,
10177 							       local->sum_nr_running + 1,
10178 							       env->sd->imb_numa_nr);
10179 		}
10180 #endif
10181 
10182 		/* Number of tasks to move to restore balance */
10183 		env->imbalance >>= 1;
10184 
10185 		return;
10186 	}
10187 
10188 	/*
10189 	 * Local is fully busy but has to take more load to relieve the
10190 	 * busiest group
10191 	 */
10192 	if (local->group_type < group_overloaded) {
10193 		/*
10194 		 * Local will become overloaded so the avg_load metrics are
10195 		 * finally needed.
10196 		 */
10197 
10198 		local->avg_load = (local->group_load * SCHED_CAPACITY_SCALE) /
10199 				  local->group_capacity;
10200 
10201 		/*
10202 		 * If the local group is more loaded than the selected
10203 		 * busiest group don't try to pull any tasks.
10204 		 */
10205 		if (local->avg_load >= busiest->avg_load) {
10206 			env->imbalance = 0;
10207 			return;
10208 		}
10209 
10210 		sds->avg_load = (sds->total_load * SCHED_CAPACITY_SCALE) /
10211 				sds->total_capacity;
10212 	}
10213 
10214 	/*
10215 	 * Both group are or will become overloaded and we're trying to get all
10216 	 * the CPUs to the average_load, so we don't want to push ourselves
10217 	 * above the average load, nor do we wish to reduce the max loaded CPU
10218 	 * below the average load. At the same time, we also don't want to
10219 	 * reduce the group load below the group capacity. Thus we look for
10220 	 * the minimum possible imbalance.
10221 	 */
10222 	env->migration_type = migrate_load;
10223 	env->imbalance = min(
10224 		(busiest->avg_load - sds->avg_load) * busiest->group_capacity,
10225 		(sds->avg_load - local->avg_load) * local->group_capacity
10226 	) / SCHED_CAPACITY_SCALE;
10227 }
10228 
10229 /******* find_busiest_group() helpers end here *********************/
10230 
10231 /*
10232  * Decision matrix according to the local and busiest group type:
10233  *
10234  * busiest \ local has_spare fully_busy misfit asym imbalanced overloaded
10235  * has_spare        nr_idle   balanced   N/A    N/A  balanced   balanced
10236  * fully_busy       nr_idle   nr_idle    N/A    N/A  balanced   balanced
10237  * misfit_task      force     N/A        N/A    N/A  N/A        N/A
10238  * asym_packing     force     force      N/A    N/A  force      force
10239  * imbalanced       force     force      N/A    N/A  force      force
10240  * overloaded       force     force      N/A    N/A  force      avg_load
10241  *
10242  * N/A :      Not Applicable because already filtered while updating
10243  *            statistics.
10244  * balanced : The system is balanced for these 2 groups.
10245  * force :    Calculate the imbalance as load migration is probably needed.
10246  * avg_load : Only if imbalance is significant enough.
10247  * nr_idle :  dst_cpu is not busy and the number of idle CPUs is quite
10248  *            different in groups.
10249  */
10250 
10251 /**
10252  * find_busiest_group - Returns the busiest group within the sched_domain
10253  * if there is an imbalance.
10254  * @env: The load balancing environment.
10255  *
10256  * Also calculates the amount of runnable load which should be moved
10257  * to restore balance.
10258  *
10259  * Return:	- The busiest group if imbalance exists.
10260  */
10261 static struct sched_group *find_busiest_group(struct lb_env *env)
10262 {
10263 	struct sg_lb_stats *local, *busiest;
10264 	struct sd_lb_stats sds;
10265 
10266 	init_sd_lb_stats(&sds);
10267 
10268 	/*
10269 	 * Compute the various statistics relevant for load balancing at
10270 	 * this level.
10271 	 */
10272 	update_sd_lb_stats(env, &sds);
10273 
10274 	/* There is no busy sibling group to pull tasks from */
10275 	if (!sds.busiest)
10276 		goto out_balanced;
10277 
10278 	busiest = &sds.busiest_stat;
10279 
10280 	/* Misfit tasks should be dealt with regardless of the avg load */
10281 	if (busiest->group_type == group_misfit_task)
10282 		goto force_balance;
10283 
10284 	if (sched_energy_enabled()) {
10285 		struct root_domain *rd = env->dst_rq->rd;
10286 
10287 		if (rcu_dereference(rd->pd) && !READ_ONCE(rd->overutilized))
10288 			goto out_balanced;
10289 	}
10290 
10291 	/* ASYM feature bypasses nice load balance check */
10292 	if (busiest->group_type == group_asym_packing)
10293 		goto force_balance;
10294 
10295 	/*
10296 	 * If the busiest group is imbalanced the below checks don't
10297 	 * work because they assume all things are equal, which typically
10298 	 * isn't true due to cpus_ptr constraints and the like.
10299 	 */
10300 	if (busiest->group_type == group_imbalanced)
10301 		goto force_balance;
10302 
10303 	local = &sds.local_stat;
10304 	/*
10305 	 * If the local group is busier than the selected busiest group
10306 	 * don't try and pull any tasks.
10307 	 */
10308 	if (local->group_type > busiest->group_type)
10309 		goto out_balanced;
10310 
10311 	/*
10312 	 * When groups are overloaded, use the avg_load to ensure fairness
10313 	 * between tasks.
10314 	 */
10315 	if (local->group_type == group_overloaded) {
10316 		/*
10317 		 * If the local group is more loaded than the selected
10318 		 * busiest group don't try to pull any tasks.
10319 		 */
10320 		if (local->avg_load >= busiest->avg_load)
10321 			goto out_balanced;
10322 
10323 		/* XXX broken for overlapping NUMA groups */
10324 		sds.avg_load = (sds.total_load * SCHED_CAPACITY_SCALE) /
10325 				sds.total_capacity;
10326 
10327 		/*
10328 		 * Don't pull any tasks if this group is already above the
10329 		 * domain average load.
10330 		 */
10331 		if (local->avg_load >= sds.avg_load)
10332 			goto out_balanced;
10333 
10334 		/*
10335 		 * If the busiest group is more loaded, use imbalance_pct to be
10336 		 * conservative.
10337 		 */
10338 		if (100 * busiest->avg_load <=
10339 				env->sd->imbalance_pct * local->avg_load)
10340 			goto out_balanced;
10341 	}
10342 
10343 	/* Try to move all excess tasks to child's sibling domain */
10344 	if (sds.prefer_sibling && local->group_type == group_has_spare &&
10345 	    busiest->sum_nr_running > local->sum_nr_running + 1)
10346 		goto force_balance;
10347 
10348 	if (busiest->group_type != group_overloaded) {
10349 		if (env->idle == CPU_NOT_IDLE)
10350 			/*
10351 			 * If the busiest group is not overloaded (and as a
10352 			 * result the local one too) but this CPU is already
10353 			 * busy, let another idle CPU try to pull task.
10354 			 */
10355 			goto out_balanced;
10356 
10357 		if (busiest->group_weight > 1 &&
10358 		    local->idle_cpus <= (busiest->idle_cpus + 1))
10359 			/*
10360 			 * If the busiest group is not overloaded
10361 			 * and there is no imbalance between this and busiest
10362 			 * group wrt idle CPUs, it is balanced. The imbalance
10363 			 * becomes significant if the diff is greater than 1
10364 			 * otherwise we might end up to just move the imbalance
10365 			 * on another group. Of course this applies only if
10366 			 * there is more than 1 CPU per group.
10367 			 */
10368 			goto out_balanced;
10369 
10370 		if (busiest->sum_h_nr_running == 1)
10371 			/*
10372 			 * busiest doesn't have any tasks waiting to run
10373 			 */
10374 			goto out_balanced;
10375 	}
10376 
10377 force_balance:
10378 	/* Looks like there is an imbalance. Compute it */
10379 	calculate_imbalance(env, &sds);
10380 	return env->imbalance ? sds.busiest : NULL;
10381 
10382 out_balanced:
10383 	env->imbalance = 0;
10384 	return NULL;
10385 }
10386 
10387 /*
10388  * find_busiest_queue - find the busiest runqueue among the CPUs in the group.
10389  */
10390 static struct rq *find_busiest_queue(struct lb_env *env,
10391 				     struct sched_group *group)
10392 {
10393 	struct rq *busiest = NULL, *rq;
10394 	unsigned long busiest_util = 0, busiest_load = 0, busiest_capacity = 1;
10395 	unsigned int busiest_nr = 0;
10396 	int i;
10397 
10398 	for_each_cpu_and(i, sched_group_span(group), env->cpus) {
10399 		unsigned long capacity, load, util;
10400 		unsigned int nr_running;
10401 		enum fbq_type rt;
10402 
10403 		rq = cpu_rq(i);
10404 		rt = fbq_classify_rq(rq);
10405 
10406 		/*
10407 		 * We classify groups/runqueues into three groups:
10408 		 *  - regular: there are !numa tasks
10409 		 *  - remote:  there are numa tasks that run on the 'wrong' node
10410 		 *  - all:     there is no distinction
10411 		 *
10412 		 * In order to avoid migrating ideally placed numa tasks,
10413 		 * ignore those when there's better options.
10414 		 *
10415 		 * If we ignore the actual busiest queue to migrate another
10416 		 * task, the next balance pass can still reduce the busiest
10417 		 * queue by moving tasks around inside the node.
10418 		 *
10419 		 * If we cannot move enough load due to this classification
10420 		 * the next pass will adjust the group classification and
10421 		 * allow migration of more tasks.
10422 		 *
10423 		 * Both cases only affect the total convergence complexity.
10424 		 */
10425 		if (rt > env->fbq_type)
10426 			continue;
10427 
10428 		nr_running = rq->cfs.h_nr_running;
10429 		if (!nr_running)
10430 			continue;
10431 
10432 		capacity = capacity_of(i);
10433 
10434 		/*
10435 		 * For ASYM_CPUCAPACITY domains, don't pick a CPU that could
10436 		 * eventually lead to active_balancing high->low capacity.
10437 		 * Higher per-CPU capacity is considered better than balancing
10438 		 * average load.
10439 		 */
10440 		if (env->sd->flags & SD_ASYM_CPUCAPACITY &&
10441 		    !capacity_greater(capacity_of(env->dst_cpu), capacity) &&
10442 		    nr_running == 1)
10443 			continue;
10444 
10445 		/* Make sure we only pull tasks from a CPU of lower priority */
10446 		if ((env->sd->flags & SD_ASYM_PACKING) &&
10447 		    sched_asym_prefer(i, env->dst_cpu) &&
10448 		    nr_running == 1)
10449 			continue;
10450 
10451 		switch (env->migration_type) {
10452 		case migrate_load:
10453 			/*
10454 			 * When comparing with load imbalance, use cpu_load()
10455 			 * which is not scaled with the CPU capacity.
10456 			 */
10457 			load = cpu_load(rq);
10458 
10459 			if (nr_running == 1 && load > env->imbalance &&
10460 			    !check_cpu_capacity(rq, env->sd))
10461 				break;
10462 
10463 			/*
10464 			 * For the load comparisons with the other CPUs,
10465 			 * consider the cpu_load() scaled with the CPU
10466 			 * capacity, so that the load can be moved away
10467 			 * from the CPU that is potentially running at a
10468 			 * lower capacity.
10469 			 *
10470 			 * Thus we're looking for max(load_i / capacity_i),
10471 			 * crosswise multiplication to rid ourselves of the
10472 			 * division works out to:
10473 			 * load_i * capacity_j > load_j * capacity_i;
10474 			 * where j is our previous maximum.
10475 			 */
10476 			if (load * busiest_capacity > busiest_load * capacity) {
10477 				busiest_load = load;
10478 				busiest_capacity = capacity;
10479 				busiest = rq;
10480 			}
10481 			break;
10482 
10483 		case migrate_util:
10484 			util = cpu_util_cfs(i);
10485 
10486 			/*
10487 			 * Don't try to pull utilization from a CPU with one
10488 			 * running task. Whatever its utilization, we will fail
10489 			 * detach the task.
10490 			 */
10491 			if (nr_running <= 1)
10492 				continue;
10493 
10494 			if (busiest_util < util) {
10495 				busiest_util = util;
10496 				busiest = rq;
10497 			}
10498 			break;
10499 
10500 		case migrate_task:
10501 			if (busiest_nr < nr_running) {
10502 				busiest_nr = nr_running;
10503 				busiest = rq;
10504 			}
10505 			break;
10506 
10507 		case migrate_misfit:
10508 			/*
10509 			 * For ASYM_CPUCAPACITY domains with misfit tasks we
10510 			 * simply seek the "biggest" misfit task.
10511 			 */
10512 			if (rq->misfit_task_load > busiest_load) {
10513 				busiest_load = rq->misfit_task_load;
10514 				busiest = rq;
10515 			}
10516 
10517 			break;
10518 
10519 		}
10520 	}
10521 
10522 	return busiest;
10523 }
10524 
10525 /*
10526  * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
10527  * so long as it is large enough.
10528  */
10529 #define MAX_PINNED_INTERVAL	512
10530 
10531 static inline bool
10532 asym_active_balance(struct lb_env *env)
10533 {
10534 	/*
10535 	 * ASYM_PACKING needs to force migrate tasks from busy but
10536 	 * lower priority CPUs in order to pack all tasks in the
10537 	 * highest priority CPUs.
10538 	 */
10539 	return env->idle != CPU_NOT_IDLE && (env->sd->flags & SD_ASYM_PACKING) &&
10540 	       sched_asym_prefer(env->dst_cpu, env->src_cpu);
10541 }
10542 
10543 static inline bool
10544 imbalanced_active_balance(struct lb_env *env)
10545 {
10546 	struct sched_domain *sd = env->sd;
10547 
10548 	/*
10549 	 * The imbalanced case includes the case of pinned tasks preventing a fair
10550 	 * distribution of the load on the system but also the even distribution of the
10551 	 * threads on a system with spare capacity
10552 	 */
10553 	if ((env->migration_type == migrate_task) &&
10554 	    (sd->nr_balance_failed > sd->cache_nice_tries+2))
10555 		return 1;
10556 
10557 	return 0;
10558 }
10559 
10560 static int need_active_balance(struct lb_env *env)
10561 {
10562 	struct sched_domain *sd = env->sd;
10563 
10564 	if (asym_active_balance(env))
10565 		return 1;
10566 
10567 	if (imbalanced_active_balance(env))
10568 		return 1;
10569 
10570 	/*
10571 	 * The dst_cpu is idle and the src_cpu CPU has only 1 CFS task.
10572 	 * It's worth migrating the task if the src_cpu's capacity is reduced
10573 	 * because of other sched_class or IRQs if more capacity stays
10574 	 * available on dst_cpu.
10575 	 */
10576 	if ((env->idle != CPU_NOT_IDLE) &&
10577 	    (env->src_rq->cfs.h_nr_running == 1)) {
10578 		if ((check_cpu_capacity(env->src_rq, sd)) &&
10579 		    (capacity_of(env->src_cpu)*sd->imbalance_pct < capacity_of(env->dst_cpu)*100))
10580 			return 1;
10581 	}
10582 
10583 	if (env->migration_type == migrate_misfit)
10584 		return 1;
10585 
10586 	return 0;
10587 }
10588 
10589 static int active_load_balance_cpu_stop(void *data);
10590 
10591 static int should_we_balance(struct lb_env *env)
10592 {
10593 	struct sched_group *sg = env->sd->groups;
10594 	int cpu;
10595 
10596 	/*
10597 	 * Ensure the balancing environment is consistent; can happen
10598 	 * when the softirq triggers 'during' hotplug.
10599 	 */
10600 	if (!cpumask_test_cpu(env->dst_cpu, env->cpus))
10601 		return 0;
10602 
10603 	/*
10604 	 * In the newly idle case, we will allow all the CPUs
10605 	 * to do the newly idle load balance.
10606 	 *
10607 	 * However, we bail out if we already have tasks or a wakeup pending,
10608 	 * to optimize wakeup latency.
10609 	 */
10610 	if (env->idle == CPU_NEWLY_IDLE) {
10611 		if (env->dst_rq->nr_running > 0 || env->dst_rq->ttwu_pending)
10612 			return 0;
10613 		return 1;
10614 	}
10615 
10616 	/* Try to find first idle CPU */
10617 	for_each_cpu_and(cpu, group_balance_mask(sg), env->cpus) {
10618 		if (!idle_cpu(cpu))
10619 			continue;
10620 
10621 		/* Are we the first idle CPU? */
10622 		return cpu == env->dst_cpu;
10623 	}
10624 
10625 	/* Are we the first CPU of this group ? */
10626 	return group_balance_cpu(sg) == env->dst_cpu;
10627 }
10628 
10629 /*
10630  * Check this_cpu to ensure it is balanced within domain. Attempt to move
10631  * tasks if there is an imbalance.
10632  */
10633 static int load_balance(int this_cpu, struct rq *this_rq,
10634 			struct sched_domain *sd, enum cpu_idle_type idle,
10635 			int *continue_balancing)
10636 {
10637 	int ld_moved, cur_ld_moved, active_balance = 0;
10638 	struct sched_domain *sd_parent = sd->parent;
10639 	struct sched_group *group;
10640 	struct rq *busiest;
10641 	struct rq_flags rf;
10642 	struct cpumask *cpus = this_cpu_cpumask_var_ptr(load_balance_mask);
10643 	struct lb_env env = {
10644 		.sd		= sd,
10645 		.dst_cpu	= this_cpu,
10646 		.dst_rq		= this_rq,
10647 		.dst_grpmask    = sched_group_span(sd->groups),
10648 		.idle		= idle,
10649 		.loop_break	= SCHED_NR_MIGRATE_BREAK,
10650 		.cpus		= cpus,
10651 		.fbq_type	= all,
10652 		.tasks		= LIST_HEAD_INIT(env.tasks),
10653 	};
10654 
10655 	cpumask_and(cpus, sched_domain_span(sd), cpu_active_mask);
10656 
10657 	schedstat_inc(sd->lb_count[idle]);
10658 
10659 redo:
10660 	if (!should_we_balance(&env)) {
10661 		*continue_balancing = 0;
10662 		goto out_balanced;
10663 	}
10664 
10665 	group = find_busiest_group(&env);
10666 	if (!group) {
10667 		schedstat_inc(sd->lb_nobusyg[idle]);
10668 		goto out_balanced;
10669 	}
10670 
10671 	busiest = find_busiest_queue(&env, group);
10672 	if (!busiest) {
10673 		schedstat_inc(sd->lb_nobusyq[idle]);
10674 		goto out_balanced;
10675 	}
10676 
10677 	WARN_ON_ONCE(busiest == env.dst_rq);
10678 
10679 	schedstat_add(sd->lb_imbalance[idle], env.imbalance);
10680 
10681 	env.src_cpu = busiest->cpu;
10682 	env.src_rq = busiest;
10683 
10684 	ld_moved = 0;
10685 	/* Clear this flag as soon as we find a pullable task */
10686 	env.flags |= LBF_ALL_PINNED;
10687 	if (busiest->nr_running > 1) {
10688 		/*
10689 		 * Attempt to move tasks. If find_busiest_group has found
10690 		 * an imbalance but busiest->nr_running <= 1, the group is
10691 		 * still unbalanced. ld_moved simply stays zero, so it is
10692 		 * correctly treated as an imbalance.
10693 		 */
10694 		env.loop_max  = min(sysctl_sched_nr_migrate, busiest->nr_running);
10695 
10696 more_balance:
10697 		rq_lock_irqsave(busiest, &rf);
10698 		update_rq_clock(busiest);
10699 
10700 		/*
10701 		 * cur_ld_moved - load moved in current iteration
10702 		 * ld_moved     - cumulative load moved across iterations
10703 		 */
10704 		cur_ld_moved = detach_tasks(&env);
10705 
10706 		/*
10707 		 * We've detached some tasks from busiest_rq. Every
10708 		 * task is masked "TASK_ON_RQ_MIGRATING", so we can safely
10709 		 * unlock busiest->lock, and we are able to be sure
10710 		 * that nobody can manipulate the tasks in parallel.
10711 		 * See task_rq_lock() family for the details.
10712 		 */
10713 
10714 		rq_unlock(busiest, &rf);
10715 
10716 		if (cur_ld_moved) {
10717 			attach_tasks(&env);
10718 			ld_moved += cur_ld_moved;
10719 		}
10720 
10721 		local_irq_restore(rf.flags);
10722 
10723 		if (env.flags & LBF_NEED_BREAK) {
10724 			env.flags &= ~LBF_NEED_BREAK;
10725 			/* Stop if we tried all running tasks */
10726 			if (env.loop < busiest->nr_running)
10727 				goto more_balance;
10728 		}
10729 
10730 		/*
10731 		 * Revisit (affine) tasks on src_cpu that couldn't be moved to
10732 		 * us and move them to an alternate dst_cpu in our sched_group
10733 		 * where they can run. The upper limit on how many times we
10734 		 * iterate on same src_cpu is dependent on number of CPUs in our
10735 		 * sched_group.
10736 		 *
10737 		 * This changes load balance semantics a bit on who can move
10738 		 * load to a given_cpu. In addition to the given_cpu itself
10739 		 * (or a ilb_cpu acting on its behalf where given_cpu is
10740 		 * nohz-idle), we now have balance_cpu in a position to move
10741 		 * load to given_cpu. In rare situations, this may cause
10742 		 * conflicts (balance_cpu and given_cpu/ilb_cpu deciding
10743 		 * _independently_ and at _same_ time to move some load to
10744 		 * given_cpu) causing excess load to be moved to given_cpu.
10745 		 * This however should not happen so much in practice and
10746 		 * moreover subsequent load balance cycles should correct the
10747 		 * excess load moved.
10748 		 */
10749 		if ((env.flags & LBF_DST_PINNED) && env.imbalance > 0) {
10750 
10751 			/* Prevent to re-select dst_cpu via env's CPUs */
10752 			__cpumask_clear_cpu(env.dst_cpu, env.cpus);
10753 
10754 			env.dst_rq	 = cpu_rq(env.new_dst_cpu);
10755 			env.dst_cpu	 = env.new_dst_cpu;
10756 			env.flags	&= ~LBF_DST_PINNED;
10757 			env.loop	 = 0;
10758 			env.loop_break	 = SCHED_NR_MIGRATE_BREAK;
10759 
10760 			/*
10761 			 * Go back to "more_balance" rather than "redo" since we
10762 			 * need to continue with same src_cpu.
10763 			 */
10764 			goto more_balance;
10765 		}
10766 
10767 		/*
10768 		 * We failed to reach balance because of affinity.
10769 		 */
10770 		if (sd_parent) {
10771 			int *group_imbalance = &sd_parent->groups->sgc->imbalance;
10772 
10773 			if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0)
10774 				*group_imbalance = 1;
10775 		}
10776 
10777 		/* All tasks on this runqueue were pinned by CPU affinity */
10778 		if (unlikely(env.flags & LBF_ALL_PINNED)) {
10779 			__cpumask_clear_cpu(cpu_of(busiest), cpus);
10780 			/*
10781 			 * Attempting to continue load balancing at the current
10782 			 * sched_domain level only makes sense if there are
10783 			 * active CPUs remaining as possible busiest CPUs to
10784 			 * pull load from which are not contained within the
10785 			 * destination group that is receiving any migrated
10786 			 * load.
10787 			 */
10788 			if (!cpumask_subset(cpus, env.dst_grpmask)) {
10789 				env.loop = 0;
10790 				env.loop_break = SCHED_NR_MIGRATE_BREAK;
10791 				goto redo;
10792 			}
10793 			goto out_all_pinned;
10794 		}
10795 	}
10796 
10797 	if (!ld_moved) {
10798 		schedstat_inc(sd->lb_failed[idle]);
10799 		/*
10800 		 * Increment the failure counter only on periodic balance.
10801 		 * We do not want newidle balance, which can be very
10802 		 * frequent, pollute the failure counter causing
10803 		 * excessive cache_hot migrations and active balances.
10804 		 */
10805 		if (idle != CPU_NEWLY_IDLE)
10806 			sd->nr_balance_failed++;
10807 
10808 		if (need_active_balance(&env)) {
10809 			unsigned long flags;
10810 
10811 			raw_spin_rq_lock_irqsave(busiest, flags);
10812 
10813 			/*
10814 			 * Don't kick the active_load_balance_cpu_stop,
10815 			 * if the curr task on busiest CPU can't be
10816 			 * moved to this_cpu:
10817 			 */
10818 			if (!cpumask_test_cpu(this_cpu, busiest->curr->cpus_ptr)) {
10819 				raw_spin_rq_unlock_irqrestore(busiest, flags);
10820 				goto out_one_pinned;
10821 			}
10822 
10823 			/* Record that we found at least one task that could run on this_cpu */
10824 			env.flags &= ~LBF_ALL_PINNED;
10825 
10826 			/*
10827 			 * ->active_balance synchronizes accesses to
10828 			 * ->active_balance_work.  Once set, it's cleared
10829 			 * only after active load balance is finished.
10830 			 */
10831 			if (!busiest->active_balance) {
10832 				busiest->active_balance = 1;
10833 				busiest->push_cpu = this_cpu;
10834 				active_balance = 1;
10835 			}
10836 			raw_spin_rq_unlock_irqrestore(busiest, flags);
10837 
10838 			if (active_balance) {
10839 				stop_one_cpu_nowait(cpu_of(busiest),
10840 					active_load_balance_cpu_stop, busiest,
10841 					&busiest->active_balance_work);
10842 			}
10843 		}
10844 	} else {
10845 		sd->nr_balance_failed = 0;
10846 	}
10847 
10848 	if (likely(!active_balance) || need_active_balance(&env)) {
10849 		/* We were unbalanced, so reset the balancing interval */
10850 		sd->balance_interval = sd->min_interval;
10851 	}
10852 
10853 	goto out;
10854 
10855 out_balanced:
10856 	/*
10857 	 * We reach balance although we may have faced some affinity
10858 	 * constraints. Clear the imbalance flag only if other tasks got
10859 	 * a chance to move and fix the imbalance.
10860 	 */
10861 	if (sd_parent && !(env.flags & LBF_ALL_PINNED)) {
10862 		int *group_imbalance = &sd_parent->groups->sgc->imbalance;
10863 
10864 		if (*group_imbalance)
10865 			*group_imbalance = 0;
10866 	}
10867 
10868 out_all_pinned:
10869 	/*
10870 	 * We reach balance because all tasks are pinned at this level so
10871 	 * we can't migrate them. Let the imbalance flag set so parent level
10872 	 * can try to migrate them.
10873 	 */
10874 	schedstat_inc(sd->lb_balanced[idle]);
10875 
10876 	sd->nr_balance_failed = 0;
10877 
10878 out_one_pinned:
10879 	ld_moved = 0;
10880 
10881 	/*
10882 	 * newidle_balance() disregards balance intervals, so we could
10883 	 * repeatedly reach this code, which would lead to balance_interval
10884 	 * skyrocketing in a short amount of time. Skip the balance_interval
10885 	 * increase logic to avoid that.
10886 	 */
10887 	if (env.idle == CPU_NEWLY_IDLE)
10888 		goto out;
10889 
10890 	/* tune up the balancing interval */
10891 	if ((env.flags & LBF_ALL_PINNED &&
10892 	     sd->balance_interval < MAX_PINNED_INTERVAL) ||
10893 	    sd->balance_interval < sd->max_interval)
10894 		sd->balance_interval *= 2;
10895 out:
10896 	return ld_moved;
10897 }
10898 
10899 static inline unsigned long
10900 get_sd_balance_interval(struct sched_domain *sd, int cpu_busy)
10901 {
10902 	unsigned long interval = sd->balance_interval;
10903 
10904 	if (cpu_busy)
10905 		interval *= sd->busy_factor;
10906 
10907 	/* scale ms to jiffies */
10908 	interval = msecs_to_jiffies(interval);
10909 
10910 	/*
10911 	 * Reduce likelihood of busy balancing at higher domains racing with
10912 	 * balancing at lower domains by preventing their balancing periods
10913 	 * from being multiples of each other.
10914 	 */
10915 	if (cpu_busy)
10916 		interval -= 1;
10917 
10918 	interval = clamp(interval, 1UL, max_load_balance_interval);
10919 
10920 	return interval;
10921 }
10922 
10923 static inline void
10924 update_next_balance(struct sched_domain *sd, unsigned long *next_balance)
10925 {
10926 	unsigned long interval, next;
10927 
10928 	/* used by idle balance, so cpu_busy = 0 */
10929 	interval = get_sd_balance_interval(sd, 0);
10930 	next = sd->last_balance + interval;
10931 
10932 	if (time_after(*next_balance, next))
10933 		*next_balance = next;
10934 }
10935 
10936 /*
10937  * active_load_balance_cpu_stop is run by the CPU stopper. It pushes
10938  * running tasks off the busiest CPU onto idle CPUs. It requires at
10939  * least 1 task to be running on each physical CPU where possible, and
10940  * avoids physical / logical imbalances.
10941  */
10942 static int active_load_balance_cpu_stop(void *data)
10943 {
10944 	struct rq *busiest_rq = data;
10945 	int busiest_cpu = cpu_of(busiest_rq);
10946 	int target_cpu = busiest_rq->push_cpu;
10947 	struct rq *target_rq = cpu_rq(target_cpu);
10948 	struct sched_domain *sd;
10949 	struct task_struct *p = NULL;
10950 	struct rq_flags rf;
10951 
10952 	rq_lock_irq(busiest_rq, &rf);
10953 	/*
10954 	 * Between queueing the stop-work and running it is a hole in which
10955 	 * CPUs can become inactive. We should not move tasks from or to
10956 	 * inactive CPUs.
10957 	 */
10958 	if (!cpu_active(busiest_cpu) || !cpu_active(target_cpu))
10959 		goto out_unlock;
10960 
10961 	/* Make sure the requested CPU hasn't gone down in the meantime: */
10962 	if (unlikely(busiest_cpu != smp_processor_id() ||
10963 		     !busiest_rq->active_balance))
10964 		goto out_unlock;
10965 
10966 	/* Is there any task to move? */
10967 	if (busiest_rq->nr_running <= 1)
10968 		goto out_unlock;
10969 
10970 	/*
10971 	 * This condition is "impossible", if it occurs
10972 	 * we need to fix it. Originally reported by
10973 	 * Bjorn Helgaas on a 128-CPU setup.
10974 	 */
10975 	WARN_ON_ONCE(busiest_rq == target_rq);
10976 
10977 	/* Search for an sd spanning us and the target CPU. */
10978 	rcu_read_lock();
10979 	for_each_domain(target_cpu, sd) {
10980 		if (cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
10981 			break;
10982 	}
10983 
10984 	if (likely(sd)) {
10985 		struct lb_env env = {
10986 			.sd		= sd,
10987 			.dst_cpu	= target_cpu,
10988 			.dst_rq		= target_rq,
10989 			.src_cpu	= busiest_rq->cpu,
10990 			.src_rq		= busiest_rq,
10991 			.idle		= CPU_IDLE,
10992 			.flags		= LBF_ACTIVE_LB,
10993 		};
10994 
10995 		schedstat_inc(sd->alb_count);
10996 		update_rq_clock(busiest_rq);
10997 
10998 		p = detach_one_task(&env);
10999 		if (p) {
11000 			schedstat_inc(sd->alb_pushed);
11001 			/* Active balancing done, reset the failure counter. */
11002 			sd->nr_balance_failed = 0;
11003 		} else {
11004 			schedstat_inc(sd->alb_failed);
11005 		}
11006 	}
11007 	rcu_read_unlock();
11008 out_unlock:
11009 	busiest_rq->active_balance = 0;
11010 	rq_unlock(busiest_rq, &rf);
11011 
11012 	if (p)
11013 		attach_one_task(target_rq, p);
11014 
11015 	local_irq_enable();
11016 
11017 	return 0;
11018 }
11019 
11020 static DEFINE_SPINLOCK(balancing);
11021 
11022 /*
11023  * Scale the max load_balance interval with the number of CPUs in the system.
11024  * This trades load-balance latency on larger machines for less cross talk.
11025  */
11026 void update_max_interval(void)
11027 {
11028 	max_load_balance_interval = HZ*num_online_cpus()/10;
11029 }
11030 
11031 static inline bool update_newidle_cost(struct sched_domain *sd, u64 cost)
11032 {
11033 	if (cost > sd->max_newidle_lb_cost) {
11034 		/*
11035 		 * Track max cost of a domain to make sure to not delay the
11036 		 * next wakeup on the CPU.
11037 		 */
11038 		sd->max_newidle_lb_cost = cost;
11039 		sd->last_decay_max_lb_cost = jiffies;
11040 	} else if (time_after(jiffies, sd->last_decay_max_lb_cost + HZ)) {
11041 		/*
11042 		 * Decay the newidle max times by ~1% per second to ensure that
11043 		 * it is not outdated and the current max cost is actually
11044 		 * shorter.
11045 		 */
11046 		sd->max_newidle_lb_cost = (sd->max_newidle_lb_cost * 253) / 256;
11047 		sd->last_decay_max_lb_cost = jiffies;
11048 
11049 		return true;
11050 	}
11051 
11052 	return false;
11053 }
11054 
11055 /*
11056  * It checks each scheduling domain to see if it is due to be balanced,
11057  * and initiates a balancing operation if so.
11058  *
11059  * Balancing parameters are set up in init_sched_domains.
11060  */
11061 static void rebalance_domains(struct rq *rq, enum cpu_idle_type idle)
11062 {
11063 	int continue_balancing = 1;
11064 	int cpu = rq->cpu;
11065 	int busy = idle != CPU_IDLE && !sched_idle_cpu(cpu);
11066 	unsigned long interval;
11067 	struct sched_domain *sd;
11068 	/* Earliest time when we have to do rebalance again */
11069 	unsigned long next_balance = jiffies + 60*HZ;
11070 	int update_next_balance = 0;
11071 	int need_serialize, need_decay = 0;
11072 	u64 max_cost = 0;
11073 
11074 	rcu_read_lock();
11075 	for_each_domain(cpu, sd) {
11076 		/*
11077 		 * Decay the newidle max times here because this is a regular
11078 		 * visit to all the domains.
11079 		 */
11080 		need_decay = update_newidle_cost(sd, 0);
11081 		max_cost += sd->max_newidle_lb_cost;
11082 
11083 		/*
11084 		 * Stop the load balance at this level. There is another
11085 		 * CPU in our sched group which is doing load balancing more
11086 		 * actively.
11087 		 */
11088 		if (!continue_balancing) {
11089 			if (need_decay)
11090 				continue;
11091 			break;
11092 		}
11093 
11094 		interval = get_sd_balance_interval(sd, busy);
11095 
11096 		need_serialize = sd->flags & SD_SERIALIZE;
11097 		if (need_serialize) {
11098 			if (!spin_trylock(&balancing))
11099 				goto out;
11100 		}
11101 
11102 		if (time_after_eq(jiffies, sd->last_balance + interval)) {
11103 			if (load_balance(cpu, rq, sd, idle, &continue_balancing)) {
11104 				/*
11105 				 * The LBF_DST_PINNED logic could have changed
11106 				 * env->dst_cpu, so we can't know our idle
11107 				 * state even if we migrated tasks. Update it.
11108 				 */
11109 				idle = idle_cpu(cpu) ? CPU_IDLE : CPU_NOT_IDLE;
11110 				busy = idle != CPU_IDLE && !sched_idle_cpu(cpu);
11111 			}
11112 			sd->last_balance = jiffies;
11113 			interval = get_sd_balance_interval(sd, busy);
11114 		}
11115 		if (need_serialize)
11116 			spin_unlock(&balancing);
11117 out:
11118 		if (time_after(next_balance, sd->last_balance + interval)) {
11119 			next_balance = sd->last_balance + interval;
11120 			update_next_balance = 1;
11121 		}
11122 	}
11123 	if (need_decay) {
11124 		/*
11125 		 * Ensure the rq-wide value also decays but keep it at a
11126 		 * reasonable floor to avoid funnies with rq->avg_idle.
11127 		 */
11128 		rq->max_idle_balance_cost =
11129 			max((u64)sysctl_sched_migration_cost, max_cost);
11130 	}
11131 	rcu_read_unlock();
11132 
11133 	/*
11134 	 * next_balance will be updated only when there is a need.
11135 	 * When the cpu is attached to null domain for ex, it will not be
11136 	 * updated.
11137 	 */
11138 	if (likely(update_next_balance))
11139 		rq->next_balance = next_balance;
11140 
11141 }
11142 
11143 static inline int on_null_domain(struct rq *rq)
11144 {
11145 	return unlikely(!rcu_dereference_sched(rq->sd));
11146 }
11147 
11148 #ifdef CONFIG_NO_HZ_COMMON
11149 /*
11150  * idle load balancing details
11151  * - When one of the busy CPUs notice that there may be an idle rebalancing
11152  *   needed, they will kick the idle load balancer, which then does idle
11153  *   load balancing for all the idle CPUs.
11154  * - HK_TYPE_MISC CPUs are used for this task, because HK_TYPE_SCHED not set
11155  *   anywhere yet.
11156  */
11157 
11158 static inline int find_new_ilb(void)
11159 {
11160 	int ilb;
11161 	const struct cpumask *hk_mask;
11162 
11163 	hk_mask = housekeeping_cpumask(HK_TYPE_MISC);
11164 
11165 	for_each_cpu_and(ilb, nohz.idle_cpus_mask, hk_mask) {
11166 
11167 		if (ilb == smp_processor_id())
11168 			continue;
11169 
11170 		if (idle_cpu(ilb))
11171 			return ilb;
11172 	}
11173 
11174 	return nr_cpu_ids;
11175 }
11176 
11177 /*
11178  * Kick a CPU to do the nohz balancing, if it is time for it. We pick any
11179  * idle CPU in the HK_TYPE_MISC housekeeping set (if there is one).
11180  */
11181 static void kick_ilb(unsigned int flags)
11182 {
11183 	int ilb_cpu;
11184 
11185 	/*
11186 	 * Increase nohz.next_balance only when if full ilb is triggered but
11187 	 * not if we only update stats.
11188 	 */
11189 	if (flags & NOHZ_BALANCE_KICK)
11190 		nohz.next_balance = jiffies+1;
11191 
11192 	ilb_cpu = find_new_ilb();
11193 
11194 	if (ilb_cpu >= nr_cpu_ids)
11195 		return;
11196 
11197 	/*
11198 	 * Access to rq::nohz_csd is serialized by NOHZ_KICK_MASK; he who sets
11199 	 * the first flag owns it; cleared by nohz_csd_func().
11200 	 */
11201 	flags = atomic_fetch_or(flags, nohz_flags(ilb_cpu));
11202 	if (flags & NOHZ_KICK_MASK)
11203 		return;
11204 
11205 	/*
11206 	 * This way we generate an IPI on the target CPU which
11207 	 * is idle. And the softirq performing nohz idle load balance
11208 	 * will be run before returning from the IPI.
11209 	 */
11210 	smp_call_function_single_async(ilb_cpu, &cpu_rq(ilb_cpu)->nohz_csd);
11211 }
11212 
11213 /*
11214  * Current decision point for kicking the idle load balancer in the presence
11215  * of idle CPUs in the system.
11216  */
11217 static void nohz_balancer_kick(struct rq *rq)
11218 {
11219 	unsigned long now = jiffies;
11220 	struct sched_domain_shared *sds;
11221 	struct sched_domain *sd;
11222 	int nr_busy, i, cpu = rq->cpu;
11223 	unsigned int flags = 0;
11224 
11225 	if (unlikely(rq->idle_balance))
11226 		return;
11227 
11228 	/*
11229 	 * We may be recently in ticked or tickless idle mode. At the first
11230 	 * busy tick after returning from idle, we will update the busy stats.
11231 	 */
11232 	nohz_balance_exit_idle(rq);
11233 
11234 	/*
11235 	 * None are in tickless mode and hence no need for NOHZ idle load
11236 	 * balancing.
11237 	 */
11238 	if (likely(!atomic_read(&nohz.nr_cpus)))
11239 		return;
11240 
11241 	if (READ_ONCE(nohz.has_blocked) &&
11242 	    time_after(now, READ_ONCE(nohz.next_blocked)))
11243 		flags = NOHZ_STATS_KICK;
11244 
11245 	if (time_before(now, nohz.next_balance))
11246 		goto out;
11247 
11248 	if (rq->nr_running >= 2) {
11249 		flags = NOHZ_STATS_KICK | NOHZ_BALANCE_KICK;
11250 		goto out;
11251 	}
11252 
11253 	rcu_read_lock();
11254 
11255 	sd = rcu_dereference(rq->sd);
11256 	if (sd) {
11257 		/*
11258 		 * If there's a CFS task and the current CPU has reduced
11259 		 * capacity; kick the ILB to see if there's a better CPU to run
11260 		 * on.
11261 		 */
11262 		if (rq->cfs.h_nr_running >= 1 && check_cpu_capacity(rq, sd)) {
11263 			flags = NOHZ_STATS_KICK | NOHZ_BALANCE_KICK;
11264 			goto unlock;
11265 		}
11266 	}
11267 
11268 	sd = rcu_dereference(per_cpu(sd_asym_packing, cpu));
11269 	if (sd) {
11270 		/*
11271 		 * When ASYM_PACKING; see if there's a more preferred CPU
11272 		 * currently idle; in which case, kick the ILB to move tasks
11273 		 * around.
11274 		 */
11275 		for_each_cpu_and(i, sched_domain_span(sd), nohz.idle_cpus_mask) {
11276 			if (sched_asym_prefer(i, cpu)) {
11277 				flags = NOHZ_STATS_KICK | NOHZ_BALANCE_KICK;
11278 				goto unlock;
11279 			}
11280 		}
11281 	}
11282 
11283 	sd = rcu_dereference(per_cpu(sd_asym_cpucapacity, cpu));
11284 	if (sd) {
11285 		/*
11286 		 * When ASYM_CPUCAPACITY; see if there's a higher capacity CPU
11287 		 * to run the misfit task on.
11288 		 */
11289 		if (check_misfit_status(rq, sd)) {
11290 			flags = NOHZ_STATS_KICK | NOHZ_BALANCE_KICK;
11291 			goto unlock;
11292 		}
11293 
11294 		/*
11295 		 * For asymmetric systems, we do not want to nicely balance
11296 		 * cache use, instead we want to embrace asymmetry and only
11297 		 * ensure tasks have enough CPU capacity.
11298 		 *
11299 		 * Skip the LLC logic because it's not relevant in that case.
11300 		 */
11301 		goto unlock;
11302 	}
11303 
11304 	sds = rcu_dereference(per_cpu(sd_llc_shared, cpu));
11305 	if (sds) {
11306 		/*
11307 		 * If there is an imbalance between LLC domains (IOW we could
11308 		 * increase the overall cache use), we need some less-loaded LLC
11309 		 * domain to pull some load. Likewise, we may need to spread
11310 		 * load within the current LLC domain (e.g. packed SMT cores but
11311 		 * other CPUs are idle). We can't really know from here how busy
11312 		 * the others are - so just get a nohz balance going if it looks
11313 		 * like this LLC domain has tasks we could move.
11314 		 */
11315 		nr_busy = atomic_read(&sds->nr_busy_cpus);
11316 		if (nr_busy > 1) {
11317 			flags = NOHZ_STATS_KICK | NOHZ_BALANCE_KICK;
11318 			goto unlock;
11319 		}
11320 	}
11321 unlock:
11322 	rcu_read_unlock();
11323 out:
11324 	if (READ_ONCE(nohz.needs_update))
11325 		flags |= NOHZ_NEXT_KICK;
11326 
11327 	if (flags)
11328 		kick_ilb(flags);
11329 }
11330 
11331 static void set_cpu_sd_state_busy(int cpu)
11332 {
11333 	struct sched_domain *sd;
11334 
11335 	rcu_read_lock();
11336 	sd = rcu_dereference(per_cpu(sd_llc, cpu));
11337 
11338 	if (!sd || !sd->nohz_idle)
11339 		goto unlock;
11340 	sd->nohz_idle = 0;
11341 
11342 	atomic_inc(&sd->shared->nr_busy_cpus);
11343 unlock:
11344 	rcu_read_unlock();
11345 }
11346 
11347 void nohz_balance_exit_idle(struct rq *rq)
11348 {
11349 	SCHED_WARN_ON(rq != this_rq());
11350 
11351 	if (likely(!rq->nohz_tick_stopped))
11352 		return;
11353 
11354 	rq->nohz_tick_stopped = 0;
11355 	cpumask_clear_cpu(rq->cpu, nohz.idle_cpus_mask);
11356 	atomic_dec(&nohz.nr_cpus);
11357 
11358 	set_cpu_sd_state_busy(rq->cpu);
11359 }
11360 
11361 static void set_cpu_sd_state_idle(int cpu)
11362 {
11363 	struct sched_domain *sd;
11364 
11365 	rcu_read_lock();
11366 	sd = rcu_dereference(per_cpu(sd_llc, cpu));
11367 
11368 	if (!sd || sd->nohz_idle)
11369 		goto unlock;
11370 	sd->nohz_idle = 1;
11371 
11372 	atomic_dec(&sd->shared->nr_busy_cpus);
11373 unlock:
11374 	rcu_read_unlock();
11375 }
11376 
11377 /*
11378  * This routine will record that the CPU is going idle with tick stopped.
11379  * This info will be used in performing idle load balancing in the future.
11380  */
11381 void nohz_balance_enter_idle(int cpu)
11382 {
11383 	struct rq *rq = cpu_rq(cpu);
11384 
11385 	SCHED_WARN_ON(cpu != smp_processor_id());
11386 
11387 	/* If this CPU is going down, then nothing needs to be done: */
11388 	if (!cpu_active(cpu))
11389 		return;
11390 
11391 	/* Spare idle load balancing on CPUs that don't want to be disturbed: */
11392 	if (!housekeeping_cpu(cpu, HK_TYPE_SCHED))
11393 		return;
11394 
11395 	/*
11396 	 * Can be set safely without rq->lock held
11397 	 * If a clear happens, it will have evaluated last additions because
11398 	 * rq->lock is held during the check and the clear
11399 	 */
11400 	rq->has_blocked_load = 1;
11401 
11402 	/*
11403 	 * The tick is still stopped but load could have been added in the
11404 	 * meantime. We set the nohz.has_blocked flag to trig a check of the
11405 	 * *_avg. The CPU is already part of nohz.idle_cpus_mask so the clear
11406 	 * of nohz.has_blocked can only happen after checking the new load
11407 	 */
11408 	if (rq->nohz_tick_stopped)
11409 		goto out;
11410 
11411 	/* If we're a completely isolated CPU, we don't play: */
11412 	if (on_null_domain(rq))
11413 		return;
11414 
11415 	rq->nohz_tick_stopped = 1;
11416 
11417 	cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
11418 	atomic_inc(&nohz.nr_cpus);
11419 
11420 	/*
11421 	 * Ensures that if nohz_idle_balance() fails to observe our
11422 	 * @idle_cpus_mask store, it must observe the @has_blocked
11423 	 * and @needs_update stores.
11424 	 */
11425 	smp_mb__after_atomic();
11426 
11427 	set_cpu_sd_state_idle(cpu);
11428 
11429 	WRITE_ONCE(nohz.needs_update, 1);
11430 out:
11431 	/*
11432 	 * Each time a cpu enter idle, we assume that it has blocked load and
11433 	 * enable the periodic update of the load of idle cpus
11434 	 */
11435 	WRITE_ONCE(nohz.has_blocked, 1);
11436 }
11437 
11438 static bool update_nohz_stats(struct rq *rq)
11439 {
11440 	unsigned int cpu = rq->cpu;
11441 
11442 	if (!rq->has_blocked_load)
11443 		return false;
11444 
11445 	if (!cpumask_test_cpu(cpu, nohz.idle_cpus_mask))
11446 		return false;
11447 
11448 	if (!time_after(jiffies, READ_ONCE(rq->last_blocked_load_update_tick)))
11449 		return true;
11450 
11451 	update_blocked_averages(cpu);
11452 
11453 	return rq->has_blocked_load;
11454 }
11455 
11456 /*
11457  * Internal function that runs load balance for all idle cpus. The load balance
11458  * can be a simple update of blocked load or a complete load balance with
11459  * tasks movement depending of flags.
11460  */
11461 static void _nohz_idle_balance(struct rq *this_rq, unsigned int flags)
11462 {
11463 	/* Earliest time when we have to do rebalance again */
11464 	unsigned long now = jiffies;
11465 	unsigned long next_balance = now + 60*HZ;
11466 	bool has_blocked_load = false;
11467 	int update_next_balance = 0;
11468 	int this_cpu = this_rq->cpu;
11469 	int balance_cpu;
11470 	struct rq *rq;
11471 
11472 	SCHED_WARN_ON((flags & NOHZ_KICK_MASK) == NOHZ_BALANCE_KICK);
11473 
11474 	/*
11475 	 * We assume there will be no idle load after this update and clear
11476 	 * the has_blocked flag. If a cpu enters idle in the mean time, it will
11477 	 * set the has_blocked flag and trigger another update of idle load.
11478 	 * Because a cpu that becomes idle, is added to idle_cpus_mask before
11479 	 * setting the flag, we are sure to not clear the state and not
11480 	 * check the load of an idle cpu.
11481 	 *
11482 	 * Same applies to idle_cpus_mask vs needs_update.
11483 	 */
11484 	if (flags & NOHZ_STATS_KICK)
11485 		WRITE_ONCE(nohz.has_blocked, 0);
11486 	if (flags & NOHZ_NEXT_KICK)
11487 		WRITE_ONCE(nohz.needs_update, 0);
11488 
11489 	/*
11490 	 * Ensures that if we miss the CPU, we must see the has_blocked
11491 	 * store from nohz_balance_enter_idle().
11492 	 */
11493 	smp_mb();
11494 
11495 	/*
11496 	 * Start with the next CPU after this_cpu so we will end with this_cpu and let a
11497 	 * chance for other idle cpu to pull load.
11498 	 */
11499 	for_each_cpu_wrap(balance_cpu,  nohz.idle_cpus_mask, this_cpu+1) {
11500 		if (!idle_cpu(balance_cpu))
11501 			continue;
11502 
11503 		/*
11504 		 * If this CPU gets work to do, stop the load balancing
11505 		 * work being done for other CPUs. Next load
11506 		 * balancing owner will pick it up.
11507 		 */
11508 		if (need_resched()) {
11509 			if (flags & NOHZ_STATS_KICK)
11510 				has_blocked_load = true;
11511 			if (flags & NOHZ_NEXT_KICK)
11512 				WRITE_ONCE(nohz.needs_update, 1);
11513 			goto abort;
11514 		}
11515 
11516 		rq = cpu_rq(balance_cpu);
11517 
11518 		if (flags & NOHZ_STATS_KICK)
11519 			has_blocked_load |= update_nohz_stats(rq);
11520 
11521 		/*
11522 		 * If time for next balance is due,
11523 		 * do the balance.
11524 		 */
11525 		if (time_after_eq(jiffies, rq->next_balance)) {
11526 			struct rq_flags rf;
11527 
11528 			rq_lock_irqsave(rq, &rf);
11529 			update_rq_clock(rq);
11530 			rq_unlock_irqrestore(rq, &rf);
11531 
11532 			if (flags & NOHZ_BALANCE_KICK)
11533 				rebalance_domains(rq, CPU_IDLE);
11534 		}
11535 
11536 		if (time_after(next_balance, rq->next_balance)) {
11537 			next_balance = rq->next_balance;
11538 			update_next_balance = 1;
11539 		}
11540 	}
11541 
11542 	/*
11543 	 * next_balance will be updated only when there is a need.
11544 	 * When the CPU is attached to null domain for ex, it will not be
11545 	 * updated.
11546 	 */
11547 	if (likely(update_next_balance))
11548 		nohz.next_balance = next_balance;
11549 
11550 	if (flags & NOHZ_STATS_KICK)
11551 		WRITE_ONCE(nohz.next_blocked,
11552 			   now + msecs_to_jiffies(LOAD_AVG_PERIOD));
11553 
11554 abort:
11555 	/* There is still blocked load, enable periodic update */
11556 	if (has_blocked_load)
11557 		WRITE_ONCE(nohz.has_blocked, 1);
11558 }
11559 
11560 /*
11561  * In CONFIG_NO_HZ_COMMON case, the idle balance kickee will do the
11562  * rebalancing for all the cpus for whom scheduler ticks are stopped.
11563  */
11564 static bool nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
11565 {
11566 	unsigned int flags = this_rq->nohz_idle_balance;
11567 
11568 	if (!flags)
11569 		return false;
11570 
11571 	this_rq->nohz_idle_balance = 0;
11572 
11573 	if (idle != CPU_IDLE)
11574 		return false;
11575 
11576 	_nohz_idle_balance(this_rq, flags);
11577 
11578 	return true;
11579 }
11580 
11581 /*
11582  * Check if we need to run the ILB for updating blocked load before entering
11583  * idle state.
11584  */
11585 void nohz_run_idle_balance(int cpu)
11586 {
11587 	unsigned int flags;
11588 
11589 	flags = atomic_fetch_andnot(NOHZ_NEWILB_KICK, nohz_flags(cpu));
11590 
11591 	/*
11592 	 * Update the blocked load only if no SCHED_SOFTIRQ is about to happen
11593 	 * (ie NOHZ_STATS_KICK set) and will do the same.
11594 	 */
11595 	if ((flags == NOHZ_NEWILB_KICK) && !need_resched())
11596 		_nohz_idle_balance(cpu_rq(cpu), NOHZ_STATS_KICK);
11597 }
11598 
11599 static void nohz_newidle_balance(struct rq *this_rq)
11600 {
11601 	int this_cpu = this_rq->cpu;
11602 
11603 	/*
11604 	 * This CPU doesn't want to be disturbed by scheduler
11605 	 * housekeeping
11606 	 */
11607 	if (!housekeeping_cpu(this_cpu, HK_TYPE_SCHED))
11608 		return;
11609 
11610 	/* Will wake up very soon. No time for doing anything else*/
11611 	if (this_rq->avg_idle < sysctl_sched_migration_cost)
11612 		return;
11613 
11614 	/* Don't need to update blocked load of idle CPUs*/
11615 	if (!READ_ONCE(nohz.has_blocked) ||
11616 	    time_before(jiffies, READ_ONCE(nohz.next_blocked)))
11617 		return;
11618 
11619 	/*
11620 	 * Set the need to trigger ILB in order to update blocked load
11621 	 * before entering idle state.
11622 	 */
11623 	atomic_or(NOHZ_NEWILB_KICK, nohz_flags(this_cpu));
11624 }
11625 
11626 #else /* !CONFIG_NO_HZ_COMMON */
11627 static inline void nohz_balancer_kick(struct rq *rq) { }
11628 
11629 static inline bool nohz_idle_balance(struct rq *this_rq, enum cpu_idle_type idle)
11630 {
11631 	return false;
11632 }
11633 
11634 static inline void nohz_newidle_balance(struct rq *this_rq) { }
11635 #endif /* CONFIG_NO_HZ_COMMON */
11636 
11637 /*
11638  * newidle_balance is called by schedule() if this_cpu is about to become
11639  * idle. Attempts to pull tasks from other CPUs.
11640  *
11641  * Returns:
11642  *   < 0 - we released the lock and there are !fair tasks present
11643  *     0 - failed, no new tasks
11644  *   > 0 - success, new (fair) tasks present
11645  */
11646 static int newidle_balance(struct rq *this_rq, struct rq_flags *rf)
11647 {
11648 	unsigned long next_balance = jiffies + HZ;
11649 	int this_cpu = this_rq->cpu;
11650 	u64 t0, t1, curr_cost = 0;
11651 	struct sched_domain *sd;
11652 	int pulled_task = 0;
11653 
11654 	update_misfit_status(NULL, this_rq);
11655 
11656 	/*
11657 	 * There is a task waiting to run. No need to search for one.
11658 	 * Return 0; the task will be enqueued when switching to idle.
11659 	 */
11660 	if (this_rq->ttwu_pending)
11661 		return 0;
11662 
11663 	/*
11664 	 * We must set idle_stamp _before_ calling idle_balance(), such that we
11665 	 * measure the duration of idle_balance() as idle time.
11666 	 */
11667 	this_rq->idle_stamp = rq_clock(this_rq);
11668 
11669 	/*
11670 	 * Do not pull tasks towards !active CPUs...
11671 	 */
11672 	if (!cpu_active(this_cpu))
11673 		return 0;
11674 
11675 	/*
11676 	 * This is OK, because current is on_cpu, which avoids it being picked
11677 	 * for load-balance and preemption/IRQs are still disabled avoiding
11678 	 * further scheduler activity on it and we're being very careful to
11679 	 * re-start the picking loop.
11680 	 */
11681 	rq_unpin_lock(this_rq, rf);
11682 
11683 	rcu_read_lock();
11684 	sd = rcu_dereference_check_sched_domain(this_rq->sd);
11685 
11686 	if (!READ_ONCE(this_rq->rd->overload) ||
11687 	    (sd && this_rq->avg_idle < sd->max_newidle_lb_cost)) {
11688 
11689 		if (sd)
11690 			update_next_balance(sd, &next_balance);
11691 		rcu_read_unlock();
11692 
11693 		goto out;
11694 	}
11695 	rcu_read_unlock();
11696 
11697 	raw_spin_rq_unlock(this_rq);
11698 
11699 	t0 = sched_clock_cpu(this_cpu);
11700 	update_blocked_averages(this_cpu);
11701 
11702 	rcu_read_lock();
11703 	for_each_domain(this_cpu, sd) {
11704 		int continue_balancing = 1;
11705 		u64 domain_cost;
11706 
11707 		update_next_balance(sd, &next_balance);
11708 
11709 		if (this_rq->avg_idle < curr_cost + sd->max_newidle_lb_cost)
11710 			break;
11711 
11712 		if (sd->flags & SD_BALANCE_NEWIDLE) {
11713 
11714 			pulled_task = load_balance(this_cpu, this_rq,
11715 						   sd, CPU_NEWLY_IDLE,
11716 						   &continue_balancing);
11717 
11718 			t1 = sched_clock_cpu(this_cpu);
11719 			domain_cost = t1 - t0;
11720 			update_newidle_cost(sd, domain_cost);
11721 
11722 			curr_cost += domain_cost;
11723 			t0 = t1;
11724 		}
11725 
11726 		/*
11727 		 * Stop searching for tasks to pull if there are
11728 		 * now runnable tasks on this rq.
11729 		 */
11730 		if (pulled_task || this_rq->nr_running > 0 ||
11731 		    this_rq->ttwu_pending)
11732 			break;
11733 	}
11734 	rcu_read_unlock();
11735 
11736 	raw_spin_rq_lock(this_rq);
11737 
11738 	if (curr_cost > this_rq->max_idle_balance_cost)
11739 		this_rq->max_idle_balance_cost = curr_cost;
11740 
11741 	/*
11742 	 * While browsing the domains, we released the rq lock, a task could
11743 	 * have been enqueued in the meantime. Since we're not going idle,
11744 	 * pretend we pulled a task.
11745 	 */
11746 	if (this_rq->cfs.h_nr_running && !pulled_task)
11747 		pulled_task = 1;
11748 
11749 	/* Is there a task of a high priority class? */
11750 	if (this_rq->nr_running != this_rq->cfs.h_nr_running)
11751 		pulled_task = -1;
11752 
11753 out:
11754 	/* Move the next balance forward */
11755 	if (time_after(this_rq->next_balance, next_balance))
11756 		this_rq->next_balance = next_balance;
11757 
11758 	if (pulled_task)
11759 		this_rq->idle_stamp = 0;
11760 	else
11761 		nohz_newidle_balance(this_rq);
11762 
11763 	rq_repin_lock(this_rq, rf);
11764 
11765 	return pulled_task;
11766 }
11767 
11768 /*
11769  * run_rebalance_domains is triggered when needed from the scheduler tick.
11770  * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
11771  */
11772 static __latent_entropy void run_rebalance_domains(struct softirq_action *h)
11773 {
11774 	struct rq *this_rq = this_rq();
11775 	enum cpu_idle_type idle = this_rq->idle_balance ?
11776 						CPU_IDLE : CPU_NOT_IDLE;
11777 
11778 	/*
11779 	 * If this CPU has a pending nohz_balance_kick, then do the
11780 	 * balancing on behalf of the other idle CPUs whose ticks are
11781 	 * stopped. Do nohz_idle_balance *before* rebalance_domains to
11782 	 * give the idle CPUs a chance to load balance. Else we may
11783 	 * load balance only within the local sched_domain hierarchy
11784 	 * and abort nohz_idle_balance altogether if we pull some load.
11785 	 */
11786 	if (nohz_idle_balance(this_rq, idle))
11787 		return;
11788 
11789 	/* normal load balance */
11790 	update_blocked_averages(this_rq->cpu);
11791 	rebalance_domains(this_rq, idle);
11792 }
11793 
11794 /*
11795  * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
11796  */
11797 void trigger_load_balance(struct rq *rq)
11798 {
11799 	/*
11800 	 * Don't need to rebalance while attached to NULL domain or
11801 	 * runqueue CPU is not active
11802 	 */
11803 	if (unlikely(on_null_domain(rq) || !cpu_active(cpu_of(rq))))
11804 		return;
11805 
11806 	if (time_after_eq(jiffies, rq->next_balance))
11807 		raise_softirq(SCHED_SOFTIRQ);
11808 
11809 	nohz_balancer_kick(rq);
11810 }
11811 
11812 static void rq_online_fair(struct rq *rq)
11813 {
11814 	update_sysctl();
11815 
11816 	update_runtime_enabled(rq);
11817 }
11818 
11819 static void rq_offline_fair(struct rq *rq)
11820 {
11821 	update_sysctl();
11822 
11823 	/* Ensure any throttled groups are reachable by pick_next_task */
11824 	unthrottle_offline_cfs_rqs(rq);
11825 }
11826 
11827 #endif /* CONFIG_SMP */
11828 
11829 #ifdef CONFIG_SCHED_CORE
11830 static inline bool
11831 __entity_slice_used(struct sched_entity *se, int min_nr_tasks)
11832 {
11833 	u64 slice = sched_slice(cfs_rq_of(se), se);
11834 	u64 rtime = se->sum_exec_runtime - se->prev_sum_exec_runtime;
11835 
11836 	return (rtime * min_nr_tasks > slice);
11837 }
11838 
11839 #define MIN_NR_TASKS_DURING_FORCEIDLE	2
11840 static inline void task_tick_core(struct rq *rq, struct task_struct *curr)
11841 {
11842 	if (!sched_core_enabled(rq))
11843 		return;
11844 
11845 	/*
11846 	 * If runqueue has only one task which used up its slice and
11847 	 * if the sibling is forced idle, then trigger schedule to
11848 	 * give forced idle task a chance.
11849 	 *
11850 	 * sched_slice() considers only this active rq and it gets the
11851 	 * whole slice. But during force idle, we have siblings acting
11852 	 * like a single runqueue and hence we need to consider runnable
11853 	 * tasks on this CPU and the forced idle CPU. Ideally, we should
11854 	 * go through the forced idle rq, but that would be a perf hit.
11855 	 * We can assume that the forced idle CPU has at least
11856 	 * MIN_NR_TASKS_DURING_FORCEIDLE - 1 tasks and use that to check
11857 	 * if we need to give up the CPU.
11858 	 */
11859 	if (rq->core->core_forceidle_count && rq->cfs.nr_running == 1 &&
11860 	    __entity_slice_used(&curr->se, MIN_NR_TASKS_DURING_FORCEIDLE))
11861 		resched_curr(rq);
11862 }
11863 
11864 /*
11865  * se_fi_update - Update the cfs_rq->min_vruntime_fi in a CFS hierarchy if needed.
11866  */
11867 static void se_fi_update(const struct sched_entity *se, unsigned int fi_seq,
11868 			 bool forceidle)
11869 {
11870 	for_each_sched_entity(se) {
11871 		struct cfs_rq *cfs_rq = cfs_rq_of(se);
11872 
11873 		if (forceidle) {
11874 			if (cfs_rq->forceidle_seq == fi_seq)
11875 				break;
11876 			cfs_rq->forceidle_seq = fi_seq;
11877 		}
11878 
11879 		cfs_rq->min_vruntime_fi = cfs_rq->min_vruntime;
11880 	}
11881 }
11882 
11883 void task_vruntime_update(struct rq *rq, struct task_struct *p, bool in_fi)
11884 {
11885 	struct sched_entity *se = &p->se;
11886 
11887 	if (p->sched_class != &fair_sched_class)
11888 		return;
11889 
11890 	se_fi_update(se, rq->core->core_forceidle_seq, in_fi);
11891 }
11892 
11893 bool cfs_prio_less(const struct task_struct *a, const struct task_struct *b,
11894 			bool in_fi)
11895 {
11896 	struct rq *rq = task_rq(a);
11897 	const struct sched_entity *sea = &a->se;
11898 	const struct sched_entity *seb = &b->se;
11899 	struct cfs_rq *cfs_rqa;
11900 	struct cfs_rq *cfs_rqb;
11901 	s64 delta;
11902 
11903 	SCHED_WARN_ON(task_rq(b)->core != rq->core);
11904 
11905 #ifdef CONFIG_FAIR_GROUP_SCHED
11906 	/*
11907 	 * Find an se in the hierarchy for tasks a and b, such that the se's
11908 	 * are immediate siblings.
11909 	 */
11910 	while (sea->cfs_rq->tg != seb->cfs_rq->tg) {
11911 		int sea_depth = sea->depth;
11912 		int seb_depth = seb->depth;
11913 
11914 		if (sea_depth >= seb_depth)
11915 			sea = parent_entity(sea);
11916 		if (sea_depth <= seb_depth)
11917 			seb = parent_entity(seb);
11918 	}
11919 
11920 	se_fi_update(sea, rq->core->core_forceidle_seq, in_fi);
11921 	se_fi_update(seb, rq->core->core_forceidle_seq, in_fi);
11922 
11923 	cfs_rqa = sea->cfs_rq;
11924 	cfs_rqb = seb->cfs_rq;
11925 #else
11926 	cfs_rqa = &task_rq(a)->cfs;
11927 	cfs_rqb = &task_rq(b)->cfs;
11928 #endif
11929 
11930 	/*
11931 	 * Find delta after normalizing se's vruntime with its cfs_rq's
11932 	 * min_vruntime_fi, which would have been updated in prior calls
11933 	 * to se_fi_update().
11934 	 */
11935 	delta = (s64)(sea->vruntime - seb->vruntime) +
11936 		(s64)(cfs_rqb->min_vruntime_fi - cfs_rqa->min_vruntime_fi);
11937 
11938 	return delta > 0;
11939 }
11940 
11941 static int task_is_throttled_fair(struct task_struct *p, int cpu)
11942 {
11943 	struct cfs_rq *cfs_rq;
11944 
11945 #ifdef CONFIG_FAIR_GROUP_SCHED
11946 	cfs_rq = task_group(p)->cfs_rq[cpu];
11947 #else
11948 	cfs_rq = &cpu_rq(cpu)->cfs;
11949 #endif
11950 	return throttled_hierarchy(cfs_rq);
11951 }
11952 #else
11953 static inline void task_tick_core(struct rq *rq, struct task_struct *curr) {}
11954 #endif
11955 
11956 /*
11957  * scheduler tick hitting a task of our scheduling class.
11958  *
11959  * NOTE: This function can be called remotely by the tick offload that
11960  * goes along full dynticks. Therefore no local assumption can be made
11961  * and everything must be accessed through the @rq and @curr passed in
11962  * parameters.
11963  */
11964 static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
11965 {
11966 	struct cfs_rq *cfs_rq;
11967 	struct sched_entity *se = &curr->se;
11968 
11969 	for_each_sched_entity(se) {
11970 		cfs_rq = cfs_rq_of(se);
11971 		entity_tick(cfs_rq, se, queued);
11972 	}
11973 
11974 	if (static_branch_unlikely(&sched_numa_balancing))
11975 		task_tick_numa(rq, curr);
11976 
11977 	update_misfit_status(curr, rq);
11978 	update_overutilized_status(task_rq(curr));
11979 
11980 	task_tick_core(rq, curr);
11981 }
11982 
11983 /*
11984  * called on fork with the child task as argument from the parent's context
11985  *  - child not yet on the tasklist
11986  *  - preemption disabled
11987  */
11988 static void task_fork_fair(struct task_struct *p)
11989 {
11990 	struct cfs_rq *cfs_rq;
11991 	struct sched_entity *se = &p->se, *curr;
11992 	struct rq *rq = this_rq();
11993 	struct rq_flags rf;
11994 
11995 	rq_lock(rq, &rf);
11996 	update_rq_clock(rq);
11997 
11998 	cfs_rq = task_cfs_rq(current);
11999 	curr = cfs_rq->curr;
12000 	if (curr) {
12001 		update_curr(cfs_rq);
12002 		se->vruntime = curr->vruntime;
12003 	}
12004 	place_entity(cfs_rq, se, 1);
12005 
12006 	if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
12007 		/*
12008 		 * Upon rescheduling, sched_class::put_prev_task() will place
12009 		 * 'current' within the tree based on its new key value.
12010 		 */
12011 		swap(curr->vruntime, se->vruntime);
12012 		resched_curr(rq);
12013 	}
12014 
12015 	se->vruntime -= cfs_rq->min_vruntime;
12016 	rq_unlock(rq, &rf);
12017 }
12018 
12019 /*
12020  * Priority of the task has changed. Check to see if we preempt
12021  * the current task.
12022  */
12023 static void
12024 prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
12025 {
12026 	if (!task_on_rq_queued(p))
12027 		return;
12028 
12029 	if (rq->cfs.nr_running == 1)
12030 		return;
12031 
12032 	/*
12033 	 * Reschedule if we are currently running on this runqueue and
12034 	 * our priority decreased, or if we are not currently running on
12035 	 * this runqueue and our priority is higher than the current's
12036 	 */
12037 	if (task_current(rq, p)) {
12038 		if (p->prio > oldprio)
12039 			resched_curr(rq);
12040 	} else
12041 		check_preempt_curr(rq, p, 0);
12042 }
12043 
12044 static inline bool vruntime_normalized(struct task_struct *p)
12045 {
12046 	struct sched_entity *se = &p->se;
12047 
12048 	/*
12049 	 * In both the TASK_ON_RQ_QUEUED and TASK_ON_RQ_MIGRATING cases,
12050 	 * the dequeue_entity(.flags=0) will already have normalized the
12051 	 * vruntime.
12052 	 */
12053 	if (p->on_rq)
12054 		return true;
12055 
12056 	/*
12057 	 * When !on_rq, vruntime of the task has usually NOT been normalized.
12058 	 * But there are some cases where it has already been normalized:
12059 	 *
12060 	 * - A forked child which is waiting for being woken up by
12061 	 *   wake_up_new_task().
12062 	 * - A task which has been woken up by try_to_wake_up() and
12063 	 *   waiting for actually being woken up by sched_ttwu_pending().
12064 	 */
12065 	if (!se->sum_exec_runtime ||
12066 	    (READ_ONCE(p->__state) == TASK_WAKING && p->sched_remote_wakeup))
12067 		return true;
12068 
12069 	return false;
12070 }
12071 
12072 #ifdef CONFIG_FAIR_GROUP_SCHED
12073 /*
12074  * Propagate the changes of the sched_entity across the tg tree to make it
12075  * visible to the root
12076  */
12077 static void propagate_entity_cfs_rq(struct sched_entity *se)
12078 {
12079 	struct cfs_rq *cfs_rq = cfs_rq_of(se);
12080 
12081 	if (cfs_rq_throttled(cfs_rq))
12082 		return;
12083 
12084 	if (!throttled_hierarchy(cfs_rq))
12085 		list_add_leaf_cfs_rq(cfs_rq);
12086 
12087 	/* Start to propagate at parent */
12088 	se = se->parent;
12089 
12090 	for_each_sched_entity(se) {
12091 		cfs_rq = cfs_rq_of(se);
12092 
12093 		update_load_avg(cfs_rq, se, UPDATE_TG);
12094 
12095 		if (cfs_rq_throttled(cfs_rq))
12096 			break;
12097 
12098 		if (!throttled_hierarchy(cfs_rq))
12099 			list_add_leaf_cfs_rq(cfs_rq);
12100 	}
12101 }
12102 #else
12103 static void propagate_entity_cfs_rq(struct sched_entity *se) { }
12104 #endif
12105 
12106 static void detach_entity_cfs_rq(struct sched_entity *se)
12107 {
12108 	struct cfs_rq *cfs_rq = cfs_rq_of(se);
12109 
12110 #ifdef CONFIG_SMP
12111 	/*
12112 	 * In case the task sched_avg hasn't been attached:
12113 	 * - A forked task which hasn't been woken up by wake_up_new_task().
12114 	 * - A task which has been woken up by try_to_wake_up() but is
12115 	 *   waiting for actually being woken up by sched_ttwu_pending().
12116 	 */
12117 	if (!se->avg.last_update_time)
12118 		return;
12119 #endif
12120 
12121 	/* Catch up with the cfs_rq and remove our load when we leave */
12122 	update_load_avg(cfs_rq, se, 0);
12123 	detach_entity_load_avg(cfs_rq, se);
12124 	update_tg_load_avg(cfs_rq);
12125 	propagate_entity_cfs_rq(se);
12126 }
12127 
12128 static void attach_entity_cfs_rq(struct sched_entity *se)
12129 {
12130 	struct cfs_rq *cfs_rq = cfs_rq_of(se);
12131 
12132 	/* Synchronize entity with its cfs_rq */
12133 	update_load_avg(cfs_rq, se, sched_feat(ATTACH_AGE_LOAD) ? 0 : SKIP_AGE_LOAD);
12134 	attach_entity_load_avg(cfs_rq, se);
12135 	update_tg_load_avg(cfs_rq);
12136 	propagate_entity_cfs_rq(se);
12137 }
12138 
12139 static void detach_task_cfs_rq(struct task_struct *p)
12140 {
12141 	struct sched_entity *se = &p->se;
12142 	struct cfs_rq *cfs_rq = cfs_rq_of(se);
12143 
12144 	if (!vruntime_normalized(p)) {
12145 		/*
12146 		 * Fix up our vruntime so that the current sleep doesn't
12147 		 * cause 'unlimited' sleep bonus.
12148 		 */
12149 		place_entity(cfs_rq, se, 0);
12150 		se->vruntime -= cfs_rq->min_vruntime;
12151 	}
12152 
12153 	detach_entity_cfs_rq(se);
12154 }
12155 
12156 static void attach_task_cfs_rq(struct task_struct *p)
12157 {
12158 	struct sched_entity *se = &p->se;
12159 	struct cfs_rq *cfs_rq = cfs_rq_of(se);
12160 
12161 	attach_entity_cfs_rq(se);
12162 
12163 	if (!vruntime_normalized(p))
12164 		se->vruntime += cfs_rq->min_vruntime;
12165 }
12166 
12167 static void switched_from_fair(struct rq *rq, struct task_struct *p)
12168 {
12169 	detach_task_cfs_rq(p);
12170 }
12171 
12172 static void switched_to_fair(struct rq *rq, struct task_struct *p)
12173 {
12174 	attach_task_cfs_rq(p);
12175 
12176 	if (task_on_rq_queued(p)) {
12177 		/*
12178 		 * We were most likely switched from sched_rt, so
12179 		 * kick off the schedule if running, otherwise just see
12180 		 * if we can still preempt the current task.
12181 		 */
12182 		if (task_current(rq, p))
12183 			resched_curr(rq);
12184 		else
12185 			check_preempt_curr(rq, p, 0);
12186 	}
12187 }
12188 
12189 /* Account for a task changing its policy or group.
12190  *
12191  * This routine is mostly called to set cfs_rq->curr field when a task
12192  * migrates between groups/classes.
12193  */
12194 static void set_next_task_fair(struct rq *rq, struct task_struct *p, bool first)
12195 {
12196 	struct sched_entity *se = &p->se;
12197 
12198 #ifdef CONFIG_SMP
12199 	if (task_on_rq_queued(p)) {
12200 		/*
12201 		 * Move the next running task to the front of the list, so our
12202 		 * cfs_tasks list becomes MRU one.
12203 		 */
12204 		list_move(&se->group_node, &rq->cfs_tasks);
12205 	}
12206 #endif
12207 
12208 	for_each_sched_entity(se) {
12209 		struct cfs_rq *cfs_rq = cfs_rq_of(se);
12210 
12211 		set_next_entity(cfs_rq, se);
12212 		/* ensure bandwidth has been allocated on our new cfs_rq */
12213 		account_cfs_rq_runtime(cfs_rq, 0);
12214 	}
12215 }
12216 
12217 void init_cfs_rq(struct cfs_rq *cfs_rq)
12218 {
12219 	cfs_rq->tasks_timeline = RB_ROOT_CACHED;
12220 	u64_u32_store(cfs_rq->min_vruntime, (u64)(-(1LL << 20)));
12221 #ifdef CONFIG_SMP
12222 	raw_spin_lock_init(&cfs_rq->removed.lock);
12223 #endif
12224 }
12225 
12226 #ifdef CONFIG_FAIR_GROUP_SCHED
12227 static void task_change_group_fair(struct task_struct *p)
12228 {
12229 	/*
12230 	 * We couldn't detach or attach a forked task which
12231 	 * hasn't been woken up by wake_up_new_task().
12232 	 */
12233 	if (READ_ONCE(p->__state) == TASK_NEW)
12234 		return;
12235 
12236 	detach_task_cfs_rq(p);
12237 
12238 #ifdef CONFIG_SMP
12239 	/* Tell se's cfs_rq has been changed -- migrated */
12240 	p->se.avg.last_update_time = 0;
12241 #endif
12242 	set_task_rq(p, task_cpu(p));
12243 	attach_task_cfs_rq(p);
12244 }
12245 
12246 void free_fair_sched_group(struct task_group *tg)
12247 {
12248 	int i;
12249 
12250 	for_each_possible_cpu(i) {
12251 		if (tg->cfs_rq)
12252 			kfree(tg->cfs_rq[i]);
12253 		if (tg->se)
12254 			kfree(tg->se[i]);
12255 	}
12256 
12257 	kfree(tg->cfs_rq);
12258 	kfree(tg->se);
12259 }
12260 
12261 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
12262 {
12263 	struct sched_entity *se;
12264 	struct cfs_rq *cfs_rq;
12265 	int i;
12266 
12267 	tg->cfs_rq = kcalloc(nr_cpu_ids, sizeof(cfs_rq), GFP_KERNEL);
12268 	if (!tg->cfs_rq)
12269 		goto err;
12270 	tg->se = kcalloc(nr_cpu_ids, sizeof(se), GFP_KERNEL);
12271 	if (!tg->se)
12272 		goto err;
12273 
12274 	tg->shares = NICE_0_LOAD;
12275 
12276 	init_cfs_bandwidth(tg_cfs_bandwidth(tg));
12277 
12278 	for_each_possible_cpu(i) {
12279 		cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
12280 				      GFP_KERNEL, cpu_to_node(i));
12281 		if (!cfs_rq)
12282 			goto err;
12283 
12284 		se = kzalloc_node(sizeof(struct sched_entity_stats),
12285 				  GFP_KERNEL, cpu_to_node(i));
12286 		if (!se)
12287 			goto err_free_rq;
12288 
12289 		init_cfs_rq(cfs_rq);
12290 		init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
12291 		init_entity_runnable_average(se);
12292 	}
12293 
12294 	return 1;
12295 
12296 err_free_rq:
12297 	kfree(cfs_rq);
12298 err:
12299 	return 0;
12300 }
12301 
12302 void online_fair_sched_group(struct task_group *tg)
12303 {
12304 	struct sched_entity *se;
12305 	struct rq_flags rf;
12306 	struct rq *rq;
12307 	int i;
12308 
12309 	for_each_possible_cpu(i) {
12310 		rq = cpu_rq(i);
12311 		se = tg->se[i];
12312 		rq_lock_irq(rq, &rf);
12313 		update_rq_clock(rq);
12314 		attach_entity_cfs_rq(se);
12315 		sync_throttle(tg, i);
12316 		rq_unlock_irq(rq, &rf);
12317 	}
12318 }
12319 
12320 void unregister_fair_sched_group(struct task_group *tg)
12321 {
12322 	unsigned long flags;
12323 	struct rq *rq;
12324 	int cpu;
12325 
12326 	destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
12327 
12328 	for_each_possible_cpu(cpu) {
12329 		if (tg->se[cpu])
12330 			remove_entity_load_avg(tg->se[cpu]);
12331 
12332 		/*
12333 		 * Only empty task groups can be destroyed; so we can speculatively
12334 		 * check on_list without danger of it being re-added.
12335 		 */
12336 		if (!tg->cfs_rq[cpu]->on_list)
12337 			continue;
12338 
12339 		rq = cpu_rq(cpu);
12340 
12341 		raw_spin_rq_lock_irqsave(rq, flags);
12342 		list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
12343 		raw_spin_rq_unlock_irqrestore(rq, flags);
12344 	}
12345 }
12346 
12347 void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
12348 			struct sched_entity *se, int cpu,
12349 			struct sched_entity *parent)
12350 {
12351 	struct rq *rq = cpu_rq(cpu);
12352 
12353 	cfs_rq->tg = tg;
12354 	cfs_rq->rq = rq;
12355 	init_cfs_rq_runtime(cfs_rq);
12356 
12357 	tg->cfs_rq[cpu] = cfs_rq;
12358 	tg->se[cpu] = se;
12359 
12360 	/* se could be NULL for root_task_group */
12361 	if (!se)
12362 		return;
12363 
12364 	if (!parent) {
12365 		se->cfs_rq = &rq->cfs;
12366 		se->depth = 0;
12367 	} else {
12368 		se->cfs_rq = parent->my_q;
12369 		se->depth = parent->depth + 1;
12370 	}
12371 
12372 	se->my_q = cfs_rq;
12373 	/* guarantee group entities always have weight */
12374 	update_load_set(&se->load, NICE_0_LOAD);
12375 	se->parent = parent;
12376 }
12377 
12378 static DEFINE_MUTEX(shares_mutex);
12379 
12380 static int __sched_group_set_shares(struct task_group *tg, unsigned long shares)
12381 {
12382 	int i;
12383 
12384 	lockdep_assert_held(&shares_mutex);
12385 
12386 	/*
12387 	 * We can't change the weight of the root cgroup.
12388 	 */
12389 	if (!tg->se[0])
12390 		return -EINVAL;
12391 
12392 	shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
12393 
12394 	if (tg->shares == shares)
12395 		return 0;
12396 
12397 	tg->shares = shares;
12398 	for_each_possible_cpu(i) {
12399 		struct rq *rq = cpu_rq(i);
12400 		struct sched_entity *se = tg->se[i];
12401 		struct rq_flags rf;
12402 
12403 		/* Propagate contribution to hierarchy */
12404 		rq_lock_irqsave(rq, &rf);
12405 		update_rq_clock(rq);
12406 		for_each_sched_entity(se) {
12407 			update_load_avg(cfs_rq_of(se), se, UPDATE_TG);
12408 			update_cfs_group(se);
12409 		}
12410 		rq_unlock_irqrestore(rq, &rf);
12411 	}
12412 
12413 	return 0;
12414 }
12415 
12416 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
12417 {
12418 	int ret;
12419 
12420 	mutex_lock(&shares_mutex);
12421 	if (tg_is_idle(tg))
12422 		ret = -EINVAL;
12423 	else
12424 		ret = __sched_group_set_shares(tg, shares);
12425 	mutex_unlock(&shares_mutex);
12426 
12427 	return ret;
12428 }
12429 
12430 int sched_group_set_idle(struct task_group *tg, long idle)
12431 {
12432 	int i;
12433 
12434 	if (tg == &root_task_group)
12435 		return -EINVAL;
12436 
12437 	if (idle < 0 || idle > 1)
12438 		return -EINVAL;
12439 
12440 	mutex_lock(&shares_mutex);
12441 
12442 	if (tg->idle == idle) {
12443 		mutex_unlock(&shares_mutex);
12444 		return 0;
12445 	}
12446 
12447 	tg->idle = idle;
12448 
12449 	for_each_possible_cpu(i) {
12450 		struct rq *rq = cpu_rq(i);
12451 		struct sched_entity *se = tg->se[i];
12452 		struct cfs_rq *parent_cfs_rq, *grp_cfs_rq = tg->cfs_rq[i];
12453 		bool was_idle = cfs_rq_is_idle(grp_cfs_rq);
12454 		long idle_task_delta;
12455 		struct rq_flags rf;
12456 
12457 		rq_lock_irqsave(rq, &rf);
12458 
12459 		grp_cfs_rq->idle = idle;
12460 		if (WARN_ON_ONCE(was_idle == cfs_rq_is_idle(grp_cfs_rq)))
12461 			goto next_cpu;
12462 
12463 		if (se->on_rq) {
12464 			parent_cfs_rq = cfs_rq_of(se);
12465 			if (cfs_rq_is_idle(grp_cfs_rq))
12466 				parent_cfs_rq->idle_nr_running++;
12467 			else
12468 				parent_cfs_rq->idle_nr_running--;
12469 		}
12470 
12471 		idle_task_delta = grp_cfs_rq->h_nr_running -
12472 				  grp_cfs_rq->idle_h_nr_running;
12473 		if (!cfs_rq_is_idle(grp_cfs_rq))
12474 			idle_task_delta *= -1;
12475 
12476 		for_each_sched_entity(se) {
12477 			struct cfs_rq *cfs_rq = cfs_rq_of(se);
12478 
12479 			if (!se->on_rq)
12480 				break;
12481 
12482 			cfs_rq->idle_h_nr_running += idle_task_delta;
12483 
12484 			/* Already accounted at parent level and above. */
12485 			if (cfs_rq_is_idle(cfs_rq))
12486 				break;
12487 		}
12488 
12489 next_cpu:
12490 		rq_unlock_irqrestore(rq, &rf);
12491 	}
12492 
12493 	/* Idle groups have minimum weight. */
12494 	if (tg_is_idle(tg))
12495 		__sched_group_set_shares(tg, scale_load(WEIGHT_IDLEPRIO));
12496 	else
12497 		__sched_group_set_shares(tg, NICE_0_LOAD);
12498 
12499 	mutex_unlock(&shares_mutex);
12500 	return 0;
12501 }
12502 
12503 #else /* CONFIG_FAIR_GROUP_SCHED */
12504 
12505 void free_fair_sched_group(struct task_group *tg) { }
12506 
12507 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
12508 {
12509 	return 1;
12510 }
12511 
12512 void online_fair_sched_group(struct task_group *tg) { }
12513 
12514 void unregister_fair_sched_group(struct task_group *tg) { }
12515 
12516 #endif /* CONFIG_FAIR_GROUP_SCHED */
12517 
12518 
12519 static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
12520 {
12521 	struct sched_entity *se = &task->se;
12522 	unsigned int rr_interval = 0;
12523 
12524 	/*
12525 	 * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
12526 	 * idle runqueue:
12527 	 */
12528 	if (rq->cfs.load.weight)
12529 		rr_interval = NS_TO_JIFFIES(sched_slice(cfs_rq_of(se), se));
12530 
12531 	return rr_interval;
12532 }
12533 
12534 /*
12535  * All the scheduling class methods:
12536  */
12537 DEFINE_SCHED_CLASS(fair) = {
12538 
12539 	.enqueue_task		= enqueue_task_fair,
12540 	.dequeue_task		= dequeue_task_fair,
12541 	.yield_task		= yield_task_fair,
12542 	.yield_to_task		= yield_to_task_fair,
12543 
12544 	.check_preempt_curr	= check_preempt_wakeup,
12545 
12546 	.pick_next_task		= __pick_next_task_fair,
12547 	.put_prev_task		= put_prev_task_fair,
12548 	.set_next_task          = set_next_task_fair,
12549 
12550 #ifdef CONFIG_SMP
12551 	.balance		= balance_fair,
12552 	.pick_task		= pick_task_fair,
12553 	.select_task_rq		= select_task_rq_fair,
12554 	.migrate_task_rq	= migrate_task_rq_fair,
12555 
12556 	.rq_online		= rq_online_fair,
12557 	.rq_offline		= rq_offline_fair,
12558 
12559 	.task_dead		= task_dead_fair,
12560 	.set_cpus_allowed	= set_cpus_allowed_common,
12561 #endif
12562 
12563 	.task_tick		= task_tick_fair,
12564 	.task_fork		= task_fork_fair,
12565 
12566 	.prio_changed		= prio_changed_fair,
12567 	.switched_from		= switched_from_fair,
12568 	.switched_to		= switched_to_fair,
12569 
12570 	.get_rr_interval	= get_rr_interval_fair,
12571 
12572 	.update_curr		= update_curr_fair,
12573 
12574 #ifdef CONFIG_FAIR_GROUP_SCHED
12575 	.task_change_group	= task_change_group_fair,
12576 #endif
12577 
12578 #ifdef CONFIG_SCHED_CORE
12579 	.task_is_throttled	= task_is_throttled_fair,
12580 #endif
12581 
12582 #ifdef CONFIG_UCLAMP_TASK
12583 	.uclamp_enabled		= 1,
12584 #endif
12585 };
12586 
12587 #ifdef CONFIG_SCHED_DEBUG
12588 void print_cfs_stats(struct seq_file *m, int cpu)
12589 {
12590 	struct cfs_rq *cfs_rq, *pos;
12591 
12592 	rcu_read_lock();
12593 	for_each_leaf_cfs_rq_safe(cpu_rq(cpu), cfs_rq, pos)
12594 		print_cfs_rq(m, cpu, cfs_rq);
12595 	rcu_read_unlock();
12596 }
12597 
12598 #ifdef CONFIG_NUMA_BALANCING
12599 void show_numa_stats(struct task_struct *p, struct seq_file *m)
12600 {
12601 	int node;
12602 	unsigned long tsf = 0, tpf = 0, gsf = 0, gpf = 0;
12603 	struct numa_group *ng;
12604 
12605 	rcu_read_lock();
12606 	ng = rcu_dereference(p->numa_group);
12607 	for_each_online_node(node) {
12608 		if (p->numa_faults) {
12609 			tsf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 0)];
12610 			tpf = p->numa_faults[task_faults_idx(NUMA_MEM, node, 1)];
12611 		}
12612 		if (ng) {
12613 			gsf = ng->faults[task_faults_idx(NUMA_MEM, node, 0)],
12614 			gpf = ng->faults[task_faults_idx(NUMA_MEM, node, 1)];
12615 		}
12616 		print_numa_stats(m, node, tsf, tpf, gsf, gpf);
12617 	}
12618 	rcu_read_unlock();
12619 }
12620 #endif /* CONFIG_NUMA_BALANCING */
12621 #endif /* CONFIG_SCHED_DEBUG */
12622 
12623 __init void init_sched_fair_class(void)
12624 {
12625 #ifdef CONFIG_SMP
12626 	int i;
12627 
12628 	for_each_possible_cpu(i) {
12629 		zalloc_cpumask_var_node(&per_cpu(load_balance_mask, i), GFP_KERNEL, cpu_to_node(i));
12630 		zalloc_cpumask_var_node(&per_cpu(select_rq_mask,    i), GFP_KERNEL, cpu_to_node(i));
12631 
12632 #ifdef CONFIG_CFS_BANDWIDTH
12633 		INIT_CSD(&cpu_rq(i)->cfsb_csd, __cfsb_csd_unthrottle, cpu_rq(i));
12634 		INIT_LIST_HEAD(&cpu_rq(i)->cfsb_csd_list);
12635 #endif
12636 	}
12637 
12638 	open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
12639 
12640 #ifdef CONFIG_NO_HZ_COMMON
12641 	nohz.next_balance = jiffies;
12642 	nohz.next_blocked = jiffies;
12643 	zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
12644 #endif
12645 #endif /* SMP */
12646 
12647 }
12648